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.
44689 Commits
32 Branches
824MB
Tree: c5714b51aa
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'c5714b51aa'
${ noResults }
FFmpeg/libavcodec/vp9block.c

1750 lines
74KB
Raw Blame History 

  1. /*

  2.  * VP9 compatible video decoder

  3.  *

  4.  * Copyright (C) 2013 Ronald S. Bultje <rsbultje gmail com>

  5.  * Copyright (C) 2013 Clément Bœsch <u pkh me>

  6.  *

  7.  * This file is part of Libav.

  8.  *

  9.  * Libav is free software; you can redistribute it and/or

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

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

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

  13.  *

  14.  * Libav is distributed in the hope that it will be useful,

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

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

  17.  * Lesser General Public License for more details.

  18.  *

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

  20.  * License along with Libav; if not, write to the Free Software

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

  22.  */

  23. 

  24. #include "libavutil/avassert.h"

  25. 

  26. #include "avcodec.h"

  27. #include "internal.h"

  28. #include "videodsp.h"

  29. #include "vp56.h"

  30. #include "vp9.h"

  31. #include "vp9data.h"

  32. 

  33. static const uint8_t bwh_tab[2][N_BS_SIZES][2] = {

  34.     {

  35.         { 16, 16 }, { 16, 8 }, { 8, 16 }, { 8, 8 }, { 8, 4 }, { 4, 8 },

  36.         {  4,  4 }, {  4, 2 }, { 2,  4 }, { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 },

  37.     },  {

  38.         {  8,  8 }, {  8, 4 }, { 4,  8 }, { 4, 4 }, { 4, 2 }, { 2, 4 },

  39.         {  2,  2 }, {  2, 1 }, { 1,  2 }, { 1, 1 }, { 1, 1 }, { 1, 1 }, { 1, 1 },

  40.     }

  41. };

  42. 

  43. // differential forward probability updates

  44. static void decode_mode(VP9Context *s, VP9Block *const b)

  45. {

  46.     static const uint8_t left_ctx[N_BS_SIZES] = {

  47.         0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf

  48.     };

  49.     static const uint8_t above_ctx[N_BS_SIZES] = {

  50.         0x0, 0x0, 0x8, 0x8, 0x8, 0xc, 0xc, 0xc, 0xe, 0xe, 0xe, 0xf, 0xf

  51.     };

  52.     static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {

  53.         TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,

  54.         TX_16X16, TX_8X8,   TX_8X8,   TX_8X8,   TX_4X4,   TX_4X4,  TX_4X4

  55.     };

  56.     int row = b->row, col = b->col, row7 = b->row7;

  57.     enum TxfmMode max_tx = max_tx_for_bl_bp[b->bs];

  58.     int w4 = FFMIN(s->cols - col, bwh_tab[1][b->bs][0]);

  59.     int h4 = FFMIN(s->rows - row, bwh_tab[1][b->bs][1]);

  60.     int have_a = row > 0, have_l = col > s->tiling.tile_col_start;

  61.     int y;

  62. 

  63.     if (!s->segmentation.enabled) {

  64.         b->seg_id = 0;

  65.     } else if (s->keyframe || s->intraonly) {

  66.         b->seg_id = s->segmentation.update_map ?

  67.                     vp8_rac_get_tree(&s->c, ff_vp9_segmentation_tree, s->prob.seg) : 0;

  68.     } else if (!s->segmentation.update_map ||

  69.                (s->segmentation.temporal &&

  70.                 vp56_rac_get_prob_branchy(&s->c,

  71.                                           s->prob.segpred[s->above_segpred_ctx[col] +

  72.                                                           s->left_segpred_ctx[row7]]))) {

  73.         if (!s->errorres) {

  74.             uint8_t *refsegmap = s->frames[LAST_FRAME].segmentation_map;

  75.             int pred = MAX_SEGMENT - 1;

  76.             int x;

  77. 

  78.             if (!s->last_uses_2pass)

  79.                 ff_thread_await_progress(&s->frames[LAST_FRAME].tf, row >> 3, 0);

  80. 

  81.             for (y = 0; y < h4; y++)

  82.                 for (x = 0; x < w4; x++)

  83.                     pred = FFMIN(pred,

  84.                                  refsegmap[(y + row) * 8 * s->sb_cols + x + col]);

  85.             b->seg_id = pred;

  86.         } else {

  87.             b->seg_id = 0;

  88.         }

  89. 

  90.         memset(&s->above_segpred_ctx[col], 1, w4);

  91.         memset(&s->left_segpred_ctx[row7], 1, h4);

  92.     } else {

  93.         b->seg_id = vp8_rac_get_tree(&s->c, ff_vp9_segmentation_tree,

  94.                                      s->prob.seg);

  95. 

  96.         memset(&s->above_segpred_ctx[col], 0, w4);

  97.         memset(&s->left_segpred_ctx[row7], 0, h4);

  98.     }

  99.     if ((s->segmentation.enabled && s->segmentation.update_map) || s->keyframe) {

  100.         uint8_t *segmap = s->frames[CUR_FRAME].segmentation_map;

  101. 

  102.         for (y = 0; y < h4; y++)

  103.             memset(&segmap[(y + row) * 8 * s->sb_cols + col],

  104.                    b->seg_id, w4);

  105.     }

  106. 

  107.     b->skip = s->segmentation.enabled &&

  108.               s->segmentation.feat[b->seg_id].skip_enabled;

  109.     if (!b->skip) {

  110.         int c = s->left_skip_ctx[row7] + s->above_skip_ctx[col];

  111.         b->skip = vp56_rac_get_prob(&s->c, s->prob.p.skip[c]);

  112.         s->counts.skip[c][b->skip]++;

  113.     }

  114. 

  115.     if (s->keyframe || s->intraonly) {

  116.         b->intra = 1;

  117.     } else if (s->segmentation.feat[b->seg_id].ref_enabled) {

  118.         b->intra = !s->segmentation.feat[b->seg_id].ref_val;

  119.     } else {

  120.         int c, bit;

  121. 

  122.         if (have_a && have_l) {

  123.             c  = s->above_intra_ctx[col] + s->left_intra_ctx[row7];

  124.             c += (c == 2);

  125.         } else {

  126.             c = have_a ? 2 * s->above_intra_ctx[col] :

  127.                 have_l ? 2 * s->left_intra_ctx[row7] : 0;

  128.         }

  129.         bit = vp56_rac_get_prob(&s->c, s->prob.p.intra[c]);

  130.         s->counts.intra[c][bit]++;

  131.         b->intra = !bit;

  132.     }

  133. 

  134.     if ((b->intra || !b->skip) && s->txfmmode == TX_SWITCHABLE) {

  135.         int c;

  136.         if (have_a) {

  137.             if (have_l) {

  138.                 c = (s->above_skip_ctx[col] ? max_tx :

  139.                      s->above_txfm_ctx[col]) +

  140.                     (s->left_skip_ctx[row7] ? max_tx :

  141.                      s->left_txfm_ctx[row7]) > max_tx;

  142.             } else {

  143.                 c = s->above_skip_ctx[col] ? 1 :

  144.                     (s->above_txfm_ctx[col] * 2 > max_tx);

  145.             }

  146.         } else if (have_l) {

  147.             c = s->left_skip_ctx[row7] ? 1 :

  148.                 (s->left_txfm_ctx[row7] * 2 > max_tx);

  149.         } else {

  150.             c = 1;

  151.         }

  152.         switch (max_tx) {

  153.         case TX_32X32:

  154.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][0]);

  155.             if (b->tx) {

  156.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][1]);

  157.                 if (b->tx == 2)

  158.                     b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][2]);

  159.             }

  160.             s->counts.tx32p[c][b->tx]++;

  161.             break;

  162.         case TX_16X16:

  163.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][0]);

  164.             if (b->tx)

  165.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][1]);

  166.             s->counts.tx16p[c][b->tx]++;

  167.             break;

  168.         case TX_8X8:

  169.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx8p[c]);

  170.             s->counts.tx8p[c][b->tx]++;

  171.             break;

  172.         case TX_4X4:

  173.             b->tx = TX_4X4;

  174.             break;

  175.         }

  176.     } else {

  177.         b->tx = FFMIN(max_tx, s->txfmmode);

  178.     }

  179. 

  180.     if (s->keyframe || s->intraonly) {

  181.         uint8_t *a = &s->above_mode_ctx[col * 2];

  182.         uint8_t *l = &s->left_mode_ctx[(row7) << 1];

  183. 

  184.         b->comp = 0;

  185.         if (b->bs > BS_8x8) {

  186.             // FIXME the memory storage intermediates here aren't really

  187.             // necessary, they're just there to make the code slightly

  188.             // simpler for now

  189.             b->mode[0] =

  190.             a[0]       = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,

  191.                                           ff_vp9_default_kf_ymode_probs[a[0]][l[0]]);

  192.             if (b->bs != BS_8x4) {

  193.                 b->mode[1] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,

  194.                                               ff_vp9_default_kf_ymode_probs[a[1]][b->mode[0]]);

  195.                 l[0]       =

  196.                 a[1]       = b->mode[1];

  197.             } else {

  198.                 l[0]       =

  199.                 a[1]       =

  200.                 b->mode[1] = b->mode[0];

  201.             }

  202.             if (b->bs != BS_4x8) {

  203.                 b->mode[2] =

  204.                 a[0]       = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,

  205.                                               ff_vp9_default_kf_ymode_probs[a[0]][l[1]]);

  206.                 if (b->bs != BS_8x4) {

  207.                     b->mode[3] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,

  208.                                                   ff_vp9_default_kf_ymode_probs[a[1]][b->mode[2]]);

  209.                     l[1]       =

  210.                     a[1]       = b->mode[3];

  211.                 } else {

  212.                     l[1]       =

  213.                     a[1]       =

  214.                     b->mode[3] = b->mode[2];

  215.                 }

  216.             } else {

  217.                 b->mode[2] = b->mode[0];

  218.                 l[1]       =

  219.                 a[1]       =

  220.                 b->mode[3] = b->mode[1];

  221.             }

  222.         } else {

  223.             b->mode[0] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,

  224.                                           ff_vp9_default_kf_ymode_probs[*a][*l]);

  225.             b->mode[3] =

  226.             b->mode[2] =

  227.             b->mode[1] = b->mode[0];

  228.             // FIXME this can probably be optimized

  229.             memset(a, b->mode[0], bwh_tab[0][b->bs][0]);

  230.             memset(l, b->mode[0], bwh_tab[0][b->bs][1]);

  231.         }

  232.         b->uvmode = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,

  233.                                      ff_vp9_default_kf_uvmode_probs[b->mode[3]]);

  234.     } else if (b->intra) {

  235.         b->comp = 0;

  236.         if (b->bs > BS_8x8) {

  237.             b->mode[0] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,

  238.                                           s->prob.p.y_mode[0]);

  239.             s->counts.y_mode[0][b->mode[0]]++;

  240.             if (b->bs != BS_8x4) {

  241.                 b->mode[1] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,

  242.                                               s->prob.p.y_mode[0]);

  243.                 s->counts.y_mode[0][b->mode[1]]++;

  244.             } else {

  245.                 b->mode[1] = b->mode[0];

  246.             }

  247.             if (b->bs != BS_4x8) {

  248.                 b->mode[2] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,

  249.                                               s->prob.p.y_mode[0]);

  250.                 s->counts.y_mode[0][b->mode[2]]++;

  251.                 if (b->bs != BS_8x4) {

  252.                     b->mode[3] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,

  253.                                                   s->prob.p.y_mode[0]);

  254.                     s->counts.y_mode[0][b->mode[3]]++;

  255.                 } else {

  256.                     b->mode[3] = b->mode[2];

  257.                 }

  258.             } else {

  259.                 b->mode[2] = b->mode[0];

  260.                 b->mode[3] = b->mode[1];

  261.             }

  262.         } else {

  263.             static const uint8_t size_group[10] = {

  264.                 3, 3, 3, 3, 2, 2, 2, 1, 1, 1

  265.             };

  266.             int sz = size_group[b->bs];

  267. 

  268.             b->mode[0] = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,

  269.                                           s->prob.p.y_mode[sz]);

  270.             b->mode[1] =

  271.             b->mode[2] =

  272.             b->mode[3] = b->mode[0];

  273.             s->counts.y_mode[sz][b->mode[3]]++;

  274.         }

  275.         b->uvmode = vp8_rac_get_tree(&s->c, ff_vp9_intramode_tree,

  276.                                      s->prob.p.uv_mode[b->mode[3]]);

  277.         s->counts.uv_mode[b->mode[3]][b->uvmode]++;

  278.     } else {

  279.         static const uint8_t inter_mode_ctx_lut[14][14] = {

  280.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  281.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  282.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  283.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  284.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  285.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  286.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  287.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  288.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  289.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  290.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  291.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  292.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },

  293.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },

  294.         };

  295. 

  296.         if (s->segmentation.feat[b->seg_id].ref_enabled) {

  297.             av_assert2(s->segmentation.feat[b->seg_id].ref_val != 0);

  298.             b->comp   = 0;

  299.             b->ref[0] = s->segmentation.feat[b->seg_id].ref_val - 1;

  300.         } else {

  301.             // read comp_pred flag

  302.             if (s->comppredmode != PRED_SWITCHABLE) {

  303.                 b->comp = s->comppredmode == PRED_COMPREF;

  304.             } else {

  305.                 int c;

  306. 

  307.                 // FIXME add intra as ref=0xff (or -1) to make these easier?

  308.                 if (have_a) {

  309.                     if (have_l) {

  310.                         if (s->above_comp_ctx[col] && s->left_comp_ctx[row7]) {

  311.                             c = 4;

  312.                         } else if (s->above_comp_ctx[col]) {

  313.                             c = 2 + (s->left_intra_ctx[row7] ||

  314.                                      s->left_ref_ctx[row7] == s->fixcompref);

  315.                         } else if (s->left_comp_ctx[row7]) {

  316.                             c = 2 + (s->above_intra_ctx[col] ||

  317.                                      s->above_ref_ctx[col] == s->fixcompref);

  318.                         } else {

  319.                             c = (!s->above_intra_ctx[col] &&

  320.                                  s->above_ref_ctx[col] == s->fixcompref) ^

  321.                                 (!s->left_intra_ctx[row7] &&

  322.                                  s->left_ref_ctx[row & 7] == s->fixcompref);

  323.                         }

  324.                     } else {

  325.                         c = s->above_comp_ctx[col] ? 3 :

  326.                             (!s->above_intra_ctx[col] && s->above_ref_ctx[col] == s->fixcompref);

  327.                     }

  328.                 } else if (have_l) {

  329.                     c = s->left_comp_ctx[row7] ? 3 :

  330.                         (!s->left_intra_ctx[row7] && s->left_ref_ctx[row7] == s->fixcompref);

  331.                 } else {

  332.                     c = 1;

  333.                 }

  334.                 b->comp = vp56_rac_get_prob(&s->c, s->prob.p.comp[c]);

  335.                 s->counts.comp[c][b->comp]++;

  336.             }

  337. 

  338.             // read actual references

  339.             // FIXME probably cache a few variables here to prevent repetitive

  340.             // memory accesses below

  341.             if (b->comp) { /* two references */

  342.                 int fix_idx = s->signbias[s->fixcompref], var_idx = !fix_idx, c, bit;

  343. 

  344.                 b->ref[fix_idx] = s->fixcompref;

  345.                 // FIXME can this codeblob be replaced by some sort of LUT?

  346.                 if (have_a) {

  347.                     if (have_l) {

  348.                         if (s->above_intra_ctx[col]) {

  349.                             if (s->left_intra_ctx[row7]) {

  350.                                 c = 2;

  351.                             } else {

  352.                                 c = 1 + 2 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  353.                             }

  354.                         } else if (s->left_intra_ctx[row7]) {

  355.                             c = 1 + 2 * (s->above_ref_ctx[col] != s->varcompref[1]);

  356.                         } else {

  357.                             int refl = s->left_ref_ctx[row7], refa = s->above_ref_ctx[col];

  358. 

  359.                             if (refl == refa && refa == s->varcompref[1]) {

  360.                                 c = 0;

  361.                             } else if (!s->left_comp_ctx[row7] && !s->above_comp_ctx[col]) {

  362.                                 if ((refa == s->fixcompref && refl == s->varcompref[0]) ||

  363.                                     (refl == s->fixcompref && refa == s->varcompref[0])) {

  364.                                     c = 4;

  365.                                 } else {

  366.                                     c = (refa == refl) ? 3 : 1;

  367.                                 }

  368.                             } else if (!s->left_comp_ctx[row7]) {

  369.                                 if (refa == s->varcompref[1] && refl != s->varcompref[1]) {

  370.                                     c = 1;

  371.                                 } else {

  372.                                     c = (refl == s->varcompref[1] &&

  373.                                          refa != s->varcompref[1]) ? 2 : 4;

  374.                                 }

  375.                             } else if (!s->above_comp_ctx[col]) {

  376.                                 if (refl == s->varcompref[1] && refa != s->varcompref[1]) {

  377.                                     c = 1;

  378.                                 } else {

  379.                                     c = (refa == s->varcompref[1] &&

  380.                                          refl != s->varcompref[1]) ? 2 : 4;

  381.                                 }

  382.                             } else {

  383.                                 c = (refl == refa) ? 4 : 2;

  384.                             }

  385.                         }

  386.                     } else {

  387.                         if (s->above_intra_ctx[col]) {

  388.                             c = 2;

  389.                         } else if (s->above_comp_ctx[col]) {

  390.                             c = 4 * (s->above_ref_ctx[col] != s->varcompref[1]);

  391.                         } else {

  392.                             c = 3 * (s->above_ref_ctx[col] != s->varcompref[1]);

  393.                         }

  394.                     }

  395.                 } else if (have_l) {

  396.                     if (s->left_intra_ctx[row7]) {

  397.                         c = 2;

  398.                     } else if (s->left_comp_ctx[row7]) {

  399.                         c = 4 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  400.                     } else {

  401.                         c = 3 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  402.                     }

  403.                 } else {

  404.                     c = 2;

  405.                 }

  406.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.comp_ref[c]);

  407.                 b->ref[var_idx] = s->varcompref[bit];

  408.                 s->counts.comp_ref[c][bit]++;

  409.             } else { /* single reference */

  410.                 int bit, c;

  411. 

  412.                 if (have_a && !s->above_intra_ctx[col]) {

  413.                     if (have_l && !s->left_intra_ctx[row7]) {

  414.                         if (s->left_comp_ctx[row7]) {

  415.                             if (s->above_comp_ctx[col]) {

  416.                                 c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7] ||

  417.                                          !s->above_ref_ctx[col]);

  418.                             } else {

  419.                                 c = (3 * !s->above_ref_ctx[col]) +

  420.                                     (!s->fixcompref || !s->left_ref_ctx[row7]);

  421.                             }

  422.                         } else if (s->above_comp_ctx[col]) {

  423.                             c = (3 * !s->left_ref_ctx[row7]) +

  424.                                 (!s->fixcompref || !s->above_ref_ctx[col]);

  425.                         } else {

  426.                             c = 2 * !s->left_ref_ctx[row7] + 2 * !s->above_ref_ctx[col];

  427.                         }

  428.                     } else if (s->above_intra_ctx[col]) {

  429.                         c = 2;

  430.                     } else if (s->above_comp_ctx[col]) {

  431.                         c = 1 + (!s->fixcompref || !s->above_ref_ctx[col]);

  432.                     } else {

  433.                         c = 4 * (!s->above_ref_ctx[col]);

  434.                     }

  435.                 } else if (have_l && !s->left_intra_ctx[row7]) {

  436.                     if (s->left_intra_ctx[row7]) {

  437.                         c = 2;

  438.                     } else if (s->left_comp_ctx[row7]) {

  439.                         c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7]);

  440.                     } else {

  441.                         c = 4 * (!s->left_ref_ctx[row7]);

  442.                     }

  443.                 } else {

  444.                     c = 2;

  445.                 }

  446.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][0]);

  447.                 s->counts.single_ref[c][0][bit]++;

  448.                 if (!bit) {

  449.                     b->ref[0] = 0;

  450.                 } else {

  451.                     // FIXME can this codeblob be replaced by some sort of LUT?

  452.                     if (have_a) {

  453.                         if (have_l) {

  454.                             if (s->left_intra_ctx[row7]) {

  455.                                 if (s->above_intra_ctx[col]) {

  456.                                     c = 2;

  457.                                 } else if (s->above_comp_ctx[col]) {

  458.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  459.                                                  s->above_ref_ctx[col] == 1);

  460.                                 } else if (!s->above_ref_ctx[col]) {

  461.                                     c = 3;

  462.                                 } else {

  463.                                     c = 4 * (s->above_ref_ctx[col] == 1);

  464.                                 }

  465.                             } else if (s->above_intra_ctx[col]) {

  466.                                 if (s->left_intra_ctx[row7]) {

  467.                                     c = 2;

  468.                                 } else if (s->left_comp_ctx[row7]) {

  469.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  470.                                                  s->left_ref_ctx[row7] == 1);

  471.                                 } else if (!s->left_ref_ctx[row7]) {

  472.                                     c = 3;

  473.                                 } else {

  474.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  475.                                 }

  476.                             } else if (s->above_comp_ctx[col]) {

  477.                                 if (s->left_comp_ctx[row7]) {

  478.                                     if (s->left_ref_ctx[row7] == s->above_ref_ctx[col]) {

  479.                                         c = 3 * (s->fixcompref == 1 ||

  480.                                                  s->left_ref_ctx[row7] == 1);

  481.                                     } else {

  482.                                         c = 2;

  483.                                     }

  484.                                 } else if (!s->left_ref_ctx[row7]) {

  485.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  486.                                                  s->above_ref_ctx[col] == 1);

  487.                                 } else {

  488.                                     c = 3 * (s->left_ref_ctx[row7] == 1) +

  489.                                         (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  490.                                 }

  491.                             } else if (s->left_comp_ctx[row7]) {

  492.                                 if (!s->above_ref_ctx[col]) {

  493.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  494.                                                  s->left_ref_ctx[row7] == 1);

  495.                                 } else {

  496.                                     c = 3 * (s->above_ref_ctx[col] == 1) +

  497.                                         (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  498.                                 }

  499.                             } else if (!s->above_ref_ctx[col]) {

  500.                                 if (!s->left_ref_ctx[row7]) {

  501.                                     c = 3;

  502.                                 } else {

  503.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  504.                                 }

  505.                             } else if (!s->left_ref_ctx[row7]) {

  506.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  507.                             } else {

  508.                                 c = 2 * (s->left_ref_ctx[row7] == 1) +

  509.                                     2 * (s->above_ref_ctx[col] == 1);

  510.                             }

  511.                         } else {

  512.                             if (s->above_intra_ctx[col] ||

  513.                                 (!s->above_comp_ctx[col] && !s->above_ref_ctx[col])) {

  514.                                 c = 2;

  515.                             } else if (s->above_comp_ctx[col]) {

  516.                                 c = 3 * (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  517.                             } else {

  518.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  519.                             }

  520.                         }

  521.                     } else if (have_l) {

  522.                         if (s->left_intra_ctx[row7] ||

  523.                             (!s->left_comp_ctx[row7] && !s->left_ref_ctx[row7])) {

  524.                             c = 2;

  525.                         } else if (s->left_comp_ctx[row7]) {

  526.                             c = 3 * (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  527.                         } else {

  528.                             c = 4 * (s->left_ref_ctx[row7] == 1);

  529.                         }

  530.                     } else {

  531.                         c = 2;

  532.                     }

  533.                     bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][1]);

  534.                     s->counts.single_ref[c][1][bit]++;

  535.                     b->ref[0] = 1 + bit;

  536.                 }

  537.             }

  538.         }

  539. 

  540.         if (b->bs <= BS_8x8) {

  541.             if (s->segmentation.feat[b->seg_id].skip_enabled) {

  542.                 b->mode[0] =

  543.                 b->mode[1] =

  544.                 b->mode[2] =

  545.                 b->mode[3] = ZEROMV;

  546.             } else {

  547.                 static const uint8_t off[10] = {

  548.                     3, 0, 0, 1, 0, 0, 0, 0, 0, 0

  549.                 };

  550. 

  551.                 // FIXME this needs to use the LUT tables from find_ref_mvs

  552.                 // because not all are -1,0/0,-1

  553.                 int c = inter_mode_ctx_lut[s->above_mode_ctx[col + off[b->bs]]]

  554.                                           [s->left_mode_ctx[row7 + off[b->bs]]];

  555. 

  556.                 b->mode[0] = vp8_rac_get_tree(&s->c, ff_vp9_inter_mode_tree,

  557.                                               s->prob.p.mv_mode[c]);

  558.                 b->mode[1] =

  559.                 b->mode[2] =

  560.                 b->mode[3] = b->mode[0];

  561.                 s->counts.mv_mode[c][b->mode[0] - 10]++;

  562.             }

  563.         }

  564. 

  565.         if (s->filtermode == FILTER_SWITCHABLE) {

  566.             int c;

  567. 

  568.             if (have_a && s->above_mode_ctx[col] >= NEARESTMV) {

  569.                 if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  570.                     c = s->above_filter_ctx[col] == s->left_filter_ctx[row7] ?

  571.                         s->left_filter_ctx[row7] : 3;

  572.                 } else {

  573.                     c = s->above_filter_ctx[col];

  574.                 }

  575.             } else if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  576.                 c = s->left_filter_ctx[row7];

  577.             } else {

  578.                 c = 3;

  579.             }

  580. 

  581.             b->filter = vp8_rac_get_tree(&s->c, ff_vp9_filter_tree,

  582.                                          s->prob.p.filter[c]);

  583.             s->counts.filter[c][b->filter]++;

  584.         } else {

  585.             b->filter = s->filtermode;

  586.         }

  587. 

  588.         if (b->bs > BS_8x8) {

  589.             int c = inter_mode_ctx_lut[s->above_mode_ctx[col]][s->left_mode_ctx[row7]];

  590. 

  591.             b->mode[0] = vp8_rac_get_tree(&s->c, ff_vp9_inter_mode_tree,

  592.                                           s->prob.p.mv_mode[c]);

  593.             s->counts.mv_mode[c][b->mode[0] - 10]++;

  594.             ff_vp9_fill_mv(s, b->mv[0], b->mode[0], 0);

  595. 

  596.             if (b->bs != BS_8x4) {

  597.                 b->mode[1] = vp8_rac_get_tree(&s->c, ff_vp9_inter_mode_tree,

  598.                                               s->prob.p.mv_mode[c]);

  599.                 s->counts.mv_mode[c][b->mode[1] - 10]++;

  600.                 ff_vp9_fill_mv(s, b->mv[1], b->mode[1], 1);

  601.             } else {

  602.                 b->mode[1] = b->mode[0];

  603.                 AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  604.                 AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  605.             }

  606. 

  607.             if (b->bs != BS_4x8) {

  608.                 b->mode[2] = vp8_rac_get_tree(&s->c, ff_vp9_inter_mode_tree,

  609.                                               s->prob.p.mv_mode[c]);

  610.                 s->counts.mv_mode[c][b->mode[2] - 10]++;

  611.                 ff_vp9_fill_mv(s, b->mv[2], b->mode[2], 2);

  612. 

  613.                 if (b->bs != BS_8x4) {

  614.                     b->mode[3] = vp8_rac_get_tree(&s->c, ff_vp9_inter_mode_tree,

  615.                                                   s->prob.p.mv_mode[c]);

  616.                     s->counts.mv_mode[c][b->mode[3] - 10]++;

  617.                     ff_vp9_fill_mv(s, b->mv[3], b->mode[3], 3);

  618.                 } else {

  619.                     b->mode[3] = b->mode[2];

  620.                     AV_COPY32(&b->mv[3][0], &b->mv[2][0]);

  621.                     AV_COPY32(&b->mv[3][1], &b->mv[2][1]);

  622.                 }

  623.             } else {

  624.                 b->mode[2] = b->mode[0];

  625.                 AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  626.                 AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  627.                 b->mode[3] = b->mode[1];

  628.                 AV_COPY32(&b->mv[3][0], &b->mv[1][0]);

  629.                 AV_COPY32(&b->mv[3][1], &b->mv[1][1]);

  630.             }

  631.         } else {

  632.             ff_vp9_fill_mv(s, b->mv[0], b->mode[0], -1);

  633.             AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  634.             AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  635.             AV_COPY32(&b->mv[3][0], &b->mv[0][0]);

  636.             AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  637.             AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  638.             AV_COPY32(&b->mv[3][1], &b->mv[0][1]);

  639.         }

  640.     }

  641. 

  642.     // FIXME this can probably be optimized

  643.     memset(&s->above_skip_ctx[col], b->skip, w4);

  644.     memset(&s->left_skip_ctx[row7], b->skip, h4);

  645.     memset(&s->above_txfm_ctx[col], b->tx, w4);

  646.     memset(&s->left_txfm_ctx[row7], b->tx, h4);

  647.     memset(&s->above_partition_ctx[col], above_ctx[b->bs], w4);

  648.     memset(&s->left_partition_ctx[row7], left_ctx[b->bs], h4);

  649.     if (!s->keyframe && !s->intraonly) {

  650.         memset(&s->above_intra_ctx[col], b->intra, w4);

  651.         memset(&s->left_intra_ctx[row7], b->intra, h4);

  652.         memset(&s->above_comp_ctx[col], b->comp, w4);

  653.         memset(&s->left_comp_ctx[row7], b->comp, h4);

  654.         memset(&s->above_mode_ctx[col], b->mode[3], w4);

  655.         memset(&s->left_mode_ctx[row7], b->mode[3], h4);

  656.         if (s->filtermode == FILTER_SWITCHABLE && !b->intra) {

  657.             memset(&s->above_filter_ctx[col], b->filter, w4);

  658.             memset(&s->left_filter_ctx[row7], b->filter, h4);

  659.             b->filter = ff_vp9_filter_lut[b->filter];

  660.         }

  661.         if (b->bs > BS_8x8) {

  662.             int mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  663. 

  664.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][0], &b->mv[1][0]);

  665.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][1], &b->mv[1][1]);

  666.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][0], mv0);

  667.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][1], mv1);

  668.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][0], &b->mv[2][0]);

  669.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][1], &b->mv[2][1]);

  670.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][0], mv0);

  671.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][1], mv1);

  672.         } else {

  673.             int n, mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  674. 

  675.             for (n = 0; n < w4 * 2; n++) {

  676.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][0], mv0);

  677.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][1], mv1);

  678.             }

  679.             for (n = 0; n < h4 * 2; n++) {

  680.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][0], mv0);

  681.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][1], mv1);

  682.             }

  683.         }

  684. 

  685.         if (!b->intra) { // FIXME write 0xff or -1 if intra, so we can use this

  686.                          // as a direct check in above branches

  687.             int vref = b->ref[b->comp ? s->signbias[s->varcompref[0]] : 0];

  688. 

  689.             memset(&s->above_ref_ctx[col], vref, w4);

  690.             memset(&s->left_ref_ctx[row7], vref, h4);

  691.         }

  692.     }

  693. 

  694.     // FIXME kinda ugly

  695.     for (y = 0; y < h4; y++) {

  696.         int x, o = (row + y) * s->sb_cols * 8 + col;

  697.         VP9MVRefPair *mv = &s->frames[CUR_FRAME].mv[o];

  698. 

  699.         if (b->intra) {

  700.             for (x = 0; x < w4; x++) {

  701.                 mv[x].ref[0] =

  702.                 mv[x].ref[1] = -1;

  703.             }

  704.         } else if (b->comp) {

  705.             for (x = 0; x < w4; x++) {

  706.                 mv[x].ref[0] = b->ref[0];

  707.                 mv[x].ref[1] = b->ref[1];

  708.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  709.                 AV_COPY32(&mv[x].mv[1], &b->mv[3][1]);

  710.             }

  711.         } else {

  712.             for (x = 0; x < w4; x++) {

  713.                 mv[x].ref[0] = b->ref[0];

  714.                 mv[x].ref[1] = -1;

  715.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  716.             }

  717.         }

  718.     }

  719. }

  720. 

  721. // FIXME remove tx argument, and merge cnt/eob arguments?

  722. static int decode_block_coeffs(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  723.                                enum TxfmMode tx, unsigned (*cnt)[6][3],

  724.                                unsigned (*eob)[6][2], uint8_t(*p)[6][11],

  725.                                int nnz, const int16_t *scan,

  726.                                const int16_t(*nb)[2],

  727.                                const int16_t *band_counts, const int16_t *qmul)

  728. {

  729.     int i = 0, band = 0, band_left = band_counts[band];

  730.     uint8_t *tp = p[0][nnz];

  731.     uint8_t cache[1024];

  732. 

  733.     do {

  734.         int val, rc;

  735. 

  736.         val = vp56_rac_get_prob_branchy(c, tp[0]); // eob

  737.         eob[band][nnz][val]++;

  738.         if (!val)

  739.             break;

  740. 

  741. skip_eob:

  742.         if (!vp56_rac_get_prob_branchy(c, tp[1])) { // zero

  743.             cnt[band][nnz][0]++;

  744.             if (!--band_left)

  745.                 band_left = band_counts[++band];

  746.             cache[scan[i]] = 0;

  747.             nnz            = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  748.             tp             = p[band][nnz];

  749.             if (++i == n_coeffs)

  750.                 break;  //invalid input; blocks should end with EOB

  751.             goto skip_eob;

  752.         }

  753. 

  754.         rc = scan[i];

  755.         if (!vp56_rac_get_prob_branchy(c, tp[2])) { // one

  756.             cnt[band][nnz][1]++;

  757.             val       = 1;

  758.             cache[rc] = 1;

  759.         } else {

  760.             // fill in p[3-10] (model fill) - only once per frame for each pos

  761.             if (!tp[3])

  762.                 memcpy(&tp[3], ff_vp9_model_pareto8[tp[2]], 8);

  763. 

  764.             cnt[band][nnz][2]++;

  765.             if (!vp56_rac_get_prob_branchy(c, tp[3])) { // 2, 3, 4

  766.                 if (!vp56_rac_get_prob_branchy(c, tp[4])) {

  767.                     cache[rc] = val = 2;

  768.                 } else {

  769.                     val       = 3 + vp56_rac_get_prob(c, tp[5]);

  770.                     cache[rc] = 3;

  771.                 }

  772.             } else if (!vp56_rac_get_prob_branchy(c, tp[6])) { // cat1/2

  773.                 cache[rc] = 4;

  774.                 if (!vp56_rac_get_prob_branchy(c, tp[7])) {

  775.                     val  =  vp56_rac_get_prob(c, 159) + 5;

  776.                 } else {

  777.                     val  = (vp56_rac_get_prob(c, 165) << 1) + 7;

  778.                     val +=  vp56_rac_get_prob(c, 145);

  779.                 }

  780.             } else { // cat 3-6

  781.                 cache[rc] = 5;

  782.                 if (!vp56_rac_get_prob_branchy(c, tp[8])) {

  783.                     if (!vp56_rac_get_prob_branchy(c, tp[9])) {

  784.                         val  = (vp56_rac_get_prob(c, 173) << 2) + 11;

  785.                         val += (vp56_rac_get_prob(c, 148) << 1);

  786.                         val +=  vp56_rac_get_prob(c, 140);

  787.                     } else {

  788.                         val  = (vp56_rac_get_prob(c, 176) << 3) + 19;

  789.                         val += (vp56_rac_get_prob(c, 155) << 2);

  790.                         val += (vp56_rac_get_prob(c, 140) << 1);

  791.                         val +=  vp56_rac_get_prob(c, 135);

  792.                     }

  793.                 } else if (!vp56_rac_get_prob_branchy(c, tp[10])) {

  794.                     val  = (vp56_rac_get_prob(c, 180) << 4) + 35;

  795.                     val += (vp56_rac_get_prob(c, 157) << 3);

  796.                     val += (vp56_rac_get_prob(c, 141) << 2);

  797.                     val += (vp56_rac_get_prob(c, 134) << 1);

  798.                     val +=  vp56_rac_get_prob(c, 130);

  799.                 } else {

  800.                     val  = (vp56_rac_get_prob(c, 254) << 13) + 67;

  801.                     val += (vp56_rac_get_prob(c, 254) << 12);

  802.                     val += (vp56_rac_get_prob(c, 254) << 11);

  803.                     val += (vp56_rac_get_prob(c, 252) << 10);

  804.                     val += (vp56_rac_get_prob(c, 249) << 9);

  805.                     val += (vp56_rac_get_prob(c, 243) << 8);

  806.                     val += (vp56_rac_get_prob(c, 230) << 7);

  807.                     val += (vp56_rac_get_prob(c, 196) << 6);

  808.                     val += (vp56_rac_get_prob(c, 177) << 5);

  809.                     val += (vp56_rac_get_prob(c, 153) << 4);

  810.                     val += (vp56_rac_get_prob(c, 140) << 3);

  811.                     val += (vp56_rac_get_prob(c, 133) << 2);

  812.                     val += (vp56_rac_get_prob(c, 130) << 1);

  813.                     val +=  vp56_rac_get_prob(c, 129);

  814.                 }

  815.             }

  816.         }

  817.         if (!--band_left)

  818.             band_left = band_counts[++band];

  819.         if (tx == TX_32X32) // FIXME slow

  820.             coef[rc] = ((vp8_rac_get(c) ? -val : val) * qmul[!!i]) / 2;

  821.         else

  822.             coef[rc] = (vp8_rac_get(c) ? -val : val) * qmul[!!i];

  823.         nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  824.         tp  = p[band][nnz];

  825.     } while (++i < n_coeffs);

  826. 

  827.     return i;

  828. }

  829. 

  830. static int decode_coeffs(AVCodecContext *avctx)

  831. {

  832.     VP9Context *s = avctx->priv_data;

  833.     VP9Block *b = s->b;

  834.     int row = b->row, col = b->col;

  835.     uint8_t (*p)[6][11] = s->prob.coef[b->tx][0 /* y */][!b->intra];

  836.     unsigned (*c)[6][3] = s->counts.coef[b->tx][0 /* y */][!b->intra];

  837.     unsigned (*e)[6][2] = s->counts.eob[b->tx][0 /* y */][!b->intra];

  838.     int w4 = bwh_tab[1][b->bs][0] << 1, h4 = bwh_tab[1][b->bs][1] << 1;

  839.     int end_x = FFMIN(2 * (s->cols - col), w4);

  840.     int end_y = FFMIN(2 * (s->rows - row), h4);

  841.     int n, pl, x, y, step1d = 1 << b->tx, step = 1 << (b->tx * 2);

  842.     int uvstep1d = 1 << b->uvtx, uvstep = 1 << (b->uvtx * 2), ret;

  843.     int16_t (*qmul)[2] = s->segmentation.feat[b->seg_id].qmul;

  844.     int tx = 4 * s->lossless + b->tx;

  845.     const int16_t * const *yscans = ff_vp9_scans[tx];

  846.     const int16_t (* const * ynbs)[2] = ff_vp9_scans_nb[tx];

  847.     const int16_t *uvscan = ff_vp9_scans[b->uvtx][DCT_DCT];

  848.     const int16_t (*uvnb)[2] = ff_vp9_scans_nb[b->uvtx][DCT_DCT];

  849.     uint8_t *a = &s->above_y_nnz_ctx[col * 2];

  850.     uint8_t *l = &s->left_y_nnz_ctx[(row & 7) << 1];

  851.     static const int16_t band_counts[4][8] = {

  852.         { 1, 2, 3, 4,  3,   16 - 13, 0 },

  853.         { 1, 2, 3, 4, 11,   64 - 21, 0 },

  854.         { 1, 2, 3, 4, 11,  256 - 21, 0 },

  855.         { 1, 2, 3, 4, 11, 1024 - 21, 0 },

  856.     };

  857.     const int16_t *y_band_counts  = band_counts[b->tx];

  858.     const int16_t *uv_band_counts = band_counts[b->uvtx];

  859. 

  860.     /* y tokens */

  861.     if (b->tx > TX_4X4) { // FIXME slow

  862.         for (y = 0; y < end_y; y += step1d)

  863.             for (x = 1; x < step1d; x++)

  864.                 l[y] |= l[y + x];

  865.         for (x = 0; x < end_x; x += step1d)

  866.             for (y = 1; y < step1d; y++)

  867.                 a[x] |= a[x + y];

  868.     }

  869.     for (n = 0, y = 0; y < end_y; y += step1d) {

  870.         for (x = 0; x < end_x; x += step1d, n += step) {

  871.             enum TxfmType txtp = ff_vp9_intra_txfm_type[b->mode[b->tx == TX_4X4 &&

  872.                                                                 b->bs > BS_8x8 ?

  873.                                                                 n : 0]];

  874.             int nnz = a[x] + l[y];

  875.             if ((ret = decode_block_coeffs(&s->c, s->block + 16 * n, 16 * step,

  876.                                            b->tx, c, e, p, nnz, yscans[txtp],

  877.                                            ynbs[txtp], y_band_counts,

  878.                                            qmul[0])) < 0)

  879.                 return ret;

  880.             a[x] = l[y] = !!ret;

  881.             if (b->tx > TX_8X8)

  882.                 AV_WN16A(&s->eob[n], ret);

  883.             else

  884.                 s->eob[n] = ret;

  885.         }

  886.     }

  887.     if (b->tx > TX_4X4) { // FIXME slow

  888.         for (y = 0; y < end_y; y += step1d)

  889.             memset(&l[y + 1], l[y], FFMIN(end_y - y - 1, step1d - 1));

  890.         for (x = 0; x < end_x; x += step1d)

  891.             memset(&a[x + 1], a[x], FFMIN(end_x - x - 1, step1d - 1));

  892.     }

  893. 

  894.     p = s->prob.coef[b->uvtx][1 /* uv */][!b->intra];

  895.     c = s->counts.coef[b->uvtx][1 /* uv */][!b->intra];

  896.     e = s->counts.eob[b->uvtx][1 /* uv */][!b->intra];

  897.     w4    >>= 1;

  898.     h4    >>= 1;

  899.     end_x >>= 1;

  900.     end_y >>= 1;

  901.     for (pl = 0; pl < 2; pl++) {

  902.         a = &s->above_uv_nnz_ctx[pl][col];

  903.         l = &s->left_uv_nnz_ctx[pl][row & 7];

  904.         if (b->uvtx > TX_4X4) { // FIXME slow

  905.             for (y = 0; y < end_y; y += uvstep1d)

  906.                 for (x = 1; x < uvstep1d; x++)

  907.                     l[y] |= l[y + x];

  908.             for (x = 0; x < end_x; x += uvstep1d)

  909.                 for (y = 1; y < uvstep1d; y++)

  910.                     a[x] |= a[x + y];

  911.         }

  912.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  913.             for (x = 0; x < end_x; x += uvstep1d, n += uvstep) {

  914.                 int nnz = a[x] + l[y];

  915.                 if ((ret = decode_block_coeffs(&s->c, s->uvblock[pl] + 16 * n,

  916.                                                16 * uvstep, b->uvtx, c, e, p,

  917.                                                nnz, uvscan, uvnb,

  918.                                                uv_band_counts, qmul[1])) < 0)

  919.                     return ret;

  920.                 a[x] = l[y] = !!ret;

  921.                 if (b->uvtx > TX_8X8)

  922.                     AV_WN16A(&s->uveob[pl][n], ret);

  923.                 else

  924.                     s->uveob[pl][n] = ret;

  925.             }

  926.         }

  927.         if (b->uvtx > TX_4X4) { // FIXME slow

  928.             for (y = 0; y < end_y; y += uvstep1d)

  929.                 memset(&l[y + 1], l[y], FFMIN(end_y - y - 1, uvstep1d - 1));

  930.             for (x = 0; x < end_x; x += uvstep1d)

  931.                 memset(&a[x + 1], a[x], FFMIN(end_x - x - 1, uvstep1d - 1));

  932.         }

  933.     }

  934. 

  935.     return 0;

  936. }

  937. 

  938. static av_always_inline int check_intra_mode(VP9Context *s, int mode,

  939.                                              uint8_t **a,

  940.                                              uint8_t *dst_edge,

  941.                                              ptrdiff_t stride_edge,

  942.                                              uint8_t *dst_inner,

  943.                                              ptrdiff_t stride_inner,

  944.                                              uint8_t *l, int col, int x, int w,

  945.                                              int row, int y, enum TxfmMode tx,

  946.                                              int p)

  947. {

  948.     int have_top   = row > 0 || y > 0;

  949.     int have_left  = col > s->tiling.tile_col_start || x > 0;

  950.     int have_right = x < w - 1;

  951.     static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {

  952.         [VERT_PRED]            = { { DC_127_PRED,          VERT_PRED            },

  953.                                    { DC_127_PRED,          VERT_PRED            } },

  954.         [HOR_PRED]             = { { DC_129_PRED,          DC_129_PRED          },

  955.                                    { HOR_PRED,             HOR_PRED             } },

  956.         [DC_PRED]              = { { DC_128_PRED,          TOP_DC_PRED          },

  957.                                    { LEFT_DC_PRED,         DC_PRED              } },

  958.         [DIAG_DOWN_LEFT_PRED]  = { { DC_127_PRED,          DIAG_DOWN_LEFT_PRED  },

  959.                                    { DC_127_PRED,          DIAG_DOWN_LEFT_PRED  } },

  960.         [DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },

  961.                                    { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },

  962.         [VERT_RIGHT_PRED]      = { { VERT_RIGHT_PRED,      VERT_RIGHT_PRED      },

  963.                                    { VERT_RIGHT_PRED,      VERT_RIGHT_PRED      } },

  964.         [HOR_DOWN_PRED]        = { { HOR_DOWN_PRED,        HOR_DOWN_PRED        },

  965.                                    { HOR_DOWN_PRED,        HOR_DOWN_PRED        } },

  966.         [VERT_LEFT_PRED]       = { { DC_127_PRED,          VERT_LEFT_PRED       },

  967.                                    { DC_127_PRED,          VERT_LEFT_PRED       } },

  968.         [HOR_UP_PRED]          = { { DC_129_PRED,          DC_129_PRED          },

  969.                                    { HOR_UP_PRED,          HOR_UP_PRED          } },

  970.         [TM_VP8_PRED]          = { { DC_129_PRED,          VERT_PRED            },

  971.                                    { HOR_PRED,             TM_VP8_PRED          } },

  972.     };

  973.     static const struct {

  974.         uint8_t needs_left:1;

  975.         uint8_t needs_top:1;

  976.         uint8_t needs_topleft:1;

  977.         uint8_t needs_topright:1;

  978.     } edges[N_INTRA_PRED_MODES] = {

  979.         [VERT_PRED]            = { .needs_top  = 1 },

  980.         [HOR_PRED]             = { .needs_left = 1 },

  981.         [DC_PRED]              = { .needs_top  = 1, .needs_left = 1 },

  982.         [DIAG_DOWN_LEFT_PRED]  = { .needs_top  = 1, .needs_topright = 1 },

  983.         [DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1,

  984.                                    .needs_topleft = 1 },

  985.         [VERT_RIGHT_PRED]      = { .needs_left = 1, .needs_top = 1,

  986.                                    .needs_topleft = 1 },

  987.         [HOR_DOWN_PRED]        = { .needs_left = 1, .needs_top = 1,

  988.                                    .needs_topleft = 1 },

  989.         [VERT_LEFT_PRED]       = { .needs_top  = 1, .needs_topright = 1 },

  990.         [HOR_UP_PRED]          = { .needs_left = 1 },

  991.         [TM_VP8_PRED]          = { .needs_left = 1, .needs_top = 1,

  992.                                    .needs_topleft = 1 },

  993.         [LEFT_DC_PRED]         = { .needs_left = 1 },

  994.         [TOP_DC_PRED]          = { .needs_top  = 1 },

  995.         [DC_128_PRED]          = { 0 },

  996.         [DC_127_PRED]          = { 0 },

  997.         [DC_129_PRED]          = { 0 }

  998.     };

  999. 

  1000.     av_assert2(mode >= 0 && mode < 10);

  1001.     mode = mode_conv[mode][have_left][have_top];

  1002.     if (edges[mode].needs_top) {

  1003.         uint8_t *top = NULL, *topleft = NULL;

  1004.         int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !p) - x) * 4;

  1005.         int n_px_need_tr = 0;

  1006. 

  1007.         if (tx == TX_4X4 && edges[mode].needs_topright && have_right)

  1008.             n_px_need_tr = 4;

  1009. 

  1010.         // if top of sb64-row, use s->intra_pred_data[] instead of

  1011.         // dst[-stride] for intra prediction (it contains pre- instead of

  1012.         // post-loopfilter data)

  1013.         if (have_top) {

  1014.             top = !(row & 7) && !y ?

  1015.                   s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  1016.                   y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];

  1017.             if (have_left)

  1018.                 topleft = !(row & 7) && !y ?

  1019.                           s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  1020.                           y == 0 || x == 0 ? &dst_edge[-stride_edge] :

  1021.                           &dst_inner[-stride_inner];

  1022.         }

  1023. 

  1024.         if (have_top &&

  1025.             (!edges[mode].needs_topleft || (have_left && top == topleft)) &&

  1026.             (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&

  1027.             n_px_need + n_px_need_tr <= n_px_have) {

  1028.             *a = top;

  1029.         } else {

  1030.             if (have_top) {

  1031.                 if (n_px_need <= n_px_have) {

  1032.                     memcpy(*a, top, n_px_need);

  1033.                 } else {

  1034.                     memcpy(*a, top, n_px_have);

  1035.                     memset(&(*a)[n_px_have], (*a)[n_px_have - 1],

  1036.                            n_px_need - n_px_have);

  1037.                 }

  1038.             } else {

  1039.                 memset(*a, 127, n_px_need);

  1040.             }

  1041.             if (edges[mode].needs_topleft) {

  1042.                 if (have_left && have_top)

  1043.                     (*a)[-1] = topleft[-1];

  1044.                 else

  1045.                     (*a)[-1] = have_top ? 129 : 127;

  1046.             }

  1047.             if (tx == TX_4X4 && edges[mode].needs_topright) {

  1048.                 if (have_top && have_right &&

  1049.                     n_px_need + n_px_need_tr <= n_px_have) {

  1050.                     memcpy(&(*a)[4], &top[4], 4);

  1051.                 } else {

  1052.                     memset(&(*a)[4], (*a)[3], 4);

  1053.                 }

  1054.             }

  1055.         }

  1056.     }

  1057.     if (edges[mode].needs_left) {

  1058.         if (have_left) {

  1059.             int i;

  1060.             int n_px_need = 4 << tx;

  1061.             int n_px_have = (((s->rows - row) << !p) - y) * 4;

  1062.             uint8_t *dst     = x == 0 ? dst_edge : dst_inner;

  1063.             ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;

  1064. 

  1065.             if (n_px_need <= n_px_have) {

  1066.                 for (i = 0; i < n_px_need; i++)

  1067.                     l[i] = dst[i * stride - 1];

  1068.             } else {

  1069.                 for (i = 0; i < n_px_have; i++)

  1070.                     l[i] = dst[i * stride - 1];

  1071.                 memset(&l[i], l[i - 1], n_px_need - n_px_have);

  1072.             }

  1073.         } else {

  1074.             memset(l, 129, 4 << tx);

  1075.         }

  1076.     }

  1077. 

  1078.     return mode;

  1079. }

  1080. 

  1081. static void intra_recon(AVCodecContext *avctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  1082. {

  1083.     VP9Context *s = avctx->priv_data;

  1084.     VP9Block *b = s->b;

  1085.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  1086.     int row = b->row, col = b->col;

  1087.     int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  1088.     int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  1089.     int end_x = FFMIN(2 * (s->cols - col), w4);

  1090.     int end_y = FFMIN(2 * (s->rows - row), h4);

  1091.     int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  1092.     int uvstep1d = 1 << b->uvtx, p;

  1093.     uint8_t *dst = b->dst[0], *dst_r = f->data[0] + y_off;

  1094. 

  1095.     for (n = 0, y = 0; y < end_y; y += step1d) {

  1096.         uint8_t *ptr = dst, *ptr_r = dst_r;

  1097.         for (x = 0; x < end_x;

  1098.              x += step1d, ptr += 4 * step1d, ptr_r += 4 * step1d, n += step) {

  1099.             int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?

  1100.                                y * 2 + x : 0];

  1101.             LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);

  1102.             uint8_t *a = &a_buf[16], l[32];

  1103.             enum TxfmType txtp = ff_vp9_intra_txfm_type[mode];

  1104.             int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  1105. 

  1106.             mode = check_intra_mode(s, mode, &a, ptr_r,

  1107.                                     f->linesize[0],

  1108.                                     ptr, b->y_stride, l,

  1109.                                     col, x, w4, row, y, b->tx, 0);

  1110.             s->dsp.intra_pred[b->tx][mode](ptr, b->y_stride, l, a);

  1111.             if (eob)

  1112.                 s->dsp.itxfm_add[tx][txtp](ptr, b->y_stride,

  1113.                                            s->block + 16 * n, eob);

  1114.         }

  1115.         dst_r += 4 * f->linesize[0] * step1d;

  1116.         dst   += 4 * b->y_stride * step1d;

  1117.     }

  1118. 

  1119.     // U/V

  1120.     h4    >>= 1;

  1121.     w4    >>= 1;

  1122.     end_x >>= 1;

  1123.     end_y >>= 1;

  1124.     step    = 1 << (b->uvtx * 2);

  1125.     for (p = 0; p < 2; p++) {

  1126.         dst   = b->dst[1 + p];

  1127.         dst_r = f->data[1 + p] + uv_off;

  1128.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  1129.             uint8_t *ptr = dst, *ptr_r = dst_r;

  1130.             for (x = 0; x < end_x;

  1131.                  x += uvstep1d, ptr += 4 * uvstep1d,

  1132.                  ptr_r += 4 * uvstep1d, n += step) {

  1133.                 int mode = b->uvmode;

  1134.                 LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);

  1135.                 uint8_t *a = &a_buf[16], l[32];

  1136.                 int eob    = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n])

  1137.                                               : s->uveob[p][n];

  1138. 

  1139.                 mode = check_intra_mode(s, mode, &a, ptr_r,

  1140.                                         f->linesize[1],

  1141.                                         ptr, b->uv_stride, l,

  1142.                                         col, x, w4, row, y, b->uvtx, p + 1);

  1143.                 s->dsp.intra_pred[b->uvtx][mode](ptr, b->uv_stride, l, a);

  1144.                 if (eob)

  1145.                     s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, b->uv_stride,

  1146.                                                     s->uvblock[p] + 16 * n,

  1147.                                                     eob);

  1148.             }

  1149.             dst_r += 4 * uvstep1d * f->linesize[1];

  1150.             dst   += 4 * uvstep1d * b->uv_stride;

  1151.         }

  1152.     }

  1153. }

  1154. 

  1155. static av_always_inline void mc_luma_dir(VP9Context *s, vp9_mc_func(*mc)[2],

  1156.                                          uint8_t *dst, ptrdiff_t dst_stride,

  1157.                                          const uint8_t *ref,

  1158.                                          ptrdiff_t ref_stride,

  1159.                                          ThreadFrame *ref_frame,

  1160.                                          ptrdiff_t y, ptrdiff_t x,

  1161.                                          const VP56mv *mv,

  1162.                                          int bw, int bh, int w, int h)

  1163. {

  1164.     int mx = mv->x, my = mv->y;

  1165.     int th;

  1166. 

  1167.     y   += my >> 3;

  1168.     x   += mx >> 3;

  1169.     ref += y * ref_stride + x;

  1170.     mx  &= 7;

  1171.     my  &= 7;

  1172. 

  1173.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  1174.     // the longest loopfilter of the next sbrow

  1175.     th = (y + bh + 4 * !!my + 7) >> 6;

  1176.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  1177. 

  1178.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  1179.     // The arm/aarch64 _hv filters read one more row than what actually is

  1180.     // needed, so switch to emulated edge one pixel sooner vertically

  1181.     // (!!my * 5) than horizontally (!!mx * 4).

  1182.     if (x < !!mx * 3 || y < !!my * 3 ||

  1183.         x + !!mx * 4 > w - bw || y + !!my * 5 > h - bh) {

  1184.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1185.                                  ref - !!my * 3 * ref_stride - !!mx * 3,

  1186.                                  80,

  1187.                                  ref_stride,

  1188.                                  bw + !!mx * 7, bh + !!my * 7,

  1189.                                  x - !!mx * 3, y - !!my * 3, w, h);

  1190.         ref        = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  1191.         ref_stride = 80;

  1192.     }

  1193.     mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);

  1194. }

  1195. 

  1196. static av_always_inline void mc_chroma_dir(VP9Context *s, vp9_mc_func(*mc)[2],

  1197.                                            uint8_t *dst_u, uint8_t *dst_v,

  1198.                                            ptrdiff_t dst_stride,

  1199.                                            const uint8_t *ref_u,

  1200.                                            ptrdiff_t src_stride_u,

  1201.                                            const uint8_t *ref_v,

  1202.                                            ptrdiff_t src_stride_v,

  1203.                                            ThreadFrame *ref_frame,

  1204.                                            ptrdiff_t y, ptrdiff_t x,

  1205.                                            const VP56mv *mv,

  1206.                                            int bw, int bh, int w, int h)

  1207. {

  1208.     int mx = mv->x, my = mv->y;

  1209.     int th;

  1210. 

  1211.     y     += my >> 4;

  1212.     x     += mx >> 4;

  1213.     ref_u += y * src_stride_u + x;

  1214.     ref_v += y * src_stride_v + x;

  1215.     mx    &= 15;

  1216.     my    &= 15;

  1217. 

  1218.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  1219.     // the longest loopfilter of the next sbrow

  1220.     th = (y + bh + 4 * !!my + 7) >> 5;

  1221.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  1222. 

  1223.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  1224.     // The arm/aarch64 _hv filters read one more row than what actually is

  1225.     // needed, so switch to emulated edge one pixel sooner vertically

  1226.     // (!!my * 5) than horizontally (!!mx * 4).

  1227.     if (x < !!mx * 3 || y < !!my * 3 ||

  1228.         x + !!mx * 4 > w - bw || y + !!my * 5 > h - bh) {

  1229.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1230.                                  ref_u - !!my * 3 * src_stride_u - !!mx * 3,

  1231.                                  80,

  1232.                                  src_stride_u,

  1233.                                  bw + !!mx * 7, bh + !!my * 7,

  1234.                                  x - !!mx * 3, y - !!my * 3, w, h);

  1235.         ref_u = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  1236.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, 80, bh, mx, my);

  1237. 

  1238.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1239.                                  ref_v - !!my * 3 * src_stride_v - !!mx * 3,

  1240.                                  80,

  1241.                                  src_stride_v,

  1242.                                  bw + !!mx * 7, bh + !!my * 7,

  1243.                                  x - !!mx * 3, y - !!my * 3, w, h);

  1244.         ref_v = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  1245.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, 80, bh, mx, my);

  1246.     } else {

  1247.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);

  1248.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);

  1249.     }

  1250. }

  1251. 

  1252. static int inter_recon(AVCodecContext *avctx)

  1253. {

  1254.     static const uint8_t bwlog_tab[2][N_BS_SIZES] = {

  1255.         { 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4 },

  1256.         { 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 4, 4 },

  1257.     };

  1258.     VP9Context *s = avctx->priv_data;

  1259.     VP9Block *b = s->b;

  1260.     int row = b->row, col = b->col;

  1261. 

  1262.     ThreadFrame *tref1 = &s->refs[s->refidx[b->ref[0]]];

  1263.     ThreadFrame *tref2 = b->comp ? &s->refs[s->refidx[b->ref[1]]] : NULL;

  1264.     AVFrame      *ref1 = tref1->f;

  1265.     AVFrame      *ref2 = tref2 ? tref2->f : NULL;

  1266. 

  1267.     int w = avctx->width, h = avctx->height;

  1268.     ptrdiff_t ls_y = b->y_stride, ls_uv = b->uv_stride;

  1269. 

  1270.     if (!ref1->data[0] || (b->comp && !ref2->data[0]))

  1271.         return AVERROR_INVALIDDATA;

  1272. 

  1273.     // y inter pred

  1274.     if (b->bs > BS_8x8) {

  1275.         if (b->bs == BS_8x4) {

  1276.             mc_luma_dir(s, s->dsp.mc[3][b->filter][0], b->dst[0], ls_y,

  1277.                         ref1->data[0], ref1->linesize[0], tref1,

  1278.                         row << 3, col << 3, &b->mv[0][0], 8, 4, w, h);

  1279.             mc_luma_dir(s, s->dsp.mc[3][b->filter][0],

  1280.                         b->dst[0] + 4 * ls_y, ls_y,

  1281.                         ref1->data[0], ref1->linesize[0], tref1,

  1282.                         (row << 3) + 4, col << 3, &b->mv[2][0], 8, 4, w, h);

  1283. 

  1284.             if (b->comp) {

  1285.                 mc_luma_dir(s, s->dsp.mc[3][b->filter][1], b->dst[0], ls_y,

  1286.                             ref2->data[0], ref2->linesize[0], tref2,

  1287.                             row << 3, col << 3, &b->mv[0][1], 8, 4, w, h);

  1288.                 mc_luma_dir(s, s->dsp.mc[3][b->filter][1],

  1289.                             b->dst[0] + 4 * ls_y, ls_y,

  1290.                             ref2->data[0], ref2->linesize[0], tref2,

  1291.                             (row << 3) + 4, col << 3, &b->mv[2][1], 8, 4, w, h);

  1292.             }

  1293.         } else if (b->bs == BS_4x8) {

  1294.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], b->dst[0], ls_y,

  1295.                         ref1->data[0], ref1->linesize[0], tref1,

  1296.                         row << 3, col << 3, &b->mv[0][0], 4, 8, w, h);

  1297.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], b->dst[0] + 4, ls_y,

  1298.                         ref1->data[0], ref1->linesize[0], tref1,

  1299.                         row << 3, (col << 3) + 4, &b->mv[1][0], 4, 8, w, h);

  1300. 

  1301.             if (b->comp) {

  1302.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], b->dst[0], ls_y,

  1303.                             ref2->data[0], ref2->linesize[0], tref2,

  1304.                             row << 3, col << 3, &b->mv[0][1], 4, 8, w, h);

  1305.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], b->dst[0] + 4, ls_y,

  1306.                             ref2->data[0], ref2->linesize[0], tref2,

  1307.                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 8, w, h);

  1308.             }

  1309.         } else {

  1310.             av_assert2(b->bs == BS_4x4);

  1311. 

  1312.             // FIXME if two horizontally adjacent blocks have the same MV,

  1313.             // do a w8 instead of a w4 call

  1314.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], b->dst[0], ls_y,

  1315.                         ref1->data[0], ref1->linesize[0], tref1,

  1316.                         row << 3, col << 3, &b->mv[0][0], 4, 4, w, h);

  1317.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], b->dst[0] + 4, ls_y,

  1318.                         ref1->data[0], ref1->linesize[0], tref1,

  1319.                         row << 3, (col << 3) + 4, &b->mv[1][0], 4, 4, w, h);

  1320.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0],

  1321.                         b->dst[0] + 4 * ls_y, ls_y,

  1322.                         ref1->data[0], ref1->linesize[0], tref1,

  1323.                         (row << 3) + 4, col << 3, &b->mv[2][0], 4, 4, w, h);

  1324.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0],

  1325.                         b->dst[0] + 4 * ls_y + 4, ls_y,

  1326.                         ref1->data[0], ref1->linesize[0], tref1,

  1327.                         (row << 3) + 4, (col << 3) + 4, &b->mv[3][0], 4, 4, w, h);

  1328. 

  1329.             if (b->comp) {

  1330.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], b->dst[0], ls_y,

  1331.                             ref2->data[0], ref2->linesize[0], tref2,

  1332.                             row << 3, col << 3, &b->mv[0][1], 4, 4, w, h);

  1333.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], b->dst[0] + 4, ls_y,

  1334.                             ref2->data[0], ref2->linesize[0], tref2,

  1335.                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 4, w, h);

  1336.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1],

  1337.                             b->dst[0] + 4 * ls_y, ls_y,

  1338.                             ref2->data[0], ref2->linesize[0], tref2,

  1339.                             (row << 3) + 4, col << 3, &b->mv[2][1], 4, 4, w, h);

  1340.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1],

  1341.                             b->dst[0] + 4 * ls_y + 4, ls_y,

  1342.                             ref2->data[0], ref2->linesize[0], tref2,

  1343.                             (row << 3) + 4, (col << 3) + 4, &b->mv[3][1], 4, 4, w, h);

  1344.             }

  1345.         }

  1346.     } else {

  1347.         int bwl = bwlog_tab[0][b->bs];

  1348.         int bw  = bwh_tab[0][b->bs][0] * 4;

  1349.         int bh  = bwh_tab[0][b->bs][1] * 4;

  1350. 

  1351.         mc_luma_dir(s, s->dsp.mc[bwl][b->filter][0], b->dst[0], ls_y,

  1352.                     ref1->data[0], ref1->linesize[0], tref1,

  1353.                     row << 3, col << 3, &b->mv[0][0], bw, bh, w, h);

  1354. 

  1355.         if (b->comp)

  1356.             mc_luma_dir(s, s->dsp.mc[bwl][b->filter][1], b->dst[0], ls_y,

  1357.                         ref2->data[0], ref2->linesize[0], tref2,

  1358.                         row << 3, col << 3, &b->mv[0][1], bw, bh, w, h);

  1359.     }

  1360. 

  1361.     // uv inter pred

  1362.     {

  1363.         int bwl = bwlog_tab[1][b->bs];

  1364.         int bw  = bwh_tab[1][b->bs][0] * 4, bh = bwh_tab[1][b->bs][1] * 4;

  1365.         VP56mv mvuv;

  1366. 

  1367.         w = (w + 1) >> 1;

  1368.         h = (h + 1) >> 1;

  1369.         if (b->bs > BS_8x8) {

  1370.             mvuv.x = ROUNDED_DIV(b->mv[0][0].x + b->mv[1][0].x +

  1371.                                  b->mv[2][0].x + b->mv[3][0].x, 4);

  1372.             mvuv.y = ROUNDED_DIV(b->mv[0][0].y + b->mv[1][0].y +

  1373.                                  b->mv[2][0].y + b->mv[3][0].y, 4);

  1374.         } else {

  1375.             mvuv = b->mv[0][0];

  1376.         }

  1377. 

  1378.         mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][0],

  1379.                       b->dst[1], b->dst[2], ls_uv,

  1380.                       ref1->data[1], ref1->linesize[1],

  1381.                       ref1->data[2], ref1->linesize[2], tref1,

  1382.                       row << 2, col << 2, &mvuv, bw, bh, w, h);

  1383. 

  1384.         if (b->comp) {

  1385.             if (b->bs > BS_8x8) {

  1386.                 mvuv.x = ROUNDED_DIV(b->mv[0][1].x + b->mv[1][1].x +

  1387.                                      b->mv[2][1].x + b->mv[3][1].x, 4);

  1388.                 mvuv.y = ROUNDED_DIV(b->mv[0][1].y + b->mv[1][1].y +

  1389.                                      b->mv[2][1].y + b->mv[3][1].y, 4);

  1390.             } else {

  1391.                 mvuv = b->mv[0][1];

  1392.             }

  1393.             mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][1],

  1394.                           b->dst[1], b->dst[2], ls_uv,

  1395.                           ref2->data[1], ref2->linesize[1],

  1396.                           ref2->data[2], ref2->linesize[2], tref2,

  1397.                           row << 2, col << 2, &mvuv, bw, bh, w, h);

  1398.         }

  1399.     }

  1400. 

  1401.     if (!b->skip) {

  1402.         /* mostly copied intra_reconn() */

  1403. 

  1404.         int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  1405.         int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  1406.         int end_x = FFMIN(2 * (s->cols - col), w4);

  1407.         int end_y = FFMIN(2 * (s->rows - row), h4);

  1408.         int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  1409.         int uvstep1d = 1 << b->uvtx, p;

  1410.         uint8_t *dst = b->dst[0];

  1411. 

  1412.         // y itxfm add

  1413.         for (n = 0, y = 0; y < end_y; y += step1d) {

  1414.             uint8_t *ptr = dst;

  1415.             for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d, n += step) {

  1416.                 int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  1417. 

  1418.                 if (eob)

  1419.                     s->dsp.itxfm_add[tx][DCT_DCT](ptr, b->y_stride,

  1420.                                                   s->block + 16 * n, eob);

  1421.             }

  1422.             dst += 4 * b->y_stride * step1d;

  1423.         }

  1424. 

  1425.         // uv itxfm add

  1426.         h4    >>= 1;

  1427.         w4    >>= 1;

  1428.         end_x >>= 1;

  1429.         end_y >>= 1;

  1430.         step    = 1 << (b->uvtx * 2);

  1431.         for (p = 0; p < 2; p++) {

  1432.             dst = b->dst[p + 1];

  1433.             for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  1434.                 uint8_t *ptr = dst;

  1435.                 for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d, n += step) {

  1436.                     int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n])

  1437.                                                : s->uveob[p][n];

  1438.                     if (eob)

  1439.                         s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, b->uv_stride,

  1440.                                                         s->uvblock[p] + 16 * n, eob);

  1441.                 }

  1442.                 dst += 4 * uvstep1d * b->uv_stride;

  1443.             }

  1444.         }

  1445.     }

  1446.     return 0;

  1447. }

  1448. 

  1449. static av_always_inline void mask_edges(VP9Filter *lflvl, int is_uv,

  1450.                                         int row_and_7, int col_and_7,

  1451.                                         int w, int h, int col_end, int row_end,

  1452.                                         enum TxfmMode tx, int skip_inter)

  1453. {

  1454.     // FIXME I'm pretty sure all loops can be replaced by a single LUT if

  1455.     // we make VP9Filter.mask uint64_t (i.e. row/col all single variable)

  1456.     // and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then

  1457.     // use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)

  1458. 

  1459.     // the intended behaviour of the vp9 loopfilter is to work on 8-pixel

  1460.     // edges. This means that for UV, we work on two subsampled blocks at

  1461.     // a time, and we only use the topleft block's mode information to set

  1462.     // things like block strength. Thus, for any block size smaller than

  1463.     // 16x16, ignore the odd portion of the block.

  1464.     if (tx == TX_4X4 && is_uv) {

  1465.         if (h == 1) {

  1466.             if (row_and_7 & 1)

  1467.                 return;

  1468.             if (!row_end)

  1469.                 h += 1;

  1470.         }

  1471.         if (w == 1) {

  1472.             if (col_and_7 & 1)

  1473.                 return;

  1474.             if (!col_end)

  1475.                 w += 1;

  1476.         }

  1477.     }

  1478. 

  1479.     if (tx == TX_4X4 && !skip_inter) {

  1480.         int t = 1 << col_and_7, m_col = (t << w) - t, y;

  1481.         int m_col_odd = (t << (w - 1)) - t;

  1482. 

  1483.         // on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide

  1484.         if (is_uv) {

  1485.             int m_row_8 = m_col & 0x01, m_row_4 = m_col - m_row_8;

  1486. 

  1487.             for (y = row_and_7; y < h + row_and_7; y++) {

  1488.                 int col_mask_id = 2 - !(y & 7);

  1489. 

  1490.                 lflvl->mask[is_uv][0][y][1] |= m_row_8;

  1491.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  1492.                 // for odd lines, if the odd col is not being filtered,

  1493.                 // skip odd row also:

  1494.                 // .---. <-- a

  1495.                 // |   |

  1496.                 // |___| <-- b

  1497.                 // ^   ^

  1498.                 // c   d

  1499.                 //

  1500.                 // if a/c are even row/col and b/d are odd, and d is skipped,

  1501.                 // e.g. right edge of size-66x66.webm, then skip b also (bug)

  1502.                 if ((col_end & 1) && (y & 1)) {

  1503.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col_odd;

  1504.                 } else {

  1505.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col;

  1506.                 }

  1507.             }

  1508.         } else {

  1509.             int m_row_8 = m_col & 0x11, m_row_4 = m_col - m_row_8;

  1510. 

  1511.             for (y = row_and_7; y < h + row_and_7; y++) {

  1512.                 int col_mask_id = 2 - !(y & 3);

  1513. 

  1514.                 lflvl->mask[is_uv][0][y][1]           |= m_row_8; // row edge

  1515.                 lflvl->mask[is_uv][0][y][2]           |= m_row_4;

  1516.                 lflvl->mask[is_uv][1][y][col_mask_id] |= m_col; // col edge

  1517.                 lflvl->mask[is_uv][0][y][3]           |= m_col;

  1518.                 lflvl->mask[is_uv][1][y][3]           |= m_col;

  1519.             }

  1520.         }

  1521.     } else {

  1522.         int y, t = 1 << col_and_7, m_col = (t << w) - t;

  1523. 

  1524.         if (!skip_inter) {

  1525.             int mask_id = (tx == TX_8X8);

  1526.             int l2 = tx + is_uv - 1, step1d = 1 << l2;

  1527.             static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };

  1528.             int m_row = m_col & masks[l2];

  1529. 

  1530.             // at odd UV col/row edges tx16/tx32 loopfilter edges, force

  1531.             // 8wd loopfilter to prevent going off the visible edge.

  1532.             if (is_uv && tx > TX_8X8 && (w ^ (w - 1)) == 1) {

  1533.                 int m_row_16 = ((t << (w - 1)) - t) & masks[l2];

  1534.                 int m_row_8  = m_row - m_row_16;

  1535. 

  1536.                 for (y = row_and_7; y < h + row_and_7; y++) {

  1537.                     lflvl->mask[is_uv][0][y][0] |= m_row_16;

  1538.                     lflvl->mask[is_uv][0][y][1] |= m_row_8;

  1539.                 }

  1540.             } else {

  1541.                 for (y = row_and_7; y < h + row_and_7; y++)

  1542.                     lflvl->mask[is_uv][0][y][mask_id] |= m_row;

  1543.             }

  1544. 

  1545.             if (is_uv && tx > TX_8X8 && (h ^ (h - 1)) == 1) {

  1546.                 for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)

  1547.                     lflvl->mask[is_uv][1][y][0] |= m_col;

  1548.                 if (y - row_and_7 == h - 1)

  1549.                     lflvl->mask[is_uv][1][y][1] |= m_col;

  1550.             } else {

  1551.                 for (y = row_and_7; y < h + row_and_7; y += step1d)

  1552.                     lflvl->mask[is_uv][1][y][mask_id] |= m_col;

  1553.             }

  1554.         } else if (tx != TX_4X4) {

  1555.             int mask_id;

  1556. 

  1557.             mask_id = (tx == TX_8X8) || (is_uv && h == 1);

  1558.             lflvl->mask[is_uv][1][row_and_7][mask_id] |= m_col;

  1559.             mask_id = (tx == TX_8X8) || (is_uv && w == 1);

  1560.             for (y = row_and_7; y < h + row_and_7; y++)

  1561.                 lflvl->mask[is_uv][0][y][mask_id] |= t;

  1562.         } else if (is_uv) {

  1563.             int t8 = t & 0x01, t4 = t - t8;

  1564. 

  1565.             for (y = row_and_7; y < h + row_and_7; y++) {

  1566.                 lflvl->mask[is_uv][0][y][2] |= t4;

  1567.                 lflvl->mask[is_uv][0][y][1] |= t8;

  1568.             }

  1569.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 7)] |= m_col;

  1570.         } else {

  1571.             int t8 = t & 0x11, t4 = t - t8;

  1572. 

  1573.             for (y = row_and_7; y < h + row_and_7; y++) {

  1574.                 lflvl->mask[is_uv][0][y][2] |= t4;

  1575.                 lflvl->mask[is_uv][0][y][1] |= t8;

  1576.             }

  1577.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 3)] |= m_col;

  1578.         }

  1579.     }

  1580. }

  1581. 

  1582. int ff_vp9_decode_block(AVCodecContext *avctx, int row, int col,

  1583.                         VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,

  1584.                         enum BlockLevel bl, enum BlockPartition bp)

  1585. {

  1586.     VP9Context *s = avctx->priv_data;

  1587.     VP9Block *b = s->b;

  1588.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  1589.     enum BlockSize bs = bl * 3 + bp;

  1590.     int ret, y, w4 = bwh_tab[1][bs][0], h4 = bwh_tab[1][bs][1], lvl;

  1591.     int emu[2];

  1592. 

  1593.     b->row  = row;

  1594.     b->row7 = row & 7;

  1595.     b->col  = col;

  1596.     b->col7 = col & 7;

  1597. 

  1598.     s->min_mv.x = -(128 + col * 64);

  1599.     s->min_mv.y = -(128 + row * 64);

  1600.     s->max_mv.x = 128 + (s->cols - col - w4) * 64;

  1601.     s->max_mv.y = 128 + (s->rows - row - h4) * 64;

  1602. 

  1603.     if (s->pass < 2) {

  1604.         b->bs = bs;

  1605.         b->bl = bl;

  1606.         b->bp = bp;

  1607.         decode_mode(s, b);

  1608.         b->uvtx = b->tx - (w4 * 2 == (1 << b->tx) || h4 * 2 == (1 << b->tx));

  1609. 

  1610.         if (!b->skip) {

  1611.             if ((ret = decode_coeffs(avctx)) < 0)

  1612.                 return ret;

  1613.         } else {

  1614.             int pl;

  1615. 

  1616.             memset(&s->above_y_nnz_ctx[col * 2], 0, w4 * 2);

  1617.             memset(&s->left_y_nnz_ctx[(row & 7) << 1], 0, h4 * 2);

  1618.             for (pl = 0; pl < 2; pl++) {

  1619.                 memset(&s->above_uv_nnz_ctx[pl][col], 0, w4);

  1620.                 memset(&s->left_uv_nnz_ctx[pl][row & 7], 0, h4);

  1621.             }

  1622.         }

  1623. 

  1624.         if (s->pass == 1) {

  1625.             s->b++;

  1626.             s->block      += w4 * h4 * 64;

  1627.             s->uvblock[0] += w4 * h4 * 16;

  1628.             s->uvblock[1] += w4 * h4 * 16;

  1629.             s->eob        += w4 * h4 * 4;

  1630.             s->uveob[0]   += w4 * h4;

  1631.             s->uveob[1]   += w4 * h4;

  1632. 

  1633.             return 0;

  1634.         }

  1635.     }

  1636. 

  1637.     /* Emulated overhangs if the stride of the target buffer can't hold.

  1638.      * This allows to support emu-edge and so on even if we have large

  1639.      * block overhangs. */

  1640.     emu[0] = (col + w4) * 8 > f->linesize[0] ||

  1641.              (row + h4) > s->rows;

  1642.     emu[1] = (col + w4) * 4 > f->linesize[1] ||

  1643.              (row + h4) > s->rows;

  1644.     if (emu[0]) {

  1645.         b->dst[0]   = s->tmp_y;

  1646.         b->y_stride = 64;

  1647.     } else {

  1648.         b->dst[0]   = f->data[0] + yoff;

  1649.         b->y_stride = f->linesize[0];

  1650.     }

  1651.     if (emu[1]) {

  1652.         b->dst[1]    = s->tmp_uv[0];

  1653.         b->dst[2]    = s->tmp_uv[1];

  1654.         b->uv_stride = 32;

  1655.     } else {

  1656.         b->dst[1]    = f->data[1] + uvoff;

  1657.         b->dst[2]    = f->data[2] + uvoff;

  1658.         b->uv_stride = f->linesize[1];

  1659.     }

  1660.     if (b->intra) {

  1661.         intra_recon(avctx, yoff, uvoff);

  1662.     } else {

  1663.         if ((ret = inter_recon(avctx)) < 0)

  1664.             return ret;

  1665.     }

  1666.     if (emu[0]) {

  1667.         int w = FFMIN(s->cols - col, w4) * 8;

  1668.         int h = FFMIN(s->rows - row, h4) * 8;

  1669.         int n, o = 0;

  1670. 

  1671.         for (n = 0; o < w; n++) {

  1672.             int bw = 64 >> n;

  1673. 

  1674.             av_assert2(n <= 4);

  1675.             if (w & bw) {

  1676.                 s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o,

  1677.                                          f->linesize[0],

  1678.                                          s->tmp_y + o,

  1679.                                          64, h, 0, 0);

  1680.                 o += bw;

  1681.             }

  1682.         }

  1683.     }

  1684.     if (emu[1]) {

  1685.         int w = FFMIN(s->cols - col, w4) * 4;

  1686.         int h = FFMIN(s->rows - row, h4) * 4;

  1687.         int n, o = 0;

  1688. 

  1689.         for (n = 1; o < w; n++) {

  1690.             int bw = 64 >> n;

  1691. 

  1692.             av_assert2(n <= 4);

  1693.             if (w & bw) {

  1694.                 s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o,

  1695.                                          f->linesize[1],

  1696.                                          s->tmp_uv[0] + o,

  1697.                                          32, h, 0, 0);

  1698.                 s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o,

  1699.                                          f->linesize[2],

  1700.                                          s->tmp_uv[1] + o,

  1701.                                          32, h, 0, 0);

  1702.                 o += bw;

  1703.             }

  1704.         }

  1705.     }

  1706. 

  1707.     // pick filter level and find edges to apply filter to

  1708.     if (s->filter.level &&

  1709.         (lvl = s->segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]

  1710.                                                     [b->mode[3] != ZEROMV]) > 0) {

  1711.         int x_end = FFMIN(s->cols - col, w4);

  1712.         int y_end = FFMIN(s->rows - row, h4);

  1713.         int skip_inter = !b->intra && b->skip;

  1714. 

  1715.         for (y = 0; y < h4; y++)

  1716.             memset(&lflvl->level[((row & 7) + y) * 8 + (col & 7)], lvl, w4);

  1717.         mask_edges(lflvl, 0, row & 7, col & 7, x_end, y_end, 0, 0, b->tx, skip_inter);

  1718.         mask_edges(lflvl, 1, row & 7, col & 7, x_end, y_end,

  1719.                    s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,

  1720.                    s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,

  1721.                    b->uvtx, skip_inter);

  1722. 

  1723.         if (!s->filter.lim_lut[lvl]) {

  1724.             int sharp = s->filter.sharpness;

  1725.             int limit = lvl;

  1726. 

  1727.             if (sharp > 0) {

  1728.                 limit >>= (sharp + 3) >> 2;

  1729.                 limit   = FFMIN(limit, 9 - sharp);

  1730.             }

  1731.             limit = FFMAX(limit, 1);

  1732. 

  1733.             s->filter.lim_lut[lvl]   = limit;

  1734.             s->filter.mblim_lut[lvl] = 2 * (lvl + 2) + limit;

  1735.         }

  1736.     }

  1737. 

  1738.     if (s->pass == 2) {

  1739.         s->b++;

  1740.         s->block      += w4 * h4 * 64;

  1741.         s->uvblock[0] += w4 * h4 * 16;

  1742.         s->uvblock[1] += w4 * h4 * 16;

  1743.         s->eob        += w4 * h4 * 4;

  1744.         s->uveob[0]   += w4 * h4;

  1745.         s->uveob[1]   += w4 * h4;

  1746.     }

  1747. 

  1748.     return 0;

  1749. }