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.
33469 Commits
32 Branches
283MB
Tree: 4b4a6ebd68
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '4b4a6ebd68'
${ noResults }
FFmpeg/libavcodec/vp8.c

1781 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 vp8_decode_flush(AVCodecContext *avctx)

  37. {

  38.     VP8Context *s = avctx->priv_data;

  39.     int i;

  40. 

  41.     if (!avctx->is_copy) {

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

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

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

  45.     }

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

  47. 

  48.     av_freep(&s->macroblocks_base);

  49.     av_freep(&s->filter_strength);

  50.     av_freep(&s->intra4x4_pred_mode_top);

  51.     av_freep(&s->top_nnz);

  52.     av_freep(&s->edge_emu_buffer);

  53.     av_freep(&s->top_border);

  54.     av_freep(&s->segmentation_map);

  55. 

  56.     s->macroblocks        = NULL;

  57. }

  58. 

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

  60. {

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

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

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

  64.             return AVERROR_INVALIDDATA;

  65. 

  66.         vp8_decode_flush(s->avctx);

  67. 

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

  69.     }

  70. 

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

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

  73. 

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

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

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

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

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

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

  80. 

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

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

  83.         return AVERROR(ENOMEM);

  84. 

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

  86. 

  87.     return 0;

  88. }

  89. 

  90. static void parse_segment_info(VP8Context *s)

  91. {

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

  93.     int i;

  94. 

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

  96. 

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

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

  99. 

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

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

  102. 

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

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

  105.     }

  106.     if (s->segmentation.update_map)

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

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

  109. }

  110. 

  111. static void update_lf_deltas(VP8Context *s)

  112. {

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

  114.     int i;

  115. 

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

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

  118. 

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

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

  121. }

  122. 

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

  124. {

  125.     const uint8_t *sizes = buf;

  126.     int i;

  127. 

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

  129. 

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

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

  132.     if (buf_size < 0)

  133.         return -1;

  134. 

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

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

  137.         if (buf_size - size < 0)

  138.             return -1;

  139. 

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

  141.         buf      += size;

  142.         buf_size -= size;

  143.     }

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

  145. 

  146.     return 0;

  147. }

  148. 

  149. static void get_quants(VP8Context *s)

  150. {

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

  152.     int i, base_qi;

  153. 

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

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

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

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

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

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

  160. 

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

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

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

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

  165.                 base_qi += yac_qi;

  166.         } else

  167.             base_qi = yac_qi;

  168. 

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

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

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

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

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

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

  175. 

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

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

  178.     }

  179. }

  180. 

  181. /**

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

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

  184.  *

  185.  * Intra frames update all 3 references

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

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

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

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

  190.  *      0: no update

  191.  *      1: VP56_FRAME_PREVIOUS

  192.  *      2: update golden with altref, or update altref with golden

  193.  */

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

  195. {

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

  197. 

  198.     if (update)

  199.         return VP56_FRAME_CURRENT;

  200. 

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

  202.     case 1:

  203.         return VP56_FRAME_PREVIOUS;

  204.     case 2:

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

  206.     }

  207.     return VP56_FRAME_NONE;

  208. }

  209. 

  210. static void update_refs(VP8Context *s)

  211. {

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

  213. 

  214.     int update_golden = vp8_rac_get(c);

  215.     int update_altref = vp8_rac_get(c);

  216. 

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

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

  219. }

  220. 

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

  222. {

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

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

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

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

  227. 

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

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

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

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

  232.     buf      += 3;

  233.     buf_size -= 3;

  234. 

  235.     if (s->profile > 3)

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

  237. 

  238.     if (!s->profile)

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

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

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

  242. 

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

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

  245.         return AVERROR_INVALIDDATA;

  246.     }

  247. 

  248.     if (s->keyframe) {

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

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

  251.             return AVERROR_INVALIDDATA;

  252.         }

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

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

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

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

  257.         buf      += 7;

  258.         buf_size -= 7;

  259. 

  260.         if (hscale || vscale)

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

  262. 

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

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

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

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

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

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

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

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

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

  272.     }

  273. 

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

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

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

  277.             return ret;

  278.     }

  279. 

  280.     ff_vp56_init_range_decoder(c, buf, header_size);

  281.     buf      += header_size;

  282.     buf_size -= header_size;

  283. 

  284.     if (s->keyframe) {

  285.         if (vp8_rac_get(c))

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

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

  288.     }

  289. 

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

  291.         parse_segment_info(s);

  292.     else

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

  294. 

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

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

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

  298. 

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

  300.         if (vp8_rac_get(c))

  301.             update_lf_deltas(s);

  302. 

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

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

  305.         return AVERROR_INVALIDDATA;

  306.     }

  307. 

  308.     get_quants(s);

  309. 

  310.     if (!s->keyframe) {

  311.         update_refs(s);

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

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

  314.     }

  315. 

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

  317.     // make a copy of the current probabilities

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

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

  320. 

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

  322. 

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

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

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

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

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

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

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

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

  331.                     }

  332. 

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

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

  335. 

  336.     if (!s->keyframe) {

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

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

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

  340. 

  341.         if (vp8_rac_get(c))

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

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

  344.         if (vp8_rac_get(c))

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

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

  347. 

  348.         // 17.2 MV probability update

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

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

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

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

  353.     }

  354. 

  355.     return 0;

  356. }

  357. 

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

  359. {

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

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

  362. }

  363. 

  364. /**

  365.  * Motion vector coding, 17.1.

  366.  */

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

  368. {

  369.     int bit, x = 0;

  370. 

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

  372.         int i;

  373. 

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

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

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

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

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

  379.             x += 8;

  380.     } else {

  381.         // small_mvtree

  382.         const uint8_t *ps = p+2;

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

  384.         ps += 1 + 3*bit;

  385.         x  += 4*bit;

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

  387.         ps += 1 + bit;

  388.         x  += 2*bit;

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

  390.     }

  391. 

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

  393. }

  394. 

  395. static av_always_inline

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

  397. {

  398.     if (left == top)

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

  400.     if (!top)

  401.         return vp8_submv_prob[2];

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

  403. }

  404. 

  405. /**

  406.  * Split motion vector prediction, 16.4.

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

  408.  */

  409. static av_always_inline

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

  411. {

  412.     int part_idx;

  413.     int n, num;

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

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

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

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

  418.                   *mbsplits_cur, *firstidx;

  419.     VP56mv *top_mv  = top_mb->bmv;

  420.     VP56mv *left_mv = left_mb->bmv;

  421.     VP56mv *cur_mv  = mb->bmv;

  422. 

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

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

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

  426.         } else {

  427.             part_idx = VP8_SPLITMVMODE_8x8;

  428.         }

  429.     } else {

  430.         part_idx = VP8_SPLITMVMODE_4x4;

  431.     }

  432. 

  433.     num = vp8_mbsplit_count[part_idx];

  434.     mbsplits_cur = vp8_mbsplits[part_idx],

  435.     firstidx = vp8_mbfirstidx[part_idx];

  436.     mb->partitioning = part_idx;

  437. 

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

  439.         int k = firstidx[n];

  440.         uint32_t left, above;

  441.         const uint8_t *submv_prob;

  442. 

  443.         if (!(k & 3))

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

  445.         else

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

  447.         if (k <= 3)

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

  449.         else

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

  451. 

  452.         submv_prob = get_submv_prob(left, above);

  453. 

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

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

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

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

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

  459.                 } else {

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

  461.                 }

  462.             } else {

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

  464.             }

  465.         } else {

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

  467.         }

  468.     }

  469. 

  470.     return num;

  471. }

  472. 

  473. static av_always_inline

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

  475. {

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

  477.                                   mb - 1 /* left */,

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

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

  480.     enum { VP8_EDGE_TOP, VP8_EDGE_LEFT, VP8_EDGE_TOPLEFT };

  481.     int idx = CNT_ZERO;

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

  483.     int8_t *sign_bias = s->sign_bias;

  484.     VP56mv near_mv[4];

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

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

  487. 

  488.     AV_ZERO32(&near_mv[0]);

  489.     AV_ZERO32(&near_mv[1]);

  490. 

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

  492.     #define MV_EDGE_CHECK(n)\

  493.     {\

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

  495.         int edge_ref = edge->ref_frame;\

  496.         if (edge_ref != VP56_FRAME_CURRENT) {\

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

  498.             if (mv) {\

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

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

  501.                     mv = ~mv;\

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

  503.                 }\

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

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

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

  507.             } else\

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

  509.         }\

  510.     }

  511. 

  512.     MV_EDGE_CHECK(0)

  513.     MV_EDGE_CHECK(1)

  514.     MV_EDGE_CHECK(2)

  515. 

  516.     mb->partitioning = VP8_SPLITMVMODE_NONE;

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

  518.         mb->mode = VP8_MVMODE_MV;

  519. 

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

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

  522.             cnt[CNT_NEAREST] += 1;

  523. 

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

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

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

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

  528.         }

  529. 

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

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

  532. 

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

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

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

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

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

  538. 

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

  540.                     mb->mode = VP8_MVMODE_SPLIT;

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

  542.                 } else {

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

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

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

  546.                 }

  547.             } else {

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

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

  550.             }

  551.         } else {

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

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

  554.         }

  555.     } else {

  556.         mb->mode = VP8_MVMODE_ZERO;

  557.         AV_ZERO32(&mb->mv);

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

  559.     }

  560. }

  561. 

  562. static av_always_inline

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

  564.                            int mb_x, int keyframe)

  565. {

  566.     uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;

  567.     if (keyframe) {

  568.         int x, y;

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

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

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

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

  573.                 const uint8_t *ctx;

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

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

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

  577.                 intra4x4++;

  578.             }

  579.         }

  580.     } else {

  581.         int i;

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

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

  584.     }

  585. }

  586. 

  587. static av_always_inline

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

  589. {

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

  591. 

  592.     if (s->segmentation.update_map)

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

  594.     else

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

  596.     s->segment = *segment;

  597. 

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

  599. 

  600.     if (s->keyframe) {

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

  602. 

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

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

  605.         } else {

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

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

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

  609.         }

  610. 

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

  612.         mb->ref_frame = VP56_FRAME_CURRENT;

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

  614.         // inter MB, 16.2

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

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

  617.                 VP56_FRAME_GOLDEN2 /* altref */ : VP56_FRAME_GOLDEN;

  618.         else

  619.             mb->ref_frame = VP56_FRAME_PREVIOUS;

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

  621. 

  622.         // motion vectors, 16.3

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

  624.     } else {

  625.         // intra MB, 16.1

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

  627. 

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

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

  630. 

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

  632.         mb->ref_frame = VP56_FRAME_CURRENT;

  633.         mb->partitioning = VP8_SPLITMVMODE_NONE;

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

  635.     }

  636. }

  637. 

  638. #ifndef decode_block_coeffs_internal

  639. /**

  640.  * @param c arithmetic bitstream reader context

  641.  * @param block destination for block coefficients

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

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

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

  645.  * @return 0 if no coeffs were decoded

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

  647.  */

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

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

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

  651. {

  652.     goto skip_eob;

  653.     do {

  654.         int coeff;

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

  656.             return i;

  657. 

  658. skip_eob:

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

  660.             if (++i == 16)

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

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

  663.             goto skip_eob;

  664.         }

  665. 

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

  667.             coeff = 1;

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

  669.         } else {

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

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

  672.                 if (coeff)

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

  674.                 coeff += 2;

  675.             } else {

  676.                 // DCT_CAT*

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

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

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

  680.                     } else {                                    // DCT_CAT2

  681.                         coeff  = 7;

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

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

  684.                     }

  685.                 } else {    // DCT_CAT3 and up

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

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

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

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

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

  691.                 }

  692.             }

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

  694.         }

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

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

  697. 

  698.     return i;

  699. }

  700. #endif

  701. 

  702. /**

  703.  * @param c arithmetic bitstream reader context

  704.  * @param block destination for block coefficients

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

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

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

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

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

  710.  * @return 0 if no coeffs were decoded

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

  712.  */

  713. static av_always_inline

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

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

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

  717. {

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

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

  720.         return 0;

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

  722. }

  723. 

  724. static av_always_inline

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

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

  727. {

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

  729.     int nnz_pred, nnz, nnz_total = 0;

  730.     int segment = s->segment;

  731.     int block_dc = 0;

  732. 

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

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

  735. 

  736.         // decode DC values and do hadamard

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

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

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

  740.         if (nnz) {

  741.             nnz_total += nnz;

  742.             block_dc = 1;

  743.             if (nnz == 1)

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

  745.             else

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

  747.         }

  748.         luma_start = 1;

  749.         luma_ctx = 0;

  750.     }

  751. 

  752.     // luma blocks

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

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

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

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

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

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

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

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

  761.             nnz_total += nnz;

  762.         }

  763. 

  764.     // chroma blocks

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

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

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

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

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

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

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

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

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

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

  775.                 nnz_total += nnz;

  776.             }

  777. 

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

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

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

  781.     if (!nnz_total)

  782.         mb->skip = 1;

  783. }

  784. 

  785. static av_always_inline

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

  787.                       int linesize, int uvlinesize, int simple)

  788. {

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

  790.     if (!simple) {

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

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

  793.     }

  794. }

  795. 

  796. static av_always_inline

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

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

  799.                     int simple, int xchg)

  800. {

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

  802.     src_y  -=   linesize;

  803.     src_cb -= uvlinesize;

  804.     src_cr -= uvlinesize;

  805. 

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

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

  808.         else      AV_COPY64(b,a);               \

  809.     } while (0)

  810. 

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

  812.     XCHG(top_border,      src_y,   xchg);

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

  814.     if (mb_x < mb_width-1)

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

  816. 

  817.     // only copy chroma for normal loop filter

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

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

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

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

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

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

  824.     }

  825. }

  826. 

  827. static av_always_inline

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

  829. {

  830.     if (!mb_x) {

  831.         return mb_y ? TOP_DC_PRED8x8 : DC_128_PRED8x8;

  832.     } else {

  833.         return mb_y ? mode : LEFT_DC_PRED8x8;

  834.     }

  835. }

  836. 

  837. static av_always_inline

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

  839. {

  840.     if (!mb_x) {

  841.         return mb_y ? VERT_PRED8x8 : DC_129_PRED8x8;

  842.     } else {

  843.         return mb_y ? mode : HOR_PRED8x8;

  844.     }

  845. }

  846. 

  847. static av_always_inline

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

  849. {

  850.     if (mode == DC_PRED8x8) {

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

  852.     } else {

  853.         return mode;

  854.     }

  855. }

  856. 

  857. static av_always_inline

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

  859. {

  860.     switch (mode) {

  861.     case DC_PRED8x8:

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

  863.     case VERT_PRED8x8:

  864.         return !mb_y ? DC_127_PRED8x8 : mode;

  865.     case HOR_PRED8x8:

  866.         return !mb_x ? DC_129_PRED8x8 : mode;

  867.     case PLANE_PRED8x8 /*TM*/:

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

  869.     }

  870.     return mode;

  871. }

  872. 

  873. static av_always_inline

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

  875. {

  876.     if (!mb_x) {

  877.         return mb_y ? VERT_VP8_PRED : DC_129_PRED;

  878.     } else {

  879.         return mb_y ? mode : HOR_VP8_PRED;

  880.     }

  881. }

  882. 

  883. static av_always_inline

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

  885. {

  886.     switch (mode) {

  887.     case VERT_PRED:

  888.         if (!mb_x && mb_y) {

  889.             *copy_buf = 1;

  890.             return mode;

  891.         }

  892.         /* fall-through */

  893.     case DIAG_DOWN_LEFT_PRED:

  894.     case VERT_LEFT_PRED:

  895.         return !mb_y ? DC_127_PRED : mode;

  896.     case HOR_PRED:

  897.         if (!mb_y) {

  898.             *copy_buf = 1;

  899.             return mode;

  900.         }

  901.         /* fall-through */

  902.     case HOR_UP_PRED:

  903.         return !mb_x ? DC_129_PRED : mode;

  904.     case TM_VP8_PRED:

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

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

  907.     case DIAG_DOWN_RIGHT_PRED:

  908.     case VERT_RIGHT_PRED:

  909.     case HOR_DOWN_PRED:

  910.         if (!mb_y || !mb_x)

  911.             *copy_buf = 1;

  912.         return mode;

  913.     }

  914.     return mode;

  915. }

  916. 

  917. static av_always_inline

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

  919.                    int mb_x, int mb_y)

  920. {

  921.     AVCodecContext *avctx = s->avctx;

  922.     int x, y, mode, nnz;

  923.     uint32_t tr;

  924. 

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

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

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

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

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

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

  931. 

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

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

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

  935.         } else {

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

  937.         }

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

  939.     } else {

  940.         uint8_t *ptr = dst[0];

  941.         uint8_t *intra4x4 = s->intra4x4_pred_mode_mb;

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

  943. 

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

  945.         // the top macroblock for their topright edge

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

  947. 

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

  949.         // from the top macroblock

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

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

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

  953.             tr_right = (uint8_t *)&tr;

  954.         }

  955. 

  956.         if (mb->skip)

  957.             AV_ZERO128(s->non_zero_count_cache);

  958. 

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

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

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

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

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

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

  965. 

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

  967.                     topright = tr_top;

  968.                 } else if (x == 3)

  969.                     topright = tr_right;

  970. 

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

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

  973.                     if (copy) {

  974.                         dst = copy_dst + 12;

  975.                         linesize = 8;

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

  977.                             copy_dst[3] = 127U;

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

  979.                         } else {

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

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

  982.                                 copy_dst[3] = 129U;

  983.                             } else {

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

  985.                             }

  986.                         }

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

  988.                             copy_dst[11] =

  989.                             copy_dst[19] =

  990.                             copy_dst[27] =

  991.                             copy_dst[35] = 129U;

  992.                         } else {

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

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

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

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

  997.                         }

  998.                     }

  999.                 } else {

  1000.                     mode = intra4x4[x];

  1001.                 }

  1002.                 s->hpc.pred4x4[mode](dst, topright, linesize);

  1003.                 if (copy) {

  1004.                     AV_COPY32(ptr+4*x              , copy_dst+12);

  1005.                     AV_COPY32(ptr+4*x+s->linesize  , copy_dst+20);

  1006.                     AV_COPY32(ptr+4*x+s->linesize*2, copy_dst+28);

  1007.                     AV_COPY32(ptr+4*x+s->linesize*3, copy_dst+36);

  1008.                 }

  1009. 

  1010.                 nnz = s->non_zero_count_cache[y][x];

  1011.                 if (nnz) {

  1012.                     if (nnz == 1)

  1013.                         s->vp8dsp.vp8_idct_dc_add(ptr+4*x, s->block[y][x], s->linesize);

  1014.                     else

  1015.                         s->vp8dsp.vp8_idct_add(ptr+4*x, s->block[y][x], s->linesize);

  1016.                 }

  1017.                 topright += 4;

  1018.             }

  1019. 

  1020.             ptr   += 4*s->linesize;

  1021.             intra4x4 += 4;

  1022.         }

  1023.     }

  1024. 

  1025.     if (avctx->flags & CODEC_FLAG_EMU_EDGE) {

  1026.         mode = check_intra_pred8x8_mode_emuedge(s->chroma_pred_mode, mb_x, mb_y);

  1027.     } else {

  1028.         mode = check_intra_pred8x8_mode(s->chroma_pred_mode, mb_x, mb_y);

  1029.     }

  1030.     s->hpc.pred8x8[mode](dst[1], s->uvlinesize);

  1031.     s->hpc.pred8x8[mode](dst[2], s->uvlinesize);

  1032. 

  1033.     if (!(avctx->flags & CODEC_FLAG_EMU_EDGE && !mb_y) && (s->deblock_filter || !mb_y))

  1034.         xchg_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2],

  1035.                        s->linesize, s->uvlinesize, mb_x, mb_y, s->mb_width,

  1036.                        s->filter.simple, 0);

  1037. }

  1038. 

  1039. static const uint8_t subpel_idx[3][8] = {

  1040.     { 0, 1, 2, 1, 2, 1, 2, 1 }, // nr. of left extra pixels,

  1041.                                 // also function pointer index

  1042.     { 0, 3, 5, 3, 5, 3, 5, 3 }, // nr. of extra pixels required

  1043.     { 0, 2, 3, 2, 3, 2, 3, 2 }, // nr. of right extra pixels

  1044. };

  1045. 

  1046. /**

  1047.  * luma MC function

  1048.  *

  1049.  * @param s VP8 decoding context

  1050.  * @param dst target buffer for block data at block position

  1051.  * @param ref reference picture buffer at origin (0, 0)

  1052.  * @param mv motion vector (relative to block position) to get pixel data from

  1053.  * @param x_off horizontal position of block from origin (0, 0)

  1054.  * @param y_off vertical position of block from origin (0, 0)

  1055.  * @param block_w width of block (16, 8 or 4)

  1056.  * @param block_h height of block (always same as block_w)

  1057.  * @param width width of src/dst plane data

  1058.  * @param height height of src/dst plane data

  1059.  * @param linesize size of a single line of plane data, including padding

  1060.  * @param mc_func motion compensation function pointers (bilinear or sixtap MC)

  1061.  */

  1062. static av_always_inline

  1063. void vp8_mc_luma(VP8Context *s, uint8_t *dst, AVFrame *ref, const VP56mv *mv,

  1064.                  int x_off, int y_off, int block_w, int block_h,

  1065.                  int width, int height, int linesize,

  1066.                  vp8_mc_func mc_func[3][3])

  1067. {

  1068.     uint8_t *src = ref->data[0];

  1069. 

  1070.     if (AV_RN32A(mv)) {

  1071. 

  1072.         int mx = (mv->x << 1)&7, mx_idx = subpel_idx[0][mx];

  1073.         int my = (mv->y << 1)&7, my_idx = subpel_idx[0][my];

  1074. 

  1075.         x_off += mv->x >> 2;

  1076.         y_off += mv->y >> 2;

  1077. 

  1078.         // edge emulation

  1079.         ff_thread_await_progress(ref, (3 + y_off + block_h + subpel_idx[2][my]) >> 4, 0);

  1080.         src += y_off * linesize + x_off;

  1081.         if (x_off < mx_idx || x_off >= width  - block_w - subpel_idx[2][mx] ||

  1082.             y_off < my_idx || y_off >= height - block_h - subpel_idx[2][my]) {

  1083.             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src - my_idx * linesize - mx_idx, linesize,

  1084.                                     block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],

  1085.                                     x_off - mx_idx, y_off - my_idx, width, height);

  1086.             src = s->edge_emu_buffer + mx_idx + linesize * my_idx;

  1087.         }

  1088.         mc_func[my_idx][mx_idx](dst, linesize, src, linesize, block_h, mx, my);

  1089.     } else {

  1090.         ff_thread_await_progress(ref, (3 + y_off + block_h) >> 4, 0);

  1091.         mc_func[0][0](dst, linesize, src + y_off * linesize + x_off, linesize, block_h, 0, 0);

  1092.     }

  1093. }

  1094. 

  1095. /**

  1096.  * chroma MC function

  1097.  *

  1098.  * @param s VP8 decoding context

  1099.  * @param dst1 target buffer for block data at block position (U plane)

  1100.  * @param dst2 target buffer for block data at block position (V plane)

  1101.  * @param ref reference picture buffer at origin (0, 0)

  1102.  * @param mv motion vector (relative to block position) to get pixel data from

  1103.  * @param x_off horizontal position of block from origin (0, 0)

  1104.  * @param y_off vertical position of block from origin (0, 0)

  1105.  * @param block_w width of block (16, 8 or 4)

  1106.  * @param block_h height of block (always same as block_w)

  1107.  * @param width width of src/dst plane data

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

  1109.  * @param linesize size of a single line of plane data, including padding

  1110.  * @param mc_func motion compensation function pointers (bilinear or sixtap MC)

  1111.  */

  1112. static av_always_inline

  1113. void vp8_mc_chroma(VP8Context *s, uint8_t *dst1, uint8_t *dst2, AVFrame *ref,

  1114.                    const VP56mv *mv, int x_off, int y_off,

  1115.                    int block_w, int block_h, int width, int height, int linesize,

  1116.                    vp8_mc_func mc_func[3][3])

  1117. {

  1118.     uint8_t *src1 = ref->data[1], *src2 = ref->data[2];

  1119. 

  1120.     if (AV_RN32A(mv)) {

  1121.         int mx = mv->x&7, mx_idx = subpel_idx[0][mx];

  1122.         int my = mv->y&7, my_idx = subpel_idx[0][my];

  1123. 

  1124.         x_off += mv->x >> 3;

  1125.         y_off += mv->y >> 3;

  1126. 

  1127.         // edge emulation

  1128.         src1 += y_off * linesize + x_off;

  1129.         src2 += y_off * linesize + x_off;

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

  1131.         if (x_off < mx_idx || x_off >= width  - block_w - subpel_idx[2][mx] ||

  1132.             y_off < my_idx || y_off >= height - block_h - subpel_idx[2][my]) {

  1133.             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src1 - my_idx * linesize - mx_idx, linesize,

  1134.                                     block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],

  1135.                                     x_off - mx_idx, y_off - my_idx, width, height);

  1136.             src1 = s->edge_emu_buffer + mx_idx + linesize * my_idx;

  1137.             mc_func[my_idx][mx_idx](dst1, linesize, src1, linesize, block_h, mx, my);

  1138. 

  1139.             s->dsp.emulated_edge_mc(s->edge_emu_buffer, src2 - my_idx * linesize - mx_idx, linesize,

  1140.                                     block_w + subpel_idx[1][mx], block_h + subpel_idx[1][my],

  1141.                                     x_off - mx_idx, y_off - my_idx, width, height);

  1142.             src2 = s->edge_emu_buffer + mx_idx + linesize * my_idx;

  1143.             mc_func[my_idx][mx_idx](dst2, linesize, src2, linesize, block_h, mx, my);

  1144.         } else {

  1145.             mc_func[my_idx][mx_idx](dst1, linesize, src1, linesize, block_h, mx, my);

  1146.             mc_func[my_idx][mx_idx](dst2, linesize, src2, linesize, block_h, mx, my);

  1147.         }

  1148.     } else {

  1149.         ff_thread_await_progress(ref, (3 + y_off + block_h) >> 3, 0);

  1150.         mc_func[0][0](dst1, linesize, src1 + y_off * linesize + x_off, linesize, block_h, 0, 0);

  1151.         mc_func[0][0](dst2, linesize, src2 + y_off * linesize + x_off, linesize, block_h, 0, 0);

  1152.     }

  1153. }

  1154. 

  1155. static av_always_inline

  1156. void vp8_mc_part(VP8Context *s, uint8_t *dst[3],

  1157.                  AVFrame *ref_frame, int x_off, int y_off,

  1158.                  int bx_off, int by_off,

  1159.                  int block_w, int block_h,

  1160.                  int width, int height, VP56mv *mv)

  1161. {

  1162.     VP56mv uvmv = *mv;

  1163. 

  1164.     /* Y */

  1165.     vp8_mc_luma(s, dst[0] + by_off * s->linesize + bx_off,

  1166.                 ref_frame, mv, x_off + bx_off, y_off + by_off,

  1167.                 block_w, block_h, width, height, s->linesize,

  1168.                 s->put_pixels_tab[block_w == 8]);

  1169. 

  1170.     /* U/V */

  1171.     if (s->profile == 3) {

  1172.         uvmv.x &= ~7;

  1173.         uvmv.y &= ~7;

  1174.     }

  1175.     x_off   >>= 1; y_off   >>= 1;

  1176.     bx_off  >>= 1; by_off  >>= 1;

  1177.     width   >>= 1; height  >>= 1;

  1178.     block_w >>= 1; block_h >>= 1;

  1179.     vp8_mc_chroma(s, dst[1] + by_off * s->uvlinesize + bx_off,

  1180.                   dst[2] + by_off * s->uvlinesize + bx_off, ref_frame,

  1181.                   &uvmv, x_off + bx_off, y_off + by_off,

  1182.                   block_w, block_h, width, height, s->uvlinesize,

  1183.                   s->put_pixels_tab[1 + (block_w == 4)]);

  1184. }

  1185. 

  1186. /* Fetch pixels for estimated mv 4 macroblocks ahead.

  1187.  * Optimized for 64-byte cache lines.  Inspired by ffh264 prefetch_motion. */

  1188. static av_always_inline void prefetch_motion(VP8Context *s, VP8Macroblock *mb, int mb_x, int mb_y, int mb_xy, int ref)

  1189. {

  1190.     /* Don't prefetch refs that haven't been used very often this frame. */

  1191.     if (s->ref_count[ref-1] > (mb_xy >> 5)) {

  1192.         int x_off = mb_x << 4, y_off = mb_y << 4;

  1193.         int mx = (mb->mv.x>>2) + x_off + 8;

  1194.         int my = (mb->mv.y>>2) + y_off;

  1195.         uint8_t **src= s->framep[ref]->data;

  1196.         int off= mx + (my + (mb_x&3)*4)*s->linesize + 64;

  1197.         /* For threading, a ff_thread_await_progress here might be useful, but

  1198.          * it actually slows down the decoder. Since a bad prefetch doesn't

  1199.          * generate bad decoder output, we don't run it here. */

  1200.         s->dsp.prefetch(src[0]+off, s->linesize, 4);

  1201.         off= (mx>>1) + ((my>>1) + (mb_x&7))*s->uvlinesize + 64;

  1202.         s->dsp.prefetch(src[1]+off, src[2]-src[1], 2);

  1203.     }

  1204. }

  1205. 

  1206. /**

  1207.  * Apply motion vectors to prediction buffer, chapter 18.

  1208.  */

  1209. static av_always_inline

  1210. void inter_predict(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb,

  1211.                    int mb_x, int mb_y)

  1212. {

  1213.     int x_off = mb_x << 4, y_off = mb_y << 4;

  1214.     int width = 16*s->mb_width, height = 16*s->mb_height;

  1215.     AVFrame *ref = s->framep[mb->ref_frame];

  1216.     VP56mv *bmv = mb->bmv;

  1217. 

  1218.     switch (mb->partitioning) {

  1219.     case VP8_SPLITMVMODE_NONE:

  1220.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1221.                     0, 0, 16, 16, width, height, &mb->mv);

  1222.         break;

  1223.     case VP8_SPLITMVMODE_4x4: {

  1224.         int x, y;

  1225.         VP56mv uvmv;

  1226. 

  1227.         /* Y */

  1228.         for (y = 0; y < 4; y++) {

  1229.             for (x = 0; x < 4; x++) {

  1230.                 vp8_mc_luma(s, dst[0] + 4*y*s->linesize + x*4,

  1231.                             ref, &bmv[4*y + x],

  1232.                             4*x + x_off, 4*y + y_off, 4, 4,

  1233.                             width, height, s->linesize,

  1234.                             s->put_pixels_tab[2]);

  1235.             }

  1236.         }

  1237. 

  1238.         /* U/V */

  1239.         x_off >>= 1; y_off >>= 1; width >>= 1; height >>= 1;

  1240.         for (y = 0; y < 2; y++) {

  1241.             for (x = 0; x < 2; x++) {

  1242.                 uvmv.x = mb->bmv[ 2*y    * 4 + 2*x  ].x +

  1243.                          mb->bmv[ 2*y    * 4 + 2*x+1].x +

  1244.                          mb->bmv[(2*y+1) * 4 + 2*x  ].x +

  1245.                          mb->bmv[(2*y+1) * 4 + 2*x+1].x;

  1246.                 uvmv.y = mb->bmv[ 2*y    * 4 + 2*x  ].y +

  1247.                          mb->bmv[ 2*y    * 4 + 2*x+1].y +

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

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

  1250.                 uvmv.x = (uvmv.x + 2 + (uvmv.x >> (INT_BIT-1))) >> 2;

  1251.                 uvmv.y = (uvmv.y + 2 + (uvmv.y >> (INT_BIT-1))) >> 2;

  1252.                 if (s->profile == 3) {

  1253.                     uvmv.x &= ~7;

  1254.                     uvmv.y &= ~7;

  1255.                 }

  1256.                 vp8_mc_chroma(s, dst[1] + 4*y*s->uvlinesize + x*4,

  1257.                               dst[2] + 4*y*s->uvlinesize + x*4, ref, &uvmv,

  1258.                               4*x + x_off, 4*y + y_off, 4, 4,

  1259.                               width, height, s->uvlinesize,

  1260.                               s->put_pixels_tab[2]);

  1261.             }

  1262.         }

  1263.         break;

  1264.     }

  1265.     case VP8_SPLITMVMODE_16x8:

  1266.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1267.                     0, 0, 16, 8, width, height, &bmv[0]);

  1268.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1269.                     0, 8, 16, 8, width, height, &bmv[1]);

  1270.         break;

  1271.     case VP8_SPLITMVMODE_8x16:

  1272.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1273.                     0, 0, 8, 16, width, height, &bmv[0]);

  1274.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1275.                     8, 0, 8, 16, width, height, &bmv[1]);

  1276.         break;

  1277.     case VP8_SPLITMVMODE_8x8:

  1278.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1279.                     0, 0, 8, 8, width, height, &bmv[0]);

  1280.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1281.                     8, 0, 8, 8, width, height, &bmv[1]);

  1282.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1283.                     0, 8, 8, 8, width, height, &bmv[2]);

  1284.         vp8_mc_part(s, dst, ref, x_off, y_off,

  1285.                     8, 8, 8, 8, width, height, &bmv[3]);

  1286.         break;

  1287.     }

  1288. }

  1289. 

  1290. static av_always_inline void idct_mb(VP8Context *s, uint8_t *dst[3], VP8Macroblock *mb)

  1291. {

  1292.     int x, y, ch;

  1293. 

  1294.     if (mb->mode != MODE_I4x4) {

  1295.         uint8_t *y_dst = dst[0];

  1296.         for (y = 0; y < 4; y++) {

  1297.             uint32_t nnz4 = AV_RL32(s->non_zero_count_cache[y]);

  1298.             if (nnz4) {

  1299.                 if (nnz4&~0x01010101) {

  1300.                     for (x = 0; x < 4; x++) {

  1301.                         if ((uint8_t)nnz4 == 1)

  1302.                             s->vp8dsp.vp8_idct_dc_add(y_dst+4*x, s->block[y][x], s->linesize);

  1303.                         else if((uint8_t)nnz4 > 1)

  1304.                             s->vp8dsp.vp8_idct_add(y_dst+4*x, s->block[y][x], s->linesize);

  1305.                         nnz4 >>= 8;

  1306.                         if (!nnz4)

  1307.                             break;

  1308.                     }

  1309.                 } else {

  1310.                     s->vp8dsp.vp8_idct_dc_add4y(y_dst, s->block[y], s->linesize);

  1311.                 }

  1312.             }

  1313.             y_dst += 4*s->linesize;

  1314.         }

  1315.     }

  1316. 

  1317.     for (ch = 0; ch < 2; ch++) {

  1318.         uint32_t nnz4 = AV_RL32(s->non_zero_count_cache[4+ch]);

  1319.         if (nnz4) {

  1320.             uint8_t *ch_dst = dst[1+ch];

  1321.             if (nnz4&~0x01010101) {

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

  1323.                     for (x = 0; x < 2; x++) {

  1324.                         if ((uint8_t)nnz4 == 1)

  1325.                             s->vp8dsp.vp8_idct_dc_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);

  1326.                         else if((uint8_t)nnz4 > 1)

  1327.                             s->vp8dsp.vp8_idct_add(ch_dst+4*x, s->block[4+ch][(y<<1)+x], s->uvlinesize);

  1328.                         nnz4 >>= 8;

  1329.                         if (!nnz4)

  1330.                             goto chroma_idct_end;

  1331.                     }

  1332.                     ch_dst += 4*s->uvlinesize;

  1333.                 }

  1334.             } else {

  1335.                 s->vp8dsp.vp8_idct_dc_add4uv(ch_dst, s->block[4+ch], s->uvlinesize);

  1336.             }

  1337.         }

  1338. chroma_idct_end: ;

  1339.     }

  1340. }

  1341. 

  1342. static av_always_inline void filter_level_for_mb(VP8Context *s, VP8Macroblock *mb, VP8FilterStrength *f )

  1343. {

  1344.     int interior_limit, filter_level;

  1345. 

  1346.     if (s->segmentation.enabled) {

  1347.         filter_level = s->segmentation.filter_level[s->segment];

  1348.         if (!s->segmentation.absolute_vals)

  1349.             filter_level += s->filter.level;

  1350.     } else

  1351.         filter_level = s->filter.level;

  1352. 

  1353.     if (s->lf_delta.enabled) {

  1354.         filter_level += s->lf_delta.ref[mb->ref_frame];

  1355.         filter_level += s->lf_delta.mode[mb->mode];

  1356.     }

  1357. 

  1358.     filter_level = av_clip_uintp2(filter_level, 6);

  1359. 

  1360.     interior_limit = filter_level;

  1361.     if (s->filter.sharpness) {

  1362.         interior_limit >>= (s->filter.sharpness + 3) >> 2;

  1363.         interior_limit = FFMIN(interior_limit, 9 - s->filter.sharpness);

  1364.     }

  1365.     interior_limit = FFMAX(interior_limit, 1);

  1366. 

  1367.     f->filter_level = filter_level;

  1368.     f->inner_limit = interior_limit;

  1369.     f->inner_filter = !mb->skip || mb->mode == MODE_I4x4 || mb->mode == VP8_MVMODE_SPLIT;

  1370. }

  1371. 

  1372. static av_always_inline void filter_mb(VP8Context *s, uint8_t *dst[3], VP8FilterStrength *f, int mb_x, int mb_y)

  1373. {

  1374.     int mbedge_lim, bedge_lim, hev_thresh;

  1375.     int filter_level = f->filter_level;

  1376.     int inner_limit = f->inner_limit;

  1377.     int inner_filter = f->inner_filter;

  1378.     int linesize = s->linesize;

  1379.     int uvlinesize = s->uvlinesize;

  1380.     static const uint8_t hev_thresh_lut[2][64] = {

  1381.         { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1,

  1382.           2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,

  1383.           3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3,

  1384.           3, 3, 3, 3 },

  1385.         { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1,

  1386.           1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 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.           2, 2, 2, 2 }

  1389.     };

  1390. 

  1391.     if (!filter_level)

  1392.         return;

  1393. 

  1394.      bedge_lim = 2*filter_level + inner_limit;

  1395.     mbedge_lim = bedge_lim + 4;

  1396. 

  1397.     hev_thresh = hev_thresh_lut[s->keyframe][filter_level];

  1398. 

  1399.     if (mb_x) {

  1400.         s->vp8dsp.vp8_h_loop_filter16y(dst[0],     linesize,

  1401.                                        mbedge_lim, inner_limit, hev_thresh);

  1402.         s->vp8dsp.vp8_h_loop_filter8uv(dst[1],     dst[2],      uvlinesize,

  1403.                                        mbedge_lim, inner_limit, hev_thresh);

  1404.     }

  1405. 

  1406.     if (inner_filter) {

  1407.         s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 4, linesize, bedge_lim,

  1408.                                              inner_limit, hev_thresh);

  1409.         s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+ 8, linesize, bedge_lim,

  1410.                                              inner_limit, hev_thresh);

  1411.         s->vp8dsp.vp8_h_loop_filter16y_inner(dst[0]+12, linesize, bedge_lim,

  1412.                                              inner_limit, hev_thresh);

  1413.         s->vp8dsp.vp8_h_loop_filter8uv_inner(dst[1] + 4, dst[2] + 4,

  1414.                                              uvlinesize,  bedge_lim,

  1415.                                              inner_limit, hev_thresh);

  1416.     }

  1417. 

  1418.     if (mb_y) {

  1419.         s->vp8dsp.vp8_v_loop_filter16y(dst[0],     linesize,

  1420.                                        mbedge_lim, inner_limit, hev_thresh);

  1421.         s->vp8dsp.vp8_v_loop_filter8uv(dst[1],     dst[2],      uvlinesize,

  1422.                                        mbedge_lim, inner_limit, hev_thresh);

  1423.     }

  1424. 

  1425.     if (inner_filter) {

  1426.         s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 4*linesize,

  1427.                                              linesize,    bedge_lim,

  1428.                                              inner_limit, hev_thresh);

  1429.         s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+ 8*linesize,

  1430.                                              linesize,    bedge_lim,

  1431.                                              inner_limit, hev_thresh);

  1432.         s->vp8dsp.vp8_v_loop_filter16y_inner(dst[0]+12*linesize,

  1433.                                              linesize,    bedge_lim,

  1434.                                              inner_limit, hev_thresh);

  1435.         s->vp8dsp.vp8_v_loop_filter8uv_inner(dst[1] + 4 * uvlinesize,

  1436.                                              dst[2] + 4 * uvlinesize,

  1437.                                              uvlinesize,  bedge_lim,

  1438.                                              inner_limit, hev_thresh);

  1439.     }

  1440. }

  1441. 

  1442. static av_always_inline void filter_mb_simple(VP8Context *s, uint8_t *dst, VP8FilterStrength *f, int mb_x, int mb_y)

  1443. {

  1444.     int mbedge_lim, bedge_lim;

  1445.     int filter_level = f->filter_level;

  1446.     int inner_limit = f->inner_limit;

  1447.     int inner_filter = f->inner_filter;

  1448.     int linesize = s->linesize;

  1449. 

  1450.     if (!filter_level)

  1451.         return;

  1452. 

  1453.      bedge_lim = 2*filter_level + inner_limit;

  1454.     mbedge_lim = bedge_lim + 4;

  1455. 

  1456.     if (mb_x)

  1457.         s->vp8dsp.vp8_h_loop_filter_simple(dst, linesize, mbedge_lim);

  1458.     if (inner_filter) {

  1459.         s->vp8dsp.vp8_h_loop_filter_simple(dst+ 4, linesize, bedge_lim);

  1460.         s->vp8dsp.vp8_h_loop_filter_simple(dst+ 8, linesize, bedge_lim);

  1461.         s->vp8dsp.vp8_h_loop_filter_simple(dst+12, linesize, bedge_lim);

  1462.     }

  1463. 

  1464.     if (mb_y)

  1465.         s->vp8dsp.vp8_v_loop_filter_simple(dst, linesize, mbedge_lim);

  1466.     if (inner_filter) {

  1467.         s->vp8dsp.vp8_v_loop_filter_simple(dst+ 4*linesize, linesize, bedge_lim);

  1468.         s->vp8dsp.vp8_v_loop_filter_simple(dst+ 8*linesize, linesize, bedge_lim);

  1469.         s->vp8dsp.vp8_v_loop_filter_simple(dst+12*linesize, linesize, bedge_lim);

  1470.     }

  1471. }

  1472. 

  1473. static void filter_mb_row(VP8Context *s, AVFrame *curframe, int mb_y)

  1474. {

  1475.     VP8FilterStrength *f = s->filter_strength;

  1476.     uint8_t *dst[3] = {

  1477.         curframe->data[0] + 16*mb_y*s->linesize,

  1478.         curframe->data[1] +  8*mb_y*s->uvlinesize,

  1479.         curframe->data[2] +  8*mb_y*s->uvlinesize

  1480.     };

  1481.     int mb_x;

  1482. 

  1483.     for (mb_x = 0; mb_x < s->mb_width; mb_x++) {

  1484.         backup_mb_border(s->top_border[mb_x+1], dst[0], dst[1], dst[2], s->linesize, s->uvlinesize, 0);

  1485.         filter_mb(s, dst, f++, mb_x, mb_y);

  1486.         dst[0] += 16;

  1487.         dst[1] += 8;

  1488.         dst[2] += 8;

  1489.     }

  1490. }

  1491. 

  1492. static void filter_mb_row_simple(VP8Context *s, AVFrame *curframe, int mb_y)

  1493. {

  1494.     VP8FilterStrength *f = s->filter_strength;

  1495.     uint8_t *dst = curframe->data[0] + 16*mb_y*s->linesize;

  1496.     int mb_x;

  1497. 

  1498.     for (mb_x = 0; mb_x < s->mb_width; mb_x++) {

  1499.         backup_mb_border(s->top_border[mb_x+1], dst, NULL, NULL, s->linesize, 0, 1);

  1500.         filter_mb_simple(s, dst, f++, mb_x, mb_y);

  1501.         dst += 16;

  1502.     }

  1503. }

  1504. 

  1505. static int vp8_decode_frame(AVCodecContext *avctx, void *data, int *data_size,

  1506.                             AVPacket *avpkt)

  1507. {

  1508.     VP8Context *s = avctx->priv_data;

  1509.     int ret, mb_x, mb_y, i, y, referenced;

  1510.     enum AVDiscard skip_thresh;

  1511.     AVFrame *av_uninit(curframe), *prev_frame = s->framep[VP56_FRAME_CURRENT];

  1512. 

  1513.     if ((ret = decode_frame_header(s, avpkt->data, avpkt->size)) < 0)

  1514.         return ret;

  1515. 

  1516.     referenced = s->update_last || s->update_golden == VP56_FRAME_CURRENT

  1517.                                 || s->update_altref == VP56_FRAME_CURRENT;

  1518. 

  1519.     skip_thresh = !referenced ? AVDISCARD_NONREF :

  1520.                     !s->keyframe ? AVDISCARD_NONKEY : AVDISCARD_ALL;

  1521. 

  1522.     if (avctx->skip_frame >= skip_thresh) {

  1523.         s->invisible = 1;

  1524.         goto skip_decode;

  1525.     }

  1526.     s->deblock_filter = s->filter.level && avctx->skip_loop_filter < skip_thresh;

  1527. 

  1528.     // release no longer referenced frames

  1529.     for (i = 0; i < 5; i++)

  1530.         if (s->frames[i].data[0] &&

  1531.             &s->frames[i] != prev_frame &&

  1532.             &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&

  1533.             &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&

  1534.             &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2])

  1535.             ff_thread_release_buffer(avctx, &s->frames[i]);

  1536. 

  1537.     // find a free buffer

  1538.     for (i = 0; i < 5; i++)

  1539.         if (&s->frames[i] != prev_frame &&

  1540.             &s->frames[i] != s->framep[VP56_FRAME_PREVIOUS] &&

  1541.             &s->frames[i] != s->framep[VP56_FRAME_GOLDEN] &&

  1542.             &s->frames[i] != s->framep[VP56_FRAME_GOLDEN2]) {

  1543.             curframe = s->framep[VP56_FRAME_CURRENT] = &s->frames[i];

  1544.             break;

  1545.         }

  1546.     if (i == 5) {

  1547.         av_log(avctx, AV_LOG_FATAL, "Ran out of free frames!\n");

  1548.         abort();

  1549.     }

  1550.     if (curframe->data[0])

  1551.         ff_thread_release_buffer(avctx, curframe);

  1552. 

  1553.     curframe->key_frame = s->keyframe;

  1554.     curframe->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;

  1555.     curframe->reference = referenced ? 3 : 0;

  1556.     curframe->ref_index[0] = s->segmentation_map;

  1557.     if ((ret = ff_thread_get_buffer(avctx, curframe))) {

  1558.         av_log(avctx, AV_LOG_ERROR, "get_buffer() failed!\n");

  1559.         return ret;

  1560.     }

  1561. 

  1562.     // check if golden and altref are swapped

  1563.     if (s->update_altref != VP56_FRAME_NONE) {

  1564.         s->next_framep[VP56_FRAME_GOLDEN2]  = s->framep[s->update_altref];

  1565.     } else {

  1566.         s->next_framep[VP56_FRAME_GOLDEN2]  = s->framep[VP56_FRAME_GOLDEN2];

  1567.     }

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

  1569.         s->next_framep[VP56_FRAME_GOLDEN]   = s->framep[s->update_golden];

  1570.     } else {

  1571.         s->next_framep[VP56_FRAME_GOLDEN]   = s->framep[VP56_FRAME_GOLDEN];

  1572.     }

  1573.     if (s->update_last) {

  1574.         s->next_framep[VP56_FRAME_PREVIOUS] = curframe;

  1575.     } else {

  1576.         s->next_framep[VP56_FRAME_PREVIOUS] = s->framep[VP56_FRAME_PREVIOUS];

  1577.     }

  1578.     s->next_framep[VP56_FRAME_CURRENT]      = curframe;

  1579. 

  1580.     ff_thread_finish_setup(avctx);

  1581. 

  1582.     // Given that arithmetic probabilities are updated every frame, it's quite likely

  1583.     // that the values we have on a random interframe are complete junk if we didn't

  1584.     // start decode on a keyframe. So just don't display anything rather than junk.

  1585.     if (!s->keyframe && (!s->framep[VP56_FRAME_PREVIOUS] ||

  1586.                          !s->framep[VP56_FRAME_GOLDEN] ||

  1587.                          !s->framep[VP56_FRAME_GOLDEN2])) {

  1588.         av_log(avctx, AV_LOG_WARNING, "Discarding interframe without a prior keyframe!\n");

  1589.         return AVERROR_INVALIDDATA;

  1590.     }

  1591. 

  1592.     s->linesize   = curframe->linesize[0];

  1593.     s->uvlinesize = curframe->linesize[1];

  1594. 

  1595.     if (!s->edge_emu_buffer)

  1596.         s->edge_emu_buffer = av_malloc(21*s->linesize);

  1597. 

  1598.     memset(s->top_nnz, 0, s->mb_width*sizeof(*s->top_nnz));

  1599. 

  1600.     /* Zero macroblock structures for top/top-left prediction from outside the frame. */

  1601.     memset(s->macroblocks + s->mb_height*2 - 1, 0, (s->mb_width+1)*sizeof(*s->macroblocks));

  1602. 

  1603.     // top edge of 127 for intra prediction

  1604.     if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {

  1605.         s->top_border[0][15] = s->top_border[0][23] = 127;

  1606.         memset(s->top_border[1]-1, 127, s->mb_width*sizeof(*s->top_border)+1);

  1607.     }

  1608.     memset(s->ref_count, 0, sizeof(s->ref_count));

  1609.     if (s->keyframe)

  1610.         memset(s->intra4x4_pred_mode_top, DC_PRED, s->mb_width*4);

  1611. 

  1612. #define MARGIN (16 << 2)

  1613.     s->mv_min.y = -MARGIN;

  1614.     s->mv_max.y = ((s->mb_height - 1) << 6) + MARGIN;

  1615. 

  1616.     for (mb_y = 0; mb_y < s->mb_height; mb_y++) {

  1617.         VP56RangeCoder *c = &s->coeff_partition[mb_y & (s->num_coeff_partitions-1)];

  1618.         VP8Macroblock *mb = s->macroblocks + (s->mb_height - mb_y - 1)*2;

  1619.         int mb_xy = mb_y*s->mb_width;

  1620.         uint8_t *dst[3] = {

  1621.             curframe->data[0] + 16*mb_y*s->linesize,

  1622.             curframe->data[1] +  8*mb_y*s->uvlinesize,

  1623.             curframe->data[2] +  8*mb_y*s->uvlinesize

  1624.         };

  1625. 

  1626.         memset(mb - 1, 0, sizeof(*mb));   // zero left macroblock

  1627.         memset(s->left_nnz, 0, sizeof(s->left_nnz));

  1628.         AV_WN32A(s->intra4x4_pred_mode_left, DC_PRED*0x01010101);

  1629. 

  1630.         // left edge of 129 for intra prediction

  1631.         if (!(avctx->flags & CODEC_FLAG_EMU_EDGE)) {

  1632.             for (i = 0; i < 3; i++)

  1633.                 for (y = 0; y < 16>>!!i; y++)

  1634.                     dst[i][y*curframe->linesize[i]-1] = 129;

  1635.             if (mb_y == 1) // top left edge is also 129

  1636.                 s->top_border[0][15] = s->top_border[0][23] = s->top_border[0][31] = 129;

  1637.         }

  1638. 

  1639.         s->mv_min.x = -MARGIN;

  1640.         s->mv_max.x = ((s->mb_width  - 1) << 6) + MARGIN;

  1641.         if (prev_frame && s->segmentation.enabled && !s->segmentation.update_map)

  1642.             ff_thread_await_progress(prev_frame, mb_y, 0);

  1643. 

  1644.         for (mb_x = 0; mb_x < s->mb_width; mb_x++, mb_xy++, mb++) {

  1645.             /* Prefetch the current frame, 4 MBs ahead */

  1646.             s->dsp.prefetch(dst[0] + (mb_x&3)*4*s->linesize + 64, s->linesize, 4);

  1647.             s->dsp.prefetch(dst[1] + (mb_x&7)*s->uvlinesize + 64, dst[2] - dst[1], 2);

  1648. 

  1649.             decode_mb_mode(s, mb, mb_x, mb_y, s->segmentation_map + mb_xy,

  1650.                            prev_frame ? prev_frame->ref_index[0] + mb_xy : NULL);

  1651. 

  1652.             prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_PREVIOUS);

  1653. 

  1654.             if (!mb->skip)

  1655.                 decode_mb_coeffs(s, c, mb, s->top_nnz[mb_x], s->left_nnz);

  1656. 

  1657.             if (mb->mode <= MODE_I4x4)

  1658.                 intra_predict(s, dst, mb, mb_x, mb_y);

  1659.             else

  1660.                 inter_predict(s, dst, mb, mb_x, mb_y);

  1661. 

  1662.             prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN);

  1663. 

  1664.             if (!mb->skip) {

  1665.                 idct_mb(s, dst, mb);

  1666.             } else {

  1667.                 AV_ZERO64(s->left_nnz);

  1668.                 AV_WN64(s->top_nnz[mb_x], 0);   // array of 9, so unaligned

  1669. 

  1670.                 // Reset DC block predictors if they would exist if the mb had coefficients

  1671.                 if (mb->mode != MODE_I4x4 && mb->mode != VP8_MVMODE_SPLIT) {

  1672.                     s->left_nnz[8]      = 0;

  1673.                     s->top_nnz[mb_x][8] = 0;

  1674.                 }

  1675.             }

  1676. 

  1677.             if (s->deblock_filter)

  1678.                 filter_level_for_mb(s, mb, &s->filter_strength[mb_x]);

  1679. 

  1680.             prefetch_motion(s, mb, mb_x, mb_y, mb_xy, VP56_FRAME_GOLDEN2);

  1681. 

  1682.             dst[0] += 16;

  1683.             dst[1] += 8;

  1684.             dst[2] += 8;

  1685.             s->mv_min.x -= 64;

  1686.             s->mv_max.x -= 64;

  1687.         }

  1688.         if (s->deblock_filter) {

  1689.             if (s->filter.simple)

  1690.                 filter_mb_row_simple(s, curframe, mb_y);

  1691.             else

  1692.                 filter_mb_row(s, curframe, mb_y);

  1693.         }

  1694.         s->mv_min.y -= 64;

  1695.         s->mv_max.y -= 64;

  1696. 

  1697.         ff_thread_report_progress(curframe, mb_y, 0);

  1698.     }

  1699. 

  1700.     ff_thread_report_progress(curframe, INT_MAX, 0);

  1701. skip_decode:

  1702.     // if future frames don't use the updated probabilities,

  1703.     // reset them to the values we saved

  1704.     if (!s->update_probabilities)

  1705.         s->prob[0] = s->prob[1];

  1706. 

  1707.     memcpy(&s->framep[0], &s->next_framep[0], sizeof(s->framep[0]) * 4);

  1708. 

  1709.     if (!s->invisible) {

  1710.         *(AVFrame*)data = *curframe;

  1711.         *data_size = sizeof(AVFrame);

  1712.     }

  1713. 

  1714.     return avpkt->size;

  1715. }

  1716. 

  1717. static av_cold int vp8_decode_init(AVCodecContext *avctx)

  1718. {

  1719.     VP8Context *s = avctx->priv_data;

  1720. 

  1721.     s->avctx = avctx;

  1722.     avctx->pix_fmt = PIX_FMT_YUV420P;

  1723. 

  1724.     dsputil_init(&s->dsp, avctx);

  1725.     ff_h264_pred_init(&s->hpc, CODEC_ID_VP8, 8, 1);

  1726.     ff_vp8dsp_init(&s->vp8dsp);

  1727. 

  1728.     return 0;

  1729. }

  1730. 

  1731. static av_cold int vp8_decode_free(AVCodecContext *avctx)

  1732. {

  1733.     vp8_decode_flush(avctx);

  1734.     return 0;

  1735. }

  1736. 

  1737. static av_cold int vp8_decode_init_thread_copy(AVCodecContext *avctx)

  1738. {

  1739.     VP8Context *s = avctx->priv_data;

  1740. 

  1741.     s->avctx = avctx;

  1742. 

  1743.     return 0;

  1744. }

  1745. 

  1746. #define REBASE(pic) \

  1747.     pic ? pic - &s_src->frames[0] + &s->frames[0] : NULL

  1748. 

  1749. static int vp8_decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src)

  1750. {

  1751.     VP8Context *s = dst->priv_data, *s_src = src->priv_data;

  1752. 

  1753.     s->prob[0] = s_src->prob[!s_src->update_probabilities];

  1754.     s->segmentation = s_src->segmentation;

  1755.     s->lf_delta = s_src->lf_delta;

  1756.     memcpy(s->sign_bias, s_src->sign_bias, sizeof(s->sign_bias));

  1757. 

  1758.     memcpy(&s->frames, &s_src->frames, sizeof(s->frames));

  1759.     s->framep[0] = REBASE(s_src->next_framep[0]);

  1760.     s->framep[1] = REBASE(s_src->next_framep[1]);

  1761.     s->framep[2] = REBASE(s_src->next_framep[2]);

  1762.     s->framep[3] = REBASE(s_src->next_framep[3]);

  1763. 

  1764.     return 0;

  1765. }

  1766. 

  1767. AVCodec ff_vp8_decoder = {

  1768.     .name           = "vp8",

  1769.     .type           = AVMEDIA_TYPE_VIDEO,

  1770.     .id             = CODEC_ID_VP8,

  1771.     .priv_data_size = sizeof(VP8Context),

  1772.     .init           = vp8_decode_init,

  1773.     .close          = vp8_decode_free,

  1774.     .decode         = vp8_decode_frame,

  1775.     .capabilities   = CODEC_CAP_DR1 | CODEC_CAP_FRAME_THREADS,

  1776.     .flush = vp8_decode_flush,

  1777.     .long_name = NULL_IF_CONFIG_SMALL("On2 VP8"),

  1778.     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(vp8_decode_init_thread_copy),

  1779.     .update_thread_context = ONLY_IF_THREADS_ENABLED(vp8_decode_update_thread_context),

  1780. };