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"
30 #include "dualinput.h"
37 DITHERING_FLOYD_STEINBERG,
43 enum color_search_method {
44 COLOR_SEARCH_NNS_ITERATIVE,
45 COLOR_SEARCH_NNS_RECURSIVE,
46 COLOR_SEARCH_BRUTEFORCE,
60 int left_id, right_id;
64 #define CACHE_SIZE (1<<(3*NBITS))
72 struct cached_color *entries;
76 struct PaletteUseContext;
78 typedef int (*set_frame_func)(struct PaletteUseContext *s, AVFrame *out, AVFrame *in,
79 int x_start, int y_start, int width, int height);
81 typedef struct PaletteUseContext {
83 FFDualInputContext dinput;
84 struct cache_node cache[CACHE_SIZE]; /* lookup cache */
85 struct color_node map[AVPALETTE_COUNT]; /* 3D-Tree (KD-Tree with K=3) for reverse colormap */
86 uint32_t palette[AVPALETTE_COUNT];
89 set_frame_func set_frame;
91 int ordered_dither[8*8];
98 int color_search_method;
100 uint64_t total_mean_err;
104 #define OFFSET(x) offsetof(PaletteUseContext, x)
105 #define FLAGS AV_OPT_FLAG_FILTERING_PARAM|AV_OPT_FLAG_VIDEO_PARAM
106 static const AVOption paletteuse_options[] = {
107 { "dither", "select dithering mode", OFFSET(dither), AV_OPT_TYPE_INT, {.i64=DITHERING_SIERRA2_4A}, 0, NB_DITHERING-1, FLAGS, "dithering_mode" },
108 { "bayer", "ordered 8x8 bayer dithering (deterministic)", 0, AV_OPT_TYPE_CONST, {.i64=DITHERING_BAYER}, INT_MIN, INT_MAX, FLAGS, "dithering_mode" },
109 { "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" },
110 { "floyd_steinberg", "Floyd and Steingberg dithering (error diffusion)", 0, AV_OPT_TYPE_CONST, {.i64=DITHERING_FLOYD_STEINBERG}, INT_MIN, INT_MAX, FLAGS, "dithering_mode" },
111 { "sierra2", "Frankie Sierra dithering v2 (error diffusion)", 0, AV_OPT_TYPE_CONST, {.i64=DITHERING_SIERRA2}, INT_MIN, INT_MAX, FLAGS, "dithering_mode" },
112 { "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" },
113 { "bayer_scale", "set scale for bayer dithering", OFFSET(bayer_scale), AV_OPT_TYPE_INT, {.i64=2}, 0, 5, FLAGS },
114 { "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" },
115 { "rectangle", "process smallest different rectangle", 0, AV_OPT_TYPE_CONST, {.i64=DIFF_MODE_RECTANGLE}, INT_MIN, INT_MAX, FLAGS, "diff_mode" },
117 /* following are the debug options, not part of the official API */
118 { "debug_kdtree", "save Graphviz graph of the kdtree in specified file", OFFSET(dot_filename), AV_OPT_TYPE_STRING, {.str=NULL}, CHAR_MIN, CHAR_MAX, FLAGS },
119 { "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" },
120 { "nns_iterative", "iterative search", 0, AV_OPT_TYPE_CONST, {.i64=COLOR_SEARCH_NNS_ITERATIVE}, INT_MIN, INT_MAX, FLAGS, "search" },
121 { "nns_recursive", "recursive search", 0, AV_OPT_TYPE_CONST, {.i64=COLOR_SEARCH_NNS_RECURSIVE}, INT_MIN, INT_MAX, FLAGS, "search" },
122 { "bruteforce", "brute-force into the palette", 0, AV_OPT_TYPE_CONST, {.i64=COLOR_SEARCH_BRUTEFORCE}, INT_MIN, INT_MAX, FLAGS, "search" },
123 { "mean_err", "compute and print mean error", OFFSET(calc_mean_err), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, FLAGS },
124 { "debug_accuracy", "test color search accuracy", OFFSET(debug_accuracy), AV_OPT_TYPE_BOOL, {.i64=0}, 0, 1, FLAGS },
128 AVFILTER_DEFINE_CLASS(paletteuse);
130 static int query_formats(AVFilterContext *ctx)
132 static const enum AVPixelFormat in_fmts[] = {AV_PIX_FMT_RGB32, AV_PIX_FMT_NONE};
133 static const enum AVPixelFormat inpal_fmts[] = {AV_PIX_FMT_RGB32, AV_PIX_FMT_NONE};
134 static const enum AVPixelFormat out_fmts[] = {AV_PIX_FMT_PAL8, AV_PIX_FMT_NONE};
136 AVFilterFormats *in = ff_make_format_list(in_fmts);
137 AVFilterFormats *inpal = ff_make_format_list(inpal_fmts);
138 AVFilterFormats *out = ff_make_format_list(out_fmts);
139 if (!in || !inpal || !out) {
143 return AVERROR(ENOMEM);
145 if ((ret = ff_formats_ref(in , &ctx->inputs[0]->out_formats)) < 0 ||
146 (ret = ff_formats_ref(inpal, &ctx->inputs[1]->out_formats)) < 0 ||
147 (ret = ff_formats_ref(out , &ctx->outputs[0]->in_formats)) < 0)
152 static av_always_inline int dither_color(uint32_t px, int er, int eg, int eb, int scale, int shift)
154 return av_clip_uint8((px >> 16 & 0xff) + ((er * scale) / (1<<shift))) << 16
155 | av_clip_uint8((px >> 8 & 0xff) + ((eg * scale) / (1<<shift))) << 8
156 | av_clip_uint8((px & 0xff) + ((eb * scale) / (1<<shift)));
159 static av_always_inline int diff(const uint8_t *c1, const uint8_t *c2)
161 // XXX: try L*a*b with CIE76 (dL*dL + da*da + db*db)
162 const int dr = c1[0] - c2[0];
163 const int dg = c1[1] - c2[1];
164 const int db = c1[2] - c2[2];
165 return dr*dr + dg*dg + db*db;
168 static av_always_inline uint8_t colormap_nearest_bruteforce(const uint32_t *palette, const uint8_t *rgb)
170 int i, pal_id = -1, min_dist = INT_MAX;
172 for (i = 0; i < AVPALETTE_COUNT; i++) {
173 const uint32_t c = palette[i];
175 if ((c & 0xff000000) == 0xff000000) { // ignore transparent entry
176 const uint8_t palrgb[] = {
177 palette[i]>>16 & 0xff,
178 palette[i]>> 8 & 0xff,
181 const int d = diff(palrgb, rgb);
191 /* Recursive form, simpler but a bit slower. Kept for reference. */
192 struct nearest_color {
197 static void colormap_nearest_node(const struct color_node *map,
199 const uint8_t *target,
200 struct nearest_color *nearest)
202 const struct color_node *kd = map + node_pos;
203 const int s = kd->split;
204 int dx, nearer_kd_id, further_kd_id;
205 const uint8_t *current = kd->val;
206 const int current_to_target = diff(target, current);
208 if (current_to_target < nearest->dist_sqd) {
209 nearest->node_pos = node_pos;
210 nearest->dist_sqd = current_to_target;
213 if (kd->left_id != -1 || kd->right_id != -1) {
214 dx = target[s] - current[s];
216 if (dx <= 0) nearer_kd_id = kd->left_id, further_kd_id = kd->right_id;
217 else nearer_kd_id = kd->right_id, further_kd_id = kd->left_id;
219 if (nearer_kd_id != -1)
220 colormap_nearest_node(map, nearer_kd_id, target, nearest);
222 if (further_kd_id != -1 && dx*dx < nearest->dist_sqd)
223 colormap_nearest_node(map, further_kd_id, target, nearest);
227 static av_always_inline uint8_t colormap_nearest_recursive(const struct color_node *node, const uint8_t *rgb)
229 struct nearest_color res = {.dist_sqd = INT_MAX, .node_pos = -1};
230 colormap_nearest_node(node, 0, rgb, &res);
231 return node[res.node_pos].palette_id;
239 static av_always_inline uint8_t colormap_nearest_iterative(const struct color_node *root, const uint8_t *target)
241 int pos = 0, best_node_id = -1, best_dist = INT_MAX, cur_color_id = 0;
242 struct stack_node nodes[16];
243 struct stack_node *node = &nodes[0];
247 const struct color_node *kd = &root[cur_color_id];
248 const uint8_t *current = kd->val;
249 const int current_to_target = diff(target, current);
251 /* Compare current color node to the target and update our best node if
252 * it's actually better. */
253 if (current_to_target < best_dist) {
254 best_node_id = cur_color_id;
255 if (!current_to_target)
256 goto end; // exact match, we can return immediately
257 best_dist = current_to_target;
260 /* Check if it's not a leaf */
261 if (kd->left_id != -1 || kd->right_id != -1) {
262 const int split = kd->split;
263 const int dx = target[split] - current[split];
264 int nearer_kd_id, further_kd_id;
266 /* Define which side is the most interesting. */
267 if (dx <= 0) nearer_kd_id = kd->left_id, further_kd_id = kd->right_id;
268 else nearer_kd_id = kd->right_id, further_kd_id = kd->left_id;
270 if (nearer_kd_id != -1) {
271 if (further_kd_id != -1) {
272 /* Here, both paths are defined, so we push a state for
273 * when we are going back. */
274 node->color_id = further_kd_id;
279 /* We can now update current color with the most probable path
280 * (no need to create a state since there is nothing to save
282 cur_color_id = nearer_kd_id;
284 } else if (dx*dx < best_dist) {
285 /* The nearest path isn't available, so there is only one path
286 * possible and it's the least probable. We enter it only if the
287 * distance from the current point to the hyper rectangle is
288 * less than our best distance. */
289 cur_color_id = further_kd_id;
294 /* Unstack as much as we can, typically as long as the least probable
295 * branch aren't actually probable. */
300 } while (node->dx2 >= best_dist);
302 /* We got a node where the least probable branch might actually contain
303 * a relevant color. */
304 cur_color_id = node->color_id;
308 return root[best_node_id].palette_id;
311 #define COLORMAP_NEAREST(search, palette, root, target) \
312 search == COLOR_SEARCH_NNS_ITERATIVE ? colormap_nearest_iterative(root, target) : \
313 search == COLOR_SEARCH_NNS_RECURSIVE ? colormap_nearest_recursive(root, target) : \
314 colormap_nearest_bruteforce(palette, target)
317 * Check if the requested color is in the cache already. If not, find it in the
318 * color tree and cache it.
319 * Note: r, g, and b are the component of c but are passed as well to avoid
320 * recomputing them (they are generally computed by the caller for other uses).
322 static av_always_inline int color_get(struct cache_node *cache, uint32_t color,
323 uint8_t r, uint8_t g, uint8_t b,
324 const struct color_node *map,
325 const uint32_t *palette,
326 const enum color_search_method search_method)
329 const uint8_t rgb[] = {r, g, b};
330 const uint8_t rhash = r & ((1<<NBITS)-1);
331 const uint8_t ghash = g & ((1<<NBITS)-1);
332 const uint8_t bhash = b & ((1<<NBITS)-1);
333 const unsigned hash = rhash<<(NBITS*2) | ghash<<NBITS | bhash;
334 struct cache_node *node = &cache[hash];
335 struct cached_color *e;
337 for (i = 0; i < node->nb_entries; i++) {
338 e = &node->entries[i];
339 if (e->color == color)
343 e = av_dynarray2_add((void**)&node->entries, &node->nb_entries,
344 sizeof(*node->entries), NULL);
346 return AVERROR(ENOMEM);
348 e->pal_entry = COLORMAP_NEAREST(search_method, palette, map, rgb);
352 static av_always_inline int get_dst_color_err(struct cache_node *cache,
353 uint32_t c, const struct color_node *map,
354 const uint32_t *palette,
355 int *er, int *eg, int *eb,
356 const enum color_search_method search_method)
358 const uint8_t r = c >> 16 & 0xff;
359 const uint8_t g = c >> 8 & 0xff;
360 const uint8_t b = c & 0xff;
361 const int dstx = color_get(cache, c, r, g, b, map, palette, search_method);
362 const uint32_t dstc = palette[dstx];
363 *er = r - (dstc >> 16 & 0xff);
364 *eg = g - (dstc >> 8 & 0xff);
365 *eb = b - (dstc & 0xff);
369 static av_always_inline int set_frame(PaletteUseContext *s, AVFrame *out, AVFrame *in,
370 int x_start, int y_start, int w, int h,
371 enum dithering_mode dither,
372 const enum color_search_method search_method)
375 const struct color_node *map = s->map;
376 struct cache_node *cache = s->cache;
377 const uint32_t *palette = s->palette;
378 const int src_linesize = in ->linesize[0] >> 2;
379 const int dst_linesize = out->linesize[0];
380 uint32_t *src = ((uint32_t *)in ->data[0]) + y_start*src_linesize;
381 uint8_t *dst = out->data[0] + y_start*dst_linesize;
386 for (y = y_start; y < h; y++) {
387 for (x = x_start; x < w; x++) {
390 if (dither == DITHERING_BAYER) {
391 const int d = s->ordered_dither[(y & 7)<<3 | (x & 7)];
392 const uint8_t r8 = src[x] >> 16 & 0xff;
393 const uint8_t g8 = src[x] >> 8 & 0xff;
394 const uint8_t b8 = src[x] & 0xff;
395 const uint8_t r = av_clip_uint8(r8 + d);
396 const uint8_t g = av_clip_uint8(g8 + d);
397 const uint8_t b = av_clip_uint8(b8 + d);
398 const uint32_t c = r<<16 | g<<8 | b;
399 const int color = color_get(cache, c, r, g, b, map, palette, search_method);
405 } else if (dither == DITHERING_HECKBERT) {
406 const int right = x < w - 1, down = y < h - 1;
407 const int color = get_dst_color_err(cache, src[x], map, palette, &er, &eg, &eb, search_method);
413 if (right) src[ x + 1] = dither_color(src[ x + 1], er, eg, eb, 3, 3);
414 if ( down) src[src_linesize + x ] = dither_color(src[src_linesize + x ], er, eg, eb, 3, 3);
415 if (right && down) src[src_linesize + x + 1] = dither_color(src[src_linesize + x + 1], er, eg, eb, 2, 3);
417 } else if (dither == DITHERING_FLOYD_STEINBERG) {
418 const int right = x < w - 1, down = y < h - 1, left = x > x_start;
419 const int color = get_dst_color_err(cache, src[x], map, palette, &er, &eg, &eb, search_method);
425 if (right) src[ x + 1] = dither_color(src[ x + 1], er, eg, eb, 7, 4);
426 if (left && down) src[src_linesize + x - 1] = dither_color(src[src_linesize + x - 1], er, eg, eb, 3, 4);
427 if ( down) src[src_linesize + x ] = dither_color(src[src_linesize + x ], er, eg, eb, 5, 4);
428 if (right && down) src[src_linesize + x + 1] = dither_color(src[src_linesize + x + 1], er, eg, eb, 1, 4);
430 } else if (dither == DITHERING_SIERRA2) {
431 const int right = x < w - 1, down = y < h - 1, left = x > x_start;
432 const int right2 = x < w - 2, left2 = x > x_start + 1;
433 const int color = get_dst_color_err(cache, src[x], map, palette, &er, &eg, &eb, search_method);
439 if (right) src[ x + 1] = dither_color(src[ x + 1], er, eg, eb, 4, 4);
440 if (right2) src[ x + 2] = dither_color(src[ x + 2], er, eg, eb, 3, 4);
443 if (left2) src[ src_linesize + x - 2] = dither_color(src[ src_linesize + x - 2], er, eg, eb, 1, 4);
444 if (left) src[ src_linesize + x - 1] = dither_color(src[ src_linesize + x - 1], er, eg, eb, 2, 4);
445 src[ src_linesize + x ] = dither_color(src[ src_linesize + x ], er, eg, eb, 3, 4);
446 if (right) src[ src_linesize + x + 1] = dither_color(src[ src_linesize + x + 1], er, eg, eb, 2, 4);
447 if (right2) src[ src_linesize + x + 2] = dither_color(src[ src_linesize + x + 2], er, eg, eb, 1, 4);
450 } else if (dither == DITHERING_SIERRA2_4A) {
451 const int right = x < w - 1, down = y < h - 1, left = x > x_start;
452 const int color = get_dst_color_err(cache, src[x], map, palette, &er, &eg, &eb, search_method);
458 if (right) src[ x + 1] = dither_color(src[ x + 1], er, eg, eb, 2, 2);
459 if (left && down) src[src_linesize + x - 1] = dither_color(src[src_linesize + x - 1], er, eg, eb, 1, 2);
460 if ( down) src[src_linesize + x ] = dither_color(src[src_linesize + x ], er, eg, eb, 1, 2);
463 const uint8_t r = src[x] >> 16 & 0xff;
464 const uint8_t g = src[x] >> 8 & 0xff;
465 const uint8_t b = src[x] & 0xff;
466 const int color = color_get(cache, src[x] & 0xffffff, r, g, b, map, palette, search_method);
480 static void disp_node(AVBPrint *buf,
481 const struct color_node *map,
482 int parent_id, int node_id,
485 const struct color_node *node = &map[node_id];
486 const uint32_t fontcolor = node->val[0] > 0x50 &&
487 node->val[1] > 0x50 &&
488 node->val[2] > 0x50 ? 0 : 0xffffff;
489 av_bprintf(buf, "%*cnode%d ["
490 "label=\"%c%02X%c%02X%c%02X%c\" "
491 "fillcolor=\"#%02x%02x%02x\" "
492 "fontcolor=\"#%06X\"]\n",
493 depth*INDENT, ' ', node->palette_id,
494 "[ "[node->split], node->val[0],
495 "][ "[node->split], node->val[1],
496 " ]["[node->split], node->val[2],
498 node->val[0], node->val[1], node->val[2],
501 av_bprintf(buf, "%*cnode%d -> node%d\n", depth*INDENT, ' ',
502 map[parent_id].palette_id, node->palette_id);
503 if (node->left_id != -1) disp_node(buf, map, node_id, node->left_id, depth + 1);
504 if (node->right_id != -1) disp_node(buf, map, node_id, node->right_id, depth + 1);
507 // debug_kdtree=kdtree.dot -> dot -Tpng kdtree.dot > kdtree.png
508 static int disp_tree(const struct color_node *node, const char *fname)
511 FILE *f = av_fopen_utf8(fname, "w");
514 int ret = AVERROR(errno);
515 av_log(NULL, AV_LOG_ERROR, "Cannot open file '%s' for writing: %s\n",
516 fname, av_err2str(ret));
520 av_bprint_init(&buf, 0, AV_BPRINT_SIZE_UNLIMITED);
522 av_bprintf(&buf, "digraph {\n");
523 av_bprintf(&buf, " node [style=filled fontsize=10 shape=box]\n");
524 disp_node(&buf, node, -1, 0, 0);
525 av_bprintf(&buf, "}\n");
527 fwrite(buf.str, 1, buf.len, f);
529 av_bprint_finalize(&buf, NULL);
533 static int debug_accuracy(const struct color_node *node, const uint32_t *palette,
534 const enum color_search_method search_method)
536 int r, g, b, ret = 0;
538 for (r = 0; r < 256; r++) {
539 for (g = 0; g < 256; g++) {
540 for (b = 0; b < 256; b++) {
541 const uint8_t rgb[] = {r, g, b};
542 const int r1 = COLORMAP_NEAREST(search_method, palette, node, rgb);
543 const int r2 = colormap_nearest_bruteforce(palette, rgb);
545 const uint32_t c1 = palette[r1];
546 const uint32_t c2 = palette[r2];
547 const uint8_t palrgb1[] = { c1>>16 & 0xff, c1>> 8 & 0xff, c1 & 0xff };
548 const uint8_t palrgb2[] = { c2>>16 & 0xff, c2>> 8 & 0xff, c2 & 0xff };
549 const int d1 = diff(palrgb1, rgb);
550 const int d2 = diff(palrgb2, rgb);
552 av_log(NULL, AV_LOG_ERROR,
553 "/!\\ %02X%02X%02X: %d ! %d (%06X ! %06X) / dist: %d ! %d\n",
554 r, g, b, r1, r2, c1 & 0xffffff, c2 & 0xffffff, d1, d2);
574 typedef int (*cmp_func)(const void *, const void *);
576 #define DECLARE_CMP_FUNC(name, pos) \
577 static int cmp_##name(const void *pa, const void *pb) \
579 const struct color *a = pa; \
580 const struct color *b = pb; \
581 return (a->value >> (8 * (2 - (pos))) & 0xff) \
582 - (b->value >> (8 * (2 - (pos))) & 0xff); \
585 DECLARE_CMP_FUNC(r, 0)
586 DECLARE_CMP_FUNC(g, 1)
587 DECLARE_CMP_FUNC(b, 2)
589 static const cmp_func cmp_funcs[] = {cmp_r, cmp_g, cmp_b};
591 static int get_next_color(const uint8_t *color_used, const uint32_t *palette,
592 int *component, const struct color_rect *box)
596 unsigned nb_color = 0;
597 struct color_rect ranges;
598 struct color tmp_pal[256];
601 ranges.min[0] = ranges.min[1] = ranges.min[2] = 0xff;
602 ranges.max[0] = ranges.max[1] = ranges.max[2] = 0x00;
604 for (i = 0; i < AVPALETTE_COUNT; i++) {
605 const uint32_t c = palette[i];
606 const uint8_t r = c >> 16 & 0xff;
607 const uint8_t g = c >> 8 & 0xff;
608 const uint8_t b = c & 0xff;
611 r < box->min[0] || g < box->min[1] || b < box->min[2] ||
612 r > box->max[0] || g > box->max[1] || b > box->max[2])
615 if (r < ranges.min[0]) ranges.min[0] = r;
616 if (g < ranges.min[1]) ranges.min[1] = g;
617 if (b < ranges.min[2]) ranges.min[2] = b;
619 if (r > ranges.max[0]) ranges.max[0] = r;
620 if (g > ranges.max[1]) ranges.max[1] = g;
621 if (b > ranges.max[2]) ranges.max[2] = b;
623 tmp_pal[nb_color].value = c;
624 tmp_pal[nb_color].pal_id = i;
632 /* define longest axis that will be the split component */
633 wr = ranges.max[0] - ranges.min[0];
634 wg = ranges.max[1] - ranges.min[1];
635 wb = ranges.max[2] - ranges.min[2];
636 if (wr >= wg && wr >= wb) longest = 0;
637 if (wg >= wr && wg >= wb) longest = 1;
638 if (wb >= wr && wb >= wg) longest = 2;
639 cmpf = cmp_funcs[longest];
640 *component = longest;
642 /* sort along this axis to get median */
643 AV_QSORT(tmp_pal, nb_color, struct color, cmpf);
645 return tmp_pal[nb_color >> 1].pal_id;
648 static int colormap_insert(struct color_node *map,
651 const uint32_t *palette,
652 const struct color_rect *box)
655 int component, cur_id;
656 int node_left_id = -1, node_right_id = -1;
657 struct color_node *node;
658 struct color_rect box1, box2;
659 const int pal_id = get_next_color(color_used, palette, &component, box);
664 /* create new node with that color */
665 cur_id = (*nb_used)++;
668 node->split = component;
669 node->palette_id = pal_id;
670 node->val[0] = c>>16 & 0xff;
671 node->val[1] = c>> 8 & 0xff;
672 node->val[2] = c & 0xff;
674 color_used[pal_id] = 1;
676 /* get the two boxes this node creates */
678 box1.max[component] = node->val[component];
679 box2.min[component] = node->val[component] + 1;
681 node_left_id = colormap_insert(map, color_used, nb_used, palette, &box1);
683 if (box2.min[component] <= box2.max[component])
684 node_right_id = colormap_insert(map, color_used, nb_used, palette, &box2);
686 node->left_id = node_left_id;
687 node->right_id = node_right_id;
692 static int cmp_pal_entry(const void *a, const void *b)
694 const int c1 = *(const uint32_t *)a & 0xffffff;
695 const int c2 = *(const uint32_t *)b & 0xffffff;
699 static void load_colormap(PaletteUseContext *s)
702 uint8_t color_used[AVPALETTE_COUNT] = {0};
703 uint32_t last_color = 0;
704 struct color_rect box;
706 /* disable transparent colors and dups */
707 qsort(s->palette, AVPALETTE_COUNT, sizeof(*s->palette), cmp_pal_entry);
708 for (i = 0; i < AVPALETTE_COUNT; i++) {
709 const uint32_t c = s->palette[i];
710 if (i != 0 && c == last_color) {
715 if ((c & 0xff000000) != 0xff000000) {
716 color_used[i] = 1; // ignore transparent color(s)
721 box.min[0] = box.min[1] = box.min[2] = 0x00;
722 box.max[0] = box.max[1] = box.max[2] = 0xff;
724 colormap_insert(s->map, color_used, &nb_used, s->palette, &box);
727 disp_tree(s->map, s->dot_filename);
729 if (s->debug_accuracy) {
730 if (!debug_accuracy(s->map, s->palette, s->color_search_method))
731 av_log(NULL, AV_LOG_INFO, "Accuracy check passed\n");
735 static void debug_mean_error(PaletteUseContext *s, const AVFrame *in1,
736 const AVFrame *in2, int frame_count)
739 const uint32_t *palette = s->palette;
740 uint32_t *src1 = (uint32_t *)in1->data[0];
741 uint8_t *src2 = in2->data[0];
742 const int src1_linesize = in1->linesize[0] >> 2;
743 const int src2_linesize = in2->linesize[0];
744 const float div = in1->width * in1->height * 3;
745 unsigned mean_err = 0;
747 for (y = 0; y < in1->height; y++) {
748 for (x = 0; x < in1->width; x++) {
749 const uint32_t c1 = src1[x];
750 const uint32_t c2 = palette[src2[x]];
751 const uint8_t rgb1[] = {c1 >> 16 & 0xff, c1 >> 8 & 0xff, c1 & 0xff};
752 const uint8_t rgb2[] = {c2 >> 16 & 0xff, c2 >> 8 & 0xff, c2 & 0xff};
753 mean_err += diff(rgb1, rgb2);
755 src1 += src1_linesize;
756 src2 += src2_linesize;
759 s->total_mean_err += mean_err;
761 av_log(NULL, AV_LOG_INFO, "MEP:%.3f TotalMEP:%.3f\n",
762 mean_err / div, s->total_mean_err / (div * frame_count));
765 static void set_processing_window(enum diff_mode diff_mode,
766 const AVFrame *prv_src, const AVFrame *cur_src,
767 const AVFrame *prv_dst, AVFrame *cur_dst,
768 int *xp, int *yp, int *wp, int *hp)
770 int x_start = 0, y_start = 0;
771 int width = cur_src->width;
772 int height = cur_src->height;
774 if (prv_src && diff_mode == DIFF_MODE_RECTANGLE) {
776 int x_end = cur_src->width - 1,
777 y_end = cur_src->height - 1;
778 const uint32_t *prv_srcp = (const uint32_t *)prv_src->data[0];
779 const uint32_t *cur_srcp = (const uint32_t *)cur_src->data[0];
780 const uint8_t *prv_dstp = prv_dst->data[0];
781 uint8_t *cur_dstp = cur_dst->data[0];
783 const int prv_src_linesize = prv_src->linesize[0] >> 2;
784 const int cur_src_linesize = cur_src->linesize[0] >> 2;
785 const int prv_dst_linesize = prv_dst->linesize[0];
786 const int cur_dst_linesize = cur_dst->linesize[0];
788 /* skip common lines */
789 while (y_start < y_end && !memcmp(prv_srcp + y_start*prv_src_linesize,
790 cur_srcp + y_start*cur_src_linesize,
791 cur_src->width * 4)) {
792 memcpy(cur_dstp + y_start*cur_dst_linesize,
793 prv_dstp + y_start*prv_dst_linesize,
797 while (y_end > y_start && !memcmp(prv_srcp + y_end*prv_src_linesize,
798 cur_srcp + y_end*cur_src_linesize,
799 cur_src->width * 4)) {
800 memcpy(cur_dstp + y_end*cur_dst_linesize,
801 prv_dstp + y_end*prv_dst_linesize,
806 height = y_end + 1 - y_start;
808 /* skip common columns */
809 while (x_start < x_end) {
811 for (y = y_start; y <= y_end; y++) {
812 if (prv_srcp[y*prv_src_linesize + x_start] != cur_srcp[y*cur_src_linesize + x_start]) {
821 while (x_end > x_start) {
823 for (y = y_start; y <= y_end; y++) {
824 if (prv_srcp[y*prv_src_linesize + x_end] != cur_srcp[y*cur_src_linesize + x_end]) {
833 width = x_end + 1 - x_start;
836 for (y = y_start; y <= y_end; y++)
837 memcpy(cur_dstp + y*cur_dst_linesize,
838 prv_dstp + y*prv_dst_linesize, x_start);
840 if (x_end != cur_src->width - 1) {
841 const int copy_len = cur_src->width - 1 - x_end;
842 for (y = y_start; y <= y_end; y++)
843 memcpy(cur_dstp + y*cur_dst_linesize + x_end + 1,
844 prv_dstp + y*prv_dst_linesize + x_end + 1,
854 static AVFrame *apply_palette(AVFilterLink *inlink, AVFrame *in)
857 AVFilterContext *ctx = inlink->dst;
858 PaletteUseContext *s = ctx->priv;
859 AVFilterLink *outlink = inlink->dst->outputs[0];
861 AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
866 av_frame_copy_props(out, in);
868 set_processing_window(s->diff_mode, s->last_in, in,
869 s->last_out, out, &x, &y, &w, &h);
870 av_frame_free(&s->last_in);
871 av_frame_free(&s->last_out);
872 s->last_in = av_frame_clone(in);
873 s->last_out = av_frame_clone(out);
874 if (!s->last_in || !s->last_out ||
875 av_frame_make_writable(s->last_in) < 0) {
881 ff_dlog(ctx, "%dx%d rect: (%d;%d) -> (%d,%d) [area:%dx%d]\n",
882 w, h, x, y, x+w, y+h, in->width, in->height);
884 if (s->set_frame(s, out, in, x, y, w, h) < 0) {
888 memcpy(out->data[1], s->palette, AVPALETTE_SIZE);
889 if (s->calc_mean_err)
890 debug_mean_error(s, in, out, inlink->frame_count);
895 static int config_output(AVFilterLink *outlink)
898 AVFilterContext *ctx = outlink->src;
899 PaletteUseContext *s = ctx->priv;
901 outlink->w = ctx->inputs[0]->w;
902 outlink->h = ctx->inputs[0]->h;
904 outlink->time_base = ctx->inputs[0]->time_base;
905 if ((ret = ff_dualinput_init(ctx, &s->dinput)) < 0)
910 static int config_input_palette(AVFilterLink *inlink)
912 AVFilterContext *ctx = inlink->dst;
914 if (inlink->w * inlink->h != AVPALETTE_COUNT) {
915 av_log(ctx, AV_LOG_ERROR,
916 "Palette input must contain exactly %d pixels. "
917 "Specified input has %dx%d=%d pixels\n",
918 AVPALETTE_COUNT, inlink->w, inlink->h,
919 inlink->w * inlink->h);
920 return AVERROR(EINVAL);
925 static void load_palette(PaletteUseContext *s, const AVFrame *palette_frame)
928 const uint32_t *p = (const uint32_t *)palette_frame->data[0];
929 const int p_linesize = palette_frame->linesize[0] >> 2;
932 for (y = 0; y < palette_frame->height; y++) {
933 for (x = 0; x < palette_frame->width; x++)
934 s->palette[i++] = p[x];
940 s->palette_loaded = 1;
943 static AVFrame *load_apply_palette(AVFilterContext *ctx, AVFrame *main,
944 const AVFrame *second)
946 AVFilterLink *inlink = ctx->inputs[0];
947 PaletteUseContext *s = ctx->priv;
948 if (!s->palette_loaded) {
949 load_palette(s, second);
951 return apply_palette(inlink, main);
954 static int filter_frame(AVFilterLink *inlink, AVFrame *in)
956 PaletteUseContext *s = inlink->dst->priv;
957 return ff_dualinput_filter_frame(&s->dinput, inlink, in);
960 #define DEFINE_SET_FRAME(color_search, name, value) \
961 static int set_frame_##name(PaletteUseContext *s, AVFrame *out, AVFrame *in, \
962 int x_start, int y_start, int w, int h) \
964 return set_frame(s, out, in, x_start, y_start, w, h, value, color_search); \
967 #define DEFINE_SET_FRAME_COLOR_SEARCH(color_search, color_search_macro) \
968 DEFINE_SET_FRAME(color_search_macro, color_search##_##none, DITHERING_NONE) \
969 DEFINE_SET_FRAME(color_search_macro, color_search##_##bayer, DITHERING_BAYER) \
970 DEFINE_SET_FRAME(color_search_macro, color_search##_##heckbert, DITHERING_HECKBERT) \
971 DEFINE_SET_FRAME(color_search_macro, color_search##_##floyd_steinberg, DITHERING_FLOYD_STEINBERG) \
972 DEFINE_SET_FRAME(color_search_macro, color_search##_##sierra2, DITHERING_SIERRA2) \
973 DEFINE_SET_FRAME(color_search_macro, color_search##_##sierra2_4a, DITHERING_SIERRA2_4A) \
975 DEFINE_SET_FRAME_COLOR_SEARCH(nns_iterative, COLOR_SEARCH_NNS_ITERATIVE)
976 DEFINE_SET_FRAME_COLOR_SEARCH(nns_recursive, COLOR_SEARCH_NNS_RECURSIVE)
977 DEFINE_SET_FRAME_COLOR_SEARCH(bruteforce, COLOR_SEARCH_BRUTEFORCE)
979 #define DITHERING_ENTRIES(color_search) { \
980 set_frame_##color_search##_none, \
981 set_frame_##color_search##_bayer, \
982 set_frame_##color_search##_heckbert, \
983 set_frame_##color_search##_floyd_steinberg, \
984 set_frame_##color_search##_sierra2, \
985 set_frame_##color_search##_sierra2_4a, \
988 static const set_frame_func set_frame_lut[NB_COLOR_SEARCHES][NB_DITHERING] = {
989 DITHERING_ENTRIES(nns_iterative),
990 DITHERING_ENTRIES(nns_recursive),
991 DITHERING_ENTRIES(bruteforce),
994 static int dither_value(int p)
996 const int q = p ^ (p >> 3);
997 return (p & 4) >> 2 | (q & 4) >> 1 \
998 | (p & 2) << 1 | (q & 2) << 2 \
999 | (p & 1) << 4 | (q & 1) << 5;
1002 static av_cold int init(AVFilterContext *ctx)
1004 PaletteUseContext *s = ctx->priv;
1005 s->dinput.repeatlast = 1; // only 1 frame in the palette
1006 s->dinput.skip_initial_unpaired = 1;
1007 s->dinput.process = load_apply_palette;
1009 s->set_frame = set_frame_lut[s->color_search_method][s->dither];
1011 if (s->dither == DITHERING_BAYER) {
1013 const int delta = 1 << (5 - s->bayer_scale); // to avoid too much luma
1015 for (i = 0; i < FF_ARRAY_ELEMS(s->ordered_dither); i++)
1016 s->ordered_dither[i] = (dither_value(i) >> s->bayer_scale) - delta;
1022 static int request_frame(AVFilterLink *outlink)
1024 PaletteUseContext *s = outlink->src->priv;
1025 return ff_dualinput_request_frame(&s->dinput, outlink);
1028 static av_cold void uninit(AVFilterContext *ctx)
1031 PaletteUseContext *s = ctx->priv;
1033 ff_dualinput_uninit(&s->dinput);
1034 for (i = 0; i < CACHE_SIZE; i++)
1035 av_freep(&s->cache[i].entries);
1036 av_frame_free(&s->last_in);
1037 av_frame_free(&s->last_out);
1040 static const AVFilterPad paletteuse_inputs[] = {
1043 .type = AVMEDIA_TYPE_VIDEO,
1044 .filter_frame = filter_frame,
1045 .needs_writable = 1, // for error diffusal dithering
1048 .type = AVMEDIA_TYPE_VIDEO,
1049 .config_props = config_input_palette,
1050 .filter_frame = filter_frame,
1055 static const AVFilterPad paletteuse_outputs[] = {
1058 .type = AVMEDIA_TYPE_VIDEO,
1059 .config_props = config_output,
1060 .request_frame = request_frame,
1065 AVFilter ff_vf_paletteuse = {
1066 .name = "paletteuse",
1067 .description = NULL_IF_CONFIG_SMALL("Use a palette to downsample an input video stream."),
1068 .priv_size = sizeof(PaletteUseContext),
1069 .query_formats = query_formats,
1072 .inputs = paletteuse_inputs,
1073 .outputs = paletteuse_outputs,
1074 .priv_class = &paletteuse_class,