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.
42730 Commits
32 Branches
283MB
Tree: b06cb15b9d
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'b06cb15b9d'
${ noResults }
FFmpeg/libavcodec/vp9block.c

1686 lines
72KB
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 "get_bits.h"

  28. #include "internal.h"

  29. #include "videodsp.h"

  30. #include "vp56.h"

  31. #include "vp9.h"

  32. #include "vp9data.h"

  33. 

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

  35.     {

  36.         { 16, 16 }, { 16, 8 }, { 8, 16 }, { 8, 8 }, { 8, 4 }, { 4, 8 },

  37.         {  4,  4 }, {  4, 2 }, { 2,  4 }, { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 },

  38.     },  {

  39.         {  8,  8 }, {  8, 4 }, { 4,  8 }, { 4, 4 }, { 4, 2 }, { 2, 4 },

  40.         {  2,  2 }, {  2, 1 }, { 1,  2 }, { 1, 1 }, { 1, 1 }, { 1, 1 }, { 1, 1 },

  41.     }

  42. };

  43. 

  44. // differential forward probability updates

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

  46. {

  47.     static const uint8_t left_ctx[N_BS_SIZES] = {

  48.         0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf

  49.     };

  50.     static const uint8_t above_ctx[N_BS_SIZES] = {

  51.         0x0, 0x0, 0x8, 0x8, 0x8, 0xc, 0xc, 0xc, 0xe, 0xe, 0xe, 0xf, 0xf

  52.     };

  53.     static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {

  54.         TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,

  55.         TX_16X16, TX_8X8,   TX_8X8,   TX_8X8,   TX_4X4,   TX_4X4,  TX_4X4

  56.     };

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

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

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

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

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

  62.     int y;

  63. 

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

  65.         b->seg_id = 0;

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

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

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

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

  70.                (s->segmentation.temporal &&

  71.                 vp56_rac_get_prob_branchy(&s->c,

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

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

  74.         int pred = MAX_SEGMENT - 1;

  75.         int x;

  76. 

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

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

  79.                 pred = FFMIN(pred,

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

  81.         b->seg_id = pred;

  82. 

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

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

  85.     } else {

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

  87.                                      s->prob.seg);

  88. 

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

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

  91.     }

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

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

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

  95.                    b->seg_id, w4);

  96.     }

  97. 

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

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

  100.     if (!b->skip) {

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

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

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

  104.     }

  105. 

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

  107.         b->intra = 1;

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

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

  110.     } else {

  111.         int c, bit;

  112. 

  113.         if (have_a && have_l) {

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

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

  116.         } else {

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

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

  119.         }

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

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

  122.         b->intra = !bit;

  123.     }

  124. 

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

  126.         int c;

  127.         if (have_a) {

  128.             if (have_l) {

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

  130.                      s->above_txfm_ctx[col]) +

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

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

  133.             } else {

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

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

  136.             }

  137.         } else if (have_l) {

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

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

  140.         } else {

  141.             c = 1;

  142.         }

  143.         switch (max_tx) {

  144.         case TX_32X32:

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

  146.             if (b->tx) {

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

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

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

  150.             }

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

  152.             break;

  153.         case TX_16X16:

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

  155.             if (b->tx)

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

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

  158.             break;

  159.         case TX_8X8:

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

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

  162.             break;

  163.         case TX_4X4:

  164.             b->tx = TX_4X4;

  165.             break;

  166.         }

  167.     } else {

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

  169.     }

  170. 

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

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

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

  174. 

  175.         b->comp = 0;

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

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

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

  179.             // simpler for now

  180.             b->mode[0] =

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

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

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

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

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

  186.                 l[0]       =

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

  188.             } else {

  189.                 l[0]       =

  190.                 a[1]       =

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

  192.             }

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

  194.                 b->mode[2] =

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

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

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

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

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

  200.                     l[1]       =

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

  202.                 } else {

  203.                     l[1]       =

  204.                     a[1]       =

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

  206.                 }

  207.             } else {

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

  209.                 l[1]       =

  210.                 a[1]       =

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

  212.             }

  213.         } else {

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

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

  216.             b->mode[3] =

  217.             b->mode[2] =

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

  219.             // FIXME this can probably be optimized

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

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

  222.         }

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

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

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

  226.         b->comp = 0;

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

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

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

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

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

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

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

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

  235.             } else {

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

  237.             }

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

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

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

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

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

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

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

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

  246.                 } else {

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

  248.                 }

  249.             } else {

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

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

  252.             }

  253.         } else {

  254.             static const uint8_t size_group[10] = {

  255.                 3, 3, 3, 3, 2, 2, 2, 1, 1, 1

  256.             };

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

  258. 

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

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

  261.             b->mode[1] =

  262.             b->mode[2] =

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

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

  265.         }

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

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

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

  269.     } else {

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

  271.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  272.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  273.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  274.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  275.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  276.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  277.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  278.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  279.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

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

  281.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  282.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  283.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },

  284.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },

  285.         };

  286. 

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

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

  289.             b->comp   = 0;

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

  291.         } else {

  292.             // read comp_pred flag

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

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

  295.             } else {

  296.                 int c;

  297. 

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

  299.                 if (have_a) {

  300.                     if (have_l) {

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

  302.                             c = 4;

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

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

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

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

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

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

  309.                         } else {

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

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

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

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

  314.                         }

  315.                     } else {

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

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

  318.                     }

  319.                 } else if (have_l) {

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

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

  322.                 } else {

  323.                     c = 1;

  324.                 }

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

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

  327.             }

  328. 

  329.             // read actual references

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

  331.             // memory accesses below

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

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

  334. 

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

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

  337.                 if (have_a) {

  338.                     if (have_l) {

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

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

  341.                                 c = 2;

  342.                             } else {

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

  344.                             }

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

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

  347.                         } else {

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

  349. 

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

  351.                                 c = 0;

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

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

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

  355.                                     c = 4;

  356.                                 } else {

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

  358.                                 }

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

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

  361.                                     c = 1;

  362.                                 } else {

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

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

  365.                                 }

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

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

  368.                                     c = 1;

  369.                                 } else {

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

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

  372.                                 }

  373.                             } else {

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

  375.                             }

  376.                         }

  377.                     } else {

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

  379.                             c = 2;

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

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

  382.                         } else {

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

  384.                         }

  385.                     }

  386.                 } else if (have_l) {

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

  388.                         c = 2;

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

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

  391.                     } else {

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

  393.                     }

  394.                 } else {

  395.                     c = 2;

  396.                 }

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

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

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

  400.             } else { /* single reference */

  401.                 int bit, c;

  402. 

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

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

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

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

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

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

  409.                             } else {

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

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

  412.                             }

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

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

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

  416.                         } else {

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

  418.                         }

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

  420.                         c = 2;

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

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

  423.                     } else {

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

  425.                     }

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

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

  428.                         c = 2;

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

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

  431.                     } else {

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

  433.                     }

  434.                 } else {

  435.                     c = 2;

  436.                 }

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

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

  439.                 if (!bit) {

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

  441.                 } else {

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

  443.                     if (have_a) {

  444.                         if (have_l) {

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

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

  447.                                     c = 2;

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

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

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

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

  452.                                     c = 3;

  453.                                 } else {

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

  455.                                 }

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

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

  458.                                     c = 2;

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

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

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

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

  463.                                     c = 3;

  464.                                 } else {

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

  466.                                 }

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

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

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

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

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

  472.                                     } else {

  473.                                         c = 2;

  474.                                     }

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

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

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

  478.                                 } else {

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

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

  481.                                 }

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

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

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

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

  486.                                 } else {

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

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

  489.                                 }

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

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

  492.                                     c = 3;

  493.                                 } else {

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

  495.                                 }

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

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

  498.                             } else {

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

  500.                                     2 * (s->above_ref_ctx[col] == 1);

  501.                             }

  502.                         } else {

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

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

  505.                                 c = 2;

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

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

  508.                             } else {

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

  510.                             }

  511.                         }

  512.                     } else if (have_l) {

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

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

  515.                             c = 2;

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

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

  518.                         } else {

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

  520.                         }

  521.                     } else {

  522.                         c = 2;

  523.                     }

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

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

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

  527.                 }

  528.             }

  529.         }

  530. 

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

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

  533.                 b->mode[0] =

  534.                 b->mode[1] =

  535.                 b->mode[2] =

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

  537.             } else {

  538.                 static const uint8_t off[10] = {

  539.                     3, 0, 0, 1, 0, 0, 0, 0, 0, 0

  540.                 };

  541. 

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

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

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

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

  546. 

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

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

  549.                 b->mode[1] =

  550.                 b->mode[2] =

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

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

  553.             }

  554.         }

  555. 

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

  557.             int c;

  558. 

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

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

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

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

  563.                 } else {

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

  565.                 }

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

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

  568.             } else {

  569.                 c = 3;

  570.             }

  571. 

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

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

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

  575.         } else {

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

  577.         }

  578. 

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

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

  581. 

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

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

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

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

  586. 

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

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

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

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

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

  592.             } else {

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

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

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

  596.             }

  597. 

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

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

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

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

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

  603. 

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

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

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

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

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

  609.                 } else {

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

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

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

  613.                 }

  614.             } else {

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

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

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

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

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

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

  621.             }

  622.         } else {

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

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

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

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

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

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

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

  630.         }

  631.     }

  632. 

  633.     // FIXME this can probably be optimized

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  651.         }

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

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

  654. 

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

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

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

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

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

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

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

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

  663.         } else {

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

  665. 

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

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

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

  669.             }

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

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

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

  673.             }

  674.         }

  675. 

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

  677.                          // as a direct check in above branches

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

  679. 

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

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

  682.         }

  683.     }

  684. 

  685.     // FIXME kinda ugly

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

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

  688. 

  689.         if (b->intra) {

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

  691.                 s->mv[0][o + x].ref[0] =

  692.                 s->mv[0][o + x].ref[1] = -1;

  693.             }

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

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

  696.                 s->mv[0][o + x].ref[0] = b->ref[0];

  697.                 s->mv[0][o + x].ref[1] = b->ref[1];

  698.                 AV_COPY32(&s->mv[0][o + x].mv[0], &b->mv[3][0]);

  699.                 AV_COPY32(&s->mv[0][o + x].mv[1], &b->mv[3][1]);

  700.             }

  701.         } else {

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

  703.                 s->mv[0][o + x].ref[0] = b->ref[0];

  704.                 s->mv[0][o + x].ref[1] = -1;

  705.                 AV_COPY32(&s->mv[0][o + x].mv[0], &b->mv[3][0]);

  706.             }

  707.         }

  708.     }

  709. }

  710. 

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

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

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

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

  715.                                int nnz, const int16_t *scan,

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

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

  718. {

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

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

  721.     uint8_t cache[1024];

  722. 

  723.     do {

  724.         int val, rc;

  725. 

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

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

  728.         if (!val)

  729.             break;

  730. 

  731. skip_eob:

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

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

  734.             if (!--band_left)

  735.                 band_left = band_counts[++band];

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

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

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

  739.             if (++i == n_coeffs)

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

  741.             goto skip_eob;

  742.         }

  743. 

  744.         rc = scan[i];

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

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

  747.             val       = 1;

  748.             cache[rc] = 1;

  749.         } else {

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

  751.             if (!tp[3])

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

  753. 

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

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

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

  757.                     cache[rc] = val = 2;

  758.                 } else {

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

  760.                     cache[rc] = 3;

  761.                 }

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

  763.                 cache[rc] = 4;

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

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

  766.                 } else {

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

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

  769.                 }

  770.             } else { // cat 3-6

  771.                 cache[rc] = 5;

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

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

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

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

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

  777.                     } else {

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

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

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

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

  782.                     }

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

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

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

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

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

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

  789.                 } else {

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  804.                 }

  805.             }

  806.         }

  807.         if (!--band_left)

  808.             band_left = band_counts[++band];

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

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

  811.         else

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

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

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

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

  816. 

  817.     return i;

  818. }

  819. 

  820. static int decode_coeffs(AVCodecContext *avctx)

  821. {

  822.     VP9Context *s = avctx->priv_data;

  823.     VP9Block *const b = &s->b;

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  842.         { 1, 2, 3, 4,  3,   16 - 13, 0 },

  843.         { 1, 2, 3, 4, 11,   64 - 21, 0 },

  844.         { 1, 2, 3, 4, 11,  256 - 21, 0 },

  845.         { 1, 2, 3, 4, 11, 1024 - 21, 0 },

  846.     };

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

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

  849. 

  850.     /* y tokens */

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

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

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

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

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

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

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

  858.     }

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

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

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

  862.                                                                 b->bs > BS_8x8 ?

  863.                                                                 n : 0]];

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

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

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

  867.                                            ynbs[txtp], y_band_counts,

  868.                                            qmul[0])) < 0)

  869.                 return ret;

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

  871.             if (b->tx > TX_8X8)

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

  873.             else

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

  875.         }

  876.     }

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

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

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

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

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

  882.     }

  883. 

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

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

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

  887.     w4    >>= 1;

  888.     h4    >>= 1;

  889.     end_x >>= 1;

  890.     end_y >>= 1;

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

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

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

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

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

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

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

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

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

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

  901.         }

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

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

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

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

  906.                                                16 * uvstep, b->uvtx, c, e, p,

  907.                                                nnz, uvscan, uvnb,

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

  909.                     return ret;

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

  911.                 if (b->uvtx > TX_8X8)

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

  913.                 else

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

  915.             }

  916.         }

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

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

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

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

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

  922.         }

  923.     }

  924. 

  925.     return 0;

  926. }

  927. 

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

  929.                                              uint8_t **a,

  930.                                              uint8_t *dst_edge,

  931.                                              ptrdiff_t stride_edge,

  932.                                              uint8_t *dst_inner,

  933.                                              ptrdiff_t stride_inner,

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

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

  936.                                              int p)

  937. {

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

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

  940.     int have_right = x < w - 1;

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

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

  943.                                    { DC_127_PRED,          VERT_PRED            } },

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

  945.                                    { HOR_PRED,             HOR_PRED             } },

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

  947.                                    { LEFT_DC_PRED,         DC_PRED              } },

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

  949.                                    { DC_127_PRED,          DIAG_DOWN_LEFT_PRED  } },

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

  951.                                    { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },

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

  953.                                    { VERT_RIGHT_PRED,      VERT_RIGHT_PRED      } },

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

  955.                                    { HOR_DOWN_PRED,        HOR_DOWN_PRED        } },

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

  957.                                    { DC_127_PRED,          VERT_LEFT_PRED       } },

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

  959.                                    { HOR_UP_PRED,          HOR_UP_PRED          } },

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

  961.                                    { HOR_PRED,             TM_VP8_PRED          } },

  962.     };

  963.     static const struct {

  964.         uint8_t needs_left:1;

  965.         uint8_t needs_top:1;

  966.         uint8_t needs_topleft:1;

  967.         uint8_t needs_topright:1;

  968.     } edges[N_INTRA_PRED_MODES] = {

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

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

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

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

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

  974.                                    .needs_topleft = 1 },

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

  976.                                    .needs_topleft = 1 },

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

  978.                                    .needs_topleft = 1 },

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

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

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

  982.                                    .needs_topleft = 1 },

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

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

  985.         [DC_128_PRED]          = { 0 },

  986.         [DC_127_PRED]          = { 0 },

  987.         [DC_129_PRED]          = { 0 }

  988.     };

  989. 

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

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

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

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

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

  995.         int n_px_need_tr = 0;

  996. 

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

  998.             n_px_need_tr = 4;

  999. 

  1000.         // if top of sb64-row, use s->intra_pred_data[] instead of

  1001.         // dst[-stride] for intra prediction (it contains pre- instead of

  1002.         // post-loopfilter data)

  1003.         if (have_top) {

  1004.             top = !(row & 7) && !y ?

  1005.                   s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  1006.                   y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];

  1007.             if (have_left)

  1008.                 topleft = !(row & 7) && !y ?

  1009.                           s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  1010.                           y == 0 || x == 0 ? &dst_edge[-stride_edge] :

  1011.                           &dst_inner[-stride_inner];

  1012.         }

  1013. 

  1014.         if (have_top &&

  1015.             (!edges[mode].needs_topleft || (have_left && top == topleft)) &&

  1016.             (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&

  1017.             n_px_need + n_px_need_tr <= n_px_have) {

  1018.             *a = top;

  1019.         } else {

  1020.             if (have_top) {

  1021.                 if (n_px_need <= n_px_have) {

  1022.                     memcpy(*a, top, n_px_need);

  1023.                 } else {

  1024.                     memcpy(*a, top, n_px_have);

  1025.                     memset(&(*a)[n_px_have], (*a)[n_px_have - 1],

  1026.                            n_px_need - n_px_have);

  1027.                 }

  1028.             } else {

  1029.                 memset(*a, 127, n_px_need);

  1030.             }

  1031.             if (edges[mode].needs_topleft) {

  1032.                 if (have_left && have_top)

  1033.                     (*a)[-1] = topleft[-1];

  1034.                 else

  1035.                     (*a)[-1] = have_top ? 129 : 127;

  1036.             }

  1037.             if (tx == TX_4X4 && edges[mode].needs_topright) {

  1038.                 if (have_top && have_right &&

  1039.                     n_px_need + n_px_need_tr <= n_px_have) {

  1040.                     memcpy(&(*a)[4], &top[4], 4);

  1041.                 } else {

  1042.                     memset(&(*a)[4], (*a)[3], 4);

  1043.                 }

  1044.             }

  1045.         }

  1046.     }

  1047.     if (edges[mode].needs_left) {

  1048.         if (have_left) {

  1049.             int i;

  1050.             int n_px_need = 4 << tx;

  1051.             int n_px_have = (((s->rows - row) << !p) - y) * 4;

  1052.             uint8_t *dst     = x == 0 ? dst_edge : dst_inner;

  1053.             ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;

  1054. 

  1055.             if (n_px_need <= n_px_have) {

  1056.                 for (i = 0; i < n_px_need; i++)

  1057.                     l[i] = dst[i * stride - 1];

  1058.             } else {

  1059.                 for (i = 0; i < n_px_have; i++)

  1060.                     l[i] = dst[i * stride - 1];

  1061.                 memset(&l[i], l[i - 1], n_px_need - n_px_have);

  1062.             }

  1063.         } else {

  1064.             memset(l, 129, 4 << tx);

  1065.         }

  1066.     }

  1067. 

  1068.     return mode;

  1069. }

  1070. 

  1071. static void intra_recon(AVCodecContext *avctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  1072. {

  1073.     VP9Context *s = avctx->priv_data;

  1074.     VP9Block *const b = &s->b;

  1075.     int row = b->row, col = b->col;

  1076.     int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  1077.     int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  1078.     int end_x = FFMIN(2 * (s->cols - col), w4);

  1079.     int end_y = FFMIN(2 * (s->rows - row), h4);

  1080.     int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  1081.     int uvstep1d = 1 << b->uvtx, p;

  1082.     uint8_t *dst = b->dst[0], *dst_r = s->cur_frame->data[0] + y_off;

  1083. 

  1084.     for (n = 0, y = 0; y < end_y; y += step1d) {

  1085.         uint8_t *ptr = dst, *ptr_r = dst_r;

  1086.         for (x = 0; x < end_x;

  1087.              x += step1d, ptr += 4 * step1d, ptr_r += 4 * step1d, n += step) {

  1088.             int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?

  1089.                                y * 2 + x : 0];

  1090.             LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);

  1091.             uint8_t *a = &a_buf[16], l[32];

  1092.             enum TxfmType txtp = ff_vp9_intra_txfm_type[mode];

  1093.             int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  1094. 

  1095.             mode = check_intra_mode(s, mode, &a, ptr_r,

  1096.                                     s->cur_frame->linesize[0],

  1097.                                     ptr, b->y_stride, l,

  1098.                                     col, x, w4, row, y, b->tx, 0);

  1099.             s->dsp.intra_pred[b->tx][mode](ptr, b->y_stride, l, a);

  1100.             if (eob)

  1101.                 s->dsp.itxfm_add[tx][txtp](ptr, b->y_stride,

  1102.                                            s->block + 16 * n, eob);

  1103.         }

  1104.         dst_r += 4 * s->cur_frame->linesize[0] * step1d;

  1105.         dst   += 4 * b->y_stride * step1d;

  1106.     }

  1107. 

  1108.     // U/V

  1109.     h4    >>= 1;

  1110.     w4    >>= 1;

  1111.     end_x >>= 1;

  1112.     end_y >>= 1;

  1113.     step    = 1 << (b->uvtx * 2);

  1114.     for (p = 0; p < 2; p++) {

  1115.         dst   = b->dst[1 + p];

  1116.         dst_r = s->cur_frame->data[1 + p] + uv_off;

  1117.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  1118.             uint8_t *ptr = dst, *ptr_r = dst_r;

  1119.             for (x = 0; x < end_x;

  1120.                  x += uvstep1d, ptr += 4 * uvstep1d,

  1121.                  ptr_r += 4 * uvstep1d, n += step) {

  1122.                 int mode = b->uvmode;

  1123.                 LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);

  1124.                 uint8_t *a = &a_buf[16], l[32];

  1125.                 int eob    = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n])

  1126.                                               : s->uveob[p][n];

  1127. 

  1128.                 mode = check_intra_mode(s, mode, &a, ptr_r,

  1129.                                         s->cur_frame->linesize[1],

  1130.                                         ptr, b->uv_stride, l,

  1131.                                         col, x, w4, row, y, b->uvtx, p + 1);

  1132.                 s->dsp.intra_pred[b->uvtx][mode](ptr, b->uv_stride, l, a);

  1133.                 if (eob)

  1134.                     s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, b->uv_stride,

  1135.                                                     s->uvblock[p] + 16 * n,

  1136.                                                     eob);

  1137.             }

  1138.             dst_r += 4 * uvstep1d * s->cur_frame->linesize[1];

  1139.             dst   += 4 * uvstep1d * b->uv_stride;

  1140.         }

  1141.     }

  1142. }

  1143. 

  1144. static av_always_inline void mc_luma_dir(VP9Context *s, vp9_mc_func(*mc)[2],

  1145.                                          uint8_t *dst, ptrdiff_t dst_stride,

  1146.                                          const uint8_t *ref,

  1147.                                          ptrdiff_t ref_stride,

  1148.                                          ptrdiff_t y, ptrdiff_t x,

  1149.                                          const VP56mv *mv,

  1150.                                          int bw, int bh, int w, int h)

  1151. {

  1152.     int mx = mv->x, my = mv->y;

  1153. 

  1154.     y   += my >> 3;

  1155.     x   += mx >> 3;

  1156.     ref += y * ref_stride + x;

  1157.     mx  &= 7;

  1158.     my  &= 7;

  1159.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  1160.     if (x < !!mx * 3 || y < !!my * 3 ||

  1161.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  1162.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1163.                                  ref - !!my * 3 * ref_stride - !!mx * 3,

  1164.                                  80,

  1165.                                  ref_stride,

  1166.                                  bw + !!mx * 7, bh + !!my * 7,

  1167.                                  x - !!mx * 3, y - !!my * 3, w, h);

  1168.         ref        = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  1169.         ref_stride = 80;

  1170.     }

  1171.     mc[!!mx][!!my](dst, ref, dst_stride, ref_stride, bh, mx << 1, my << 1);

  1172. }

  1173. 

  1174. static av_always_inline void mc_chroma_dir(VP9Context *s, vp9_mc_func(*mc)[2],

  1175.                                            uint8_t *dst_u, uint8_t *dst_v,

  1176.                                            ptrdiff_t dst_stride,

  1177.                                            const uint8_t *ref_u,

  1178.                                            ptrdiff_t src_stride_u,

  1179.                                            const uint8_t *ref_v,

  1180.                                            ptrdiff_t src_stride_v,

  1181.                                            ptrdiff_t y, ptrdiff_t x,

  1182.                                            const VP56mv *mv,

  1183.                                            int bw, int bh, int w, int h)

  1184. {

  1185.     int mx = mv->x, my = mv->y;

  1186. 

  1187.     y     += my >> 4;

  1188.     x     += mx >> 4;

  1189.     ref_u += y * src_stride_u + x;

  1190.     ref_v += y * src_stride_v + x;

  1191.     mx    &= 15;

  1192.     my    &= 15;

  1193.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  1194.     if (x < !!mx * 3 || y < !!my * 3 ||

  1195.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  1196.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1197.                                  ref_u - !!my * 3 * src_stride_u - !!mx * 3,

  1198.                                  80,

  1199.                                  src_stride_u,

  1200.                                  bw + !!mx * 7, bh + !!my * 7,

  1201.                                  x - !!mx * 3, y - !!my * 3, w, h);

  1202.         ref_u = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  1203.         mc[!!mx][!!my](dst_u, ref_u, dst_stride, 80, bh, mx, my);

  1204. 

  1205.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1206.                                  ref_v - !!my * 3 * src_stride_v - !!mx * 3,

  1207.                                  80,

  1208.                                  src_stride_v,

  1209.                                  bw + !!mx * 7, bh + !!my * 7,

  1210.                                  x - !!mx * 3, y - !!my * 3, w, h);

  1211.         ref_v = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  1212.         mc[!!mx][!!my](dst_v, ref_v, dst_stride, 80, bh, mx, my);

  1213.     } else {

  1214.         mc[!!mx][!!my](dst_u, ref_u, dst_stride, src_stride_u, bh, mx, my);

  1215.         mc[!!mx][!!my](dst_v, ref_v, dst_stride, src_stride_v, bh, mx, my);

  1216.     }

  1217. }

  1218. 

  1219. static int inter_recon(AVCodecContext *avctx)

  1220. {

  1221.     static const uint8_t bwlog_tab[2][N_BS_SIZES] = {

  1222.         { 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4 },

  1223.         { 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 4, 4 },

  1224.     };

  1225.     VP9Context *s = avctx->priv_data;

  1226.     VP9Block *const b = &s->b;

  1227.     int row = b->row, col = b->col;

  1228.     AVFrame *ref1 = s->refs[s->refidx[b->ref[0]]];

  1229.     AVFrame *ref2 = b->comp ? s->refs[s->refidx[b->ref[1]]] : NULL;

  1230.     int w = avctx->width, h = avctx->height;

  1231.     ptrdiff_t ls_y = b->y_stride, ls_uv = b->uv_stride;

  1232. 

  1233.     if (!ref1->data[0] || (b->comp && !ref2->data[0]))

  1234.         return AVERROR_INVALIDDATA;

  1235. 

  1236.     // y inter pred

  1237.     if (b->bs > BS_8x8) {

  1238.         if (b->bs == BS_8x4) {

  1239.             mc_luma_dir(s, s->dsp.mc[3][b->filter][0], b->dst[0], ls_y,

  1240.                         ref1->data[0], ref1->linesize[0],

  1241.                         row << 3, col << 3, &b->mv[0][0], 8, 4, w, h);

  1242.             mc_luma_dir(s, s->dsp.mc[3][b->filter][0],

  1243.                         b->dst[0] + 4 * ls_y, ls_y,

  1244.                         ref1->data[0], ref1->linesize[0],

  1245.                         (row << 3) + 4, col << 3, &b->mv[2][0], 8, 4, w, h);

  1246. 

  1247.             if (b->comp) {

  1248.                 mc_luma_dir(s, s->dsp.mc[3][b->filter][1], b->dst[0], ls_y,

  1249.                             ref2->data[0], ref2->linesize[0],

  1250.                             row << 3, col << 3, &b->mv[0][1], 8, 4, w, h);

  1251.                 mc_luma_dir(s, s->dsp.mc[3][b->filter][1],

  1252.                             b->dst[0] + 4 * ls_y, ls_y,

  1253.                             ref2->data[0], ref2->linesize[0],

  1254.                             (row << 3) + 4, col << 3, &b->mv[2][1], 8, 4, w, h);

  1255.             }

  1256.         } else if (b->bs == BS_4x8) {

  1257.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], b->dst[0], ls_y,

  1258.                         ref1->data[0], ref1->linesize[0],

  1259.                         row << 3, col << 3, &b->mv[0][0], 4, 8, w, h);

  1260.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], b->dst[0] + 4, ls_y,

  1261.                         ref1->data[0], ref1->linesize[0],

  1262.                         row << 3, (col << 3) + 4, &b->mv[1][0], 4, 8, w, h);

  1263. 

  1264.             if (b->comp) {

  1265.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], b->dst[0], ls_y,

  1266.                             ref2->data[0], ref2->linesize[0],

  1267.                             row << 3, col << 3, &b->mv[0][1], 4, 8, w, h);

  1268.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], b->dst[0] + 4, ls_y,

  1269.                             ref2->data[0], ref2->linesize[0],

  1270.                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 8, w, h);

  1271.             }

  1272.         } else {

  1273.             av_assert2(b->bs == BS_4x4);

  1274. 

  1275.             // FIXME if two horizontally adjacent blocks have the same MV,

  1276.             // do a w8 instead of a w4 call

  1277.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], b->dst[0], ls_y,

  1278.                         ref1->data[0], ref1->linesize[0],

  1279.                         row << 3, col << 3, &b->mv[0][0], 4, 4, w, h);

  1280.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], b->dst[0] + 4, ls_y,

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

  1282.                         row << 3, (col << 3) + 4, &b->mv[1][0], 4, 4, w, h);

  1283.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0],

  1284.                         b->dst[0] + 4 * ls_y, ls_y,

  1285.                         ref1->data[0], ref1->linesize[0],

  1286.                         (row << 3) + 4, col << 3, &b->mv[2][0], 4, 4, w, h);

  1287.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0],

  1288.                         b->dst[0] + 4 * ls_y + 4, ls_y,

  1289.                         ref1->data[0], ref1->linesize[0],

  1290.                         (row << 3) + 4, (col << 3) + 4, &b->mv[3][0], 4, 4, w, h);

  1291. 

  1292.             if (b->comp) {

  1293.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], b->dst[0], ls_y,

  1294.                             ref2->data[0], ref2->linesize[0],

  1295.                             row << 3, col << 3, &b->mv[0][1], 4, 4, w, h);

  1296.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], b->dst[0] + 4, ls_y,

  1297.                             ref2->data[0], ref2->linesize[0],

  1298.                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 4, w, h);

  1299.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1],

  1300.                             b->dst[0] + 4 * ls_y, ls_y,

  1301.                             ref2->data[0], ref2->linesize[0],

  1302.                             (row << 3) + 4, col << 3, &b->mv[2][1], 4, 4, w, h);

  1303.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1],

  1304.                             b->dst[0] + 4 * ls_y + 4, ls_y,

  1305.                             ref2->data[0], ref2->linesize[0],

  1306.                             (row << 3) + 4, (col << 3) + 4, &b->mv[3][1], 4, 4, w, h);

  1307.             }

  1308.         }

  1309.     } else {

  1310.         int bwl = bwlog_tab[0][b->bs];

  1311.         int bw  = bwh_tab[0][b->bs][0] * 4;

  1312.         int bh  = bwh_tab[0][b->bs][1] * 4;

  1313. 

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

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

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

  1317. 

  1318.         if (b->comp)

  1319.             mc_luma_dir(s, s->dsp.mc[bwl][b->filter][1], b->dst[0], ls_y,

  1320.                         ref2->data[0], ref2->linesize[0],

  1321.                         row << 3, col << 3, &b->mv[0][1], bw, bh, w, h);

  1322.     }

  1323. 

  1324.     // uv inter pred

  1325.     {

  1326.         int bwl = bwlog_tab[1][b->bs];

  1327.         int bw  = bwh_tab[1][b->bs][0] * 4, bh = bwh_tab[1][b->bs][1] * 4;

  1328.         VP56mv mvuv;

  1329. 

  1330.         w = (w + 1) >> 1;

  1331.         h = (h + 1) >> 1;

  1332.         if (b->bs > BS_8x8) {

  1333.             mvuv.x = ROUNDED_DIV(b->mv[0][0].x + b->mv[1][0].x +

  1334.                                  b->mv[2][0].x + b->mv[3][0].x, 4);

  1335.             mvuv.y = ROUNDED_DIV(b->mv[0][0].y + b->mv[1][0].y +

  1336.                                  b->mv[2][0].y + b->mv[3][0].y, 4);

  1337.         } else {

  1338.             mvuv = b->mv[0][0];

  1339.         }

  1340. 

  1341.         mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][0],

  1342.                       b->dst[1], b->dst[2], ls_uv,

  1343.                       ref1->data[1], ref1->linesize[1],

  1344.                       ref1->data[2], ref1->linesize[2],

  1345.                       row << 2, col << 2, &mvuv, bw, bh, w, h);

  1346. 

  1347.         if (b->comp) {

  1348.             if (b->bs > BS_8x8) {

  1349.                 mvuv.x = ROUNDED_DIV(b->mv[0][1].x + b->mv[1][1].x +

  1350.                                      b->mv[2][1].x + b->mv[3][1].x, 4);

  1351.                 mvuv.y = ROUNDED_DIV(b->mv[0][1].y + b->mv[1][1].y +

  1352.                                      b->mv[2][1].y + b->mv[3][1].y, 4);

  1353.             } else {

  1354.                 mvuv = b->mv[0][1];

  1355.             }

  1356.             mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][1],

  1357.                           b->dst[1], b->dst[2], ls_uv,

  1358.                           ref2->data[1], ref2->linesize[1],

  1359.                           ref2->data[2], ref2->linesize[2],

  1360.                           row << 2, col << 2, &mvuv, bw, bh, w, h);

  1361.         }

  1362.     }

  1363. 

  1364.     if (!b->skip) {

  1365.         /* mostly copied intra_reconn() */

  1366. 

  1367.         int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  1368.         int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  1369.         int end_x = FFMIN(2 * (s->cols - col), w4);

  1370.         int end_y = FFMIN(2 * (s->rows - row), h4);

  1371.         int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  1372.         int uvstep1d = 1 << b->uvtx, p;

  1373.         uint8_t *dst = b->dst[0];

  1374. 

  1375.         // y itxfm add

  1376.         for (n = 0, y = 0; y < end_y; y += step1d) {

  1377.             uint8_t *ptr = dst;

  1378.             for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d, n += step) {

  1379.                 int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  1380. 

  1381.                 if (eob)

  1382.                     s->dsp.itxfm_add[tx][DCT_DCT](ptr, b->y_stride,

  1383.                                                   s->block + 16 * n, eob);

  1384.             }

  1385.             dst += 4 * b->y_stride * step1d;

  1386.         }

  1387. 

  1388.         // uv itxfm add

  1389.         h4    >>= 1;

  1390.         w4    >>= 1;

  1391.         end_x >>= 1;

  1392.         end_y >>= 1;

  1393.         step    = 1 << (b->uvtx * 2);

  1394.         for (p = 0; p < 2; p++) {

  1395.             dst = b->dst[p + 1];

  1396.             for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  1397.                 uint8_t *ptr = dst;

  1398.                 for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d, n += step) {

  1399.                     int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n])

  1400.                                                : s->uveob[p][n];

  1401.                     if (eob)

  1402.                         s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, b->uv_stride,

  1403.                                                         s->uvblock[p] + 16 * n, eob);

  1404.                 }

  1405.                 dst += 4 * uvstep1d * b->uv_stride;

  1406.             }

  1407.         }

  1408.     }

  1409.     return 0;

  1410. }

  1411. 

  1412. static av_always_inline void mask_edges(VP9Filter *lflvl, int is_uv,

  1413.                                         int row_and_7, int col_and_7,

  1414.                                         int w, int h, int col_end, int row_end,

  1415.                                         enum TxfmMode tx, int skip_inter)

  1416. {

  1417.     // FIXME I'm pretty sure all loops can be replaced by a single LUT if

  1418.     // we make VP9Filter.mask uint64_t (i.e. row/col all single variable)

  1419.     // and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then

  1420.     // use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)

  1421. 

  1422.     // the intended behaviour of the vp9 loopfilter is to work on 8-pixel

  1423.     // edges. This means that for UV, we work on two subsampled blocks at

  1424.     // a time, and we only use the topleft block's mode information to set

  1425.     // things like block strength. Thus, for any block size smaller than

  1426.     // 16x16, ignore the odd portion of the block.

  1427.     if (tx == TX_4X4 && is_uv) {

  1428.         if (h == 1) {

  1429.             if (row_and_7 & 1)

  1430.                 return;

  1431.             if (!row_end)

  1432.                 h += 1;

  1433.         }

  1434.         if (w == 1) {

  1435.             if (col_and_7 & 1)

  1436.                 return;

  1437.             if (!col_end)

  1438.                 w += 1;

  1439.         }

  1440.     }

  1441. 

  1442.     if (tx == TX_4X4 && !skip_inter) {

  1443.         int t = 1 << col_and_7, m_col = (t << w) - t, y;

  1444.         int m_col_odd = (t << (w - 1)) - t;

  1445. 

  1446.         // on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide

  1447.         if (is_uv) {

  1448.             int m_row_8 = m_col & 0x01, m_row_4 = m_col - m_row_8;

  1449. 

  1450.             for (y = row_and_7; y < h + row_and_7; y++) {

  1451.                 int col_mask_id = 2 - !(y & 7);

  1452. 

  1453.                 lflvl->mask[is_uv][0][y][1] |= m_row_8;

  1454.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  1455.                 // for odd lines, if the odd col is not being filtered,

  1456.                 // skip odd row also:

  1457.                 // .---. <-- a

  1458.                 // |   |

  1459.                 // |___| <-- b

  1460.                 // ^   ^

  1461.                 // c   d

  1462.                 //

  1463.                 // if a/c are even row/col and b/d are odd, and d is skipped,

  1464.                 // e.g. right edge of size-66x66.webm, then skip b also (bug)

  1465.                 if ((col_end & 1) && (y & 1)) {

  1466.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col_odd;

  1467.                 } else {

  1468.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col;

  1469.                 }

  1470.             }

  1471.         } else {

  1472.             int m_row_8 = m_col & 0x11, m_row_4 = m_col - m_row_8;

  1473. 

  1474.             for (y = row_and_7; y < h + row_and_7; y++) {

  1475.                 int col_mask_id = 2 - !(y & 3);

  1476. 

  1477.                 lflvl->mask[is_uv][0][y][1]           |= m_row_8; // row edge

  1478.                 lflvl->mask[is_uv][0][y][2]           |= m_row_4;

  1479.                 lflvl->mask[is_uv][1][y][col_mask_id] |= m_col; // col edge

  1480.                 lflvl->mask[is_uv][0][y][3]           |= m_col;

  1481.                 lflvl->mask[is_uv][1][y][3]           |= m_col;

  1482.             }

  1483.         }

  1484.     } else {

  1485.         int y, t = 1 << col_and_7, m_col = (t << w) - t;

  1486. 

  1487.         if (!skip_inter) {

  1488.             int mask_id = (tx == TX_8X8);

  1489.             int l2 = tx + is_uv - 1, step1d = 1 << l2;

  1490.             static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };

  1491.             int m_row = m_col & masks[l2];

  1492. 

  1493.             // at odd UV col/row edges tx16/tx32 loopfilter edges, force

  1494.             // 8wd loopfilter to prevent going off the visible edge.

  1495.             if (is_uv && tx > TX_8X8 && (w ^ (w - 1)) == 1) {

  1496.                 int m_row_16 = ((t << (w - 1)) - t) & masks[l2];

  1497.                 int m_row_8  = m_row - m_row_16;

  1498. 

  1499.                 for (y = row_and_7; y < h + row_and_7; y++) {

  1500.                     lflvl->mask[is_uv][0][y][0] |= m_row_16;

  1501.                     lflvl->mask[is_uv][0][y][1] |= m_row_8;

  1502.                 }

  1503.             } else {

  1504.                 for (y = row_and_7; y < h + row_and_7; y++)

  1505.                     lflvl->mask[is_uv][0][y][mask_id] |= m_row;

  1506.             }

  1507. 

  1508.             if (is_uv && tx > TX_8X8 && (h ^ (h - 1)) == 1) {

  1509.                 for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)

  1510.                     lflvl->mask[is_uv][1][y][0] |= m_col;

  1511.                 if (y - row_and_7 == h - 1)

  1512.                     lflvl->mask[is_uv][1][y][1] |= m_col;

  1513.             } else {

  1514.                 for (y = row_and_7; y < h + row_and_7; y += step1d)

  1515.                     lflvl->mask[is_uv][1][y][mask_id] |= m_col;

  1516.             }

  1517.         } else if (tx != TX_4X4) {

  1518.             int mask_id;

  1519. 

  1520.             mask_id = (tx == TX_8X8) || (is_uv && h == 1);

  1521.             lflvl->mask[is_uv][1][row_and_7][mask_id] |= m_col;

  1522.             mask_id = (tx == TX_8X8) || (is_uv && w == 1);

  1523.             for (y = row_and_7; y < h + row_and_7; y++)

  1524.                 lflvl->mask[is_uv][0][y][mask_id] |= t;

  1525.         } else if (is_uv) {

  1526.             int t8 = t & 0x01, t4 = t - t8;

  1527. 

  1528.             for (y = row_and_7; y < h + row_and_7; y++) {

  1529.                 lflvl->mask[is_uv][0][y][2] |= t4;

  1530.                 lflvl->mask[is_uv][0][y][1] |= t8;

  1531.             }

  1532.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 7)] |= m_col;

  1533.         } else {

  1534.             int t8 = t & 0x11, t4 = t - t8;

  1535. 

  1536.             for (y = row_and_7; y < h + row_and_7; y++) {

  1537.                 lflvl->mask[is_uv][0][y][2] |= t4;

  1538.                 lflvl->mask[is_uv][0][y][1] |= t8;

  1539.             }

  1540.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 3)] |= m_col;

  1541.         }

  1542.     }

  1543. }

  1544. 

  1545. int ff_vp9_decode_block(AVCodecContext *avctx, int row, int col,

  1546.                         VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,

  1547.                         enum BlockLevel bl, enum BlockPartition bp)

  1548. {

  1549.     VP9Context *s = avctx->priv_data;

  1550.     VP9Block *const b = &s->b;

  1551.     enum BlockSize bs = bl * 3 + bp;

  1552.     int ret, y, w4 = bwh_tab[1][bs][0], h4 = bwh_tab[1][bs][1], lvl;

  1553.     int emu[2];

  1554. 

  1555.     b->row  = row;

  1556.     b->row7 = row & 7;

  1557.     b->col  = col;

  1558.     b->col7 = col & 7;

  1559. 

  1560.     s->min_mv.x = -(128 + col * 64);

  1561.     s->min_mv.y = -(128 + row * 64);

  1562.     s->max_mv.x = 128 + (s->cols - col - w4) * 64;

  1563.     s->max_mv.y = 128 + (s->rows - row - h4) * 64;

  1564. 

  1565.     b->bs = bs;

  1566.     decode_mode(s, b);

  1567.     b->uvtx = b->tx - (w4 * 2 == (1 << b->tx) || h4 * 2 == (1 << b->tx));

  1568. 

  1569.     if (!b->skip) {

  1570.         if ((ret = decode_coeffs(avctx)) < 0)

  1571.             return ret;

  1572.     } else {

  1573.         int pl;

  1574. 

  1575.         memset(&s->above_y_nnz_ctx[col * 2], 0, w4 * 2);

  1576.         memset(&s->left_y_nnz_ctx[(row & 7) << 1], 0, h4 * 2);

  1577.         for (pl = 0; pl < 2; pl++) {

  1578.             memset(&s->above_uv_nnz_ctx[pl][col], 0, w4);

  1579.             memset(&s->left_uv_nnz_ctx[pl][row & 7], 0, h4);

  1580.         }

  1581.     }

  1582. 

  1583.     /* Emulated overhangs if the stride of the target buffer can't hold.

  1584.      * This allows to support emu-edge and so on even if we have large

  1585.      * block overhangs. */

  1586.     emu[0] = (col + w4) * 8 > s->cur_frame->linesize[0] ||

  1587.              (row + h4) > s->rows;

  1588.     emu[1] = (col + w4) * 4 > s->cur_frame->linesize[1] ||

  1589.              (row + h4) > s->rows;

  1590.     if (emu[0]) {

  1591.         b->dst[0]   = s->tmp_y;

  1592.         b->y_stride = 64;

  1593.     } else {

  1594.         b->dst[0]   = s->cur_frame->data[0] + yoff;

  1595.         b->y_stride = s->cur_frame->linesize[0];

  1596.     }

  1597.     if (emu[1]) {

  1598.         b->dst[1]    = s->tmp_uv[0];

  1599.         b->dst[2]    = s->tmp_uv[1];

  1600.         b->uv_stride = 32;

  1601.     } else {

  1602.         b->dst[1]    = s->cur_frame->data[1] + uvoff;

  1603.         b->dst[2]    = s->cur_frame->data[2] + uvoff;

  1604.         b->uv_stride = s->cur_frame->linesize[1];

  1605.     }

  1606.     if (b->intra) {

  1607.         intra_recon(avctx, yoff, uvoff);

  1608.     } else {

  1609.         if ((ret = inter_recon(avctx)) < 0)

  1610.             return ret;

  1611.     }

  1612.     if (emu[0]) {

  1613.         int w = FFMIN(s->cols - col, w4) * 8;

  1614.         int h = FFMIN(s->rows - row, h4) * 8;

  1615.         int n, o = 0;

  1616. 

  1617.         for (n = 0; o < w; n++) {

  1618.             int bw = 64 >> n;

  1619. 

  1620.             av_assert2(n <= 4);

  1621.             if (w & bw) {

  1622.                 s->dsp.mc[n][0][0][0][0](s->cur_frame->data[0] + yoff + o,

  1623.                                          s->tmp_y + o,

  1624.                                          s->cur_frame->linesize[0],

  1625.                                          64, h, 0, 0);

  1626.                 o += bw;

  1627.             }

  1628.         }

  1629.     }

  1630.     if (emu[1]) {

  1631.         int w = FFMIN(s->cols - col, w4) * 4;

  1632.         int h = FFMIN(s->rows - row, h4) * 4;

  1633.         int n, o = 0;

  1634. 

  1635.         for (n = 1; o < w; n++) {

  1636.             int bw = 64 >> n;

  1637. 

  1638.             av_assert2(n <= 4);

  1639.             if (w & bw) {

  1640.                 s->dsp.mc[n][0][0][0][0](s->cur_frame->data[1] + uvoff + o,

  1641.                                          s->tmp_uv[0] + o,

  1642.                                          s->cur_frame->linesize[1],

  1643.                                          32, h, 0, 0);

  1644.                 s->dsp.mc[n][0][0][0][0](s->cur_frame->data[2] + uvoff + o,

  1645.                                          s->tmp_uv[1] + o,

  1646.                                          s->cur_frame->linesize[2],

  1647.                                          32, h, 0, 0);

  1648.                 o += bw;

  1649.             }

  1650.         }

  1651.     }

  1652. 

  1653.     // pick filter level and find edges to apply filter to

  1654.     if (s->filter.level &&

  1655.         (lvl = s->segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]

  1656.                                                     [b->mode[3] != ZEROMV]) > 0) {

  1657.         int x_end = FFMIN(s->cols - col, w4);

  1658.         int y_end = FFMIN(s->rows - row, h4);

  1659.         int skip_inter = !b->intra && b->skip;

  1660. 

  1661.         for (y = 0; y < h4; y++)

  1662.             memset(&lflvl->level[((row & 7) + y) * 8 + (col & 7)], lvl, w4);

  1663.         mask_edges(lflvl, 0, row & 7, col & 7, x_end, y_end, 0, 0, b->tx, skip_inter);

  1664.         mask_edges(lflvl, 1, row & 7, col & 7, x_end, y_end,

  1665.                    s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,

  1666.                    s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,

  1667.                    b->uvtx, skip_inter);

  1668. 

  1669.         if (!s->filter.lim_lut[lvl]) {

  1670.             int sharp = s->filter.sharpness;

  1671.             int limit = lvl;

  1672. 

  1673.             if (sharp > 0) {

  1674.                 limit >>= (sharp + 3) >> 2;

  1675.                 limit   = FFMIN(limit, 9 - sharp);

  1676.             }

  1677.             limit = FFMAX(limit, 1);

  1678. 

  1679.             s->filter.lim_lut[lvl]   = limit;

  1680.             s->filter.mblim_lut[lvl] = 2 * (lvl + 2) + limit;

  1681.         }

  1682.     }

  1683. 

  1684.     return 0;

  1685. }