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.
31446 Commits
32 Branches
1.0GB
Tree: 5f5f36b52e
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '5f5f36b52e'
${ noResults }
FFmpeg/libavcodec/vp8.c

1792 lines
64KB
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.     av_freep(&s->segmentation_map);

  45. 

  46.     s->macroblocks = NULL;

  47. }

  48. 

  49. static void vp8_decode_flush(AVCodecContext *avctx)

  50. {

  51.     VP8Context *s = avctx->priv_data;

  52.     int i;

  53. 

  54.     if (!avctx->is_copy) {

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

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

  57.                 ff_thread_release_buffer(avctx, &s->frames[i]);

  58.     }

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

  60. 

  61.     free_buffers(s);

  62. }

  63. 

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

  65. {

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

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

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

  69.             return AVERROR_INVALIDDATA;

  70. 

  71.         vp8_decode_flush(s->avctx);

  72. 

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

  74.     }

  75. 

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

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

  78. 

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

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

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

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

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

  84.     s->segmentation_map        = av_mallocz(s->mb_width*s->mb_height);

  85. 

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

  87.         !s->top_nnz || !s->top_border || !s->segmentation_map)

  88.         return AVERROR(ENOMEM);

  89. 

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

  91. 

  92.     return 0;

  93. }

  94. 

  95. static void parse_segment_info(VP8Context *s)

  96. {

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

  98.     int i;

  99. 

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

  101. 

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

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

  104. 

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

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

  107. 

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

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

  110.     }

  111.     if (s->segmentation.update_map)

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

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

  114. }

  115. 

  116. static void update_lf_deltas(VP8Context *s)

  117. {

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

  119.     int i;

  120. 

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

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

  123. 

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

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

  126. }

  127. 

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

  129. {

  130.     const uint8_t *sizes = buf;

  131.     int i;

  132. 

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

  134. 

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

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

  137.     if (buf_size < 0)

  138.         return -1;

  139. 

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

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

  142.         if (buf_size - size < 0)

  143.             return -1;

  144. 

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

  146.         buf      += size;

  147.         buf_size -= size;

  148.     }

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

  150. 

  151.     return 0;

  152. }

  153. 

  154. static void get_quants(VP8Context *s)

  155. {

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

  157.     int i, base_qi;

  158. 

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

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

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

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

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

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

  165. 

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

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

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

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

  170.                 base_qi += yac_qi;

  171.         } else

  172.             base_qi = yac_qi;

  173. 

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

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

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

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

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

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

  180. 

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

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

  183.     }

  184. }

  185. 

  186. /**

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

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

  189.  *

  190.  * Intra frames update all 3 references

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

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

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

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

  195.  *      0: no update

  196.  *      1: VP56_FRAME_PREVIOUS

  197.  *      2: update golden with altref, or update altref with golden

  198.  */

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

  200. {

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

  202. 

  203.     if (update)

  204.         return VP56_FRAME_CURRENT;

  205. 

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

  207.     case 1:

  208.         return VP56_FRAME_PREVIOUS;

  209.     case 2:

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

  211.     }

  212.     return VP56_FRAME_NONE;

  213. }

  214. 

  215. static void update_refs(VP8Context *s)

  216. {

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

  218. 

  219.     int update_golden = vp8_rac_get(c);

  220.     int update_altref = vp8_rac_get(c);

  221. 

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

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

  224. }

  225. 

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

  227. {

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

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

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

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

  232. 

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

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

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

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

  237.     buf      += 3;

  238.     buf_size -= 3;

  239. 

  240.     if (s->profile > 3)

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

  242. 

  243.     if (!s->profile)

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

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

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

  247. 

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

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

  250.         return AVERROR_INVALIDDATA;

  251.     }

  252. 

  253.     if (s->keyframe) {

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

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

  256.             return AVERROR_INVALIDDATA;

  257.         }

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

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

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

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

  262.         buf      += 7;

  263.         buf_size -= 7;

  264. 

  265.         if (hscale || vscale)

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

  267. 

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

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

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

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

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

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

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

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

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

  277.     }

  278. 

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

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

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

  282.             return ret;

  283.     }

  284. 

  285.     ff_vp56_init_range_decoder(c, buf, header_size);

  286.     buf      += header_size;

  287.     buf_size -= header_size;

  288. 

  289.     if (s->keyframe) {

  290.         if (vp8_rac_get(c))

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

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

  293.     }

  294. 

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

  296.         parse_segment_info(s);

  297.     else

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

  299. 

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

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

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

  303. 

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

  305.         if (vp8_rac_get(c))

  306.             update_lf_deltas(s);

  307. 

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

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

  310.         return AVERROR_INVALIDDATA;

  311.     }

  312. 

  313.     get_quants(s);

  314. 

  315.     if (!s->keyframe) {

  316.         update_refs(s);

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

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

  319.     }

  320. 

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

  322.     // make a copy of the current probabilities

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

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

  325. 

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

  327. 

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

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

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

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

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

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

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

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

  336.                     }

  337. 

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

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

  340. 

  341.     if (!s->keyframe) {

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

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

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

  345. 

  346.         if (vp8_rac_get(c))

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

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

  349.         if (vp8_rac_get(c))

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

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

  352. 

  353.         // 17.2 MV probability update

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

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

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

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

  358.     }

  359. 

  360.     return 0;

  361. }

  362. 

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

  364. {

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

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

  367. }

  368. 

  369. /**

  370.  * Motion vector coding, 17.1.

  371.  */

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

  373. {

  374.     int bit, x = 0;

  375. 

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

  377.         int i;

  378. 

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

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

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

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

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

  384.             x += 8;

  385.     } else {

  386.         // small_mvtree

  387.         const uint8_t *ps = p+2;

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

  389.         ps += 1 + 3*bit;

  390.         x  += 4*bit;

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

  392.         ps += 1 + bit;

  393.         x  += 2*bit;

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

  395.     }

  396. 

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

  398. }

  399. 

  400. static av_always_inline

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

  402. {

  403.     if (left == top)

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

  405.     if (!top)

  406.         return vp8_submv_prob[2];

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

  408. }

  409. 

  410. /**

  411.  * Split motion vector prediction, 16.4.

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

  413.  */

  414. static av_always_inline

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

  416. {

  417.     int part_idx;

  418.     int n, num;

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

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

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

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

  423.                   *mbsplits_cur, *firstidx;

  424.     VP56mv *top_mv  = top_mb->bmv;

  425.     VP56mv *left_mv = left_mb->bmv;

  426.     VP56mv *cur_mv  = mb->bmv;

  427. 

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

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

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

  431.         } else {

  432.             part_idx = VP8_SPLITMVMODE_8x8;

  433.         }

  434.     } else {

  435.         part_idx = VP8_SPLITMVMODE_4x4;

  436.     }

  437. 

  438.     num = vp8_mbsplit_count[part_idx];

  439.     mbsplits_cur = vp8_mbsplits[part_idx],

  440.     firstidx = vp8_mbfirstidx[part_idx];

  441.     mb->partitioning = part_idx;

  442. 

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

  444.         int k = firstidx[n];

  445.         uint32_t left, above;

  446.         const uint8_t *submv_prob;

  447. 

  448.         if (!(k & 3))

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

  450.         else

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

  452.         if (k <= 3)

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

  454.         else

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

  456. 

  457.         submv_prob = get_submv_prob(left, above);

  458. 

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

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

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

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

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

  464.                 } else {

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

  466.                 }

  467.             } else {

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

  469.             }

  470.         } else {

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

  472.         }

  473.     }

  474. 

  475.     return num;

  476. }

  477. 

  478. static av_always_inline

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

  480. {

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

  482.                                   mb - 1 /* left */,

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

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

  485.     enum { VP8_EDGE_TOP, VP8_EDGE_LEFT, VP8_EDGE_TOPLEFT };

  486.     int idx = CNT_ZERO;

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

  488.     int8_t *sign_bias = s->sign_bias;

  489.     VP56mv near_mv[4];

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

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

  492. 

  493.     AV_ZERO32(&near_mv[0]);

  494.     AV_ZERO32(&near_mv[1]);

  495. 

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

  497.     #define MV_EDGE_CHECK(n)\

  498.     {\

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

  500.         int edge_ref = edge->ref_frame;\

  501.         if (edge_ref != VP56_FRAME_CURRENT) {\

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

  503.             if (mv) {\

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

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

  506.                     mv = ~mv;\

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

  508.                 }\

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

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

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

  512.             } else\

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

  514.         }\

  515.     }

  516. 

  517.     MV_EDGE_CHECK(0)

  518.     MV_EDGE_CHECK(1)

  519.     MV_EDGE_CHECK(2)

  520. 

  521.     mb->partitioning = VP8_SPLITMVMODE_NONE;

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

  523.         mb->mode = VP8_MVMODE_MV;

  524. 

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

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

  527.             cnt[CNT_NEAREST] += 1;

  528. 

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

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

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

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

  533.         }

  534. 

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

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

  537. 

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

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

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

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

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

  543. 

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

  545.                     mb->mode = VP8_MVMODE_SPLIT;

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

  547.                 } else {

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

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

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

  551.                 }

  552.             } else {

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

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

  555.             }

  556.         } else {

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

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

  559.         }

  560.     } else {

  561.         mb->mode = VP8_MVMODE_ZERO;

  562.         AV_ZERO32(&mb->mv);

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

  564.     }

  565. }

  566. 

  567. static av_always_inline

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

  569.                            int mb_x, int keyframe)

  570. {

  571.     uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;

  572.     if (keyframe) {

  573.         int x, y;

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

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

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

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

  578.                 const uint8_t *ctx;

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

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

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

  582.                 intra4x4++;

  583.             }

  584.         }

  585.     } else {

  586.         int i;

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

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

  589.     }

  590. }

  591. 

  592. static av_always_inline

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

  594. {

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

  596. 

  597.     if (s->segmentation.update_map)

  598.         *segment = vp8_rac_get_tree(c, vp8_segmentid_tree, s->prob->segmentid);

  599.     else

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

  601.     s->segment = *segment;

  602. 

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

  604. 

  605.     if (s->keyframe) {

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

  607. 

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

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

  610.         } else {

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

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

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

  614.         }

  615. 

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

  617.         mb->ref_frame = VP56_FRAME_CURRENT;

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

  619.         // inter MB, 16.2

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

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

  622.                 VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;

  623.         else

  624.             mb->ref_frame = VP56_FRAME_PREVIOUS;

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

  626. 

  627.         // motion vectors, 16.3

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

  629.     } else {

  630.         // intra MB, 16.1

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

  632. 

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

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

  635. 

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

  637.         mb->ref_frame = VP56_FRAME_CURRENT;

  638.         mb->partitioning = VP8_SPLITMVMODE_NONE;

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

  640.     }

  641. }

  642. 

  643. #ifndef decode_block_coeffs_internal

  644. /**

  645.  * @param c arithmetic bitstream reader context

  646.  * @param block destination for block coefficients

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

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

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

  650.  * @return 0 if no coeffs were decoded

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

  652.  */

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

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

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

  656. {

  657.     goto skip_eob;

  658.     do {

  659.         int coeff;

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

  661.             return i;

  662. 

  663. skip_eob:

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

  665.             if (++i == 16)

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

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

  668.             goto skip_eob;

  669.         }

  670. 

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

  672.             coeff = 1;

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

  674.         } else {

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

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

  677.                 if (coeff)

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

  679.                 coeff += 2;

  680.             } else {

  681.                 // DCT_CAT*

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

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

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

  685.                     } else {                                    // DCT_CAT2

  686.                         coeff  = 7;

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

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

  689.                     }

  690.                 } else {    // DCT_CAT3 and up

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

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

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

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

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

  696.                 }

  697.             }

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

  699.         }

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

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

  702. 

  703.     return i;

  704. }

  705. #endif

  706. 

  707. /**

  708.  * @param c arithmetic bitstream reader context

  709.  * @param block destination for block coefficients

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

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

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

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

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

  715.  * @return 0 if no coeffs were decoded

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

  717.  */

  718. static av_always_inline

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

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

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

  722. {

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

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

  725.         return 0;

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

  727. }

  728. 

  729. static av_always_inline

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

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

  732. {

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

  734.     int nnz_pred, nnz, nnz_total = 0;

  735.     int segment = s->segment;

  736.     int block_dc = 0;

  737. 

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

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

  740. 

  741.         // decode DC values and do hadamard

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

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

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

  745.         if (nnz) {

  746.             nnz_total += nnz;

  747.             block_dc = 1;

  748.             if (nnz == 1)

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

  750.             else

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

  752.         }

  753.         luma_start = 1;

  754.         luma_ctx = 0;

  755.     }

  756. 

  757.     // luma blocks

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

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

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

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

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

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

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

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

  766.             nnz_total += nnz;

  767.         }

  768. 

  769.     // chroma blocks

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

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

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

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

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

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

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

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

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

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

  780.                 nnz_total += nnz;

  781.             }

  782. 

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

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

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

  786.     if (!nnz_total)

  787.         mb->skip = 1;

  788. }

  789. 

  790. static av_always_inline

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

  792.                       int linesize, int uvlinesize, int simple)

  793. {

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

  795.     if (!simple) {

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

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

  798.     }

  799. }

  800. 

  801. static av_always_inline

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

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

  804.                     int simple, int xchg)

  805. {

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

  807.     src_y  -=   linesize;

  808.     src_cb -= uvlinesize;

  809.     src_cr -= uvlinesize;

  810. 

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

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

  813.         else      AV_COPY64(b,a);               \

  814.     } while (0)

  815. 

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

  817.     XCHG(top_border,      src_y,   xchg);

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

  819.     if (mb_x < mb_width-1)

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

  821. 

  822.     // only copy chroma for normal loop filter

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

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

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

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

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

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

  829.     }

  830. }

  831. 

  832. static av_always_inline

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

  834. {

  835.     if (!mb_x) {

  836.         return mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;

  837.     } else {

  838.         return mb_y ? mode : LEFT_DC_PRED8x8;

  839.     }

  840. }

  841. 

  842. static av_always_inline

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

  844. {

  845.     if (!mb_x) {

  846.         return mb_y ? VERT_PRED8x8 : DC_129_PRED8x8;

  847.     } else {

  848.         return mb_y ? mode : HOR_PRED8x8;

  849.     }

  850. }

  851. 

  852. static av_always_inline

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

  854. {

  855.     if (mode == DC_PRED8x8) {

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

  857.     } else {

  858.         return mode;

  859.     }

  860. }

  861. 

  862. static av_always_inline

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

  864. {

  865.     switch (mode) {

  866.     case DC_PRED8x8:

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

  868.     case VERT_PRED8x8:

  869.         return !mb_y ? DC_127_PRED8x8 : mode;

  870.     case HOR_PRED8x8:

  871.         return !mb_x ? DC_129_PRED8x8 : mode;

  872.     case PLANE_PRED8x8 /*TM*/:

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

  874.     }

  875.     return mode;

  876. }

  877. 

  878. static av_always_inline

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

  880. {

  881.     if (!mb_x) {

  882.         return mb_y ? VERT_VP8_PRED : DC_129_PRED;

  883.     } else {

  884.         return mb_y ? mode : HOR_VP8_PRED;

  885.     }

  886. }

  887. 

  888. static av_always_inline

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

  890. {

  891.     switch (mode) {

  892.     case VERT_PRED:

  893.         if (!mb_x && mb_y) {

  894.             *copy_buf = 1;

  895.             return mode;

  896.         }

  897.         /* fall-through */

  898.     case DIAG_DOWN_LEFT_PRED:

  899.     case VERT_LEFT_PRED:

  900.         return !mb_y ? DC_127_PRED : mode;

  901.     case HOR_PRED:

  902.         if (!mb_y) {

  903.             *copy_buf = 1;

  904.             return mode;

  905.         }

  906.         /* fall-through */

  907.     case HOR_UP_PRED:

  908.         return !mb_x ? DC_129_PRED : mode;

  909.     case TM_VP8_PRED:

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

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

  912.     case DIAG_DOWN_RIGHT_PRED:

  913.     case VERT_RIGHT_PRED:

  914.     case HOR_DOWN_PRED:

  915.         if (!mb_y || !mb_x)

  916.             *copy_buf = 1;

  917.         return mode;

  918.     }

  919.     return mode;

  920. }

  921. 

  922. static av_always_inline

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

  924.                    int mb_x, int mb_y)

  925. {

  926.     AVCodecContext *avctx = s->avctx;

  927.     int x, y, mode, nnz;

  928.     uint32_t tr;

  929. 

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

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

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

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

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

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

  936. 

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

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

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

  940.         } else {

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

  942.         }

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

  944.     } else {

  945.         uint8_t *ptr = dst[0];

  946.         uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;

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

  948. 

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

  950.         // the top macroblock for their topright edge

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

  952. 

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

  954.         // from the top macroblock

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

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

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

  958.             tr_right = (uint8_t *)&tr;

  959.         }

  960. 

  961.         if (mb->skip)

  962.             AV_ZERO128(s->non_zero_count_cache);

  963. 

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

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

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

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

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

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

  970. 

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

  972.                     topright = tr_top;

  973.                 } else if (x == 3)

  974.                     topright = tr_right;

  975. 

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

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

  978.                     if (copy) {

  979.                         dst = copy_dst + 12;

  980.                         linesize = 8;

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

  982.                             copy_dst[3] = 127U;

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

  984.                         } else {

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

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

  987.                                 copy_dst[3] = 129U;

  988.                             } else {

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

  990.                             }

  991.                         }

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

  993.                             copy_dst[11] =

  994.                             copy_dst[19] =

  995.                             copy_dst[27] =

  996.                             copy_dst[35] = 129U;

  997.                         } else {

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

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

  1000.                             copy_dst[27] = ptr[4*x+s->linesize*2-1];

  1001.                             copy_dst[35] = ptr[4*x+s->linesize*3-1];

  1002.                         }

  1003.                     }

  1004.                 } else {

  1005.                     mode = intra4x4[x];

  1006.                 }

  1007.                 s->hpc.pred4x4[mode](dst, topright, linesize);

  1008.                 if (copy) {

  1009.                     AV_COPY32(ptr+4*x              , copy_dst+12);

  1010.                     AV_COPY32(ptr+4*x+s->linesize  , copy_dst+20);

  1011.                     AV_COPY32(ptr+4*x+s->linesize*2, copy_dst+28);

  1012.                     AV_COPY32(ptr+4*x+s->linesize*3, copy_dst+36);

  1013.                 }

  1014. 

  1015.                 nnz = s->non_zero_count_cache[y][x];

  1016.                 if (nnz) {

  1017.                     if (nnz == 1)

  1018.                         s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);

  1019.                     else

  1020.                         s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);

  1021.                 }

  1022.                 topright += 4;

  1023.             }

  1024. 

  1025.             ptr   += 4*s->linesize;

  1026.             intra4x4 += 4;

  1027.         }

  1028.     }

  1029. 

  1030.     if (avctx->flags & CODEC_FLAG_EMU_EDGE) {

  1031.         mode = check_intra_pred8x8_mode_emuedge(s->chroma_pred_mode, mb_x, mb_y);

  1032.     } else {

  1033.         mode = check_intra_pred8x8_mode(s->chroma_pred_mode, mb_x, mb_y);

  1034.     }

  1035.     s->hpc.pred8x8[mode](dst[1], s->uvlinesize);

  1036.     s->hpc.pred8x8[mode](dst[2], s->uvlinesize);

  1037. 

  1038.     if (!(avctx->flags & CODEC_FLAG_EMU_EDGE && !mb_y) && (s->deblock_filter || !mb_y))

  1039.         xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],

  1040.                        s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,

  1041.                        s->filter.simple, 0);

  1042. }

  1043. 

  1044. static const uint8_t subpel_idx[3][8] = {

  1045.     { 0, 1, 2, 1, 2, 1, 2, 1 }, // nr. of left extra pixels,

  1046.                                 // also function pointer index

  1047.     { 0, 3, 5, 3, 5, 3, 5, 3 }, // nr. of extra pixels required

  1048.     { 0, 2, 3, 2, 3, 2, 3, 2 }, // nr. of right extra pixels

  1049. };

  1050. 

  1051. /**

  1052.  * luma MC function

  1053.  *

  1054.  * @param s VP8 decoding context

  1055.  * @param dst target buffer for block data at block position

  1056.  * @param src reference picture buffer at origin (0, 0)

  1057.  * @param mv motion vector (relative to block position) to get pixel data from

  1058.  * @param x_off horizontal position of block from origin (0, 0)

  1059.  * @param y_off vertical position of block from origin (0, 0)

  1060.  * @param block_w width of block (16, 8 or 4)

  1061.  * @param block_h height of block (always same as block_w)

  1062.  * @param width width of src/dst plane data

  1063.  * @param height height of src/dst plane data

  1064.  * @param linesize size of a single line of plane data, including padding

  1065.  * @param mc_func motion compensation function pointers (bilinear or sixtap MC)

  1066.  */

  1067. static av_always_inline

  1068. void vp8_mc_luma(VP8Context *s, uint8_t *dst, AVFrame *ref, const VP56mv *mv,

  1069.                  int x_off, int y_off, int block_w, int block_h,

  1070.                  int width, int height, int linesize,

  1071.                  vp8_mc_func mc_func[3][3])

  1072. {

  1073.     uint8_t *src = ref->data[0];

  1074. 

  1075.     if (AV_RN32A(mv)) {

  1076. 

  1077.         int mx = (mv->x << 1)&7, mx_idx = subpel_idx[0][mx];

  1078.         int my = (mv->y << 1)&7, my_idx = subpel_idx[0][my];

  1079. 

  1080.         x_off += mv->x >> 2;

  1081.         y_off += mv->y >> 2;

  1082. 

  1083.         // edge emulation

  1084.         ff_thread_await_progress(ref, (3 + y_off + block_h + subpel_idx[2][my]) >> 4, 0);

  1085.         src += y_off * linesize + x_off;

  1086.         if (x_off < mx_idx || x_off >= width  - block_w - subpel_idx[2][mx] ||

  1087.             y_off < my_idx || y_off >= height - block_h - subpel_idx[2][my]) {

  1088.             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src - my_idx * linesize - mx_idx, linesize,

  1089.                                     block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],

  1090.                                     x_off - mx_idx, y_off - my_idx, width, height);

  1091.             src = s->edge_emu_buffer + mx_idx + linesize * my_idx;

  1092.         }

  1093.         mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);

  1094.     } else {

  1095.         ff_thread_await_progress(ref, (3 + y_off + block_h) >> 4, 0);

  1096.         mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);

  1097.     }

  1098. }

  1099. 

  1100. /**

  1101.  * chroma MC function

  1102.  *

  1103.  * @param s VP8 decoding context

  1104.  * @param dst1 target buffer for block data at block position (U plane)

  1105.  * @param dst2 target buffer for block data at block position (V plane)

  1106.  * @param ref reference picture buffer at origin (0, 0)

  1107.  * @param mv motion vector (relative to block position) to get pixel data from

  1108.  * @param x_off horizontal position of block from origin (0, 0)

  1109.  * @param y_off vertical position of block from origin (0, 0)

  1110.  * @param block_w width of block (16, 8 or 4)

  1111.  * @param block_h height of block (always same as block_w)

  1112.  * @param width width of src/dst plane data

  1113.  * @param height height of src/dst plane data

  1114.  * @param linesize size of a single line of plane data, including padding

  1115.  * @param mc_func motion compensation function pointers (bilinear or sixtap MC)

  1116.  */

  1117. static av_always_inline

  1118. void vp8_mc_chroma(VP8Context *s, uint8_t *dst1, uint8_t *dst2, AVFrame *ref,

  1119.                    const VP56mv *mv, int x_off, int y_off,

  1120.                    int block_w, int block_h, int width, int height, int linesize,

  1121.                    vp8_mc_func mc_func[3][3])

  1122. {

  1123.     uint8_t *src1 = ref->data[1], *src2 = ref->data[2];

  1124. 

  1125.     if (AV_RN32A(mv)) {

  1126.         int mx = mv->x&7, mx_idx = subpel_idx[0][mx];

  1127.         int my = mv->y&7, my_idx = subpel_idx[0][my];

  1128. 

  1129.         x_off += mv->x >> 3;

  1130.         y_off += mv->y >> 3;

  1131. 

  1132.         // edge emulation

  1133.         src1 += y_off * linesize + x_off;

  1134.         src2 += y_off * linesize + x_off;

  1135.         ff_thread_await_progress(ref, (3 + y_off + block_h + subpel_idx[2][my]) >> 3, 0);

  1136.         if (x_off < mx_idx || x_off >= width  - block_w - subpel_idx[2][mx] ||

  1137.             y_off < my_idx || y_off >= height - block_h - subpel_idx[2][my]) {

  1138.             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src1 - my_idx * linesize - mx_idx, linesize,

  1139.                                     block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],

  1140.                                     x_off - mx_idx, y_off - my_idx, width, height);

  1141.             src1 = s->edge_emu_buffer + mx_idx + linesize * my_idx;

  1142.             mc_func[my_idx][mx_idx](dst1, linesize, src1, linesize, block_h, mx, my);

  1143. 

  1144.             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src2 - my_idx * linesize - mx_idx, linesize,

  1145.                                     block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],

  1146.                                     x_off - mx_idx, y_off - my_idx, width, height);

  1147.             src2 = s->edge_emu_buffer + mx_idx + linesize * my_idx;

  1148.             mc_func[my_idx][mx_idx](dst2, linesize, src2, linesize, block_h, mx, my);

  1149.         } else {

  1150.             mc_func[my_idx][mx_idx](dst1, linesize, src1, linesize, block_h, mx, my);

  1151.             mc_func[my_idx][mx_idx](dst2, linesize, src2, linesize, block_h, mx, my);

  1152.         }

  1153.     } else {

  1154.         ff_thread_await_progress(ref, (3 + y_off + block_h) >> 3, 0);

  1155.         mc_func[0][0](dst1, linesize, src1 + y_off * linesize + x_off, linesize, block_h, 0, 0);

  1156.         mc_func[0][0](dst2, linesize, src2 + y_off * linesize + x_off, linesize, block_h, 0, 0);

  1157.     }

  1158. }

  1159. 

  1160. static av_always_inline

  1161. void vp8_mc_part(VP8Context *s, uint8_t *dst[3],

  1162.                  AVFrame *ref_frame, int x_off, int y_off,

  1163.                  int bx_off, int by_off,

  1164.                  int block_w, int block_h,

  1165.                  int width, int height, VP56mv *mv)

  1166. {

  1167.     VP56mv uvmv = *mv;

  1168. 

  1169.     /* Y */

  1170.     vp8_mc_luma(s, dst[0] + by_off * s->linesize + bx_off,

  1171.                 ref_frame, mv, x_off + bx_off, y_off + by_off,

  1172.                 block_w, block_h, width, height, s->linesize,

  1173.                 s->put_pixels_tab[block_w == 8]);

  1174. 

  1175.     /* U/V */

  1176.     if (s->profile == 3) {

  1177.         uvmv.x &= ~7;

  1178.         uvmv.y &= ~7;

  1179.     }

  1180.     x_off   >>= 1; y_off   >>= 1;

  1181.     bx_off  >>= 1; by_off  >>= 1;

  1182.     width   >>= 1; height  >>= 1;

  1183.     block_w >>= 1; block_h >>= 1;

  1184.     vp8_mc_chroma(s, dst[1] + by_off * s->uvlinesize + bx_off,

  1185.                   dst[2] + by_off * s->uvlinesize + bx_off, ref_frame,

  1186.                   &uvmv, x_off + bx_off, y_off + by_off,

  1187.                   block_w, block_h, width, height, s->uvlinesize,

  1188.                   s->put_pixels_tab[1 + (block_w == 4)]);

  1189. }

  1190. 

  1191. /* Fetch pixels for estimated mv 4 macroblocks ahead.

  1192.  * Optimized for 64-byte cache lines.  Inspired by ffh264 prefetch_motion. */

  1193. static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref)

  1194. {

  1195.     /* Don't prefetch refs that haven't been used very often this frame. */

  1196.     if (s->ref_count[ref-1] > (mb_xy >> 5)) {

  1197.         int x_off = mb_x << 4, y_off = mb_y << 4;

  1198.         int mx = (mb->mv.x>>2) + x_off + 8;

  1199.         int my = (mb->mv.y>>2) + y_off;

  1200.         uint8_t **src= s->framep[ref]->data;

  1201.         int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;

  1202.         /* For threading, a ff_thread_await_progress here might be useful, but

  1203.          * it actually slows down the decoder. Since a bad prefetch doesn't

  1204.          * generate bad decoder output, we don't run it here. */

  1205.         s->dsp.prefetch(src[0]+off, s->linesize, 4);

  1206.         off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;

  1207.         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);

  1208.     }

  1209. }

  1210. 

  1211. /**

  1212.  * Apply motion vectors to prediction buffer, chapter 18.

  1213.  */

  1214. static av_always_inline

  1215. void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,

  1216.                    int mb_x, int mb_y)

  1217. {

  1218.     int x_off = mb_x << 4, y_off = mb_y << 4;

  1219.     int width = 16*s->mb_width, height = 16*s->mb_height;

  1220.     AVFrame *ref = s->framep[mb->ref_frame];

  1221.     VP56mv *bmv = mb->bmv;

  1222. 

  1223.     switch (mb->partitioning) {

  1224.     case VP8_SPLITMVMODE_NONE:

  1225.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1226.                     0, 0, 16, 16, width, height, &mb->mv);

  1227.         break;

  1228.     case VP8_SPLITMVMODE_4x4: {

  1229.         int x, y;

  1230.         VP56mv uvmv;

  1231. 

  1232.         /* Y */

  1233.         for (y = 0; y < 4; y++) {

  1234.             for (x = 0; x < 4; x++) {

  1235.                 vp8_mc_luma(s, dst[0] + 4*y*s->linesize + x*4,

  1236.                             ref, &bmv[4*y + x],

  1237.                             4*x + x_off, 4*y + y_off, 4, 4,

  1238.                             width, height, s->linesize,

  1239.                             s->put_pixels_tab[2]);

  1240.             }

  1241.         }

  1242. 

  1243.         /* U/V */

  1244.         x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;

  1245.         for (y = 0; y < 2; y++) {

  1246.             for (x = 0; x < 2; x++) {

  1247.                 uvmv.x = mb->bmv[ 2*y    * 4 + 2*x  ].x +

  1248.                          mb->bmv[ 2*y    * 4 + 2*x+1].x +

  1249.                          mb->bmv[(2*y+1) * 4 + 2*x  ].x +

  1250.                          mb->bmv[(2*y+1) * 4 + 2*x+1].x;

  1251.                 uvmv.y = mb->bmv[ 2*y    * 4 + 2*x  ].y +

  1252.                          mb->bmv[ 2*y    * 4 + 2*x+1].y +

  1253.                          mb->bmv[(2*y+1) * 4 + 2*x  ].y +

  1254.                          mb->bmv[(2*y+1) * 4 + 2*x+1].y;

  1255.                 uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;

  1256.                 uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;

  1257.                 if (s->profile == 3) {

  1258.                     uvmv.x &= ~7;

  1259.                     uvmv.y &= ~7;

  1260.                 }

  1261.                 vp8_mc_chroma(s, dst[1] + 4*y*s->uvlinesize + x*4,

  1262.                               dst[2] + 4*y*s->uvlinesize + x*4, ref, &uvmv,

  1263.                               4*x + x_off, 4*y + y_off, 4, 4,

  1264.                               width, height, s->uvlinesize,

  1265.                               s->put_pixels_tab[2]);

  1266.             }

  1267.         }

  1268.         break;

  1269.     }

  1270.     case VP8_SPLITMVMODE_16x8:

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

  1272.                     0, 0, 16, 8, width, height, &bmv[0]);

  1273.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1274.                     0, 8, 16, 8, width, height, &bmv[1]);

  1275.         break;

  1276.     case VP8_SPLITMVMODE_8x16:

  1277.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1278.                     0, 0, 8, 16, width, height, &bmv[0]);

  1279.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1280.                     8, 0, 8, 16, width, height, &bmv[1]);

  1281.         break;

  1282.     case VP8_SPLITMVMODE_8x8:

  1283.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1284.                     0, 0, 8, 8, width, height, &bmv[0]);

  1285.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1286.                     8, 0, 8, 8, width, height, &bmv[1]);

  1287.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1288.                     0, 8, 8, 8, width, height, &bmv[2]);

  1289.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1290.                     8, 8, 8, 8, width, height, &bmv[3]);

  1291.         break;

  1292.     }

  1293. }

  1294. 

  1295. static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb)

  1296. {

  1297.     int x, y, ch;

  1298. 

  1299.     if (mb->mode != MODE_I4x4) {

  1300.         uint8_t *y_dst = dst[0];

  1301.         for (y = 0; y < 4; y++) {

  1302.             uint32_t nnz4 = AV_RL32(s->non_zero_count_cache[y]);

  1303.             if (nnz4) {

  1304.                 if (nnz4&~0x01010101) {

  1305.                     for (x = 0; x < 4; x++) {

  1306.                         if ((uint8_t)nnz4 == 1)

  1307.                             s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);

  1308.                         else if((uint8_t)nnz4 > 1)

  1309.                             s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);

  1310.                         nnz4 >>= 8;

  1311.                         if (!nnz4)

  1312.                             break;

  1313.                     }

  1314.                 } else {

  1315.                     s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize);

  1316.                 }

  1317.             }

  1318.             y_dst += 4*s->linesize;

  1319.         }

  1320.     }

  1321. 

  1322.     for (ch = 0; ch < 2; ch++) {

  1323.         uint32_t nnz4 = AV_RL32(s->non_zero_count_cache[4+ch]);

  1324.         if (nnz4) {

  1325.             uint8_t *ch_dst = dst[1+ch];

  1326.             if (nnz4&~0x01010101) {

  1327.                 for (y = 0; y < 2; y++) {

  1328.                     for (x = 0; x < 2; x++) {

  1329.                         if ((uint8_t)nnz4 == 1)

  1330.                             s->vp8dsp.vp8_idct_dc_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);

  1331.                         else if((uint8_t)nnz4 > 1)

  1332.                             s->vp8dsp.vp8_idct_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);

  1333.                         nnz4 >>= 8;

  1334.                         if (!nnz4)

  1335.                             goto chroma_idct_end;

  1336.                     }

  1337.                     ch_dst += 4*s->uvlinesize;

  1338.                 }

  1339.             } else {

  1340.                 s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize);

  1341.             }

  1342.         }

  1343. chroma_idct_end: ;

  1344.     }

  1345. }

  1346. 

  1347. static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )

  1348. {

  1349.     int interior_limit, filter_level;

  1350. 

  1351.     if (s->segmentation.enabled) {

  1352.         filter_level = s->segmentation.filter_level[s->segment];

  1353.         if (!s->segmentation.absolute_vals)

  1354.             filter_level += s->filter.level;

  1355.     } else

  1356.         filter_level = s->filter.level;

  1357. 

  1358.     if (s->lf_delta.enabled) {

  1359.         filter_level += s->lf_delta.ref[mb->ref_frame];

  1360.         filter_level += s->lf_delta.mode[mb->mode];

  1361.     }

  1362. 

  1363.     filter_level = av_clip_uintp2(filter_level, 6);

  1364. 

  1365.     interior_limit = filter_level;

  1366.     if (s->filter.sharpness) {

  1367.         interior_limit >>= (s->filter.sharpness + 3) >> 2;

  1368.         interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);

  1369.     }

  1370.     interior_limit = FFMAX(interior_limit, 1);

  1371. 

  1372.     f->filter_level = filter_level;

  1373.     f->inner_limit = interior_limit;

  1374.     f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;

  1375. }

  1376. 

  1377. static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)

  1378. {

  1379.     int mbedge_lim, bedge_lim, hev_thresh;

  1380.     int filter_level = f->filter_level;

  1381.     int inner_limit = f->inner_limit;

  1382.     int inner_filter = f->inner_filter;

  1383.     int linesize = s->linesize;

  1384.     int uvlinesize = s->uvlinesize;

  1385.     static const uint8_t hev_thresh_lut[2][64] = {

  1386.         { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1,

  1387.           2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,

  1388.           3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,

  1389.           3, 3, 3, 3 },

  1390.         { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1,

  1391.           1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,

  1392.           2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,

  1393.           2, 2, 2, 2 }

  1394.     };

  1395. 

  1396.     if (!filter_level)

  1397.         return;

  1398. 

  1399.      bedge_lim = 2*filter_level + inner_limit;

  1400.     mbedge_lim = bedge_lim + 4;

  1401. 

  1402.     hev_thresh = hev_thresh_lut[s->keyframe][filter_level];

  1403. 

  1404.     if (mb_x) {

  1405.         s->vp8dsp.vp8_h_loop_filter16y(dst[0],     linesize,

  1406.                                        mbedge_lim, inner_limit, hev_thresh);

  1407.         s->vp8dsp.vp8_h_loop_filter8uv(dst[1],     dst[2],      uvlinesize,

  1408.                                        mbedge_lim, inner_limit, hev_thresh);

  1409.     }

  1410. 

  1411.     if (inner_filter) {

  1412.         s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, linesize, bedge_lim,

  1413.                                              inner_limit, hev_thresh);

  1414.         s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, linesize, bedge_lim,

  1415.                                              inner_limit, hev_thresh);

  1416.         s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, linesize, bedge_lim,

  1417.                                              inner_limit, hev_thresh);

  1418.         s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4,

  1419.                                              uvlinesize,  bedge_lim,

  1420.                                              inner_limit, hev_thresh);

  1421.     }

  1422. 

  1423.     if (mb_y) {

  1424.         s->vp8dsp.vp8_v_loop_filter16y(dst[0],     linesize,

  1425.                                        mbedge_lim, inner_limit, hev_thresh);

  1426.         s->vp8dsp.vp8_v_loop_filter8uv(dst[1],     dst[2],      uvlinesize,

  1427.                                        mbedge_lim, inner_limit, hev_thresh);

  1428.     }

  1429. 

  1430.     if (inner_filter) {

  1431.         s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*linesize,

  1432.                                              linesize,    bedge_lim,

  1433.                                              inner_limit, hev_thresh);

  1434.         s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*linesize,

  1435.                                              linesize,    bedge_lim,

  1436.                                              inner_limit, hev_thresh);

  1437.         s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*linesize,

  1438.                                              linesize,    bedge_lim,

  1439.                                              inner_limit, hev_thresh);

  1440.         s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * uvlinesize,

  1441.                                              dst[2] + 4 * uvlinesize,

  1442.                                              uvlinesize,  bedge_lim,

  1443.                                              inner_limit, hev_thresh);

  1444.     }

  1445. }

  1446. 

  1447. static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)

  1448. {

  1449.     int mbedge_lim, bedge_lim;

  1450.     int filter_level = f->filter_level;

  1451.     int inner_limit = f->inner_limit;

  1452.     int inner_filter = f->inner_filter;

  1453.     int linesize = s->linesize;

  1454. 

  1455.     if (!filter_level)

  1456.         return;

  1457. 

  1458.      bedge_lim = 2*filter_level + inner_limit;

  1459.     mbedge_lim = bedge_lim + 4;

  1460. 

  1461.     if (mb_x)

  1462.         s->vp8dsp.vp8_h_loop_filter_simple(dst, linesize, mbedge_lim);

  1463.     if (inner_filter) {

  1464.         s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, linesize, bedge_lim);

  1465.         s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, linesize, bedge_lim);

  1466.         s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim);

  1467.     }

  1468. 

  1469.     if (mb_y)

  1470.         s->vp8dsp.vp8_v_loop_filter_simple(dst, linesize, mbedge_lim);

  1471.     if (inner_filter) {

  1472.         s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*linesize, linesize, bedge_lim);

  1473.         s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*linesize, linesize, bedge_lim);

  1474.         s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim);

  1475.     }

  1476. }

  1477. 

  1478. static void filter_mb_row(VP8Context *s, AVFrame *curframe, int mb_y)

  1479. {

  1480.     VP8FilterStrength *f = s->filter_strength;

  1481.     uint8_t *dst[3] = {

  1482.         curframe->data[0] + 16*mb_y*s->linesize,

  1483.         curframe->data[1] +  8*mb_y*s->uvlinesize,

  1484.         curframe->data[2] +  8*mb_y*s->uvlinesize

  1485.     };

  1486.     int mb_x;

  1487. 

  1488.     for (mb_x = 0; mb_x < s->mb_width; mb_x++) {

  1489.         backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);

  1490.         filter_mb(s, dst, f++, mb_x, mb_y);

  1491.         dst[0] += 16;

  1492.         dst[1] += 8;

  1493.         dst[2] += 8;

  1494.     }

  1495. }

  1496. 

  1497. static void filter_mb_row_simple(VP8Context *s, AVFrame *curframe, int mb_y)

  1498. {

  1499.     VP8FilterStrength *f = s->filter_strength;

  1500.     uint8_t *dst = curframe->data[0] + 16*mb_y*s->linesize;

  1501.     int mb_x;

  1502. 

  1503.     for (mb_x = 0; mb_x < s->mb_width; mb_x++) {

  1504.         backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1);

  1505.         filter_mb_simple(s, dst, f++, mb_x, mb_y);

  1506.         dst += 16;

  1507.     }

  1508. }

  1509. 

  1510. static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,

  1511.                             AVPacket *avpkt)

  1512. {

  1513.     VP8Context *s = avctx->priv_data;

  1514.     int ret, mb_x, mb_y, i, y, referenced;

  1515.     enum AVDiscard skip_thresh;

  1516.     AVFrame *av_uninit(curframe), *prev_frame = s->framep[VP56_FRAME_CURRENT];

  1517. 

  1518.     if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)

  1519.         return ret;

  1520. 

  1521.     referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT

  1522.                                 || s->update_altref == VP56_FRAME_CURRENT;

  1523. 

  1524.     skip_thresh = !referenced ? AVDISCARD_NONREF :

  1525.                     !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;

  1526. 

  1527.     if (avctx->skip_frame >= skip_thresh) {

  1528.         s->invisible = 1;

  1529.         goto skip_decode;

  1530.     }

  1531.     s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;

  1532. 

  1533.     // release no longer referenced frames

  1534.     for (i = 0; i < 5; i++)

  1535.         if (s->frames[i].data[0] &&

  1536.             &s->frames[i] != prev_frame &&

  1537.             &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&

  1538.             &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&

  1539.             &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])

  1540.             ff_thread_release_buffer(avctx, &s->frames[i]);

  1541. 

  1542.     // find a free buffer

  1543.     for (i = 0; i < 5; i++)

  1544.         if (&s->frames[i] != prev_frame &&

  1545.             &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&

  1546.             &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&

  1547.             &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {

  1548.             curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];

  1549.             break;

  1550.         }

  1551.     if (i == 5) {

  1552.         av_log(avctx, AV_LOG_FATAL, "Ran out of free frames!\n");

  1553.         abort();

  1554.     }

  1555.     if (curframe->data[0])

  1556.         ff_thread_release_buffer(avctx, curframe);

  1557. 

  1558.     curframe->key_frame = s->keyframe;

  1559.     curframe->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;

  1560.     curframe->reference = referenced ? 3 : 0;

  1561.     curframe->ref_index[0] = s->segmentation_map;

  1562.     if ((ret = ff_thread_get_buffer(avctx, curframe))) {

  1563.         av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");

  1564.         return ret;

  1565.     }

  1566. 

  1567.     // check if golden and altref are swapped

  1568.     if (s->update_altref != VP56_FRAME_NONE) {

  1569.         s->next_framep[VP56_FRAME_GOLDEN2]  = s->framep[s->update_altref];

  1570.     } else {

  1571.         s->next_framep[VP56_FRAME_GOLDEN2]  = s->framep[VP56_FRAME_GOLDEN2];

  1572.     }

  1573.     if (s->update_golden != VP56_FRAME_NONE) {

  1574.         s->next_framep[VP56_FRAME_GOLDEN]   = s->framep[s->update_golden];

  1575.     } else {

  1576.         s->next_framep[VP56_FRAME_GOLDEN]   = s->framep[VP56_FRAME_GOLDEN];

  1577.     }

  1578.     if (s->update_last) {

  1579.         s->next_framep[VP56_FRAME_PREVIOUS] = curframe;

  1580.     } else {

  1581.         s->next_framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_PREVIOUS];

  1582.     }

  1583.     s->next_framep[VP56_FRAME_CURRENT]      = curframe;

  1584. 

  1585.     ff_thread_finish_setup(avctx);

  1586. 

  1587.     // Given that arithmetic probabilities are updated every frame, it's quite likely

  1588.     // that the values we have on a random interframe are complete junk if we didn't

  1589.     // start decode on a keyframe. So just don't display anything rather than junk.

  1590.     if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||

  1591.                          !s->framep[VP56_FRAME_GOLDEN] ||

  1592.                          !s->framep[VP56_FRAME_GOLDEN2])) {

  1593.         av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");

  1594.         return AVERROR_INVALIDDATA;

  1595.     }

  1596. 

  1597.     s->linesize   = curframe->linesize[0];

  1598.     s->uvlinesize = curframe->linesize[1];

  1599. 

  1600.     if (!s->edge_emu_buffer)

  1601.         s->edge_emu_buffer = av_malloc(21*s->linesize);

  1602. 

  1603.     memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));

  1604. 

  1605.     /* Zero macroblock structures for top/top-left prediction from outside the frame. */

  1606.     memset(s->macroblocks + s->mb_height*2 - 1, 0, (s->mb_width+1)*sizeof(*s->macroblocks));

  1607. 

  1608.     // top edge of 127 for intra prediction

  1609.     if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {

  1610.         s->top_border[0][15] = s->top_border[0][23] = 127;

  1611.         memset(s->top_border[1]-1, 127, s->mb_width*sizeof(*s->top_border)+1);

  1612.     }

  1613.     memset(s->ref_count, 0, sizeof(s->ref_count));

  1614.     if (s->keyframe)

  1615.         memset(s->intra4x4_pred_mode_top, DC_PRED, s->mb_width*4);

  1616. 

  1617. #define MARGIN (16 << 2)

  1618.     s->mv_min.y = -MARGIN;

  1619.     s->mv_max.y = ((s->mb_height - 1) << 6) + MARGIN;

  1620. 

  1621.     for (mb_y = 0; mb_y < s->mb_height; mb_y++) {

  1622.         VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];

  1623.         VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;

  1624.         int mb_xy = mb_y*s->mb_width;

  1625.         uint8_t *dst[3] = {

  1626.             curframe->data[0] + 16*mb_y*s->linesize,

  1627.             curframe->data[1] +  8*mb_y*s->uvlinesize,

  1628.             curframe->data[2] +  8*mb_y*s->uvlinesize

  1629.         };

  1630. 

  1631.         memset(mb - 1, 0, sizeof(*mb));   // zero left macroblock

  1632.         memset(s->left_nnz, 0, sizeof(s->left_nnz));

  1633.         AV_WN32A(s->intra4x4_pred_mode_left, DC_PRED*0x01010101);

  1634. 

  1635.         // left edge of 129 for intra prediction

  1636.         if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {

  1637.             for (i = 0; i < 3; i++)

  1638.                 for (y = 0; y < 16>>!!i; y++)

  1639.                     dst[i][y*curframe->linesize[i]-1] = 129;

  1640.             if (mb_y == 1) // top left edge is also 129

  1641.                 s->top_border[0][15] = s->top_border[0][23] = s->top_border[0][31] = 129;

  1642.         }

  1643. 

  1644.         s->mv_min.x = -MARGIN;

  1645.         s->mv_max.x = ((s->mb_width  - 1) << 6) + MARGIN;

  1646.         if (prev_frame && s->segmentation.enabled && !s->segmentation.update_map)

  1647.             ff_thread_await_progress(prev_frame, mb_y, 0);

  1648. 

  1649.         for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {

  1650.             /* Prefetch the current frame, 4 MBs ahead */

  1651.             s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);

  1652.             s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);

  1653. 

  1654.             decode_mb_mode(s, mb, mb_x, mb_y, s->segmentation_map + mb_xy,

  1655.                            prev_frame ? prev_frame->ref_index[0] + mb_xy : NULL);

  1656. 

  1657.             prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS);

  1658. 

  1659.             if (!mb->skip)

  1660.                 decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);

  1661. 

  1662.             if (mb->mode <= MODE_I4x4)

  1663.                 intra_predict(s, dst, mb, mb_x, mb_y);

  1664.             else

  1665.                 inter_predict(s, dst, mb, mb_x, mb_y);

  1666. 

  1667.             prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN);

  1668. 

  1669.             if (!mb->skip) {

  1670.                 idct_mb(s, dst, mb);

  1671.             } else {

  1672.                 AV_ZERO64(s->left_nnz);

  1673.                 AV_WN64(s->top_nnz[mb_x], 0);   // array of 9, so unaligned

  1674. 

  1675.                 // Reset DC block predictors if they would exist if the mb had coefficients

  1676.                 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {

  1677.                     s->left_nnz[8]      = 0;

  1678.                     s->top_nnz[mb_x][8] = 0;

  1679.                 }

  1680.             }

  1681. 

  1682.             if (s->deblock_filter)

  1683.                 filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);

  1684. 

  1685.             prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2);

  1686. 

  1687.             dst[0] += 16;

  1688.             dst[1] += 8;

  1689.             dst[2] += 8;

  1690.             s->mv_min.x -= 64;

  1691.             s->mv_max.x -= 64;

  1692.         }

  1693.         if (s->deblock_filter) {

  1694.             if (s->filter.simple)

  1695.                 filter_mb_row_simple(s, curframe, mb_y);

  1696.             else

  1697.                 filter_mb_row(s, curframe, mb_y);

  1698.         }

  1699.         s->mv_min.y -= 64;

  1700.         s->mv_max.y -= 64;

  1701. 

  1702.         ff_thread_report_progress(curframe, mb_y, 0);

  1703.     }

  1704. 

  1705.     ff_thread_report_progress(curframe, INT_MAX, 0);

  1706. skip_decode:

  1707.     // if future frames don't use the updated probabilities,

  1708.     // reset them to the values we saved

  1709.     if (!s->update_probabilities)

  1710.         s->prob[0] = s->prob[1];

  1711. 

  1712.     memcpy(&s->framep[0], &s->next_framep[0], sizeof(s->framep[0]) * 4);

  1713. 

  1714.     if (!s->invisible) {

  1715.         *(AVFrame*)data = *curframe;

  1716.         *data_size = sizeof(AVFrame);

  1717.     }

  1718. 

  1719.     return avpkt->size;

  1720. }

  1721. 

  1722. static av_cold int vp8_decode_init(AVCodecContext *avctx)

  1723. {

  1724.     VP8Context *s = avctx->priv_data;

  1725. 

  1726.     s->avctx = avctx;

  1727.     avctx->pix_fmt = PIX_FMT_YUV420P;

  1728. 

  1729.     dsputil_init(&s->dsp, avctx);

  1730.     ff_h264_pred_init(&s->hpc, CODEC_ID_VP8, 8);

  1731.     ff_vp8dsp_init(&s->vp8dsp);

  1732. 

  1733.     return 0;

  1734. }

  1735. 

  1736. static av_cold int vp8_decode_free(AVCodecContext *avctx)

  1737. {

  1738.     vp8_decode_flush(avctx);

  1739.     return 0;

  1740. }

  1741. 

  1742. static av_cold int vp8_decode_init_thread_copy(AVCodecContext *avctx)

  1743. {

  1744.     VP8Context *s = avctx->priv_data;

  1745. 

  1746.     s->avctx = avctx;

  1747. 

  1748.     return 0;

  1749. }

  1750. 

  1751. #define REBASE(pic) \

  1752.     pic ? pic - &s_src->frames[0] + &s->frames[0] : NULL

  1753. 

  1754. static int vp8_decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src)

  1755. {

  1756.     VP8Context *s = dst->priv_data, *s_src = src->priv_data;

  1757. 

  1758.     if (s->macroblocks_base &&

  1759.         (s_src->mb_width != s->mb_width || s_src->mb_height != s->mb_height)) {

  1760.         free_buffers(s);

  1761.     }

  1762. 

  1763.     s->prob[0] = s_src->prob[!s_src->update_probabilities];

  1764.     s->segmentation = s_src->segmentation;

  1765.     s->lf_delta = s_src->lf_delta;

  1766.     memcpy(s->sign_bias, s_src->sign_bias, sizeof(s->sign_bias));

  1767. 

  1768.     memcpy(&s->frames, &s_src->frames, sizeof(s->frames));

  1769.     s->framep[0] = REBASE(s_src->next_framep[0]);

  1770.     s->framep[1] = REBASE(s_src->next_framep[1]);

  1771.     s->framep[2] = REBASE(s_src->next_framep[2]);

  1772.     s->framep[3] = REBASE(s_src->next_framep[3]);

  1773. 

  1774.     return 0;

  1775. }

  1776. 

  1777. AVCodec ff_vp8_decoder = {

  1778.     "vp8",

  1779.     AVMEDIA_TYPE_VIDEO,

  1780.     CODEC_ID_VP8,

  1781.     sizeof(VP8Context),

  1782.     vp8_decode_init,

  1783.     NULL,

  1784.     vp8_decode_free,

  1785.     vp8_decode_frame,

  1786.     CODEC_CAP_DR1 | CODEC_CAP_FRAME_THREADS,

  1787.     .flush = vp8_decode_flush,

  1788.     .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),

  1789.     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(vp8_decode_init_thread_copy),

  1790.     .update_thread_context = ONLY_IF_THREADS_ENABLED(vp8_decode_update_thread_context),

  1791. };