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;
56 #define SB_NOT_CODED 0
57 #define SB_PARTIALLY_CODED 1
58 #define SB_FULLY_CODED 2
60 // This is the maximum length of a single long bit run that can be encoded
61 // for superblock coding or block qps. Theora special-cases this to read a
62 // bit instead of flipping the current bit to allow for runs longer than 4129.
63 #define MAXIMUM_LONG_BIT_RUN 4129
65 #define MODE_INTER_NO_MV 0
67 #define MODE_INTER_PLUS_MV 2
68 #define MODE_INTER_LAST_MV 3
69 #define MODE_INTER_PRIOR_LAST 4
70 #define MODE_USING_GOLDEN 5
71 #define MODE_GOLDEN_MV 6
72 #define MODE_INTER_FOURMV 7
73 #define CODING_MODE_COUNT 8
75 /* special internal mode */
78 /* There are 6 preset schemes, plus a free-form scheme */
79 static const int ModeAlphabet[6][CODING_MODE_COUNT] =
81 /* scheme 1: Last motion vector dominates */
82 { MODE_INTER_LAST_MV, MODE_INTER_PRIOR_LAST,
83 MODE_INTER_PLUS_MV, MODE_INTER_NO_MV,
84 MODE_INTRA, MODE_USING_GOLDEN,
85 MODE_GOLDEN_MV, MODE_INTER_FOURMV },
88 { MODE_INTER_LAST_MV, MODE_INTER_PRIOR_LAST,
89 MODE_INTER_NO_MV, MODE_INTER_PLUS_MV,
90 MODE_INTRA, MODE_USING_GOLDEN,
91 MODE_GOLDEN_MV, MODE_INTER_FOURMV },
94 { MODE_INTER_LAST_MV, MODE_INTER_PLUS_MV,
95 MODE_INTER_PRIOR_LAST, MODE_INTER_NO_MV,
96 MODE_INTRA, MODE_USING_GOLDEN,
97 MODE_GOLDEN_MV, MODE_INTER_FOURMV },
100 { MODE_INTER_LAST_MV, MODE_INTER_PLUS_MV,
101 MODE_INTER_NO_MV, MODE_INTER_PRIOR_LAST,
102 MODE_INTRA, MODE_USING_GOLDEN,
103 MODE_GOLDEN_MV, MODE_INTER_FOURMV },
105 /* scheme 5: No motion vector dominates */
106 { MODE_INTER_NO_MV, MODE_INTER_LAST_MV,
107 MODE_INTER_PRIOR_LAST, MODE_INTER_PLUS_MV,
108 MODE_INTRA, MODE_USING_GOLDEN,
109 MODE_GOLDEN_MV, MODE_INTER_FOURMV },
112 { MODE_INTER_NO_MV, MODE_USING_GOLDEN,
113 MODE_INTER_LAST_MV, MODE_INTER_PRIOR_LAST,
114 MODE_INTER_PLUS_MV, MODE_INTRA,
115 MODE_GOLDEN_MV, MODE_INTER_FOURMV },
119 static const uint8_t hilbert_offset[16][2] = {
120 {0,0}, {1,0}, {1,1}, {0,1},
121 {0,2}, {0,3}, {1,3}, {1,2},
122 {2,2}, {2,3}, {3,3}, {3,2},
123 {3,1}, {2,1}, {2,0}, {3,0}
126 #define MIN_DEQUANT_VAL 2
128 typedef struct Vp3DecodeContext {
129 AVCodecContext *avctx;
130 int theora, theora_tables;
133 AVFrame golden_frame;
135 AVFrame current_frame;
145 int superblock_count;
146 int y_superblock_width;
147 int y_superblock_height;
148 int y_superblock_count;
149 int c_superblock_width;
150 int c_superblock_height;
151 int c_superblock_count;
152 int u_superblock_start;
153 int v_superblock_start;
154 unsigned char *superblock_coding;
156 int macroblock_count;
157 int macroblock_width;
158 int macroblock_height;
164 Vp3Fragment *all_fragments;
165 int fragment_start[3];
171 uint16_t coded_dc_scale_factor[64];
172 uint32_t coded_ac_scale_factor[64];
173 uint8_t base_matrix[384][64];
174 uint8_t qr_count[2][3];
175 uint8_t qr_size [2][3][64];
176 uint16_t qr_base[2][3][64];
179 * This is a list of all tokens in bitstream order. Reordering takes place
180 * by pulling from each level during IDCT. As a consequence, IDCT must be
181 * in Hilbert order, making the minimum slice height 64 for 4:2:0 and 32
182 * otherwise. The 32 different tokens with up to 12 bits of extradata are
183 * collapsed into 3 types, packed as follows:
184 * (from the low to high bits)
186 * 2 bits: type (0,1,2)
187 * 0: EOB run, 14 bits for run length (12 needed)
188 * 1: zero run, 7 bits for run length
189 * 7 bits for the next coefficient (3 needed)
190 * 2: coefficient, 14 bits (11 needed)
192 * Coefficients are signed, so are packed in the highest bits for automatic
195 int16_t *dct_tokens[3][64];
196 int16_t *dct_tokens_base;
197 #define TOKEN_EOB(eob_run) ((eob_run) << 2)
198 #define TOKEN_ZERO_RUN(coeff, zero_run) (((coeff) << 9) + ((zero_run) << 2) + 1)
199 #define TOKEN_COEFF(coeff) (((coeff) << 2) + 2)
202 * number of blocks that contain DCT coefficients at the given level or higher
204 int num_coded_frags[3][64];
205 int total_num_coded_frags;
207 /* this is a list of indexes into the all_fragments array indicating
208 * which of the fragments are coded */
209 int *coded_fragment_list[3];
217 VLC superblock_run_length_vlc;
218 VLC fragment_run_length_vlc;
220 VLC motion_vector_vlc;
222 /* these arrays need to be on 16-byte boundaries since SSE2 operations
224 DECLARE_ALIGNED(16, int16_t, qmat)[3][2][3][64]; //<qmat[qpi][is_inter][plane]
226 /* This table contains superblock_count * 16 entries. Each set of 16
227 * numbers corresponds to the fragment indexes 0..15 of the superblock.
228 * An entry will be -1 to indicate that no entry corresponds to that
230 int *superblock_fragments;
232 /* This is an array that indicates how a particular macroblock
234 unsigned char *macroblock_coding;
236 uint8_t edge_emu_buffer[9*2048]; //FIXME dynamic alloc
237 int8_t qscale_table[2048]; //FIXME dynamic alloc (width+15)/16
244 uint16_t huffman_table[80][32][2];
246 uint8_t filter_limit_values[64];
247 DECLARE_ALIGNED(8, int, bounding_values_array)[256+2];
250 /************************************************************************
251 * VP3 specific functions
252 ************************************************************************/
255 * This function sets up all of the various blocks mappings:
256 * superblocks <-> fragments, macroblocks <-> fragments,
257 * superblocks <-> macroblocks
259 * Returns 0 is successful; returns 1 if *anything* went wrong.
261 static int init_block_mapping(Vp3DecodeContext *s)
264 signed int hilbert_walk_mb[4];
266 int current_fragment = 0;
267 int current_width = 0;
268 int current_height = 0;
271 int superblock_row_inc = 0;
272 int mapping_index = 0;
274 static const signed char travel_width[16] = {
281 static const signed char travel_height[16] = {
288 hilbert_walk_mb[0] = 1;
289 hilbert_walk_mb[1] = s->macroblock_width;
290 hilbert_walk_mb[2] = 1;
291 hilbert_walk_mb[3] = -s->macroblock_width;
293 /* iterate through each superblock (all planes) and map the fragments */
294 for (i = 0; i < s->superblock_count; i++) {
295 /* time to re-assign the limits? */
298 /* start of Y superblocks */
299 right_edge = s->fragment_width;
300 bottom_edge = s->fragment_height;
303 superblock_row_inc = 3 * s->fragment_width -
304 (s->y_superblock_width * 4 - s->fragment_width);
306 /* the first operation for this variable is to advance by 1 */
307 current_fragment = -1;
309 } else if (i == s->u_superblock_start) {
311 /* start of U superblocks */
312 right_edge = s->fragment_width / 2;
313 bottom_edge = s->fragment_height / 2;
316 superblock_row_inc = 3 * (s->fragment_width / 2) -
317 (s->c_superblock_width * 4 - s->fragment_width / 2);
319 /* the first operation for this variable is to advance by 1 */
320 current_fragment = s->fragment_start[1] - 1;
322 } else if (i == s->v_superblock_start) {
324 /* start of V superblocks */
325 right_edge = s->fragment_width / 2;
326 bottom_edge = s->fragment_height / 2;
329 superblock_row_inc = 3 * (s->fragment_width / 2) -
330 (s->c_superblock_width * 4 - s->fragment_width / 2);
332 /* the first operation for this variable is to advance by 1 */
333 current_fragment = s->fragment_start[2] - 1;
337 if (current_width >= right_edge - 1) {
338 /* reset width and move to next superblock row */
342 /* fragment is now at the start of a new superblock row */
343 current_fragment += superblock_row_inc;
346 /* iterate through all 16 fragments in a superblock */
347 for (j = 0; j < 16; j++) {
348 current_fragment += travel_width[j] + right_edge * travel_height[j];
349 current_width += travel_width[j];
350 current_height += travel_height[j];
352 /* check if the fragment is in bounds */
353 if ((current_width < right_edge) &&
354 (current_height < bottom_edge)) {
355 s->superblock_fragments[mapping_index] = current_fragment;
357 s->superblock_fragments[mapping_index] = -1;
364 return 0; /* successful path out */
368 * This function wipes out all of the fragment data.
370 static void init_frame(Vp3DecodeContext *s, GetBitContext *gb)
374 /* zero out all of the fragment information */
375 for (i = 0; i < s->fragment_count; i++) {
376 s->all_fragments[i].motion_x = 127;
377 s->all_fragments[i].motion_y = 127;
378 s->all_fragments[i].dc = 0;
379 s->all_fragments[i].qpi = 0;
384 * This function sets up the dequantization tables used for a particular
387 static void init_dequantizer(Vp3DecodeContext *s, int qpi)
389 int ac_scale_factor = s->coded_ac_scale_factor[s->qps[qpi]];
390 int dc_scale_factor = s->coded_dc_scale_factor[s->qps[qpi]];
391 int i, plane, inter, qri, bmi, bmj, qistart;
393 for(inter=0; inter<2; inter++){
394 for(plane=0; plane<3; plane++){
396 for(qri=0; qri<s->qr_count[inter][plane]; qri++){
397 sum+= s->qr_size[inter][plane][qri];
398 if(s->qps[qpi] <= sum)
401 qistart= sum - s->qr_size[inter][plane][qri];
402 bmi= s->qr_base[inter][plane][qri ];
403 bmj= s->qr_base[inter][plane][qri+1];
405 int coeff= ( 2*(sum -s->qps[qpi])*s->base_matrix[bmi][i]
406 - 2*(qistart-s->qps[qpi])*s->base_matrix[bmj][i]
407 + s->qr_size[inter][plane][qri])
408 / (2*s->qr_size[inter][plane][qri]);
410 int qmin= 8<<(inter + !i);
411 int qscale= i ? ac_scale_factor : dc_scale_factor;
413 s->qmat[qpi][inter][plane][s->dsp.idct_permutation[i]]= av_clip((qscale * coeff)/100 * 4, qmin, 4096);
415 // all DC coefficients use the same quant so as not to interfere with DC prediction
416 s->qmat[qpi][inter][plane][0] = s->qmat[0][inter][plane][0];
420 memset(s->qscale_table, (FFMAX(s->qmat[0][0][0][1], s->qmat[0][0][1][1])+8)/16, 512); //FIXME finetune
424 * This function initializes the loop filter boundary limits if the frame's
425 * quality index is different from the previous frame's.
427 * The filter_limit_values may not be larger than 127.
429 static void init_loop_filter(Vp3DecodeContext *s)
431 int *bounding_values= s->bounding_values_array+127;
436 filter_limit = s->filter_limit_values[s->qps[0]];
438 /* set up the bounding values */
439 memset(s->bounding_values_array, 0, 256 * sizeof(int));
440 for (x = 0; x < filter_limit; x++) {
441 bounding_values[-x] = -x;
442 bounding_values[x] = x;
444 for (x = value = filter_limit; x < 128 && value; x++, value--) {
445 bounding_values[ x] = value;
446 bounding_values[-x] = -value;
449 bounding_values[128] = value;
450 bounding_values[129] = bounding_values[130] = filter_limit * 0x02020202;
454 * This function unpacks all of the superblock/macroblock/fragment coding
455 * information from the bitstream.
457 static int unpack_superblocks(Vp3DecodeContext *s, GetBitContext *gb)
459 int superblock_starts[3] = { 0, s->u_superblock_start, s->v_superblock_start };
461 int current_superblock = 0;
463 int num_partial_superblocks = 0;
466 int current_fragment;
470 memset(s->superblock_coding, SB_FULLY_CODED, s->superblock_count);
474 /* unpack the list of partially-coded superblocks */
476 while (current_superblock < s->superblock_count) {
477 current_run = get_vlc2(gb,
478 s->superblock_run_length_vlc.table, 6, 2) + 1;
479 if (current_run == 34)
480 current_run += get_bits(gb, 12);
482 if (current_superblock + current_run > s->superblock_count) {
483 av_log(s->avctx, AV_LOG_ERROR, "Invalid partially coded superblock run length\n");
487 memset(s->superblock_coding + current_superblock, bit, current_run);
489 current_superblock += current_run;
491 num_partial_superblocks += current_run;
493 if (s->theora && current_run == MAXIMUM_LONG_BIT_RUN)
499 /* unpack the list of fully coded superblocks if any of the blocks were
500 * not marked as partially coded in the previous step */
501 if (num_partial_superblocks < s->superblock_count) {
502 int superblocks_decoded = 0;
504 current_superblock = 0;
506 while (superblocks_decoded < s->superblock_count - num_partial_superblocks) {
507 current_run = get_vlc2(gb,
508 s->superblock_run_length_vlc.table, 6, 2) + 1;
509 if (current_run == 34)
510 current_run += get_bits(gb, 12);
512 for (j = 0; j < current_run; current_superblock++) {
513 if (current_superblock >= s->superblock_count) {
514 av_log(s->avctx, AV_LOG_ERROR, "Invalid fully coded superblock run length\n");
518 /* skip any superblocks already marked as partially coded */
519 if (s->superblock_coding[current_superblock] == SB_NOT_CODED) {
520 s->superblock_coding[current_superblock] = 2*bit;
524 superblocks_decoded += current_run;
526 if (s->theora && current_run == MAXIMUM_LONG_BIT_RUN)
533 /* if there were partial blocks, initialize bitstream for
534 * unpacking fragment codings */
535 if (num_partial_superblocks) {
539 /* toggle the bit because as soon as the first run length is
540 * fetched the bit will be toggled again */
545 /* figure out which fragments are coded; iterate through each
546 * superblock (all planes) */
547 s->total_num_coded_frags = 0;
548 memset(s->macroblock_coding, MODE_COPY, s->macroblock_count);
550 for (plane = 0; plane < 3; plane++) {
551 int sb_start = superblock_starts[plane];
552 int sb_end = sb_start + (plane ? s->c_superblock_count : s->y_superblock_count);
553 int num_coded_frags = 0;
555 for (i = sb_start; i < sb_end; i++) {
557 /* iterate through all 16 fragments in a superblock */
558 for (j = 0; j < 16; j++) {
560 /* if the fragment is in bounds, check its coding status */
561 current_fragment = s->superblock_fragments[i * 16 + j];
562 if (current_fragment >= s->fragment_count) {
563 av_log(s->avctx, AV_LOG_ERROR, " vp3:unpack_superblocks(): bad fragment number (%d >= %d)\n",
564 current_fragment, s->fragment_count);
567 if (current_fragment != -1) {
568 int coded = s->superblock_coding[i];
570 if (s->superblock_coding[i] == SB_PARTIALLY_CODED) {
572 /* fragment may or may not be coded; this is the case
573 * that cares about the fragment coding runs */
574 if (current_run-- == 0) {
576 current_run = get_vlc2(gb,
577 s->fragment_run_length_vlc.table, 5, 2);
583 /* default mode; actual mode will be decoded in
585 s->all_fragments[current_fragment].coding_method =
587 s->coded_fragment_list[plane][num_coded_frags++] =
590 /* not coded; copy this fragment from the prior frame */
591 s->all_fragments[current_fragment].coding_method =
597 s->total_num_coded_frags += num_coded_frags;
598 for (i = 0; i < 64; i++)
599 s->num_coded_frags[plane][i] = num_coded_frags;
601 s->coded_fragment_list[plane+1] = s->coded_fragment_list[plane] + num_coded_frags;
607 * This function unpacks all the coding mode data for individual macroblocks
608 * from the bitstream.
610 static int unpack_modes(Vp3DecodeContext *s, GetBitContext *gb)
612 int i, j, k, sb_x, sb_y;
614 int current_macroblock;
615 int current_fragment;
617 int custom_mode_alphabet[CODING_MODE_COUNT];
621 for (i = 0; i < s->fragment_count; i++)
622 s->all_fragments[i].coding_method = MODE_INTRA;
626 /* fetch the mode coding scheme for this frame */
627 scheme = get_bits(gb, 3);
629 /* is it a custom coding scheme? */
631 for (i = 0; i < 8; i++)
632 custom_mode_alphabet[i] = MODE_INTER_NO_MV;
633 for (i = 0; i < 8; i++)
634 custom_mode_alphabet[get_bits(gb, 3)] = i;
635 alphabet = custom_mode_alphabet;
637 alphabet = ModeAlphabet[scheme-1];
639 /* iterate through all of the macroblocks that contain 1 or more
641 for (sb_y = 0; sb_y < s->y_superblock_height; sb_y++) {
642 for (sb_x = 0; sb_x < s->y_superblock_width; sb_x++) {
644 for (j = 0; j < 4; j++) {
645 int mb_x = 2*sb_x + (j>>1);
646 int mb_y = 2*sb_y + (((j>>1)+j)&1);
647 current_macroblock = mb_y * s->macroblock_width + mb_x;
649 if (mb_x >= s->macroblock_width || mb_y >= s->macroblock_height)
652 #define BLOCK_X (2*mb_x + (k&1))
653 #define BLOCK_Y (2*mb_y + (k>>1))
654 /* coding modes are only stored if the macroblock has at least one
655 * luma block coded, otherwise it must be INTER_NO_MV */
656 for (k = 0; k < 4; k++) {
657 current_fragment = BLOCK_Y*s->fragment_width + BLOCK_X;
658 if (s->all_fragments[current_fragment].coding_method != MODE_COPY)
662 s->macroblock_coding[current_macroblock] = MODE_INTER_NO_MV;
666 /* mode 7 means get 3 bits for each coding mode */
668 coding_mode = get_bits(gb, 3);
670 coding_mode = alphabet
671 [get_vlc2(gb, s->mode_code_vlc.table, 3, 3)];
673 s->macroblock_coding[current_macroblock] = coding_mode;
674 for (k = 0; k < 4; k++) {
676 BLOCK_Y*s->fragment_width + BLOCK_X;
677 if (s->all_fragments[current_fragment].coding_method !=
679 s->all_fragments[current_fragment].coding_method =
682 for (k = 0; k < 2; k++) {
683 current_fragment = s->fragment_start[k+1] +
684 mb_y*(s->fragment_width>>1) + mb_x;
685 if (s->all_fragments[current_fragment].coding_method !=
687 s->all_fragments[current_fragment].coding_method =
699 * This function unpacks all the motion vectors for the individual
700 * macroblocks from the bitstream.
702 static int unpack_vectors(Vp3DecodeContext *s, GetBitContext *gb)
704 int j, k, sb_x, sb_y;
708 int last_motion_x = 0;
709 int last_motion_y = 0;
710 int prior_last_motion_x = 0;
711 int prior_last_motion_y = 0;
712 int current_macroblock;
713 int current_fragment;
718 memset(motion_x, 0, 6 * sizeof(int));
719 memset(motion_y, 0, 6 * sizeof(int));
721 /* coding mode 0 is the VLC scheme; 1 is the fixed code scheme */
722 coding_mode = get_bits1(gb);
724 /* iterate through all of the macroblocks that contain 1 or more
726 for (sb_y = 0; sb_y < s->y_superblock_height; sb_y++) {
727 for (sb_x = 0; sb_x < s->y_superblock_width; sb_x++) {
729 for (j = 0; j < 4; j++) {
730 int mb_x = 2*sb_x + (j>>1);
731 int mb_y = 2*sb_y + (((j>>1)+j)&1);
732 current_macroblock = mb_y * s->macroblock_width + mb_x;
734 if (mb_x >= s->macroblock_width || mb_y >= s->macroblock_height ||
735 (s->macroblock_coding[current_macroblock] == MODE_COPY))
738 switch (s->macroblock_coding[current_macroblock]) {
740 case MODE_INTER_PLUS_MV:
742 /* all 6 fragments use the same motion vector */
743 if (coding_mode == 0) {
744 motion_x[0] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
745 motion_y[0] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
747 motion_x[0] = fixed_motion_vector_table[get_bits(gb, 6)];
748 motion_y[0] = fixed_motion_vector_table[get_bits(gb, 6)];
751 /* vector maintenance, only on MODE_INTER_PLUS_MV */
752 if (s->macroblock_coding[current_macroblock] ==
753 MODE_INTER_PLUS_MV) {
754 prior_last_motion_x = last_motion_x;
755 prior_last_motion_y = last_motion_y;
756 last_motion_x = motion_x[0];
757 last_motion_y = motion_y[0];
761 case MODE_INTER_FOURMV:
762 /* vector maintenance */
763 prior_last_motion_x = last_motion_x;
764 prior_last_motion_y = last_motion_y;
766 /* fetch 4 vectors from the bitstream, one for each
767 * Y fragment, then average for the C fragment vectors */
768 motion_x[4] = motion_y[4] = 0;
769 for (k = 0; k < 4; k++) {
770 current_fragment = BLOCK_Y*s->fragment_width + BLOCK_X;
771 if (s->all_fragments[current_fragment].coding_method != MODE_COPY) {
772 if (coding_mode == 0) {
773 motion_x[k] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
774 motion_y[k] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
776 motion_x[k] = fixed_motion_vector_table[get_bits(gb, 6)];
777 motion_y[k] = fixed_motion_vector_table[get_bits(gb, 6)];
779 last_motion_x = motion_x[k];
780 last_motion_y = motion_y[k];
785 motion_x[4] += motion_x[k];
786 motion_y[4] += motion_y[k];
790 motion_x[4]= RSHIFT(motion_x[4], 2);
792 motion_y[4]= RSHIFT(motion_y[4], 2);
795 case MODE_INTER_LAST_MV:
796 /* all 6 fragments use the last motion vector */
797 motion_x[0] = last_motion_x;
798 motion_y[0] = last_motion_y;
800 /* no vector maintenance (last vector remains the
804 case MODE_INTER_PRIOR_LAST:
805 /* all 6 fragments use the motion vector prior to the
806 * last motion vector */
807 motion_x[0] = prior_last_motion_x;
808 motion_y[0] = prior_last_motion_y;
810 /* vector maintenance */
811 prior_last_motion_x = last_motion_x;
812 prior_last_motion_y = last_motion_y;
813 last_motion_x = motion_x[0];
814 last_motion_y = motion_y[0];
818 /* covers intra, inter without MV, golden without MV */
822 /* no vector maintenance */
826 /* assign the motion vectors to the correct fragments */
827 for (k = 0; k < 4; k++) {
829 BLOCK_Y*s->fragment_width + BLOCK_X;
830 if (s->macroblock_coding[current_macroblock] == MODE_INTER_FOURMV) {
831 s->all_fragments[current_fragment].motion_x = motion_x[k];
832 s->all_fragments[current_fragment].motion_y = motion_y[k];
834 s->all_fragments[current_fragment].motion_x = motion_x[0];
835 s->all_fragments[current_fragment].motion_y = motion_y[0];
838 for (k = 0; k < 2; k++) {
839 current_fragment = s->fragment_start[k+1] +
840 mb_y*(s->fragment_width>>1) + mb_x;
841 if (s->macroblock_coding[current_macroblock] == MODE_INTER_FOURMV) {
842 s->all_fragments[current_fragment].motion_x = motion_x[k+4];
843 s->all_fragments[current_fragment].motion_y = motion_y[k+4];
845 s->all_fragments[current_fragment].motion_x = motion_x[0];
846 s->all_fragments[current_fragment].motion_y = motion_y[0];
856 static int unpack_block_qpis(Vp3DecodeContext *s, GetBitContext *gb)
858 int qpi, i, j, bit, run_length, blocks_decoded, num_blocks_at_qpi;
859 int num_blocks = s->total_num_coded_frags;
861 for (qpi = 0; qpi < s->nqps-1 && num_blocks > 0; qpi++) {
862 i = blocks_decoded = num_blocks_at_qpi = 0;
867 run_length = get_vlc2(gb, s->superblock_run_length_vlc.table, 6, 2) + 1;
868 if (run_length == 34)
869 run_length += get_bits(gb, 12);
870 blocks_decoded += run_length;
873 num_blocks_at_qpi += run_length;
875 for (j = 0; j < run_length; i++) {
876 if (i >= s->total_num_coded_frags)
879 if (s->all_fragments[s->coded_fragment_list[0][i]].qpi == qpi) {
880 s->all_fragments[s->coded_fragment_list[0][i]].qpi += bit;
885 if (run_length == MAXIMUM_LONG_BIT_RUN)
889 } while (blocks_decoded < num_blocks);
891 num_blocks -= num_blocks_at_qpi;
898 * This function is called by unpack_dct_coeffs() to extract the VLCs from
899 * the bitstream. The VLCs encode tokens which are used to unpack DCT
900 * data. This function unpacks all the VLCs for either the Y plane or both
901 * C planes, and is called for DC coefficients or different AC coefficient
902 * levels (since different coefficient types require different VLC tables.
904 * This function returns a residual eob run. E.g, if a particular token gave
905 * instructions to EOB the next 5 fragments and there were only 2 fragments
906 * left in the current fragment range, 3 would be returned so that it could
907 * be passed into the next call to this same function.
909 static int unpack_vlcs(Vp3DecodeContext *s, GetBitContext *gb,
910 VLC *table, int coeff_index,
921 int num_coeffs = s->num_coded_frags[plane][coeff_index];
922 int16_t *dct_tokens = s->dct_tokens[plane][coeff_index];
924 /* local references to structure members to avoid repeated deferences */
925 int *coded_fragment_list = s->coded_fragment_list[plane];
926 Vp3Fragment *all_fragments = s->all_fragments;
927 VLC_TYPE (*vlc_table)[2] = table->table;
930 av_log(s->avctx, AV_LOG_ERROR, "Invalid number of coefficents at level %d\n", coeff_index);
932 if (eob_run > num_coeffs) {
933 coeff_i = blocks_ended = num_coeffs;
934 eob_run -= num_coeffs;
936 coeff_i = blocks_ended = eob_run;
940 // insert fake EOB token to cover the split between planes or zzi
942 dct_tokens[j++] = blocks_ended << 2;
944 while (coeff_i < num_coeffs) {
945 /* decode a VLC into a token */
946 token = get_vlc2(gb, vlc_table, 5, 3);
947 /* use the token to get a zero run, a coefficient, and an eob run */
949 eob_run = eob_run_base[token];
950 if (eob_run_get_bits[token])
951 eob_run += get_bits(gb, eob_run_get_bits[token]);
953 // record only the number of blocks ended in this plane,
954 // any spill will be recorded in the next plane.
955 if (eob_run > num_coeffs - coeff_i) {
956 dct_tokens[j++] = TOKEN_EOB(num_coeffs - coeff_i);
957 blocks_ended += num_coeffs - coeff_i;
958 eob_run -= num_coeffs - coeff_i;
959 coeff_i = num_coeffs;
961 dct_tokens[j++] = TOKEN_EOB(eob_run);
962 blocks_ended += eob_run;
967 bits_to_get = coeff_get_bits[token];
969 bits_to_get = get_bits(gb, bits_to_get);
970 coeff = coeff_tables[token][bits_to_get];
972 zero_run = zero_run_base[token];
973 if (zero_run_get_bits[token])
974 zero_run += get_bits(gb, zero_run_get_bits[token]);
977 dct_tokens[j++] = TOKEN_ZERO_RUN(coeff, zero_run);
979 // Save DC into the fragment structure. DC prediction is
980 // done in raster order, so the actual DC can't be in with
981 // other tokens. We still need the token in dct_tokens[]
982 // however, or else the structure collapses on itself.
984 all_fragments[coded_fragment_list[coeff_i]].dc = coeff;
986 dct_tokens[j++] = TOKEN_COEFF(coeff);
989 if (coeff_index + zero_run > 64) {
990 av_log(s->avctx, AV_LOG_DEBUG, "Invalid zero run of %d with"
991 " %d coeffs left\n", zero_run, 64-coeff_index);
992 zero_run = 64 - coeff_index;
995 // zero runs code multiple coefficients,
996 // so don't try to decode coeffs for those higher levels
997 for (i = coeff_index+1; i <= coeff_index+zero_run; i++)
998 s->num_coded_frags[plane][i]--;
1003 if (blocks_ended > s->num_coded_frags[plane][coeff_index])
1004 av_log(s->avctx, AV_LOG_ERROR, "More blocks ended than coded!\n");
1006 // decrement the number of blocks that have higher coeffecients for each
1007 // EOB run at this level
1009 for (i = coeff_index+1; i < 64; i++)
1010 s->num_coded_frags[plane][i] -= blocks_ended;
1012 // setup the next buffer
1014 s->dct_tokens[plane+1][coeff_index] = dct_tokens + j;
1015 else if (coeff_index < 63)
1016 s->dct_tokens[0][coeff_index+1] = dct_tokens + j;
1021 static void reverse_dc_prediction(Vp3DecodeContext *s,
1024 int fragment_height);
1026 * This function unpacks all of the DCT coefficient data from the
1029 static int unpack_dct_coeffs(Vp3DecodeContext *s, GetBitContext *gb)
1036 int residual_eob_run = 0;
1040 s->dct_tokens[0][0] = s->dct_tokens_base;
1042 /* fetch the DC table indexes */
1043 dc_y_table = get_bits(gb, 4);
1044 dc_c_table = get_bits(gb, 4);
1046 /* unpack the Y plane DC coefficients */
1047 residual_eob_run = unpack_vlcs(s, gb, &s->dc_vlc[dc_y_table], 0,
1048 0, residual_eob_run);
1050 /* reverse prediction of the Y-plane DC coefficients */
1051 reverse_dc_prediction(s, 0, s->fragment_width, s->fragment_height);
1053 /* unpack the C plane DC coefficients */
1054 residual_eob_run = unpack_vlcs(s, gb, &s->dc_vlc[dc_c_table], 0,
1055 1, residual_eob_run);
1056 residual_eob_run = unpack_vlcs(s, gb, &s->dc_vlc[dc_c_table], 0,
1057 2, residual_eob_run);
1059 /* reverse prediction of the C-plane DC coefficients */
1060 if (!(s->avctx->flags & CODEC_FLAG_GRAY))
1062 reverse_dc_prediction(s, s->fragment_start[1],
1063 s->fragment_width / 2, s->fragment_height / 2);
1064 reverse_dc_prediction(s, s->fragment_start[2],
1065 s->fragment_width / 2, s->fragment_height / 2);
1068 /* fetch the AC table indexes */
1069 ac_y_table = get_bits(gb, 4);
1070 ac_c_table = get_bits(gb, 4);
1072 /* build tables of AC VLC tables */
1073 for (i = 1; i <= 5; i++) {
1074 y_tables[i] = &s->ac_vlc_1[ac_y_table];
1075 c_tables[i] = &s->ac_vlc_1[ac_c_table];
1077 for (i = 6; i <= 14; i++) {
1078 y_tables[i] = &s->ac_vlc_2[ac_y_table];
1079 c_tables[i] = &s->ac_vlc_2[ac_c_table];
1081 for (i = 15; i <= 27; i++) {
1082 y_tables[i] = &s->ac_vlc_3[ac_y_table];
1083 c_tables[i] = &s->ac_vlc_3[ac_c_table];
1085 for (i = 28; i <= 63; i++) {
1086 y_tables[i] = &s->ac_vlc_4[ac_y_table];
1087 c_tables[i] = &s->ac_vlc_4[ac_c_table];
1090 /* decode all AC coefficents */
1091 for (i = 1; i <= 63; i++) {
1092 residual_eob_run = unpack_vlcs(s, gb, y_tables[i], i,
1093 0, residual_eob_run);
1095 residual_eob_run = unpack_vlcs(s, gb, c_tables[i], i,
1096 1, residual_eob_run);
1097 residual_eob_run = unpack_vlcs(s, gb, c_tables[i], i,
1098 2, residual_eob_run);
1105 * This function reverses the DC prediction for each coded fragment in
1106 * the frame. Much of this function is adapted directly from the original
1109 #define COMPATIBLE_FRAME(x) \
1110 (compatible_frame[s->all_fragments[x].coding_method] == current_frame_type)
1111 #define DC_COEFF(u) s->all_fragments[u].dc
1113 static void reverse_dc_prediction(Vp3DecodeContext *s,
1116 int fragment_height)
1125 int i = first_fragment;
1129 /* DC values for the left, up-left, up, and up-right fragments */
1130 int vl, vul, vu, vur;
1132 /* indexes for the left, up-left, up, and up-right fragments */
1136 * The 6 fields mean:
1137 * 0: up-left multiplier
1139 * 2: up-right multiplier
1140 * 3: left multiplier
1142 static const int predictor_transform[16][4] = {
1144 { 0, 0, 0,128}, // PL
1145 { 0, 0,128, 0}, // PUR
1146 { 0, 0, 53, 75}, // PUR|PL
1147 { 0,128, 0, 0}, // PU
1148 { 0, 64, 0, 64}, // PU|PL
1149 { 0,128, 0, 0}, // PU|PUR
1150 { 0, 0, 53, 75}, // PU|PUR|PL
1151 {128, 0, 0, 0}, // PUL
1152 { 0, 0, 0,128}, // PUL|PL
1153 { 64, 0, 64, 0}, // PUL|PUR
1154 { 0, 0, 53, 75}, // PUL|PUR|PL
1155 { 0,128, 0, 0}, // PUL|PU
1156 {-104,116, 0,116}, // PUL|PU|PL
1157 { 24, 80, 24, 0}, // PUL|PU|PUR
1158 {-104,116, 0,116} // PUL|PU|PUR|PL
1161 /* This table shows which types of blocks can use other blocks for
1162 * prediction. For example, INTRA is the only mode in this table to
1163 * have a frame number of 0. That means INTRA blocks can only predict
1164 * from other INTRA blocks. There are 2 golden frame coding types;
1165 * blocks encoding in these modes can only predict from other blocks
1166 * that were encoded with these 1 of these 2 modes. */
1167 static const unsigned char compatible_frame[9] = {
1168 1, /* MODE_INTER_NO_MV */
1170 1, /* MODE_INTER_PLUS_MV */
1171 1, /* MODE_INTER_LAST_MV */
1172 1, /* MODE_INTER_PRIOR_MV */
1173 2, /* MODE_USING_GOLDEN */
1174 2, /* MODE_GOLDEN_MV */
1175 1, /* MODE_INTER_FOUR_MV */
1178 int current_frame_type;
1180 /* there is a last DC predictor for each of the 3 frame types */
1185 vul = vu = vur = vl = 0;
1186 last_dc[0] = last_dc[1] = last_dc[2] = 0;
1188 /* for each fragment row... */
1189 for (y = 0; y < fragment_height; y++) {
1191 /* for each fragment in a row... */
1192 for (x = 0; x < fragment_width; x++, i++) {
1194 /* reverse prediction if this block was coded */
1195 if (s->all_fragments[i].coding_method != MODE_COPY) {
1197 current_frame_type =
1198 compatible_frame[s->all_fragments[i].coding_method];
1204 if(COMPATIBLE_FRAME(l))
1208 u= i-fragment_width;
1210 if(COMPATIBLE_FRAME(u))
1213 ul= i-fragment_width-1;
1215 if(COMPATIBLE_FRAME(ul))
1218 if(x + 1 < fragment_width){
1219 ur= i-fragment_width+1;
1221 if(COMPATIBLE_FRAME(ur))
1226 if (transform == 0) {
1228 /* if there were no fragments to predict from, use last
1230 predicted_dc = last_dc[current_frame_type];
1233 /* apply the appropriate predictor transform */
1235 (predictor_transform[transform][0] * vul) +
1236 (predictor_transform[transform][1] * vu) +
1237 (predictor_transform[transform][2] * vur) +
1238 (predictor_transform[transform][3] * vl);
1240 predicted_dc /= 128;
1242 /* check for outranging on the [ul u l] and
1243 * [ul u ur l] predictors */
1244 if ((transform == 15) || (transform == 13)) {
1245 if (FFABS(predicted_dc - vu) > 128)
1247 else if (FFABS(predicted_dc - vl) > 128)
1249 else if (FFABS(predicted_dc - vul) > 128)
1254 /* at long last, apply the predictor */
1255 DC_COEFF(i) += predicted_dc;
1257 last_dc[current_frame_type] = DC_COEFF(i);
1263 static void apply_loop_filter(Vp3DecodeContext *s, int plane, int ystart, int yend)
1266 int *bounding_values= s->bounding_values_array+127;
1268 int width = s->fragment_width >> !!plane;
1269 int height = s->fragment_height >> !!plane;
1270 int fragment = s->fragment_start [plane] + ystart * width;
1271 int stride = s->current_frame.linesize[plane];
1272 uint8_t *plane_data = s->current_frame.data [plane];
1273 if (!s->flipped_image) stride = -stride;
1274 plane_data += s->data_offset[plane] + 8*ystart*stride;
1276 for (y = ystart; y < yend; y++) {
1278 for (x = 0; x < width; x++) {
1279 /* This code basically just deblocks on the edges of coded blocks.
1280 * However, it has to be much more complicated because of the
1281 * braindamaged deblock ordering used in VP3/Theora. Order matters
1282 * because some pixels get filtered twice. */
1283 if( s->all_fragments[fragment].coding_method != MODE_COPY )
1285 /* do not perform left edge filter for left columns frags */
1287 s->dsp.vp3_h_loop_filter(
1289 stride, bounding_values);
1292 /* do not perform top edge filter for top row fragments */
1294 s->dsp.vp3_v_loop_filter(
1296 stride, bounding_values);
1299 /* do not perform right edge filter for right column
1300 * fragments or if right fragment neighbor is also coded
1301 * in this frame (it will be filtered in next iteration) */
1302 if ((x < width - 1) &&
1303 (s->all_fragments[fragment + 1].coding_method == MODE_COPY)) {
1304 s->dsp.vp3_h_loop_filter(
1305 plane_data + 8*x + 8,
1306 stride, bounding_values);
1309 /* do not perform bottom edge filter for bottom row
1310 * fragments or if bottom fragment neighbor is also coded
1311 * in this frame (it will be filtered in the next row) */
1312 if ((y < height - 1) &&
1313 (s->all_fragments[fragment + width].coding_method == MODE_COPY)) {
1314 s->dsp.vp3_v_loop_filter(
1315 plane_data + 8*x + 8*stride,
1316 stride, bounding_values);
1322 plane_data += 8*stride;
1327 * Pulls DCT tokens from the 64 levels to decode and dequant the coefficients
1328 * for the next block in coding order
1330 static inline int vp3_dequant(Vp3DecodeContext *s, Vp3Fragment *frag,
1331 int plane, int inter, DCTELEM block[64])
1333 int16_t *dequantizer = s->qmat[frag->qpi][inter][plane];
1334 uint8_t *perm = s->scantable.permutated;
1338 int token = *s->dct_tokens[plane][i];
1339 switch (token & 3) {
1341 if (--token < 4) // 0-3 are token types, so the EOB run must now be 0
1342 s->dct_tokens[plane][i]++;
1344 *s->dct_tokens[plane][i] = token & ~3;
1347 s->dct_tokens[plane][i]++;
1348 i += (token >> 2) & 0x7f;
1349 block[perm[i]] = (token >> 9) * dequantizer[perm[i]];
1353 block[perm[i]] = (token >> 2) * dequantizer[perm[i]];
1354 s->dct_tokens[plane][i++]++;
1357 av_log(s->avctx, AV_LOG_ERROR, "internal: invalid token type\n");
1362 // the actual DC+prediction is in the fragment structure
1363 block[0] = frag->dc * s->qmat[0][inter][plane][0];
1368 * called when all pixels up to row y are complete
1370 static void vp3_draw_horiz_band(Vp3DecodeContext *s, int y)
1375 if(s->avctx->draw_horiz_band==NULL)
1378 h= y - s->last_slice_end;
1381 if (!s->flipped_image) {
1383 h -= s->height - s->avctx->height; // account for non-mod16
1384 y = s->height - y - h;
1388 offset[0] = s->current_frame.linesize[0]*y;
1389 offset[1] = s->current_frame.linesize[1]*cy;
1390 offset[2] = s->current_frame.linesize[2]*cy;
1394 s->avctx->draw_horiz_band(s->avctx, &s->current_frame, offset, y, 3, h);
1395 s->last_slice_end= y + h;
1399 * Perform the final rendering for a particular slice of data.
1400 * The slice number ranges from 0..(c_superblock_height - 1).
1402 static void render_slice(Vp3DecodeContext *s, int slice)
1405 LOCAL_ALIGNED_16(DCTELEM, block, [64]);
1406 int motion_x = 0xdeadbeef, motion_y = 0xdeadbeef;
1407 int motion_halfpel_index;
1408 uint8_t *motion_source;
1409 int plane, first_pixel;
1411 if (slice >= s->c_superblock_height)
1414 for (plane = 0; plane < 3; plane++) {
1415 uint8_t *output_plane = s->current_frame.data [plane] + s->data_offset[plane];
1416 uint8_t * last_plane = s-> last_frame.data [plane] + s->data_offset[plane];
1417 uint8_t *golden_plane = s-> golden_frame.data [plane] + s->data_offset[plane];
1418 int stride = s->current_frame.linesize[plane];
1419 int plane_width = s->width >> !!plane;
1420 int plane_height = s->height >> !!plane;
1422 int sb_x, sb_y = slice << !plane;
1423 int slice_height = sb_y + (plane ? 1 : 2);
1424 int slice_width = plane ? s->c_superblock_width : s->y_superblock_width;
1426 int fragment_width = s->fragment_width >> !!plane;
1427 int fragment_height = s->fragment_height >> !!plane;
1428 int fragment_start = s->fragment_start[plane];
1430 if (!s->flipped_image) stride = -stride;
1431 if (CONFIG_GRAY && plane && (s->avctx->flags & CODEC_FLAG_GRAY))
1435 if(FFABS(stride) > 2048)
1436 return; //various tables are fixed size
1438 /* for each superblock row in the slice (both of them)... */
1439 for (; sb_y < slice_height; sb_y++) {
1441 /* for each superblock in a row... */
1442 for (sb_x = 0; sb_x < slice_width; sb_x++) {
1444 /* for each block in a superblock... */
1445 for (j = 0; j < 16; j++) {
1446 x = 4*sb_x + hilbert_offset[j][0];
1447 y = 4*sb_y + hilbert_offset[j][1];
1449 i = fragment_start + y*fragment_width + x;
1452 if (x >= fragment_width || y >= fragment_height)
1455 first_pixel = 8*y*stride + 8*x;
1457 /* transform if this block was coded */
1458 if (s->all_fragments[i].coding_method != MODE_COPY) {
1459 int intra = s->all_fragments[i].coding_method == MODE_INTRA;
1461 if ((s->all_fragments[i].coding_method == MODE_USING_GOLDEN) ||
1462 (s->all_fragments[i].coding_method == MODE_GOLDEN_MV))
1463 motion_source= golden_plane;
1465 motion_source= last_plane;
1467 motion_source += first_pixel;
1468 motion_halfpel_index = 0;
1470 /* sort out the motion vector if this fragment is coded
1471 * using a motion vector method */
1472 if ((s->all_fragments[i].coding_method > MODE_INTRA) &&
1473 (s->all_fragments[i].coding_method != MODE_USING_GOLDEN)) {
1475 motion_x = s->all_fragments[i].motion_x;
1476 motion_y = s->all_fragments[i].motion_y;
1478 motion_x= (motion_x>>1) | (motion_x&1);
1479 motion_y= (motion_y>>1) | (motion_y&1);
1482 src_x= (motion_x>>1) + 8*x;
1483 src_y= (motion_y>>1) + 8*y;
1484 if ((motion_x == 127) || (motion_y == 127))
1485 av_log(s->avctx, AV_LOG_ERROR, " help! got invalid motion vector! (%X, %X)\n", motion_x, motion_y);
1487 motion_halfpel_index = motion_x & 0x01;
1488 motion_source += (motion_x >> 1);
1490 motion_halfpel_index |= (motion_y & 0x01) << 1;
1491 motion_source += ((motion_y >> 1) * stride);
1493 if(src_x<0 || src_y<0 || src_x + 9 >= plane_width || src_y + 9 >= plane_height){
1494 uint8_t *temp= s->edge_emu_buffer;
1495 if(stride<0) temp -= 9*stride;
1496 else temp += 9*stride;
1498 ff_emulated_edge_mc(temp, motion_source, stride, 9, 9, src_x, src_y, plane_width, plane_height);
1499 motion_source= temp;
1504 /* first, take care of copying a block from either the
1505 * previous or the golden frame */
1506 if (s->all_fragments[i].coding_method != MODE_INTRA) {
1507 /* Note, it is possible to implement all MC cases with
1508 put_no_rnd_pixels_l2 which would look more like the
1509 VP3 source but this would be slower as
1510 put_no_rnd_pixels_tab is better optimzed */
1511 if(motion_halfpel_index != 3){
1512 s->dsp.put_no_rnd_pixels_tab[1][motion_halfpel_index](
1513 output_plane + first_pixel,
1514 motion_source, stride, 8);
1516 int d= (motion_x ^ motion_y)>>31; // d is 0 if motion_x and _y have the same sign, else -1
1517 s->dsp.put_no_rnd_pixels_l2[1](
1518 output_plane + first_pixel,
1520 motion_source + stride + 1 + d,
1525 s->dsp.clear_block(block);
1526 vp3_dequant(s, s->all_fragments + i, plane, !intra, block);
1528 /* invert DCT and place (or add) in final output */
1530 if (s->all_fragments[i].coding_method == MODE_INTRA) {
1531 if(s->avctx->idct_algo!=FF_IDCT_VP3)
1534 output_plane + first_pixel,
1539 output_plane + first_pixel,
1545 /* copy directly from the previous frame */
1546 s->dsp.put_pixels_tab[1][0](
1547 output_plane + first_pixel,
1548 last_plane + first_pixel,
1555 // Filter up to the last row in the superblock row
1556 apply_loop_filter(s, plane, 4*sb_y - !!sb_y, FFMIN(4*sb_y+3, fragment_height-1));
1560 /* this looks like a good place for slice dispatch... */
1562 * if (slice == s->macroblock_height - 1)
1563 * dispatch (both last slice & 2nd-to-last slice);
1564 * else if (slice > 0)
1565 * dispatch (slice - 1);
1568 vp3_draw_horiz_band(s, 64*slice + 64-16);
1572 * This is the ffmpeg/libavcodec API init function.
1574 static av_cold int vp3_decode_init(AVCodecContext *avctx)
1576 Vp3DecodeContext *s = avctx->priv_data;
1577 int i, inter, plane;
1581 if (avctx->codec_tag == MKTAG('V','P','3','0'))
1587 s->width = FFALIGN(avctx->width, 16);
1588 s->height = FFALIGN(avctx->height, 16);
1589 avctx->pix_fmt = PIX_FMT_YUV420P;
1590 avctx->chroma_sample_location = AVCHROMA_LOC_CENTER;
1591 if(avctx->idct_algo==FF_IDCT_AUTO)
1592 avctx->idct_algo=FF_IDCT_VP3;
1593 dsputil_init(&s->dsp, avctx);
1595 ff_init_scantable(s->dsp.idct_permutation, &s->scantable, ff_zigzag_direct);
1597 /* initialize to an impossible value which will force a recalculation
1598 * in the first frame decode */
1599 for (i = 0; i < 3; i++)
1602 s->y_superblock_width = (s->width + 31) / 32;
1603 s->y_superblock_height = (s->height + 31) / 32;
1604 s->y_superblock_count = s->y_superblock_width * s->y_superblock_height;
1606 /* work out the dimensions for the C planes */
1607 c_width = s->width / 2;
1608 c_height = s->height / 2;
1609 s->c_superblock_width = (c_width + 31) / 32;
1610 s->c_superblock_height = (c_height + 31) / 32;
1611 s->c_superblock_count = s->c_superblock_width * s->c_superblock_height;
1613 s->superblock_count = s->y_superblock_count + (s->c_superblock_count * 2);
1614 s->u_superblock_start = s->y_superblock_count;
1615 s->v_superblock_start = s->u_superblock_start + s->c_superblock_count;
1616 s->superblock_coding = av_malloc(s->superblock_count);
1618 s->macroblock_width = (s->width + 15) / 16;
1619 s->macroblock_height = (s->height + 15) / 16;
1620 s->macroblock_count = s->macroblock_width * s->macroblock_height;
1622 s->fragment_width = s->width / FRAGMENT_PIXELS;
1623 s->fragment_height = s->height / FRAGMENT_PIXELS;
1625 /* fragment count covers all 8x8 blocks for all 3 planes */
1626 s->fragment_count = s->fragment_width * s->fragment_height * 3 / 2;
1627 s->fragment_start[1] = s->fragment_width * s->fragment_height;
1628 s->fragment_start[2] = s->fragment_width * s->fragment_height * 5 / 4;
1630 s->all_fragments = av_malloc(s->fragment_count * sizeof(Vp3Fragment));
1631 s->coded_fragment_list[0] = av_malloc(s->fragment_count * sizeof(int));
1632 s->dct_tokens_base = av_malloc(64*s->fragment_count * sizeof(*s->dct_tokens_base));
1633 if (!s->superblock_coding || !s->all_fragments || !s->dct_tokens_base ||
1634 !s->coded_fragment_list[0]) {
1635 vp3_decode_end(avctx);
1639 if (!s->theora_tables)
1641 for (i = 0; i < 64; i++) {
1642 s->coded_dc_scale_factor[i] = vp31_dc_scale_factor[i];
1643 s->coded_ac_scale_factor[i] = vp31_ac_scale_factor[i];
1644 s->base_matrix[0][i] = vp31_intra_y_dequant[i];
1645 s->base_matrix[1][i] = vp31_intra_c_dequant[i];
1646 s->base_matrix[2][i] = vp31_inter_dequant[i];
1647 s->filter_limit_values[i] = vp31_filter_limit_values[i];
1650 for(inter=0; inter<2; inter++){
1651 for(plane=0; plane<3; plane++){
1652 s->qr_count[inter][plane]= 1;
1653 s->qr_size [inter][plane][0]= 63;
1654 s->qr_base [inter][plane][0]=
1655 s->qr_base [inter][plane][1]= 2*inter + (!!plane)*!inter;
1659 /* init VLC tables */
1660 for (i = 0; i < 16; i++) {
1663 init_vlc(&s->dc_vlc[i], 5, 32,
1664 &dc_bias[i][0][1], 4, 2,
1665 &dc_bias[i][0][0], 4, 2, 0);
1667 /* group 1 AC histograms */
1668 init_vlc(&s->ac_vlc_1[i], 5, 32,
1669 &ac_bias_0[i][0][1], 4, 2,
1670 &ac_bias_0[i][0][0], 4, 2, 0);
1672 /* group 2 AC histograms */
1673 init_vlc(&s->ac_vlc_2[i], 5, 32,
1674 &ac_bias_1[i][0][1], 4, 2,
1675 &ac_bias_1[i][0][0], 4, 2, 0);
1677 /* group 3 AC histograms */
1678 init_vlc(&s->ac_vlc_3[i], 5, 32,
1679 &ac_bias_2[i][0][1], 4, 2,
1680 &ac_bias_2[i][0][0], 4, 2, 0);
1682 /* group 4 AC histograms */
1683 init_vlc(&s->ac_vlc_4[i], 5, 32,
1684 &ac_bias_3[i][0][1], 4, 2,
1685 &ac_bias_3[i][0][0], 4, 2, 0);
1688 for (i = 0; i < 16; i++) {
1691 if (init_vlc(&s->dc_vlc[i], 5, 32,
1692 &s->huffman_table[i][0][1], 4, 2,
1693 &s->huffman_table[i][0][0], 4, 2, 0) < 0)
1696 /* group 1 AC histograms */
1697 if (init_vlc(&s->ac_vlc_1[i], 5, 32,
1698 &s->huffman_table[i+16][0][1], 4, 2,
1699 &s->huffman_table[i+16][0][0], 4, 2, 0) < 0)
1702 /* group 2 AC histograms */
1703 if (init_vlc(&s->ac_vlc_2[i], 5, 32,
1704 &s->huffman_table[i+16*2][0][1], 4, 2,
1705 &s->huffman_table[i+16*2][0][0], 4, 2, 0) < 0)
1708 /* group 3 AC histograms */
1709 if (init_vlc(&s->ac_vlc_3[i], 5, 32,
1710 &s->huffman_table[i+16*3][0][1], 4, 2,
1711 &s->huffman_table[i+16*3][0][0], 4, 2, 0) < 0)
1714 /* group 4 AC histograms */
1715 if (init_vlc(&s->ac_vlc_4[i], 5, 32,
1716 &s->huffman_table[i+16*4][0][1], 4, 2,
1717 &s->huffman_table[i+16*4][0][0], 4, 2, 0) < 0)
1722 init_vlc(&s->superblock_run_length_vlc, 6, 34,
1723 &superblock_run_length_vlc_table[0][1], 4, 2,
1724 &superblock_run_length_vlc_table[0][0], 4, 2, 0);
1726 init_vlc(&s->fragment_run_length_vlc, 5, 30,
1727 &fragment_run_length_vlc_table[0][1], 4, 2,
1728 &fragment_run_length_vlc_table[0][0], 4, 2, 0);
1730 init_vlc(&s->mode_code_vlc, 3, 8,
1731 &mode_code_vlc_table[0][1], 2, 1,
1732 &mode_code_vlc_table[0][0], 2, 1, 0);
1734 init_vlc(&s->motion_vector_vlc, 6, 63,
1735 &motion_vector_vlc_table[0][1], 2, 1,
1736 &motion_vector_vlc_table[0][0], 2, 1, 0);
1738 /* work out the block mapping tables */
1739 s->superblock_fragments = av_malloc(s->superblock_count * 16 * sizeof(int));
1740 s->macroblock_coding = av_malloc(s->macroblock_count + 1);
1741 if (!s->superblock_fragments || !s->macroblock_coding) {
1742 vp3_decode_end(avctx);
1745 init_block_mapping(s);
1747 for (i = 0; i < 3; i++) {
1748 s->current_frame.data[i] = NULL;
1749 s->last_frame.data[i] = NULL;
1750 s->golden_frame.data[i] = NULL;
1756 av_log(avctx, AV_LOG_FATAL, "Invalid huffman table\n");
1761 * This is the ffmpeg/libavcodec API frame decode function.
1763 static int vp3_decode_frame(AVCodecContext *avctx,
1764 void *data, int *data_size,
1767 const uint8_t *buf = avpkt->data;
1768 int buf_size = avpkt->size;
1769 Vp3DecodeContext *s = avctx->priv_data;
1771 static int counter = 0;
1774 init_get_bits(&gb, buf, buf_size * 8);
1776 if (s->theora && get_bits1(&gb))
1778 av_log(avctx, AV_LOG_ERROR, "Header packet passed to frame decoder, skipping\n");
1782 s->keyframe = !get_bits1(&gb);
1785 for (i = 0; i < 3; i++)
1786 s->last_qps[i] = s->qps[i];
1790 s->qps[s->nqps++]= get_bits(&gb, 6);
1791 } while(s->theora >= 0x030200 && s->nqps<3 && get_bits1(&gb));
1792 for (i = s->nqps; i < 3; i++)
1795 if (s->avctx->debug & FF_DEBUG_PICT_INFO)
1796 av_log(s->avctx, AV_LOG_INFO, " VP3 %sframe #%d: Q index = %d\n",
1797 s->keyframe?"key":"", counter, s->qps[0]);
1800 if (s->qps[0] != s->last_qps[0])
1801 init_loop_filter(s);
1803 for (i = 0; i < s->nqps; i++)
1804 // reinit all dequantizers if the first one changed, because
1805 // the DC of the first quantizer must be used for all matrices
1806 if (s->qps[i] != s->last_qps[i] || s->qps[0] != s->last_qps[0])
1807 init_dequantizer(s, i);
1809 if (avctx->skip_frame >= AVDISCARD_NONKEY && !s->keyframe)
1812 s->current_frame.reference = 3;
1813 s->current_frame.pict_type = s->keyframe ? FF_I_TYPE : FF_P_TYPE;
1814 if (avctx->get_buffer(avctx, &s->current_frame) < 0) {
1815 av_log(s->avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1822 skip_bits(&gb, 4); /* width code */
1823 skip_bits(&gb, 4); /* height code */
1826 s->version = get_bits(&gb, 5);
1828 av_log(s->avctx, AV_LOG_DEBUG, "VP version: %d\n", s->version);
1831 if (s->version || s->theora)
1834 av_log(s->avctx, AV_LOG_ERROR, "Warning, unsupported keyframe coding type?!\n");
1835 skip_bits(&gb, 2); /* reserved? */
1838 if (!s->golden_frame.data[0]) {
1839 av_log(s->avctx, AV_LOG_WARNING, "vp3: first frame not a keyframe\n");
1841 s->golden_frame.reference = 3;
1842 s->golden_frame.pict_type = FF_I_TYPE;
1843 if (avctx->get_buffer(avctx, &s->golden_frame) < 0) {
1844 av_log(s->avctx, AV_LOG_ERROR, "get_buffer() failed\n");
1847 s->last_frame = s->golden_frame;
1848 s->last_frame.type = FF_BUFFER_TYPE_COPY;
1852 s->current_frame.qscale_table= s->qscale_table; //FIXME allocate individual tables per AVFrame
1853 s->current_frame.qstride= 0;
1857 if (unpack_superblocks(s, &gb)){
1858 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_superblocks\n");
1861 if (unpack_modes(s, &gb)){
1862 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_modes\n");
1865 if (unpack_vectors(s, &gb)){
1866 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_vectors\n");
1869 if (unpack_block_qpis(s, &gb)){
1870 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_block_qpis\n");
1873 if (unpack_dct_coeffs(s, &gb)){
1874 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_dct_coeffs\n");
1878 for (i = 0; i < 3; i++) {
1879 if (s->flipped_image)
1880 s->data_offset[i] = 0;
1882 s->data_offset[i] = ((s->height>>!!i)-1) * s->current_frame.linesize[i];
1885 s->last_slice_end = 0;
1886 for (i = 0; i < s->c_superblock_height; i++)
1889 // filter the last row
1890 for (i = 0; i < 3; i++) {
1891 int row = (s->height >> (3+!!i)) - 1;
1892 apply_loop_filter(s, i, row, row+1);
1894 vp3_draw_horiz_band(s, s->height);
1896 *data_size=sizeof(AVFrame);
1897 *(AVFrame*)data= s->current_frame;
1899 /* release the last frame, if it is allocated and if it is not the
1901 if (s->last_frame.data[0] && s->last_frame.type != FF_BUFFER_TYPE_COPY)
1902 avctx->release_buffer(avctx, &s->last_frame);
1904 /* shuffle frames (last = current) */
1905 s->last_frame= s->current_frame;
1908 if (s->golden_frame.data[0])
1909 avctx->release_buffer(avctx, &s->golden_frame);
1910 s->golden_frame = s->current_frame;
1911 s->last_frame.type = FF_BUFFER_TYPE_COPY;
1914 s->current_frame.data[0]= NULL; /* ensure that we catch any access to this released frame */
1919 if (s->current_frame.data[0])
1920 avctx->release_buffer(avctx, &s->current_frame);
1925 * This is the ffmpeg/libavcodec API module cleanup function.
1927 static av_cold int vp3_decode_end(AVCodecContext *avctx)
1929 Vp3DecodeContext *s = avctx->priv_data;
1932 av_free(s->superblock_coding);
1933 av_free(s->all_fragments);
1934 av_free(s->coded_fragment_list[0]);
1935 av_free(s->dct_tokens_base);
1936 av_free(s->superblock_fragments);
1937 av_free(s->macroblock_coding);
1939 for (i = 0; i < 16; i++) {
1940 free_vlc(&s->dc_vlc[i]);
1941 free_vlc(&s->ac_vlc_1[i]);
1942 free_vlc(&s->ac_vlc_2[i]);
1943 free_vlc(&s->ac_vlc_3[i]);
1944 free_vlc(&s->ac_vlc_4[i]);
1947 free_vlc(&s->superblock_run_length_vlc);
1948 free_vlc(&s->fragment_run_length_vlc);
1949 free_vlc(&s->mode_code_vlc);
1950 free_vlc(&s->motion_vector_vlc);
1952 /* release all frames */
1953 if (s->golden_frame.data[0])
1954 avctx->release_buffer(avctx, &s->golden_frame);
1955 if (s->last_frame.data[0] && s->last_frame.type != FF_BUFFER_TYPE_COPY)
1956 avctx->release_buffer(avctx, &s->last_frame);
1957 /* no need to release the current_frame since it will always be pointing
1958 * to the same frame as either the golden or last frame */
1963 static int read_huffman_tree(AVCodecContext *avctx, GetBitContext *gb)
1965 Vp3DecodeContext *s = avctx->priv_data;
1967 if (get_bits1(gb)) {
1969 if (s->entries >= 32) { /* overflow */
1970 av_log(avctx, AV_LOG_ERROR, "huffman tree overflow\n");
1973 token = get_bits(gb, 5);
1974 //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);
1975 s->huffman_table[s->hti][token][0] = s->hbits;
1976 s->huffman_table[s->hti][token][1] = s->huff_code_size;
1980 if (s->huff_code_size >= 32) {/* overflow */
1981 av_log(avctx, AV_LOG_ERROR, "huffman tree overflow\n");
1984 s->huff_code_size++;
1986 if (read_huffman_tree(avctx, gb))
1989 if (read_huffman_tree(avctx, gb))
1992 s->huff_code_size--;
1997 #if CONFIG_THEORA_DECODER
1998 static int theora_decode_header(AVCodecContext *avctx, GetBitContext *gb)
2000 Vp3DecodeContext *s = avctx->priv_data;
2001 int visible_width, visible_height, colorspace;
2003 s->theora = get_bits_long(gb, 24);
2004 av_log(avctx, AV_LOG_DEBUG, "Theora bitstream version %X\n", s->theora);
2006 /* 3.2.0 aka alpha3 has the same frame orientation as original vp3 */
2007 /* but previous versions have the image flipped relative to vp3 */
2008 if (s->theora < 0x030200)
2010 s->flipped_image = 1;
2011 av_log(avctx, AV_LOG_DEBUG, "Old (<alpha3) Theora bitstream, flipped image\n");
2014 visible_width = s->width = get_bits(gb, 16) << 4;
2015 visible_height = s->height = get_bits(gb, 16) << 4;
2017 if(avcodec_check_dimensions(avctx, s->width, s->height)){
2018 av_log(avctx, AV_LOG_ERROR, "Invalid dimensions (%dx%d)\n", s->width, s->height);
2019 s->width= s->height= 0;
2023 if (s->theora >= 0x030200) {
2024 visible_width = get_bits_long(gb, 24);
2025 visible_height = get_bits_long(gb, 24);
2027 skip_bits(gb, 8); /* offset x */
2028 skip_bits(gb, 8); /* offset y */
2031 skip_bits(gb, 32); /* fps numerator */
2032 skip_bits(gb, 32); /* fps denumerator */
2033 skip_bits(gb, 24); /* aspect numerator */
2034 skip_bits(gb, 24); /* aspect denumerator */
2036 if (s->theora < 0x030200)
2037 skip_bits(gb, 5); /* keyframe frequency force */
2038 colorspace = get_bits(gb, 8);
2039 skip_bits(gb, 24); /* bitrate */
2041 skip_bits(gb, 6); /* quality hint */
2043 if (s->theora >= 0x030200)
2045 skip_bits(gb, 5); /* keyframe frequency force */
2046 skip_bits(gb, 2); /* pixel format: 420,res,422,444 */
2047 skip_bits(gb, 3); /* reserved */
2050 // align_get_bits(gb);
2052 if ( visible_width <= s->width && visible_width > s->width-16
2053 && visible_height <= s->height && visible_height > s->height-16)
2054 avcodec_set_dimensions(avctx, visible_width, visible_height);
2056 avcodec_set_dimensions(avctx, s->width, s->height);
2058 if (colorspace == 1) {
2059 avctx->color_primaries = AVCOL_PRI_BT470M;
2060 } else if (colorspace == 2) {
2061 avctx->color_primaries = AVCOL_PRI_BT470BG;
2063 if (colorspace == 1 || colorspace == 2) {
2064 avctx->colorspace = AVCOL_SPC_BT470BG;
2065 avctx->color_trc = AVCOL_TRC_BT709;
2071 static int theora_decode_tables(AVCodecContext *avctx, GetBitContext *gb)
2073 Vp3DecodeContext *s = avctx->priv_data;
2074 int i, n, matrices, inter, plane;
2076 if (s->theora >= 0x030200) {
2077 n = get_bits(gb, 3);
2078 /* loop filter limit values table */
2079 for (i = 0; i < 64; i++) {
2080 s->filter_limit_values[i] = get_bits(gb, n);
2081 if (s->filter_limit_values[i] > 127) {
2082 av_log(avctx, AV_LOG_ERROR, "filter limit value too large (%i > 127), clamping\n", s->filter_limit_values[i]);
2083 s->filter_limit_values[i] = 127;
2088 if (s->theora >= 0x030200)
2089 n = get_bits(gb, 4) + 1;
2092 /* quality threshold table */
2093 for (i = 0; i < 64; i++)
2094 s->coded_ac_scale_factor[i] = get_bits(gb, n);
2096 if (s->theora >= 0x030200)
2097 n = get_bits(gb, 4) + 1;
2100 /* dc scale factor table */
2101 for (i = 0; i < 64; i++)
2102 s->coded_dc_scale_factor[i] = get_bits(gb, n);
2104 if (s->theora >= 0x030200)
2105 matrices = get_bits(gb, 9) + 1;
2110 av_log(avctx, AV_LOG_ERROR, "invalid number of base matrixes\n");
2114 for(n=0; n<matrices; n++){
2115 for (i = 0; i < 64; i++)
2116 s->base_matrix[n][i]= get_bits(gb, 8);
2119 for (inter = 0; inter <= 1; inter++) {
2120 for (plane = 0; plane <= 2; plane++) {
2122 if (inter || plane > 0)
2123 newqr = get_bits1(gb);
2126 if(inter && get_bits1(gb)){
2130 qtj= (3*inter + plane - 1) / 3;
2131 plj= (plane + 2) % 3;
2133 s->qr_count[inter][plane]= s->qr_count[qtj][plj];
2134 memcpy(s->qr_size[inter][plane], s->qr_size[qtj][plj], sizeof(s->qr_size[0][0]));
2135 memcpy(s->qr_base[inter][plane], s->qr_base[qtj][plj], sizeof(s->qr_base[0][0]));
2141 i= get_bits(gb, av_log2(matrices-1)+1);
2143 av_log(avctx, AV_LOG_ERROR, "invalid base matrix index\n");
2146 s->qr_base[inter][plane][qri]= i;
2149 i = get_bits(gb, av_log2(63-qi)+1) + 1;
2150 s->qr_size[inter][plane][qri++]= i;
2155 av_log(avctx, AV_LOG_ERROR, "invalid qi %d > 63\n", qi);
2158 s->qr_count[inter][plane]= qri;
2163 /* Huffman tables */
2164 for (s->hti = 0; s->hti < 80; s->hti++) {
2166 s->huff_code_size = 1;
2167 if (!get_bits1(gb)) {
2169 if(read_huffman_tree(avctx, gb))
2172 if(read_huffman_tree(avctx, gb))
2177 s->theora_tables = 1;
2182 static av_cold int theora_decode_init(AVCodecContext *avctx)
2184 Vp3DecodeContext *s = avctx->priv_data;
2187 uint8_t *header_start[3];
2193 if (!avctx->extradata_size)
2195 av_log(avctx, AV_LOG_ERROR, "Missing extradata!\n");
2199 if (ff_split_xiph_headers(avctx->extradata, avctx->extradata_size,
2200 42, header_start, header_len) < 0) {
2201 av_log(avctx, AV_LOG_ERROR, "Corrupt extradata\n");
2206 init_get_bits(&gb, header_start[i], header_len[i] * 8);
2208 ptype = get_bits(&gb, 8);
2210 if (!(ptype & 0x80))
2212 av_log(avctx, AV_LOG_ERROR, "Invalid extradata!\n");
2216 // FIXME: Check for this as well.
2217 skip_bits_long(&gb, 6*8); /* "theora" */
2222 theora_decode_header(avctx, &gb);
2225 // FIXME: is this needed? it breaks sometimes
2226 // theora_decode_comments(avctx, gb);
2229 if (theora_decode_tables(avctx, &gb))
2233 av_log(avctx, AV_LOG_ERROR, "Unknown Theora config packet: %d\n", ptype&~0x80);
2236 if(ptype != 0x81 && 8*header_len[i] != get_bits_count(&gb))
2237 av_log(avctx, AV_LOG_WARNING, "%d bits left in packet %X\n", 8*header_len[i] - get_bits_count(&gb), ptype);
2238 if (s->theora < 0x030200)
2242 return vp3_decode_init(avctx);
2245 AVCodec theora_decoder = {
2249 sizeof(Vp3DecodeContext),
2254 CODEC_CAP_DR1 | CODEC_CAP_DRAW_HORIZ_BAND,
2256 .long_name = NULL_IF_CONFIG_SMALL("Theora"),
2260 AVCodec vp3_decoder = {
2264 sizeof(Vp3DecodeContext),
2269 CODEC_CAP_DR1 | CODEC_CAP_DRAW_HORIZ_BAND,
2271 .long_name = NULL_IF_CONFIG_SMALL("On2 VP3"),