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 typedef struct Coeff {
53 //FIXME split things out into their own arrays
54 typedef struct Vp3Fragment {
56 uint8_t coding_method;
62 #define SB_NOT_CODED 0
63 #define SB_PARTIALLY_CODED 1
64 #define SB_FULLY_CODED 2
66 // This is the maximum length of a single long bit run that can be encoded
67 // for superblock coding or block qps. Theora special-cases this to read a
68 // bit instead of flipping the current bit to allow for runs longer than 4129.
69 #define MAXIMUM_LONG_BIT_RUN 4129
71 #define MODE_INTER_NO_MV 0
73 #define MODE_INTER_PLUS_MV 2
74 #define MODE_INTER_LAST_MV 3
75 #define MODE_INTER_PRIOR_LAST 4
76 #define MODE_USING_GOLDEN 5
77 #define MODE_GOLDEN_MV 6
78 #define MODE_INTER_FOURMV 7
79 #define CODING_MODE_COUNT 8
81 /* special internal mode */
84 /* There are 6 preset schemes, plus a free-form scheme */
85 static const int ModeAlphabet[6][CODING_MODE_COUNT] =
87 /* scheme 1: Last motion vector dominates */
88 { MODE_INTER_LAST_MV, MODE_INTER_PRIOR_LAST,
89 MODE_INTER_PLUS_MV, MODE_INTER_NO_MV,
90 MODE_INTRA, MODE_USING_GOLDEN,
91 MODE_GOLDEN_MV, MODE_INTER_FOURMV },
94 { MODE_INTER_LAST_MV, MODE_INTER_PRIOR_LAST,
95 MODE_INTER_NO_MV, MODE_INTER_PLUS_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_PRIOR_LAST, MODE_INTER_NO_MV,
102 MODE_INTRA, MODE_USING_GOLDEN,
103 MODE_GOLDEN_MV, MODE_INTER_FOURMV },
106 { MODE_INTER_LAST_MV, MODE_INTER_PLUS_MV,
107 MODE_INTER_NO_MV, MODE_INTER_PRIOR_LAST,
108 MODE_INTRA, MODE_USING_GOLDEN,
109 MODE_GOLDEN_MV, MODE_INTER_FOURMV },
111 /* scheme 5: No motion vector dominates */
112 { MODE_INTER_NO_MV, MODE_INTER_LAST_MV,
113 MODE_INTER_PRIOR_LAST, MODE_INTER_PLUS_MV,
114 MODE_INTRA, MODE_USING_GOLDEN,
115 MODE_GOLDEN_MV, MODE_INTER_FOURMV },
118 { MODE_INTER_NO_MV, MODE_USING_GOLDEN,
119 MODE_INTER_LAST_MV, MODE_INTER_PRIOR_LAST,
120 MODE_INTER_PLUS_MV, MODE_INTRA,
121 MODE_GOLDEN_MV, MODE_INTER_FOURMV },
125 #define MIN_DEQUANT_VAL 2
127 typedef struct Vp3DecodeContext {
128 AVCodecContext *avctx;
129 int theora, theora_tables;
132 AVFrame golden_frame;
134 AVFrame current_frame;
144 int superblock_count;
145 int y_superblock_width;
146 int y_superblock_height;
147 int c_superblock_width;
148 int c_superblock_height;
149 int u_superblock_start;
150 int v_superblock_start;
151 unsigned char *superblock_coding;
153 int macroblock_count;
154 int macroblock_width;
155 int macroblock_height;
161 Vp3Fragment *all_fragments;
162 uint8_t *coeff_counts;
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];
178 /* this is a list of indexes into the all_fragments array indicating
179 * which of the fragments are coded */
180 int *coded_fragment_list;
181 int coded_fragment_list_index;
183 /* track which fragments have already been decoded; called 'fast'
184 * because this data structure avoids having to iterate through every
185 * fragment in coded_fragment_list; once a fragment has been fully
186 * decoded, it is removed from this list */
187 int *fast_fragment_list;
188 int fragment_list_y_head;
189 int fragment_list_c_head;
197 VLC superblock_run_length_vlc;
198 VLC fragment_run_length_vlc;
200 VLC motion_vector_vlc;
202 /* these arrays need to be on 16-byte boundaries since SSE2 operations
204 DECLARE_ALIGNED_16(int16_t, qmat)[3][2][3][64]; //<qmat[qpi][is_inter][plane]
206 /* This table contains superblock_count * 16 entries. Each set of 16
207 * numbers corresponds to the fragment indexes 0..15 of the superblock.
208 * An entry will be -1 to indicate that no entry corresponds to that
210 int *superblock_fragments;
212 /* This is an array that indicates how a particular macroblock
214 unsigned char *macroblock_coding;
216 int first_coded_y_fragment;
217 int first_coded_c_fragment;
218 int last_coded_y_fragment;
219 int last_coded_c_fragment;
221 uint8_t edge_emu_buffer[9*2048]; //FIXME dynamic alloc
222 int8_t qscale_table[2048]; //FIXME dynamic alloc (width+15)/16
229 uint16_t huffman_table[80][32][2];
231 uint8_t filter_limit_values[64];
232 DECLARE_ALIGNED_8(int, bounding_values_array)[256+2];
235 /************************************************************************
236 * VP3 specific functions
237 ************************************************************************/
240 * This function sets up all of the various blocks mappings:
241 * superblocks <-> fragments, macroblocks <-> fragments,
242 * superblocks <-> macroblocks
244 * Returns 0 is successful; returns 1 if *anything* went wrong.
246 static int init_block_mapping(Vp3DecodeContext *s)
249 signed int hilbert_walk_mb[4];
251 int current_fragment = 0;
252 int current_width = 0;
253 int current_height = 0;
256 int superblock_row_inc = 0;
257 int mapping_index = 0;
259 int current_macroblock;
262 static const signed char travel_width[16] = {
269 static const signed char travel_height[16] = {
276 hilbert_walk_mb[0] = 1;
277 hilbert_walk_mb[1] = s->macroblock_width;
278 hilbert_walk_mb[2] = 1;
279 hilbert_walk_mb[3] = -s->macroblock_width;
281 /* iterate through each superblock (all planes) and map the fragments */
282 for (i = 0; i < s->superblock_count; i++) {
283 /* time to re-assign the limits? */
286 /* start of Y superblocks */
287 right_edge = s->fragment_width;
288 bottom_edge = s->fragment_height;
291 superblock_row_inc = 3 * s->fragment_width -
292 (s->y_superblock_width * 4 - s->fragment_width);
294 /* the first operation for this variable is to advance by 1 */
295 current_fragment = -1;
297 } else if (i == s->u_superblock_start) {
299 /* start of U superblocks */
300 right_edge = s->fragment_width / 2;
301 bottom_edge = s->fragment_height / 2;
304 superblock_row_inc = 3 * (s->fragment_width / 2) -
305 (s->c_superblock_width * 4 - s->fragment_width / 2);
307 /* the first operation for this variable is to advance by 1 */
308 current_fragment = s->fragment_start[1] - 1;
310 } else if (i == s->v_superblock_start) {
312 /* start of V superblocks */
313 right_edge = s->fragment_width / 2;
314 bottom_edge = s->fragment_height / 2;
317 superblock_row_inc = 3 * (s->fragment_width / 2) -
318 (s->c_superblock_width * 4 - s->fragment_width / 2);
320 /* the first operation for this variable is to advance by 1 */
321 current_fragment = s->fragment_start[2] - 1;
325 if (current_width >= right_edge - 1) {
326 /* reset width and move to next superblock row */
330 /* fragment is now at the start of a new superblock row */
331 current_fragment += superblock_row_inc;
334 /* iterate through all 16 fragments in a superblock */
335 for (j = 0; j < 16; j++) {
336 current_fragment += travel_width[j] + right_edge * travel_height[j];
337 current_width += travel_width[j];
338 current_height += travel_height[j];
340 /* check if the fragment is in bounds */
341 if ((current_width < right_edge) &&
342 (current_height < bottom_edge)) {
343 s->superblock_fragments[mapping_index] = current_fragment;
345 s->superblock_fragments[mapping_index] = -1;
352 return 0; /* successful path out */
356 * This function wipes out all of the fragment data.
358 static void init_frame(Vp3DecodeContext *s, GetBitContext *gb)
362 /* zero out all of the fragment information */
363 s->coded_fragment_list_index = 0;
364 for (i = 0; i < s->fragment_count; i++) {
365 s->coeff_counts[i] = 0;
366 s->all_fragments[i].motion_x = 127;
367 s->all_fragments[i].motion_y = 127;
368 s->all_fragments[i].next_coeff= NULL;
369 s->all_fragments[i].qpi = 0;
371 s->coeffs[i].coeff=0;
372 s->coeffs[i].next= NULL;
377 * This function sets up the dequantization tables used for a particular
380 static void init_dequantizer(Vp3DecodeContext *s, int qpi)
382 int ac_scale_factor = s->coded_ac_scale_factor[s->qps[qpi]];
383 int dc_scale_factor = s->coded_dc_scale_factor[s->qps[qpi]];
384 int i, plane, inter, qri, bmi, bmj, qistart;
386 for(inter=0; inter<2; inter++){
387 for(plane=0; plane<3; plane++){
389 for(qri=0; qri<s->qr_count[inter][plane]; qri++){
390 sum+= s->qr_size[inter][plane][qri];
391 if(s->qps[qpi] <= sum)
394 qistart= sum - s->qr_size[inter][plane][qri];
395 bmi= s->qr_base[inter][plane][qri ];
396 bmj= s->qr_base[inter][plane][qri+1];
398 int coeff= ( 2*(sum -s->qps[qpi])*s->base_matrix[bmi][i]
399 - 2*(qistart-s->qps[qpi])*s->base_matrix[bmj][i]
400 + s->qr_size[inter][plane][qri])
401 / (2*s->qr_size[inter][plane][qri]);
403 int qmin= 8<<(inter + !i);
404 int qscale= i ? ac_scale_factor : dc_scale_factor;
406 s->qmat[qpi][inter][plane][s->dsp.idct_permutation[i]]= av_clip((qscale * coeff)/100 * 4, qmin, 4096);
408 // all DC coefficients use the same quant so as not to interfere with DC prediction
409 s->qmat[qpi][inter][plane][0] = s->qmat[0][inter][plane][0];
413 memset(s->qscale_table, (FFMAX(s->qmat[0][0][0][1], s->qmat[0][0][1][1])+8)/16, 512); //FIXME finetune
417 * This function initializes the loop filter boundary limits if the frame's
418 * quality index is different from the previous frame's.
420 * The filter_limit_values may not be larger than 127.
422 static void init_loop_filter(Vp3DecodeContext *s)
424 int *bounding_values= s->bounding_values_array+127;
429 filter_limit = s->filter_limit_values[s->qps[0]];
431 /* set up the bounding values */
432 memset(s->bounding_values_array, 0, 256 * sizeof(int));
433 for (x = 0; x < filter_limit; x++) {
434 bounding_values[-x] = -x;
435 bounding_values[x] = x;
437 for (x = value = filter_limit; x < 128 && value; x++, value--) {
438 bounding_values[ x] = value;
439 bounding_values[-x] = -value;
442 bounding_values[128] = value;
443 bounding_values[129] = bounding_values[130] = filter_limit * 0x02020202;
447 * This function unpacks all of the superblock/macroblock/fragment coding
448 * information from the bitstream.
450 static int unpack_superblocks(Vp3DecodeContext *s, GetBitContext *gb)
453 int current_superblock = 0;
455 int num_partial_superblocks = 0;
456 int first_c_fragment_seen;
459 int current_fragment;
462 memset(s->superblock_coding, SB_FULLY_CODED, s->superblock_count);
466 /* unpack the list of partially-coded superblocks */
468 while (current_superblock < s->superblock_count) {
469 current_run = get_vlc2(gb,
470 s->superblock_run_length_vlc.table, 6, 2) + 1;
471 if (current_run == 34)
472 current_run += get_bits(gb, 12);
474 if (current_superblock + current_run > s->superblock_count) {
475 av_log(s->avctx, AV_LOG_ERROR, "Invalid partially coded superblock run length\n");
479 memset(s->superblock_coding + current_superblock, bit, current_run);
481 current_superblock += current_run;
483 num_partial_superblocks += current_run;
485 if (s->theora && current_run == MAXIMUM_LONG_BIT_RUN)
491 /* unpack the list of fully coded superblocks if any of the blocks were
492 * not marked as partially coded in the previous step */
493 if (num_partial_superblocks < s->superblock_count) {
494 int superblocks_decoded = 0;
496 current_superblock = 0;
498 while (superblocks_decoded < s->superblock_count - num_partial_superblocks) {
499 current_run = get_vlc2(gb,
500 s->superblock_run_length_vlc.table, 6, 2) + 1;
501 if (current_run == 34)
502 current_run += get_bits(gb, 12);
504 for (j = 0; j < current_run; current_superblock++) {
505 if (current_superblock >= s->superblock_count) {
506 av_log(s->avctx, AV_LOG_ERROR, "Invalid fully coded superblock run length\n");
510 /* skip any superblocks already marked as partially coded */
511 if (s->superblock_coding[current_superblock] == SB_NOT_CODED) {
512 s->superblock_coding[current_superblock] = 2*bit;
516 superblocks_decoded += current_run;
518 if (s->theora && current_run == MAXIMUM_LONG_BIT_RUN)
525 /* if there were partial blocks, initialize bitstream for
526 * unpacking fragment codings */
527 if (num_partial_superblocks) {
531 /* toggle the bit because as soon as the first run length is
532 * fetched the bit will be toggled again */
537 /* figure out which fragments are coded; iterate through each
538 * superblock (all planes) */
539 s->coded_fragment_list_index = 0;
540 s->next_coeff= s->coeffs + s->fragment_count;
541 s->first_coded_y_fragment = s->first_coded_c_fragment = 0;
542 s->last_coded_y_fragment = s->last_coded_c_fragment = -1;
543 first_c_fragment_seen = 0;
544 memset(s->macroblock_coding, MODE_COPY, s->macroblock_count);
545 for (i = 0; i < s->superblock_count; i++) {
547 /* iterate through all 16 fragments in a superblock */
548 for (j = 0; j < 16; j++) {
550 /* if the fragment is in bounds, check its coding status */
551 current_fragment = s->superblock_fragments[i * 16 + j];
552 if (current_fragment >= s->fragment_count) {
553 av_log(s->avctx, AV_LOG_ERROR, " vp3:unpack_superblocks(): bad fragment number (%d >= %d)\n",
554 current_fragment, s->fragment_count);
557 if (current_fragment != -1) {
558 int coded = s->superblock_coding[i];
560 if (s->superblock_coding[i] == SB_PARTIALLY_CODED) {
562 /* fragment may or may not be coded; this is the case
563 * that cares about the fragment coding runs */
564 if (current_run-- == 0) {
566 current_run = get_vlc2(gb,
567 s->fragment_run_length_vlc.table, 5, 2);
573 /* default mode; actual mode will be decoded in
575 s->all_fragments[current_fragment].coding_method =
577 s->all_fragments[current_fragment].next_coeff= s->coeffs + current_fragment;
578 s->coded_fragment_list[s->coded_fragment_list_index] =
580 if ((current_fragment >= s->fragment_start[1]) &&
581 (s->last_coded_y_fragment == -1) &&
582 (!first_c_fragment_seen)) {
583 s->first_coded_c_fragment = s->coded_fragment_list_index;
584 s->last_coded_y_fragment = s->first_coded_c_fragment - 1;
585 first_c_fragment_seen = 1;
587 s->coded_fragment_list_index++;
589 /* not coded; copy this fragment from the prior frame */
590 s->all_fragments[current_fragment].coding_method =
597 if (!first_c_fragment_seen)
598 /* only Y fragments coded in this frame */
599 s->last_coded_y_fragment = s->coded_fragment_list_index - 1;
601 /* end the list of coded C fragments */
602 s->last_coded_c_fragment = s->coded_fragment_list_index - 1;
604 for (i = 0; i < s->fragment_count - 1; i++) {
605 s->fast_fragment_list[i] = i + 1;
607 s->fast_fragment_list[s->fragment_count - 1] = -1;
609 if (s->last_coded_y_fragment == -1)
610 s->fragment_list_y_head = -1;
612 s->fragment_list_y_head = s->first_coded_y_fragment;
613 s->fast_fragment_list[s->last_coded_y_fragment] = -1;
616 if (s->last_coded_c_fragment == -1)
617 s->fragment_list_c_head = -1;
619 s->fragment_list_c_head = s->first_coded_c_fragment;
620 s->fast_fragment_list[s->last_coded_c_fragment] = -1;
627 * This function unpacks all the coding mode data for individual macroblocks
628 * from the bitstream.
630 static int unpack_modes(Vp3DecodeContext *s, GetBitContext *gb)
632 int i, j, k, sb_x, sb_y;
634 int current_macroblock;
635 int current_fragment;
637 int custom_mode_alphabet[CODING_MODE_COUNT];
641 for (i = 0; i < s->fragment_count; i++)
642 s->all_fragments[i].coding_method = MODE_INTRA;
646 /* fetch the mode coding scheme for this frame */
647 scheme = get_bits(gb, 3);
649 /* is it a custom coding scheme? */
651 for (i = 0; i < 8; i++)
652 custom_mode_alphabet[i] = MODE_INTER_NO_MV;
653 for (i = 0; i < 8; i++)
654 custom_mode_alphabet[get_bits(gb, 3)] = i;
655 alphabet = custom_mode_alphabet;
657 alphabet = ModeAlphabet[scheme-1];
659 /* iterate through all of the macroblocks that contain 1 or more
661 for (sb_y = 0; sb_y < s->y_superblock_height; sb_y++) {
662 for (sb_x = 0; sb_x < s->y_superblock_width; sb_x++) {
664 for (j = 0; j < 4; j++) {
665 int mb_x = 2*sb_x + (j>>1);
666 int mb_y = 2*sb_y + (((j>>1)+j)&1);
668 current_macroblock = mb_y * s->macroblock_width + mb_x;
670 if (mb_x >= s->macroblock_width || mb_y >= s->macroblock_height)
673 #define BLOCK_X (2*mb_x + (k&1))
674 #define BLOCK_Y (2*mb_y + (k>>1))
675 /* coding modes are only stored if the macroblock has at least one
676 * luma block coded, otherwise it must be INTER_NO_MV */
677 for (k = 0; k < 4; k++) {
678 current_fragment = BLOCK_Y*s->fragment_width + BLOCK_X;
679 if (s->all_fragments[current_fragment].coding_method != MODE_COPY)
683 s->macroblock_coding[current_macroblock] = MODE_INTER_NO_MV;
687 /* mode 7 means get 3 bits for each coding mode */
689 coding_mode = get_bits(gb, 3);
691 coding_mode = alphabet
692 [get_vlc2(gb, s->mode_code_vlc.table, 3, 3)];
694 s->macroblock_coding[current_macroblock] = coding_mode;
695 for (k = 0; k < 4; k++) {
697 BLOCK_Y*s->fragment_width + BLOCK_X;
698 if (s->all_fragments[current_fragment].coding_method !=
700 s->all_fragments[current_fragment].coding_method =
703 for (k = 0; k < 2; k++) {
704 current_fragment = s->fragment_start[k+1] +
705 mb_y*(s->fragment_width>>1) + mb_x;
706 if (s->all_fragments[current_fragment].coding_method !=
708 s->all_fragments[current_fragment].coding_method =
720 * This function unpacks all the motion vectors for the individual
721 * macroblocks from the bitstream.
723 static int unpack_vectors(Vp3DecodeContext *s, GetBitContext *gb)
725 int j, k, sb_x, sb_y;
729 int last_motion_x = 0;
730 int last_motion_y = 0;
731 int prior_last_motion_x = 0;
732 int prior_last_motion_y = 0;
733 int current_macroblock;
734 int current_fragment;
739 memset(motion_x, 0, 6 * sizeof(int));
740 memset(motion_y, 0, 6 * sizeof(int));
742 /* coding mode 0 is the VLC scheme; 1 is the fixed code scheme */
743 coding_mode = get_bits1(gb);
745 /* iterate through all of the macroblocks that contain 1 or more
747 for (sb_y = 0; sb_y < s->y_superblock_height; sb_y++) {
748 for (sb_x = 0; sb_x < s->y_superblock_width; sb_x++) {
750 for (j = 0; j < 4; j++) {
751 int mb_x = 2*sb_x + (j>>1);
752 int mb_y = 2*sb_y + (((j>>1)+j)&1);
753 current_macroblock = mb_y * s->macroblock_width + mb_x;
755 if (mb_x >= s->macroblock_width || mb_y >= s->macroblock_height ||
756 (s->macroblock_coding[current_macroblock] == MODE_COPY))
759 switch (s->macroblock_coding[current_macroblock]) {
761 case MODE_INTER_PLUS_MV:
763 /* all 6 fragments use the same motion vector */
764 if (coding_mode == 0) {
765 motion_x[0] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
766 motion_y[0] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
768 motion_x[0] = fixed_motion_vector_table[get_bits(gb, 6)];
769 motion_y[0] = fixed_motion_vector_table[get_bits(gb, 6)];
772 /* vector maintenance, only on MODE_INTER_PLUS_MV */
773 if (s->macroblock_coding[current_macroblock] ==
774 MODE_INTER_PLUS_MV) {
775 prior_last_motion_x = last_motion_x;
776 prior_last_motion_y = last_motion_y;
777 last_motion_x = motion_x[0];
778 last_motion_y = motion_y[0];
782 case MODE_INTER_FOURMV:
783 /* vector maintenance */
784 prior_last_motion_x = last_motion_x;
785 prior_last_motion_y = last_motion_y;
787 /* fetch 4 vectors from the bitstream, one for each
788 * Y fragment, then average for the C fragment vectors */
789 motion_x[4] = motion_y[4] = 0;
790 for (k = 0; k < 4; k++) {
791 current_fragment = BLOCK_Y*s->fragment_width + BLOCK_X;
792 if (s->all_fragments[current_fragment].coding_method != MODE_COPY) {
793 if (coding_mode == 0) {
794 motion_x[k] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
795 motion_y[k] = motion_vector_table[get_vlc2(gb, s->motion_vector_vlc.table, 6, 2)];
797 motion_x[k] = fixed_motion_vector_table[get_bits(gb, 6)];
798 motion_y[k] = fixed_motion_vector_table[get_bits(gb, 6)];
800 last_motion_x = motion_x[k];
801 last_motion_y = motion_y[k];
806 motion_x[4] += motion_x[k];
807 motion_y[4] += motion_y[k];
811 motion_x[4]= RSHIFT(motion_x[4], 2);
813 motion_y[4]= RSHIFT(motion_y[4], 2);
816 case MODE_INTER_LAST_MV:
817 /* all 6 fragments use the last motion vector */
818 motion_x[0] = last_motion_x;
819 motion_y[0] = last_motion_y;
821 /* no vector maintenance (last vector remains the
825 case MODE_INTER_PRIOR_LAST:
826 /* all 6 fragments use the motion vector prior to the
827 * last motion vector */
828 motion_x[0] = prior_last_motion_x;
829 motion_y[0] = prior_last_motion_y;
831 /* vector maintenance */
832 prior_last_motion_x = last_motion_x;
833 prior_last_motion_y = last_motion_y;
834 last_motion_x = motion_x[0];
835 last_motion_y = motion_y[0];
839 /* covers intra, inter without MV, golden without MV */
843 /* no vector maintenance */
847 /* assign the motion vectors to the correct fragments */
848 for (k = 0; k < 4; k++) {
850 BLOCK_Y*s->fragment_width + BLOCK_X;
851 if (s->macroblock_coding[current_macroblock] == MODE_INTER_FOURMV) {
852 s->all_fragments[current_fragment].motion_x = motion_x[k];
853 s->all_fragments[current_fragment].motion_y = motion_y[k];
855 s->all_fragments[current_fragment].motion_x = motion_x[0];
856 s->all_fragments[current_fragment].motion_y = motion_y[0];
859 for (k = 0; k < 2; k++) {
860 current_fragment = s->fragment_start[k+1] +
861 mb_y*(s->fragment_width>>1) + mb_x;
862 if (s->macroblock_coding[current_macroblock] == MODE_INTER_FOURMV) {
863 s->all_fragments[current_fragment].motion_x = motion_x[k+4];
864 s->all_fragments[current_fragment].motion_y = motion_y[k+4];
866 s->all_fragments[current_fragment].motion_x = motion_x[0];
867 s->all_fragments[current_fragment].motion_y = motion_y[0];
877 static int unpack_block_qpis(Vp3DecodeContext *s, GetBitContext *gb)
879 int qpi, i, j, bit, run_length, blocks_decoded, num_blocks_at_qpi;
880 int num_blocks = s->coded_fragment_list_index;
882 for (qpi = 0; qpi < s->nqps-1 && num_blocks > 0; qpi++) {
883 i = blocks_decoded = num_blocks_at_qpi = 0;
888 run_length = get_vlc2(gb, s->superblock_run_length_vlc.table, 6, 2) + 1;
889 if (run_length == 34)
890 run_length += get_bits(gb, 12);
891 blocks_decoded += run_length;
894 num_blocks_at_qpi += run_length;
896 for (j = 0; j < run_length; i++) {
897 if (i >= s->coded_fragment_list_index)
900 if (s->all_fragments[s->coded_fragment_list[i]].qpi == qpi) {
901 s->all_fragments[s->coded_fragment_list[i]].qpi += bit;
906 if (run_length == MAXIMUM_LONG_BIT_RUN)
910 } while (blocks_decoded < num_blocks);
912 num_blocks -= num_blocks_at_qpi;
919 * This function is called by unpack_dct_coeffs() to extract the VLCs from
920 * the bitstream. The VLCs encode tokens which are used to unpack DCT
921 * data. This function unpacks all the VLCs for either the Y plane or both
922 * C planes, and is called for DC coefficients or different AC coefficient
923 * levels (since different coefficient types require different VLC tables.
925 * This function returns a residual eob run. E.g, if a particular token gave
926 * instructions to EOB the next 5 fragments and there were only 2 fragments
927 * left in the current fragment range, 3 would be returned so that it could
928 * be passed into the next call to this same function.
930 static int unpack_vlcs(Vp3DecodeContext *s, GetBitContext *gb,
931 VLC *table, int coeff_index,
939 Vp3Fragment *fragment;
942 int previous_fragment;
946 /* local references to structure members to avoid repeated deferences */
947 uint8_t *perm= s->scantable.permutated;
948 int *coded_fragment_list = s->coded_fragment_list;
949 Vp3Fragment *all_fragments = s->all_fragments;
950 uint8_t *coeff_counts = s->coeff_counts;
951 VLC_TYPE (*vlc_table)[2] = table->table;
952 int *fast_fragment_list = s->fast_fragment_list;
955 next_fragment = s->fragment_list_y_head;
956 list_head = &s->fragment_list_y_head;
958 next_fragment = s->fragment_list_c_head;
959 list_head = &s->fragment_list_c_head;
963 previous_fragment = -1; /* this indicates that the previous fragment is actually the list head */
965 fragment_num = coded_fragment_list[i];
967 if (coeff_counts[fragment_num] > coeff_index) {
968 previous_fragment = i;
969 i = fast_fragment_list[i];
972 fragment = &all_fragments[fragment_num];
975 /* decode a VLC into a token */
976 token = get_vlc2(gb, vlc_table, 5, 3);
977 /* use the token to get a zero run, a coefficient, and an eob run */
979 eob_run = eob_run_base[token];
980 if (eob_run_get_bits[token])
981 eob_run += get_bits(gb, eob_run_get_bits[token]);
982 coeff = zero_run = 0;
984 bits_to_get = coeff_get_bits[token];
986 bits_to_get = get_bits(gb, bits_to_get);
987 coeff = coeff_tables[token][bits_to_get];
989 zero_run = zero_run_base[token];
990 if (zero_run_get_bits[token])
991 zero_run += get_bits(gb, zero_run_get_bits[token]);
996 coeff_counts[fragment_num] += zero_run;
997 if (coeff_counts[fragment_num] < 64){
998 fragment->next_coeff->coeff= coeff;
999 fragment->next_coeff->index= perm[coeff_counts[fragment_num]++]; //FIXME perm here already?
1000 fragment->next_coeff->next= s->next_coeff;
1001 s->next_coeff->next=NULL;
1002 fragment->next_coeff= s->next_coeff++;
1004 /* previous fragment is now this fragment */
1005 previous_fragment = i;
1007 coeff_counts[fragment_num] |= 128;
1009 /* remove this fragment from the list */
1010 if (previous_fragment != -1)
1011 fast_fragment_list[previous_fragment] = fast_fragment_list[i];
1013 *list_head = fast_fragment_list[i];
1014 /* previous fragment remains unchanged */
1017 i = fast_fragment_list[i];
1023 static void reverse_dc_prediction(Vp3DecodeContext *s,
1026 int fragment_height);
1028 * This function unpacks all of the DCT coefficient data from the
1031 static int unpack_dct_coeffs(Vp3DecodeContext *s, GetBitContext *gb)
1038 int residual_eob_run = 0;
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 1, 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 0, residual_eob_run);
1057 /* reverse prediction of the C-plane DC coefficients */
1058 if (!(s->avctx->flags & CODEC_FLAG_GRAY))
1060 reverse_dc_prediction(s, s->fragment_start[1],
1061 s->fragment_width / 2, s->fragment_height / 2);
1062 reverse_dc_prediction(s, s->fragment_start[2],
1063 s->fragment_width / 2, s->fragment_height / 2);
1066 /* fetch the AC table indexes */
1067 ac_y_table = get_bits(gb, 4);
1068 ac_c_table = get_bits(gb, 4);
1070 /* build tables of AC VLC tables */
1071 for (i = 1; i <= 5; i++) {
1072 y_tables[i] = &s->ac_vlc_1[ac_y_table];
1073 c_tables[i] = &s->ac_vlc_1[ac_c_table];
1075 for (i = 6; i <= 14; i++) {
1076 y_tables[i] = &s->ac_vlc_2[ac_y_table];
1077 c_tables[i] = &s->ac_vlc_2[ac_c_table];
1079 for (i = 15; i <= 27; i++) {
1080 y_tables[i] = &s->ac_vlc_3[ac_y_table];
1081 c_tables[i] = &s->ac_vlc_3[ac_c_table];
1083 for (i = 28; i <= 63; i++) {
1084 y_tables[i] = &s->ac_vlc_4[ac_y_table];
1085 c_tables[i] = &s->ac_vlc_4[ac_c_table];
1088 /* decode all AC coefficents */
1089 for (i = 1; i <= 63; i++) {
1090 if (s->fragment_list_y_head != -1)
1091 residual_eob_run = unpack_vlcs(s, gb, y_tables[i], i,
1092 1, residual_eob_run);
1094 if (s->fragment_list_c_head != -1)
1095 residual_eob_run = unpack_vlcs(s, gb, c_tables[i], i,
1096 0, residual_eob_run);
1103 * This function reverses the DC prediction for each coded fragment in
1104 * the frame. Much of this function is adapted directly from the original
1107 #define COMPATIBLE_FRAME(x) \
1108 (compatible_frame[s->all_fragments[x].coding_method] == current_frame_type)
1109 #define DC_COEFF(u) (s->coeffs[u].index ? 0 : s->coeffs[u].coeff) //FIXME do somethin to simplify this
1111 static void reverse_dc_prediction(Vp3DecodeContext *s,
1114 int fragment_height)
1123 int i = first_fragment;
1127 /* DC values for the left, up-left, up, and up-right fragments */
1128 int vl, vul, vu, vur;
1130 /* indexes for the left, up-left, up, and up-right fragments */
1134 * The 6 fields mean:
1135 * 0: up-left multiplier
1137 * 2: up-right multiplier
1138 * 3: left multiplier
1140 static const int predictor_transform[16][4] = {
1142 { 0, 0, 0,128}, // PL
1143 { 0, 0,128, 0}, // PUR
1144 { 0, 0, 53, 75}, // PUR|PL
1145 { 0,128, 0, 0}, // PU
1146 { 0, 64, 0, 64}, // PU|PL
1147 { 0,128, 0, 0}, // PU|PUR
1148 { 0, 0, 53, 75}, // PU|PUR|PL
1149 {128, 0, 0, 0}, // PUL
1150 { 0, 0, 0,128}, // PUL|PL
1151 { 64, 0, 64, 0}, // PUL|PUR
1152 { 0, 0, 53, 75}, // PUL|PUR|PL
1153 { 0,128, 0, 0}, // PUL|PU
1154 {-104,116, 0,116}, // PUL|PU|PL
1155 { 24, 80, 24, 0}, // PUL|PU|PUR
1156 {-104,116, 0,116} // PUL|PU|PUR|PL
1159 /* This table shows which types of blocks can use other blocks for
1160 * prediction. For example, INTRA is the only mode in this table to
1161 * have a frame number of 0. That means INTRA blocks can only predict
1162 * from other INTRA blocks. There are 2 golden frame coding types;
1163 * blocks encoding in these modes can only predict from other blocks
1164 * that were encoded with these 1 of these 2 modes. */
1165 static const unsigned char compatible_frame[9] = {
1166 1, /* MODE_INTER_NO_MV */
1168 1, /* MODE_INTER_PLUS_MV */
1169 1, /* MODE_INTER_LAST_MV */
1170 1, /* MODE_INTER_PRIOR_MV */
1171 2, /* MODE_USING_GOLDEN */
1172 2, /* MODE_GOLDEN_MV */
1173 1, /* MODE_INTER_FOUR_MV */
1176 int current_frame_type;
1178 /* there is a last DC predictor for each of the 3 frame types */
1183 vul = vu = vur = vl = 0;
1184 last_dc[0] = last_dc[1] = last_dc[2] = 0;
1186 /* for each fragment row... */
1187 for (y = 0; y < fragment_height; y++) {
1189 /* for each fragment in a row... */
1190 for (x = 0; x < fragment_width; x++, i++) {
1192 /* reverse prediction if this block was coded */
1193 if (s->all_fragments[i].coding_method != MODE_COPY) {
1195 current_frame_type =
1196 compatible_frame[s->all_fragments[i].coding_method];
1202 if(COMPATIBLE_FRAME(l))
1206 u= i-fragment_width;
1208 if(COMPATIBLE_FRAME(u))
1211 ul= i-fragment_width-1;
1213 if(COMPATIBLE_FRAME(ul))
1216 if(x + 1 < fragment_width){
1217 ur= i-fragment_width+1;
1219 if(COMPATIBLE_FRAME(ur))
1224 if (transform == 0) {
1226 /* if there were no fragments to predict from, use last
1228 predicted_dc = last_dc[current_frame_type];
1231 /* apply the appropriate predictor transform */
1233 (predictor_transform[transform][0] * vul) +
1234 (predictor_transform[transform][1] * vu) +
1235 (predictor_transform[transform][2] * vur) +
1236 (predictor_transform[transform][3] * vl);
1238 predicted_dc /= 128;
1240 /* check for outranging on the [ul u l] and
1241 * [ul u ur l] predictors */
1242 if ((transform == 15) || (transform == 13)) {
1243 if (FFABS(predicted_dc - vu) > 128)
1245 else if (FFABS(predicted_dc - vl) > 128)
1247 else if (FFABS(predicted_dc - vul) > 128)
1252 /* at long last, apply the predictor */
1253 if(s->coeffs[i].index){
1254 *s->next_coeff= s->coeffs[i];
1255 s->coeffs[i].index=0;
1256 s->coeffs[i].coeff=0;
1257 s->coeffs[i].next= s->next_coeff++;
1259 s->coeffs[i].coeff += predicted_dc;
1261 last_dc[current_frame_type] = DC_COEFF(i);
1262 if(DC_COEFF(i) && !(s->coeff_counts[i]&127)){
1263 s->coeff_counts[i]= 129;
1264 // s->all_fragments[i].next_coeff= s->next_coeff;
1265 s->coeffs[i].next= s->next_coeff;
1266 (s->next_coeff++)->next=NULL;
1273 static void apply_loop_filter(Vp3DecodeContext *s, int plane, int ystart, int yend)
1276 int *bounding_values= s->bounding_values_array+127;
1278 int width = s->fragment_width >> !!plane;
1279 int height = s->fragment_height >> !!plane;
1280 int fragment = s->fragment_start [plane] + ystart * width;
1281 int stride = s->current_frame.linesize[plane];
1282 uint8_t *plane_data = s->current_frame.data [plane];
1283 if (!s->flipped_image) stride = -stride;
1284 plane_data += s->data_offset[plane] + 8*ystart*stride;
1286 for (y = ystart; y < yend; y++) {
1288 for (x = 0; x < width; x++) {
1289 /* This code basically just deblocks on the edges of coded blocks.
1290 * However, it has to be much more complicated because of the
1291 * braindamaged deblock ordering used in VP3/Theora. Order matters
1292 * because some pixels get filtered twice. */
1293 if( s->all_fragments[fragment].coding_method != MODE_COPY )
1295 /* do not perform left edge filter for left columns frags */
1297 s->dsp.vp3_h_loop_filter(
1299 stride, bounding_values);
1302 /* do not perform top edge filter for top row fragments */
1304 s->dsp.vp3_v_loop_filter(
1306 stride, bounding_values);
1309 /* do not perform right edge filter for right column
1310 * fragments or if right fragment neighbor is also coded
1311 * in this frame (it will be filtered in next iteration) */
1312 if ((x < width - 1) &&
1313 (s->all_fragments[fragment + 1].coding_method == MODE_COPY)) {
1314 s->dsp.vp3_h_loop_filter(
1315 plane_data + 8*x + 8,
1316 stride, bounding_values);
1319 /* do not perform bottom edge filter for bottom row
1320 * fragments or if bottom fragment neighbor is also coded
1321 * in this frame (it will be filtered in the next row) */
1322 if ((y < height - 1) &&
1323 (s->all_fragments[fragment + width].coding_method == MODE_COPY)) {
1324 s->dsp.vp3_v_loop_filter(
1325 plane_data + 8*x + 8*stride,
1326 stride, bounding_values);
1332 plane_data += 8*stride;
1337 * called when all pixels up to row y are complete
1339 static void vp3_draw_horiz_band(Vp3DecodeContext *s, int y)
1344 if(s->avctx->draw_horiz_band==NULL)
1347 h= y - s->last_slice_end;
1350 if (!s->flipped_image) {
1352 h -= s->height - s->avctx->height; // account for non-mod16
1353 y = s->height - y - h;
1357 offset[0] = s->current_frame.linesize[0]*y;
1358 offset[1] = s->current_frame.linesize[1]*cy;
1359 offset[2] = s->current_frame.linesize[2]*cy;
1363 s->avctx->draw_horiz_band(s->avctx, &s->current_frame, offset, y, 3, h);
1364 s->last_slice_end= y + h;
1368 * Perform the final rendering for a particular slice of data.
1369 * The slice number ranges from 0..(macroblock_height - 1).
1371 static void render_slice(Vp3DecodeContext *s, int slice)
1374 int16_t *dequantizer;
1375 LOCAL_ALIGNED_16(DCTELEM, block, [64]);
1376 int motion_x = 0xdeadbeef, motion_y = 0xdeadbeef;
1377 int motion_halfpel_index;
1378 uint8_t *motion_source;
1381 if (slice >= s->macroblock_height)
1384 for (plane = 0; plane < 3; plane++) {
1385 uint8_t *output_plane = s->current_frame.data [plane] + s->data_offset[plane];
1386 uint8_t * last_plane = s-> last_frame.data [plane] + s->data_offset[plane];
1387 uint8_t *golden_plane = s-> golden_frame.data [plane] + s->data_offset[plane];
1388 int stride = s->current_frame.linesize[plane];
1389 int plane_width = s->width >> !!plane;
1390 int plane_height = s->height >> !!plane;
1391 int y = slice * FRAGMENT_PIXELS << !plane ;
1392 int slice_height = y + (FRAGMENT_PIXELS << !plane);
1393 int i = s->fragment_start[plane] + (y>>3)*(s->fragment_width>>!!plane);
1395 if (!s->flipped_image) stride = -stride;
1396 if (CONFIG_GRAY && plane && (s->avctx->flags & CODEC_FLAG_GRAY))
1400 if(FFABS(stride) > 2048)
1401 return; //various tables are fixed size
1403 /* for each fragment row in the slice (both of them)... */
1404 for (; y < slice_height; y += 8) {
1406 /* for each fragment in a row... */
1407 for (x = 0; x < plane_width; x += 8, i++) {
1408 int first_pixel = y*stride + x;
1410 if ((i < 0) || (i >= s->fragment_count)) {
1411 av_log(s->avctx, AV_LOG_ERROR, " vp3:render_slice(): bad fragment number (%d)\n", i);
1415 /* transform if this block was coded */
1416 if (s->all_fragments[i].coding_method != MODE_COPY) {
1418 if ((s->all_fragments[i].coding_method == MODE_USING_GOLDEN) ||
1419 (s->all_fragments[i].coding_method == MODE_GOLDEN_MV))
1420 motion_source= golden_plane;
1422 motion_source= last_plane;
1424 motion_source += first_pixel;
1425 motion_halfpel_index = 0;
1427 /* sort out the motion vector if this fragment is coded
1428 * using a motion vector method */
1429 if ((s->all_fragments[i].coding_method > MODE_INTRA) &&
1430 (s->all_fragments[i].coding_method != MODE_USING_GOLDEN)) {
1432 motion_x = s->all_fragments[i].motion_x;
1433 motion_y = s->all_fragments[i].motion_y;
1435 motion_x= (motion_x>>1) | (motion_x&1);
1436 motion_y= (motion_y>>1) | (motion_y&1);
1439 src_x= (motion_x>>1) + x;
1440 src_y= (motion_y>>1) + y;
1441 if ((motion_x == 127) || (motion_y == 127))
1442 av_log(s->avctx, AV_LOG_ERROR, " help! got invalid motion vector! (%X, %X)\n", motion_x, motion_y);
1444 motion_halfpel_index = motion_x & 0x01;
1445 motion_source += (motion_x >> 1);
1447 motion_halfpel_index |= (motion_y & 0x01) << 1;
1448 motion_source += ((motion_y >> 1) * stride);
1450 if(src_x<0 || src_y<0 || src_x + 9 >= plane_width || src_y + 9 >= plane_height){
1451 uint8_t *temp= s->edge_emu_buffer;
1452 if(stride<0) temp -= 9*stride;
1453 else temp += 9*stride;
1455 ff_emulated_edge_mc(temp, motion_source, stride, 9, 9, src_x, src_y, plane_width, plane_height);
1456 motion_source= temp;
1461 /* first, take care of copying a block from either the
1462 * previous or the golden frame */
1463 if (s->all_fragments[i].coding_method != MODE_INTRA) {
1464 /* Note, it is possible to implement all MC cases with
1465 put_no_rnd_pixels_l2 which would look more like the
1466 VP3 source but this would be slower as
1467 put_no_rnd_pixels_tab is better optimzed */
1468 if(motion_halfpel_index != 3){
1469 s->dsp.put_no_rnd_pixels_tab[1][motion_halfpel_index](
1470 output_plane + first_pixel,
1471 motion_source, stride, 8);
1473 int d= (motion_x ^ motion_y)>>31; // d is 0 if motion_x and _y have the same sign, else -1
1474 s->dsp.put_no_rnd_pixels_l2[1](
1475 output_plane + first_pixel,
1477 motion_source + stride + 1 + d,
1480 dequantizer = s->qmat[s->all_fragments[i].qpi][1][plane];
1482 dequantizer = s->qmat[s->all_fragments[i].qpi][0][plane];
1485 /* dequantize the DCT coefficients */
1486 if(s->avctx->idct_algo==FF_IDCT_VP3){
1487 Coeff *coeff= s->coeffs + i;
1488 s->dsp.clear_block(block);
1490 block[coeff->index]= coeff->coeff * dequantizer[coeff->index];
1494 Coeff *coeff= s->coeffs + i;
1495 s->dsp.clear_block(block);
1497 block[coeff->index]= (coeff->coeff * dequantizer[coeff->index] + 2)>>2;
1502 /* invert DCT and place (or add) in final output */
1504 if (s->all_fragments[i].coding_method == MODE_INTRA) {
1505 if(s->avctx->idct_algo!=FF_IDCT_VP3)
1508 output_plane + first_pixel,
1513 output_plane + first_pixel,
1519 /* copy directly from the previous frame */
1520 s->dsp.put_pixels_tab[1][0](
1521 output_plane + first_pixel,
1522 last_plane + first_pixel,
1527 // Filter the previous block row. We can't filter the current row yet
1528 // since it needs pixels from the next row
1530 apply_loop_filter(s, plane, (y>>3)-1, (y>>3));
1534 /* this looks like a good place for slice dispatch... */
1536 * if (slice == s->macroblock_height - 1)
1537 * dispatch (both last slice & 2nd-to-last slice);
1538 * else if (slice > 0)
1539 * dispatch (slice - 1);
1542 // now that we've filtered the last rows, they're safe to display
1544 vp3_draw_horiz_band(s, 16*slice);
1548 * This is the ffmpeg/libavcodec API init function.
1550 static av_cold int vp3_decode_init(AVCodecContext *avctx)
1552 Vp3DecodeContext *s = avctx->priv_data;
1553 int i, inter, plane;
1556 int y_superblock_count;
1557 int c_superblock_count;
1559 if (avctx->codec_tag == MKTAG('V','P','3','0'))
1565 s->width = FFALIGN(avctx->width, 16);
1566 s->height = FFALIGN(avctx->height, 16);
1567 avctx->pix_fmt = PIX_FMT_YUV420P;
1568 avctx->chroma_sample_location = AVCHROMA_LOC_CENTER;
1569 if(avctx->idct_algo==FF_IDCT_AUTO)
1570 avctx->idct_algo=FF_IDCT_VP3;
1571 dsputil_init(&s->dsp, avctx);
1573 ff_init_scantable(s->dsp.idct_permutation, &s->scantable, ff_zigzag_direct);
1575 /* initialize to an impossible value which will force a recalculation
1576 * in the first frame decode */
1577 for (i = 0; i < 3; i++)
1580 s->y_superblock_width = (s->width + 31) / 32;
1581 s->y_superblock_height = (s->height + 31) / 32;
1582 y_superblock_count = s->y_superblock_width * s->y_superblock_height;
1584 /* work out the dimensions for the C planes */
1585 c_width = s->width / 2;
1586 c_height = s->height / 2;
1587 s->c_superblock_width = (c_width + 31) / 32;
1588 s->c_superblock_height = (c_height + 31) / 32;
1589 c_superblock_count = s->c_superblock_width * s->c_superblock_height;
1591 s->superblock_count = y_superblock_count + (c_superblock_count * 2);
1592 s->u_superblock_start = y_superblock_count;
1593 s->v_superblock_start = s->u_superblock_start + c_superblock_count;
1594 s->superblock_coding = av_malloc(s->superblock_count);
1596 s->macroblock_width = (s->width + 15) / 16;
1597 s->macroblock_height = (s->height + 15) / 16;
1598 s->macroblock_count = s->macroblock_width * s->macroblock_height;
1600 s->fragment_width = s->width / FRAGMENT_PIXELS;
1601 s->fragment_height = s->height / FRAGMENT_PIXELS;
1603 /* fragment count covers all 8x8 blocks for all 3 planes */
1604 s->fragment_count = s->fragment_width * s->fragment_height * 3 / 2;
1605 s->fragment_start[1] = s->fragment_width * s->fragment_height;
1606 s->fragment_start[2] = s->fragment_width * s->fragment_height * 5 / 4;
1608 s->all_fragments = av_malloc(s->fragment_count * sizeof(Vp3Fragment));
1609 s->coeff_counts = av_malloc(s->fragment_count * sizeof(*s->coeff_counts));
1610 s->coeffs = av_malloc(s->fragment_count * sizeof(Coeff) * 65);
1611 s->coded_fragment_list = av_malloc(s->fragment_count * sizeof(int));
1612 s->fast_fragment_list = av_malloc(s->fragment_count * sizeof(int));
1613 if (!s->superblock_coding || !s->all_fragments || !s->coeff_counts ||
1614 !s->coeffs || !s->coded_fragment_list || !s->fast_fragment_list) {
1615 vp3_decode_end(avctx);
1619 if (!s->theora_tables)
1621 for (i = 0; i < 64; i++) {
1622 s->coded_dc_scale_factor[i] = vp31_dc_scale_factor[i];
1623 s->coded_ac_scale_factor[i] = vp31_ac_scale_factor[i];
1624 s->base_matrix[0][i] = vp31_intra_y_dequant[i];
1625 s->base_matrix[1][i] = vp31_intra_c_dequant[i];
1626 s->base_matrix[2][i] = vp31_inter_dequant[i];
1627 s->filter_limit_values[i] = vp31_filter_limit_values[i];
1630 for(inter=0; inter<2; inter++){
1631 for(plane=0; plane<3; plane++){
1632 s->qr_count[inter][plane]= 1;
1633 s->qr_size [inter][plane][0]= 63;
1634 s->qr_base [inter][plane][0]=
1635 s->qr_base [inter][plane][1]= 2*inter + (!!plane)*!inter;
1639 /* init VLC tables */
1640 for (i = 0; i < 16; i++) {
1643 init_vlc(&s->dc_vlc[i], 5, 32,
1644 &dc_bias[i][0][1], 4, 2,
1645 &dc_bias[i][0][0], 4, 2, 0);
1647 /* group 1 AC histograms */
1648 init_vlc(&s->ac_vlc_1[i], 5, 32,
1649 &ac_bias_0[i][0][1], 4, 2,
1650 &ac_bias_0[i][0][0], 4, 2, 0);
1652 /* group 2 AC histograms */
1653 init_vlc(&s->ac_vlc_2[i], 5, 32,
1654 &ac_bias_1[i][0][1], 4, 2,
1655 &ac_bias_1[i][0][0], 4, 2, 0);
1657 /* group 3 AC histograms */
1658 init_vlc(&s->ac_vlc_3[i], 5, 32,
1659 &ac_bias_2[i][0][1], 4, 2,
1660 &ac_bias_2[i][0][0], 4, 2, 0);
1662 /* group 4 AC histograms */
1663 init_vlc(&s->ac_vlc_4[i], 5, 32,
1664 &ac_bias_3[i][0][1], 4, 2,
1665 &ac_bias_3[i][0][0], 4, 2, 0);
1668 for (i = 0; i < 16; i++) {
1671 if (init_vlc(&s->dc_vlc[i], 5, 32,
1672 &s->huffman_table[i][0][1], 4, 2,
1673 &s->huffman_table[i][0][0], 4, 2, 0) < 0)
1676 /* group 1 AC histograms */
1677 if (init_vlc(&s->ac_vlc_1[i], 5, 32,
1678 &s->huffman_table[i+16][0][1], 4, 2,
1679 &s->huffman_table[i+16][0][0], 4, 2, 0) < 0)
1682 /* group 2 AC histograms */
1683 if (init_vlc(&s->ac_vlc_2[i], 5, 32,
1684 &s->huffman_table[i+16*2][0][1], 4, 2,
1685 &s->huffman_table[i+16*2][0][0], 4, 2, 0) < 0)
1688 /* group 3 AC histograms */
1689 if (init_vlc(&s->ac_vlc_3[i], 5, 32,
1690 &s->huffman_table[i+16*3][0][1], 4, 2,
1691 &s->huffman_table[i+16*3][0][0], 4, 2, 0) < 0)
1694 /* group 4 AC histograms */
1695 if (init_vlc(&s->ac_vlc_4[i], 5, 32,
1696 &s->huffman_table[i+16*4][0][1], 4, 2,
1697 &s->huffman_table[i+16*4][0][0], 4, 2, 0) < 0)
1702 init_vlc(&s->superblock_run_length_vlc, 6, 34,
1703 &superblock_run_length_vlc_table[0][1], 4, 2,
1704 &superblock_run_length_vlc_table[0][0], 4, 2, 0);
1706 init_vlc(&s->fragment_run_length_vlc, 5, 30,
1707 &fragment_run_length_vlc_table[0][1], 4, 2,
1708 &fragment_run_length_vlc_table[0][0], 4, 2, 0);
1710 init_vlc(&s->mode_code_vlc, 3, 8,
1711 &mode_code_vlc_table[0][1], 2, 1,
1712 &mode_code_vlc_table[0][0], 2, 1, 0);
1714 init_vlc(&s->motion_vector_vlc, 6, 63,
1715 &motion_vector_vlc_table[0][1], 2, 1,
1716 &motion_vector_vlc_table[0][0], 2, 1, 0);
1718 /* work out the block mapping tables */
1719 s->superblock_fragments = av_malloc(s->superblock_count * 16 * sizeof(int));
1720 s->macroblock_coding = av_malloc(s->macroblock_count + 1);
1721 if (!s->superblock_fragments || !s->macroblock_coding) {
1722 vp3_decode_end(avctx);
1725 init_block_mapping(s);
1727 for (i = 0; i < 3; i++) {
1728 s->current_frame.data[i] = NULL;
1729 s->last_frame.data[i] = NULL;
1730 s->golden_frame.data[i] = NULL;
1736 av_log(avctx, AV_LOG_FATAL, "Invalid huffman table\n");
1741 * This is the ffmpeg/libavcodec API frame decode function.
1743 static int vp3_decode_frame(AVCodecContext *avctx,
1744 void *data, int *data_size,
1747 const uint8_t *buf = avpkt->data;
1748 int buf_size = avpkt->size;
1749 Vp3DecodeContext *s = avctx->priv_data;
1751 static int counter = 0;
1754 init_get_bits(&gb, buf, buf_size * 8);
1756 if (s->theora && get_bits1(&gb))
1758 av_log(avctx, AV_LOG_ERROR, "Header packet passed to frame decoder, skipping\n");
1762 s->keyframe = !get_bits1(&gb);
1765 for (i = 0; i < 3; i++)
1766 s->last_qps[i] = s->qps[i];
1770 s->qps[s->nqps++]= get_bits(&gb, 6);
1771 } while(s->theora >= 0x030200 && s->nqps<3 && get_bits1(&gb));
1772 for (i = s->nqps; i < 3; i++)
1775 if (s->avctx->debug & FF_DEBUG_PICT_INFO)
1776 av_log(s->avctx, AV_LOG_INFO, " VP3 %sframe #%d: Q index = %d\n",
1777 s->keyframe?"key":"", counter, s->qps[0]);
1780 if (s->qps[0] != s->last_qps[0])
1781 init_loop_filter(s);
1783 for (i = 0; i < s->nqps; i++)
1784 // reinit all dequantizers if the first one changed, because
1785 // the DC of the first quantizer must be used for all matrices
1786 if (s->qps[i] != s->last_qps[i] || s->qps[0] != s->last_qps[0])
1787 init_dequantizer(s, i);
1789 if (avctx->skip_frame >= AVDISCARD_NONKEY && !s->keyframe)
1795 skip_bits(&gb, 4); /* width code */
1796 skip_bits(&gb, 4); /* height code */
1799 s->version = get_bits(&gb, 5);
1801 av_log(s->avctx, AV_LOG_DEBUG, "VP version: %d\n", s->version);
1804 if (s->version || s->theora)
1807 av_log(s->avctx, AV_LOG_ERROR, "Warning, unsupported keyframe coding type?!\n");
1808 skip_bits(&gb, 2); /* reserved? */
1811 if (s->last_frame.data[0] == s->golden_frame.data[0]) {
1812 if (s->golden_frame.data[0])
1813 avctx->release_buffer(avctx, &s->golden_frame);
1814 s->last_frame= s->golden_frame; /* ensure that we catch any access to this released frame */
1816 if (s->golden_frame.data[0])
1817 avctx->release_buffer(avctx, &s->golden_frame);
1818 if (s->last_frame.data[0])
1819 avctx->release_buffer(avctx, &s->last_frame);
1822 s->golden_frame.reference = 3;
1823 if(avctx->get_buffer(avctx, &s->golden_frame) < 0) {
1824 av_log(s->avctx, AV_LOG_ERROR, "vp3: get_buffer() failed\n");
1828 /* golden frame is also the current frame */
1829 s->current_frame= s->golden_frame;
1831 /* allocate a new current frame */
1832 s->current_frame.reference = 3;
1833 if (!s->golden_frame.data[0]) {
1834 av_log(s->avctx, AV_LOG_ERROR, "vp3: first frame not a keyframe\n");
1837 if(avctx->get_buffer(avctx, &s->current_frame) < 0) {
1838 av_log(s->avctx, AV_LOG_ERROR, "vp3: get_buffer() failed\n");
1843 s->current_frame.qscale_table= s->qscale_table; //FIXME allocate individual tables per AVFrame
1844 s->current_frame.qstride= 0;
1848 if (unpack_superblocks(s, &gb)){
1849 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_superblocks\n");
1852 if (unpack_modes(s, &gb)){
1853 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_modes\n");
1856 if (unpack_vectors(s, &gb)){
1857 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_vectors\n");
1860 if (unpack_block_qpis(s, &gb)){
1861 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_block_qpis\n");
1864 if (unpack_dct_coeffs(s, &gb)){
1865 av_log(s->avctx, AV_LOG_ERROR, "error in unpack_dct_coeffs\n");
1869 for (i = 0; i < 3; i++) {
1870 if (s->flipped_image)
1871 s->data_offset[i] = 0;
1873 s->data_offset[i] = ((s->height>>!!i)-1) * s->current_frame.linesize[i];
1876 s->last_slice_end = 0;
1877 for (i = 0; i < s->macroblock_height; i++)
1880 // filter the last row
1881 for (i = 0; i < 3; i++) {
1882 int row = (s->height >> (3+!!i)) - 1;
1883 apply_loop_filter(s, i, row, row+1);
1885 vp3_draw_horiz_band(s, s->height);
1887 *data_size=sizeof(AVFrame);
1888 *(AVFrame*)data= s->current_frame;
1890 /* release the last frame, if it is allocated and if it is not the
1892 if ((s->last_frame.data[0]) &&
1893 (s->last_frame.data[0] != s->golden_frame.data[0]))
1894 avctx->release_buffer(avctx, &s->last_frame);
1896 /* shuffle frames (last = current) */
1897 s->last_frame= s->current_frame;
1898 s->current_frame.data[0]= NULL; /* ensure that we catch any access to this released frame */
1904 * This is the ffmpeg/libavcodec API module cleanup function.
1906 static av_cold int vp3_decode_end(AVCodecContext *avctx)
1908 Vp3DecodeContext *s = avctx->priv_data;
1911 av_free(s->superblock_coding);
1912 av_free(s->all_fragments);
1913 av_free(s->coeff_counts);
1915 av_free(s->coded_fragment_list);
1916 av_free(s->fast_fragment_list);
1917 av_free(s->superblock_fragments);
1918 av_free(s->macroblock_coding);
1920 for (i = 0; i < 16; i++) {
1921 free_vlc(&s->dc_vlc[i]);
1922 free_vlc(&s->ac_vlc_1[i]);
1923 free_vlc(&s->ac_vlc_2[i]);
1924 free_vlc(&s->ac_vlc_3[i]);
1925 free_vlc(&s->ac_vlc_4[i]);
1928 free_vlc(&s->superblock_run_length_vlc);
1929 free_vlc(&s->fragment_run_length_vlc);
1930 free_vlc(&s->mode_code_vlc);
1931 free_vlc(&s->motion_vector_vlc);
1933 /* release all frames */
1934 if (s->golden_frame.data[0] && s->golden_frame.data[0] != s->last_frame.data[0])
1935 avctx->release_buffer(avctx, &s->golden_frame);
1936 if (s->last_frame.data[0])
1937 avctx->release_buffer(avctx, &s->last_frame);
1938 /* no need to release the current_frame since it will always be pointing
1939 * to the same frame as either the golden or last frame */
1944 static int read_huffman_tree(AVCodecContext *avctx, GetBitContext *gb)
1946 Vp3DecodeContext *s = avctx->priv_data;
1948 if (get_bits1(gb)) {
1950 if (s->entries >= 32) { /* overflow */
1951 av_log(avctx, AV_LOG_ERROR, "huffman tree overflow\n");
1954 token = get_bits(gb, 5);
1955 //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);
1956 s->huffman_table[s->hti][token][0] = s->hbits;
1957 s->huffman_table[s->hti][token][1] = s->huff_code_size;
1961 if (s->huff_code_size >= 32) {/* overflow */
1962 av_log(avctx, AV_LOG_ERROR, "huffman tree overflow\n");
1965 s->huff_code_size++;
1967 if (read_huffman_tree(avctx, gb))
1970 if (read_huffman_tree(avctx, gb))
1973 s->huff_code_size--;
1978 #if CONFIG_THEORA_DECODER
1979 static int theora_decode_header(AVCodecContext *avctx, GetBitContext *gb)
1981 Vp3DecodeContext *s = avctx->priv_data;
1982 int visible_width, visible_height, colorspace;
1984 s->theora = get_bits_long(gb, 24);
1985 av_log(avctx, AV_LOG_DEBUG, "Theora bitstream version %X\n", s->theora);
1987 /* 3.2.0 aka alpha3 has the same frame orientation as original vp3 */
1988 /* but previous versions have the image flipped relative to vp3 */
1989 if (s->theora < 0x030200)
1991 s->flipped_image = 1;
1992 av_log(avctx, AV_LOG_DEBUG, "Old (<alpha3) Theora bitstream, flipped image\n");
1995 visible_width = s->width = get_bits(gb, 16) << 4;
1996 visible_height = s->height = get_bits(gb, 16) << 4;
1998 if(avcodec_check_dimensions(avctx, s->width, s->height)){
1999 av_log(avctx, AV_LOG_ERROR, "Invalid dimensions (%dx%d)\n", s->width, s->height);
2000 s->width= s->height= 0;
2004 if (s->theora >= 0x030200) {
2005 visible_width = get_bits_long(gb, 24);
2006 visible_height = get_bits_long(gb, 24);
2008 skip_bits(gb, 8); /* offset x */
2009 skip_bits(gb, 8); /* offset y */
2012 skip_bits(gb, 32); /* fps numerator */
2013 skip_bits(gb, 32); /* fps denumerator */
2014 skip_bits(gb, 24); /* aspect numerator */
2015 skip_bits(gb, 24); /* aspect denumerator */
2017 if (s->theora < 0x030200)
2018 skip_bits(gb, 5); /* keyframe frequency force */
2019 colorspace = get_bits(gb, 8);
2020 skip_bits(gb, 24); /* bitrate */
2022 skip_bits(gb, 6); /* quality hint */
2024 if (s->theora >= 0x030200)
2026 skip_bits(gb, 5); /* keyframe frequency force */
2027 skip_bits(gb, 2); /* pixel format: 420,res,422,444 */
2028 skip_bits(gb, 3); /* reserved */
2031 // align_get_bits(gb);
2033 if ( visible_width <= s->width && visible_width > s->width-16
2034 && visible_height <= s->height && visible_height > s->height-16)
2035 avcodec_set_dimensions(avctx, visible_width, visible_height);
2037 avcodec_set_dimensions(avctx, s->width, s->height);
2039 if (colorspace == 1) {
2040 avctx->color_primaries = AVCOL_PRI_BT470M;
2041 } else if (colorspace == 2) {
2042 avctx->color_primaries = AVCOL_PRI_BT470BG;
2044 if (colorspace == 1 || colorspace == 2) {
2045 avctx->colorspace = AVCOL_SPC_BT470BG;
2046 avctx->color_trc = AVCOL_TRC_BT709;
2052 static int theora_decode_tables(AVCodecContext *avctx, GetBitContext *gb)
2054 Vp3DecodeContext *s = avctx->priv_data;
2055 int i, n, matrices, inter, plane;
2057 if (s->theora >= 0x030200) {
2058 n = get_bits(gb, 3);
2059 /* loop filter limit values table */
2060 for (i = 0; i < 64; i++) {
2061 s->filter_limit_values[i] = get_bits(gb, n);
2062 if (s->filter_limit_values[i] > 127) {
2063 av_log(avctx, AV_LOG_ERROR, "filter limit value too large (%i > 127), clamping\n", s->filter_limit_values[i]);
2064 s->filter_limit_values[i] = 127;
2069 if (s->theora >= 0x030200)
2070 n = get_bits(gb, 4) + 1;
2073 /* quality threshold table */
2074 for (i = 0; i < 64; i++)
2075 s->coded_ac_scale_factor[i] = get_bits(gb, n);
2077 if (s->theora >= 0x030200)
2078 n = get_bits(gb, 4) + 1;
2081 /* dc scale factor table */
2082 for (i = 0; i < 64; i++)
2083 s->coded_dc_scale_factor[i] = get_bits(gb, n);
2085 if (s->theora >= 0x030200)
2086 matrices = get_bits(gb, 9) + 1;
2091 av_log(avctx, AV_LOG_ERROR, "invalid number of base matrixes\n");
2095 for(n=0; n<matrices; n++){
2096 for (i = 0; i < 64; i++)
2097 s->base_matrix[n][i]= get_bits(gb, 8);
2100 for (inter = 0; inter <= 1; inter++) {
2101 for (plane = 0; plane <= 2; plane++) {
2103 if (inter || plane > 0)
2104 newqr = get_bits1(gb);
2107 if(inter && get_bits1(gb)){
2111 qtj= (3*inter + plane - 1) / 3;
2112 plj= (plane + 2) % 3;
2114 s->qr_count[inter][plane]= s->qr_count[qtj][plj];
2115 memcpy(s->qr_size[inter][plane], s->qr_size[qtj][plj], sizeof(s->qr_size[0][0]));
2116 memcpy(s->qr_base[inter][plane], s->qr_base[qtj][plj], sizeof(s->qr_base[0][0]));
2122 i= get_bits(gb, av_log2(matrices-1)+1);
2124 av_log(avctx, AV_LOG_ERROR, "invalid base matrix index\n");
2127 s->qr_base[inter][plane][qri]= i;
2130 i = get_bits(gb, av_log2(63-qi)+1) + 1;
2131 s->qr_size[inter][plane][qri++]= i;
2136 av_log(avctx, AV_LOG_ERROR, "invalid qi %d > 63\n", qi);
2139 s->qr_count[inter][plane]= qri;
2144 /* Huffman tables */
2145 for (s->hti = 0; s->hti < 80; s->hti++) {
2147 s->huff_code_size = 1;
2148 if (!get_bits1(gb)) {
2150 if(read_huffman_tree(avctx, gb))
2153 if(read_huffman_tree(avctx, gb))
2158 s->theora_tables = 1;
2163 static av_cold int theora_decode_init(AVCodecContext *avctx)
2165 Vp3DecodeContext *s = avctx->priv_data;
2168 uint8_t *header_start[3];
2174 if (!avctx->extradata_size)
2176 av_log(avctx, AV_LOG_ERROR, "Missing extradata!\n");
2180 if (ff_split_xiph_headers(avctx->extradata, avctx->extradata_size,
2181 42, header_start, header_len) < 0) {
2182 av_log(avctx, AV_LOG_ERROR, "Corrupt extradata\n");
2187 init_get_bits(&gb, header_start[i], header_len[i] * 8);
2189 ptype = get_bits(&gb, 8);
2191 if (!(ptype & 0x80))
2193 av_log(avctx, AV_LOG_ERROR, "Invalid extradata!\n");
2197 // FIXME: Check for this as well.
2198 skip_bits_long(&gb, 6*8); /* "theora" */
2203 theora_decode_header(avctx, &gb);
2206 // FIXME: is this needed? it breaks sometimes
2207 // theora_decode_comments(avctx, gb);
2210 if (theora_decode_tables(avctx, &gb))
2214 av_log(avctx, AV_LOG_ERROR, "Unknown Theora config packet: %d\n", ptype&~0x80);
2217 if(ptype != 0x81 && 8*header_len[i] != get_bits_count(&gb))
2218 av_log(avctx, AV_LOG_WARNING, "%d bits left in packet %X\n", 8*header_len[i] - get_bits_count(&gb), ptype);
2219 if (s->theora < 0x030200)
2223 return vp3_decode_init(avctx);
2226 AVCodec theora_decoder = {
2230 sizeof(Vp3DecodeContext),
2235 CODEC_CAP_DR1 | CODEC_CAP_DRAW_HORIZ_BAND,
2237 .long_name = NULL_IF_CONFIG_SMALL("Theora"),
2241 AVCodec vp3_decoder = {
2245 sizeof(Vp3DecodeContext),
2250 CODEC_CAP_DR1 | CODEC_CAP_DRAW_HORIZ_BAND,
2252 .long_name = NULL_IF_CONFIG_SMALL("On2 VP3"),