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 },
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}, CHAR_MIN, CHAR_MAX, 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 AVFilterFormats *in = ff_make_format_list(in_fmts);
146 AVFilterFormats *inpal = ff_make_format_list(inpal_fmts);
147 AVFilterFormats *out = ff_make_format_list(out_fmts);
148 if (!in || !inpal || !out) {
152 return AVERROR(ENOMEM);
154 if ((ret = ff_formats_ref(in , &ctx->inputs[0]->out_formats)) < 0 ||
155 (ret = ff_formats_ref(inpal, &ctx->inputs[1]->out_formats)) < 0 ||
156 (ret = ff_formats_ref(out , &ctx->outputs[0]->in_formats)) < 0)
161 static av_always_inline int dither_color(uint32_t px, int er, int eg, int eb, int scale, int shift)
163 return av_clip_uint8( px >> 24 ) << 24
164 | av_clip_uint8((px >> 16 & 0xff) + ((er * scale) / (1<<shift))) << 16
165 | av_clip_uint8((px >> 8 & 0xff) + ((eg * scale) / (1<<shift))) << 8
166 | av_clip_uint8((px & 0xff) + ((eb * scale) / (1<<shift)));
169 static av_always_inline int diff(const uint8_t *c1, const uint8_t *c2, const int trans_thresh)
171 // XXX: try L*a*b with CIE76 (dL*dL + da*da + db*db)
172 const int dr = c1[1] - c2[1];
173 const int dg = c1[2] - c2[2];
174 const int db = c1[3] - c2[3];
176 if (c1[0] < trans_thresh && c2[0] < trans_thresh) {
178 } else if (c1[0] >= trans_thresh && c2[0] >= trans_thresh) {
179 return dr*dr + dg*dg + db*db;
181 return 255*255 + 255*255 + 255*255;
185 static av_always_inline uint8_t colormap_nearest_bruteforce(const uint32_t *palette, const uint8_t *argb, const int trans_thresh)
187 int i, pal_id = -1, min_dist = INT_MAX;
189 for (i = 0; i < AVPALETTE_COUNT; i++) {
190 const uint32_t c = palette[i];
192 if (c >> 24 >= trans_thresh) { // ignore transparent entry
193 const uint8_t palargb[] = {
194 palette[i]>>24 & 0xff,
195 palette[i]>>16 & 0xff,
196 palette[i]>> 8 & 0xff,
199 const int d = diff(palargb, argb, trans_thresh);
209 /* Recursive form, simpler but a bit slower. Kept for reference. */
210 struct nearest_color {
215 static void colormap_nearest_node(const struct color_node *map,
217 const uint8_t *target,
218 const int trans_thresh,
219 struct nearest_color *nearest)
221 const struct color_node *kd = map + node_pos;
222 const int s = kd->split;
223 int dx, nearer_kd_id, further_kd_id;
224 const uint8_t *current = kd->val;
225 const int current_to_target = diff(target, current, trans_thresh);
227 if (current_to_target < nearest->dist_sqd) {
228 nearest->node_pos = node_pos;
229 nearest->dist_sqd = current_to_target;
232 if (kd->left_id != -1 || kd->right_id != -1) {
233 dx = target[s] - current[s];
235 if (dx <= 0) nearer_kd_id = kd->left_id, further_kd_id = kd->right_id;
236 else nearer_kd_id = kd->right_id, further_kd_id = kd->left_id;
238 if (nearer_kd_id != -1)
239 colormap_nearest_node(map, nearer_kd_id, target, trans_thresh, nearest);
241 if (further_kd_id != -1 && dx*dx < nearest->dist_sqd)
242 colormap_nearest_node(map, further_kd_id, target, trans_thresh, nearest);
246 static av_always_inline uint8_t colormap_nearest_recursive(const struct color_node *node, const uint8_t *rgb, const int trans_thresh)
248 struct nearest_color res = {.dist_sqd = INT_MAX, .node_pos = -1};
249 colormap_nearest_node(node, 0, rgb, trans_thresh, &res);
250 return node[res.node_pos].palette_id;
258 static av_always_inline uint8_t colormap_nearest_iterative(const struct color_node *root, const uint8_t *target, const int trans_thresh)
260 int pos = 0, best_node_id = -1, best_dist = INT_MAX, cur_color_id = 0;
261 struct stack_node nodes[16];
262 struct stack_node *node = &nodes[0];
266 const struct color_node *kd = &root[cur_color_id];
267 const uint8_t *current = kd->val;
268 const int current_to_target = diff(target, current, trans_thresh);
270 /* Compare current color node to the target and update our best node if
271 * it's actually better. */
272 if (current_to_target < best_dist) {
273 best_node_id = cur_color_id;
274 if (!current_to_target)
275 goto end; // exact match, we can return immediately
276 best_dist = current_to_target;
279 /* Check if it's not a leaf */
280 if (kd->left_id != -1 || kd->right_id != -1) {
281 const int split = kd->split;
282 const int dx = target[split] - current[split];
283 int nearer_kd_id, further_kd_id;
285 /* Define which side is the most interesting. */
286 if (dx <= 0) nearer_kd_id = kd->left_id, further_kd_id = kd->right_id;
287 else nearer_kd_id = kd->right_id, further_kd_id = kd->left_id;
289 if (nearer_kd_id != -1) {
290 if (further_kd_id != -1) {
291 /* Here, both paths are defined, so we push a state for
292 * when we are going back. */
293 node->color_id = further_kd_id;
298 /* We can now update current color with the most probable path
299 * (no need to create a state since there is nothing to save
301 cur_color_id = nearer_kd_id;
303 } else if (dx*dx < best_dist) {
304 /* The nearest path isn't available, so there is only one path
305 * possible and it's the least probable. We enter it only if the
306 * distance from the current point to the hyper rectangle is
307 * less than our best distance. */
308 cur_color_id = further_kd_id;
313 /* Unstack as much as we can, typically as long as the least probable
314 * branch aren't actually probable. */
319 } while (node->dx2 >= best_dist);
321 /* We got a node where the least probable branch might actually contain
322 * a relevant color. */
323 cur_color_id = node->color_id;
327 return root[best_node_id].palette_id;
330 #define COLORMAP_NEAREST(search, palette, root, target, trans_thresh) \
331 search == COLOR_SEARCH_NNS_ITERATIVE ? colormap_nearest_iterative(root, target, trans_thresh) : \
332 search == COLOR_SEARCH_NNS_RECURSIVE ? colormap_nearest_recursive(root, target, trans_thresh) : \
333 colormap_nearest_bruteforce(palette, target, trans_thresh)
336 * Check if the requested color is in the cache already. If not, find it in the
337 * color tree and cache it.
338 * Note: a, r, g, and b are the components of color, but are passed as well to avoid
339 * recomputing them (they are generally computed by the caller for other uses).
341 static av_always_inline int color_get(PaletteUseContext *s, uint32_t color,
342 uint8_t a, uint8_t r, uint8_t g, uint8_t b,
343 const enum color_search_method search_method)
346 const uint8_t argb_elts[] = {a, r, g, b};
347 const uint8_t rhash = r & ((1<<NBITS)-1);
348 const uint8_t ghash = g & ((1<<NBITS)-1);
349 const uint8_t bhash = b & ((1<<NBITS)-1);
350 const unsigned hash = rhash<<(NBITS*2) | ghash<<NBITS | bhash;
351 struct cache_node *node = &s->cache[hash];
352 struct cached_color *e;
354 // first, check for transparency
355 if (a < s->trans_thresh && s->transparency_index >= 0) {
356 return s->transparency_index;
359 for (i = 0; i < node->nb_entries; i++) {
360 e = &node->entries[i];
361 if (e->color == color)
365 e = av_dynarray2_add((void**)&node->entries, &node->nb_entries,
366 sizeof(*node->entries), NULL);
368 return AVERROR(ENOMEM);
370 e->pal_entry = COLORMAP_NEAREST(search_method, s->palette, s->map, argb_elts, s->trans_thresh);
375 static av_always_inline int get_dst_color_err(PaletteUseContext *s,
376 uint32_t c, int *er, int *eg, int *eb,
377 const enum color_search_method search_method)
379 const uint8_t a = c >> 24 & 0xff;
380 const uint8_t r = c >> 16 & 0xff;
381 const uint8_t g = c >> 8 & 0xff;
382 const uint8_t b = c & 0xff;
384 const int dstx = color_get(s, c, a, r, g, b, search_method);
387 dstc = s->palette[dstx];
388 *er = r - (dstc >> 16 & 0xff);
389 *eg = g - (dstc >> 8 & 0xff);
390 *eb = b - (dstc & 0xff);
394 static av_always_inline int set_frame(PaletteUseContext *s, AVFrame *out, AVFrame *in,
395 int x_start, int y_start, int w, int h,
396 enum dithering_mode dither,
397 const enum color_search_method search_method)
400 const int src_linesize = in ->linesize[0] >> 2;
401 const int dst_linesize = out->linesize[0];
402 uint32_t *src = ((uint32_t *)in ->data[0]) + y_start*src_linesize;
403 uint8_t *dst = out->data[0] + y_start*dst_linesize;
408 for (y = y_start; y < h; y++) {
409 for (x = x_start; x < w; x++) {
412 if (dither == DITHERING_BAYER) {
413 const int d = s->ordered_dither[(y & 7)<<3 | (x & 7)];
414 const uint8_t a8 = src[x] >> 24 & 0xff;
415 const uint8_t r8 = src[x] >> 16 & 0xff;
416 const uint8_t g8 = src[x] >> 8 & 0xff;
417 const uint8_t b8 = src[x] & 0xff;
418 const uint8_t r = av_clip_uint8(r8 + d);
419 const uint8_t g = av_clip_uint8(g8 + d);
420 const uint8_t b = av_clip_uint8(b8 + d);
421 const int color = color_get(s, src[x], a8, r, g, b, search_method);
427 } else if (dither == DITHERING_HECKBERT) {
428 const int right = x < w - 1, down = y < h - 1;
429 const int color = get_dst_color_err(s, src[x], &er, &eg, &eb, search_method);
435 if (right) src[ x + 1] = dither_color(src[ x + 1], er, eg, eb, 3, 3);
436 if ( down) src[src_linesize + x ] = dither_color(src[src_linesize + x ], er, eg, eb, 3, 3);
437 if (right && down) src[src_linesize + x + 1] = dither_color(src[src_linesize + x + 1], er, eg, eb, 2, 3);
439 } else if (dither == DITHERING_FLOYD_STEINBERG) {
440 const int right = x < w - 1, down = y < h - 1, left = x > x_start;
441 const int color = get_dst_color_err(s, src[x], &er, &eg, &eb, search_method);
447 if (right) src[ x + 1] = dither_color(src[ x + 1], er, eg, eb, 7, 4);
448 if (left && down) src[src_linesize + x - 1] = dither_color(src[src_linesize + x - 1], er, eg, eb, 3, 4);
449 if ( down) src[src_linesize + x ] = dither_color(src[src_linesize + x ], er, eg, eb, 5, 4);
450 if (right && down) src[src_linesize + x + 1] = dither_color(src[src_linesize + x + 1], er, eg, eb, 1, 4);
452 } else if (dither == DITHERING_SIERRA2) {
453 const int right = x < w - 1, down = y < h - 1, left = x > x_start;
454 const int right2 = x < w - 2, left2 = x > x_start + 1;
455 const int color = get_dst_color_err(s, src[x], &er, &eg, &eb, search_method);
461 if (right) src[ x + 1] = dither_color(src[ x + 1], er, eg, eb, 4, 4);
462 if (right2) src[ x + 2] = dither_color(src[ x + 2], er, eg, eb, 3, 4);
465 if (left2) src[ src_linesize + x - 2] = dither_color(src[ src_linesize + x - 2], er, eg, eb, 1, 4);
466 if (left) src[ src_linesize + x - 1] = dither_color(src[ src_linesize + x - 1], er, eg, eb, 2, 4);
467 if (1) src[ src_linesize + x ] = dither_color(src[ src_linesize + x ], er, eg, eb, 3, 4);
468 if (right) src[ src_linesize + x + 1] = dither_color(src[ src_linesize + x + 1], er, eg, eb, 2, 4);
469 if (right2) src[ src_linesize + x + 2] = dither_color(src[ src_linesize + x + 2], er, eg, eb, 1, 4);
472 } else if (dither == DITHERING_SIERRA2_4A) {
473 const int right = x < w - 1, down = y < h - 1, left = x > x_start;
474 const int color = get_dst_color_err(s, src[x], &er, &eg, &eb, search_method);
480 if (right) src[ x + 1] = dither_color(src[ x + 1], er, eg, eb, 2, 2);
481 if (left && down) src[src_linesize + x - 1] = dither_color(src[src_linesize + x - 1], er, eg, eb, 1, 2);
482 if ( down) src[src_linesize + x ] = dither_color(src[src_linesize + x ], er, eg, eb, 1, 2);
485 const uint8_t a = src[x] >> 24 & 0xff;
486 const uint8_t r = src[x] >> 16 & 0xff;
487 const uint8_t g = src[x] >> 8 & 0xff;
488 const uint8_t b = src[x] & 0xff;
489 const int color = color_get(s, src[x], a, r, g, b, search_method);
503 static void disp_node(AVBPrint *buf,
504 const struct color_node *map,
505 int parent_id, int node_id,
508 const struct color_node *node = &map[node_id];
509 const uint32_t fontcolor = node->val[1] > 0x50 &&
510 node->val[2] > 0x50 &&
511 node->val[3] > 0x50 ? 0 : 0xffffff;
512 const int rgb_comp = node->split - 1;
513 av_bprintf(buf, "%*cnode%d ["
514 "label=\"%c%02X%c%02X%c%02X%c\" "
515 "fillcolor=\"#%02x%02x%02x\" "
516 "fontcolor=\"#%06"PRIX32"\"]\n",
517 depth*INDENT, ' ', node->palette_id,
518 "[ "[rgb_comp], node->val[1],
519 "][ "[rgb_comp], node->val[2],
520 " ]["[rgb_comp], node->val[3],
522 node->val[1], node->val[2], node->val[3],
525 av_bprintf(buf, "%*cnode%d -> node%d\n", depth*INDENT, ' ',
526 map[parent_id].palette_id, node->palette_id);
527 if (node->left_id != -1) disp_node(buf, map, node_id, node->left_id, depth + 1);
528 if (node->right_id != -1) disp_node(buf, map, node_id, node->right_id, depth + 1);
531 // debug_kdtree=kdtree.dot -> dot -Tpng kdtree.dot > kdtree.png
532 static int disp_tree(const struct color_node *node, const char *fname)
535 FILE *f = av_fopen_utf8(fname, "w");
538 int ret = AVERROR(errno);
539 av_log(NULL, AV_LOG_ERROR, "Cannot open file '%s' for writing: %s\n",
540 fname, av_err2str(ret));
544 av_bprint_init(&buf, 0, AV_BPRINT_SIZE_UNLIMITED);
546 av_bprintf(&buf, "digraph {\n");
547 av_bprintf(&buf, " node [style=filled fontsize=10 shape=box]\n");
548 disp_node(&buf, node, -1, 0, 0);
549 av_bprintf(&buf, "}\n");
551 fwrite(buf.str, 1, buf.len, f);
553 av_bprint_finalize(&buf, NULL);
557 static int debug_accuracy(const struct color_node *node, const uint32_t *palette, const int trans_thresh,
558 const enum color_search_method search_method)
560 int r, g, b, ret = 0;
562 for (r = 0; r < 256; r++) {
563 for (g = 0; g < 256; g++) {
564 for (b = 0; b < 256; b++) {
565 const uint8_t argb[] = {0xff, r, g, b};
566 const int r1 = COLORMAP_NEAREST(search_method, palette, node, argb, trans_thresh);
567 const int r2 = colormap_nearest_bruteforce(palette, argb, trans_thresh);
569 const uint32_t c1 = palette[r1];
570 const uint32_t c2 = palette[r2];
571 const uint8_t palargb1[] = { 0xff, c1>>16 & 0xff, c1>> 8 & 0xff, c1 & 0xff };
572 const uint8_t palargb2[] = { 0xff, c2>>16 & 0xff, c2>> 8 & 0xff, c2 & 0xff };
573 const int d1 = diff(palargb1, argb, trans_thresh);
574 const int d2 = diff(palargb2, argb, trans_thresh);
576 av_log(NULL, AV_LOG_ERROR,
577 "/!\\ %02X%02X%02X: %d ! %d (%06"PRIX32" ! %06"PRIX32") / dist: %d ! %d\n",
578 r, g, b, r1, r2, c1 & 0xffffff, c2 & 0xffffff, d1, d2);
598 typedef int (*cmp_func)(const void *, const void *);
600 #define DECLARE_CMP_FUNC(name, pos) \
601 static int cmp_##name(const void *pa, const void *pb) \
603 const struct color *a = pa; \
604 const struct color *b = pb; \
605 return (a->value >> (8 * (3 - (pos))) & 0xff) \
606 - (b->value >> (8 * (3 - (pos))) & 0xff); \
609 DECLARE_CMP_FUNC(a, 0)
610 DECLARE_CMP_FUNC(r, 1)
611 DECLARE_CMP_FUNC(g, 2)
612 DECLARE_CMP_FUNC(b, 3)
614 static const cmp_func cmp_funcs[] = {cmp_a, cmp_r, cmp_g, cmp_b};
616 static int get_next_color(const uint8_t *color_used, const uint32_t *palette,
617 const int trans_thresh,
618 int *component, const struct color_rect *box)
622 unsigned nb_color = 0;
623 struct color_rect ranges;
624 struct color tmp_pal[256];
627 ranges.min[0] = ranges.min[1] = ranges.min[2] = 0xff;
628 ranges.max[0] = ranges.max[1] = ranges.max[2] = 0x00;
630 for (i = 0; i < AVPALETTE_COUNT; i++) {
631 const uint32_t c = palette[i];
632 const uint8_t a = c >> 24 & 0xff;
633 const uint8_t r = c >> 16 & 0xff;
634 const uint8_t g = c >> 8 & 0xff;
635 const uint8_t b = c & 0xff;
637 if (a < trans_thresh) {
641 if (color_used[i] || (a != 0xff) ||
642 r < box->min[0] || g < box->min[1] || b < box->min[2] ||
643 r > box->max[0] || g > box->max[1] || b > box->max[2])
646 if (r < ranges.min[0]) ranges.min[0] = r;
647 if (g < ranges.min[1]) ranges.min[1] = g;
648 if (b < ranges.min[2]) ranges.min[2] = b;
650 if (r > ranges.max[0]) ranges.max[0] = r;
651 if (g > ranges.max[1]) ranges.max[1] = g;
652 if (b > ranges.max[2]) ranges.max[2] = b;
654 tmp_pal[nb_color].value = c;
655 tmp_pal[nb_color].pal_id = i;
663 /* define longest axis that will be the split component */
664 wr = ranges.max[0] - ranges.min[0];
665 wg = ranges.max[1] - ranges.min[1];
666 wb = ranges.max[2] - ranges.min[2];
667 if (wr >= wg && wr >= wb) longest = 1;
668 if (wg >= wr && wg >= wb) longest = 2;
669 if (wb >= wr && wb >= wg) longest = 3;
670 cmpf = cmp_funcs[longest];
671 *component = longest;
673 /* sort along this axis to get median */
674 AV_QSORT(tmp_pal, nb_color, struct color, cmpf);
676 return tmp_pal[nb_color >> 1].pal_id;
679 static int colormap_insert(struct color_node *map,
682 const uint32_t *palette,
683 const int trans_thresh,
684 const struct color_rect *box)
687 int component, cur_id;
688 int node_left_id = -1, node_right_id = -1;
689 struct color_node *node;
690 struct color_rect box1, box2;
691 const int pal_id = get_next_color(color_used, palette, trans_thresh, &component, box);
696 /* create new node with that color */
697 cur_id = (*nb_used)++;
700 node->split = component;
701 node->palette_id = pal_id;
702 node->val[0] = c>>24 & 0xff;
703 node->val[1] = c>>16 & 0xff;
704 node->val[2] = c>> 8 & 0xff;
705 node->val[3] = c & 0xff;
707 color_used[pal_id] = 1;
709 /* get the two boxes this node creates */
711 box1.max[component-1] = node->val[component];
712 box2.min[component-1] = node->val[component] + 1;
714 node_left_id = colormap_insert(map, color_used, nb_used, palette, trans_thresh, &box1);
716 if (box2.min[component-1] <= box2.max[component-1])
717 node_right_id = colormap_insert(map, color_used, nb_used, palette, trans_thresh, &box2);
719 node->left_id = node_left_id;
720 node->right_id = node_right_id;
725 static int cmp_pal_entry(const void *a, const void *b)
727 const int c1 = *(const uint32_t *)a & 0xffffff;
728 const int c2 = *(const uint32_t *)b & 0xffffff;
732 static void load_colormap(PaletteUseContext *s)
735 uint8_t color_used[AVPALETTE_COUNT] = {0};
736 uint32_t last_color = 0;
737 struct color_rect box;
739 /* disable transparent colors and dups */
740 qsort(s->palette, AVPALETTE_COUNT, sizeof(*s->palette), cmp_pal_entry);
741 // update transparency index:
742 if (s->transparency_index >= 0) {
743 for (i = 0; i < AVPALETTE_COUNT; i++) {
744 if ((s->palette[i]>>24 & 0xff) == 0) {
745 s->transparency_index = i; // we are assuming at most one transparent color in palette
751 for (i = 0; i < AVPALETTE_COUNT; i++) {
752 const uint32_t c = s->palette[i];
753 if (i != 0 && c == last_color) {
758 if (c >> 24 < s->trans_thresh) {
759 color_used[i] = 1; // ignore transparent color(s)
764 box.min[0] = box.min[1] = box.min[2] = 0x00;
765 box.max[0] = box.max[1] = box.max[2] = 0xff;
767 colormap_insert(s->map, color_used, &nb_used, s->palette, s->trans_thresh, &box);
770 disp_tree(s->map, s->dot_filename);
772 if (s->debug_accuracy) {
773 if (!debug_accuracy(s->map, s->palette, s->trans_thresh, s->color_search_method))
774 av_log(NULL, AV_LOG_INFO, "Accuracy check passed\n");
778 static void debug_mean_error(PaletteUseContext *s, const AVFrame *in1,
779 const AVFrame *in2, int frame_count)
782 const uint32_t *palette = s->palette;
783 uint32_t *src1 = (uint32_t *)in1->data[0];
784 uint8_t *src2 = in2->data[0];
785 const int src1_linesize = in1->linesize[0] >> 2;
786 const int src2_linesize = in2->linesize[0];
787 const float div = in1->width * in1->height * 3;
788 unsigned mean_err = 0;
790 for (y = 0; y < in1->height; y++) {
791 for (x = 0; x < in1->width; x++) {
792 const uint32_t c1 = src1[x];
793 const uint32_t c2 = palette[src2[x]];
794 const uint8_t argb1[] = {0xff, c1 >> 16 & 0xff, c1 >> 8 & 0xff, c1 & 0xff};
795 const uint8_t argb2[] = {0xff, c2 >> 16 & 0xff, c2 >> 8 & 0xff, c2 & 0xff};
796 mean_err += diff(argb1, argb2, s->trans_thresh);
798 src1 += src1_linesize;
799 src2 += src2_linesize;
802 s->total_mean_err += mean_err;
804 av_log(NULL, AV_LOG_INFO, "MEP:%.3f TotalMEP:%.3f\n",
805 mean_err / div, s->total_mean_err / (div * frame_count));
808 static void set_processing_window(enum diff_mode diff_mode,
809 const AVFrame *prv_src, const AVFrame *cur_src,
810 const AVFrame *prv_dst, AVFrame *cur_dst,
811 int *xp, int *yp, int *wp, int *hp)
813 int x_start = 0, y_start = 0;
814 int width = cur_src->width;
815 int height = cur_src->height;
817 if (prv_src && diff_mode == DIFF_MODE_RECTANGLE) {
819 int x_end = cur_src->width - 1,
820 y_end = cur_src->height - 1;
821 const uint32_t *prv_srcp = (const uint32_t *)prv_src->data[0];
822 const uint32_t *cur_srcp = (const uint32_t *)cur_src->data[0];
823 const uint8_t *prv_dstp = prv_dst->data[0];
824 uint8_t *cur_dstp = cur_dst->data[0];
826 const int prv_src_linesize = prv_src->linesize[0] >> 2;
827 const int cur_src_linesize = cur_src->linesize[0] >> 2;
828 const int prv_dst_linesize = prv_dst->linesize[0];
829 const int cur_dst_linesize = cur_dst->linesize[0];
831 /* skip common lines */
832 while (y_start < y_end && !memcmp(prv_srcp + y_start*prv_src_linesize,
833 cur_srcp + y_start*cur_src_linesize,
834 cur_src->width * 4)) {
835 memcpy(cur_dstp + y_start*cur_dst_linesize,
836 prv_dstp + y_start*prv_dst_linesize,
840 while (y_end > y_start && !memcmp(prv_srcp + y_end*prv_src_linesize,
841 cur_srcp + y_end*cur_src_linesize,
842 cur_src->width * 4)) {
843 memcpy(cur_dstp + y_end*cur_dst_linesize,
844 prv_dstp + y_end*prv_dst_linesize,
849 height = y_end + 1 - y_start;
851 /* skip common columns */
852 while (x_start < x_end) {
854 for (y = y_start; y <= y_end; y++) {
855 if (prv_srcp[y*prv_src_linesize + x_start] != cur_srcp[y*cur_src_linesize + x_start]) {
864 while (x_end > x_start) {
866 for (y = y_start; y <= y_end; y++) {
867 if (prv_srcp[y*prv_src_linesize + x_end] != cur_srcp[y*cur_src_linesize + x_end]) {
876 width = x_end + 1 - x_start;
879 for (y = y_start; y <= y_end; y++)
880 memcpy(cur_dstp + y*cur_dst_linesize,
881 prv_dstp + y*prv_dst_linesize, x_start);
883 if (x_end != cur_src->width - 1) {
884 const int copy_len = cur_src->width - 1 - x_end;
885 for (y = y_start; y <= y_end; y++)
886 memcpy(cur_dstp + y*cur_dst_linesize + x_end + 1,
887 prv_dstp + y*prv_dst_linesize + x_end + 1,
897 static int apply_palette(AVFilterLink *inlink, AVFrame *in, AVFrame **outf)
900 AVFilterContext *ctx = inlink->dst;
901 PaletteUseContext *s = ctx->priv;
902 AVFilterLink *outlink = inlink->dst->outputs[0];
904 AVFrame *out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
908 return AVERROR(ENOMEM);
910 av_frame_copy_props(out, in);
912 set_processing_window(s->diff_mode, s->last_in, in,
913 s->last_out, out, &x, &y, &w, &h);
914 av_frame_free(&s->last_in);
915 av_frame_free(&s->last_out);
916 s->last_in = av_frame_clone(in);
917 s->last_out = av_frame_clone(out);
918 if (!s->last_in || !s->last_out ||
919 av_frame_make_writable(s->last_in) < 0) {
923 return AVERROR(ENOMEM);
926 ff_dlog(ctx, "%dx%d rect: (%d;%d) -> (%d,%d) [area:%dx%d]\n",
927 w, h, x, y, x+w, y+h, in->width, in->height);
929 ret = s->set_frame(s, out, in, x, y, w, h);
935 memcpy(out->data[1], s->palette, AVPALETTE_SIZE);
936 if (s->calc_mean_err)
937 debug_mean_error(s, in, out, inlink->frame_count_out);
943 static int config_output(AVFilterLink *outlink)
946 AVFilterContext *ctx = outlink->src;
947 PaletteUseContext *s = ctx->priv;
949 ret = ff_framesync_init_dualinput(&s->fs, ctx);
952 s->fs.opt_repeatlast = 1; // only 1 frame in the palette
953 s->fs.in[1].before = s->fs.in[1].after = EXT_INFINITY;
954 s->fs.on_event = load_apply_palette;
956 outlink->w = ctx->inputs[0]->w;
957 outlink->h = ctx->inputs[0]->h;
959 outlink->time_base = ctx->inputs[0]->time_base;
960 if ((ret = ff_framesync_configure(&s->fs)) < 0)
965 static int config_input_palette(AVFilterLink *inlink)
967 AVFilterContext *ctx = inlink->dst;
969 if (inlink->w * inlink->h != AVPALETTE_COUNT) {
970 av_log(ctx, AV_LOG_ERROR,
971 "Palette input must contain exactly %d pixels. "
972 "Specified input has %dx%d=%d pixels\n",
973 AVPALETTE_COUNT, inlink->w, inlink->h,
974 inlink->w * inlink->h);
975 return AVERROR(EINVAL);
980 static void load_palette(PaletteUseContext *s, const AVFrame *palette_frame)
983 const uint32_t *p = (const uint32_t *)palette_frame->data[0];
984 const int p_linesize = palette_frame->linesize[0] >> 2;
986 s->transparency_index = -1;
989 memset(s->palette, 0, sizeof(s->palette));
990 memset(s->map, 0, sizeof(s->map));
991 for (i = 0; i < CACHE_SIZE; i++)
992 av_freep(&s->cache[i].entries);
993 memset(s->cache, 0, sizeof(s->cache));
997 for (y = 0; y < palette_frame->height; y++) {
998 for (x = 0; x < palette_frame->width; x++) {
999 s->palette[i] = p[x];
1000 if (p[x]>>24 < s->trans_thresh) {
1001 s->transparency_index = i; // we are assuming at most one transparent color in palette
1011 s->palette_loaded = 1;
1014 static int load_apply_palette(FFFrameSync *fs)
1016 AVFilterContext *ctx = fs->parent;
1017 AVFilterLink *inlink = ctx->inputs[0];
1018 PaletteUseContext *s = ctx->priv;
1019 AVFrame *master, *second, *out = NULL;
1022 // writable for error diffusal dithering
1023 ret = ff_framesync_dualinput_get_writable(fs, &master, &second);
1026 if (!master || !second) {
1030 if (!s->palette_loaded) {
1031 load_palette(s, second);
1033 ret = apply_palette(inlink, master, &out);
1036 return ff_filter_frame(ctx->outputs[0], out);
1039 av_frame_free(&master);
1043 #define DEFINE_SET_FRAME(color_search, name, value) \
1044 static int set_frame_##name(PaletteUseContext *s, AVFrame *out, AVFrame *in, \
1045 int x_start, int y_start, int w, int h) \
1047 return set_frame(s, out, in, x_start, y_start, w, h, value, color_search); \
1050 #define DEFINE_SET_FRAME_COLOR_SEARCH(color_search, color_search_macro) \
1051 DEFINE_SET_FRAME(color_search_macro, color_search##_##none, DITHERING_NONE) \
1052 DEFINE_SET_FRAME(color_search_macro, color_search##_##bayer, DITHERING_BAYER) \
1053 DEFINE_SET_FRAME(color_search_macro, color_search##_##heckbert, DITHERING_HECKBERT) \
1054 DEFINE_SET_FRAME(color_search_macro, color_search##_##floyd_steinberg, DITHERING_FLOYD_STEINBERG) \
1055 DEFINE_SET_FRAME(color_search_macro, color_search##_##sierra2, DITHERING_SIERRA2) \
1056 DEFINE_SET_FRAME(color_search_macro, color_search##_##sierra2_4a, DITHERING_SIERRA2_4A) \
1058 DEFINE_SET_FRAME_COLOR_SEARCH(nns_iterative, COLOR_SEARCH_NNS_ITERATIVE)
1059 DEFINE_SET_FRAME_COLOR_SEARCH(nns_recursive, COLOR_SEARCH_NNS_RECURSIVE)
1060 DEFINE_SET_FRAME_COLOR_SEARCH(bruteforce, COLOR_SEARCH_BRUTEFORCE)
1062 #define DITHERING_ENTRIES(color_search) { \
1063 set_frame_##color_search##_none, \
1064 set_frame_##color_search##_bayer, \
1065 set_frame_##color_search##_heckbert, \
1066 set_frame_##color_search##_floyd_steinberg, \
1067 set_frame_##color_search##_sierra2, \
1068 set_frame_##color_search##_sierra2_4a, \
1071 static const set_frame_func set_frame_lut[NB_COLOR_SEARCHES][NB_DITHERING] = {
1072 DITHERING_ENTRIES(nns_iterative),
1073 DITHERING_ENTRIES(nns_recursive),
1074 DITHERING_ENTRIES(bruteforce),
1077 static int dither_value(int p)
1079 const int q = p ^ (p >> 3);
1080 return (p & 4) >> 2 | (q & 4) >> 1 \
1081 | (p & 2) << 1 | (q & 2) << 2 \
1082 | (p & 1) << 4 | (q & 1) << 5;
1085 static av_cold int init(AVFilterContext *ctx)
1087 PaletteUseContext *s = ctx->priv;
1089 s->set_frame = set_frame_lut[s->color_search_method][s->dither];
1091 if (s->dither == DITHERING_BAYER) {
1093 const int delta = 1 << (5 - s->bayer_scale); // to avoid too much luma
1095 for (i = 0; i < FF_ARRAY_ELEMS(s->ordered_dither); i++)
1096 s->ordered_dither[i] = (dither_value(i) >> s->bayer_scale) - delta;
1102 static int activate(AVFilterContext *ctx)
1104 PaletteUseContext *s = ctx->priv;
1105 return ff_framesync_activate(&s->fs);
1108 static av_cold void uninit(AVFilterContext *ctx)
1111 PaletteUseContext *s = ctx->priv;
1113 ff_framesync_uninit(&s->fs);
1114 for (i = 0; i < CACHE_SIZE; i++)
1115 av_freep(&s->cache[i].entries);
1116 av_frame_free(&s->last_in);
1117 av_frame_free(&s->last_out);
1120 static const AVFilterPad paletteuse_inputs[] = {
1123 .type = AVMEDIA_TYPE_VIDEO,
1126 .type = AVMEDIA_TYPE_VIDEO,
1127 .config_props = config_input_palette,
1132 static const AVFilterPad paletteuse_outputs[] = {
1135 .type = AVMEDIA_TYPE_VIDEO,
1136 .config_props = config_output,
1141 AVFilter ff_vf_paletteuse = {
1142 .name = "paletteuse",
1143 .description = NULL_IF_CONFIG_SMALL("Use a palette to downsample an input video stream."),
1144 .priv_size = sizeof(PaletteUseContext),
1145 .query_formats = query_formats,
1148 .activate = activate,
1149 .inputs = paletteuse_inputs,
1150 .outputs = paletteuse_outputs,
1151 .priv_class = &paletteuse_class,