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.
43764 Commits
32 Branches
2.4GB
Tree: c19830aa2c
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'c19830aa2c'
${ noResults }
FFmpeg/libavcodec/vp9block.c

1740 lines
73KB
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.         uint8_t *refsegmap = s->frames[LAST_FRAME].segmentation_map;

  74.         int pred = MAX_SEGMENT - 1;

  75.         int x;

  76. 

  77.         if (!s->last_uses_2pass)

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

  79. 

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

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

  82.                 pred = FFMIN(pred,

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

  84.         b->seg_id = pred;

  85. 

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

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

  88.     } else {

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

  90.                                      s->prob.seg);

  91. 

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

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

  94.     }

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

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

  97. 

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

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

  100.                    b->seg_id, w4);

  101.     }

  102. 

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

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

  105.     if (!b->skip) {

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

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

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

  109.     }

  110. 

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

  112.         b->intra = 1;

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

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

  115.     } else {

  116.         int c, bit;

  117. 

  118.         if (have_a && have_l) {

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

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

  121.         } else {

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

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

  124.         }

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

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

  127.         b->intra = !bit;

  128.     }

  129. 

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

  131.         int c;

  132.         if (have_a) {

  133.             if (have_l) {

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

  135.                      s->above_txfm_ctx[col]) +

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

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

  138.             } else {

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

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

  141.             }

  142.         } else if (have_l) {

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

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

  145.         } else {

  146.             c = 1;

  147.         }

  148.         switch (max_tx) {

  149.         case TX_32X32:

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

  151.             if (b->tx) {

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

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

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

  155.             }

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

  157.             break;

  158.         case TX_16X16:

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

  160.             if (b->tx)

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

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

  163.             break;

  164.         case TX_8X8:

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

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

  167.             break;

  168.         case TX_4X4:

  169.             b->tx = TX_4X4;

  170.             break;

  171.         }

  172.     } else {

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

  174.     }

  175. 

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

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

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

  179. 

  180.         b->comp = 0;

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

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

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

  184.             // simpler for now

  185.             b->mode[0] =

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

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

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

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

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

  191.                 l[0]       =

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

  193.             } else {

  194.                 l[0]       =

  195.                 a[1]       =

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

  197.             }

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

  199.                 b->mode[2] =

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

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

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

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

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

  205.                     l[1]       =

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

  207.                 } else {

  208.                     l[1]       =

  209.                     a[1]       =

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

  211.                 }

  212.             } else {

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

  214.                 l[1]       =

  215.                 a[1]       =

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

  217.             }

  218.         } else {

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

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

  221.             b->mode[3] =

  222.             b->mode[2] =

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

  224.             // FIXME this can probably be optimized

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

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

  227.         }

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

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

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

  231.         b->comp = 0;

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

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

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

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

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

  237.                 b->mode[1] = 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[1]]++;

  240.             } else {

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

  242.             }

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

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

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

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

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

  248.                     b->mode[3] = 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[3]]++;

  251.                 } else {

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

  253.                 }

  254.             } else {

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

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

  257.             }

  258.         } else {

  259.             static const uint8_t size_group[10] = {

  260.                 3, 3, 3, 3, 2, 2, 2, 1, 1, 1

  261.             };

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

  263. 

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

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

  266.             b->mode[1] =

  267.             b->mode[2] =

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

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

  270.         }

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

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

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

  274.     } else {

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

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

  287.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  288.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },

  289.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },

  290.         };

  291. 

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

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

  294.             b->comp   = 0;

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

  296.         } else {

  297.             // read comp_pred flag

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

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

  300.             } else {

  301.                 int c;

  302. 

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

  304.                 if (have_a) {

  305.                     if (have_l) {

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

  307.                             c = 4;

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

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

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

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

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

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

  314.                         } else {

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

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

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

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

  319.                         }

  320.                     } else {

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

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

  323.                     }

  324.                 } else if (have_l) {

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

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

  327.                 } else {

  328.                     c = 1;

  329.                 }

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

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

  332.             }

  333. 

  334.             // read actual references

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

  336.             // memory accesses below

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

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

  339. 

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

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

  342.                 if (have_a) {

  343.                     if (have_l) {

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

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

  346.                                 c = 2;

  347.                             } else {

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

  349.                             }

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

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

  352.                         } else {

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

  354. 

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

  356.                                 c = 0;

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

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

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

  360.                                     c = 4;

  361.                                 } else {

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

  363.                                 }

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

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

  366.                                     c = 1;

  367.                                 } else {

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

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

  370.                                 }

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

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

  373.                                     c = 1;

  374.                                 } else {

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

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

  377.                                 }

  378.                             } else {

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

  380.                             }

  381.                         }

  382.                     } else {

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

  384.                             c = 2;

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

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

  387.                         } else {

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

  389.                         }

  390.                     }

  391.                 } else if (have_l) {

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

  393.                         c = 2;

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

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

  396.                     } else {

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

  398.                     }

  399.                 } else {

  400.                     c = 2;

  401.                 }

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

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

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

  405.             } else { /* single reference */

  406.                 int bit, c;

  407. 

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

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

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

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

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

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

  414.                             } else {

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

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

  417.                             }

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

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

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

  421.                         } else {

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

  423.                         }

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

  425.                         c = 2;

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

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

  428.                     } else {

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

  430.                     }

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

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

  433.                         c = 2;

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

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

  436.                     } else {

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

  438.                     }

  439.                 } else {

  440.                     c = 2;

  441.                 }

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

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

  444.                 if (!bit) {

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

  446.                 } else {

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

  448.                     if (have_a) {

  449.                         if (have_l) {

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

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

  452.                                     c = 2;

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

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

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

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

  457.                                     c = 3;

  458.                                 } else {

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

  460.                                 }

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

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

  463.                                     c = 2;

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

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

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

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

  468.                                     c = 3;

  469.                                 } else {

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

  471.                                 }

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

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

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

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

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

  477.                                     } else {

  478.                                         c = 2;

  479.                                     }

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

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

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

  483.                                 } else {

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

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

  486.                                 }

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

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

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

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

  491.                                 } else {

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

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

  494.                                 }

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

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

  497.                                     c = 3;

  498.                                 } else {

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

  500.                                 }

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

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

  503.                             } else {

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

  505.                                     2 * (s->above_ref_ctx[col] == 1);

  506.                             }

  507.                         } else {

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

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

  510.                                 c = 2;

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

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

  513.                             } else {

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

  515.                             }

  516.                         }

  517.                     } else if (have_l) {

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

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

  520.                             c = 2;

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

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

  523.                         } else {

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

  525.                         }

  526.                     } else {

  527.                         c = 2;

  528.                     }

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

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

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

  532.                 }

  533.             }

  534.         }

  535. 

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

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

  538.                 b->mode[0] =

  539.                 b->mode[1] =

  540.                 b->mode[2] =

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

  542.             } else {

  543.                 static const uint8_t off[10] = {

  544.                     3, 0, 0, 1, 0, 0, 0, 0, 0, 0

  545.                 };

  546. 

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

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

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

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

  551. 

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

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

  554.                 b->mode[1] =

  555.                 b->mode[2] =

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

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

  558.             }

  559.         }

  560. 

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

  562.             int c;

  563. 

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

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

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

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

  568.                 } else {

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

  570.                 }

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

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

  573.             } else {

  574.                 c = 3;

  575.             }

  576. 

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

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

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

  580.         } else {

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

  582.         }

  583. 

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

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

  586. 

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

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

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

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

  591. 

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

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

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

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

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

  597.             } else {

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

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

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

  601.             }

  602. 

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

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

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

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

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

  608. 

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

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

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

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

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

  614.                 } else {

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

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

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

  618.                 }

  619.             } else {

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

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

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

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

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

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

  626.             }

  627.         } else {

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

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

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

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

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

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

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

  635.         }

  636.     }

  637. 

  638.     // FIXME this can probably be optimized

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  656.         }

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

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

  659. 

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

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

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

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

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

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

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

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

  668.         } else {

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

  670. 

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

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

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

  674.             }

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

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

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

  678.             }

  679.         }

  680. 

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

  682.                          // as a direct check in above branches

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

  684. 

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

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

  687.         }

  688.     }

  689. 

  690.     // FIXME kinda ugly

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

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

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

  694. 

  695.         if (b->intra) {

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

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

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

  699.             }

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

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

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

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

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

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

  706.             }

  707.         } else {

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

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

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

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

  712.             }

  713.         }

  714.     }

  715. }

  716. 

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

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

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

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

  721.                                int nnz, const int16_t *scan,

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

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

  724. {

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

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

  727.     uint8_t cache[1024];

  728. 

  729.     do {

  730.         int val, rc;

  731. 

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

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

  734.         if (!val)

  735.             break;

  736. 

  737. skip_eob:

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

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

  740.             if (!--band_left)

  741.                 band_left = band_counts[++band];

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

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

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

  745.             if (++i == n_coeffs)

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

  747.             goto skip_eob;

  748.         }

  749. 

  750.         rc = scan[i];

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

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

  753.             val       = 1;

  754.             cache[rc] = 1;

  755.         } else {

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

  757.             if (!tp[3])

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

  759. 

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

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

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

  763.                     cache[rc] = val = 2;

  764.                 } else {

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

  766.                     cache[rc] = 3;

  767.                 }

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

  769.                 cache[rc] = 4;

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

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

  772.                 } else {

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

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

  775.                 }

  776.             } else { // cat 3-6

  777.                 cache[rc] = 5;

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

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

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

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

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

  783.                     } else {

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

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

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

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

  788.                     }

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

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

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

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

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

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

  795.                 } else {

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  810.                 }

  811.             }

  812.         }

  813.         if (!--band_left)

  814.             band_left = band_counts[++band];

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

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

  817.         else

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

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

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

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

  822. 

  823.     return i;

  824. }

  825. 

  826. static int decode_coeffs(AVCodecContext *avctx)

  827. {

  828.     VP9Context *s = avctx->priv_data;

  829.     VP9Block *b = s->b;

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  848.         { 1, 2, 3, 4,  3,   16 - 13, 0 },

  849.         { 1, 2, 3, 4, 11,   64 - 21, 0 },

  850.         { 1, 2, 3, 4, 11,  256 - 21, 0 },

  851.         { 1, 2, 3, 4, 11, 1024 - 21, 0 },

  852.     };

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

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

  855. 

  856.     /* y tokens */

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

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

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

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

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

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

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

  864.     }

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

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

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

  868.                                                                 b->bs > BS_8x8 ?

  869.                                                                 n : 0]];

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

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

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

  873.                                            ynbs[txtp], y_band_counts,

  874.                                            qmul[0])) < 0)

  875.                 return ret;

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

  877.             if (b->tx > TX_8X8)

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

  879.             else

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

  881.         }

  882.     }

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

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

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

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

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

  888.     }

  889. 

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

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

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

  893.     w4    >>= 1;

  894.     h4    >>= 1;

  895.     end_x >>= 1;

  896.     end_y >>= 1;

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

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

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

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

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

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

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

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

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

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

  907.         }

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

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

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

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

  912.                                                16 * uvstep, b->uvtx, c, e, p,

  913.                                                nnz, uvscan, uvnb,

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

  915.                     return ret;

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

  917.                 if (b->uvtx > TX_8X8)

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

  919.                 else

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

  921.             }

  922.         }

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

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

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

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

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

  928.         }

  929.     }

  930. 

  931.     return 0;

  932. }

  933. 

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

  935.                                              uint8_t **a,

  936.                                              uint8_t *dst_edge,

  937.                                              ptrdiff_t stride_edge,

  938.                                              uint8_t *dst_inner,

  939.                                              ptrdiff_t stride_inner,

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

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

  942.                                              int p)

  943. {

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

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

  946.     int have_right = x < w - 1;

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

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

  949.                                    { DC_127_PRED,          VERT_PRED            } },

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

  951.                                    { HOR_PRED,             HOR_PRED             } },

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

  953.                                    { LEFT_DC_PRED,         DC_PRED              } },

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

  955.                                    { DC_127_PRED,          DIAG_DOWN_LEFT_PRED  } },

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

  957.                                    { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },

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

  959.                                    { VERT_RIGHT_PRED,      VERT_RIGHT_PRED      } },

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

  961.                                    { HOR_DOWN_PRED,        HOR_DOWN_PRED        } },

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

  963.                                    { DC_127_PRED,          VERT_LEFT_PRED       } },

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

  965.                                    { HOR_UP_PRED,          HOR_UP_PRED          } },

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

  967.                                    { HOR_PRED,             TM_VP8_PRED          } },

  968.     };

  969.     static const struct {

  970.         uint8_t needs_left:1;

  971.         uint8_t needs_top:1;

  972.         uint8_t needs_topleft:1;

  973.         uint8_t needs_topright:1;

  974.     } edges[N_INTRA_PRED_MODES] = {

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

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

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

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

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

  980.                                    .needs_topleft = 1 },

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

  982.                                    .needs_topleft = 1 },

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

  984.                                    .needs_topleft = 1 },

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

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

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

  988.                                    .needs_topleft = 1 },

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

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

  991.         [DC_128_PRED]          = { 0 },

  992.         [DC_127_PRED]          = { 0 },

  993.         [DC_129_PRED]          = { 0 }

  994.     };

  995. 

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

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

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

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

  1000.         int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !p) - x) * 4;

  1001.         int n_px_need_tr = 0;

  1002. 

  1003.         if (tx == TX_4X4 && edges[mode].needs_topright && have_right)

  1004.             n_px_need_tr = 4;

  1005. 

  1006.         // if top of sb64-row, use s->intra_pred_data[] instead of

  1007.         // dst[-stride] for intra prediction (it contains pre- instead of

  1008.         // post-loopfilter data)

  1009.         if (have_top) {

  1010.             top = !(row & 7) && !y ?

  1011.                   s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  1012.                   y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];

  1013.             if (have_left)

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

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

  1016.                           y == 0 || x == 0 ? &dst_edge[-stride_edge] :

  1017.                           &dst_inner[-stride_inner];

  1018.         }

  1019. 

  1020.         if (have_top &&

  1021.             (!edges[mode].needs_topleft || (have_left && top == topleft)) &&

  1022.             (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&

  1023.             n_px_need + n_px_need_tr <= n_px_have) {

  1024.             *a = top;

  1025.         } else {

  1026.             if (have_top) {

  1027.                 if (n_px_need <= n_px_have) {

  1028.                     memcpy(*a, top, n_px_need);

  1029.                 } else {

  1030.                     memcpy(*a, top, n_px_have);

  1031.                     memset(&(*a)[n_px_have], (*a)[n_px_have - 1],

  1032.                            n_px_need - n_px_have);

  1033.                 }

  1034.             } else {

  1035.                 memset(*a, 127, n_px_need);

  1036.             }

  1037.             if (edges[mode].needs_topleft) {

  1038.                 if (have_left && have_top)

  1039.                     (*a)[-1] = topleft[-1];

  1040.                 else

  1041.                     (*a)[-1] = have_top ? 129 : 127;

  1042.             }

  1043.             if (tx == TX_4X4 && edges[mode].needs_topright) {

  1044.                 if (have_top && have_right &&

  1045.                     n_px_need + n_px_need_tr <= n_px_have) {

  1046.                     memcpy(&(*a)[4], &top[4], 4);

  1047.                 } else {

  1048.                     memset(&(*a)[4], (*a)[3], 4);

  1049.                 }

  1050.             }

  1051.         }

  1052.     }

  1053.     if (edges[mode].needs_left) {

  1054.         if (have_left) {

  1055.             int i;

  1056.             int n_px_need = 4 << tx;

  1057.             int n_px_have = (((s->rows - row) << !p) - y) * 4;

  1058.             uint8_t *dst     = x == 0 ? dst_edge : dst_inner;

  1059.             ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;

  1060. 

  1061.             if (n_px_need <= n_px_have) {

  1062.                 for (i = 0; i < n_px_need; i++)

  1063.                     l[i] = dst[i * stride - 1];

  1064.             } else {

  1065.                 for (i = 0; i < n_px_have; i++)

  1066.                     l[i] = dst[i * stride - 1];

  1067.                 memset(&l[i], l[i - 1], n_px_need - n_px_have);

  1068.             }

  1069.         } else {

  1070.             memset(l, 129, 4 << tx);

  1071.         }

  1072.     }

  1073. 

  1074.     return mode;

  1075. }

  1076. 

  1077. static void intra_recon(AVCodecContext *avctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  1078. {

  1079.     VP9Context *s = avctx->priv_data;

  1080.     VP9Block *b = s->b;

  1081.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  1082.     int row = b->row, col = b->col;

  1083.     int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  1084.     int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  1085.     int end_x = FFMIN(2 * (s->cols - col), w4);

  1086.     int end_y = FFMIN(2 * (s->rows - row), h4);

  1087.     int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  1088.     int uvstep1d = 1 << b->uvtx, p;

  1089.     uint8_t *dst = b->dst[0], *dst_r = f->data[0] + y_off;

  1090. 

  1091.     for (n = 0, y = 0; y < end_y; y += step1d) {

  1092.         uint8_t *ptr = dst, *ptr_r = dst_r;

  1093.         for (x = 0; x < end_x;

  1094.              x += step1d, ptr += 4 * step1d, ptr_r += 4 * step1d, n += step) {

  1095.             int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?

  1096.                                y * 2 + x : 0];

  1097.             LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);

  1098.             uint8_t *a = &a_buf[16], l[32];

  1099.             enum TxfmType txtp = ff_vp9_intra_txfm_type[mode];

  1100.             int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  1101. 

  1102.             mode = check_intra_mode(s, mode, &a, ptr_r,

  1103.                                     f->linesize[0],

  1104.                                     ptr, b->y_stride, l,

  1105.                                     col, x, w4, row, y, b->tx, 0);

  1106.             s->dsp.intra_pred[b->tx][mode](ptr, b->y_stride, l, a);

  1107.             if (eob)

  1108.                 s->dsp.itxfm_add[tx][txtp](ptr, b->y_stride,

  1109.                                            s->block + 16 * n, eob);

  1110.         }

  1111.         dst_r += 4 * f->linesize[0] * step1d;

  1112.         dst   += 4 * b->y_stride * step1d;

  1113.     }

  1114. 

  1115.     // U/V

  1116.     h4    >>= 1;

  1117.     w4    >>= 1;

  1118.     end_x >>= 1;

  1119.     end_y >>= 1;

  1120.     step    = 1 << (b->uvtx * 2);

  1121.     for (p = 0; p < 2; p++) {

  1122.         dst   = b->dst[1 + p];

  1123.         dst_r = f->data[1 + p] + uv_off;

  1124.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  1125.             uint8_t *ptr = dst, *ptr_r = dst_r;

  1126.             for (x = 0; x < end_x;

  1127.                  x += uvstep1d, ptr += 4 * uvstep1d,

  1128.                  ptr_r += 4 * uvstep1d, n += step) {

  1129.                 int mode = b->uvmode;

  1130.                 LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);

  1131.                 uint8_t *a = &a_buf[16], l[32];

  1132.                 int eob    = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n])

  1133.                                               : s->uveob[p][n];

  1134. 

  1135.                 mode = check_intra_mode(s, mode, &a, ptr_r,

  1136.                                         f->linesize[1],

  1137.                                         ptr, b->uv_stride, l,

  1138.                                         col, x, w4, row, y, b->uvtx, p + 1);

  1139.                 s->dsp.intra_pred[b->uvtx][mode](ptr, b->uv_stride, l, a);

  1140.                 if (eob)

  1141.                     s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, b->uv_stride,

  1142.                                                     s->uvblock[p] + 16 * n,

  1143.                                                     eob);

  1144.             }

  1145.             dst_r += 4 * uvstep1d * f->linesize[1];

  1146.             dst   += 4 * uvstep1d * b->uv_stride;

  1147.         }

  1148.     }

  1149. }

  1150. 

  1151. static av_always_inline void mc_luma_dir(VP9Context *s, vp9_mc_func(*mc)[2],

  1152.                                          uint8_t *dst, ptrdiff_t dst_stride,

  1153.                                          const uint8_t *ref,

  1154.                                          ptrdiff_t ref_stride,

  1155.                                          ThreadFrame *ref_frame,

  1156.                                          ptrdiff_t y, ptrdiff_t x,

  1157.                                          const VP56mv *mv,

  1158.                                          int bw, int bh, int w, int h)

  1159. {

  1160.     int mx = mv->x, my = mv->y;

  1161.     int th;

  1162. 

  1163.     y   += my >> 3;

  1164.     x   += mx >> 3;

  1165.     ref += y * ref_stride + x;

  1166.     mx  &= 7;

  1167.     my  &= 7;

  1168. 

  1169.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  1170.     // the longest loopfilter of the next sbrow

  1171.     th = (y + bh + 4 * !!my + 7) >> 6;

  1172.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  1173. 

  1174.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  1175.     if (x < !!mx * 3 || y < !!my * 3 ||

  1176.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  1177.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1178.                                  ref - !!my * 3 * ref_stride - !!mx * 3,

  1179.                                  80,

  1180.                                  ref_stride,

  1181.                                  bw + !!mx * 7, bh + !!my * 7,

  1182.                                  x - !!mx * 3, y - !!my * 3, w, h);

  1183.         ref        = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  1184.         ref_stride = 80;

  1185.     }

  1186.     mc[!!mx][!!my](dst, ref, dst_stride, ref_stride, bh, mx << 1, my << 1);

  1187. }

  1188. 

  1189. static av_always_inline void mc_chroma_dir(VP9Context *s, vp9_mc_func(*mc)[2],

  1190.                                            uint8_t *dst_u, uint8_t *dst_v,

  1191.                                            ptrdiff_t dst_stride,

  1192.                                            const uint8_t *ref_u,

  1193.                                            ptrdiff_t src_stride_u,

  1194.                                            const uint8_t *ref_v,

  1195.                                            ptrdiff_t src_stride_v,

  1196.                                            ThreadFrame *ref_frame,

  1197.                                            ptrdiff_t y, ptrdiff_t x,

  1198.                                            const VP56mv *mv,

  1199.                                            int bw, int bh, int w, int h)

  1200. {

  1201.     int mx = mv->x, my = mv->y;

  1202.     int th;

  1203. 

  1204.     y     += my >> 4;

  1205.     x     += mx >> 4;

  1206.     ref_u += y * src_stride_u + x;

  1207.     ref_v += y * src_stride_v + x;

  1208.     mx    &= 15;

  1209.     my    &= 15;

  1210. 

  1211.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  1212.     // the longest loopfilter of the next sbrow

  1213.     th = (y + bh + 4 * !!my + 7) >> 5;

  1214.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  1215. 

  1216.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  1217.     if (x < !!mx * 3 || y < !!my * 3 ||

  1218.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  1219.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1220.                                  ref_u - !!my * 3 * src_stride_u - !!mx * 3,

  1221.                                  80,

  1222.                                  src_stride_u,

  1223.                                  bw + !!mx * 7, bh + !!my * 7,

  1224.                                  x - !!mx * 3, y - !!my * 3, w, h);

  1225.         ref_u = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  1226.         mc[!!mx][!!my](dst_u, ref_u, dst_stride, 80, bh, mx, my);

  1227. 

  1228.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1229.                                  ref_v - !!my * 3 * src_stride_v - !!mx * 3,

  1230.                                  80,

  1231.                                  src_stride_v,

  1232.                                  bw + !!mx * 7, bh + !!my * 7,

  1233.                                  x - !!mx * 3, y - !!my * 3, w, h);

  1234.         ref_v = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  1235.         mc[!!mx][!!my](dst_v, ref_v, dst_stride, 80, bh, mx, my);

  1236.     } else {

  1237.         mc[!!mx][!!my](dst_u, ref_u, dst_stride, src_stride_u, bh, mx, my);

  1238.         mc[!!mx][!!my](dst_v, ref_v, dst_stride, src_stride_v, bh, mx, my);

  1239.     }

  1240. }

  1241. 

  1242. static int inter_recon(AVCodecContext *avctx)

  1243. {

  1244.     static const uint8_t bwlog_tab[2][N_BS_SIZES] = {

  1245.         { 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4 },

  1246.         { 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 4, 4 },

  1247.     };

  1248.     VP9Context *s = avctx->priv_data;

  1249.     VP9Block *b = s->b;

  1250.     int row = b->row, col = b->col;

  1251. 

  1252.     ThreadFrame *tref1 = &s->refs[s->refidx[b->ref[0]]];

  1253.     ThreadFrame *tref2 = b->comp ? &s->refs[s->refidx[b->ref[1]]] : NULL;

  1254.     AVFrame      *ref1 = tref1->f;

  1255.     AVFrame      *ref2 = tref2 ? tref2->f : NULL;

  1256. 

  1257.     int w = avctx->width, h = avctx->height;

  1258.     ptrdiff_t ls_y = b->y_stride, ls_uv = b->uv_stride;

  1259. 

  1260.     if (!ref1->data[0] || (b->comp && !ref2->data[0]))

  1261.         return AVERROR_INVALIDDATA;

  1262. 

  1263.     // y inter pred

  1264.     if (b->bs > BS_8x8) {

  1265.         if (b->bs == BS_8x4) {

  1266.             mc_luma_dir(s, s->dsp.mc[3][b->filter][0], b->dst[0], ls_y,

  1267.                         ref1->data[0], ref1->linesize[0], tref1,

  1268.                         row << 3, col << 3, &b->mv[0][0], 8, 4, w, h);

  1269.             mc_luma_dir(s, s->dsp.mc[3][b->filter][0],

  1270.                         b->dst[0] + 4 * ls_y, ls_y,

  1271.                         ref1->data[0], ref1->linesize[0], tref1,

  1272.                         (row << 3) + 4, col << 3, &b->mv[2][0], 8, 4, w, h);

  1273. 

  1274.             if (b->comp) {

  1275.                 mc_luma_dir(s, s->dsp.mc[3][b->filter][1], b->dst[0], ls_y,

  1276.                             ref2->data[0], ref2->linesize[0], tref2,

  1277.                             row << 3, col << 3, &b->mv[0][1], 8, 4, w, h);

  1278.                 mc_luma_dir(s, s->dsp.mc[3][b->filter][1],

  1279.                             b->dst[0] + 4 * ls_y, ls_y,

  1280.                             ref2->data[0], ref2->linesize[0], tref2,

  1281.                             (row << 3) + 4, col << 3, &b->mv[2][1], 8, 4, w, h);

  1282.             }

  1283.         } else if (b->bs == BS_4x8) {

  1284.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], b->dst[0], ls_y,

  1285.                         ref1->data[0], ref1->linesize[0], tref1,

  1286.                         row << 3, col << 3, &b->mv[0][0], 4, 8, w, h);

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

  1288.                         ref1->data[0], ref1->linesize[0], tref1,

  1289.                         row << 3, (col << 3) + 4, &b->mv[1][0], 4, 8, w, h);

  1290. 

  1291.             if (b->comp) {

  1292.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], b->dst[0], ls_y,

  1293.                             ref2->data[0], ref2->linesize[0], tref2,

  1294.                             row << 3, col << 3, &b->mv[0][1], 4, 8, w, h);

  1295.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], b->dst[0] + 4, ls_y,

  1296.                             ref2->data[0], ref2->linesize[0], tref2,

  1297.                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 8, w, h);

  1298.             }

  1299.         } else {

  1300.             av_assert2(b->bs == BS_4x4);

  1301. 

  1302.             // FIXME if two horizontally adjacent blocks have the same MV,

  1303.             // do a w8 instead of a w4 call

  1304.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], b->dst[0], ls_y,

  1305.                         ref1->data[0], ref1->linesize[0], tref1,

  1306.                         row << 3, col << 3, &b->mv[0][0], 4, 4, w, h);

  1307.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], b->dst[0] + 4, ls_y,

  1308.                         ref1->data[0], ref1->linesize[0], tref1,

  1309.                         row << 3, (col << 3) + 4, &b->mv[1][0], 4, 4, w, h);

  1310.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0],

  1311.                         b->dst[0] + 4 * ls_y, ls_y,

  1312.                         ref1->data[0], ref1->linesize[0], tref1,

  1313.                         (row << 3) + 4, col << 3, &b->mv[2][0], 4, 4, w, h);

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

  1315.                         b->dst[0] + 4 * ls_y + 4, ls_y,

  1316.                         ref1->data[0], ref1->linesize[0], tref1,

  1317.                         (row << 3) + 4, (col << 3) + 4, &b->mv[3][0], 4, 4, w, h);

  1318. 

  1319.             if (b->comp) {

  1320.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], b->dst[0], ls_y,

  1321.                             ref2->data[0], ref2->linesize[0], tref2,

  1322.                             row << 3, col << 3, &b->mv[0][1], 4, 4, w, h);

  1323.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], b->dst[0] + 4, ls_y,

  1324.                             ref2->data[0], ref2->linesize[0], tref2,

  1325.                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 4, w, h);

  1326.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1],

  1327.                             b->dst[0] + 4 * ls_y, ls_y,

  1328.                             ref2->data[0], ref2->linesize[0], tref2,

  1329.                             (row << 3) + 4, col << 3, &b->mv[2][1], 4, 4, w, h);

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

  1331.                             b->dst[0] + 4 * ls_y + 4, ls_y,

  1332.                             ref2->data[0], ref2->linesize[0], tref2,

  1333.                             (row << 3) + 4, (col << 3) + 4, &b->mv[3][1], 4, 4, w, h);

  1334.             }

  1335.         }

  1336.     } else {

  1337.         int bwl = bwlog_tab[0][b->bs];

  1338.         int bw  = bwh_tab[0][b->bs][0] * 4;

  1339.         int bh  = bwh_tab[0][b->bs][1] * 4;

  1340. 

  1341.         mc_luma_dir(s, s->dsp.mc[bwl][b->filter][0], b->dst[0], ls_y,

  1342.                     ref1->data[0], ref1->linesize[0], tref1,

  1343.                     row << 3, col << 3, &b->mv[0][0], bw, bh, w, h);

  1344. 

  1345.         if (b->comp)

  1346.             mc_luma_dir(s, s->dsp.mc[bwl][b->filter][1], b->dst[0], ls_y,

  1347.                         ref2->data[0], ref2->linesize[0], tref2,

  1348.                         row << 3, col << 3, &b->mv[0][1], bw, bh, w, h);

  1349.     }

  1350. 

  1351.     // uv inter pred

  1352.     {

  1353.         int bwl = bwlog_tab[1][b->bs];

  1354.         int bw  = bwh_tab[1][b->bs][0] * 4, bh = bwh_tab[1][b->bs][1] * 4;

  1355.         VP56mv mvuv;

  1356. 

  1357.         w = (w + 1) >> 1;

  1358.         h = (h + 1) >> 1;

  1359.         if (b->bs > BS_8x8) {

  1360.             mvuv.x = ROUNDED_DIV(b->mv[0][0].x + b->mv[1][0].x +

  1361.                                  b->mv[2][0].x + b->mv[3][0].x, 4);

  1362.             mvuv.y = ROUNDED_DIV(b->mv[0][0].y + b->mv[1][0].y +

  1363.                                  b->mv[2][0].y + b->mv[3][0].y, 4);

  1364.         } else {

  1365.             mvuv = b->mv[0][0];

  1366.         }

  1367. 

  1368.         mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][0],

  1369.                       b->dst[1], b->dst[2], ls_uv,

  1370.                       ref1->data[1], ref1->linesize[1],

  1371.                       ref1->data[2], ref1->linesize[2], tref1,

  1372.                       row << 2, col << 2, &mvuv, bw, bh, w, h);

  1373. 

  1374.         if (b->comp) {

  1375.             if (b->bs > BS_8x8) {

  1376.                 mvuv.x = ROUNDED_DIV(b->mv[0][1].x + b->mv[1][1].x +

  1377.                                      b->mv[2][1].x + b->mv[3][1].x, 4);

  1378.                 mvuv.y = ROUNDED_DIV(b->mv[0][1].y + b->mv[1][1].y +

  1379.                                      b->mv[2][1].y + b->mv[3][1].y, 4);

  1380.             } else {

  1381.                 mvuv = b->mv[0][1];

  1382.             }

  1383.             mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][1],

  1384.                           b->dst[1], b->dst[2], ls_uv,

  1385.                           ref2->data[1], ref2->linesize[1],

  1386.                           ref2->data[2], ref2->linesize[2], tref2,

  1387.                           row << 2, col << 2, &mvuv, bw, bh, w, h);

  1388.         }

  1389.     }

  1390. 

  1391.     if (!b->skip) {

  1392.         /* mostly copied intra_reconn() */

  1393. 

  1394.         int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  1395.         int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  1396.         int end_x = FFMIN(2 * (s->cols - col), w4);

  1397.         int end_y = FFMIN(2 * (s->rows - row), h4);

  1398.         int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  1399.         int uvstep1d = 1 << b->uvtx, p;

  1400.         uint8_t *dst = b->dst[0];

  1401. 

  1402.         // y itxfm add

  1403.         for (n = 0, y = 0; y < end_y; y += step1d) {

  1404.             uint8_t *ptr = dst;

  1405.             for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d, n += step) {

  1406.                 int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  1407. 

  1408.                 if (eob)

  1409.                     s->dsp.itxfm_add[tx][DCT_DCT](ptr, b->y_stride,

  1410.                                                   s->block + 16 * n, eob);

  1411.             }

  1412.             dst += 4 * b->y_stride * step1d;

  1413.         }

  1414. 

  1415.         // uv itxfm add

  1416.         h4    >>= 1;

  1417.         w4    >>= 1;

  1418.         end_x >>= 1;

  1419.         end_y >>= 1;

  1420.         step    = 1 << (b->uvtx * 2);

  1421.         for (p = 0; p < 2; p++) {

  1422.             dst = b->dst[p + 1];

  1423.             for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  1424.                 uint8_t *ptr = dst;

  1425.                 for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d, n += step) {

  1426.                     int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n])

  1427.                                                : s->uveob[p][n];

  1428.                     if (eob)

  1429.                         s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, b->uv_stride,

  1430.                                                         s->uvblock[p] + 16 * n, eob);

  1431.                 }

  1432.                 dst += 4 * uvstep1d * b->uv_stride;

  1433.             }

  1434.         }

  1435.     }

  1436.     return 0;

  1437. }

  1438. 

  1439. static av_always_inline void mask_edges(VP9Filter *lflvl, int is_uv,

  1440.                                         int row_and_7, int col_and_7,

  1441.                                         int w, int h, int col_end, int row_end,

  1442.                                         enum TxfmMode tx, int skip_inter)

  1443. {

  1444.     // FIXME I'm pretty sure all loops can be replaced by a single LUT if

  1445.     // we make VP9Filter.mask uint64_t (i.e. row/col all single variable)

  1446.     // and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then

  1447.     // use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)

  1448. 

  1449.     // the intended behaviour of the vp9 loopfilter is to work on 8-pixel

  1450.     // edges. This means that for UV, we work on two subsampled blocks at

  1451.     // a time, and we only use the topleft block's mode information to set

  1452.     // things like block strength. Thus, for any block size smaller than

  1453.     // 16x16, ignore the odd portion of the block.

  1454.     if (tx == TX_4X4 && is_uv) {

  1455.         if (h == 1) {

  1456.             if (row_and_7 & 1)

  1457.                 return;

  1458.             if (!row_end)

  1459.                 h += 1;

  1460.         }

  1461.         if (w == 1) {

  1462.             if (col_and_7 & 1)

  1463.                 return;

  1464.             if (!col_end)

  1465.                 w += 1;

  1466.         }

  1467.     }

  1468. 

  1469.     if (tx == TX_4X4 && !skip_inter) {

  1470.         int t = 1 << col_and_7, m_col = (t << w) - t, y;

  1471.         int m_col_odd = (t << (w - 1)) - t;

  1472. 

  1473.         // on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide

  1474.         if (is_uv) {

  1475.             int m_row_8 = m_col & 0x01, m_row_4 = m_col - m_row_8;

  1476. 

  1477.             for (y = row_and_7; y < h + row_and_7; y++) {

  1478.                 int col_mask_id = 2 - !(y & 7);

  1479. 

  1480.                 lflvl->mask[is_uv][0][y][1] |= m_row_8;

  1481.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  1482.                 // for odd lines, if the odd col is not being filtered,

  1483.                 // skip odd row also:

  1484.                 // .---. <-- a

  1485.                 // |   |

  1486.                 // |___| <-- b

  1487.                 // ^   ^

  1488.                 // c   d

  1489.                 //

  1490.                 // if a/c are even row/col and b/d are odd, and d is skipped,

  1491.                 // e.g. right edge of size-66x66.webm, then skip b also (bug)

  1492.                 if ((col_end & 1) && (y & 1)) {

  1493.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col_odd;

  1494.                 } else {

  1495.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col;

  1496.                 }

  1497.             }

  1498.         } else {

  1499.             int m_row_8 = m_col & 0x11, m_row_4 = m_col - m_row_8;

  1500. 

  1501.             for (y = row_and_7; y < h + row_and_7; y++) {

  1502.                 int col_mask_id = 2 - !(y & 3);

  1503. 

  1504.                 lflvl->mask[is_uv][0][y][1]           |= m_row_8; // row edge

  1505.                 lflvl->mask[is_uv][0][y][2]           |= m_row_4;

  1506.                 lflvl->mask[is_uv][1][y][col_mask_id] |= m_col; // col edge

  1507.                 lflvl->mask[is_uv][0][y][3]           |= m_col;

  1508.                 lflvl->mask[is_uv][1][y][3]           |= m_col;

  1509.             }

  1510.         }

  1511.     } else {

  1512.         int y, t = 1 << col_and_7, m_col = (t << w) - t;

  1513. 

  1514.         if (!skip_inter) {

  1515.             int mask_id = (tx == TX_8X8);

  1516.             int l2 = tx + is_uv - 1, step1d = 1 << l2;

  1517.             static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };

  1518.             int m_row = m_col & masks[l2];

  1519. 

  1520.             // at odd UV col/row edges tx16/tx32 loopfilter edges, force

  1521.             // 8wd loopfilter to prevent going off the visible edge.

  1522.             if (is_uv && tx > TX_8X8 && (w ^ (w - 1)) == 1) {

  1523.                 int m_row_16 = ((t << (w - 1)) - t) & masks[l2];

  1524.                 int m_row_8  = m_row - m_row_16;

  1525. 

  1526.                 for (y = row_and_7; y < h + row_and_7; y++) {

  1527.                     lflvl->mask[is_uv][0][y][0] |= m_row_16;

  1528.                     lflvl->mask[is_uv][0][y][1] |= m_row_8;

  1529.                 }

  1530.             } else {

  1531.                 for (y = row_and_7; y < h + row_and_7; y++)

  1532.                     lflvl->mask[is_uv][0][y][mask_id] |= m_row;

  1533.             }

  1534. 

  1535.             if (is_uv && tx > TX_8X8 && (h ^ (h - 1)) == 1) {

  1536.                 for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)

  1537.                     lflvl->mask[is_uv][1][y][0] |= m_col;

  1538.                 if (y - row_and_7 == h - 1)

  1539.                     lflvl->mask[is_uv][1][y][1] |= m_col;

  1540.             } else {

  1541.                 for (y = row_and_7; y < h + row_and_7; y += step1d)

  1542.                     lflvl->mask[is_uv][1][y][mask_id] |= m_col;

  1543.             }

  1544.         } else if (tx != TX_4X4) {

  1545.             int mask_id;

  1546. 

  1547.             mask_id = (tx == TX_8X8) || (is_uv && h == 1);

  1548.             lflvl->mask[is_uv][1][row_and_7][mask_id] |= m_col;

  1549.             mask_id = (tx == TX_8X8) || (is_uv && w == 1);

  1550.             for (y = row_and_7; y < h + row_and_7; y++)

  1551.                 lflvl->mask[is_uv][0][y][mask_id] |= t;

  1552.         } else if (is_uv) {

  1553.             int t8 = t & 0x01, t4 = t - t8;

  1554. 

  1555.             for (y = row_and_7; y < h + row_and_7; y++) {

  1556.                 lflvl->mask[is_uv][0][y][2] |= t4;

  1557.                 lflvl->mask[is_uv][0][y][1] |= t8;

  1558.             }

  1559.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 7)] |= m_col;

  1560.         } else {

  1561.             int t8 = t & 0x11, t4 = t - t8;

  1562. 

  1563.             for (y = row_and_7; y < h + row_and_7; y++) {

  1564.                 lflvl->mask[is_uv][0][y][2] |= t4;

  1565.                 lflvl->mask[is_uv][0][y][1] |= t8;

  1566.             }

  1567.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 3)] |= m_col;

  1568.         }

  1569.     }

  1570. }

  1571. 

  1572. int ff_vp9_decode_block(AVCodecContext *avctx, int row, int col,

  1573.                         VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,

  1574.                         enum BlockLevel bl, enum BlockPartition bp)

  1575. {

  1576.     VP9Context *s = avctx->priv_data;

  1577.     VP9Block *b = s->b;

  1578.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  1579.     enum BlockSize bs = bl * 3 + bp;

  1580.     int ret, y, w4 = bwh_tab[1][bs][0], h4 = bwh_tab[1][bs][1], lvl;

  1581.     int emu[2];

  1582. 

  1583.     b->row  = row;

  1584.     b->row7 = row & 7;

  1585.     b->col  = col;

  1586.     b->col7 = col & 7;

  1587. 

  1588.     s->min_mv.x = -(128 + col * 64);

  1589.     s->min_mv.y = -(128 + row * 64);

  1590.     s->max_mv.x = 128 + (s->cols - col - w4) * 64;

  1591.     s->max_mv.y = 128 + (s->rows - row - h4) * 64;

  1592. 

  1593.     if (s->pass < 2) {

  1594.         b->bs = bs;

  1595.         b->bl = bl;

  1596.         b->bp = bp;

  1597.         decode_mode(s, b);

  1598.         b->uvtx = b->tx - (w4 * 2 == (1 << b->tx) || h4 * 2 == (1 << b->tx));

  1599. 

  1600.         if (!b->skip) {

  1601.             if ((ret = decode_coeffs(avctx)) < 0)

  1602.                 return ret;

  1603.         } else {

  1604.             int pl;

  1605. 

  1606.             memset(&s->above_y_nnz_ctx[col * 2], 0, w4 * 2);

  1607.             memset(&s->left_y_nnz_ctx[(row & 7) << 1], 0, h4 * 2);

  1608.             for (pl = 0; pl < 2; pl++) {

  1609.                 memset(&s->above_uv_nnz_ctx[pl][col], 0, w4);

  1610.                 memset(&s->left_uv_nnz_ctx[pl][row & 7], 0, h4);

  1611.             }

  1612.         }

  1613. 

  1614.         if (s->pass == 1) {

  1615.             s->b++;

  1616.             s->block      += w4 * h4 * 64;

  1617.             s->uvblock[0] += w4 * h4 * 16;

  1618.             s->uvblock[1] += w4 * h4 * 16;

  1619.             s->eob        += w4 * h4 * 4;

  1620.             s->uveob[0]   += w4 * h4;

  1621.             s->uveob[1]   += w4 * h4;

  1622. 

  1623.             return 0;

  1624.         }

  1625.     }

  1626. 

  1627.     /* Emulated overhangs if the stride of the target buffer can't hold.

  1628.      * This allows to support emu-edge and so on even if we have large

  1629.      * block overhangs. */

  1630.     emu[0] = (col + w4) * 8 > f->linesize[0] ||

  1631.              (row + h4) > s->rows;

  1632.     emu[1] = (col + w4) * 4 > f->linesize[1] ||

  1633.              (row + h4) > s->rows;

  1634.     if (emu[0]) {

  1635.         b->dst[0]   = s->tmp_y;

  1636.         b->y_stride = 64;

  1637.     } else {

  1638.         b->dst[0]   = f->data[0] + yoff;

  1639.         b->y_stride = f->linesize[0];

  1640.     }

  1641.     if (emu[1]) {

  1642.         b->dst[1]    = s->tmp_uv[0];

  1643.         b->dst[2]    = s->tmp_uv[1];

  1644.         b->uv_stride = 32;

  1645.     } else {

  1646.         b->dst[1]    = f->data[1] + uvoff;

  1647.         b->dst[2]    = f->data[2] + uvoff;

  1648.         b->uv_stride = f->linesize[1];

  1649.     }

  1650.     if (b->intra) {

  1651.         intra_recon(avctx, yoff, uvoff);

  1652.     } else {

  1653.         if ((ret = inter_recon(avctx)) < 0)

  1654.             return ret;

  1655.     }

  1656.     if (emu[0]) {

  1657.         int w = FFMIN(s->cols - col, w4) * 8;

  1658.         int h = FFMIN(s->rows - row, h4) * 8;

  1659.         int n, o = 0;

  1660. 

  1661.         for (n = 0; o < w; n++) {

  1662.             int bw = 64 >> n;

  1663. 

  1664.             av_assert2(n <= 4);

  1665.             if (w & bw) {

  1666.                 s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o,

  1667.                                          s->tmp_y + o,

  1668.                                          f->linesize[0],

  1669.                                          64, h, 0, 0);

  1670.                 o += bw;

  1671.             }

  1672.         }

  1673.     }

  1674.     if (emu[1]) {

  1675.         int w = FFMIN(s->cols - col, w4) * 4;

  1676.         int h = FFMIN(s->rows - row, h4) * 4;

  1677.         int n, o = 0;

  1678. 

  1679.         for (n = 1; o < w; n++) {

  1680.             int bw = 64 >> n;

  1681. 

  1682.             av_assert2(n <= 4);

  1683.             if (w & bw) {

  1684.                 s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o,

  1685.                                          s->tmp_uv[0] + o,

  1686.                                          f->linesize[1],

  1687.                                          32, h, 0, 0);

  1688.                 s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o,

  1689.                                          s->tmp_uv[1] + o,

  1690.                                          f->linesize[2],

  1691.                                          32, h, 0, 0);

  1692.                 o += bw;

  1693.             }

  1694.         }

  1695.     }

  1696. 

  1697.     // pick filter level and find edges to apply filter to

  1698.     if (s->filter.level &&

  1699.         (lvl = s->segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]

  1700.                                                     [b->mode[3] != ZEROMV]) > 0) {

  1701.         int x_end = FFMIN(s->cols - col, w4);

  1702.         int y_end = FFMIN(s->rows - row, h4);

  1703.         int skip_inter = !b->intra && b->skip;

  1704. 

  1705.         for (y = 0; y < h4; y++)

  1706.             memset(&lflvl->level[((row & 7) + y) * 8 + (col & 7)], lvl, w4);

  1707.         mask_edges(lflvl, 0, row & 7, col & 7, x_end, y_end, 0, 0, b->tx, skip_inter);

  1708.         mask_edges(lflvl, 1, row & 7, col & 7, x_end, y_end,

  1709.                    s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,

  1710.                    s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,

  1711.                    b->uvtx, skip_inter);

  1712. 

  1713.         if (!s->filter.lim_lut[lvl]) {

  1714.             int sharp = s->filter.sharpness;

  1715.             int limit = lvl;

  1716. 

  1717.             if (sharp > 0) {

  1718.                 limit >>= (sharp + 3) >> 2;

  1719.                 limit   = FFMIN(limit, 9 - sharp);

  1720.             }

  1721.             limit = FFMAX(limit, 1);

  1722. 

  1723.             s->filter.lim_lut[lvl]   = limit;

  1724.             s->filter.mblim_lut[lvl] = 2 * (lvl + 2) + limit;

  1725.         }

  1726.     }

  1727. 

  1728.     if (s->pass == 2) {

  1729.         s->b++;

  1730.         s->block      += w4 * h4 * 64;

  1731.         s->uvblock[0] += w4 * h4 * 16;

  1732.         s->uvblock[1] += w4 * h4 * 16;

  1733.         s->eob        += w4 * h4 * 4;

  1734.         s->uveob[0]   += w4 * h4;

  1735.         s->uveob[1]   += w4 * h4;

  1736.     }

  1737. 

  1738.     return 0;

  1739. }