2 * Copyright (C) 2003-2004 the ffmpeg project
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 * On2 VP3 Video Decoder
25 * VP3 Video Decoder by Mike Melanson (mike at multimedia.cx)
26 * For more information about the VP3 coding process, visit:
27 * http://wiki.multimedia.cx/index.php?title=On2_VP3
29 * Theora decoder by Alex Beregszaszi
43 #define FRAGMENT_PIXELS 8
45 static av_cold int vp3_decode_end(AVCodecContext *avctx);
47 //FIXME split things out into their own arrays
48 typedef struct Vp3Fragment {
50 uint8_t coding_method;
54 #define SB_NOT_CODED 0
55 #define SB_PARTIALLY_CODED 1
56 #define SB_FULLY_CODED 2
58 // This is the maximum length of a single long bit run that can be encoded
59 // for superblock coding or block qps. Theora special-cases this to read a
60 // bit instead of flipping the current bit to allow for runs longer than 4129.
61 #define MAXIMUM_LONG_BIT_RUN 4129
63 #define MODE_INTER_NO_MV 0
65 #define MODE_INTER_PLUS_MV 2
66 #define MODE_INTER_LAST_MV 3
67 #define MODE_INTER_PRIOR_LAST 4
68 #define MODE_USING_GOLDEN 5
69 #define MODE_GOLDEN_MV 6
70 #define MODE_INTER_FOURMV 7
71 #define CODING_MODE_COUNT 8
73 /* special internal mode */
76 /* There are 6 preset schemes, plus a free-form scheme */
77 static const int ModeAlphabet[6][CODING_MODE_COUNT] =
79 /* scheme 1: Last motion vector dominates */
80 { MODE_INTER_LAST_MV, MODE_INTER_PRIOR_LAST,
81 MODE_INTER_PLUS_MV, MODE_INTER_NO_MV,
82 MODE_INTRA, MODE_USING_GOLDEN,
83 MODE_GOLDEN_MV, MODE_INTER_FOURMV },
86 { MODE_INTER_LAST_MV, MODE_INTER_PRIOR_LAST,
87 MODE_INTER_NO_MV, MODE_INTER_PLUS_MV,
88 MODE_INTRA, MODE_USING_GOLDEN,
89 MODE_GOLDEN_MV, MODE_INTER_FOURMV },
92 { MODE_INTER_LAST_MV, MODE_INTER_PLUS_MV,
93 MODE_INTER_PRIOR_LAST, MODE_INTER_NO_MV,
94 MODE_INTRA, MODE_USING_GOLDEN,
95 MODE_GOLDEN_MV, MODE_INTER_FOURMV },
98 { MODE_INTER_LAST_MV, MODE_INTER_PLUS_MV,
99 MODE_INTER_NO_MV, MODE_INTER_PRIOR_LAST,
100 MODE_INTRA, MODE_USING_GOLDEN,
101 MODE_GOLDEN_MV, MODE_INTER_FOURMV },
103 /* scheme 5: No motion vector dominates */
104 { MODE_INTER_NO_MV, MODE_INTER_LAST_MV,
105 MODE_INTER_PRIOR_LAST, MODE_INTER_PLUS_MV,
106 MODE_INTRA, MODE_USING_GOLDEN,
107 MODE_GOLDEN_MV, MODE_INTER_FOURMV },
110 { MODE_INTER_NO_MV, MODE_USING_GOLDEN,
111 MODE_INTER_LAST_MV, MODE_INTER_PRIOR_LAST,
112 MODE_INTER_PLUS_MV, MODE_INTRA,
113 MODE_GOLDEN_MV, MODE_INTER_FOURMV },
117 static const uint8_t hilbert_offset[16][2] = {
118 {0,0}, {1,0}, {1,1}, {0,1},
119 {0,2}, {0,3}, {1,3}, {1,2},
120 {2,2}, {2,3}, {3,3}, {3,2},
121 {3,1}, {2,1}, {2,0}, {3,0}
124 #define MIN_DEQUANT_VAL 2
126 typedef struct Vp3DecodeContext {
127 AVCodecContext *avctx;
128 int theora, theora_tables;
131 int chroma_x_shift, chroma_y_shift;
132 AVFrame golden_frame;
134 AVFrame current_frame;
144 int superblock_count;
145 int y_superblock_width;
146 int y_superblock_height;
147 int y_superblock_count;
148 int c_superblock_width;
149 int c_superblock_height;
150 int c_superblock_count;
151 int u_superblock_start;
152 int v_superblock_start;
153 unsigned char *superblock_coding;
155 int macroblock_count;
156 int macroblock_width;
157 int macroblock_height;
160 int fragment_width[2];
161 int fragment_height[2];
163 Vp3Fragment *all_fragments;
164 int fragment_start[3];
167 int8_t (*motion_val[2])[2];
172 uint16_t coded_dc_scale_factor[64];
173 uint32_t coded_ac_scale_factor[64];
174 uint8_t base_matrix[384][64];
175 uint8_t qr_count[2][3];
176 uint8_t qr_size [2][3][64];
177 uint16_t qr_base[2][3][64];
180 * This is a list of all tokens in bitstream order. Reordering takes place
181 * by pulling from each level during IDCT. As a consequence, IDCT must be
182 * in Hilbert order, making the minimum slice height 64 for 4:2:0 and 32
183 * otherwise. The 32 different tokens with up to 12 bits of extradata are
184 * collapsed into 3 types, packed as follows:
185 * (from the low to high bits)
187 * 2 bits: type (0,1,2)
188 * 0: EOB run, 14 bits for run length (12 needed)
189 * 1: zero run, 7 bits for run length
190 * 7 bits for the next coefficient (3 needed)
191 * 2: coefficient, 14 bits (11 needed)
193 * Coefficients are signed, so are packed in the highest bits for automatic
196 int16_t *dct_tokens[3][64];
197 int16_t *dct_tokens_base;
198 #define TOKEN_EOB(eob_run) ((eob_run) << 2)
199 #define TOKEN_ZERO_RUN(coeff, zero_run) (((coeff) << 9) + ((zero_run) << 2) + 1)
200 #define TOKEN_COEFF(coeff) (((coeff) << 2) + 2)
203 * number of blocks that contain DCT coefficients at the given level or higher
205 int num_coded_frags[3][64];
206 int total_num_coded_frags;
208 /* this is a list of indexes into the all_fragments array indicating
209 * which of the fragments are coded */
210 int *coded_fragment_list[3];
218 VLC superblock_run_length_vlc;
219 VLC fragment_run_length_vlc;
221 VLC motion_vector_vlc;
223 /* these arrays need to be on 16-byte boundaries since SSE2 operations
225 DECLARE_ALIGNED(16, int16_t, qmat)[3][2][3][64]; //<qmat[qpi][is_inter][plane]
227 /* This table contains superblock_count * 16 entries. Each set of 16
228 * numbers corresponds to the fragment indexes 0..15 of the superblock.
229 * An entry will be -1 to indicate that no entry corresponds to that
231 int *superblock_fragments;
233 /* This is an array that indicates how a particular macroblock
235 unsigned char *macroblock_coding;
237 uint8_t edge_emu_buffer[9*2048]; //FIXME dynamic alloc
238 int8_t qscale_table[2048]; //FIXME dynamic alloc (width+15)/16
245 uint32_t huffman_table[80][32][2];
247 uint8_t filter_limit_values[64];
248 DECLARE_ALIGNED(8, int, bounding_values_array)[256+2];
251 /************************************************************************
252 * VP3 specific functions
253 ************************************************************************/
256 * This function sets up all of the various blocks mappings:
257 * superblocks <-> fragments, macroblocks <-> fragments,
258 * superblocks <-> macroblocks
260 * Returns 0 is successful; returns 1 if *anything* went wrong.
262 static int init_block_mapping(Vp3DecodeContext *s)
264 int sb_x, sb_y, plane;
267 for (plane = 0; plane < 3; plane++) {
268 int sb_width = plane ? s->c_superblock_width : s->y_superblock_width;
269 int sb_height = plane ? s->c_superblock_height : s->y_superblock_height;
270 int frag_width = s->fragment_width[!!plane];
271 int frag_height = s->fragment_height[!!plane];
273 for (sb_y = 0; sb_y < sb_height; sb_y++)
274 for (sb_x = 0; sb_x < sb_width; sb_x++)
275 for (i = 0; i < 16; i++) {
276 x = 4*sb_x + hilbert_offset[i][0];
277 y = 4*sb_y + hilbert_offset[i][1];
279 if (x < frag_width && y < frag_height)
280 s->superblock_fragments[j++] = s->fragment_start[plane] + y*frag_width + x;
282 s->superblock_fragments[j++] = -1;
286 return 0; /* successful path out */
290 * This function sets up the dequantization tables used for a particular
293 static void init_dequantizer(Vp3DecodeContext *s, int qpi)
295 int ac_scale_factor = s->coded_ac_scale_factor[s->qps[qpi]];
296 int dc_scale_factor = s->coded_dc_scale_factor[s->qps[qpi]];
297 int i, plane, inter, qri, bmi, bmj, qistart;
299 for(inter=0; inter<2; inter++){
300 for(plane=0; plane<3; plane++){
302 for(qri=0; qri<s->qr_count[inter][plane]; qri++){
303 sum+= s->qr_size[inter][plane][qri];
304 if(s->qps[qpi] <= sum)
307 qistart= sum - s->qr_size[inter][plane][qri];
308 bmi= s->qr_base[inter][plane][qri ];
309 bmj= s->qr_base[inter][plane][qri+1];
311 int coeff= ( 2*(sum -s->qps[qpi])*s->base_matrix[bmi][i]
312 - 2*(qistart-s->qps[qpi])*s->base_matrix[bmj][i]
313 + s->qr_size[inter][plane][qri])
314 / (2*s->qr_size[inter][plane][qri]);
316 int qmin= 8<<(inter + !i);
317 int qscale= i ? ac_scale_factor : dc_scale_factor;
319 s->qmat[qpi][inter][plane][s->dsp.idct_permutation[i]]= av_clip((qscale * coeff)/100 * 4, qmin, 4096);
321 // all DC coefficients use the same quant so as not to interfere with DC prediction
322 s->qmat[qpi][inter][plane][0] = s->qmat[0][inter][plane][0];
326 memset(s->qscale_table, (FFMAX(s->qmat[0][0][0][1], s->qmat[0][0][1][1])+8)/16, 512); //FIXME finetune
330 * This function initializes the loop filter boundary limits if the frame's
331 * quality index is different from the previous frame's.
333 * The filter_limit_values may not be larger than 127.
335 static void init_loop_filter(Vp3DecodeContext *s)
337 int *bounding_values= s->bounding_values_array+127;
342 filter_limit = s->filter_limit_values[s->qps[0]];
344 /* set up the bounding values */
345 memset(s->bounding_values_array, 0, 256 * sizeof(int));
346 for (x = 0; x < filter_limit; x++) {
347 bounding_values[-x] = -x;
348 bounding_values[x] = x;
350 for (x = value = filter_limit; x < 128 && value; x++, value--) {
351 bounding_values[ x] = value;
352 bounding_values[-x] = -value;
355 bounding_values[128] = value;
356 bounding_values[129] = bounding_values[130] = filter_limit * 0x02020202;
360 * This function unpacks all of the superblock/macroblock/fragment coding
361 * information from the bitstream.
363 static int unpack_superblocks(Vp3DecodeContext *s, GetBitContext *gb)
365 int superblock_starts[3] = { 0, s->u_superblock_start, s->v_superblock_start };
367 int current_superblock = 0;
369 int num_partial_superblocks = 0;
372 int current_fragment;
376 memset(s->superblock_coding, SB_FULLY_CODED, s->superblock_count);
380 /* unpack the list of partially-coded superblocks */
382 while (current_superblock < s->superblock_count && get_bits_left(gb) > 0) {
383 current_run = get_vlc2(gb,
384 s->superblock_run_length_vlc.table, 6, 2) + 1;
385 if (current_run == 34)
386 current_run += get_bits(gb, 12);
388 if (current_superblock + current_run > s->superblock_count) {
389 av_log(s->avctx, AV_LOG_ERROR, "Invalid partially coded superblock run length\n");
393 memset(s->superblock_coding + current_superblock, bit, current_run);
395 current_superblock += current_run;
397 num_partial_superblocks += current_run;
399 if (s->theora && current_run == MAXIMUM_LONG_BIT_RUN)
405 /* unpack the list of fully coded superblocks if any of the blocks were
406 * not marked as partially coded in the previous step */
407 if (num_partial_superblocks < s->superblock_count) {
408 int superblocks_decoded = 0;
410 current_superblock = 0;
412 while (superblocks_decoded < s->superblock_count - num_partial_superblocks
413 && get_bits_left(gb) > 0) {
414 current_run = get_vlc2(gb,
415 s->superblock_run_length_vlc.table, 6, 2) + 1;
416 if (current_run == 34)
417 current_run += get_bits(gb, 12);
419 for (j = 0; j < current_run; current_superblock++) {
420 if (current_superblock >= s->superblock_count) {
421 av_log(s->avctx, AV_LOG_ERROR, "Invalid fully coded superblock run length\n");
425 /* skip any superblocks already marked as partially coded */
426 if (s->superblock_coding[current_superblock] == SB_NOT_CODED) {
427 s->superblock_coding[current_superblock] = 2*bit;
431 superblocks_decoded += current_run;
433 if (s->theora && current_run == MAXIMUM_LONG_BIT_RUN)
440 /* if there were partial blocks, initialize bitstream for
441 * unpacking fragment codings */
442 if (num_partial_superblocks) {
446 /* toggle the bit because as soon as the first run length is
447 * fetched the bit will be toggled again */
452 /* figure out which fragments are coded; iterate through each
453 * superblock (all planes) */
454 s->total_num_coded_frags = 0;
455 memset(s->macroblock_coding, MODE_COPY, s->macroblock_count);
457 for (plane = 0; plane < 3; plane++) {
458 int sb_start = superblock_starts[plane];
459 int sb_end = sb_start + (plane ? s->c_superblock_count : s->y_superblock_count);
460 int num_coded_frags = 0;
462 for (i = sb_start; i < sb_end && get_bits_left(gb) > 0; i++) {
464 /* iterate through all 16 fragments in a superblock */
465 for (j = 0; j < 16; j++) {
467 /* if the fragment is in bounds, check its coding status */
468 current_fragment = s->superblock_fragments[i * 16 + j];
469 if (current_fragment != -1) {
470 int coded = s->superblock_coding[i];
472 if (s->superblock_coding[i] == SB_PARTIALLY_CODED) {
474 /* fragment may or may not be coded; this is the case
475 * that cares about the fragment coding runs */
476 if (current_run-- == 0) {
478 current_run = get_vlc2(gb,
479 s->fragment_run_length_vlc.table, 5, 2);
485 /* default mode; actual mode will be decoded in
487 s->all_fragments[current_fragment].coding_method =
489 s->coded_fragment_list[plane][num_coded_frags++] =
492 /* not coded; copy this fragment from the prior frame */
493 s->all_fragments[current_fragment].coding_method =
499 s->total_num_coded_frags += num_coded_frags;
500 for (i = 0; i < 64; i++)
501 s->num_coded_frags[plane][i] = num_coded_frags;
503 s->coded_fragment_list[plane+1] = s->coded_fragment_list[plane] + num_coded_frags;
509 * This function unpacks all the coding mode data for individual macroblocks
510 * from the bitstream.
512 static int unpack_modes(Vp3DecodeContext *s, GetBitContext *gb)
514 int i, j, k, sb_x, sb_y;
516 int current_macroblock;
517 int current_fragment;
519 int custom_mode_alphabet[CODING_MODE_COUNT];
524 for (i = 0; i < s->fragment_count; i++)
525 s->all_fragments[i].coding_method = MODE_INTRA;
529 /* fetch the mode coding scheme for this frame */
530 scheme = get_bits(gb, 3);
532 /* is it a custom coding scheme? */
534 for (i = 0; i < 8; i++)
535 custom_mode_alphabet[i] = MODE_INTER_NO_MV;
536 for (i = 0; i < 8; i++)
537 custom_mode_alphabet[get_bits(gb, 3)] = i;
538 alphabet = custom_mode_alphabet;
540 alphabet = ModeAlphabet[scheme-1];
542 /* iterate through all of the macroblocks that contain 1 or more
544 for (sb_y = 0; sb_y < s->y_superblock_height; sb_y++) {
545 for (sb_x = 0; sb_x < s->y_superblock_width; sb_x++) {
546 if (get_bits_left(gb) <= 0)
549 for (j = 0; j < 4; j++) {
550 int mb_x = 2*sb_x + (j>>1);
551 int mb_y = 2*sb_y + (((j>>1)+j)&1);
552 current_macroblock = mb_y * s->macroblock_width + mb_x;
554 if (mb_x >= s->macroblock_width || mb_y >= s->macroblock_height)
557 #define BLOCK_X (2*mb_x + (k&1))
558 #define BLOCK_Y (2*mb_y + (k>>1))
559 /* coding modes are only stored if the macroblock has at least one
560 * luma block coded, otherwise it must be INTER_NO_MV */
561 for (k = 0; k < 4; k++) {
562 current_fragment = BLOCK_Y*s->fragment_width[0] + BLOCK_X;
563 if (s->all_fragments[current_fragment].coding_method != MODE_COPY)
567 s->macroblock_coding[current_macroblock] = MODE_INTER_NO_MV;
571 /* mode 7 means get 3 bits for each coding mode */
573 coding_mode = get_bits(gb, 3);
575 coding_mode = alphabet
576 [get_vlc2(gb, s->mode_code_vlc.table, 3, 3)];
578 s->macroblock_coding[current_macroblock] = coding_mode;
579 for (k = 0; k < 4; k++) {
580 frag = s->all_fragments + BLOCK_Y*s->fragment_width[0] + BLOCK_X;
581 if (frag->coding_method != MODE_COPY)
582 frag->coding_method = coding_mode;
585 #define SET_CHROMA_MODES \
586 if (frag[s->fragment_start[1]].coding_method != MODE_COPY) \
587 frag[s->fragment_start[1]].coding_method = coding_mode;\
588 if (frag[s->fragment_start[2]].coding_method != MODE_COPY) \
589 frag[s->fragment_start[2]].coding_method = coding_mode;
591 if (s->chroma_y_shift) {
592 frag = s->all_fragments + mb_y*s->fragment_width[1] + mb_x;
594 } else if (s->chroma_x_shift) {
595 frag = s->all_fragments + 2*mb_y*s->fragment_width[1] + mb_x;
596 for (k = 0; k < 2; k++) {
598 frag += s->fragment_width[1];
601 for (k = 0; k < 4; k++) {
602 frag = s->all_fragments + BLOCK_Y*s->fragment_width[1] + BLOCK_X;
615 * This function unpacks all the motion vectors for the individual
616 * macroblocks from the bitstream.
618 static int unpack_vectors(Vp3DecodeContext *s, GetBitContext *gb)
620 int j, k, sb_x, sb_y;
624 int last_motion_x = 0;
625 int last_motion_y = 0;
626 int prior_last_motion_x = 0;
627 int prior_last_motion_y = 0;
628 int current_macroblock;
629 int current_fragment;
635 /* coding mode 0 is the VLC scheme; 1 is the fixed code scheme */
636 coding_mode = get_bits1(gb);
638 /* iterate through all of the macroblocks that contain 1 or more
640 for (sb_y = 0; sb_y < s->y_superblock_height; sb_y++) {
641 for (sb_x = 0; sb_x < s->y_superblock_width; sb_x++) {
642 if (get_bits_left(gb) <= 0)
645 for (j = 0; j < 4; j++) {
646 int mb_x = 2*sb_x + (j>>1);
647 int mb_y = 2*sb_y + (((j>>1)+j)&1);
648 current_macroblock = mb_y * s->macroblock_width + mb_x;
650 if (mb_x >= s->macroblock_width || mb_y >= s->macroblock_height ||
651 (s->macroblock_coding[current_macroblock] == MODE_COPY))
654 switch (s->macroblock_coding[current_macroblock]) {
656 case MODE_INTER_PLUS_MV:
658 /* all 6 fragments use the same motion vector */
659 if (coding_mode == 0) {
660 motion_x[0] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
661 motion_y[0] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
663 motion_x[0] = fixed_motion_vector_table[get_bits(gb, 6)];
664 motion_y[0] = fixed_motion_vector_table[get_bits(gb, 6)];
667 /* vector maintenance, only on MODE_INTER_PLUS_MV */
668 if (s->macroblock_coding[current_macroblock] ==
669 MODE_INTER_PLUS_MV) {
670 prior_last_motion_x = last_motion_x;
671 prior_last_motion_y = last_motion_y;
672 last_motion_x = motion_x[0];
673 last_motion_y = motion_y[0];
677 case MODE_INTER_FOURMV:
678 /* vector maintenance */
679 prior_last_motion_x = last_motion_x;
680 prior_last_motion_y = last_motion_y;
682 /* fetch 4 vectors from the bitstream, one for each
683 * Y fragment, then average for the C fragment vectors */
684 for (k = 0; k < 4; k++) {
685 current_fragment = BLOCK_Y*s->fragment_width[0] + BLOCK_X;
686 if (s->all_fragments[current_fragment].coding_method != MODE_COPY) {
687 if (coding_mode == 0) {
688 motion_x[k] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
689 motion_y[k] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
691 motion_x[k] = fixed_motion_vector_table[get_bits(gb, 6)];
692 motion_y[k] = fixed_motion_vector_table[get_bits(gb, 6)];
694 last_motion_x = motion_x[k];
695 last_motion_y = motion_y[k];
703 case MODE_INTER_LAST_MV:
704 /* all 6 fragments use the last motion vector */
705 motion_x[0] = last_motion_x;
706 motion_y[0] = last_motion_y;
708 /* no vector maintenance (last vector remains the
712 case MODE_INTER_PRIOR_LAST:
713 /* all 6 fragments use the motion vector prior to the
714 * last motion vector */
715 motion_x[0] = prior_last_motion_x;
716 motion_y[0] = prior_last_motion_y;
718 /* vector maintenance */
719 prior_last_motion_x = last_motion_x;
720 prior_last_motion_y = last_motion_y;
721 last_motion_x = motion_x[0];
722 last_motion_y = motion_y[0];
726 /* covers intra, inter without MV, golden without MV */
730 /* no vector maintenance */
734 /* assign the motion vectors to the correct fragments */
735 for (k = 0; k < 4; k++) {
737 BLOCK_Y*s->fragment_width[0] + BLOCK_X;
738 if (s->macroblock_coding[current_macroblock] == MODE_INTER_FOURMV) {
739 s->motion_val[0][current_fragment][0] = motion_x[k];
740 s->motion_val[0][current_fragment][1] = motion_y[k];
742 s->motion_val[0][current_fragment][0] = motion_x[0];
743 s->motion_val[0][current_fragment][1] = motion_y[0];
747 if (s->chroma_y_shift) {
748 if (s->macroblock_coding[current_macroblock] == MODE_INTER_FOURMV) {
749 motion_x[0] = RSHIFT(motion_x[0] + motion_x[1] + motion_x[2] + motion_x[3], 2);
750 motion_y[0] = RSHIFT(motion_y[0] + motion_y[1] + motion_y[2] + motion_y[3], 2);
752 motion_x[0] = (motion_x[0]>>1) | (motion_x[0]&1);
753 motion_y[0] = (motion_y[0]>>1) | (motion_y[0]&1);
754 frag = mb_y*s->fragment_width[1] + mb_x;
755 s->motion_val[1][frag][0] = motion_x[0];
756 s->motion_val[1][frag][1] = motion_y[0];
757 } else if (s->chroma_x_shift) {
758 if (s->macroblock_coding[current_macroblock] == MODE_INTER_FOURMV) {
759 motion_x[0] = RSHIFT(motion_x[0] + motion_x[1], 1);
760 motion_y[0] = RSHIFT(motion_y[0] + motion_y[1], 1);
761 motion_x[1] = RSHIFT(motion_x[2] + motion_x[3], 1);
762 motion_y[1] = RSHIFT(motion_y[2] + motion_y[3], 1);
764 motion_x[1] = motion_x[0];
765 motion_y[1] = motion_y[0];
767 motion_x[0] = (motion_x[0]>>1) | (motion_x[0]&1);
768 motion_x[1] = (motion_x[1]>>1) | (motion_x[1]&1);
770 frag = 2*mb_y*s->fragment_width[1] + mb_x;
771 for (k = 0; k < 2; k++) {
772 s->motion_val[1][frag][0] = motion_x[k];
773 s->motion_val[1][frag][1] = motion_y[k];
774 frag += s->fragment_width[1];
777 for (k = 0; k < 4; k++) {
778 frag = BLOCK_Y*s->fragment_width[1] + BLOCK_X;
779 if (s->macroblock_coding[current_macroblock] == MODE_INTER_FOURMV) {
780 s->motion_val[1][frag][0] = motion_x[k];
781 s->motion_val[1][frag][1] = motion_y[k];
783 s->motion_val[1][frag][0] = motion_x[0];
784 s->motion_val[1][frag][1] = motion_y[0];
795 static int unpack_block_qpis(Vp3DecodeContext *s, GetBitContext *gb)
797 int qpi, i, j, bit, run_length, blocks_decoded, num_blocks_at_qpi;
798 int num_blocks = s->total_num_coded_frags;
800 for (qpi = 0; qpi < s->nqps-1 && num_blocks > 0; qpi++) {
801 i = blocks_decoded = num_blocks_at_qpi = 0;
806 run_length = get_vlc2(gb, s->superblock_run_length_vlc.table, 6, 2) + 1;
807 if (run_length == 34)
808 run_length += get_bits(gb, 12);
809 blocks_decoded += run_length;
812 num_blocks_at_qpi += run_length;
814 for (j = 0; j < run_length; i++) {
815 if (i >= s->total_num_coded_frags)
818 if (s->all_fragments[s->coded_fragment_list[0][i]].qpi == qpi) {
819 s->all_fragments[s->coded_fragment_list[0][i]].qpi += bit;
824 if (run_length == MAXIMUM_LONG_BIT_RUN)
828 } while (blocks_decoded < num_blocks && get_bits_left(gb) > 0);
830 num_blocks -= num_blocks_at_qpi;
837 * This function is called by unpack_dct_coeffs() to extract the VLCs from
838 * the bitstream. The VLCs encode tokens which are used to unpack DCT
839 * data. This function unpacks all the VLCs for either the Y plane or both
840 * C planes, and is called for DC coefficients or different AC coefficient
841 * levels (since different coefficient types require different VLC tables.
843 * This function returns a residual eob run. E.g, if a particular token gave
844 * instructions to EOB the next 5 fragments and there were only 2 fragments
845 * left in the current fragment range, 3 would be returned so that it could
846 * be passed into the next call to this same function.
848 static int unpack_vlcs(Vp3DecodeContext *s, GetBitContext *gb,
849 VLC *table, int coeff_index,
860 int num_coeffs = s->num_coded_frags[plane][coeff_index];
861 int16_t *dct_tokens = s->dct_tokens[plane][coeff_index];
863 /* local references to structure members to avoid repeated deferences */
864 int *coded_fragment_list = s->coded_fragment_list[plane];
865 Vp3Fragment *all_fragments = s->all_fragments;
866 VLC_TYPE (*vlc_table)[2] = table->table;
869 av_log(s->avctx, AV_LOG_ERROR, "Invalid number of coefficents at level %d\n", coeff_index);
871 if (eob_run > num_coeffs) {
872 coeff_i = blocks_ended = num_coeffs;
873 eob_run -= num_coeffs;
875 coeff_i = blocks_ended = eob_run;
879 // insert fake EOB token to cover the split between planes or zzi
881 dct_tokens[j++] = blocks_ended << 2;
883 while (coeff_i < num_coeffs && get_bits_left(gb) > 0) {
884 /* decode a VLC into a token */
885 token = get_vlc2(gb, vlc_table, 11, 3);
886 /* use the token to get a zero run, a coefficient, and an eob run */
888 eob_run = eob_run_base[token];
889 if (eob_run_get_bits[token])
890 eob_run += get_bits(gb, eob_run_get_bits[token]);
892 // record only the number of blocks ended in this plane,
893 // any spill will be recorded in the next plane.
894 if (eob_run > num_coeffs - coeff_i) {
895 dct_tokens[j++] = TOKEN_EOB(num_coeffs - coeff_i);
896 blocks_ended += num_coeffs - coeff_i;
897 eob_run -= num_coeffs - coeff_i;
898 coeff_i = num_coeffs;
900 dct_tokens[j++] = TOKEN_EOB(eob_run);
901 blocks_ended += eob_run;
906 bits_to_get = coeff_get_bits[token];
908 bits_to_get = get_bits(gb, bits_to_get);
909 coeff = coeff_tables[token][bits_to_get];
911 zero_run = zero_run_base[token];
912 if (zero_run_get_bits[token])
913 zero_run += get_bits(gb, zero_run_get_bits[token]);
916 dct_tokens[j++] = TOKEN_ZERO_RUN(coeff, zero_run);
918 // Save DC into the fragment structure. DC prediction is
919 // done in raster order, so the actual DC can't be in with
920 // other tokens. We still need the token in dct_tokens[]
921 // however, or else the structure collapses on itself.
923 all_fragments[coded_fragment_list[coeff_i]].dc = coeff;
925 dct_tokens[j++] = TOKEN_COEFF(coeff);
928 if (coeff_index + zero_run > 64) {
929 av_log(s->avctx, AV_LOG_DEBUG, "Invalid zero run of %d with"
930 " %d coeffs left\n", zero_run, 64-coeff_index);
931 zero_run = 64 - coeff_index;
934 // zero runs code multiple coefficients,
935 // so don't try to decode coeffs for those higher levels
936 for (i = coeff_index+1; i <= coeff_index+zero_run; i++)
937 s->num_coded_frags[plane][i]--;
942 if (blocks_ended > s->num_coded_frags[plane][coeff_index])
943 av_log(s->avctx, AV_LOG_ERROR, "More blocks ended than coded!\n");
945 // decrement the number of blocks that have higher coeffecients for each
946 // EOB run at this level
948 for (i = coeff_index+1; i < 64; i++)
949 s->num_coded_frags[plane][i] -= blocks_ended;
951 // setup the next buffer
953 s->dct_tokens[plane+1][coeff_index] = dct_tokens + j;
954 else if (coeff_index < 63)
955 s->dct_tokens[0][coeff_index+1] = dct_tokens + j;
960 static void reverse_dc_prediction(Vp3DecodeContext *s,
963 int fragment_height);
965 * This function unpacks all of the DCT coefficient data from the
968 static int unpack_dct_coeffs(Vp3DecodeContext *s, GetBitContext *gb)
975 int residual_eob_run = 0;
979 s->dct_tokens[0][0] = s->dct_tokens_base;
981 /* fetch the DC table indexes */
982 dc_y_table = get_bits(gb, 4);
983 dc_c_table = get_bits(gb, 4);
985 /* unpack the Y plane DC coefficients */
986 residual_eob_run = unpack_vlcs(s, gb, &s->dc_vlc[dc_y_table], 0,
987 0, residual_eob_run);
989 /* reverse prediction of the Y-plane DC coefficients */
990 reverse_dc_prediction(s, 0, s->fragment_width[0], s->fragment_height[0]);
992 /* unpack the C plane DC coefficients */
993 residual_eob_run = unpack_vlcs(s, gb, &s->dc_vlc[dc_c_table], 0,
994 1, residual_eob_run);
995 residual_eob_run = unpack_vlcs(s, gb, &s->dc_vlc[dc_c_table], 0,
996 2, residual_eob_run);
998 /* reverse prediction of the C-plane DC coefficients */
999 if (!(s->avctx->flags & CODEC_FLAG_GRAY))
1001 reverse_dc_prediction(s, s->fragment_start[1],
1002 s->fragment_width[1], s->fragment_height[1]);
1003 reverse_dc_prediction(s, s->fragment_start[2],
1004 s->fragment_width[1], s->fragment_height[1]);
1007 /* fetch the AC table indexes */
1008 ac_y_table = get_bits(gb, 4);
1009 ac_c_table = get_bits(gb, 4);
1011 /* build tables of AC VLC tables */
1012 for (i = 1; i <= 5; i++) {
1013 y_tables[i] = &s->ac_vlc_1[ac_y_table];
1014 c_tables[i] = &s->ac_vlc_1[ac_c_table];
1016 for (i = 6; i <= 14; i++) {
1017 y_tables[i] = &s->ac_vlc_2[ac_y_table];
1018 c_tables[i] = &s->ac_vlc_2[ac_c_table];
1020 for (i = 15; i <= 27; i++) {
1021 y_tables[i] = &s->ac_vlc_3[ac_y_table];
1022 c_tables[i] = &s->ac_vlc_3[ac_c_table];
1024 for (i = 28; i <= 63; i++) {
1025 y_tables[i] = &s->ac_vlc_4[ac_y_table];
1026 c_tables[i] = &s->ac_vlc_4[ac_c_table];
1029 /* decode all AC coefficents */
1030 for (i = 1; i <= 63; i++) {
1031 residual_eob_run = unpack_vlcs(s, gb, y_tables[i], i,
1032 0, residual_eob_run);
1034 residual_eob_run = unpack_vlcs(s, gb, c_tables[i], i,
1035 1, residual_eob_run);
1036 residual_eob_run = unpack_vlcs(s, gb, c_tables[i], i,
1037 2, residual_eob_run);
1044 * This function reverses the DC prediction for each coded fragment in
1045 * the frame. Much of this function is adapted directly from the original
1048 #define COMPATIBLE_FRAME(x) \
1049 (compatible_frame[s->all_fragments[x].coding_method] == current_frame_type)
1050 #define DC_COEFF(u) s->all_fragments[u].dc
1052 static void reverse_dc_prediction(Vp3DecodeContext *s,
1055 int fragment_height)
1064 int i = first_fragment;
1068 /* DC values for the left, up-left, up, and up-right fragments */
1069 int vl, vul, vu, vur;
1071 /* indexes for the left, up-left, up, and up-right fragments */
1075 * The 6 fields mean:
1076 * 0: up-left multiplier
1078 * 2: up-right multiplier
1079 * 3: left multiplier
1081 static const int predictor_transform[16][4] = {
1083 { 0, 0, 0,128}, // PL
1084 { 0, 0,128, 0}, // PUR
1085 { 0, 0, 53, 75}, // PUR|PL
1086 { 0,128, 0, 0}, // PU
1087 { 0, 64, 0, 64}, // PU|PL
1088 { 0,128, 0, 0}, // PU|PUR
1089 { 0, 0, 53, 75}, // PU|PUR|PL
1090 {128, 0, 0, 0}, // PUL
1091 { 0, 0, 0,128}, // PUL|PL
1092 { 64, 0, 64, 0}, // PUL|PUR
1093 { 0, 0, 53, 75}, // PUL|PUR|PL
1094 { 0,128, 0, 0}, // PUL|PU
1095 {-104,116, 0,116}, // PUL|PU|PL
1096 { 24, 80, 24, 0}, // PUL|PU|PUR
1097 {-104,116, 0,116} // PUL|PU|PUR|PL
1100 /* This table shows which types of blocks can use other blocks for
1101 * prediction. For example, INTRA is the only mode in this table to
1102 * have a frame number of 0. That means INTRA blocks can only predict
1103 * from other INTRA blocks. There are 2 golden frame coding types;
1104 * blocks encoding in these modes can only predict from other blocks
1105 * that were encoded with these 1 of these 2 modes. */
1106 static const unsigned char compatible_frame[9] = {
1107 1, /* MODE_INTER_NO_MV */
1109 1, /* MODE_INTER_PLUS_MV */
1110 1, /* MODE_INTER_LAST_MV */
1111 1, /* MODE_INTER_PRIOR_MV */
1112 2, /* MODE_USING_GOLDEN */
1113 2, /* MODE_GOLDEN_MV */
1114 1, /* MODE_INTER_FOUR_MV */
1117 int current_frame_type;
1119 /* there is a last DC predictor for each of the 3 frame types */
1124 vul = vu = vur = vl = 0;
1125 last_dc[0] = last_dc[1] = last_dc[2] = 0;
1127 /* for each fragment row... */
1128 for (y = 0; y < fragment_height; y++) {
1130 /* for each fragment in a row... */
1131 for (x = 0; x < fragment_width; x++, i++) {
1133 /* reverse prediction if this block was coded */
1134 if (s->all_fragments[i].coding_method != MODE_COPY) {
1136 current_frame_type =
1137 compatible_frame[s->all_fragments[i].coding_method];
1143 if(COMPATIBLE_FRAME(l))
1147 u= i-fragment_width;
1149 if(COMPATIBLE_FRAME(u))
1152 ul= i-fragment_width-1;
1154 if(COMPATIBLE_FRAME(ul))
1157 if(x + 1 < fragment_width){
1158 ur= i-fragment_width+1;
1160 if(COMPATIBLE_FRAME(ur))
1165 if (transform == 0) {
1167 /* if there were no fragments to predict from, use last
1169 predicted_dc = last_dc[current_frame_type];
1172 /* apply the appropriate predictor transform */
1174 (predictor_transform[transform][0] * vul) +
1175 (predictor_transform[transform][1] * vu) +
1176 (predictor_transform[transform][2] * vur) +
1177 (predictor_transform[transform][3] * vl);
1179 predicted_dc /= 128;
1181 /* check for outranging on the [ul u l] and
1182 * [ul u ur l] predictors */
1183 if ((transform == 15) || (transform == 13)) {
1184 if (FFABS(predicted_dc - vu) > 128)
1186 else if (FFABS(predicted_dc - vl) > 128)
1188 else if (FFABS(predicted_dc - vul) > 128)
1193 /* at long last, apply the predictor */
1194 DC_COEFF(i) += predicted_dc;
1196 last_dc[current_frame_type] = DC_COEFF(i);
1202 static void apply_loop_filter(Vp3DecodeContext *s, int plane, int ystart, int yend)
1205 int *bounding_values= s->bounding_values_array+127;
1207 int width = s->fragment_width[!!plane];
1208 int height = s->fragment_height[!!plane];
1209 int fragment = s->fragment_start [plane] + ystart * width;
1210 int stride = s->current_frame.linesize[plane];
1211 uint8_t *plane_data = s->current_frame.data [plane];
1212 if (!s->flipped_image) stride = -stride;
1213 plane_data += s->data_offset[plane] + 8*ystart*stride;
1215 for (y = ystart; y < yend; y++) {
1217 for (x = 0; x < width; x++) {
1218 /* This code basically just deblocks on the edges of coded blocks.
1219 * However, it has to be much more complicated because of the
1220 * braindamaged deblock ordering used in VP3/Theora. Order matters
1221 * because some pixels get filtered twice. */
1222 if( s->all_fragments[fragment].coding_method != MODE_COPY )
1224 /* do not perform left edge filter for left columns frags */
1226 s->dsp.vp3_h_loop_filter(
1228 stride, bounding_values);
1231 /* do not perform top edge filter for top row fragments */
1233 s->dsp.vp3_v_loop_filter(
1235 stride, bounding_values);
1238 /* do not perform right edge filter for right column
1239 * fragments or if right fragment neighbor is also coded
1240 * in this frame (it will be filtered in next iteration) */
1241 if ((x < width - 1) &&
1242 (s->all_fragments[fragment + 1].coding_method == MODE_COPY)) {
1243 s->dsp.vp3_h_loop_filter(
1244 plane_data + 8*x + 8,
1245 stride, bounding_values);
1248 /* do not perform bottom edge filter for bottom row
1249 * fragments or if bottom fragment neighbor is also coded
1250 * in this frame (it will be filtered in the next row) */
1251 if ((y < height - 1) &&
1252 (s->all_fragments[fragment + width].coding_method == MODE_COPY)) {
1253 s->dsp.vp3_v_loop_filter(
1254 plane_data + 8*x + 8*stride,
1255 stride, bounding_values);
1261 plane_data += 8*stride;
1266 * Pulls DCT tokens from the 64 levels to decode and dequant the coefficients
1267 * for the next block in coding order
1269 static inline int vp3_dequant(Vp3DecodeContext *s, Vp3Fragment *frag,
1270 int plane, int inter, DCTELEM block[64])
1272 int16_t *dequantizer = s->qmat[frag->qpi][inter][plane];
1273 uint8_t *perm = s->scantable.permutated;
1277 int token = *s->dct_tokens[plane][i];
1278 switch (token & 3) {
1280 if (--token < 4) // 0-3 are token types, so the EOB run must now be 0
1281 s->dct_tokens[plane][i]++;
1283 *s->dct_tokens[plane][i] = token & ~3;
1286 s->dct_tokens[plane][i]++;
1287 i += (token >> 2) & 0x7f;
1288 block[perm[i]] = (token >> 9) * dequantizer[perm[i]];
1292 block[perm[i]] = (token >> 2) * dequantizer[perm[i]];
1293 s->dct_tokens[plane][i++]++;
1295 default: // shouldn't happen
1300 // the actual DC+prediction is in the fragment structure
1301 block[0] = frag->dc * s->qmat[0][inter][plane][0];
1306 * called when all pixels up to row y are complete
1308 static void vp3_draw_horiz_band(Vp3DecodeContext *s, int y)
1313 if(s->avctx->draw_horiz_band==NULL)
1316 h= y - s->last_slice_end;
1319 if (!s->flipped_image) {
1321 h -= s->height - s->avctx->height; // account for non-mod16
1322 y = s->height - y - h;
1326 offset[0] = s->current_frame.linesize[0]*y;
1327 offset[1] = s->current_frame.linesize[1]*cy;
1328 offset[2] = s->current_frame.linesize[2]*cy;
1332 s->avctx->draw_horiz_band(s->avctx, &s->current_frame, offset, y, 3, h);
1333 s->last_slice_end= y + h;
1337 * Perform the final rendering for a particular slice of data.
1338 * The slice number ranges from 0..(c_superblock_height - 1).
1340 static void render_slice(Vp3DecodeContext *s, int slice)
1343 LOCAL_ALIGNED_16(DCTELEM, block, [64]);
1344 int motion_x = 0xdeadbeef, motion_y = 0xdeadbeef;
1345 int motion_halfpel_index;
1346 uint8_t *motion_source;
1347 int plane, first_pixel;
1349 if (slice >= s->c_superblock_height)
1352 for (plane = 0; plane < 3; plane++) {
1353 uint8_t *output_plane = s->current_frame.data [plane] + s->data_offset[plane];
1354 uint8_t * last_plane = s-> last_frame.data [plane] + s->data_offset[plane];
1355 uint8_t *golden_plane = s-> golden_frame.data [plane] + s->data_offset[plane];
1356 int stride = s->current_frame.linesize[plane];
1357 int plane_width = s->width >> (plane && s->chroma_x_shift);
1358 int plane_height = s->height >> (plane && s->chroma_y_shift);
1359 int8_t (*motion_val)[2] = s->motion_val[!!plane];
1361 int sb_x, sb_y = slice << (!plane && s->chroma_y_shift);
1362 int slice_height = sb_y + 1 + (!plane && s->chroma_y_shift);
1363 int slice_width = plane ? s->c_superblock_width : s->y_superblock_width;
1365 int fragment_width = s->fragment_width[!!plane];
1366 int fragment_height = s->fragment_height[!!plane];
1367 int fragment_start = s->fragment_start[plane];
1369 if (!s->flipped_image) stride = -stride;
1370 if (CONFIG_GRAY && plane && (s->avctx->flags & CODEC_FLAG_GRAY))
1374 if(FFABS(stride) > 2048)
1375 return; //various tables are fixed size
1377 /* for each superblock row in the slice (both of them)... */
1378 for (; sb_y < slice_height; sb_y++) {
1380 /* for each superblock in a row... */
1381 for (sb_x = 0; sb_x < slice_width; sb_x++) {
1383 /* for each block in a superblock... */
1384 for (j = 0; j < 16; j++) {
1385 x = 4*sb_x + hilbert_offset[j][0];
1386 y = 4*sb_y + hilbert_offset[j][1];
1388 i = fragment_start + y*fragment_width + x;
1391 if (x >= fragment_width || y >= fragment_height)
1394 first_pixel = 8*y*stride + 8*x;
1396 /* transform if this block was coded */
1397 if (s->all_fragments[i].coding_method != MODE_COPY) {
1398 if ((s->all_fragments[i].coding_method == MODE_USING_GOLDEN) ||
1399 (s->all_fragments[i].coding_method == MODE_GOLDEN_MV))
1400 motion_source= golden_plane;
1402 motion_source= last_plane;
1404 motion_source += first_pixel;
1405 motion_halfpel_index = 0;
1407 /* sort out the motion vector if this fragment is coded
1408 * using a motion vector method */
1409 if ((s->all_fragments[i].coding_method > MODE_INTRA) &&
1410 (s->all_fragments[i].coding_method != MODE_USING_GOLDEN)) {
1412 motion_x = motion_val[y*fragment_width + x][0];
1413 motion_y = motion_val[y*fragment_width + x][1];
1415 src_x= (motion_x>>1) + 8*x;
1416 src_y= (motion_y>>1) + 8*y;
1418 motion_halfpel_index = motion_x & 0x01;
1419 motion_source += (motion_x >> 1);
1421 motion_halfpel_index |= (motion_y & 0x01) << 1;
1422 motion_source += ((motion_y >> 1) * stride);
1424 if(src_x<0 || src_y<0 || src_x + 9 >= plane_width || src_y + 9 >= plane_height){
1425 uint8_t *temp= s->edge_emu_buffer;
1426 if(stride<0) temp -= 9*stride;
1427 else temp += 9*stride;
1429 ff_emulated_edge_mc(temp, motion_source, stride, 9, 9, src_x, src_y, plane_width, plane_height);
1430 motion_source= temp;
1435 /* first, take care of copying a block from either the
1436 * previous or the golden frame */
1437 if (s->all_fragments[i].coding_method != MODE_INTRA) {
1438 /* Note, it is possible to implement all MC cases with
1439 put_no_rnd_pixels_l2 which would look more like the
1440 VP3 source but this would be slower as
1441 put_no_rnd_pixels_tab is better optimzed */
1442 if(motion_halfpel_index != 3){
1443 s->dsp.put_no_rnd_pixels_tab[1][motion_halfpel_index](
1444 output_plane + first_pixel,
1445 motion_source, stride, 8);
1447 int d= (motion_x ^ motion_y)>>31; // d is 0 if motion_x and _y have the same sign, else -1
1448 s->dsp.put_no_rnd_pixels_l2[1](
1449 output_plane + first_pixel,
1451 motion_source + stride + 1 + d,
1456 s->dsp.clear_block(block);
1458 /* invert DCT and place (or add) in final output */
1460 if (s->all_fragments[i].coding_method == MODE_INTRA) {
1461 vp3_dequant(s, s->all_fragments + i, plane, 0, block);
1462 if(s->avctx->idct_algo!=FF_IDCT_VP3)
1465 output_plane + first_pixel,
1469 if (vp3_dequant(s, s->all_fragments + i, plane, 1, block)) {
1471 output_plane + first_pixel,
1475 s->dsp.vp3_idct_dc_add(output_plane + first_pixel, stride, block);
1480 /* copy directly from the previous frame */
1481 s->dsp.put_pixels_tab[1][0](
1482 output_plane + first_pixel,
1483 last_plane + first_pixel,
1490 // Filter up to the last row in the superblock row
1491 apply_loop_filter(s, plane, 4*sb_y - !!sb_y, FFMIN(4*sb_y+3, fragment_height-1));
1495 /* this looks like a good place for slice dispatch... */
1497 * if (slice == s->macroblock_height - 1)
1498 * dispatch (both last slice & 2nd-to-last slice);
1499 * else if (slice > 0)
1500 * dispatch (slice - 1);
1503 vp3_draw_horiz_band(s, FFMIN(64*slice + 64-16, s->height-16));
1507 * This is the ffmpeg/libavcodec API init function.
1509 static av_cold int vp3_decode_init(AVCodecContext *avctx)
1511 Vp3DecodeContext *s = avctx->priv_data;
1512 int i, inter, plane;
1515 int y_fragment_count, c_fragment_count;
1517 if (avctx->codec_tag == MKTAG('V','P','3','0'))
1523 s->width = FFALIGN(avctx->width, 16);
1524 s->height = FFALIGN(avctx->height, 16);
1525 if (avctx->pix_fmt == PIX_FMT_NONE)
1526 avctx->pix_fmt = PIX_FMT_YUV420P;
1527 avctx->chroma_sample_location = AVCHROMA_LOC_CENTER;
1528 if(avctx->idct_algo==FF_IDCT_AUTO)
1529 avctx->idct_algo=FF_IDCT_VP3;
1530 dsputil_init(&s->dsp, avctx);
1532 ff_init_scantable(s->dsp.idct_permutation, &s->scantable, ff_zigzag_direct);
1534 /* initialize to an impossible value which will force a recalculation
1535 * in the first frame decode */
1536 for (i = 0; i < 3; i++)
1539 avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_x_shift, &s->chroma_y_shift);
1541 s->y_superblock_width = (s->width + 31) / 32;
1542 s->y_superblock_height = (s->height + 31) / 32;
1543 s->y_superblock_count = s->y_superblock_width * s->y_superblock_height;
1545 /* work out the dimensions for the C planes */
1546 c_width = s->width >> s->chroma_x_shift;
1547 c_height = s->height >> s->chroma_y_shift;
1548 s->c_superblock_width = (c_width + 31) / 32;
1549 s->c_superblock_height = (c_height + 31) / 32;
1550 s->c_superblock_count = s->c_superblock_width * s->c_superblock_height;
1552 s->superblock_count = s->y_superblock_count + (s->c_superblock_count * 2);
1553 s->u_superblock_start = s->y_superblock_count;
1554 s->v_superblock_start = s->u_superblock_start + s->c_superblock_count;
1555 s->superblock_coding = av_malloc(s->superblock_count);
1557 s->macroblock_width = (s->width + 15) / 16;
1558 s->macroblock_height = (s->height + 15) / 16;
1559 s->macroblock_count = s->macroblock_width * s->macroblock_height;
1561 s->fragment_width[0] = s->width / FRAGMENT_PIXELS;
1562 s->fragment_height[0] = s->height / FRAGMENT_PIXELS;
1563 s->fragment_width[1] = s->fragment_width[0] >> s->chroma_x_shift;
1564 s->fragment_height[1] = s->fragment_height[0] >> s->chroma_y_shift;
1566 /* fragment count covers all 8x8 blocks for all 3 planes */
1567 y_fragment_count = s->fragment_width[0] * s->fragment_height[0];
1568 c_fragment_count = s->fragment_width[1] * s->fragment_height[1];
1569 s->fragment_count = y_fragment_count + 2*c_fragment_count;
1570 s->fragment_start[1] = y_fragment_count;
1571 s->fragment_start[2] = y_fragment_count + c_fragment_count;
1573 s->all_fragments = av_malloc(s->fragment_count * sizeof(Vp3Fragment));
1574 s->coded_fragment_list[0] = av_malloc(s->fragment_count * sizeof(int));
1575 s->dct_tokens_base = av_malloc(64*s->fragment_count * sizeof(*s->dct_tokens_base));
1576 s->motion_val[0] = av_malloc(y_fragment_count * sizeof(*s->motion_val[0]));
1577 s->motion_val[1] = av_malloc(c_fragment_count * sizeof(*s->motion_val[1]));
1579 if (!s->superblock_coding || !s->all_fragments || !s->dct_tokens_base ||
1580 !s->coded_fragment_list[0] || !s->motion_val[0] || !s->motion_val[1]) {
1581 vp3_decode_end(avctx);
1585 if (!s->theora_tables)
1587 for (i = 0; i < 64; i++) {
1588 s->coded_dc_scale_factor[i] = vp31_dc_scale_factor[i];
1589 s->coded_ac_scale_factor[i] = vp31_ac_scale_factor[i];
1590 s->base_matrix[0][i] = vp31_intra_y_dequant[i];
1591 s->base_matrix[1][i] = vp31_intra_c_dequant[i];
1592 s->base_matrix[2][i] = vp31_inter_dequant[i];
1593 s->filter_limit_values[i] = vp31_filter_limit_values[i];
1596 for(inter=0; inter<2; inter++){
1597 for(plane=0; plane<3; plane++){
1598 s->qr_count[inter][plane]= 1;
1599 s->qr_size [inter][plane][0]= 63;
1600 s->qr_base [inter][plane][0]=
1601 s->qr_base [inter][plane][1]= 2*inter + (!!plane)*!inter;
1605 /* init VLC tables */
1606 for (i = 0; i < 16; i++) {
1609 init_vlc(&s->dc_vlc[i], 11, 32,
1610 &dc_bias[i][0][1], 4, 2,
1611 &dc_bias[i][0][0], 4, 2, 0);
1613 /* group 1 AC histograms */
1614 init_vlc(&s->ac_vlc_1[i], 11, 32,
1615 &ac_bias_0[i][0][1], 4, 2,
1616 &ac_bias_0[i][0][0], 4, 2, 0);
1618 /* group 2 AC histograms */
1619 init_vlc(&s->ac_vlc_2[i], 11, 32,
1620 &ac_bias_1[i][0][1], 4, 2,
1621 &ac_bias_1[i][0][0], 4, 2, 0);
1623 /* group 3 AC histograms */
1624 init_vlc(&s->ac_vlc_3[i], 11, 32,
1625 &ac_bias_2[i][0][1], 4, 2,
1626 &ac_bias_2[i][0][0], 4, 2, 0);
1628 /* group 4 AC histograms */
1629 init_vlc(&s->ac_vlc_4[i], 11, 32,
1630 &ac_bias_3[i][0][1], 4, 2,
1631 &ac_bias_3[i][0][0], 4, 2, 0);
1635 for (i = 0; i < 16; i++) {
1637 if (init_vlc(&s->dc_vlc[i], 11, 32,
1638 &s->huffman_table[i][0][1], 8, 4,
1639 &s->huffman_table[i][0][0], 8, 4, 0) < 0)
1642 /* group 1 AC histograms */
1643 if (init_vlc(&s->ac_vlc_1[i], 11, 32,
1644 &s->huffman_table[i+16][0][1], 8, 4,
1645 &s->huffman_table[i+16][0][0], 8, 4, 0) < 0)
1648 /* group 2 AC histograms */
1649 if (init_vlc(&s->ac_vlc_2[i], 11, 32,
1650 &s->huffman_table[i+16*2][0][1], 8, 4,
1651 &s->huffman_table[i+16*2][0][0], 8, 4, 0) < 0)
1654 /* group 3 AC histograms */
1655 if (init_vlc(&s->ac_vlc_3[i], 11, 32,
1656 &s->huffman_table[i+16*3][0][1], 8, 4,
1657 &s->huffman_table[i+16*3][0][0], 8, 4, 0) < 0)
1660 /* group 4 AC histograms */
1661 if (init_vlc(&s->ac_vlc_4[i], 11, 32,
1662 &s->huffman_table[i+16*4][0][1], 8, 4,
1663 &s->huffman_table[i+16*4][0][0], 8, 4, 0) < 0)
1668 init_vlc(&s->superblock_run_length_vlc, 6, 34,
1669 &superblock_run_length_vlc_table[0][1], 4, 2,
1670 &superblock_run_length_vlc_table[0][0], 4, 2, 0);
1672 init_vlc(&s->fragment_run_length_vlc, 5, 30,
1673 &fragment_run_length_vlc_table[0][1], 4, 2,
1674 &fragment_run_length_vlc_table[0][0], 4, 2, 0);
1676 init_vlc(&s->mode_code_vlc, 3, 8,
1677 &mode_code_vlc_table[0][1], 2, 1,
1678 &mode_code_vlc_table[0][0], 2, 1, 0);
1680 init_vlc(&s->motion_vector_vlc, 6, 63,
1681 &motion_vector_vlc_table[0][1], 2, 1,
1682 &motion_vector_vlc_table[0][0], 2, 1, 0);
1684 /* work out the block mapping tables */
1685 s->superblock_fragments = av_malloc(s->superblock_count * 16 * sizeof(int));
1686 s->macroblock_coding = av_malloc(s->macroblock_count + 1);
1687 if (!s->superblock_fragments || !s->macroblock_coding) {
1688 vp3_decode_end(avctx);
1691 init_block_mapping(s);
1693 for (i = 0; i < 3; i++) {
1694 s->current_frame.data[i] = NULL;
1695 s->last_frame.data[i] = NULL;
1696 s->golden_frame.data[i] = NULL;
1702 av_log(avctx, AV_LOG_FATAL, "Invalid huffman table\n");
1707 * This is the ffmpeg/libavcodec API frame decode function.
1709 static int vp3_decode_frame(AVCodecContext *avctx,
1710 void *data, int *data_size,
1713 const uint8_t *buf = avpkt->data;
1714 int buf_size = avpkt->size;
1715 Vp3DecodeContext *s = avctx->priv_data;
1717 static int counter = 0;
1720 init_get_bits(&gb, buf, buf_size * 8);
1722 if (s->theora && get_bits1(&gb))
1724 av_log(avctx, AV_LOG_ERROR, "Header packet passed to frame decoder, skipping\n");
1728 s->keyframe = !get_bits1(&gb);
1731 for (i = 0; i < 3; i++)
1732 s->last_qps[i] = s->qps[i];
1736 s->qps[s->nqps++]= get_bits(&gb, 6);
1737 } while(s->theora >= 0x030200 && s->nqps<3 && get_bits1(&gb));
1738 for (i = s->nqps; i < 3; i++)
1741 if (s->avctx->debug & FF_DEBUG_PICT_INFO)
1742 av_log(s->avctx, AV_LOG_INFO, " VP3 %sframe #%d: Q index = %d\n",
1743 s->keyframe?"key":"", counter, s->qps[0]);
1746 if (s->qps[0] != s->last_qps[0])
1747 init_loop_filter(s);
1749 for (i = 0; i < s->nqps; i++)
1750 // reinit all dequantizers if the first one changed, because
1751 // the DC of the first quantizer must be used for all matrices
1752 if (s->qps[i] != s->last_qps[i] || s->qps[0] != s->last_qps[0])
1753 init_dequantizer(s, i);
1755 if (avctx->skip_frame >= AVDISCARD_NONKEY && !s->keyframe)
1758 s->current_frame.reference = 3;
1759 s->current_frame.pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1760 if (avctx->get_buffer(avctx, &s->current_frame) < 0) {
1761 av_log(s->avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1768 skip_bits(&gb, 4); /* width code */
1769 skip_bits(&gb, 4); /* height code */
1772 s->version = get_bits(&gb, 5);
1774 av_log(s->avctx, AV_LOG_DEBUG, "VP version: %d\n", s->version);
1777 if (s->version || s->theora)
1780 av_log(s->avctx, AV_LOG_ERROR, "Warning, unsupported keyframe coding type?!\n");
1781 skip_bits(&gb, 2); /* reserved? */
1784 if (!s->golden_frame.data[0]) {
1785 av_log(s->avctx, AV_LOG_WARNING, "vp3: first frame not a keyframe\n");
1787 s->golden_frame.reference = 3;
1788 s->golden_frame.pict_type = FF_I_TYPE;
1789 if (avctx->get_buffer(avctx, &s->golden_frame) < 0) {
1790 av_log(s->avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1793 s->last_frame = s->golden_frame;
1794 s->last_frame.type = FF_BUFFER_TYPE_COPY;
1798 s->current_frame.qscale_table= s->qscale_table; //FIXME allocate individual tables per AVFrame
1799 s->current_frame.qstride= 0;
1801 memset(s->all_fragments, 0, s->fragment_count * sizeof(Vp3Fragment));
1803 if (unpack_superblocks(s, &gb)){
1804 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_superblocks\n");
1807 if (unpack_modes(s, &gb)){
1808 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_modes\n");
1811 if (unpack_vectors(s, &gb)){
1812 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_vectors\n");
1815 if (unpack_block_qpis(s, &gb)){
1816 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_block_qpis\n");
1819 if (unpack_dct_coeffs(s, &gb)){
1820 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_dct_coeffs\n");
1824 for (i = 0; i < 3; i++) {
1825 int height = s->height >> (i && s->chroma_y_shift);
1826 if (s->flipped_image)
1827 s->data_offset[i] = 0;
1829 s->data_offset[i] = (height-1) * s->current_frame.linesize[i];
1832 s->last_slice_end = 0;
1833 for (i = 0; i < s->c_superblock_height; i++)
1836 // filter the last row
1837 for (i = 0; i < 3; i++) {
1838 int row = (s->height >> (3+(i && s->chroma_y_shift))) - 1;
1839 apply_loop_filter(s, i, row, row+1);
1841 vp3_draw_horiz_band(s, s->height);
1843 *data_size=sizeof(AVFrame);
1844 *(AVFrame*)data= s->current_frame;
1846 /* release the last frame, if it is allocated and if it is not the
1848 if (s->last_frame.data[0] && s->last_frame.type != FF_BUFFER_TYPE_COPY)
1849 avctx->release_buffer(avctx, &s->last_frame);
1851 /* shuffle frames (last = current) */
1852 s->last_frame= s->current_frame;
1855 if (s->golden_frame.data[0])
1856 avctx->release_buffer(avctx, &s->golden_frame);
1857 s->golden_frame = s->current_frame;
1858 s->last_frame.type = FF_BUFFER_TYPE_COPY;
1861 s->current_frame.data[0]= NULL; /* ensure that we catch any access to this released frame */
1866 if (s->current_frame.data[0])
1867 avctx->release_buffer(avctx, &s->current_frame);
1872 * This is the ffmpeg/libavcodec API module cleanup function.
1874 static av_cold int vp3_decode_end(AVCodecContext *avctx)
1876 Vp3DecodeContext *s = avctx->priv_data;
1879 av_free(s->superblock_coding);
1880 av_free(s->all_fragments);
1881 av_free(s->coded_fragment_list[0]);
1882 av_free(s->dct_tokens_base);
1883 av_free(s->superblock_fragments);
1884 av_free(s->macroblock_coding);
1885 av_free(s->motion_val[0]);
1886 av_free(s->motion_val[1]);
1888 for (i = 0; i < 16; i++) {
1889 free_vlc(&s->dc_vlc[i]);
1890 free_vlc(&s->ac_vlc_1[i]);
1891 free_vlc(&s->ac_vlc_2[i]);
1892 free_vlc(&s->ac_vlc_3[i]);
1893 free_vlc(&s->ac_vlc_4[i]);
1896 free_vlc(&s->superblock_run_length_vlc);
1897 free_vlc(&s->fragment_run_length_vlc);
1898 free_vlc(&s->mode_code_vlc);
1899 free_vlc(&s->motion_vector_vlc);
1901 /* release all frames */
1902 if (s->golden_frame.data[0])
1903 avctx->release_buffer(avctx, &s->golden_frame);
1904 if (s->last_frame.data[0] && s->last_frame.type != FF_BUFFER_TYPE_COPY)
1905 avctx->release_buffer(avctx, &s->last_frame);
1906 /* no need to release the current_frame since it will always be pointing
1907 * to the same frame as either the golden or last frame */
1912 static int read_huffman_tree(AVCodecContext *avctx, GetBitContext *gb)
1914 Vp3DecodeContext *s = avctx->priv_data;
1916 if (get_bits1(gb)) {
1918 if (s->entries >= 32) { /* overflow */
1919 av_log(avctx, AV_LOG_ERROR, "huffman tree overflow\n");
1922 token = get_bits(gb, 5);
1923 //av_log(avctx, AV_LOG_DEBUG, "hti %d hbits %x token %d entry : %d size %d\n", s->hti, s->hbits, token, s->entries, s->huff_code_size);
1924 s->huffman_table[s->hti][token][0] = s->hbits;
1925 s->huffman_table[s->hti][token][1] = s->huff_code_size;
1929 if (s->huff_code_size >= 32) {/* overflow */
1930 av_log(avctx, AV_LOG_ERROR, "huffman tree overflow\n");
1933 s->huff_code_size++;
1935 if (read_huffman_tree(avctx, gb))
1938 if (read_huffman_tree(avctx, gb))
1941 s->huff_code_size--;
1946 #if CONFIG_THEORA_DECODER
1947 static const enum PixelFormat theora_pix_fmts[4] = {
1948 PIX_FMT_YUV420P, PIX_FMT_NONE, PIX_FMT_YUV422P, PIX_FMT_YUV444P
1951 static int theora_decode_header(AVCodecContext *avctx, GetBitContext *gb)
1953 Vp3DecodeContext *s = avctx->priv_data;
1954 int visible_width, visible_height, colorspace;
1955 int offset_x = 0, offset_y = 0;
1958 s->theora = get_bits_long(gb, 24);
1959 av_log(avctx, AV_LOG_DEBUG, "Theora bitstream version %X\n", s->theora);
1961 /* 3.2.0 aka alpha3 has the same frame orientation as original vp3 */
1962 /* but previous versions have the image flipped relative to vp3 */
1963 if (s->theora < 0x030200)
1965 s->flipped_image = 1;
1966 av_log(avctx, AV_LOG_DEBUG, "Old (<alpha3) Theora bitstream, flipped image\n");
1969 visible_width = s->width = get_bits(gb, 16) << 4;
1970 visible_height = s->height = get_bits(gb, 16) << 4;
1972 if(avcodec_check_dimensions(avctx, s->width, s->height)){
1973 av_log(avctx, AV_LOG_ERROR, "Invalid dimensions (%dx%d)\n", s->width, s->height);
1974 s->width= s->height= 0;
1978 if (s->theora >= 0x030200) {
1979 visible_width = get_bits_long(gb, 24);
1980 visible_height = get_bits_long(gb, 24);
1982 offset_x = get_bits(gb, 8); /* offset x */
1983 offset_y = get_bits(gb, 8); /* offset y, from bottom */
1986 fps.num = get_bits_long(gb, 32);
1987 fps.den = get_bits_long(gb, 32);
1988 if (fps.num && fps.den) {
1989 av_reduce(&avctx->time_base.num, &avctx->time_base.den,
1990 fps.den, fps.num, 1<<30);
1993 avctx->sample_aspect_ratio.num = get_bits_long(gb, 24);
1994 avctx->sample_aspect_ratio.den = get_bits_long(gb, 24);
1996 if (s->theora < 0x030200)
1997 skip_bits(gb, 5); /* keyframe frequency force */
1998 colorspace = get_bits(gb, 8);
1999 skip_bits(gb, 24); /* bitrate */
2001 skip_bits(gb, 6); /* quality hint */
2003 if (s->theora >= 0x030200)
2005 skip_bits(gb, 5); /* keyframe frequency force */
2006 avctx->pix_fmt = theora_pix_fmts[get_bits(gb, 2)];
2007 skip_bits(gb, 3); /* reserved */
2010 // align_get_bits(gb);
2012 if ( visible_width <= s->width && visible_width > s->width-16
2013 && visible_height <= s->height && visible_height > s->height-16
2014 && !offset_x && (offset_y == s->height - visible_height))
2015 avcodec_set_dimensions(avctx, visible_width, visible_height);
2017 avcodec_set_dimensions(avctx, s->width, s->height);
2019 if (colorspace == 1) {
2020 avctx->color_primaries = AVCOL_PRI_BT470M;
2021 } else if (colorspace == 2) {
2022 avctx->color_primaries = AVCOL_PRI_BT470BG;
2024 if (colorspace == 1 || colorspace == 2) {
2025 avctx->colorspace = AVCOL_SPC_BT470BG;
2026 avctx->color_trc = AVCOL_TRC_BT709;
2032 static int theora_decode_tables(AVCodecContext *avctx, GetBitContext *gb)
2034 Vp3DecodeContext *s = avctx->priv_data;
2035 int i, n, matrices, inter, plane;
2037 if (s->theora >= 0x030200) {
2038 n = get_bits(gb, 3);
2039 /* loop filter limit values table */
2040 for (i = 0; i < 64; i++) {
2041 s->filter_limit_values[i] = get_bits(gb, n);
2042 if (s->filter_limit_values[i] > 127) {
2043 av_log(avctx, AV_LOG_ERROR, "filter limit value too large (%i > 127), clamping\n", s->filter_limit_values[i]);
2044 s->filter_limit_values[i] = 127;
2049 if (s->theora >= 0x030200)
2050 n = get_bits(gb, 4) + 1;
2053 /* quality threshold table */
2054 for (i = 0; i < 64; i++)
2055 s->coded_ac_scale_factor[i] = get_bits(gb, n);
2057 if (s->theora >= 0x030200)
2058 n = get_bits(gb, 4) + 1;
2061 /* dc scale factor table */
2062 for (i = 0; i < 64; i++)
2063 s->coded_dc_scale_factor[i] = get_bits(gb, n);
2065 if (s->theora >= 0x030200)
2066 matrices = get_bits(gb, 9) + 1;
2071 av_log(avctx, AV_LOG_ERROR, "invalid number of base matrixes\n");
2075 for(n=0; n<matrices; n++){
2076 for (i = 0; i < 64; i++)
2077 s->base_matrix[n][i]= get_bits(gb, 8);
2080 for (inter = 0; inter <= 1; inter++) {
2081 for (plane = 0; plane <= 2; plane++) {
2083 if (inter || plane > 0)
2084 newqr = get_bits1(gb);
2087 if(inter && get_bits1(gb)){
2091 qtj= (3*inter + plane - 1) / 3;
2092 plj= (plane + 2) % 3;
2094 s->qr_count[inter][plane]= s->qr_count[qtj][plj];
2095 memcpy(s->qr_size[inter][plane], s->qr_size[qtj][plj], sizeof(s->qr_size[0][0]));
2096 memcpy(s->qr_base[inter][plane], s->qr_base[qtj][plj], sizeof(s->qr_base[0][0]));
2102 i= get_bits(gb, av_log2(matrices-1)+1);
2104 av_log(avctx, AV_LOG_ERROR, "invalid base matrix index\n");
2107 s->qr_base[inter][plane][qri]= i;
2110 i = get_bits(gb, av_log2(63-qi)+1) + 1;
2111 s->qr_size[inter][plane][qri++]= i;
2116 av_log(avctx, AV_LOG_ERROR, "invalid qi %d > 63\n", qi);
2119 s->qr_count[inter][plane]= qri;
2124 /* Huffman tables */
2125 for (s->hti = 0; s->hti < 80; s->hti++) {
2127 s->huff_code_size = 1;
2128 if (!get_bits1(gb)) {
2130 if(read_huffman_tree(avctx, gb))
2133 if(read_huffman_tree(avctx, gb))
2138 s->theora_tables = 1;
2143 static av_cold int theora_decode_init(AVCodecContext *avctx)
2145 Vp3DecodeContext *s = avctx->priv_data;
2148 uint8_t *header_start[3];
2154 if (!avctx->extradata_size)
2156 av_log(avctx, AV_LOG_ERROR, "Missing extradata!\n");
2160 if (ff_split_xiph_headers(avctx->extradata, avctx->extradata_size,
2161 42, header_start, header_len) < 0) {
2162 av_log(avctx, AV_LOG_ERROR, "Corrupt extradata\n");
2167 init_get_bits(&gb, header_start[i], header_len[i] * 8);
2169 ptype = get_bits(&gb, 8);
2171 if (!(ptype & 0x80))
2173 av_log(avctx, AV_LOG_ERROR, "Invalid extradata!\n");
2177 // FIXME: Check for this as well.
2178 skip_bits_long(&gb, 6*8); /* "theora" */
2183 theora_decode_header(avctx, &gb);
2186 // FIXME: is this needed? it breaks sometimes
2187 // theora_decode_comments(avctx, gb);
2190 if (theora_decode_tables(avctx, &gb))
2194 av_log(avctx, AV_LOG_ERROR, "Unknown Theora config packet: %d\n", ptype&~0x80);
2197 if(ptype != 0x81 && 8*header_len[i] != get_bits_count(&gb))
2198 av_log(avctx, AV_LOG_WARNING, "%d bits left in packet %X\n", 8*header_len[i] - get_bits_count(&gb), ptype);
2199 if (s->theora < 0x030200)
2203 return vp3_decode_init(avctx);
2206 AVCodec theora_decoder = {
2210 sizeof(Vp3DecodeContext),
2215 CODEC_CAP_DR1 | CODEC_CAP_DRAW_HORIZ_BAND,
2217 .long_name = NULL_IF_CONFIG_SMALL("Theora"),
2221 AVCodec vp3_decoder = {
2225 sizeof(Vp3DecodeContext),
2230 CODEC_CAP_DR1 | CODEC_CAP_DRAW_HORIZ_BAND,
2232 .long_name = NULL_IF_CONFIG_SMALL("On2 VP3"),