falkTX
/
FFmpeg
mirror of https://github.com/falkTX/FFmpeg.git
1
0
Fork 0
Code Issues 0 Releases 338 Wiki Activity
You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
35015 Commits
32 Branches
1.0GB
Tree: 90c02ae142
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '90c02ae142'
${ noResults }
FFmpeg/libavcodec/vp8.c

1847 lines
67KB
Raw Blame History 

  1. /**

  2.  * VP8 compatible video decoder

  3.  *

  4.  * Copyright (C) 2010 David Conrad

  5.  * Copyright (C) 2010 Ronald S. Bultje

  6.  * Copyright (C) 2010 Jason Garrett-Glaser

  7.  *

  8.  * This file is part of FFmpeg.

  9.  *

  10.  * FFmpeg is free software; you can redistribute it and/or

  11.  * modify it under the terms of the GNU Lesser General Public

  12.  * License as published by the Free Software Foundation; either

  13.  * version 2.1 of the License, or (at your option) any later version.

  14.  *

  15.  * FFmpeg is distributed in the hope that it will be useful,

  16.  * but WITHOUT ANY WARRANTY; without even the implied warranty of

  17.  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU

  18.  * Lesser General Public License for more details.

  19.  *

  20.  * You should have received a copy of the GNU Lesser General Public

  21.  * License along with FFmpeg; if not, write to the Free Software

  22.  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA

  23.  */

  24. 

  25. #include "libavutil/imgutils.h"

  26. #include "avcodec.h"

  27. #include "vp8.h"

  28. #include "vp8data.h"

  29. #include "rectangle.h"

  30. #include "thread.h"

  31. 

  32. #if ARCH_ARM

  33. #   include "arm/vp8.h"

  34. #endif

  35. 

  36. static void free_buffers(VP8Context *s)

  37. {

  38.     av_freep(&s->macroblocks_base);

  39.     av_freep(&s->filter_strength);

  40.     av_freep(&s->intra4x4_pred_mode_top);

  41.     av_freep(&s->top_nnz);

  42.     av_freep(&s->edge_emu_buffer);

  43.     av_freep(&s->top_border);

  44. 

  45.     s->macroblocks = NULL;

  46. }

  47. 

  48. static int vp8_alloc_frame(VP8Context *s, AVFrame *f)

  49. {

  50.     int ret;

  51.     if ((ret = ff_thread_get_buffer(s->avctx, f)) < 0)

  52.         return ret;

  53.     if (s->num_maps_to_be_freed) {

  54.         assert(!s->maps_are_invalid);

  55.         f->ref_index[0] = s->segmentation_maps[--s->num_maps_to_be_freed];

  56.     } else if (!(f->ref_index[0] = av_mallocz(s->mb_width * s->mb_height))) {

  57.         ff_thread_release_buffer(s->avctx, f);

  58.         return AVERROR(ENOMEM);

  59.     }

  60.     return 0;

  61. }

  62. 

  63. static void vp8_release_frame(VP8Context *s, AVFrame *f, int prefer_delayed_free, int can_direct_free)

  64. {

  65.     if (f->ref_index[0]) {

  66.         if (prefer_delayed_free) {

  67.             /* Upon a size change, we want to free the maps but other threads may still

  68.              * be using them, so queue them. Upon a seek, all threads are inactive so

  69.              * we want to cache one to prevent re-allocation in the next decoding

  70.              * iteration, but the rest we can free directly. */

  71.             int max_queued_maps = can_direct_free ? 1 : FF_ARRAY_ELEMS(s->segmentation_maps);

  72.             if (s->num_maps_to_be_freed < max_queued_maps) {

  73.                 s->segmentation_maps[s->num_maps_to_be_freed++] = f->ref_index[0];

  74.             } else if (can_direct_free) /* vp8_decode_flush(), but our queue is full */ {

  75.                 av_free(f->ref_index[0]);

  76.             } /* else: MEMLEAK (should never happen, but better that than crash) */

  77.             f->ref_index[0] = NULL;

  78.         } else /* vp8_decode_free() */ {

  79.             av_free(f->ref_index[0]);

  80.         }

  81.     }

  82.     ff_thread_release_buffer(s->avctx, f);

  83. }

  84. 

  85. static void vp8_decode_flush_impl(AVCodecContext *avctx,

  86.                                   int prefer_delayed_free, int can_direct_free, int free_mem)

  87. {

  88.     VP8Context *s = avctx->priv_data;

  89.     int i;

  90. 

  91.     if (!avctx->is_copy) {

  92.         for (i = 0; i < 5; i++)

  93.             if (s->frames[i].data[0])

  94.                 vp8_release_frame(s, &s->frames[i], prefer_delayed_free, can_direct_free);

  95.     }

  96.     memset(s->framep, 0, sizeof(s->framep));

  97. 

  98.     if (free_mem) {

  99.         free_buffers(s);

  100.         s->maps_are_invalid = 1;

  101.     }

  102. }

  103. 

  104. static void vp8_decode_flush(AVCodecContext *avctx)

  105. {

  106.     vp8_decode_flush_impl(avctx, 1, 1, 0);

  107. }

  108. 

  109. static int update_dimensions(VP8Context *s, int width, int height)

  110. {

  111.     if (width  != s->avctx->width ||

  112.         height != s->avctx->height) {

  113.         if (av_image_check_size(width, height, 0, s->avctx))

  114.             return AVERROR_INVALIDDATA;

  115. 

  116.         vp8_decode_flush_impl(s->avctx, 1, 0, 1);

  117. 

  118.         avcodec_set_dimensions(s->avctx, width, height);

  119.     }

  120. 

  121.     s->mb_width  = (s->avctx->coded_width +15) / 16;

  122.     s->mb_height = (s->avctx->coded_height+15) / 16;

  123. 

  124.     s->macroblocks_base        = av_mallocz((s->mb_width+s->mb_height*2+1)*sizeof(*s->macroblocks));

  125.     s->filter_strength         = av_mallocz(s->mb_width*sizeof(*s->filter_strength));

  126.     s->intra4x4_pred_mode_top  = av_mallocz(s->mb_width*4);

  127.     s->top_nnz                 = av_mallocz(s->mb_width*sizeof(*s->top_nnz));

  128.     s->top_border              = av_mallocz((s->mb_width+1)*sizeof(*s->top_border));

  129. 

  130.     if (!s->macroblocks_base || !s->filter_strength || !s->intra4x4_pred_mode_top ||

  131.         !s->top_nnz || !s->top_border)

  132.         return AVERROR(ENOMEM);

  133. 

  134.     s->macroblocks        = s->macroblocks_base + 1;

  135. 

  136.     return 0;

  137. }

  138. 

  139. static void parse_segment_info(VP8Context *s)

  140. {

  141.     VP56RangeCoder *c = &s->c;

  142.     int i;

  143. 

  144.     s->segmentation.update_map = vp8_rac_get(c);

  145. 

  146.     if (vp8_rac_get(c)) { // update segment feature data

  147.         s->segmentation.absolute_vals = vp8_rac_get(c);

  148. 

  149.         for (i = 0; i < 4; i++)

  150.             s->segmentation.base_quant[i]   = vp8_rac_get_sint(c, 7);

  151. 

  152.         for (i = 0; i < 4; i++)

  153.             s->segmentation.filter_level[i] = vp8_rac_get_sint(c, 6);

  154.     }

  155.     if (s->segmentation.update_map)

  156.         for (i = 0; i < 3; i++)

  157.             s->prob->segmentid[i] = vp8_rac_get(c) ? vp8_rac_get_uint(c, 8) : 255;

  158. }

  159. 

  160. static void update_lf_deltas(VP8Context *s)

  161. {

  162.     VP56RangeCoder *c = &s->c;

  163.     int i;

  164. 

  165.     for (i = 0; i < 4; i++)

  166.         s->lf_delta.ref[i]  = vp8_rac_get_sint(c, 6);

  167. 

  168.     for (i = MODE_I4x4; i <= VP8_MVMODE_SPLIT; i++)

  169.         s->lf_delta.mode[i] = vp8_rac_get_sint(c, 6);

  170. }

  171. 

  172. static int setup_partitions(VP8Context *s, const uint8_t *buf, int buf_size)

  173. {

  174.     const uint8_t *sizes = buf;

  175.     int i;

  176. 

  177.     s->num_coeff_partitions = 1 << vp8_rac_get_uint(&s->c, 2);

  178. 

  179.     buf      += 3*(s->num_coeff_partitions-1);

  180.     buf_size -= 3*(s->num_coeff_partitions-1);

  181.     if (buf_size < 0)

  182.         return -1;

  183. 

  184.     for (i = 0; i < s->num_coeff_partitions-1; i++) {

  185.         int size = AV_RL24(sizes + 3*i);

  186.         if (buf_size - size < 0)

  187.             return -1;

  188. 

  189.         ff_vp56_init_range_decoder(&s->coeff_partition[i], buf, size);

  190.         buf      += size;

  191.         buf_size -= size;

  192.     }

  193.     ff_vp56_init_range_decoder(&s->coeff_partition[i], buf, buf_size);

  194. 

  195.     return 0;

  196. }

  197. 

  198. static void get_quants(VP8Context *s)

  199. {

  200.     VP56RangeCoder *c = &s->c;

  201.     int i, base_qi;

  202. 

  203.     int yac_qi     = vp8_rac_get_uint(c, 7);

  204.     int ydc_delta  = vp8_rac_get_sint(c, 4);

  205.     int y2dc_delta = vp8_rac_get_sint(c, 4);

  206.     int y2ac_delta = vp8_rac_get_sint(c, 4);

  207.     int uvdc_delta = vp8_rac_get_sint(c, 4);

  208.     int uvac_delta = vp8_rac_get_sint(c, 4);

  209. 

  210.     for (i = 0; i < 4; i++) {

  211.         if (s->segmentation.enabled) {

  212.             base_qi = s->segmentation.base_quant[i];

  213.             if (!s->segmentation.absolute_vals)

  214.                 base_qi += yac_qi;

  215.         } else

  216.             base_qi = yac_qi;

  217. 

  218.         s->qmat[i].luma_qmul[0]    =       vp8_dc_qlookup[av_clip_uintp2(base_qi + ydc_delta , 7)];

  219.         s->qmat[i].luma_qmul[1]    =       vp8_ac_qlookup[av_clip_uintp2(base_qi             , 7)];

  220.         s->qmat[i].luma_dc_qmul[0] =   2 * vp8_dc_qlookup[av_clip_uintp2(base_qi + y2dc_delta, 7)];

  221.         s->qmat[i].luma_dc_qmul[1] = 155 * vp8_ac_qlookup[av_clip_uintp2(base_qi + y2ac_delta, 7)] / 100;

  222.         s->qmat[i].chroma_qmul[0]  =       vp8_dc_qlookup[av_clip_uintp2(base_qi + uvdc_delta, 7)];

  223.         s->qmat[i].chroma_qmul[1]  =       vp8_ac_qlookup[av_clip_uintp2(base_qi + uvac_delta, 7)];

  224. 

  225.         s->qmat[i].luma_dc_qmul[1] = FFMAX(s->qmat[i].luma_dc_qmul[1], 8);

  226.         s->qmat[i].chroma_qmul[0]  = FFMIN(s->qmat[i].chroma_qmul[0], 132);

  227.     }

  228. }

  229. 

  230. /**

  231.  * Determine which buffers golden and altref should be updated with after this frame.

  232.  * The spec isn't clear here, so I'm going by my understanding of what libvpx does

  233.  *

  234.  * Intra frames update all 3 references

  235.  * Inter frames update VP56_FRAME_PREVIOUS if the update_last flag is set

  236.  * If the update (golden|altref) flag is set, it's updated with the current frame

  237.  *      if update_last is set, and VP56_FRAME_PREVIOUS otherwise.

  238.  * If the flag is not set, the number read means:

  239.  *      0: no update

  240.  *      1: VP56_FRAME_PREVIOUS

  241.  *      2: update golden with altref, or update altref with golden

  242.  */

  243. static VP56Frame ref_to_update(VP8Context *s, int update, VP56Frame ref)

  244. {

  245.     VP56RangeCoder *c = &s->c;

  246. 

  247.     if (update)

  248.         return VP56_FRAME_CURRENT;

  249. 

  250.     switch (vp8_rac_get_uint(c, 2)) {

  251.     case 1:

  252.         return VP56_FRAME_PREVIOUS;

  253.     case 2:

  254.         return (ref == VP56_FRAME_GOLDEN) ? VP56_FRAME_GOLDEN2 : VP56_FRAME_GOLDEN;

  255.     }

  256.     return VP56_FRAME_NONE;

  257. }

  258. 

  259. static void update_refs(VP8Context *s)

  260. {

  261.     VP56RangeCoder *c = &s->c;

  262. 

  263.     int update_golden = vp8_rac_get(c);

  264.     int update_altref = vp8_rac_get(c);

  265. 

  266.     s->update_golden = ref_to_update(s, update_golden, VP56_FRAME_GOLDEN);

  267.     s->update_altref = ref_to_update(s, update_altref, VP56_FRAME_GOLDEN2);

  268. }

  269. 

  270. static int decode_frame_header(VP8Context *s, const uint8_t *buf, int buf_size)

  271. {

  272.     VP56RangeCoder *c = &s->c;

  273.     int header_size, hscale, vscale, i, j, k, l, m, ret;

  274.     int width  = s->avctx->width;

  275.     int height = s->avctx->height;

  276. 

  277.     s->keyframe  = !(buf[0] & 1);

  278.     s->profile   =  (buf[0]>>1) & 7;

  279.     s->invisible = !(buf[0] & 0x10);

  280.     header_size  = AV_RL24(buf) >> 5;

  281.     buf      += 3;

  282.     buf_size -= 3;

  283. 

  284.     if (s->profile > 3)

  285.         av_log(s->avctx, AV_LOG_WARNING, "Unknown profile %d\n", s->profile);

  286. 

  287.     if (!s->profile)

  288.         memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_epel_pixels_tab, sizeof(s->put_pixels_tab));

  289.     else    // profile 1-3 use bilinear, 4+ aren't defined so whatever

  290.         memcpy(s->put_pixels_tab, s->vp8dsp.put_vp8_bilinear_pixels_tab, sizeof(s->put_pixels_tab));

  291. 

  292.     if (header_size > buf_size - 7*s->keyframe) {

  293.         av_log(s->avctx, AV_LOG_ERROR, "Header size larger than data provided\n");

  294.         return AVERROR_INVALIDDATA;

  295.     }

  296. 

  297.     if (s->keyframe) {

  298.         if (AV_RL24(buf) != 0x2a019d) {

  299.             av_log(s->avctx, AV_LOG_ERROR, "Invalid start code 0x%x\n", AV_RL24(buf));

  300.             return AVERROR_INVALIDDATA;

  301.         }

  302.         width  = AV_RL16(buf+3) & 0x3fff;

  303.         height = AV_RL16(buf+5) & 0x3fff;

  304.         hscale = buf[4] >> 6;

  305.         vscale = buf[6] >> 6;

  306.         buf      += 7;

  307.         buf_size -= 7;

  308. 

  309.         if (hscale || vscale)

  310.             av_log_missing_feature(s->avctx, "Upscaling", 1);

  311. 

  312.         s->update_golden = s->update_altref = VP56_FRAME_CURRENT;

  313.         for (i = 0; i < 4; i++)

  314.             for (j = 0; j < 16; j++)

  315.                 memcpy(s->prob->token[i][j], vp8_token_default_probs[i][vp8_coeff_band[j]],

  316.                        sizeof(s->prob->token[i][j]));

  317.         memcpy(s->prob->pred16x16, vp8_pred16x16_prob_inter, sizeof(s->prob->pred16x16));

  318.         memcpy(s->prob->pred8x8c , vp8_pred8x8c_prob_inter , sizeof(s->prob->pred8x8c));

  319.         memcpy(s->prob->mvc      , vp8_mv_default_prob     , sizeof(s->prob->mvc));

  320.         memset(&s->segmentation, 0, sizeof(s->segmentation));

  321.     }

  322. 

  323.     if (!s->macroblocks_base || /* first frame */

  324.         width != s->avctx->width || height != s->avctx->height) {

  325.         if ((ret = update_dimensions(s, width, height)) < 0)

  326.             return ret;

  327.     }

  328. 

  329.     ff_vp56_init_range_decoder(c, buf, header_size);

  330.     buf      += header_size;

  331.     buf_size -= header_size;

  332. 

  333.     if (s->keyframe) {

  334.         if (vp8_rac_get(c))

  335.             av_log(s->avctx, AV_LOG_WARNING, "Unspecified colorspace\n");

  336.         vp8_rac_get(c); // whether we can skip clamping in dsp functions

  337.     }

  338. 

  339.     if ((s->segmentation.enabled = vp8_rac_get(c)))

  340.         parse_segment_info(s);

  341.     else

  342.         s->segmentation.update_map = 0; // FIXME: move this to some init function?

  343. 

  344.     s->filter.simple    = vp8_rac_get(c);

  345.     s->filter.level     = vp8_rac_get_uint(c, 6);

  346.     s->filter.sharpness = vp8_rac_get_uint(c, 3);

  347. 

  348.     if ((s->lf_delta.enabled = vp8_rac_get(c)))

  349.         if (vp8_rac_get(c))

  350.             update_lf_deltas(s);

  351. 

  352.     if (setup_partitions(s, buf, buf_size)) {

  353.         av_log(s->avctx, AV_LOG_ERROR, "Invalid partitions\n");

  354.         return AVERROR_INVALIDDATA;

  355.     }

  356. 

  357.     get_quants(s);

  358. 

  359.     if (!s->keyframe) {

  360.         update_refs(s);

  361.         s->sign_bias[VP56_FRAME_GOLDEN]               = vp8_rac_get(c);

  362.         s->sign_bias[VP56_FRAME_GOLDEN2 /* altref */] = vp8_rac_get(c);

  363.     }

  364. 

  365.     // if we aren't saving this frame's probabilities for future frames,

  366.     // make a copy of the current probabilities

  367.     if (!(s->update_probabilities = vp8_rac_get(c)))

  368.         s->prob[1] = s->prob[0];

  369. 

  370.     s->update_last = s->keyframe || vp8_rac_get(c);

  371. 

  372.     for (i = 0; i < 4; i++)

  373.         for (j = 0; j < 8; j++)

  374.             for (k = 0; k < 3; k++)

  375.                 for (l = 0; l < NUM_DCT_TOKENS-1; l++)

  376.                     if (vp56_rac_get_prob_branchy(c, vp8_token_update_probs[i][j][k][l])) {

  377.                         int prob = vp8_rac_get_uint(c, 8);

  378.                         for (m = 0; vp8_coeff_band_indexes[j][m] >= 0; m++)

  379.                             s->prob->token[i][vp8_coeff_band_indexes[j][m]][k][l] = prob;

  380.                     }

  381. 

  382.     if ((s->mbskip_enabled = vp8_rac_get(c)))

  383.         s->prob->mbskip = vp8_rac_get_uint(c, 8);

  384. 

  385.     if (!s->keyframe) {

  386.         s->prob->intra  = vp8_rac_get_uint(c, 8);

  387.         s->prob->last   = vp8_rac_get_uint(c, 8);

  388.         s->prob->golden = vp8_rac_get_uint(c, 8);

  389. 

  390.         if (vp8_rac_get(c))

  391.             for (i = 0; i < 4; i++)

  392.                 s->prob->pred16x16[i] = vp8_rac_get_uint(c, 8);

  393.         if (vp8_rac_get(c))

  394.             for (i = 0; i < 3; i++)

  395.                 s->prob->pred8x8c[i]  = vp8_rac_get_uint(c, 8);

  396. 

  397.         // 17.2 MV probability update

  398.         for (i = 0; i < 2; i++)

  399.             for (j = 0; j < 19; j++)

  400.                 if (vp56_rac_get_prob_branchy(c, vp8_mv_update_prob[i][j]))

  401.                     s->prob->mvc[i][j] = vp8_rac_get_nn(c);

  402.     }

  403. 

  404.     return 0;

  405. }

  406. 

  407. static av_always_inline void clamp_mv(VP8Context *s, VP56mv *dst, const VP56mv *src)

  408. {

  409.     dst->x = av_clip(src->x, s->mv_min.x, s->mv_max.x);

  410.     dst->y = av_clip(src->y, s->mv_min.y, s->mv_max.y);

  411. }

  412. 

  413. /**

  414.  * Motion vector coding, 17.1.

  415.  */

  416. static int read_mv_component(VP56RangeCoder *c, const uint8_t *p)

  417. {

  418.     int bit, x = 0;

  419. 

  420.     if (vp56_rac_get_prob_branchy(c, p[0])) {

  421.         int i;

  422. 

  423.         for (i = 0; i < 3; i++)

  424.             x += vp56_rac_get_prob(c, p[9 + i]) << i;

  425.         for (i = 9; i > 3; i--)

  426.             x += vp56_rac_get_prob(c, p[9 + i]) << i;

  427.         if (!(x & 0xFFF0) || vp56_rac_get_prob(c, p[12]))

  428.             x += 8;

  429.     } else {

  430.         // small_mvtree

  431.         const uint8_t *ps = p+2;

  432.         bit = vp56_rac_get_prob(c, *ps);

  433.         ps += 1 + 3*bit;

  434.         x  += 4*bit;

  435.         bit = vp56_rac_get_prob(c, *ps);

  436.         ps += 1 + bit;

  437.         x  += 2*bit;

  438.         x  += vp56_rac_get_prob(c, *ps);

  439.     }

  440. 

  441.     return (x && vp56_rac_get_prob(c, p[1])) ? -x : x;

  442. }

  443. 

  444. static av_always_inline

  445. const uint8_t *get_submv_prob(uint32_t left, uint32_t top)

  446. {

  447.     if (left == top)

  448.         return vp8_submv_prob[4-!!left];

  449.     if (!top)

  450.         return vp8_submv_prob[2];

  451.     return vp8_submv_prob[1-!!left];

  452. }

  453. 

  454. /**

  455.  * Split motion vector prediction, 16.4.

  456.  * @returns the number of motion vectors parsed (2, 4 or 16)

  457.  */

  458. static av_always_inline

  459. int decode_splitmvs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb)

  460. {

  461.     int part_idx;

  462.     int n, num;

  463.     VP8Macroblock *top_mb  = &mb[2];

  464.     VP8Macroblock *left_mb = &mb[-1];

  465.     const uint8_t *mbsplits_left = vp8_mbsplits[left_mb->partitioning],

  466.                   *mbsplits_top = vp8_mbsplits[top_mb->partitioning],

  467.                   *mbsplits_cur, *firstidx;

  468.     VP56mv *top_mv  = top_mb->bmv;

  469.     VP56mv *left_mv = left_mb->bmv;

  470.     VP56mv *cur_mv  = mb->bmv;

  471. 

  472.     if (vp56_rac_get_prob_branchy(c, vp8_mbsplit_prob[0])) {

  473.         if (vp56_rac_get_prob_branchy(c, vp8_mbsplit_prob[1])) {

  474.             part_idx = VP8_SPLITMVMODE_16x8 + vp56_rac_get_prob(c, vp8_mbsplit_prob[2]);

  475.         } else {

  476.             part_idx = VP8_SPLITMVMODE_8x8;

  477.         }

  478.     } else {

  479.         part_idx = VP8_SPLITMVMODE_4x4;

  480.     }

  481. 

  482.     num = vp8_mbsplit_count[part_idx];

  483.     mbsplits_cur = vp8_mbsplits[part_idx],

  484.     firstidx = vp8_mbfirstidx[part_idx];

  485.     mb->partitioning = part_idx;

  486. 

  487.     for (n = 0; n < num; n++) {

  488.         int k = firstidx[n];

  489.         uint32_t left, above;

  490.         const uint8_t *submv_prob;

  491. 

  492.         if (!(k & 3))

  493.             left = AV_RN32A(&left_mv[mbsplits_left[k + 3]]);

  494.         else

  495.             left  = AV_RN32A(&cur_mv[mbsplits_cur[k - 1]]);

  496.         if (k <= 3)

  497.             above = AV_RN32A(&top_mv[mbsplits_top[k + 12]]);

  498.         else

  499.             above = AV_RN32A(&cur_mv[mbsplits_cur[k - 4]]);

  500. 

  501.         submv_prob = get_submv_prob(left, above);

  502. 

  503.         if (vp56_rac_get_prob_branchy(c, submv_prob[0])) {

  504.             if (vp56_rac_get_prob_branchy(c, submv_prob[1])) {

  505.                 if (vp56_rac_get_prob_branchy(c, submv_prob[2])) {

  506.                     mb->bmv[n].y = mb->mv.y + read_mv_component(c, s->prob->mvc[0]);

  507.                     mb->bmv[n].x = mb->mv.x + read_mv_component(c, s->prob->mvc[1]);

  508.                 } else {

  509.                     AV_ZERO32(&mb->bmv[n]);

  510.                 }

  511.             } else {

  512.                 AV_WN32A(&mb->bmv[n], above);

  513.             }

  514.         } else {

  515.             AV_WN32A(&mb->bmv[n], left);

  516.         }

  517.     }

  518. 

  519.     return num;

  520. }

  521. 

  522. static av_always_inline

  523. void decode_mvs(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y)

  524. {

  525.     VP8Macroblock *mb_edge[3] = { mb + 2 /* top */,

  526.                                   mb - 1 /* left */,

  527.                                   mb + 1 /* top-left */ };

  528.     enum { CNT_ZERO, CNT_NEAREST, CNT_NEAR, CNT_SPLITMV };

  529.     enum { VP8_EDGE_TOP, VP8_EDGE_LEFT, VP8_EDGE_TOPLEFT };

  530.     int idx = CNT_ZERO;

  531.     int cur_sign_bias = s->sign_bias[mb->ref_frame];

  532.     int8_t *sign_bias = s->sign_bias;

  533.     VP56mv near_mv[4];

  534.     uint8_t cnt[4] = { 0 };

  535.     VP56RangeCoder *c = &s->c;

  536. 

  537.     AV_ZERO32(&near_mv[0]);

  538.     AV_ZERO32(&near_mv[1]);

  539.     AV_ZERO32(&near_mv[2]);

  540. 

  541.     /* Process MB on top, left and top-left */

  542.     #define MV_EDGE_CHECK(n)\

  543.     {\

  544.         VP8Macroblock *edge = mb_edge[n];\

  545.         int edge_ref = edge->ref_frame;\

  546.         if (edge_ref != VP56_FRAME_CURRENT) {\

  547.             uint32_t mv = AV_RN32A(&edge->mv);\

  548.             if (mv) {\

  549.                 if (cur_sign_bias != sign_bias[edge_ref]) {\

  550.                     /* SWAR negate of the values in mv. */\

  551.                     mv = ~mv;\

  552.                     mv = ((mv&0x7fff7fff) + 0x00010001) ^ (mv&0x80008000);\

  553.                 }\

  554.                 if (!n || mv != AV_RN32A(&near_mv[idx]))\

  555.                     AV_WN32A(&near_mv[++idx], mv);\

  556.                 cnt[idx]      += 1 + (n != 2);\

  557.             } else\

  558.                 cnt[CNT_ZERO] += 1 + (n != 2);\

  559.         }\

  560.     }

  561. 

  562.     MV_EDGE_CHECK(0)

  563.     MV_EDGE_CHECK(1)

  564.     MV_EDGE_CHECK(2)

  565. 

  566.     mb->partitioning = VP8_SPLITMVMODE_NONE;

  567.     if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[CNT_ZERO]][0])) {

  568.         mb->mode = VP8_MVMODE_MV;

  569. 

  570.         /* If we have three distinct MVs, merge first and last if they're the same */

  571.         if (cnt[CNT_SPLITMV] && AV_RN32A(&near_mv[1 + VP8_EDGE_TOP]) == AV_RN32A(&near_mv[1 + VP8_EDGE_TOPLEFT]))

  572.             cnt[CNT_NEAREST] += 1;

  573. 

  574.         /* Swap near and nearest if necessary */

  575.         if (cnt[CNT_NEAR] > cnt[CNT_NEAREST]) {

  576.             FFSWAP(uint8_t,     cnt[CNT_NEAREST],     cnt[CNT_NEAR]);

  577.             FFSWAP( VP56mv, near_mv[CNT_NEAREST], near_mv[CNT_NEAR]);

  578.         }

  579. 

  580.         if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[CNT_NEAREST]][1])) {

  581.             if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[CNT_NEAR]][2])) {

  582. 

  583.                 /* Choose the best mv out of 0,0 and the nearest mv */

  584.                 clamp_mv(s, &mb->mv, &near_mv[CNT_ZERO + (cnt[CNT_NEAREST] >= cnt[CNT_ZERO])]);

  585.                 cnt[CNT_SPLITMV] = ((mb_edge[VP8_EDGE_LEFT]->mode    == VP8_MVMODE_SPLIT) +

  586.                                     (mb_edge[VP8_EDGE_TOP]->mode     == VP8_MVMODE_SPLIT)) * 2 +

  587.                                     (mb_edge[VP8_EDGE_TOPLEFT]->mode == VP8_MVMODE_SPLIT);

  588. 

  589.                 if (vp56_rac_get_prob_branchy(c, vp8_mode_contexts[cnt[CNT_SPLITMV]][3])) {

  590.                     mb->mode = VP8_MVMODE_SPLIT;

  591.                     mb->mv = mb->bmv[decode_splitmvs(s, c, mb) - 1];

  592.                 } else {

  593.                     mb->mv.y += read_mv_component(c, s->prob->mvc[0]);

  594.                     mb->mv.x += read_mv_component(c, s->prob->mvc[1]);

  595.                     mb->bmv[0] = mb->mv;

  596.                 }

  597.             } else {

  598.                 clamp_mv(s, &mb->mv, &near_mv[CNT_NEAR]);

  599.                 mb->bmv[0] = mb->mv;

  600.             }

  601.         } else {

  602.             clamp_mv(s, &mb->mv, &near_mv[CNT_NEAREST]);

  603.             mb->bmv[0] = mb->mv;

  604.         }

  605.     } else {

  606.         mb->mode = VP8_MVMODE_ZERO;

  607.         AV_ZERO32(&mb->mv);

  608.         mb->bmv[0] = mb->mv;

  609.     }

  610. }

  611. 

  612. static av_always_inline

  613. void decode_intra4x4_modes(VP8Context *s, VP56RangeCoder *c,

  614.                            int mb_x, int keyframe)

  615. {

  616.     uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;

  617.     if (keyframe) {

  618.         int x, y;

  619.         uint8_t* const top = s->intra4x4_pred_mode_top + 4 * mb_x;

  620.         uint8_t* const left = s->intra4x4_pred_mode_left;

  621.         for (y = 0; y < 4; y++) {

  622.             for (x = 0; x < 4; x++) {

  623.                 const uint8_t *ctx;

  624.                 ctx = vp8_pred4x4_prob_intra[top[x]][left[y]];

  625.                 *intra4x4 = vp8_rac_get_tree(c, vp8_pred4x4_tree, ctx);

  626.                 left[y] = top[x] = *intra4x4;

  627.                 intra4x4++;

  628.             }

  629.         }

  630.     } else {

  631.         int i;

  632.         for (i = 0; i < 16; i++)

  633.             intra4x4[i] = vp8_rac_get_tree(c, vp8_pred4x4_tree, vp8_pred4x4_prob_inter);

  634.     }

  635. }

  636. 

  637. static av_always_inline

  638. void decode_mb_mode(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, uint8_t *segment, uint8_t *ref)

  639. {

  640.     VP56RangeCoder *c = &s->c;

  641. 

  642.     if (s->segmentation.update_map) {

  643.         int bit  = vp56_rac_get_prob(c, s->prob->segmentid[0]);

  644.         *segment = vp56_rac_get_prob(c, s->prob->segmentid[1+bit]) + 2*bit;

  645.     } else

  646.         *segment = ref ? *ref : *segment;

  647.     s->segment = *segment;

  648. 

  649.     mb->skip = s->mbskip_enabled ? vp56_rac_get_prob(c, s->prob->mbskip) : 0;

  650. 

  651.     if (s->keyframe) {

  652.         mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_intra, vp8_pred16x16_prob_intra);

  653. 

  654.         if (mb->mode == MODE_I4x4) {

  655.             decode_intra4x4_modes(s, c, mb_x, 1);

  656.         } else {

  657.             const uint32_t modes = vp8_pred4x4_mode[mb->mode] * 0x01010101u;

  658.             AV_WN32A(s->intra4x4_pred_mode_top + 4 * mb_x, modes);

  659.             AV_WN32A(s->intra4x4_pred_mode_left, modes);

  660.         }

  661. 

  662.         s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, vp8_pred8x8c_prob_intra);

  663.         mb->ref_frame = VP56_FRAME_CURRENT;

  664.     } else if (vp56_rac_get_prob_branchy(c, s->prob->intra)) {

  665.         // inter MB, 16.2

  666.         if (vp56_rac_get_prob_branchy(c, s->prob->last))

  667.             mb->ref_frame = vp56_rac_get_prob(c, s->prob->golden) ?

  668.                 VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;

  669.         else

  670.             mb->ref_frame = VP56_FRAME_PREVIOUS;

  671.         s->ref_count[mb->ref_frame-1]++;

  672. 

  673.         // motion vectors, 16.3

  674.         decode_mvs(s, mb, mb_x, mb_y);

  675.     } else {

  676.         // intra MB, 16.1

  677.         mb->mode = vp8_rac_get_tree(c, vp8_pred16x16_tree_inter, s->prob->pred16x16);

  678. 

  679.         if (mb->mode == MODE_I4x4)

  680.             decode_intra4x4_modes(s, c, mb_x, 0);

  681. 

  682.         s->chroma_pred_mode = vp8_rac_get_tree(c, vp8_pred8x8c_tree, s->prob->pred8x8c);

  683.         mb->ref_frame = VP56_FRAME_CURRENT;

  684.         mb->partitioning = VP8_SPLITMVMODE_NONE;

  685.         AV_ZERO32(&mb->bmv[0]);

  686.     }

  687. }

  688. 

  689. #ifndef decode_block_coeffs_internal

  690. /**

  691.  * @param c arithmetic bitstream reader context

  692.  * @param block destination for block coefficients

  693.  * @param probs probabilities to use when reading trees from the bitstream

  694.  * @param i initial coeff index, 0 unless a separate DC block is coded

  695.  * @param qmul array holding the dc/ac dequant factor at position 0/1

  696.  * @return 0 if no coeffs were decoded

  697.  *         otherwise, the index of the last coeff decoded plus one

  698.  */

  699. static int decode_block_coeffs_internal(VP56RangeCoder *c, DCTELEM block[16],

  700.                                         uint8_t probs[16][3][NUM_DCT_TOKENS-1],

  701.                                         int i, uint8_t *token_prob, int16_t qmul[2])

  702. {

  703.     goto skip_eob;

  704.     do {

  705.         int coeff;

  706.         if (!vp56_rac_get_prob_branchy(c, token_prob[0]))   // DCT_EOB

  707.             return i;

  708. 

  709. skip_eob:

  710.         if (!vp56_rac_get_prob_branchy(c, token_prob[1])) { // DCT_0

  711.             if (++i == 16)

  712.                 return i; // invalid input; blocks should end with EOB

  713.             token_prob = probs[i][0];

  714.             goto skip_eob;

  715.         }

  716. 

  717.         if (!vp56_rac_get_prob_branchy(c, token_prob[2])) { // DCT_1

  718.             coeff = 1;

  719.             token_prob = probs[i+1][1];

  720.         } else {

  721.             if (!vp56_rac_get_prob_branchy(c, token_prob[3])) { // DCT 2,3,4

  722.                 coeff = vp56_rac_get_prob_branchy(c, token_prob[4]);

  723.                 if (coeff)

  724.                     coeff += vp56_rac_get_prob(c, token_prob[5]);

  725.                 coeff += 2;

  726.             } else {

  727.                 // DCT_CAT*

  728.                 if (!vp56_rac_get_prob_branchy(c, token_prob[6])) {

  729.                     if (!vp56_rac_get_prob_branchy(c, token_prob[7])) { // DCT_CAT1

  730.                         coeff  = 5 + vp56_rac_get_prob(c, vp8_dct_cat1_prob[0]);

  731.                     } else {                                    // DCT_CAT2

  732.                         coeff  = 7;

  733.                         coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[0]) << 1;

  734.                         coeff += vp56_rac_get_prob(c, vp8_dct_cat2_prob[1]);

  735.                     }

  736.                 } else {    // DCT_CAT3 and up

  737.                     int a = vp56_rac_get_prob(c, token_prob[8]);

  738.                     int b = vp56_rac_get_prob(c, token_prob[9+a]);

  739.                     int cat = (a<<1) + b;

  740.                     coeff  = 3 + (8<<cat);

  741.                     coeff += vp8_rac_get_coeff(c, ff_vp8_dct_cat_prob[cat]);

  742.                 }

  743.             }

  744.             token_prob = probs[i+1][2];

  745.         }

  746.         block[zigzag_scan[i]] = (vp8_rac_get(c) ? -coeff : coeff) * qmul[!!i];

  747.     } while (++i < 16);

  748. 

  749.     return i;

  750. }

  751. #endif

  752. 

  753. /**

  754.  * @param c arithmetic bitstream reader context

  755.  * @param block destination for block coefficients

  756.  * @param probs probabilities to use when reading trees from the bitstream

  757.  * @param i initial coeff index, 0 unless a separate DC block is coded

  758.  * @param zero_nhood the initial prediction context for number of surrounding

  759.  *                   all-zero blocks (only left/top, so 0-2)

  760.  * @param qmul array holding the dc/ac dequant factor at position 0/1

  761.  * @return 0 if no coeffs were decoded

  762.  *         otherwise, the index of the last coeff decoded plus one

  763.  */

  764. static av_always_inline

  765. int decode_block_coeffs(VP56RangeCoder *c, DCTELEM block[16],

  766.                         uint8_t probs[16][3][NUM_DCT_TOKENS-1],

  767.                         int i, int zero_nhood, int16_t qmul[2])

  768. {

  769.     uint8_t *token_prob = probs[i][zero_nhood];

  770.     if (!vp56_rac_get_prob_branchy(c, token_prob[0]))   // DCT_EOB

  771.         return 0;

  772.     return decode_block_coeffs_internal(c, block, probs, i, token_prob, qmul);

  773. }

  774. 

  775. static av_always_inline

  776. void decode_mb_coeffs(VP8Context *s, VP56RangeCoder *c, VP8Macroblock *mb,

  777.                       uint8_t t_nnz[9], uint8_t l_nnz[9])

  778. {

  779.     int i, x, y, luma_start = 0, luma_ctx = 3;

  780.     int nnz_pred, nnz, nnz_total = 0;

  781.     int segment = s->segment;

  782.     int block_dc = 0;

  783. 

  784.     if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {

  785.         nnz_pred = t_nnz[8] + l_nnz[8];

  786. 

  787.         // decode DC values and do hadamard

  788.         nnz = decode_block_coeffs(c, s->block_dc, s->prob->token[1], 0, nnz_pred,

  789.                                   s->qmat[segment].luma_dc_qmul);

  790.         l_nnz[8] = t_nnz[8] = !!nnz;

  791.         if (nnz) {

  792.             nnz_total += nnz;

  793.             block_dc = 1;

  794.             if (nnz == 1)

  795.                 s->vp8dsp.vp8_luma_dc_wht_dc(s->block, s->block_dc);

  796.             else

  797.                 s->vp8dsp.vp8_luma_dc_wht(s->block, s->block_dc);

  798.         }

  799.         luma_start = 1;

  800.         luma_ctx = 0;

  801.     }

  802. 

  803.     // luma blocks

  804.     for (y = 0; y < 4; y++)

  805.         for (x = 0; x < 4; x++) {

  806.             nnz_pred = l_nnz[y] + t_nnz[x];

  807.             nnz = decode_block_coeffs(c, s->block[y][x], s->prob->token[luma_ctx], luma_start,

  808.                                       nnz_pred, s->qmat[segment].luma_qmul);

  809.             // nnz+block_dc may be one more than the actual last index, but we don't care

  810.             s->non_zero_count_cache[y][x] = nnz + block_dc;

  811.             t_nnz[x] = l_nnz[y] = !!nnz;

  812.             nnz_total += nnz;

  813.         }

  814. 

  815.     // chroma blocks

  816.     // TODO: what to do about dimensions? 2nd dim for luma is x,

  817.     // but for chroma it's (y<<1)|x

  818.     for (i = 4; i < 6; i++)

  819.         for (y = 0; y < 2; y++)

  820.             for (x = 0; x < 2; x++) {

  821.                 nnz_pred = l_nnz[i+2*y] + t_nnz[i+2*x];

  822.                 nnz = decode_block_coeffs(c, s->block[i][(y<<1)+x], s->prob->token[2], 0,

  823.                                           nnz_pred, s->qmat[segment].chroma_qmul);

  824.                 s->non_zero_count_cache[i][(y<<1)+x] = nnz;

  825.                 t_nnz[i+2*x] = l_nnz[i+2*y] = !!nnz;

  826.                 nnz_total += nnz;

  827.             }

  828. 

  829.     // if there were no coded coeffs despite the macroblock not being marked skip,

  830.     // we MUST not do the inner loop filter and should not do IDCT

  831.     // Since skip isn't used for bitstream prediction, just manually set it.

  832.     if (!nnz_total)

  833.         mb->skip = 1;

  834. }

  835. 

  836. static av_always_inline

  837. void backup_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,

  838.                       int linesize, int uvlinesize, int simple)

  839. {

  840.     AV_COPY128(top_border, src_y + 15*linesize);

  841.     if (!simple) {

  842.         AV_COPY64(top_border+16, src_cb + 7*uvlinesize);

  843.         AV_COPY64(top_border+24, src_cr + 7*uvlinesize);

  844.     }

  845. }

  846. 

  847. static av_always_inline

  848. void xchg_mb_border(uint8_t *top_border, uint8_t *src_y, uint8_t *src_cb, uint8_t *src_cr,

  849.                     int linesize, int uvlinesize, int mb_x, int mb_y, int mb_width,

  850.                     int simple, int xchg)

  851. {

  852.     uint8_t *top_border_m1 = top_border-32;     // for TL prediction

  853.     src_y  -=   linesize;

  854.     src_cb -= uvlinesize;

  855.     src_cr -= uvlinesize;

  856. 

  857. #define XCHG(a,b,xchg) do {                     \

  858.         if (xchg) AV_SWAP64(b,a);               \

  859.         else      AV_COPY64(b,a);               \

  860.     } while (0)

  861. 

  862.     XCHG(top_border_m1+8, src_y-8, xchg);

  863.     XCHG(top_border,      src_y,   xchg);

  864.     XCHG(top_border+8,    src_y+8, 1);

  865.     if (mb_x < mb_width-1)

  866.         XCHG(top_border+32, src_y+16, 1);

  867. 

  868.     // only copy chroma for normal loop filter

  869.     // or to initialize the top row to 127

  870.     if (!simple || !mb_y) {

  871.         XCHG(top_border_m1+16, src_cb-8, xchg);

  872.         XCHG(top_border_m1+24, src_cr-8, xchg);

  873.         XCHG(top_border+16,    src_cb, 1);

  874.         XCHG(top_border+24,    src_cr, 1);

  875.     }

  876. }

  877. 

  878. static av_always_inline

  879. int check_dc_pred8x8_mode(int mode, int mb_x, int mb_y)

  880. {

  881.     if (!mb_x) {

  882.         return mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;

  883.     } else {

  884.         return mb_y ? mode : LEFT_DC_PRED8x8;

  885.     }

  886. }

  887. 

  888. static av_always_inline

  889. int check_tm_pred8x8_mode(int mode, int mb_x, int mb_y)

  890. {

  891.     if (!mb_x) {

  892.         return mb_y ? VERT_PRED8x8 : DC_129_PRED8x8;

  893.     } else {

  894.         return mb_y ? mode : HOR_PRED8x8;

  895.     }

  896. }

  897. 

  898. static av_always_inline

  899. int check_intra_pred8x8_mode(int mode, int mb_x, int mb_y)

  900. {

  901.     if (mode == DC_PRED8x8) {

  902.         return check_dc_pred8x8_mode(mode, mb_x, mb_y);

  903.     } else {

  904.         return mode;

  905.     }

  906. }

  907. 

  908. static av_always_inline

  909. int check_intra_pred8x8_mode_emuedge(int mode, int mb_x, int mb_y)

  910. {

  911.     switch (mode) {

  912.     case DC_PRED8x8:

  913.         return check_dc_pred8x8_mode(mode, mb_x, mb_y);

  914.     case VERT_PRED8x8:

  915.         return !mb_y ? DC_127_PRED8x8 : mode;

  916.     case HOR_PRED8x8:

  917.         return !mb_x ? DC_129_PRED8x8 : mode;

  918.     case PLANE_PRED8x8 /*TM*/:

  919.         return check_tm_pred8x8_mode(mode, mb_x, mb_y);

  920.     }

  921.     return mode;

  922. }

  923. 

  924. static av_always_inline

  925. int check_tm_pred4x4_mode(int mode, int mb_x, int mb_y)

  926. {

  927.     if (!mb_x) {

  928.         return mb_y ? VERT_VP8_PRED : DC_129_PRED;

  929.     } else {

  930.         return mb_y ? mode : HOR_VP8_PRED;

  931.     }

  932. }

  933. 

  934. static av_always_inline

  935. int check_intra_pred4x4_mode_emuedge(int mode, int mb_x, int mb_y, int *copy_buf)

  936. {

  937.     switch (mode) {

  938.     case VERT_PRED:

  939.         if (!mb_x && mb_y) {

  940.             *copy_buf = 1;

  941.             return mode;

  942.         }

  943.         /* fall-through */

  944.     case DIAG_DOWN_LEFT_PRED:

  945.     case VERT_LEFT_PRED:

  946.         return !mb_y ? DC_127_PRED : mode;

  947.     case HOR_PRED:

  948.         if (!mb_y) {

  949.             *copy_buf = 1;

  950.             return mode;

  951.         }

  952.         /* fall-through */

  953.     case HOR_UP_PRED:

  954.         return !mb_x ? DC_129_PRED : mode;

  955.     case TM_VP8_PRED:

  956.         return check_tm_pred4x4_mode(mode, mb_x, mb_y);

  957.     case DC_PRED: // 4x4 DC doesn't use the same "H.264-style" exceptions as 16x16/8x8 DC

  958.     case DIAG_DOWN_RIGHT_PRED:

  959.     case VERT_RIGHT_PRED:

  960.     case HOR_DOWN_PRED:

  961.         if (!mb_y || !mb_x)

  962.             *copy_buf = 1;

  963.         return mode;

  964.     }

  965.     return mode;

  966. }

  967. 

  968. static av_always_inline

  969. void intra_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,

  970.                    int mb_x, int mb_y)

  971. {

  972.     AVCodecContext *avctx = s->avctx;

  973.     int x, y, mode, nnz;

  974.     uint32_t tr;

  975. 

  976.     // for the first row, we need to run xchg_mb_border to init the top edge to 127

  977.     // otherwise, skip it if we aren't going to deblock

  978.     if (!(avctx->flags & CODEC_FLAG_EMU_EDGE && !mb_y) && (s->deblock_filter || !mb_y))

  979.         xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],

  980.                        s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,

  981.                        s->filter.simple, 1);

  982. 

  983.     if (mb->mode < MODE_I4x4) {

  984.         if (avctx->flags & CODEC_FLAG_EMU_EDGE) { // tested

  985.             mode = check_intra_pred8x8_mode_emuedge(mb->mode, mb_x, mb_y);

  986.         } else {

  987.             mode = check_intra_pred8x8_mode(mb->mode, mb_x, mb_y);

  988.         }

  989.         s->hpc.pred16x16[mode](dst[0], s->linesize);

  990.     } else {

  991.         uint8_t *ptr = dst[0];

  992.         uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;

  993.         uint8_t tr_top[4] = { 127, 127, 127, 127 };

  994. 

  995.         // all blocks on the right edge of the macroblock use bottom edge

  996.         // the top macroblock for their topright edge

  997.         uint8_t *tr_right = ptr - s->linesize + 16;

  998. 

  999.         // if we're on the right edge of the frame, said edge is extended

  1000.         // from the top macroblock

  1001.         if (!(!mb_y && avctx->flags & CODEC_FLAG_EMU_EDGE) &&

  1002.             mb_x == s->mb_width-1) {

  1003.             tr = tr_right[-1]*0x01010101u;

  1004.             tr_right = (uint8_t *)&tr;

  1005.         }

  1006. 

  1007.         if (mb->skip)

  1008.             AV_ZERO128(s->non_zero_count_cache);

  1009. 

  1010.         for (y = 0; y < 4; y++) {

  1011.             uint8_t *topright = ptr + 4 - s->linesize;

  1012.             for (x = 0; x < 4; x++) {

  1013.                 int copy = 0, linesize = s->linesize;

  1014.                 uint8_t *dst = ptr+4*x;

  1015.                 DECLARE_ALIGNED(4, uint8_t, copy_dst)[5*8];

  1016. 

  1017.                 if ((y == 0 || x == 3) && mb_y == 0 && avctx->flags & CODEC_FLAG_EMU_EDGE) {

  1018.                     topright = tr_top;

  1019.                 } else if (x == 3)

  1020.                     topright = tr_right;

  1021. 

  1022.                 if (avctx->flags & CODEC_FLAG_EMU_EDGE) { // mb_x+x or mb_y+y is a hack but works

  1023.                     mode = check_intra_pred4x4_mode_emuedge(intra4x4[x], mb_x + x, mb_y + y, &copy);

  1024.                     if (copy) {

  1025.                         dst = copy_dst + 12;

  1026.                         linesize = 8;

  1027.                         if (!(mb_y + y)) {

  1028.                             copy_dst[3] = 127U;

  1029.                             AV_WN32A(copy_dst+4, 127U * 0x01010101U);

  1030.                         } else {

  1031.                             AV_COPY32(copy_dst+4, ptr+4*x-s->linesize);

  1032.                             if (!(mb_x + x)) {

  1033.                                 copy_dst[3] = 129U;

  1034.                             } else {

  1035.                                 copy_dst[3] = ptr[4*x-s->linesize-1];

  1036.                             }

  1037.                         }

  1038.                         if (!(mb_x + x)) {

  1039.                             copy_dst[11] =

  1040.                             copy_dst[19] =

  1041.                             copy_dst[27] =

  1042.                             copy_dst[35] = 129U;

  1043.                         } else {

  1044.                             copy_dst[11] = ptr[4*x              -1];

  1045.                             copy_dst[19] = ptr[4*x+s->linesize  -1];

  1046.                             copy_dst[27] = ptr[4*x+s->linesize*2-1];

  1047.                             copy_dst[35] = ptr[4*x+s->linesize*3-1];

  1048.                         }

  1049.                     }

  1050.                 } else {

  1051.                     mode = intra4x4[x];

  1052.                 }

  1053.                 s->hpc.pred4x4[mode](dst, topright, linesize);

  1054.                 if (copy) {

  1055.                     AV_COPY32(ptr+4*x              , copy_dst+12);

  1056.                     AV_COPY32(ptr+4*x+s->linesize  , copy_dst+20);

  1057.                     AV_COPY32(ptr+4*x+s->linesize*2, copy_dst+28);

  1058.                     AV_COPY32(ptr+4*x+s->linesize*3, copy_dst+36);

  1059.                 }

  1060. 

  1061.                 nnz = s->non_zero_count_cache[y][x];

  1062.                 if (nnz) {

  1063.                     if (nnz == 1)

  1064.                         s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);

  1065.                     else

  1066.                         s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);

  1067.                 }

  1068.                 topright += 4;

  1069.             }

  1070. 

  1071.             ptr   += 4*s->linesize;

  1072.             intra4x4 += 4;

  1073.         }

  1074.     }

  1075. 

  1076.     if (avctx->flags & CODEC_FLAG_EMU_EDGE) {

  1077.         mode = check_intra_pred8x8_mode_emuedge(s->chroma_pred_mode, mb_x, mb_y);

  1078.     } else {

  1079.         mode = check_intra_pred8x8_mode(s->chroma_pred_mode, mb_x, mb_y);

  1080.     }

  1081.     s->hpc.pred8x8[mode](dst[1], s->uvlinesize);

  1082.     s->hpc.pred8x8[mode](dst[2], s->uvlinesize);

  1083. 

  1084.     if (!(avctx->flags & CODEC_FLAG_EMU_EDGE && !mb_y) && (s->deblock_filter || !mb_y))

  1085.         xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],

  1086.                        s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,

  1087.                        s->filter.simple, 0);

  1088. }

  1089. 

  1090. static const uint8_t subpel_idx[3][8] = {

  1091.     { 0, 1, 2, 1, 2, 1, 2, 1 }, // nr. of left extra pixels,

  1092.                                 // also function pointer index

  1093.     { 0, 3, 5, 3, 5, 3, 5, 3 }, // nr. of extra pixels required

  1094.     { 0, 2, 3, 2, 3, 2, 3, 2 }, // nr. of right extra pixels

  1095. };

  1096. 

  1097. /**

  1098.  * luma MC function

  1099.  *

  1100.  * @param s VP8 decoding context

  1101.  * @param dst target buffer for block data at block position

  1102.  * @param ref reference picture buffer at origin (0, 0)

  1103.  * @param mv motion vector (relative to block position) to get pixel data from

  1104.  * @param x_off horizontal position of block from origin (0, 0)

  1105.  * @param y_off vertical position of block from origin (0, 0)

  1106.  * @param block_w width of block (16, 8 or 4)

  1107.  * @param block_h height of block (always same as block_w)

  1108.  * @param width width of src/dst plane data

  1109.  * @param height height of src/dst plane data

  1110.  * @param linesize size of a single line of plane data, including padding

  1111.  * @param mc_func motion compensation function pointers (bilinear or sixtap MC)

  1112.  */

  1113. static av_always_inline

  1114. void vp8_mc_luma(VP8Context *s, uint8_t *dst, AVFrame *ref, const VP56mv *mv,

  1115.                  int x_off, int y_off, int block_w, int block_h,

  1116.                  int width, int height, int linesize,

  1117.                  vp8_mc_func mc_func[3][3])

  1118. {

  1119.     uint8_t *src = ref->data[0];

  1120. 

  1121.     if (AV_RN32A(mv)) {

  1122. 

  1123.         int mx = (mv->x << 1)&7, mx_idx = subpel_idx[0][mx];

  1124.         int my = (mv->y << 1)&7, my_idx = subpel_idx[0][my];

  1125. 

  1126.         x_off += mv->x >> 2;

  1127.         y_off += mv->y >> 2;

  1128. 

  1129.         // edge emulation

  1130.         ff_thread_await_progress(ref, (3 + y_off + block_h + subpel_idx[2][my]) >> 4, 0);

  1131.         src += y_off * linesize + x_off;

  1132.         if (x_off < mx_idx || x_off >= width  - block_w - subpel_idx[2][mx] ||

  1133.             y_off < my_idx || y_off >= height - block_h - subpel_idx[2][my]) {

  1134.             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src - my_idx * linesize - mx_idx, linesize,

  1135.                                     block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],

  1136.                                     x_off - mx_idx, y_off - my_idx, width, height);

  1137.             src = s->edge_emu_buffer + mx_idx + linesize * my_idx;

  1138.         }

  1139.         mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);

  1140.     } else {

  1141.         ff_thread_await_progress(ref, (3 + y_off + block_h) >> 4, 0);

  1142.         mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);

  1143.     }

  1144. }

  1145. 

  1146. /**

  1147.  * chroma MC function

  1148.  *

  1149.  * @param s VP8 decoding context

  1150.  * @param dst1 target buffer for block data at block position (U plane)

  1151.  * @param dst2 target buffer for block data at block position (V plane)

  1152.  * @param ref reference picture buffer at origin (0, 0)

  1153.  * @param mv motion vector (relative to block position) to get pixel data from

  1154.  * @param x_off horizontal position of block from origin (0, 0)

  1155.  * @param y_off vertical position of block from origin (0, 0)

  1156.  * @param block_w width of block (16, 8 or 4)

  1157.  * @param block_h height of block (always same as block_w)

  1158.  * @param width width of src/dst plane data

  1159.  * @param height height of src/dst plane data

  1160.  * @param linesize size of a single line of plane data, including padding

  1161.  * @param mc_func motion compensation function pointers (bilinear or sixtap MC)

  1162.  */

  1163. static av_always_inline

  1164. void vp8_mc_chroma(VP8Context *s, uint8_t *dst1, uint8_t *dst2, AVFrame *ref,

  1165.                    const VP56mv *mv, int x_off, int y_off,

  1166.                    int block_w, int block_h, int width, int height, int linesize,

  1167.                    vp8_mc_func mc_func[3][3])

  1168. {

  1169.     uint8_t *src1 = ref->data[1], *src2 = ref->data[2];

  1170. 

  1171.     if (AV_RN32A(mv)) {

  1172.         int mx = mv->x&7, mx_idx = subpel_idx[0][mx];

  1173.         int my = mv->y&7, my_idx = subpel_idx[0][my];

  1174. 

  1175.         x_off += mv->x >> 3;

  1176.         y_off += mv->y >> 3;

  1177. 

  1178.         // edge emulation

  1179.         src1 += y_off * linesize + x_off;

  1180.         src2 += y_off * linesize + x_off;

  1181.         ff_thread_await_progress(ref, (3 + y_off + block_h + subpel_idx[2][my]) >> 3, 0);

  1182.         if (x_off < mx_idx || x_off >= width  - block_w - subpel_idx[2][mx] ||

  1183.             y_off < my_idx || y_off >= height - block_h - subpel_idx[2][my]) {

  1184.             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src1 - my_idx * linesize - mx_idx, linesize,

  1185.                                     block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],

  1186.                                     x_off - mx_idx, y_off - my_idx, width, height);

  1187.             src1 = s->edge_emu_buffer + mx_idx + linesize * my_idx;

  1188.             mc_func[my_idx][mx_idx](dst1, linesize, src1, linesize, block_h, mx, my);

  1189. 

  1190.             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src2 - my_idx * linesize - mx_idx, linesize,

  1191.                                     block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],

  1192.                                     x_off - mx_idx, y_off - my_idx, width, height);

  1193.             src2 = s->edge_emu_buffer + mx_idx + linesize * my_idx;

  1194.             mc_func[my_idx][mx_idx](dst2, linesize, src2, linesize, block_h, mx, my);

  1195.         } else {

  1196.             mc_func[my_idx][mx_idx](dst1, linesize, src1, linesize, block_h, mx, my);

  1197.             mc_func[my_idx][mx_idx](dst2, linesize, src2, linesize, block_h, mx, my);

  1198.         }

  1199.     } else {

  1200.         ff_thread_await_progress(ref, (3 + y_off + block_h) >> 3, 0);

  1201.         mc_func[0][0](dst1, linesize, src1 + y_off * linesize + x_off, linesize, block_h, 0, 0);

  1202.         mc_func[0][0](dst2, linesize, src2 + y_off * linesize + x_off, linesize, block_h, 0, 0);

  1203.     }

  1204. }

  1205. 

  1206. static av_always_inline

  1207. void vp8_mc_part(VP8Context *s, uint8_t *dst[3],

  1208.                  AVFrame *ref_frame, int x_off, int y_off,

  1209.                  int bx_off, int by_off,

  1210.                  int block_w, int block_h,

  1211.                  int width, int height, VP56mv *mv)

  1212. {

  1213.     VP56mv uvmv = *mv;

  1214. 

  1215.     /* Y */

  1216.     vp8_mc_luma(s, dst[0] + by_off * s->linesize + bx_off,

  1217.                 ref_frame, mv, x_off + bx_off, y_off + by_off,

  1218.                 block_w, block_h, width, height, s->linesize,

  1219.                 s->put_pixels_tab[block_w == 8]);

  1220. 

  1221.     /* U/V */

  1222.     if (s->profile == 3) {

  1223.         uvmv.x &= ~7;

  1224.         uvmv.y &= ~7;

  1225.     }

  1226.     x_off   >>= 1; y_off   >>= 1;

  1227.     bx_off  >>= 1; by_off  >>= 1;

  1228.     width   >>= 1; height  >>= 1;

  1229.     block_w >>= 1; block_h >>= 1;

  1230.     vp8_mc_chroma(s, dst[1] + by_off * s->uvlinesize + bx_off,

  1231.                   dst[2] + by_off * s->uvlinesize + bx_off, ref_frame,

  1232.                   &uvmv, x_off + bx_off, y_off + by_off,

  1233.                   block_w, block_h, width, height, s->uvlinesize,

  1234.                   s->put_pixels_tab[1 + (block_w == 4)]);

  1235. }

  1236. 

  1237. /* Fetch pixels for estimated mv 4 macroblocks ahead.

  1238.  * Optimized for 64-byte cache lines.  Inspired by ffh264 prefetch_motion. */

  1239. static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref)

  1240. {

  1241.     /* Don't prefetch refs that haven't been used very often this frame. */

  1242.     if (s->ref_count[ref-1] > (mb_xy >> 5)) {

  1243.         int x_off = mb_x << 4, y_off = mb_y << 4;

  1244.         int mx = (mb->mv.x>>2) + x_off + 8;

  1245.         int my = (mb->mv.y>>2) + y_off;

  1246.         uint8_t **src= s->framep[ref]->data;

  1247.         int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;

  1248.         /* For threading, a ff_thread_await_progress here might be useful, but

  1249.          * it actually slows down the decoder. Since a bad prefetch doesn't

  1250.          * generate bad decoder output, we don't run it here. */

  1251.         s->dsp.prefetch(src[0]+off, s->linesize, 4);

  1252.         off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;

  1253.         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);

  1254.     }

  1255. }

  1256. 

  1257. /**

  1258.  * Apply motion vectors to prediction buffer, chapter 18.

  1259.  */

  1260. static av_always_inline

  1261. void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,

  1262.                    int mb_x, int mb_y)

  1263. {

  1264.     int x_off = mb_x << 4, y_off = mb_y << 4;

  1265.     int width = 16*s->mb_width, height = 16*s->mb_height;

  1266.     AVFrame *ref = s->framep[mb->ref_frame];

  1267.     VP56mv *bmv = mb->bmv;

  1268. 

  1269.     switch (mb->partitioning) {

  1270.     case VP8_SPLITMVMODE_NONE:

  1271.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1272.                     0, 0, 16, 16, width, height, &mb->mv);

  1273.         break;

  1274.     case VP8_SPLITMVMODE_4x4: {

  1275.         int x, y;

  1276.         VP56mv uvmv;

  1277. 

  1278.         /* Y */

  1279.         for (y = 0; y < 4; y++) {

  1280.             for (x = 0; x < 4; x++) {

  1281.                 vp8_mc_luma(s, dst[0] + 4*y*s->linesize + x*4,

  1282.                             ref, &bmv[4*y + x],

  1283.                             4*x + x_off, 4*y + y_off, 4, 4,

  1284.                             width, height, s->linesize,

  1285.                             s->put_pixels_tab[2]);

  1286.             }

  1287.         }

  1288. 

  1289.         /* U/V */

  1290.         x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;

  1291.         for (y = 0; y < 2; y++) {

  1292.             for (x = 0; x < 2; x++) {

  1293.                 uvmv.x = mb->bmv[ 2*y    * 4 + 2*x  ].x +

  1294.                          mb->bmv[ 2*y    * 4 + 2*x+1].x +

  1295.                          mb->bmv[(2*y+1) * 4 + 2*x  ].x +

  1296.                          mb->bmv[(2*y+1) * 4 + 2*x+1].x;

  1297.                 uvmv.y = mb->bmv[ 2*y    * 4 + 2*x  ].y +

  1298.                          mb->bmv[ 2*y    * 4 + 2*x+1].y +

  1299.                          mb->bmv[(2*y+1) * 4 + 2*x  ].y +

  1300.                          mb->bmv[(2*y+1) * 4 + 2*x+1].y;

  1301.                 uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;

  1302.                 uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;

  1303.                 if (s->profile == 3) {

  1304.                     uvmv.x &= ~7;

  1305.                     uvmv.y &= ~7;

  1306.                 }

  1307.                 vp8_mc_chroma(s, dst[1] + 4*y*s->uvlinesize + x*4,

  1308.                               dst[2] + 4*y*s->uvlinesize + x*4, ref, &uvmv,

  1309.                               4*x + x_off, 4*y + y_off, 4, 4,

  1310.                               width, height, s->uvlinesize,

  1311.                               s->put_pixels_tab[2]);

  1312.             }

  1313.         }

  1314.         break;

  1315.     }

  1316.     case VP8_SPLITMVMODE_16x8:

  1317.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1318.                     0, 0, 16, 8, width, height, &bmv[0]);

  1319.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1320.                     0, 8, 16, 8, width, height, &bmv[1]);

  1321.         break;

  1322.     case VP8_SPLITMVMODE_8x16:

  1323.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1324.                     0, 0, 8, 16, width, height, &bmv[0]);

  1325.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1326.                     8, 0, 8, 16, width, height, &bmv[1]);

  1327.         break;

  1328.     case VP8_SPLITMVMODE_8x8:

  1329.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1330.                     0, 0, 8, 8, width, height, &bmv[0]);

  1331.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1332.                     8, 0, 8, 8, width, height, &bmv[1]);

  1333.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1334.                     0, 8, 8, 8, width, height, &bmv[2]);

  1335.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1336.                     8, 8, 8, 8, width, height, &bmv[3]);

  1337.         break;

  1338.     }

  1339. }

  1340. 

  1341. static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb)

  1342. {

  1343.     int x, y, ch;

  1344. 

  1345.     if (mb->mode != MODE_I4x4) {

  1346.         uint8_t *y_dst = dst[0];

  1347.         for (y = 0; y < 4; y++) {

  1348.             uint32_t nnz4 = AV_RL32(s->non_zero_count_cache[y]);

  1349.             if (nnz4) {

  1350.                 if (nnz4&~0x01010101) {

  1351.                     for (x = 0; x < 4; x++) {

  1352.                         if ((uint8_t)nnz4 == 1)

  1353.                             s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);

  1354.                         else if((uint8_t)nnz4 > 1)

  1355.                             s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);

  1356.                         nnz4 >>= 8;

  1357.                         if (!nnz4)

  1358.                             break;

  1359.                     }

  1360.                 } else {

  1361.                     s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize);

  1362.                 }

  1363.             }

  1364.             y_dst += 4*s->linesize;

  1365.         }

  1366.     }

  1367. 

  1368.     for (ch = 0; ch < 2; ch++) {

  1369.         uint32_t nnz4 = AV_RL32(s->non_zero_count_cache[4+ch]);

  1370.         if (nnz4) {

  1371.             uint8_t *ch_dst = dst[1+ch];

  1372.             if (nnz4&~0x01010101) {

  1373.                 for (y = 0; y < 2; y++) {

  1374.                     for (x = 0; x < 2; x++) {

  1375.                         if ((uint8_t)nnz4 == 1)

  1376.                             s->vp8dsp.vp8_idct_dc_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);

  1377.                         else if((uint8_t)nnz4 > 1)

  1378.                             s->vp8dsp.vp8_idct_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);

  1379.                         nnz4 >>= 8;

  1380.                         if (!nnz4)

  1381.                             goto chroma_idct_end;

  1382.                     }

  1383.                     ch_dst += 4*s->uvlinesize;

  1384.                 }

  1385.             } else {

  1386.                 s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize);

  1387.             }

  1388.         }

  1389. chroma_idct_end: ;

  1390.     }

  1391. }

  1392. 

  1393. static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )

  1394. {

  1395.     int interior_limit, filter_level;

  1396. 

  1397.     if (s->segmentation.enabled) {

  1398.         filter_level = s->segmentation.filter_level[s->segment];

  1399.         if (!s->segmentation.absolute_vals)

  1400.             filter_level += s->filter.level;

  1401.     } else

  1402.         filter_level = s->filter.level;

  1403. 

  1404.     if (s->lf_delta.enabled) {

  1405.         filter_level += s->lf_delta.ref[mb->ref_frame];

  1406.         filter_level += s->lf_delta.mode[mb->mode];

  1407.     }

  1408. 

  1409.     filter_level = av_clip_uintp2(filter_level, 6);

  1410. 

  1411.     interior_limit = filter_level;

  1412.     if (s->filter.sharpness) {

  1413.         interior_limit >>= (s->filter.sharpness + 3) >> 2;

  1414.         interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);

  1415.     }

  1416.     interior_limit = FFMAX(interior_limit, 1);

  1417. 

  1418.     f->filter_level = filter_level;

  1419.     f->inner_limit = interior_limit;

  1420.     f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;

  1421. }

  1422. 

  1423. static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)

  1424. {

  1425.     int mbedge_lim, bedge_lim, hev_thresh;

  1426.     int filter_level = f->filter_level;

  1427.     int inner_limit = f->inner_limit;

  1428.     int inner_filter = f->inner_filter;

  1429.     int linesize = s->linesize;

  1430.     int uvlinesize = s->uvlinesize;

  1431.     static const uint8_t hev_thresh_lut[2][64] = {

  1432.         { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1,

  1433.           2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,

  1434.           3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,

  1435.           3, 3, 3, 3 },

  1436.         { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1,

  1437.           1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,

  1438.           2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,

  1439.           2, 2, 2, 2 }

  1440.     };

  1441. 

  1442.     if (!filter_level)

  1443.         return;

  1444. 

  1445.      bedge_lim = 2*filter_level + inner_limit;

  1446.     mbedge_lim = bedge_lim + 4;

  1447. 

  1448.     hev_thresh = hev_thresh_lut[s->keyframe][filter_level];

  1449. 

  1450.     if (mb_x) {

  1451.         s->vp8dsp.vp8_h_loop_filter16y(dst[0],     linesize,

  1452.                                        mbedge_lim, inner_limit, hev_thresh);

  1453.         s->vp8dsp.vp8_h_loop_filter8uv(dst[1],     dst[2],      uvlinesize,

  1454.                                        mbedge_lim, inner_limit, hev_thresh);

  1455.     }

  1456. 

  1457.     if (inner_filter) {

  1458.         s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, linesize, bedge_lim,

  1459.                                              inner_limit, hev_thresh);

  1460.         s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, linesize, bedge_lim,

  1461.                                              inner_limit, hev_thresh);

  1462.         s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, linesize, bedge_lim,

  1463.                                              inner_limit, hev_thresh);

  1464.         s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4,

  1465.                                              uvlinesize,  bedge_lim,

  1466.                                              inner_limit, hev_thresh);

  1467.     }

  1468. 

  1469.     if (mb_y) {

  1470.         s->vp8dsp.vp8_v_loop_filter16y(dst[0],     linesize,

  1471.                                        mbedge_lim, inner_limit, hev_thresh);

  1472.         s->vp8dsp.vp8_v_loop_filter8uv(dst[1],     dst[2],      uvlinesize,

  1473.                                        mbedge_lim, inner_limit, hev_thresh);

  1474.     }

  1475. 

  1476.     if (inner_filter) {

  1477.         s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*linesize,

  1478.                                              linesize,    bedge_lim,

  1479.                                              inner_limit, hev_thresh);

  1480.         s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*linesize,

  1481.                                              linesize,    bedge_lim,

  1482.                                              inner_limit, hev_thresh);

  1483.         s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*linesize,

  1484.                                              linesize,    bedge_lim,

  1485.                                              inner_limit, hev_thresh);

  1486.         s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * uvlinesize,

  1487.                                              dst[2] + 4 * uvlinesize,

  1488.                                              uvlinesize,  bedge_lim,

  1489.                                              inner_limit, hev_thresh);

  1490.     }

  1491. }

  1492. 

  1493. static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)

  1494. {

  1495.     int mbedge_lim, bedge_lim;

  1496.     int filter_level = f->filter_level;

  1497.     int inner_limit = f->inner_limit;

  1498.     int inner_filter = f->inner_filter;

  1499.     int linesize = s->linesize;

  1500. 

  1501.     if (!filter_level)

  1502.         return;

  1503. 

  1504.      bedge_lim = 2*filter_level + inner_limit;

  1505.     mbedge_lim = bedge_lim + 4;

  1506. 

  1507.     if (mb_x)

  1508.         s->vp8dsp.vp8_h_loop_filter_simple(dst, linesize, mbedge_lim);

  1509.     if (inner_filter) {

  1510.         s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, linesize, bedge_lim);

  1511.         s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, linesize, bedge_lim);

  1512.         s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim);

  1513.     }

  1514. 

  1515.     if (mb_y)

  1516.         s->vp8dsp.vp8_v_loop_filter_simple(dst, linesize, mbedge_lim);

  1517.     if (inner_filter) {

  1518.         s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*linesize, linesize, bedge_lim);

  1519.         s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*linesize, linesize, bedge_lim);

  1520.         s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim);

  1521.     }

  1522. }

  1523. 

  1524. static void filter_mb_row(VP8Context *s, AVFrame *curframe, int mb_y)

  1525. {

  1526.     VP8FilterStrength *f = s->filter_strength;

  1527.     uint8_t *dst[3] = {

  1528.         curframe->data[0] + 16*mb_y*s->linesize,

  1529.         curframe->data[1] +  8*mb_y*s->uvlinesize,

  1530.         curframe->data[2] +  8*mb_y*s->uvlinesize

  1531.     };

  1532.     int mb_x;

  1533. 

  1534.     for (mb_x = 0; mb_x < s->mb_width; mb_x++) {

  1535.         backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);

  1536.         filter_mb(s, dst, f++, mb_x, mb_y);

  1537.         dst[0] += 16;

  1538.         dst[1] += 8;

  1539.         dst[2] += 8;

  1540.     }

  1541. }

  1542. 

  1543. static void filter_mb_row_simple(VP8Context *s, AVFrame *curframe, int mb_y)

  1544. {

  1545.     VP8FilterStrength *f = s->filter_strength;

  1546.     uint8_t *dst = curframe->data[0] + 16*mb_y*s->linesize;

  1547.     int mb_x;

  1548. 

  1549.     for (mb_x = 0; mb_x < s->mb_width; mb_x++) {

  1550.         backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1);

  1551.         filter_mb_simple(s, dst, f++, mb_x, mb_y);

  1552.         dst += 16;

  1553.     }

  1554. }

  1555. 

  1556. static void release_queued_segmaps(VP8Context *s, int is_close)

  1557. {

  1558.     int leave_behind = is_close ? 0 : !s->maps_are_invalid;

  1559.     while (s->num_maps_to_be_freed > leave_behind)

  1560.         av_freep(&s->segmentation_maps[--s->num_maps_to_be_freed]);

  1561.     s->maps_are_invalid = 0;

  1562. }

  1563. 

  1564. static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,

  1565.                             AVPacket *avpkt)

  1566. {

  1567.     VP8Context *s = avctx->priv_data;

  1568.     int ret, mb_x, mb_y, i, y, referenced;

  1569.     enum AVDiscard skip_thresh;

  1570.     AVFrame *av_uninit(curframe), *prev_frame = s->framep[VP56_FRAME_CURRENT];

  1571. 

  1572.     release_queued_segmaps(s, 0);

  1573. 

  1574.     if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)

  1575.         return ret;

  1576. 

  1577.     referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT

  1578.                                 || s->update_altref == VP56_FRAME_CURRENT;

  1579. 

  1580.     skip_thresh = !referenced ? AVDISCARD_NONREF :

  1581.                     !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;

  1582. 

  1583.     if (avctx->skip_frame >= skip_thresh) {

  1584.         s->invisible = 1;

  1585.         goto skip_decode;

  1586.     }

  1587.     s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;

  1588. 

  1589.     // release no longer referenced frames

  1590.     for (i = 0; i < 5; i++)

  1591.         if (s->frames[i].data[0] &&

  1592.             &s->frames[i] != prev_frame &&

  1593.             &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&

  1594.             &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&

  1595.             &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])

  1596.             vp8_release_frame(s, &s->frames[i], 1, 0);

  1597. 

  1598.     // find a free buffer

  1599.     for (i = 0; i < 5; i++)

  1600.         if (&s->frames[i] != prev_frame &&

  1601.             &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&

  1602.             &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&

  1603.             &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {

  1604.             curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];

  1605.             break;

  1606.         }

  1607.     if (i == 5) {

  1608.         av_log(avctx, AV_LOG_FATAL, "Ran out of free frames!\n");

  1609.         abort();

  1610.     }

  1611.     if (curframe->data[0])

  1612.         vp8_release_frame(s, curframe, 1, 0);

  1613. 

  1614.     curframe->key_frame = s->keyframe;

  1615.     curframe->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;

  1616.     curframe->reference = referenced ? 3 : 0;

  1617.     if ((ret = vp8_alloc_frame(s, curframe))) {

  1618.         av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");

  1619.         return ret;

  1620.     }

  1621. 

  1622.     // check if golden and altref are swapped

  1623.     if (s->update_altref != VP56_FRAME_NONE) {

  1624.         s->next_framep[VP56_FRAME_GOLDEN2]  = s->framep[s->update_altref];

  1625.     } else {

  1626.         s->next_framep[VP56_FRAME_GOLDEN2]  = s->framep[VP56_FRAME_GOLDEN2];

  1627.     }

  1628.     if (s->update_golden != VP56_FRAME_NONE) {

  1629.         s->next_framep[VP56_FRAME_GOLDEN]   = s->framep[s->update_golden];

  1630.     } else {

  1631.         s->next_framep[VP56_FRAME_GOLDEN]   = s->framep[VP56_FRAME_GOLDEN];

  1632.     }

  1633.     if (s->update_last) {

  1634.         s->next_framep[VP56_FRAME_PREVIOUS] = curframe;

  1635.     } else {

  1636.         s->next_framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_PREVIOUS];

  1637.     }

  1638.     s->next_framep[VP56_FRAME_CURRENT]      = curframe;

  1639. 

  1640.     ff_thread_finish_setup(avctx);

  1641. 

  1642.     // Given that arithmetic probabilities are updated every frame, it's quite likely

  1643.     // that the values we have on a random interframe are complete junk if we didn't

  1644.     // start decode on a keyframe. So just don't display anything rather than junk.

  1645.     if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||

  1646.                          !s->framep[VP56_FRAME_GOLDEN] ||

  1647.                          !s->framep[VP56_FRAME_GOLDEN2])) {

  1648.         av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");

  1649.         return AVERROR_INVALIDDATA;

  1650.     }

  1651. 

  1652.     s->linesize   = curframe->linesize[0];

  1653.     s->uvlinesize = curframe->linesize[1];

  1654. 

  1655.     if (!s->edge_emu_buffer)

  1656.         s->edge_emu_buffer = av_malloc(21*s->linesize);

  1657. 

  1658.     memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));

  1659. 

  1660.     /* Zero macroblock structures for top/top-left prediction from outside the frame. */

  1661.     memset(s->macroblocks + s->mb_height*2 - 1, 0, (s->mb_width+1)*sizeof(*s->macroblocks));

  1662. 

  1663.     // top edge of 127 for intra prediction

  1664.     if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {

  1665.         s->top_border[0][15] = s->top_border[0][23] = 127;

  1666.         memset(s->top_border[1]-1, 127, s->mb_width*sizeof(*s->top_border)+1);

  1667.     }

  1668.     memset(s->ref_count, 0, sizeof(s->ref_count));

  1669.     if (s->keyframe)

  1670.         memset(s->intra4x4_pred_mode_top, DC_PRED, s->mb_width*4);

  1671. 

  1672. #define MARGIN (16 << 2)

  1673.     s->mv_min.y = -MARGIN;

  1674.     s->mv_max.y = ((s->mb_height - 1) << 6) + MARGIN;

  1675. 

  1676.     for (mb_y = 0; mb_y < s->mb_height; mb_y++) {

  1677.         VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];

  1678.         VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;

  1679.         int mb_xy = mb_y*s->mb_width;

  1680.         uint8_t *dst[3] = {

  1681.             curframe->data[0] + 16*mb_y*s->linesize,

  1682.             curframe->data[1] +  8*mb_y*s->uvlinesize,

  1683.             curframe->data[2] +  8*mb_y*s->uvlinesize

  1684.         };

  1685. 

  1686.         memset(mb - 1, 0, sizeof(*mb));   // zero left macroblock

  1687.         memset(s->left_nnz, 0, sizeof(s->left_nnz));

  1688.         AV_WN32A(s->intra4x4_pred_mode_left, DC_PRED*0x01010101);

  1689. 

  1690.         // left edge of 129 for intra prediction

  1691.         if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {

  1692.             for (i = 0; i < 3; i++)

  1693.                 for (y = 0; y < 16>>!!i; y++)

  1694.                     dst[i][y*curframe->linesize[i]-1] = 129;

  1695.             if (mb_y == 1) // top left edge is also 129

  1696.                 s->top_border[0][15] = s->top_border[0][23] = s->top_border[0][31] = 129;

  1697.         }

  1698. 

  1699.         s->mv_min.x = -MARGIN;

  1700.         s->mv_max.x = ((s->mb_width  - 1) << 6) + MARGIN;

  1701.         if (prev_frame && s->segmentation.enabled && !s->segmentation.update_map)

  1702.             ff_thread_await_progress(prev_frame, mb_y, 0);

  1703. 

  1704.         for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {

  1705.             /* Prefetch the current frame, 4 MBs ahead */

  1706.             s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);

  1707.             s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);

  1708. 

  1709.             decode_mb_mode(s, mb, mb_x, mb_y, curframe->ref_index[0] + mb_xy,

  1710.                            prev_frame && prev_frame->ref_index[0] ? prev_frame->ref_index[0] + mb_xy : NULL);

  1711. 

  1712.             prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS);

  1713. 

  1714.             if (!mb->skip)

  1715.                 decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);

  1716. 

  1717.             if (mb->mode <= MODE_I4x4)

  1718.                 intra_predict(s, dst, mb, mb_x, mb_y);

  1719.             else

  1720.                 inter_predict(s, dst, mb, mb_x, mb_y);

  1721. 

  1722.             prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN);

  1723. 

  1724.             if (!mb->skip) {

  1725.                 idct_mb(s, dst, mb);

  1726.             } else {

  1727.                 AV_ZERO64(s->left_nnz);

  1728.                 AV_WN64(s->top_nnz[mb_x], 0);   // array of 9, so unaligned

  1729. 

  1730.                 // Reset DC block predictors if they would exist if the mb had coefficients

  1731.                 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {

  1732.                     s->left_nnz[8]      = 0;

  1733.                     s->top_nnz[mb_x][8] = 0;

  1734.                 }

  1735.             }

  1736. 

  1737.             if (s->deblock_filter)

  1738.                 filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);

  1739. 

  1740.             prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2);

  1741. 

  1742.             dst[0] += 16;

  1743.             dst[1] += 8;

  1744.             dst[2] += 8;

  1745.             s->mv_min.x -= 64;

  1746.             s->mv_max.x -= 64;

  1747.         }

  1748.         if (s->deblock_filter) {

  1749.             if (s->filter.simple)

  1750.                 filter_mb_row_simple(s, curframe, mb_y);

  1751.             else

  1752.                 filter_mb_row(s, curframe, mb_y);

  1753.         }

  1754.         s->mv_min.y -= 64;

  1755.         s->mv_max.y -= 64;

  1756. 

  1757.         ff_thread_report_progress(curframe, mb_y, 0);

  1758.     }

  1759. 

  1760.     ff_thread_report_progress(curframe, INT_MAX, 0);

  1761. skip_decode:

  1762.     // if future frames don't use the updated probabilities,

  1763.     // reset them to the values we saved

  1764.     if (!s->update_probabilities)

  1765.         s->prob[0] = s->prob[1];

  1766. 

  1767.     memcpy(&s->framep[0], &s->next_framep[0], sizeof(s->framep[0]) * 4);

  1768. 

  1769.     if (!s->invisible) {

  1770.         *(AVFrame*)data = *curframe;

  1771.         *data_size = sizeof(AVFrame);

  1772.     }

  1773. 

  1774.     return avpkt->size;

  1775. }

  1776. 

  1777. static av_cold int vp8_decode_init(AVCodecContext *avctx)

  1778. {

  1779.     VP8Context *s = avctx->priv_data;

  1780. 

  1781.     s->avctx = avctx;

  1782.     avctx->pix_fmt = PIX_FMT_YUV420P;

  1783. 

  1784.     dsputil_init(&s->dsp, avctx);

  1785.     ff_h264_pred_init(&s->hpc, CODEC_ID_VP8, 8, 1);

  1786.     ff_vp8dsp_init(&s->vp8dsp);

  1787. 

  1788.     return 0;

  1789. }

  1790. 

  1791. static av_cold int vp8_decode_free(AVCodecContext *avctx)

  1792. {

  1793.     vp8_decode_flush_impl(avctx, 0, 1, 1);

  1794.     release_queued_segmaps(avctx->priv_data, 1);

  1795.     return 0;

  1796. }

  1797. 

  1798. static av_cold int vp8_decode_init_thread_copy(AVCodecContext *avctx)

  1799. {

  1800.     VP8Context *s = avctx->priv_data;

  1801. 

  1802.     s->avctx = avctx;

  1803. 

  1804.     return 0;

  1805. }

  1806. 

  1807. #define REBASE(pic) \

  1808.     pic ? pic - &s_src->frames[0] + &s->frames[0] : NULL

  1809. 

  1810. static int vp8_decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src)

  1811. {

  1812.     VP8Context *s = dst->priv_data, *s_src = src->priv_data;

  1813. 

  1814.     if (s->macroblocks_base &&

  1815.         (s_src->mb_width != s->mb_width || s_src->mb_height != s->mb_height)) {

  1816.         free_buffers(s);

  1817.     }

  1818. 

  1819.     s->prob[0] = s_src->prob[!s_src->update_probabilities];

  1820.     s->segmentation = s_src->segmentation;

  1821.     s->lf_delta = s_src->lf_delta;

  1822.     memcpy(s->sign_bias, s_src->sign_bias, sizeof(s->sign_bias));

  1823. 

  1824.     memcpy(&s->frames, &s_src->frames, sizeof(s->frames));

  1825.     s->framep[0] = REBASE(s_src->next_framep[0]);

  1826.     s->framep[1] = REBASE(s_src->next_framep[1]);

  1827.     s->framep[2] = REBASE(s_src->next_framep[2]);

  1828.     s->framep[3] = REBASE(s_src->next_framep[3]);

  1829. 

  1830.     return 0;

  1831. }

  1832. 

  1833. AVCodec ff_vp8_decoder = {

  1834.     .name           = "vp8",

  1835.     .type           = AVMEDIA_TYPE_VIDEO,

  1836.     .id             = CODEC_ID_VP8,

  1837.     .priv_data_size = sizeof(VP8Context),

  1838.     .init           = vp8_decode_init,

  1839.     .close          = vp8_decode_free,

  1840.     .decode         = vp8_decode_frame,

  1841.     .capabilities   = CODEC_CAP_DR1 | CODEC_CAP_FRAME_THREADS,

  1842.     .flush = vp8_decode_flush,

  1843.     .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),

  1844.     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(vp8_decode_init_thread_copy),

  1845.     .update_thread_context = ONLY_IF_THREADS_ENABLED(vp8_decode_update_thread_context),

  1846. };