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
22 * @file libavcodec/vp3.c
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 uint16_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) {
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 current_run = get_vlc2(gb,
414 s->superblock_run_length_vlc.table, 6, 2) + 1;
415 if (current_run == 34)
416 current_run += get_bits(gb, 12);
418 for (j = 0; j < current_run; current_superblock++) {
419 if (current_superblock >= s->superblock_count) {
420 av_log(s->avctx, AV_LOG_ERROR, "Invalid fully coded superblock run length\n");
424 /* skip any superblocks already marked as partially coded */
425 if (s->superblock_coding[current_superblock] == SB_NOT_CODED) {
426 s->superblock_coding[current_superblock] = 2*bit;
430 superblocks_decoded += current_run;
432 if (s->theora && current_run == MAXIMUM_LONG_BIT_RUN)
439 /* if there were partial blocks, initialize bitstream for
440 * unpacking fragment codings */
441 if (num_partial_superblocks) {
445 /* toggle the bit because as soon as the first run length is
446 * fetched the bit will be toggled again */
451 /* figure out which fragments are coded; iterate through each
452 * superblock (all planes) */
453 s->total_num_coded_frags = 0;
454 memset(s->macroblock_coding, MODE_COPY, s->macroblock_count);
456 for (plane = 0; plane < 3; plane++) {
457 int sb_start = superblock_starts[plane];
458 int sb_end = sb_start + (plane ? s->c_superblock_count : s->y_superblock_count);
459 int num_coded_frags = 0;
461 for (i = sb_start; i < sb_end; i++) {
463 /* iterate through all 16 fragments in a superblock */
464 for (j = 0; j < 16; j++) {
466 /* if the fragment is in bounds, check its coding status */
467 current_fragment = s->superblock_fragments[i * 16 + j];
468 if (current_fragment != -1) {
469 int coded = s->superblock_coding[i];
471 if (s->superblock_coding[i] == SB_PARTIALLY_CODED) {
473 /* fragment may or may not be coded; this is the case
474 * that cares about the fragment coding runs */
475 if (current_run-- == 0) {
477 current_run = get_vlc2(gb,
478 s->fragment_run_length_vlc.table, 5, 2);
484 /* default mode; actual mode will be decoded in
486 s->all_fragments[current_fragment].coding_method =
488 s->coded_fragment_list[plane][num_coded_frags++] =
491 /* not coded; copy this fragment from the prior frame */
492 s->all_fragments[current_fragment].coding_method =
498 s->total_num_coded_frags += num_coded_frags;
499 for (i = 0; i < 64; i++)
500 s->num_coded_frags[plane][i] = num_coded_frags;
502 s->coded_fragment_list[plane+1] = s->coded_fragment_list[plane] + num_coded_frags;
508 * This function unpacks all the coding mode data for individual macroblocks
509 * from the bitstream.
511 static int unpack_modes(Vp3DecodeContext *s, GetBitContext *gb)
513 int i, j, k, sb_x, sb_y;
515 int current_macroblock;
516 int current_fragment;
518 int custom_mode_alphabet[CODING_MODE_COUNT];
523 for (i = 0; i < s->fragment_count; i++)
524 s->all_fragments[i].coding_method = MODE_INTRA;
528 /* fetch the mode coding scheme for this frame */
529 scheme = get_bits(gb, 3);
531 /* is it a custom coding scheme? */
533 for (i = 0; i < 8; i++)
534 custom_mode_alphabet[i] = MODE_INTER_NO_MV;
535 for (i = 0; i < 8; i++)
536 custom_mode_alphabet[get_bits(gb, 3)] = i;
537 alphabet = custom_mode_alphabet;
539 alphabet = ModeAlphabet[scheme-1];
541 /* iterate through all of the macroblocks that contain 1 or more
543 for (sb_y = 0; sb_y < s->y_superblock_height; sb_y++) {
544 for (sb_x = 0; sb_x < s->y_superblock_width; sb_x++) {
546 for (j = 0; j < 4; j++) {
547 int mb_x = 2*sb_x + (j>>1);
548 int mb_y = 2*sb_y + (((j>>1)+j)&1);
549 current_macroblock = mb_y * s->macroblock_width + mb_x;
551 if (mb_x >= s->macroblock_width || mb_y >= s->macroblock_height)
554 #define BLOCK_X (2*mb_x + (k&1))
555 #define BLOCK_Y (2*mb_y + (k>>1))
556 /* coding modes are only stored if the macroblock has at least one
557 * luma block coded, otherwise it must be INTER_NO_MV */
558 for (k = 0; k < 4; k++) {
559 current_fragment = BLOCK_Y*s->fragment_width[0] + BLOCK_X;
560 if (s->all_fragments[current_fragment].coding_method != MODE_COPY)
564 s->macroblock_coding[current_macroblock] = MODE_INTER_NO_MV;
568 /* mode 7 means get 3 bits for each coding mode */
570 coding_mode = get_bits(gb, 3);
572 coding_mode = alphabet
573 [get_vlc2(gb, s->mode_code_vlc.table, 3, 3)];
575 s->macroblock_coding[current_macroblock] = coding_mode;
576 for (k = 0; k < 4; k++) {
577 frag = s->all_fragments + BLOCK_Y*s->fragment_width[0] + BLOCK_X;
578 if (frag->coding_method != MODE_COPY)
579 frag->coding_method = coding_mode;
582 #define SET_CHROMA_MODES \
583 if (frag[s->fragment_start[1]].coding_method != MODE_COPY) \
584 frag[s->fragment_start[1]].coding_method = coding_mode;\
585 if (frag[s->fragment_start[2]].coding_method != MODE_COPY) \
586 frag[s->fragment_start[2]].coding_method = coding_mode;
588 if (s->chroma_y_shift) {
589 frag = s->all_fragments + mb_y*s->fragment_width[1] + mb_x;
591 } else if (s->chroma_x_shift) {
592 frag = s->all_fragments + 2*mb_y*s->fragment_width[1] + mb_x;
593 for (k = 0; k < 2; k++) {
595 frag += s->fragment_width[1];
598 for (k = 0; k < 4; k++) {
599 frag = s->all_fragments + BLOCK_Y*s->fragment_width[1] + BLOCK_X;
612 * This function unpacks all the motion vectors for the individual
613 * macroblocks from the bitstream.
615 static int unpack_vectors(Vp3DecodeContext *s, GetBitContext *gb)
617 int j, k, sb_x, sb_y;
621 int last_motion_x = 0;
622 int last_motion_y = 0;
623 int prior_last_motion_x = 0;
624 int prior_last_motion_y = 0;
625 int current_macroblock;
626 int current_fragment;
632 /* coding mode 0 is the VLC scheme; 1 is the fixed code scheme */
633 coding_mode = get_bits1(gb);
635 /* iterate through all of the macroblocks that contain 1 or more
637 for (sb_y = 0; sb_y < s->y_superblock_height; sb_y++) {
638 for (sb_x = 0; sb_x < s->y_superblock_width; sb_x++) {
640 for (j = 0; j < 4; j++) {
641 int mb_x = 2*sb_x + (j>>1);
642 int mb_y = 2*sb_y + (((j>>1)+j)&1);
643 current_macroblock = mb_y * s->macroblock_width + mb_x;
645 if (mb_x >= s->macroblock_width || mb_y >= s->macroblock_height ||
646 (s->macroblock_coding[current_macroblock] == MODE_COPY))
649 switch (s->macroblock_coding[current_macroblock]) {
651 case MODE_INTER_PLUS_MV:
653 /* all 6 fragments use the same motion vector */
654 if (coding_mode == 0) {
655 motion_x[0] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
656 motion_y[0] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
658 motion_x[0] = fixed_motion_vector_table[get_bits(gb, 6)];
659 motion_y[0] = fixed_motion_vector_table[get_bits(gb, 6)];
662 /* vector maintenance, only on MODE_INTER_PLUS_MV */
663 if (s->macroblock_coding[current_macroblock] ==
664 MODE_INTER_PLUS_MV) {
665 prior_last_motion_x = last_motion_x;
666 prior_last_motion_y = last_motion_y;
667 last_motion_x = motion_x[0];
668 last_motion_y = motion_y[0];
672 case MODE_INTER_FOURMV:
673 /* vector maintenance */
674 prior_last_motion_x = last_motion_x;
675 prior_last_motion_y = last_motion_y;
677 /* fetch 4 vectors from the bitstream, one for each
678 * Y fragment, then average for the C fragment vectors */
679 for (k = 0; k < 4; k++) {
680 current_fragment = BLOCK_Y*s->fragment_width[0] + BLOCK_X;
681 if (s->all_fragments[current_fragment].coding_method != MODE_COPY) {
682 if (coding_mode == 0) {
683 motion_x[k] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
684 motion_y[k] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
686 motion_x[k] = fixed_motion_vector_table[get_bits(gb, 6)];
687 motion_y[k] = fixed_motion_vector_table[get_bits(gb, 6)];
689 last_motion_x = motion_x[k];
690 last_motion_y = motion_y[k];
698 case MODE_INTER_LAST_MV:
699 /* all 6 fragments use the last motion vector */
700 motion_x[0] = last_motion_x;
701 motion_y[0] = last_motion_y;
703 /* no vector maintenance (last vector remains the
707 case MODE_INTER_PRIOR_LAST:
708 /* all 6 fragments use the motion vector prior to the
709 * last motion vector */
710 motion_x[0] = prior_last_motion_x;
711 motion_y[0] = prior_last_motion_y;
713 /* vector maintenance */
714 prior_last_motion_x = last_motion_x;
715 prior_last_motion_y = last_motion_y;
716 last_motion_x = motion_x[0];
717 last_motion_y = motion_y[0];
721 /* covers intra, inter without MV, golden without MV */
725 /* no vector maintenance */
729 /* assign the motion vectors to the correct fragments */
730 for (k = 0; k < 4; k++) {
732 BLOCK_Y*s->fragment_width[0] + BLOCK_X;
733 if (s->macroblock_coding[current_macroblock] == MODE_INTER_FOURMV) {
734 s->motion_val[0][current_fragment][0] = motion_x[k];
735 s->motion_val[0][current_fragment][1] = motion_y[k];
737 s->motion_val[0][current_fragment][0] = motion_x[0];
738 s->motion_val[0][current_fragment][1] = motion_y[0];
742 if (s->chroma_y_shift) {
743 if (s->macroblock_coding[current_macroblock] == MODE_INTER_FOURMV) {
744 motion_x[0] = RSHIFT(motion_x[0] + motion_x[1] + motion_x[2] + motion_x[3], 2);
745 motion_y[0] = RSHIFT(motion_y[0] + motion_y[1] + motion_y[2] + motion_y[3], 2);
747 motion_x[0] = (motion_x[0]>>1) | (motion_x[0]&1);
748 motion_y[0] = (motion_y[0]>>1) | (motion_y[0]&1);
749 frag = mb_y*s->fragment_width[1] + mb_x;
750 s->motion_val[1][frag][0] = motion_x[0];
751 s->motion_val[1][frag][1] = motion_y[0];
752 } else if (s->chroma_x_shift) {
753 if (s->macroblock_coding[current_macroblock] == MODE_INTER_FOURMV) {
754 motion_x[0] = RSHIFT(motion_x[0] + motion_x[1], 1);
755 motion_y[0] = RSHIFT(motion_y[0] + motion_y[1], 1);
756 motion_x[1] = RSHIFT(motion_x[2] + motion_x[3], 1);
757 motion_y[1] = RSHIFT(motion_y[2] + motion_y[3], 1);
759 motion_x[1] = motion_x[0];
760 motion_y[1] = motion_y[0];
762 motion_x[0] = (motion_x[0]>>1) | (motion_x[0]&1);
763 motion_x[1] = (motion_x[1]>>1) | (motion_x[1]&1);
765 frag = 2*mb_y*s->fragment_width[1] + mb_x;
766 for (k = 0; k < 2; k++) {
767 s->motion_val[1][frag][0] = motion_x[k];
768 s->motion_val[1][frag][1] = motion_y[k];
769 frag += s->fragment_width[1];
772 for (k = 0; k < 4; k++) {
773 frag = BLOCK_Y*s->fragment_width[1] + BLOCK_X;
774 if (s->macroblock_coding[current_macroblock] == MODE_INTER_FOURMV) {
775 s->motion_val[1][frag][0] = motion_x[k];
776 s->motion_val[1][frag][1] = motion_y[k];
778 s->motion_val[1][frag][0] = motion_x[0];
779 s->motion_val[1][frag][1] = motion_y[0];
790 static int unpack_block_qpis(Vp3DecodeContext *s, GetBitContext *gb)
792 int qpi, i, j, bit, run_length, blocks_decoded, num_blocks_at_qpi;
793 int num_blocks = s->total_num_coded_frags;
795 for (qpi = 0; qpi < s->nqps-1 && num_blocks > 0; qpi++) {
796 i = blocks_decoded = num_blocks_at_qpi = 0;
801 run_length = get_vlc2(gb, s->superblock_run_length_vlc.table, 6, 2) + 1;
802 if (run_length == 34)
803 run_length += get_bits(gb, 12);
804 blocks_decoded += run_length;
807 num_blocks_at_qpi += run_length;
809 for (j = 0; j < run_length; i++) {
810 if (i >= s->total_num_coded_frags)
813 if (s->all_fragments[s->coded_fragment_list[0][i]].qpi == qpi) {
814 s->all_fragments[s->coded_fragment_list[0][i]].qpi += bit;
819 if (run_length == MAXIMUM_LONG_BIT_RUN)
823 } while (blocks_decoded < num_blocks);
825 num_blocks -= num_blocks_at_qpi;
832 * This function is called by unpack_dct_coeffs() to extract the VLCs from
833 * the bitstream. The VLCs encode tokens which are used to unpack DCT
834 * data. This function unpacks all the VLCs for either the Y plane or both
835 * C planes, and is called for DC coefficients or different AC coefficient
836 * levels (since different coefficient types require different VLC tables.
838 * This function returns a residual eob run. E.g, if a particular token gave
839 * instructions to EOB the next 5 fragments and there were only 2 fragments
840 * left in the current fragment range, 3 would be returned so that it could
841 * be passed into the next call to this same function.
843 static int unpack_vlcs(Vp3DecodeContext *s, GetBitContext *gb,
844 VLC *table, int coeff_index,
855 int num_coeffs = s->num_coded_frags[plane][coeff_index];
856 int16_t *dct_tokens = s->dct_tokens[plane][coeff_index];
858 /* local references to structure members to avoid repeated deferences */
859 int *coded_fragment_list = s->coded_fragment_list[plane];
860 Vp3Fragment *all_fragments = s->all_fragments;
861 VLC_TYPE (*vlc_table)[2] = table->table;
864 av_log(s->avctx, AV_LOG_ERROR, "Invalid number of coefficents at level %d\n", coeff_index);
866 if (eob_run > num_coeffs) {
867 coeff_i = blocks_ended = num_coeffs;
868 eob_run -= num_coeffs;
870 coeff_i = blocks_ended = eob_run;
874 // insert fake EOB token to cover the split between planes or zzi
876 dct_tokens[j++] = blocks_ended << 2;
878 while (coeff_i < num_coeffs && get_bits_left(gb) > 0) {
879 /* decode a VLC into a token */
880 token = get_vlc2(gb, vlc_table, 5, 3);
881 /* use the token to get a zero run, a coefficient, and an eob run */
883 eob_run = eob_run_base[token];
884 if (eob_run_get_bits[token])
885 eob_run += get_bits(gb, eob_run_get_bits[token]);
887 // record only the number of blocks ended in this plane,
888 // any spill will be recorded in the next plane.
889 if (eob_run > num_coeffs - coeff_i) {
890 dct_tokens[j++] = TOKEN_EOB(num_coeffs - coeff_i);
891 blocks_ended += num_coeffs - coeff_i;
892 eob_run -= num_coeffs - coeff_i;
893 coeff_i = num_coeffs;
895 dct_tokens[j++] = TOKEN_EOB(eob_run);
896 blocks_ended += eob_run;
901 bits_to_get = coeff_get_bits[token];
903 bits_to_get = get_bits(gb, bits_to_get);
904 coeff = coeff_tables[token][bits_to_get];
906 zero_run = zero_run_base[token];
907 if (zero_run_get_bits[token])
908 zero_run += get_bits(gb, zero_run_get_bits[token]);
911 dct_tokens[j++] = TOKEN_ZERO_RUN(coeff, zero_run);
913 // Save DC into the fragment structure. DC prediction is
914 // done in raster order, so the actual DC can't be in with
915 // other tokens. We still need the token in dct_tokens[]
916 // however, or else the structure collapses on itself.
918 all_fragments[coded_fragment_list[coeff_i]].dc = coeff;
920 dct_tokens[j++] = TOKEN_COEFF(coeff);
923 if (coeff_index + zero_run > 64) {
924 av_log(s->avctx, AV_LOG_DEBUG, "Invalid zero run of %d with"
925 " %d coeffs left\n", zero_run, 64-coeff_index);
926 zero_run = 64 - coeff_index;
929 // zero runs code multiple coefficients,
930 // so don't try to decode coeffs for those higher levels
931 for (i = coeff_index+1; i <= coeff_index+zero_run; i++)
932 s->num_coded_frags[plane][i]--;
937 if (blocks_ended > s->num_coded_frags[plane][coeff_index])
938 av_log(s->avctx, AV_LOG_ERROR, "More blocks ended than coded!\n");
940 // decrement the number of blocks that have higher coeffecients for each
941 // EOB run at this level
943 for (i = coeff_index+1; i < 64; i++)
944 s->num_coded_frags[plane][i] -= blocks_ended;
946 // setup the next buffer
948 s->dct_tokens[plane+1][coeff_index] = dct_tokens + j;
949 else if (coeff_index < 63)
950 s->dct_tokens[0][coeff_index+1] = dct_tokens + j;
955 static void reverse_dc_prediction(Vp3DecodeContext *s,
958 int fragment_height);
960 * This function unpacks all of the DCT coefficient data from the
963 static int unpack_dct_coeffs(Vp3DecodeContext *s, GetBitContext *gb)
970 int residual_eob_run = 0;
974 s->dct_tokens[0][0] = s->dct_tokens_base;
976 /* fetch the DC table indexes */
977 dc_y_table = get_bits(gb, 4);
978 dc_c_table = get_bits(gb, 4);
980 /* unpack the Y plane DC coefficients */
981 residual_eob_run = unpack_vlcs(s, gb, &s->dc_vlc[dc_y_table], 0,
982 0, residual_eob_run);
984 /* reverse prediction of the Y-plane DC coefficients */
985 reverse_dc_prediction(s, 0, s->fragment_width[0], s->fragment_height[0]);
987 /* unpack the C plane DC coefficients */
988 residual_eob_run = unpack_vlcs(s, gb, &s->dc_vlc[dc_c_table], 0,
989 1, residual_eob_run);
990 residual_eob_run = unpack_vlcs(s, gb, &s->dc_vlc[dc_c_table], 0,
991 2, residual_eob_run);
993 /* reverse prediction of the C-plane DC coefficients */
994 if (!(s->avctx->flags & CODEC_FLAG_GRAY))
996 reverse_dc_prediction(s, s->fragment_start[1],
997 s->fragment_width[1], s->fragment_height[1]);
998 reverse_dc_prediction(s, s->fragment_start[2],
999 s->fragment_width[1], s->fragment_height[1]);
1002 /* fetch the AC table indexes */
1003 ac_y_table = get_bits(gb, 4);
1004 ac_c_table = get_bits(gb, 4);
1006 /* build tables of AC VLC tables */
1007 for (i = 1; i <= 5; i++) {
1008 y_tables[i] = &s->ac_vlc_1[ac_y_table];
1009 c_tables[i] = &s->ac_vlc_1[ac_c_table];
1011 for (i = 6; i <= 14; i++) {
1012 y_tables[i] = &s->ac_vlc_2[ac_y_table];
1013 c_tables[i] = &s->ac_vlc_2[ac_c_table];
1015 for (i = 15; i <= 27; i++) {
1016 y_tables[i] = &s->ac_vlc_3[ac_y_table];
1017 c_tables[i] = &s->ac_vlc_3[ac_c_table];
1019 for (i = 28; i <= 63; i++) {
1020 y_tables[i] = &s->ac_vlc_4[ac_y_table];
1021 c_tables[i] = &s->ac_vlc_4[ac_c_table];
1024 /* decode all AC coefficents */
1025 for (i = 1; i <= 63; i++) {
1026 residual_eob_run = unpack_vlcs(s, gb, y_tables[i], i,
1027 0, residual_eob_run);
1029 residual_eob_run = unpack_vlcs(s, gb, c_tables[i], i,
1030 1, residual_eob_run);
1031 residual_eob_run = unpack_vlcs(s, gb, c_tables[i], i,
1032 2, residual_eob_run);
1039 * This function reverses the DC prediction for each coded fragment in
1040 * the frame. Much of this function is adapted directly from the original
1043 #define COMPATIBLE_FRAME(x) \
1044 (compatible_frame[s->all_fragments[x].coding_method] == current_frame_type)
1045 #define DC_COEFF(u) s->all_fragments[u].dc
1047 static void reverse_dc_prediction(Vp3DecodeContext *s,
1050 int fragment_height)
1059 int i = first_fragment;
1063 /* DC values for the left, up-left, up, and up-right fragments */
1064 int vl, vul, vu, vur;
1066 /* indexes for the left, up-left, up, and up-right fragments */
1070 * The 6 fields mean:
1071 * 0: up-left multiplier
1073 * 2: up-right multiplier
1074 * 3: left multiplier
1076 static const int predictor_transform[16][4] = {
1078 { 0, 0, 0,128}, // PL
1079 { 0, 0,128, 0}, // PUR
1080 { 0, 0, 53, 75}, // PUR|PL
1081 { 0,128, 0, 0}, // PU
1082 { 0, 64, 0, 64}, // PU|PL
1083 { 0,128, 0, 0}, // PU|PUR
1084 { 0, 0, 53, 75}, // PU|PUR|PL
1085 {128, 0, 0, 0}, // PUL
1086 { 0, 0, 0,128}, // PUL|PL
1087 { 64, 0, 64, 0}, // PUL|PUR
1088 { 0, 0, 53, 75}, // PUL|PUR|PL
1089 { 0,128, 0, 0}, // PUL|PU
1090 {-104,116, 0,116}, // PUL|PU|PL
1091 { 24, 80, 24, 0}, // PUL|PU|PUR
1092 {-104,116, 0,116} // PUL|PU|PUR|PL
1095 /* This table shows which types of blocks can use other blocks for
1096 * prediction. For example, INTRA is the only mode in this table to
1097 * have a frame number of 0. That means INTRA blocks can only predict
1098 * from other INTRA blocks. There are 2 golden frame coding types;
1099 * blocks encoding in these modes can only predict from other blocks
1100 * that were encoded with these 1 of these 2 modes. */
1101 static const unsigned char compatible_frame[9] = {
1102 1, /* MODE_INTER_NO_MV */
1104 1, /* MODE_INTER_PLUS_MV */
1105 1, /* MODE_INTER_LAST_MV */
1106 1, /* MODE_INTER_PRIOR_MV */
1107 2, /* MODE_USING_GOLDEN */
1108 2, /* MODE_GOLDEN_MV */
1109 1, /* MODE_INTER_FOUR_MV */
1112 int current_frame_type;
1114 /* there is a last DC predictor for each of the 3 frame types */
1119 vul = vu = vur = vl = 0;
1120 last_dc[0] = last_dc[1] = last_dc[2] = 0;
1122 /* for each fragment row... */
1123 for (y = 0; y < fragment_height; y++) {
1125 /* for each fragment in a row... */
1126 for (x = 0; x < fragment_width; x++, i++) {
1128 /* reverse prediction if this block was coded */
1129 if (s->all_fragments[i].coding_method != MODE_COPY) {
1131 current_frame_type =
1132 compatible_frame[s->all_fragments[i].coding_method];
1138 if(COMPATIBLE_FRAME(l))
1142 u= i-fragment_width;
1144 if(COMPATIBLE_FRAME(u))
1147 ul= i-fragment_width-1;
1149 if(COMPATIBLE_FRAME(ul))
1152 if(x + 1 < fragment_width){
1153 ur= i-fragment_width+1;
1155 if(COMPATIBLE_FRAME(ur))
1160 if (transform == 0) {
1162 /* if there were no fragments to predict from, use last
1164 predicted_dc = last_dc[current_frame_type];
1167 /* apply the appropriate predictor transform */
1169 (predictor_transform[transform][0] * vul) +
1170 (predictor_transform[transform][1] * vu) +
1171 (predictor_transform[transform][2] * vur) +
1172 (predictor_transform[transform][3] * vl);
1174 predicted_dc /= 128;
1176 /* check for outranging on the [ul u l] and
1177 * [ul u ur l] predictors */
1178 if ((transform == 15) || (transform == 13)) {
1179 if (FFABS(predicted_dc - vu) > 128)
1181 else if (FFABS(predicted_dc - vl) > 128)
1183 else if (FFABS(predicted_dc - vul) > 128)
1188 /* at long last, apply the predictor */
1189 DC_COEFF(i) += predicted_dc;
1191 last_dc[current_frame_type] = DC_COEFF(i);
1197 static void apply_loop_filter(Vp3DecodeContext *s, int plane, int ystart, int yend)
1200 int *bounding_values= s->bounding_values_array+127;
1202 int width = s->fragment_width[!!plane];
1203 int height = s->fragment_height[!!plane];
1204 int fragment = s->fragment_start [plane] + ystart * width;
1205 int stride = s->current_frame.linesize[plane];
1206 uint8_t *plane_data = s->current_frame.data [plane];
1207 if (!s->flipped_image) stride = -stride;
1208 plane_data += s->data_offset[plane] + 8*ystart*stride;
1210 for (y = ystart; y < yend; y++) {
1212 for (x = 0; x < width; x++) {
1213 /* This code basically just deblocks on the edges of coded blocks.
1214 * However, it has to be much more complicated because of the
1215 * braindamaged deblock ordering used in VP3/Theora. Order matters
1216 * because some pixels get filtered twice. */
1217 if( s->all_fragments[fragment].coding_method != MODE_COPY )
1219 /* do not perform left edge filter for left columns frags */
1221 s->dsp.vp3_h_loop_filter(
1223 stride, bounding_values);
1226 /* do not perform top edge filter for top row fragments */
1228 s->dsp.vp3_v_loop_filter(
1230 stride, bounding_values);
1233 /* do not perform right edge filter for right column
1234 * fragments or if right fragment neighbor is also coded
1235 * in this frame (it will be filtered in next iteration) */
1236 if ((x < width - 1) &&
1237 (s->all_fragments[fragment + 1].coding_method == MODE_COPY)) {
1238 s->dsp.vp3_h_loop_filter(
1239 plane_data + 8*x + 8,
1240 stride, bounding_values);
1243 /* do not perform bottom edge filter for bottom row
1244 * fragments or if bottom fragment neighbor is also coded
1245 * in this frame (it will be filtered in the next row) */
1246 if ((y < height - 1) &&
1247 (s->all_fragments[fragment + width].coding_method == MODE_COPY)) {
1248 s->dsp.vp3_v_loop_filter(
1249 plane_data + 8*x + 8*stride,
1250 stride, bounding_values);
1256 plane_data += 8*stride;
1261 * Pulls DCT tokens from the 64 levels to decode and dequant the coefficients
1262 * for the next block in coding order
1264 static inline int vp3_dequant(Vp3DecodeContext *s, Vp3Fragment *frag,
1265 int plane, int inter, DCTELEM block[64])
1267 int16_t *dequantizer = s->qmat[frag->qpi][inter][plane];
1268 uint8_t *perm = s->scantable.permutated;
1272 int token = *s->dct_tokens[plane][i];
1273 switch (token & 3) {
1275 if (--token < 4) // 0-3 are token types, so the EOB run must now be 0
1276 s->dct_tokens[plane][i]++;
1278 *s->dct_tokens[plane][i] = token & ~3;
1281 s->dct_tokens[plane][i]++;
1282 i += (token >> 2) & 0x7f;
1283 block[perm[i]] = (token >> 9) * dequantizer[perm[i]];
1287 block[perm[i]] = (token >> 2) * dequantizer[perm[i]];
1288 s->dct_tokens[plane][i++]++;
1291 av_log(s->avctx, AV_LOG_ERROR, "internal: invalid token type\n");
1296 // the actual DC+prediction is in the fragment structure
1297 block[0] = frag->dc * s->qmat[0][inter][plane][0];
1302 * called when all pixels up to row y are complete
1304 static void vp3_draw_horiz_band(Vp3DecodeContext *s, int y)
1309 if(s->avctx->draw_horiz_band==NULL)
1312 h= y - s->last_slice_end;
1315 if (!s->flipped_image) {
1317 h -= s->height - s->avctx->height; // account for non-mod16
1318 y = s->height - y - h;
1322 offset[0] = s->current_frame.linesize[0]*y;
1323 offset[1] = s->current_frame.linesize[1]*cy;
1324 offset[2] = s->current_frame.linesize[2]*cy;
1328 s->avctx->draw_horiz_band(s->avctx, &s->current_frame, offset, y, 3, h);
1329 s->last_slice_end= y + h;
1333 * Perform the final rendering for a particular slice of data.
1334 * The slice number ranges from 0..(c_superblock_height - 1).
1336 static void render_slice(Vp3DecodeContext *s, int slice)
1339 LOCAL_ALIGNED_16(DCTELEM, block, [64]);
1340 int motion_x = 0xdeadbeef, motion_y = 0xdeadbeef;
1341 int motion_halfpel_index;
1342 uint8_t *motion_source;
1343 int plane, first_pixel;
1345 if (slice >= s->c_superblock_height)
1348 for (plane = 0; plane < 3; plane++) {
1349 uint8_t *output_plane = s->current_frame.data [plane] + s->data_offset[plane];
1350 uint8_t * last_plane = s-> last_frame.data [plane] + s->data_offset[plane];
1351 uint8_t *golden_plane = s-> golden_frame.data [plane] + s->data_offset[plane];
1352 int stride = s->current_frame.linesize[plane];
1353 int plane_width = s->width >> (plane && s->chroma_x_shift);
1354 int plane_height = s->height >> (plane && s->chroma_y_shift);
1355 int8_t (*motion_val)[2] = s->motion_val[!!plane];
1357 int sb_x, sb_y = slice << (!plane && s->chroma_y_shift);
1358 int slice_height = sb_y + 1 + (!plane && s->chroma_y_shift);
1359 int slice_width = plane ? s->c_superblock_width : s->y_superblock_width;
1361 int fragment_width = s->fragment_width[!!plane];
1362 int fragment_height = s->fragment_height[!!plane];
1363 int fragment_start = s->fragment_start[plane];
1365 if (!s->flipped_image) stride = -stride;
1366 if (CONFIG_GRAY && plane && (s->avctx->flags & CODEC_FLAG_GRAY))
1370 if(FFABS(stride) > 2048)
1371 return; //various tables are fixed size
1373 /* for each superblock row in the slice (both of them)... */
1374 for (; sb_y < slice_height; sb_y++) {
1376 /* for each superblock in a row... */
1377 for (sb_x = 0; sb_x < slice_width; sb_x++) {
1379 /* for each block in a superblock... */
1380 for (j = 0; j < 16; j++) {
1381 x = 4*sb_x + hilbert_offset[j][0];
1382 y = 4*sb_y + hilbert_offset[j][1];
1384 i = fragment_start + y*fragment_width + x;
1387 if (x >= fragment_width || y >= fragment_height)
1390 first_pixel = 8*y*stride + 8*x;
1392 /* transform if this block was coded */
1393 if (s->all_fragments[i].coding_method != MODE_COPY) {
1394 int intra = s->all_fragments[i].coding_method == MODE_INTRA;
1396 if ((s->all_fragments[i].coding_method == MODE_USING_GOLDEN) ||
1397 (s->all_fragments[i].coding_method == MODE_GOLDEN_MV))
1398 motion_source= golden_plane;
1400 motion_source= last_plane;
1402 motion_source += first_pixel;
1403 motion_halfpel_index = 0;
1405 /* sort out the motion vector if this fragment is coded
1406 * using a motion vector method */
1407 if ((s->all_fragments[i].coding_method > MODE_INTRA) &&
1408 (s->all_fragments[i].coding_method != MODE_USING_GOLDEN)) {
1410 motion_x = motion_val[y*fragment_width + x][0];
1411 motion_y = motion_val[y*fragment_width + x][1];
1413 src_x= (motion_x>>1) + 8*x;
1414 src_y= (motion_y>>1) + 8*y;
1416 motion_halfpel_index = motion_x & 0x01;
1417 motion_source += (motion_x >> 1);
1419 motion_halfpel_index |= (motion_y & 0x01) << 1;
1420 motion_source += ((motion_y >> 1) * stride);
1422 if(src_x<0 || src_y<0 || src_x + 9 >= plane_width || src_y + 9 >= plane_height){
1423 uint8_t *temp= s->edge_emu_buffer;
1424 if(stride<0) temp -= 9*stride;
1425 else temp += 9*stride;
1427 ff_emulated_edge_mc(temp, motion_source, stride, 9, 9, src_x, src_y, plane_width, plane_height);
1428 motion_source= temp;
1433 /* first, take care of copying a block from either the
1434 * previous or the golden frame */
1435 if (s->all_fragments[i].coding_method != MODE_INTRA) {
1436 /* Note, it is possible to implement all MC cases with
1437 put_no_rnd_pixels_l2 which would look more like the
1438 VP3 source but this would be slower as
1439 put_no_rnd_pixels_tab is better optimzed */
1440 if(motion_halfpel_index != 3){
1441 s->dsp.put_no_rnd_pixels_tab[1][motion_halfpel_index](
1442 output_plane + first_pixel,
1443 motion_source, stride, 8);
1445 int d= (motion_x ^ motion_y)>>31; // d is 0 if motion_x and _y have the same sign, else -1
1446 s->dsp.put_no_rnd_pixels_l2[1](
1447 output_plane + first_pixel,
1449 motion_source + stride + 1 + d,
1454 s->dsp.clear_block(block);
1455 vp3_dequant(s, s->all_fragments + i, plane, !intra, block);
1457 /* invert DCT and place (or add) in final output */
1459 if (s->all_fragments[i].coding_method == MODE_INTRA) {
1460 if(s->avctx->idct_algo!=FF_IDCT_VP3)
1463 output_plane + first_pixel,
1468 output_plane + first_pixel,
1474 /* copy directly from the previous frame */
1475 s->dsp.put_pixels_tab[1][0](
1476 output_plane + first_pixel,
1477 last_plane + first_pixel,
1484 // Filter up to the last row in the superblock row
1485 apply_loop_filter(s, plane, 4*sb_y - !!sb_y, FFMIN(4*sb_y+3, fragment_height-1));
1489 /* this looks like a good place for slice dispatch... */
1491 * if (slice == s->macroblock_height - 1)
1492 * dispatch (both last slice & 2nd-to-last slice);
1493 * else if (slice > 0)
1494 * dispatch (slice - 1);
1497 vp3_draw_horiz_band(s, FFMIN(64*slice + 64-16, s->height-16));
1501 * This is the ffmpeg/libavcodec API init function.
1503 static av_cold int vp3_decode_init(AVCodecContext *avctx)
1505 Vp3DecodeContext *s = avctx->priv_data;
1506 int i, inter, plane;
1509 int y_fragment_count, c_fragment_count;
1511 if (avctx->codec_tag == MKTAG('V','P','3','0'))
1517 s->width = FFALIGN(avctx->width, 16);
1518 s->height = FFALIGN(avctx->height, 16);
1519 if (avctx->pix_fmt == PIX_FMT_NONE)
1520 avctx->pix_fmt = PIX_FMT_YUV420P;
1521 avctx->chroma_sample_location = AVCHROMA_LOC_CENTER;
1522 if(avctx->idct_algo==FF_IDCT_AUTO)
1523 avctx->idct_algo=FF_IDCT_VP3;
1524 dsputil_init(&s->dsp, avctx);
1526 ff_init_scantable(s->dsp.idct_permutation, &s->scantable, ff_zigzag_direct);
1528 /* initialize to an impossible value which will force a recalculation
1529 * in the first frame decode */
1530 for (i = 0; i < 3; i++)
1533 avcodec_get_chroma_sub_sample(avctx->pix_fmt, &s->chroma_x_shift, &s->chroma_y_shift);
1535 s->y_superblock_width = (s->width + 31) / 32;
1536 s->y_superblock_height = (s->height + 31) / 32;
1537 s->y_superblock_count = s->y_superblock_width * s->y_superblock_height;
1539 /* work out the dimensions for the C planes */
1540 c_width = s->width >> s->chroma_x_shift;
1541 c_height = s->height >> s->chroma_y_shift;
1542 s->c_superblock_width = (c_width + 31) / 32;
1543 s->c_superblock_height = (c_height + 31) / 32;
1544 s->c_superblock_count = s->c_superblock_width * s->c_superblock_height;
1546 s->superblock_count = s->y_superblock_count + (s->c_superblock_count * 2);
1547 s->u_superblock_start = s->y_superblock_count;
1548 s->v_superblock_start = s->u_superblock_start + s->c_superblock_count;
1549 s->superblock_coding = av_malloc(s->superblock_count);
1551 s->macroblock_width = (s->width + 15) / 16;
1552 s->macroblock_height = (s->height + 15) / 16;
1553 s->macroblock_count = s->macroblock_width * s->macroblock_height;
1555 s->fragment_width[0] = s->width / FRAGMENT_PIXELS;
1556 s->fragment_height[0] = s->height / FRAGMENT_PIXELS;
1557 s->fragment_width[1] = s->fragment_width[0] >> s->chroma_x_shift;
1558 s->fragment_height[1] = s->fragment_height[0] >> s->chroma_y_shift;
1560 /* fragment count covers all 8x8 blocks for all 3 planes */
1561 y_fragment_count = s->fragment_width[0] * s->fragment_height[0];
1562 c_fragment_count = s->fragment_width[1] * s->fragment_height[1];
1563 s->fragment_count = y_fragment_count + 2*c_fragment_count;
1564 s->fragment_start[1] = y_fragment_count;
1565 s->fragment_start[2] = y_fragment_count + c_fragment_count;
1567 s->all_fragments = av_malloc(s->fragment_count * sizeof(Vp3Fragment));
1568 s->coded_fragment_list[0] = av_malloc(s->fragment_count * sizeof(int));
1569 s->dct_tokens_base = av_malloc(64*s->fragment_count * sizeof(*s->dct_tokens_base));
1570 s->motion_val[0] = av_malloc(y_fragment_count * sizeof(*s->motion_val[0]));
1571 s->motion_val[1] = av_malloc(c_fragment_count * sizeof(*s->motion_val[1]));
1573 if (!s->superblock_coding || !s->all_fragments || !s->dct_tokens_base ||
1574 !s->coded_fragment_list[0] || !s->motion_val[0] || !s->motion_val[1]) {
1575 vp3_decode_end(avctx);
1579 if (!s->theora_tables)
1581 for (i = 0; i < 64; i++) {
1582 s->coded_dc_scale_factor[i] = vp31_dc_scale_factor[i];
1583 s->coded_ac_scale_factor[i] = vp31_ac_scale_factor[i];
1584 s->base_matrix[0][i] = vp31_intra_y_dequant[i];
1585 s->base_matrix[1][i] = vp31_intra_c_dequant[i];
1586 s->base_matrix[2][i] = vp31_inter_dequant[i];
1587 s->filter_limit_values[i] = vp31_filter_limit_values[i];
1590 for(inter=0; inter<2; inter++){
1591 for(plane=0; plane<3; plane++){
1592 s->qr_count[inter][plane]= 1;
1593 s->qr_size [inter][plane][0]= 63;
1594 s->qr_base [inter][plane][0]=
1595 s->qr_base [inter][plane][1]= 2*inter + (!!plane)*!inter;
1599 /* init VLC tables */
1600 for (i = 0; i < 16; i++) {
1603 init_vlc(&s->dc_vlc[i], 5, 32,
1604 &dc_bias[i][0][1], 4, 2,
1605 &dc_bias[i][0][0], 4, 2, 0);
1607 /* group 1 AC histograms */
1608 init_vlc(&s->ac_vlc_1[i], 5, 32,
1609 &ac_bias_0[i][0][1], 4, 2,
1610 &ac_bias_0[i][0][0], 4, 2, 0);
1612 /* group 2 AC histograms */
1613 init_vlc(&s->ac_vlc_2[i], 5, 32,
1614 &ac_bias_1[i][0][1], 4, 2,
1615 &ac_bias_1[i][0][0], 4, 2, 0);
1617 /* group 3 AC histograms */
1618 init_vlc(&s->ac_vlc_3[i], 5, 32,
1619 &ac_bias_2[i][0][1], 4, 2,
1620 &ac_bias_2[i][0][0], 4, 2, 0);
1622 /* group 4 AC histograms */
1623 init_vlc(&s->ac_vlc_4[i], 5, 32,
1624 &ac_bias_3[i][0][1], 4, 2,
1625 &ac_bias_3[i][0][0], 4, 2, 0);
1628 for (i = 0; i < 16; i++) {
1631 if (init_vlc(&s->dc_vlc[i], 5, 32,
1632 &s->huffman_table[i][0][1], 4, 2,
1633 &s->huffman_table[i][0][0], 4, 2, 0) < 0)
1636 /* group 1 AC histograms */
1637 if (init_vlc(&s->ac_vlc_1[i], 5, 32,
1638 &s->huffman_table[i+16][0][1], 4, 2,
1639 &s->huffman_table[i+16][0][0], 4, 2, 0) < 0)
1642 /* group 2 AC histograms */
1643 if (init_vlc(&s->ac_vlc_2[i], 5, 32,
1644 &s->huffman_table[i+16*2][0][1], 4, 2,
1645 &s->huffman_table[i+16*2][0][0], 4, 2, 0) < 0)
1648 /* group 3 AC histograms */
1649 if (init_vlc(&s->ac_vlc_3[i], 5, 32,
1650 &s->huffman_table[i+16*3][0][1], 4, 2,
1651 &s->huffman_table[i+16*3][0][0], 4, 2, 0) < 0)
1654 /* group 4 AC histograms */
1655 if (init_vlc(&s->ac_vlc_4[i], 5, 32,
1656 &s->huffman_table[i+16*4][0][1], 4, 2,
1657 &s->huffman_table[i+16*4][0][0], 4, 2, 0) < 0)
1662 init_vlc(&s->superblock_run_length_vlc, 6, 34,
1663 &superblock_run_length_vlc_table[0][1], 4, 2,
1664 &superblock_run_length_vlc_table[0][0], 4, 2, 0);
1666 init_vlc(&s->fragment_run_length_vlc, 5, 30,
1667 &fragment_run_length_vlc_table[0][1], 4, 2,
1668 &fragment_run_length_vlc_table[0][0], 4, 2, 0);
1670 init_vlc(&s->mode_code_vlc, 3, 8,
1671 &mode_code_vlc_table[0][1], 2, 1,
1672 &mode_code_vlc_table[0][0], 2, 1, 0);
1674 init_vlc(&s->motion_vector_vlc, 6, 63,
1675 &motion_vector_vlc_table[0][1], 2, 1,
1676 &motion_vector_vlc_table[0][0], 2, 1, 0);
1678 /* work out the block mapping tables */
1679 s->superblock_fragments = av_malloc(s->superblock_count * 16 * sizeof(int));
1680 s->macroblock_coding = av_malloc(s->macroblock_count + 1);
1681 if (!s->superblock_fragments || !s->macroblock_coding) {
1682 vp3_decode_end(avctx);
1685 init_block_mapping(s);
1687 for (i = 0; i < 3; i++) {
1688 s->current_frame.data[i] = NULL;
1689 s->last_frame.data[i] = NULL;
1690 s->golden_frame.data[i] = NULL;
1696 av_log(avctx, AV_LOG_FATAL, "Invalid huffman table\n");
1701 * This is the ffmpeg/libavcodec API frame decode function.
1703 static int vp3_decode_frame(AVCodecContext *avctx,
1704 void *data, int *data_size,
1707 const uint8_t *buf = avpkt->data;
1708 int buf_size = avpkt->size;
1709 Vp3DecodeContext *s = avctx->priv_data;
1711 static int counter = 0;
1714 init_get_bits(&gb, buf, buf_size * 8);
1716 if (s->theora && get_bits1(&gb))
1718 av_log(avctx, AV_LOG_ERROR, "Header packet passed to frame decoder, skipping\n");
1722 s->keyframe = !get_bits1(&gb);
1725 for (i = 0; i < 3; i++)
1726 s->last_qps[i] = s->qps[i];
1730 s->qps[s->nqps++]= get_bits(&gb, 6);
1731 } while(s->theora >= 0x030200 && s->nqps<3 && get_bits1(&gb));
1732 for (i = s->nqps; i < 3; i++)
1735 if (s->avctx->debug & FF_DEBUG_PICT_INFO)
1736 av_log(s->avctx, AV_LOG_INFO, " VP3 %sframe #%d: Q index = %d\n",
1737 s->keyframe?"key":"", counter, s->qps[0]);
1740 if (s->qps[0] != s->last_qps[0])
1741 init_loop_filter(s);
1743 for (i = 0; i < s->nqps; i++)
1744 // reinit all dequantizers if the first one changed, because
1745 // the DC of the first quantizer must be used for all matrices
1746 if (s->qps[i] != s->last_qps[i] || s->qps[0] != s->last_qps[0])
1747 init_dequantizer(s, i);
1749 if (avctx->skip_frame >= AVDISCARD_NONKEY && !s->keyframe)
1752 s->current_frame.reference = 3;
1753 s->current_frame.pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1754 if (avctx->get_buffer(avctx, &s->current_frame) < 0) {
1755 av_log(s->avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1762 skip_bits(&gb, 4); /* width code */
1763 skip_bits(&gb, 4); /* height code */
1766 s->version = get_bits(&gb, 5);
1768 av_log(s->avctx, AV_LOG_DEBUG, "VP version: %d\n", s->version);
1771 if (s->version || s->theora)
1774 av_log(s->avctx, AV_LOG_ERROR, "Warning, unsupported keyframe coding type?!\n");
1775 skip_bits(&gb, 2); /* reserved? */
1778 if (!s->golden_frame.data[0]) {
1779 av_log(s->avctx, AV_LOG_WARNING, "vp3: first frame not a keyframe\n");
1781 s->golden_frame.reference = 3;
1782 s->golden_frame.pict_type = FF_I_TYPE;
1783 if (avctx->get_buffer(avctx, &s->golden_frame) < 0) {
1784 av_log(s->avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1787 s->last_frame = s->golden_frame;
1788 s->last_frame.type = FF_BUFFER_TYPE_COPY;
1792 s->current_frame.qscale_table= s->qscale_table; //FIXME allocate individual tables per AVFrame
1793 s->current_frame.qstride= 0;
1795 memset(s->all_fragments, 0, s->fragment_count * sizeof(Vp3Fragment));
1797 if (unpack_superblocks(s, &gb)){
1798 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_superblocks\n");
1801 if (unpack_modes(s, &gb)){
1802 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_modes\n");
1805 if (unpack_vectors(s, &gb)){
1806 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_vectors\n");
1809 if (unpack_block_qpis(s, &gb)){
1810 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_block_qpis\n");
1813 if (unpack_dct_coeffs(s, &gb)){
1814 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_dct_coeffs\n");
1818 for (i = 0; i < 3; i++) {
1819 int height = s->height >> (i && s->chroma_y_shift);
1820 if (s->flipped_image)
1821 s->data_offset[i] = 0;
1823 s->data_offset[i] = (height-1) * s->current_frame.linesize[i];
1826 s->last_slice_end = 0;
1827 for (i = 0; i < s->c_superblock_height; i++)
1830 // filter the last row
1831 for (i = 0; i < 3; i++) {
1832 int row = (s->height >> (3+(i && s->chroma_y_shift))) - 1;
1833 apply_loop_filter(s, i, row, row+1);
1835 vp3_draw_horiz_band(s, s->height);
1837 *data_size=sizeof(AVFrame);
1838 *(AVFrame*)data= s->current_frame;
1840 /* release the last frame, if it is allocated and if it is not the
1842 if (s->last_frame.data[0] && s->last_frame.type != FF_BUFFER_TYPE_COPY)
1843 avctx->release_buffer(avctx, &s->last_frame);
1845 /* shuffle frames (last = current) */
1846 s->last_frame= s->current_frame;
1849 if (s->golden_frame.data[0])
1850 avctx->release_buffer(avctx, &s->golden_frame);
1851 s->golden_frame = s->current_frame;
1852 s->last_frame.type = FF_BUFFER_TYPE_COPY;
1855 s->current_frame.data[0]= NULL; /* ensure that we catch any access to this released frame */
1860 if (s->current_frame.data[0])
1861 avctx->release_buffer(avctx, &s->current_frame);
1866 * This is the ffmpeg/libavcodec API module cleanup function.
1868 static av_cold int vp3_decode_end(AVCodecContext *avctx)
1870 Vp3DecodeContext *s = avctx->priv_data;
1873 av_free(s->superblock_coding);
1874 av_free(s->all_fragments);
1875 av_free(s->coded_fragment_list[0]);
1876 av_free(s->dct_tokens_base);
1877 av_free(s->superblock_fragments);
1878 av_free(s->macroblock_coding);
1879 av_free(s->motion_val[0]);
1880 av_free(s->motion_val[1]);
1882 for (i = 0; i < 16; i++) {
1883 free_vlc(&s->dc_vlc[i]);
1884 free_vlc(&s->ac_vlc_1[i]);
1885 free_vlc(&s->ac_vlc_2[i]);
1886 free_vlc(&s->ac_vlc_3[i]);
1887 free_vlc(&s->ac_vlc_4[i]);
1890 free_vlc(&s->superblock_run_length_vlc);
1891 free_vlc(&s->fragment_run_length_vlc);
1892 free_vlc(&s->mode_code_vlc);
1893 free_vlc(&s->motion_vector_vlc);
1895 /* release all frames */
1896 if (s->golden_frame.data[0])
1897 avctx->release_buffer(avctx, &s->golden_frame);
1898 if (s->last_frame.data[0] && s->last_frame.type != FF_BUFFER_TYPE_COPY)
1899 avctx->release_buffer(avctx, &s->last_frame);
1900 /* no need to release the current_frame since it will always be pointing
1901 * to the same frame as either the golden or last frame */
1906 static int read_huffman_tree(AVCodecContext *avctx, GetBitContext *gb)
1908 Vp3DecodeContext *s = avctx->priv_data;
1910 if (get_bits1(gb)) {
1912 if (s->entries >= 32) { /* overflow */
1913 av_log(avctx, AV_LOG_ERROR, "huffman tree overflow\n");
1916 token = get_bits(gb, 5);
1917 //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);
1918 s->huffman_table[s->hti][token][0] = s->hbits;
1919 s->huffman_table[s->hti][token][1] = s->huff_code_size;
1923 if (s->huff_code_size >= 32) {/* overflow */
1924 av_log(avctx, AV_LOG_ERROR, "huffman tree overflow\n");
1927 s->huff_code_size++;
1929 if (read_huffman_tree(avctx, gb))
1932 if (read_huffman_tree(avctx, gb))
1935 s->huff_code_size--;
1940 #if CONFIG_THEORA_DECODER
1941 static const enum PixelFormat theora_pix_fmts[4] = {
1942 PIX_FMT_YUV420P, PIX_FMT_NONE, PIX_FMT_YUV422P, PIX_FMT_YUV444P
1945 static int theora_decode_header(AVCodecContext *avctx, GetBitContext *gb)
1947 Vp3DecodeContext *s = avctx->priv_data;
1948 int visible_width, visible_height, colorspace;
1950 s->theora = get_bits_long(gb, 24);
1951 av_log(avctx, AV_LOG_DEBUG, "Theora bitstream version %X\n", s->theora);
1953 /* 3.2.0 aka alpha3 has the same frame orientation as original vp3 */
1954 /* but previous versions have the image flipped relative to vp3 */
1955 if (s->theora < 0x030200)
1957 s->flipped_image = 1;
1958 av_log(avctx, AV_LOG_DEBUG, "Old (<alpha3) Theora bitstream, flipped image\n");
1961 visible_width = s->width = get_bits(gb, 16) << 4;
1962 visible_height = s->height = get_bits(gb, 16) << 4;
1964 if(avcodec_check_dimensions(avctx, s->width, s->height)){
1965 av_log(avctx, AV_LOG_ERROR, "Invalid dimensions (%dx%d)\n", s->width, s->height);
1966 s->width= s->height= 0;
1970 if (s->theora >= 0x030200) {
1971 visible_width = get_bits_long(gb, 24);
1972 visible_height = get_bits_long(gb, 24);
1974 skip_bits(gb, 8); /* offset x */
1975 skip_bits(gb, 8); /* offset y */
1978 skip_bits(gb, 32); /* fps numerator */
1979 skip_bits(gb, 32); /* fps denumerator */
1980 skip_bits(gb, 24); /* aspect numerator */
1981 skip_bits(gb, 24); /* aspect denumerator */
1983 if (s->theora < 0x030200)
1984 skip_bits(gb, 5); /* keyframe frequency force */
1985 colorspace = get_bits(gb, 8);
1986 skip_bits(gb, 24); /* bitrate */
1988 skip_bits(gb, 6); /* quality hint */
1990 if (s->theora >= 0x030200)
1992 skip_bits(gb, 5); /* keyframe frequency force */
1993 avctx->pix_fmt = theora_pix_fmts[get_bits(gb, 2)];
1994 skip_bits(gb, 3); /* reserved */
1997 // align_get_bits(gb);
1999 if ( visible_width <= s->width && visible_width > s->width-16
2000 && visible_height <= s->height && visible_height > s->height-16)
2001 avcodec_set_dimensions(avctx, visible_width, visible_height);
2003 avcodec_set_dimensions(avctx, s->width, s->height);
2005 if (colorspace == 1) {
2006 avctx->color_primaries = AVCOL_PRI_BT470M;
2007 } else if (colorspace == 2) {
2008 avctx->color_primaries = AVCOL_PRI_BT470BG;
2010 if (colorspace == 1 || colorspace == 2) {
2011 avctx->colorspace = AVCOL_SPC_BT470BG;
2012 avctx->color_trc = AVCOL_TRC_BT709;
2018 static int theora_decode_tables(AVCodecContext *avctx, GetBitContext *gb)
2020 Vp3DecodeContext *s = avctx->priv_data;
2021 int i, n, matrices, inter, plane;
2023 if (s->theora >= 0x030200) {
2024 n = get_bits(gb, 3);
2025 /* loop filter limit values table */
2026 for (i = 0; i < 64; i++) {
2027 s->filter_limit_values[i] = get_bits(gb, n);
2028 if (s->filter_limit_values[i] > 127) {
2029 av_log(avctx, AV_LOG_ERROR, "filter limit value too large (%i > 127), clamping\n", s->filter_limit_values[i]);
2030 s->filter_limit_values[i] = 127;
2035 if (s->theora >= 0x030200)
2036 n = get_bits(gb, 4) + 1;
2039 /* quality threshold table */
2040 for (i = 0; i < 64; i++)
2041 s->coded_ac_scale_factor[i] = get_bits(gb, n);
2043 if (s->theora >= 0x030200)
2044 n = get_bits(gb, 4) + 1;
2047 /* dc scale factor table */
2048 for (i = 0; i < 64; i++)
2049 s->coded_dc_scale_factor[i] = get_bits(gb, n);
2051 if (s->theora >= 0x030200)
2052 matrices = get_bits(gb, 9) + 1;
2057 av_log(avctx, AV_LOG_ERROR, "invalid number of base matrixes\n");
2061 for(n=0; n<matrices; n++){
2062 for (i = 0; i < 64; i++)
2063 s->base_matrix[n][i]= get_bits(gb, 8);
2066 for (inter = 0; inter <= 1; inter++) {
2067 for (plane = 0; plane <= 2; plane++) {
2069 if (inter || plane > 0)
2070 newqr = get_bits1(gb);
2073 if(inter && get_bits1(gb)){
2077 qtj= (3*inter + plane - 1) / 3;
2078 plj= (plane + 2) % 3;
2080 s->qr_count[inter][plane]= s->qr_count[qtj][plj];
2081 memcpy(s->qr_size[inter][plane], s->qr_size[qtj][plj], sizeof(s->qr_size[0][0]));
2082 memcpy(s->qr_base[inter][plane], s->qr_base[qtj][plj], sizeof(s->qr_base[0][0]));
2088 i= get_bits(gb, av_log2(matrices-1)+1);
2090 av_log(avctx, AV_LOG_ERROR, "invalid base matrix index\n");
2093 s->qr_base[inter][plane][qri]= i;
2096 i = get_bits(gb, av_log2(63-qi)+1) + 1;
2097 s->qr_size[inter][plane][qri++]= i;
2102 av_log(avctx, AV_LOG_ERROR, "invalid qi %d > 63\n", qi);
2105 s->qr_count[inter][plane]= qri;
2110 /* Huffman tables */
2111 for (s->hti = 0; s->hti < 80; s->hti++) {
2113 s->huff_code_size = 1;
2114 if (!get_bits1(gb)) {
2116 if(read_huffman_tree(avctx, gb))
2119 if(read_huffman_tree(avctx, gb))
2124 s->theora_tables = 1;
2129 static av_cold int theora_decode_init(AVCodecContext *avctx)
2131 Vp3DecodeContext *s = avctx->priv_data;
2134 uint8_t *header_start[3];
2140 if (!avctx->extradata_size)
2142 av_log(avctx, AV_LOG_ERROR, "Missing extradata!\n");
2146 if (ff_split_xiph_headers(avctx->extradata, avctx->extradata_size,
2147 42, header_start, header_len) < 0) {
2148 av_log(avctx, AV_LOG_ERROR, "Corrupt extradata\n");
2153 init_get_bits(&gb, header_start[i], header_len[i] * 8);
2155 ptype = get_bits(&gb, 8);
2157 if (!(ptype & 0x80))
2159 av_log(avctx, AV_LOG_ERROR, "Invalid extradata!\n");
2163 // FIXME: Check for this as well.
2164 skip_bits_long(&gb, 6*8); /* "theora" */
2169 theora_decode_header(avctx, &gb);
2172 // FIXME: is this needed? it breaks sometimes
2173 // theora_decode_comments(avctx, gb);
2176 if (theora_decode_tables(avctx, &gb))
2180 av_log(avctx, AV_LOG_ERROR, "Unknown Theora config packet: %d\n", ptype&~0x80);
2183 if(ptype != 0x81 && 8*header_len[i] != get_bits_count(&gb))
2184 av_log(avctx, AV_LOG_WARNING, "%d bits left in packet %X\n", 8*header_len[i] - get_bits_count(&gb), ptype);
2185 if (s->theora < 0x030200)
2189 return vp3_decode_init(avctx);
2192 AVCodec theora_decoder = {
2196 sizeof(Vp3DecodeContext),
2201 CODEC_CAP_DR1 | CODEC_CAP_DRAW_HORIZ_BAND,
2203 .long_name = NULL_IF_CONFIG_SMALL("Theora"),
2207 AVCodec vp3_decoder = {
2211 sizeof(Vp3DecodeContext),
2216 CODEC_CAP_DR1 | CODEC_CAP_DRAW_HORIZ_BAND,
2218 .long_name = NULL_IF_CONFIG_SMALL("On2 VP3"),