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.
84693 Commits
32 Branches
303MB
Tree: 37814a21cb
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '37814a21cb'
${ noResults }
FFmpeg/libavcodec/vp9block.c

2086 lines
86KB
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 FFmpeg.

  8.  *

  9.  * FFmpeg 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.  * FFmpeg 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 FFmpeg; 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. static av_always_inline void setctx_2d(uint8_t *ptr, int w, int h,

  44.                                        ptrdiff_t stride, int v)

  45. {

  46.     switch (w) {

  47.     case 1:

  48.         do {

  49.             *ptr = v;

  50.             ptr += stride;

  51.         } while (--h);

  52.         break;

  53.     case 2: {

  54.         int v16 = v * 0x0101;

  55.         do {

  56.             AV_WN16A(ptr, v16);

  57.             ptr += stride;

  58.         } while (--h);

  59.         break;

  60.     }

  61.     case 4: {

  62.         uint32_t v32 = v * 0x01010101;

  63.         do {

  64.             AV_WN32A(ptr, v32);

  65.             ptr += stride;

  66.         } while (--h);

  67.         break;

  68.     }

  69.     case 8: {

  70. #if HAVE_FAST_64BIT

  71.         uint64_t v64 = v * 0x0101010101010101ULL;

  72.         do {

  73.             AV_WN64A(ptr, v64);

  74.             ptr += stride;

  75.         } while (--h);

  76. #else

  77.         uint32_t v32 = v * 0x01010101;

  78.         do {

  79.             AV_WN32A(ptr,     v32);

  80.             AV_WN32A(ptr + 4, v32);

  81.             ptr += stride;

  82.         } while (--h);

  83. #endif

  84.         break;

  85.     }

  86.     }

  87. }

  88. 

  89. static void decode_mode(AVCodecContext *avctx)

  90. {

  91.     static const uint8_t left_ctx[N_BS_SIZES] = {

  92.         0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf

  93.     };

  94.     static const uint8_t above_ctx[N_BS_SIZES] = {

  95.         0x0, 0x0, 0x8, 0x8, 0x8, 0xc, 0xc, 0xc, 0xe, 0xe, 0xe, 0xf, 0xf

  96.     };

  97.     static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {

  98.         TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,

  99.         TX_16X16, TX_8X8,   TX_8X8,   TX_8X8,   TX_4X4,   TX_4X4,  TX_4X4

  100.     };

  101.     VP9Context *s = avctx->priv_data;

  102.     VP9Block *b = s->b;

  103.     int row = s->row, col = s->col, row7 = s->row7;

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

  105.     int bw4 = bwh_tab[1][b->bs][0], w4 = FFMIN(s->cols - col, bw4);

  106.     int bh4 = bwh_tab[1][b->bs][1], h4 = FFMIN(s->rows - row, bh4), y;

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

  108.     int vref, filter_id;

  109. 

  110.     if (!s->s.h.segmentation.enabled) {

  111.         b->seg_id = 0;

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

  113.         b->seg_id = !s->s.h.segmentation.update_map ? 0 :

  114.                     vp8_rac_get_tree(&s->c, ff_vp9_segmentation_tree, s->s.h.segmentation.prob);

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

  116.                (s->s.h.segmentation.temporal &&

  117.                 vp56_rac_get_prob_branchy(&s->c,

  118.                     s->s.h.segmentation.pred_prob[s->above_segpred_ctx[col] +

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

  120.         if (!s->s.h.errorres && s->s.frames[REF_FRAME_SEGMAP].segmentation_map) {

  121.             int pred = 8, x;

  122.             uint8_t *refsegmap = s->s.frames[REF_FRAME_SEGMAP].segmentation_map;

  123. 

  124.             if (!s->s.frames[REF_FRAME_SEGMAP].uses_2pass)

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

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

  127.                 int idx_base = (y + row) * 8 * s->sb_cols + col;

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

  129.                     pred = FFMIN(pred, refsegmap[idx_base + x]);

  130.             }

  131.             av_assert1(pred < 8);

  132.             b->seg_id = pred;

  133.         } else {

  134.             b->seg_id = 0;

  135.         }

  136. 

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

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

  139.     } else {

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

  141.                                      s->s.h.segmentation.prob);

  142. 

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

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

  145.     }

  146.     if (s->s.h.segmentation.enabled &&

  147.         (s->s.h.segmentation.update_map || s->s.h.keyframe || s->s.h.intraonly)) {

  148.         setctx_2d(&s->s.frames[CUR_FRAME].segmentation_map[row * 8 * s->sb_cols + col],

  149.                   bw4, bh4, 8 * s->sb_cols, b->seg_id);

  150.     }

  151. 

  152.     b->skip = s->s.h.segmentation.enabled &&

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

  154.     if (!b->skip) {

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

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

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

  158.     }

  159. 

  160.     if (s->s.h.keyframe || s->s.h.intraonly) {

  161.         b->intra = 1;

  162.     } else if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[b->seg_id].ref_enabled) {

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

  164.     } else {

  165.         int c, bit;

  166. 

  167.         if (have_a && have_l) {

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

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

  170.         } else {

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

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

  173.         }

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

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

  176.         b->intra = !bit;

  177.     }

  178. 

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

  180.         int c;

  181.         if (have_a) {

  182.             if (have_l) {

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

  184.                      s->above_txfm_ctx[col]) +

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

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

  187.             } else {

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

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

  190.             }

  191.         } else if (have_l) {

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

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

  194.         } else {

  195.             c = 1;

  196.         }

  197.         switch (max_tx) {

  198.         case TX_32X32:

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

  200.             if (b->tx) {

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

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

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

  204.             }

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

  206.             break;

  207.         case TX_16X16:

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

  209.             if (b->tx)

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

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

  212.             break;

  213.         case TX_8X8:

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

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

  216.             break;

  217.         case TX_4X4:

  218.             b->tx = TX_4X4;

  219.             break;

  220.         }

  221.     } else {

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

  223.     }

  224. 

  225.     if (s->s.h.keyframe || s->s.h.intraonly) {

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

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

  228. 

  229.         b->comp = 0;

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

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

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

  233.             // simpler for now

  234.             b->mode[0] =

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

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

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

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

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

  240.                 l[0]       =

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

  242.             } else {

  243.                 l[0]       =

  244.                 a[1]       =

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

  246.             }

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

  248.                 b->mode[2] =

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

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

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

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

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

  254.                     l[1]       =

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

  256.                 } else {

  257.                     l[1]       =

  258.                     a[1]       =

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

  260.                 }

  261.             } else {

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

  263.                 l[1]       =

  264.                 a[1]       =

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

  266.             }

  267.         } else {

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

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

  270.             b->mode[3] =

  271.             b->mode[2] =

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

  273.             // FIXME this can probably be optimized

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

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

  276.         }

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

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

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

  280.         b->comp = 0;

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

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

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

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

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

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

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

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

  289.             } else {

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

  291.             }

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

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

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

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

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

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

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

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

  300.                 } else {

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

  302.                 }

  303.             } else {

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

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

  306.             }

  307.         } else {

  308.             static const uint8_t size_group[10] = {

  309.                 3, 3, 3, 3, 2, 2, 2, 1, 1, 1

  310.             };

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

  312. 

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

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

  315.             b->mode[1] =

  316.             b->mode[2] =

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

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

  319.         }

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

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

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

  323.     } else {

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

  325.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  326.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  327.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  328.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  329.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  330.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  331.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  332.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  333.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  334.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  335.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  336.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  337.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },

  338.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },

  339.         };

  340. 

  341.         if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[b->seg_id].ref_enabled) {

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

  343.             b->comp = 0;

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

  345.         } else {

  346.             // read comp_pred flag

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

  348.                 b->comp = s->s.h.comppredmode == PRED_COMPREF;

  349.             } else {

  350.                 int c;

  351. 

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

  353.                 if (have_a) {

  354.                     if (have_l) {

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

  356.                             c = 4;

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

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

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

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

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

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

  363.                         } else {

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

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

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

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

  368.                         }

  369.                     } else {

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

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

  372.                     }

  373.                 } else if (have_l) {

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

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

  376.                 } else {

  377.                     c = 1;

  378.                 }

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

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

  381.             }

  382. 

  383.             // read actual references

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

  385.             // memory accesses below

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

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

  388. 

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

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

  391.                 if (have_a) {

  392.                     if (have_l) {

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

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

  395.                                 c = 2;

  396.                             } else {

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

  398.                             }

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

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

  401.                         } else {

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

  403. 

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

  405.                                 c = 0;

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

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

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

  409.                                     c = 4;

  410.                                 } else {

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

  412.                                 }

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

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

  415.                                     c = 1;

  416.                                 } else {

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

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

  419.                                 }

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

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

  422.                                     c = 1;

  423.                                 } else {

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

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

  426.                                 }

  427.                             } else {

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

  429.                             }

  430.                         }

  431.                     } else {

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

  433.                             c = 2;

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

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

  436.                         } else {

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

  438.                         }

  439.                     }

  440.                 } else if (have_l) {

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

  442.                         c = 2;

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

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

  445.                     } else {

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

  447.                     }

  448.                 } else {

  449.                     c = 2;

  450.                 }

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

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

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

  454.             } else /* single reference */ {

  455.                 int bit, c;

  456. 

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

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

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

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

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

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

  463.                             } else {

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

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

  466.                             }

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

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

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

  470.                         } else {

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

  472.                         }

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

  474.                         c = 2;

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

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

  477.                     } else {

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

  479.                     }

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

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

  482.                         c = 2;

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

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

  485.                     } else {

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

  487.                     }

  488.                 } else {

  489.                     c = 2;

  490.                 }

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

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

  493.                 if (!bit) {

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

  495.                 } else {

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

  497.                     if (have_a) {

  498.                         if (have_l) {

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

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

  501.                                     c = 2;

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

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

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

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

  506.                                     c = 3;

  507.                                 } else {

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

  509.                                 }

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

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

  512.                                     c = 2;

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

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

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

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

  517.                                     c = 3;

  518.                                 } else {

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

  520.                                 }

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

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

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

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

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

  526.                                     } else {

  527.                                         c = 2;

  528.                                     }

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

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

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

  532.                                 } else {

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

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

  535.                                 }

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

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

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

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

  540.                                 } else {

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

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

  543.                                 }

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

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

  546.                                     c = 3;

  547.                                 } else {

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

  549.                                 }

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

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

  552.                             } else {

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

  554.                                     2 * (s->above_ref_ctx[col] == 1);

  555.                             }

  556.                         } else {

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

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

  559.                                 c = 2;

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

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

  562.                             } else {

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

  564.                             }

  565.                         }

  566.                     } else if (have_l) {

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

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

  569.                             c = 2;

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

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

  572.                         } else {

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

  574.                         }

  575.                     } else {

  576.                         c = 2;

  577.                     }

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

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

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

  581.                 }

  582.             }

  583.         }

  584. 

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

  586.             if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[b->seg_id].skip_enabled) {

  587.                 b->mode[0] =

  588.                 b->mode[1] =

  589.                 b->mode[2] =

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

  591.             } else {

  592.                 static const uint8_t off[10] = {

  593.                     3, 0, 0, 1, 0, 0, 0, 0, 0, 0

  594.                 };

  595. 

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

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

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

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

  600. 

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

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

  603.                 b->mode[1] =

  604.                 b->mode[2] =

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

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

  607.             }

  608.         }

  609. 

  610.         if (s->s.h.filtermode == FILTER_SWITCHABLE) {

  611.             int c;

  612. 

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

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

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

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

  617.                 } else {

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

  619.                 }

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

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

  622.             } else {

  623.                 c = 3;

  624.             }

  625. 

  626.             filter_id = vp8_rac_get_tree(&s->c, ff_vp9_filter_tree,

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

  628.             s->counts.filter[c][filter_id]++;

  629.             b->filter = ff_vp9_filter_lut[filter_id];

  630.         } else {

  631.             b->filter = s->s.h.filtermode;

  632.         }

  633. 

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

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

  636. 

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

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

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

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

  641. 

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

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

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

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

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

  647.             } else {

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

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

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

  651.             }

  652. 

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

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

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

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

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

  658. 

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

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

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

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

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

  664.                 } else {

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

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

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

  668.                 }

  669.             } else {

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

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

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

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

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

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

  676.             }

  677.         } else {

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

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

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

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

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

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

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

  685.         }

  686. 

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

  688.     }

  689. 

  690. #if HAVE_FAST_64BIT

  691. #define SPLAT_CTX(var, val, n) \

  692.     switch (n) { \

  693.     case 1:  var = val;                                    break; \

  694.     case 2:  AV_WN16A(&var, val *             0x0101);     break; \

  695.     case 4:  AV_WN32A(&var, val *         0x01010101);     break; \

  696.     case 8:  AV_WN64A(&var, val * 0x0101010101010101ULL);  break; \

  697.     case 16: { \

  698.         uint64_t v64 = val * 0x0101010101010101ULL; \

  699.         AV_WN64A(              &var,     v64); \

  700.         AV_WN64A(&((uint8_t *) &var)[8], v64); \

  701.         break; \

  702.     } \

  703.     }

  704. #else

  705. #define SPLAT_CTX(var, val, n) \

  706.     switch (n) { \

  707.     case 1:  var = val;                         break; \

  708.     case 2:  AV_WN16A(&var, val *     0x0101);  break; \

  709.     case 4:  AV_WN32A(&var, val * 0x01010101);  break; \

  710.     case 8: { \

  711.         uint32_t v32 = val * 0x01010101; \

  712.         AV_WN32A(              &var,     v32); \

  713.         AV_WN32A(&((uint8_t *) &var)[4], v32); \

  714.         break; \

  715.     } \

  716.     case 16: { \

  717.         uint32_t v32 = val * 0x01010101; \

  718.         AV_WN32A(              &var,      v32); \

  719.         AV_WN32A(&((uint8_t *) &var)[4],  v32); \

  720.         AV_WN32A(&((uint8_t *) &var)[8],  v32); \

  721.         AV_WN32A(&((uint8_t *) &var)[12], v32); \

  722.         break; \

  723.     } \

  724.     }

  725. #endif

  726. 

  727.     switch (bwh_tab[1][b->bs][0]) {

  728. #define SET_CTXS(dir, off, n) \

  729.     do { \

  730.         SPLAT_CTX(s->dir##_skip_ctx[off],      b->skip,          n); \

  731.         SPLAT_CTX(s->dir##_txfm_ctx[off],      b->tx,            n); \

  732.         SPLAT_CTX(s->dir##_partition_ctx[off], dir##_ctx[b->bs], n); \

  733.         if (!s->s.h.keyframe && !s->s.h.intraonly) { \

  734.             SPLAT_CTX(s->dir##_intra_ctx[off], b->intra,   n); \

  735.             SPLAT_CTX(s->dir##_comp_ctx[off],  b->comp,    n); \

  736.             SPLAT_CTX(s->dir##_mode_ctx[off],  b->mode[3], n); \

  737.             if (!b->intra) { \

  738.                 SPLAT_CTX(s->dir##_ref_ctx[off], vref, n); \

  739.                 if (s->s.h.filtermode == FILTER_SWITCHABLE) { \

  740.                     SPLAT_CTX(s->dir##_filter_ctx[off], filter_id, n); \

  741.                 } \

  742.             } \

  743.         } \

  744.     } while (0)

  745.     case 1: SET_CTXS(above, col, 1); break;

  746.     case 2: SET_CTXS(above, col, 2); break;

  747.     case 4: SET_CTXS(above, col, 4); break;

  748.     case 8: SET_CTXS(above, col, 8); break;

  749.     }

  750.     switch (bwh_tab[1][b->bs][1]) {

  751.     case 1: SET_CTXS(left, row7, 1); break;

  752.     case 2: SET_CTXS(left, row7, 2); break;

  753.     case 4: SET_CTXS(left, row7, 4); break;

  754.     case 8: SET_CTXS(left, row7, 8); break;

  755.     }

  756. #undef SPLAT_CTX

  757. #undef SET_CTXS

  758. 

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

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

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

  762. 

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

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

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

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

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

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

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

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

  771.         } else {

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

  773. 

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

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

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

  777.             }

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

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

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

  781.             }

  782.         }

  783.     }

  784. 

  785.     // FIXME kinda ugly

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

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

  788.         VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[o];

  789. 

  790.         if (b->intra) {

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

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

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

  794.             }

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

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

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

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

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

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

  801.             }

  802.         } else {

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

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

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

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

  807.             }

  808.         }

  809.     }

  810. }

  811. 

  812. // FIXME merge cnt/eob arguments?

  813. static av_always_inline int

  814. decode_coeffs_b_generic(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  815.                         int is_tx32x32, int is8bitsperpixel, int bpp, unsigned (*cnt)[6][3],

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

  817.                         int nnz, const int16_t *scan, const int16_t (*nb)[2],

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

  819. {

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

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

  822.     uint8_t cache[1024];

  823. 

  824.     do {

  825.         int val, rc;

  826. 

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

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

  829.         if (!val)

  830.             break;

  831. 

  832. skip_eob:

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

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

  835.             if (!--band_left)

  836.                 band_left = band_counts[++band];

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

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

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

  840.             if (++i == n_coeffs)

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

  842.             goto skip_eob;

  843.         }

  844. 

  845.         rc = scan[i];

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

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

  848.             val       = 1;

  849.             cache[rc] = 1;

  850.         } else {

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

  852.             if (!tp[3])

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

  854. 

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

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

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

  858.                     cache[rc] = val = 2;

  859.                 } else {

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

  861.                     cache[rc] = 3;

  862.                 }

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

  864.                 cache[rc] = 4;

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

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

  867.                 } else {

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

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

  870.                 }

  871.             } else { // cat 3-6

  872.                 cache[rc] = 5;

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

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

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

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

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

  878.                     } else {

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

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

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

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

  883.                     }

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

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

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

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

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

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

  890.                 } else {

  891.                     val = 67;

  892.                     if (!is8bitsperpixel) {

  893.                         if (bpp == 12) {

  894.                             val += vp56_rac_get_prob(c, 255) << 17;

  895.                             val += vp56_rac_get_prob(c, 255) << 16;

  896.                         }

  897.                         val +=  (vp56_rac_get_prob(c, 255) << 15);

  898.                         val +=  (vp56_rac_get_prob(c, 255) << 14);

  899.                     }

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  914.                 }

  915.             }

  916.         }

  917. #define STORE_COEF(c, i, v) do { \

  918.     if (is8bitsperpixel) { \

  919.         c[i] = v; \

  920.     } else { \

  921.         AV_WN32A(&c[i * 2], v); \

  922.     } \

  923. } while (0)

  924.         if (!--band_left)

  925.             band_left = band_counts[++band];

  926.         if (is_tx32x32)

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

  928.         else

  929.             STORE_COEF(coef, rc, (vp8_rac_get(c) ? -val : val) * qmul[!!i]);

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

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

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

  933. 

  934.     return i;

  935. }

  936. 

  937. static int decode_coeffs_b_8bpp(VP9Context *s, int16_t *coef, int n_coeffs,

  938.                                 unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  939.                                 uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  940.                                 const int16_t (*nb)[2], const int16_t *band_counts,

  941.                                 const int16_t *qmul)

  942. {

  943.     return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 0, 1, 8, cnt, eob, p,

  944.                                    nnz, scan, nb, band_counts, qmul);

  945. }

  946. 

  947. static int decode_coeffs_b32_8bpp(VP9Context *s, int16_t *coef, int n_coeffs,

  948.                                   unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  949.                                   uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  950.                                   const int16_t (*nb)[2], const int16_t *band_counts,

  951.                                   const int16_t *qmul)

  952. {

  953.     return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 1, 1, 8, cnt, eob, p,

  954.                                    nnz, scan, nb, band_counts, qmul);

  955. }

  956. 

  957. static int decode_coeffs_b_16bpp(VP9Context *s, int16_t *coef, int n_coeffs,

  958.                                  unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  959.                                  uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  960.                                  const int16_t (*nb)[2], const int16_t *band_counts,

  961.                                  const int16_t *qmul)

  962. {

  963.     return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 0, 0, s->s.h.bpp, cnt, eob, p,

  964.                                    nnz, scan, nb, band_counts, qmul);

  965. }

  966. 

  967. static int decode_coeffs_b32_16bpp(VP9Context *s, int16_t *coef, int n_coeffs,

  968.                                    unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  969.                                    uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  970.                                    const int16_t (*nb)[2], const int16_t *band_counts,

  971.                                    const int16_t *qmul)

  972. {

  973.     return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 1, 0, s->s.h.bpp, cnt, eob, p,

  974.                                    nnz, scan, nb, band_counts, qmul);

  975. }

  976. 

  977. static av_always_inline int decode_coeffs(AVCodecContext *avctx, int is8bitsperpixel)

  978. {

  979.     VP9Context *s = avctx->priv_data;

  980.     VP9Block *b = s->b;

  981.     int row = s->row, col = s->col;

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

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

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

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

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

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

  988.     int n, pl, x, y, ret;

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

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

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

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

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

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

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

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

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

  998.         { 1, 2, 3, 4,  3,   16 - 13 },

  999.         { 1, 2, 3, 4, 11,   64 - 21 },

  1000.         { 1, 2, 3, 4, 11,  256 - 21 },

  1001.         { 1, 2, 3, 4, 11, 1024 - 21 },

  1002.     };

  1003.     const int16_t *y_band_counts = band_counts[b->tx];

  1004.     const int16_t *uv_band_counts = band_counts[b->uvtx];

  1005.     int bytesperpixel = is8bitsperpixel ? 1 : 2;

  1006.     int total_coeff = 0;

  1007. 

  1008. #define MERGE(la, end, step, rd) \

  1009.     for (n = 0; n < end; n += step) \

  1010.         la[n] = !!rd(&la[n])

  1011. #define MERGE_CTX(step, rd) \

  1012.     do { \

  1013.         MERGE(l, end_y, step, rd); \

  1014.         MERGE(a, end_x, step, rd); \

  1015.     } while (0)

  1016. 

  1017. #define DECODE_Y_COEF_LOOP(step, mode_index, v) \

  1018.     for (n = 0, y = 0; y < end_y; y += step) { \

  1019.         for (x = 0; x < end_x; x += step, n += step * step) { \

  1020.             enum TxfmType txtp = ff_vp9_intra_txfm_type[b->mode[mode_index]]; \

  1021.             ret = (is8bitsperpixel ? decode_coeffs_b##v##_8bpp : decode_coeffs_b##v##_16bpp) \

  1022.                                     (s, s->block + 16 * n * bytesperpixel, 16 * step * step, \

  1023.                                      c, e, p, a[x] + l[y], yscans[txtp], \

  1024.                                      ynbs[txtp], y_band_counts, qmul[0]); \

  1025.             a[x] = l[y] = !!ret; \

  1026.             total_coeff |= !!ret; \

  1027.             if (step >= 4) { \

  1028.                 AV_WN16A(&s->eob[n], ret); \

  1029.             } else { \

  1030.                 s->eob[n] = ret; \

  1031.             } \

  1032.         } \

  1033.     }

  1034. 

  1035. #define SPLAT(la, end, step, cond) \

  1036.     if (step == 2) { \

  1037.         for (n = 1; n < end; n += step) \

  1038.             la[n] = la[n - 1]; \

  1039.     } else if (step == 4) { \

  1040.         if (cond) { \

  1041.             for (n = 0; n < end; n += step) \

  1042.                 AV_WN32A(&la[n], la[n] * 0x01010101); \

  1043.         } else { \

  1044.             for (n = 0; n < end; n += step) \

  1045.                 memset(&la[n + 1], la[n], FFMIN(end - n - 1, 3)); \

  1046.         } \

  1047.     } else /* step == 8 */ { \

  1048.         if (cond) { \

  1049.             if (HAVE_FAST_64BIT) { \

  1050.                 for (n = 0; n < end; n += step) \

  1051.                     AV_WN64A(&la[n], la[n] * 0x0101010101010101ULL); \

  1052.             } else { \

  1053.                 for (n = 0; n < end; n += step) { \

  1054.                     uint32_t v32 = la[n] * 0x01010101; \

  1055.                     AV_WN32A(&la[n],     v32); \

  1056.                     AV_WN32A(&la[n + 4], v32); \

  1057.                 } \

  1058.             } \

  1059.         } else { \

  1060.             for (n = 0; n < end; n += step) \

  1061.                 memset(&la[n + 1], la[n], FFMIN(end - n - 1, 7)); \

  1062.         } \

  1063.     }

  1064. #define SPLAT_CTX(step) \

  1065.     do { \

  1066.         SPLAT(a, end_x, step, end_x == w4); \

  1067.         SPLAT(l, end_y, step, end_y == h4); \

  1068.     } while (0)

  1069. 

  1070.     /* y tokens */

  1071.     switch (b->tx) {

  1072.     case TX_4X4:

  1073.         DECODE_Y_COEF_LOOP(1, b->bs > BS_8x8 ? n : 0,);

  1074.         break;

  1075.     case TX_8X8:

  1076.         MERGE_CTX(2, AV_RN16A);

  1077.         DECODE_Y_COEF_LOOP(2, 0,);

  1078.         SPLAT_CTX(2);

  1079.         break;

  1080.     case TX_16X16:

  1081.         MERGE_CTX(4, AV_RN32A);

  1082.         DECODE_Y_COEF_LOOP(4, 0,);

  1083.         SPLAT_CTX(4);

  1084.         break;

  1085.     case TX_32X32:

  1086.         MERGE_CTX(8, AV_RN64A);

  1087.         DECODE_Y_COEF_LOOP(8, 0, 32);

  1088.         SPLAT_CTX(8);

  1089.         break;

  1090.     }

  1091. 

  1092. #define DECODE_UV_COEF_LOOP(step, v) \

  1093.     for (n = 0, y = 0; y < end_y; y += step) { \

  1094.         for (x = 0; x < end_x; x += step, n += step * step) { \

  1095.             ret = (is8bitsperpixel ? decode_coeffs_b##v##_8bpp : decode_coeffs_b##v##_16bpp) \

  1096.                                     (s, s->uvblock[pl] + 16 * n * bytesperpixel, \

  1097.                                      16 * step * step, c, e, p, a[x] + l[y], \

  1098.                                      uvscan, uvnb, uv_band_counts, qmul[1]); \

  1099.             a[x] = l[y] = !!ret; \

  1100.             total_coeff |= !!ret; \

  1101.             if (step >= 4) { \

  1102.                 AV_WN16A(&s->uveob[pl][n], ret); \

  1103.             } else { \

  1104.                 s->uveob[pl][n] = ret; \

  1105.             } \

  1106.         } \

  1107.     }

  1108. 

  1109.     p = s->prob.coef[b->uvtx][1 /* uv */][!b->intra];

  1110.     c = s->counts.coef[b->uvtx][1 /* uv */][!b->intra];

  1111.     e = s->counts.eob[b->uvtx][1 /* uv */][!b->intra];

  1112.     w4 >>= s->ss_h;

  1113.     end_x >>= s->ss_h;

  1114.     h4 >>= s->ss_v;

  1115.     end_y >>= s->ss_v;

  1116.     for (pl = 0; pl < 2; pl++) {

  1117.         a = &s->above_uv_nnz_ctx[pl][col << !s->ss_h];

  1118.         l = &s->left_uv_nnz_ctx[pl][(row & 7) << !s->ss_v];

  1119.         switch (b->uvtx) {

  1120.         case TX_4X4:

  1121.             DECODE_UV_COEF_LOOP(1,);

  1122.             break;

  1123.         case TX_8X8:

  1124.             MERGE_CTX(2, AV_RN16A);

  1125.             DECODE_UV_COEF_LOOP(2,);

  1126.             SPLAT_CTX(2);

  1127.             break;

  1128.         case TX_16X16:

  1129.             MERGE_CTX(4, AV_RN32A);

  1130.             DECODE_UV_COEF_LOOP(4,);

  1131.             SPLAT_CTX(4);

  1132.             break;

  1133.         case TX_32X32:

  1134.             MERGE_CTX(8, AV_RN64A);

  1135.             DECODE_UV_COEF_LOOP(8, 32);

  1136.             SPLAT_CTX(8);

  1137.             break;

  1138.         }

  1139.     }

  1140. 

  1141.     return total_coeff;

  1142. }

  1143. 

  1144. static int decode_coeffs_8bpp(AVCodecContext *avctx)

  1145. {

  1146.     return decode_coeffs(avctx, 1);

  1147. }

  1148. 

  1149. static int decode_coeffs_16bpp(AVCodecContext *avctx)

  1150. {

  1151.     return decode_coeffs(avctx, 0);

  1152. }

  1153. 

  1154. static av_always_inline int check_intra_mode(VP9Context *s, int mode, uint8_t **a,

  1155.                                              uint8_t *dst_edge, ptrdiff_t stride_edge,

  1156.                                              uint8_t *dst_inner, ptrdiff_t stride_inner,

  1157.                                              uint8_t *l, int col, int x, int w,

  1158.                                              int row, int y, enum TxfmMode tx,

  1159.                                              int p, int ss_h, int ss_v, int bytesperpixel)

  1160. {

  1161.     int have_top = row > 0 || y > 0;

  1162.     int have_left = col > s->tile_col_start || x > 0;

  1163.     int have_right = x < w - 1;

  1164.     int bpp = s->s.h.bpp;

  1165.     static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {

  1166.         [VERT_PRED]            = { { DC_127_PRED,          VERT_PRED            },

  1167.                                    { DC_127_PRED,          VERT_PRED            } },

  1168.         [HOR_PRED]             = { { DC_129_PRED,          DC_129_PRED          },

  1169.                                    { HOR_PRED,             HOR_PRED             } },

  1170.         [DC_PRED]              = { { DC_128_PRED,          TOP_DC_PRED          },

  1171.                                    { LEFT_DC_PRED,         DC_PRED              } },

  1172.         [DIAG_DOWN_LEFT_PRED]  = { { DC_127_PRED,          DIAG_DOWN_LEFT_PRED  },

  1173.                                    { DC_127_PRED,          DIAG_DOWN_LEFT_PRED  } },

  1174.         [DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },

  1175.                                    { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },

  1176.         [VERT_RIGHT_PRED]      = { { VERT_RIGHT_PRED,      VERT_RIGHT_PRED      },

  1177.                                    { VERT_RIGHT_PRED,      VERT_RIGHT_PRED      } },

  1178.         [HOR_DOWN_PRED]        = { { HOR_DOWN_PRED,        HOR_DOWN_PRED        },

  1179.                                    { HOR_DOWN_PRED,        HOR_DOWN_PRED        } },

  1180.         [VERT_LEFT_PRED]       = { { DC_127_PRED,          VERT_LEFT_PRED       },

  1181.                                    { DC_127_PRED,          VERT_LEFT_PRED       } },

  1182.         [HOR_UP_PRED]          = { { DC_129_PRED,          DC_129_PRED          },

  1183.                                    { HOR_UP_PRED,          HOR_UP_PRED          } },

  1184.         [TM_VP8_PRED]          = { { DC_129_PRED,          VERT_PRED            },

  1185.                                    { HOR_PRED,             TM_VP8_PRED          } },

  1186.     };

  1187.     static const struct {

  1188.         uint8_t needs_left:1;

  1189.         uint8_t needs_top:1;

  1190.         uint8_t needs_topleft:1;

  1191.         uint8_t needs_topright:1;

  1192.         uint8_t invert_left:1;

  1193.     } edges[N_INTRA_PRED_MODES] = {

  1194.         [VERT_PRED]            = { .needs_top  = 1 },

  1195.         [HOR_PRED]             = { .needs_left = 1 },

  1196.         [DC_PRED]              = { .needs_top  = 1, .needs_left = 1 },

  1197.         [DIAG_DOWN_LEFT_PRED]  = { .needs_top  = 1, .needs_topright = 1 },

  1198.         [DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1,

  1199.                                    .needs_topleft = 1 },

  1200.         [VERT_RIGHT_PRED]      = { .needs_left = 1, .needs_top = 1,

  1201.                                    .needs_topleft = 1 },

  1202.         [HOR_DOWN_PRED]        = { .needs_left = 1, .needs_top = 1,

  1203.                                    .needs_topleft = 1 },

  1204.         [VERT_LEFT_PRED]       = { .needs_top  = 1, .needs_topright = 1 },

  1205.         [HOR_UP_PRED]          = { .needs_left = 1, .invert_left = 1 },

  1206.         [TM_VP8_PRED]          = { .needs_left = 1, .needs_top = 1,

  1207.                                    .needs_topleft = 1 },

  1208.         [LEFT_DC_PRED]         = { .needs_left = 1 },

  1209.         [TOP_DC_PRED]          = { .needs_top  = 1 },

  1210.         [DC_128_PRED]          = { 0 },

  1211.         [DC_127_PRED]          = { 0 },

  1212.         [DC_129_PRED]          = { 0 }

  1213.     };

  1214. 

  1215.     av_assert2(mode >= 0 && mode < 10);

  1216.     mode = mode_conv[mode][have_left][have_top];

  1217.     if (edges[mode].needs_top) {

  1218.         uint8_t *top, *topleft;

  1219.         int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !ss_h) - x) * 4;

  1220.         int n_px_need_tr = 0;

  1221. 

  1222.         if (tx == TX_4X4 && edges[mode].needs_topright && have_right)

  1223.             n_px_need_tr = 4;

  1224. 

  1225.         // if top of sb64-row, use s->intra_pred_data[] instead of

  1226.         // dst[-stride] for intra prediction (it contains pre- instead of

  1227.         // post-loopfilter data)

  1228.         if (have_top) {

  1229.             top = !(row & 7) && !y ?

  1230.                 s->intra_pred_data[p] + (col * (8 >> ss_h) + x * 4) * bytesperpixel :

  1231.                 y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];

  1232.             if (have_left)

  1233.                 topleft = !(row & 7) && !y ?

  1234.                     s->intra_pred_data[p] + (col * (8 >> ss_h) + x * 4) * bytesperpixel :

  1235.                     y == 0 || x == 0 ? &dst_edge[-stride_edge] :

  1236.                     &dst_inner[-stride_inner];

  1237.         }

  1238. 

  1239.         if (have_top &&

  1240.             (!edges[mode].needs_topleft || (have_left && top == topleft)) &&

  1241.             (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&

  1242.             n_px_need + n_px_need_tr <= n_px_have) {

  1243.             *a = top;

  1244.         } else {

  1245.             if (have_top) {

  1246.                 if (n_px_need <= n_px_have) {

  1247.                     memcpy(*a, top, n_px_need * bytesperpixel);

  1248.                 } else {

  1249. #define memset_bpp(c, i1, v, i2, num) do { \

  1250.     if (bytesperpixel == 1) { \

  1251.         memset(&(c)[(i1)], (v)[(i2)], (num)); \

  1252.     } else { \

  1253.         int n, val = AV_RN16A(&(v)[(i2) * 2]); \

  1254.         for (n = 0; n < (num); n++) { \

  1255.             AV_WN16A(&(c)[((i1) + n) * 2], val); \

  1256.         } \

  1257.     } \

  1258. } while (0)

  1259.                     memcpy(*a, top, n_px_have * bytesperpixel);

  1260.                     memset_bpp(*a, n_px_have, (*a), n_px_have - 1, n_px_need - n_px_have);

  1261.                 }

  1262.             } else {

  1263. #define memset_val(c, val, num) do { \

  1264.     if (bytesperpixel == 1) { \

  1265.         memset((c), (val), (num)); \

  1266.     } else { \

  1267.         int n; \

  1268.         for (n = 0; n < (num); n++) { \

  1269.             AV_WN16A(&(c)[n * 2], (val)); \

  1270.         } \

  1271.     } \

  1272. } while (0)

  1273.                 memset_val(*a, (128 << (bpp - 8)) - 1, n_px_need);

  1274.             }

  1275.             if (edges[mode].needs_topleft) {

  1276.                 if (have_left && have_top) {

  1277. #define assign_bpp(c, i1, v, i2) do { \

  1278.     if (bytesperpixel == 1) { \

  1279.         (c)[(i1)] = (v)[(i2)]; \

  1280.     } else { \

  1281.         AV_COPY16(&(c)[(i1) * 2], &(v)[(i2) * 2]); \

  1282.     } \

  1283. } while (0)

  1284.                     assign_bpp(*a, -1, topleft, -1);

  1285.                 } else {

  1286. #define assign_val(c, i, v) do { \

  1287.     if (bytesperpixel == 1) { \

  1288.         (c)[(i)] = (v); \

  1289.     } else { \

  1290.         AV_WN16A(&(c)[(i) * 2], (v)); \

  1291.     } \

  1292. } while (0)

  1293.                     assign_val((*a), -1, (128 << (bpp - 8)) + (have_top ? +1 : -1));

  1294.                 }

  1295.             }

  1296.             if (tx == TX_4X4 && edges[mode].needs_topright) {

  1297.                 if (have_top && have_right &&

  1298.                     n_px_need + n_px_need_tr <= n_px_have) {

  1299.                     memcpy(&(*a)[4 * bytesperpixel], &top[4 * bytesperpixel], 4 * bytesperpixel);

  1300.                 } else {

  1301.                     memset_bpp(*a, 4, *a, 3, 4);

  1302.                 }

  1303.             }

  1304.         }

  1305.     }

  1306.     if (edges[mode].needs_left) {

  1307.         if (have_left) {

  1308.             int n_px_need = 4 << tx, i, n_px_have = (((s->rows - row) << !ss_v) - y) * 4;

  1309.             uint8_t *dst = x == 0 ? dst_edge : dst_inner;

  1310.             ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;

  1311. 

  1312.             if (edges[mode].invert_left) {

  1313.                 if (n_px_need <= n_px_have) {

  1314.                     for (i = 0; i < n_px_need; i++)

  1315.                         assign_bpp(l, i, &dst[i * stride], -1);

  1316.                 } else {

  1317.                     for (i = 0; i < n_px_have; i++)

  1318.                         assign_bpp(l, i, &dst[i * stride], -1);

  1319.                     memset_bpp(l, n_px_have, l, n_px_have - 1, n_px_need - n_px_have);

  1320.                 }

  1321.             } else {

  1322.                 if (n_px_need <= n_px_have) {

  1323.                     for (i = 0; i < n_px_need; i++)

  1324.                         assign_bpp(l, n_px_need - 1 - i, &dst[i * stride], -1);

  1325.                 } else {

  1326.                     for (i = 0; i < n_px_have; i++)

  1327.                         assign_bpp(l, n_px_need - 1 - i, &dst[i * stride], -1);

  1328.                     memset_bpp(l, 0, l, n_px_need - n_px_have, n_px_need - n_px_have);

  1329.                 }

  1330.             }

  1331.         } else {

  1332.             memset_val(l, (128 << (bpp - 8)) + 1, 4 << tx);

  1333.         }

  1334.     }

  1335. 

  1336.     return mode;

  1337. }

  1338. 

  1339. static av_always_inline void intra_recon(AVCodecContext *avctx, ptrdiff_t y_off,

  1340.                                          ptrdiff_t uv_off, int bytesperpixel)

  1341. {

  1342.     VP9Context *s = avctx->priv_data;

  1343.     VP9Block *b = s->b;

  1344.     int row = s->row, col = s->col;

  1345.     int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  1346.     int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  1347.     int end_x = FFMIN(2 * (s->cols - col), w4);

  1348.     int end_y = FFMIN(2 * (s->rows - row), h4);

  1349.     int tx = 4 * s->s.h.lossless + b->tx, uvtx = b->uvtx + 4 * s->s.h.lossless;

  1350.     int uvstep1d = 1 << b->uvtx, p;

  1351.     uint8_t *dst = s->dst[0], *dst_r = s->s.frames[CUR_FRAME].tf.f->data[0] + y_off;

  1352.     LOCAL_ALIGNED_32(uint8_t, a_buf, [96]);

  1353.     LOCAL_ALIGNED_32(uint8_t, l, [64]);

  1354. 

  1355.     for (n = 0, y = 0; y < end_y; y += step1d) {

  1356.         uint8_t *ptr = dst, *ptr_r = dst_r;

  1357.         for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d * bytesperpixel,

  1358.                                ptr_r += 4 * step1d * bytesperpixel, n += step) {

  1359.             int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?

  1360.                                y * 2 + x : 0];

  1361.             uint8_t *a = &a_buf[32];

  1362.             enum TxfmType txtp = ff_vp9_intra_txfm_type[mode];

  1363.             int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  1364. 

  1365.             mode = check_intra_mode(s, mode, &a, ptr_r,

  1366.                                     s->s.frames[CUR_FRAME].tf.f->linesize[0],

  1367.                                     ptr, s->y_stride, l,

  1368.                                     col, x, w4, row, y, b->tx, 0, 0, 0, bytesperpixel);

  1369.             s->dsp.intra_pred[b->tx][mode](ptr, s->y_stride, l, a);

  1370.             if (eob)

  1371.                 s->dsp.itxfm_add[tx][txtp](ptr, s->y_stride,

  1372.                                            s->block + 16 * n * bytesperpixel, eob);

  1373.         }

  1374.         dst_r += 4 * step1d * s->s.frames[CUR_FRAME].tf.f->linesize[0];

  1375.         dst   += 4 * step1d * s->y_stride;

  1376.     }

  1377. 

  1378.     // U/V

  1379.     w4    >>= s->ss_h;

  1380.     end_x >>= s->ss_h;

  1381.     end_y >>= s->ss_v;

  1382.     step = 1 << (b->uvtx * 2);

  1383.     for (p = 0; p < 2; p++) {

  1384.         dst   = s->dst[1 + p];

  1385.         dst_r = s->s.frames[CUR_FRAME].tf.f->data[1 + p] + uv_off;

  1386.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  1387.             uint8_t *ptr = dst, *ptr_r = dst_r;

  1388.             for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d * bytesperpixel,

  1389.                                    ptr_r += 4 * uvstep1d * bytesperpixel, n += step) {

  1390.                 int mode = b->uvmode;

  1391.                 uint8_t *a = &a_buf[32];

  1392.                 int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  1393. 

  1394.                 mode = check_intra_mode(s, mode, &a, ptr_r,

  1395.                                         s->s.frames[CUR_FRAME].tf.f->linesize[1],

  1396.                                         ptr, s->uv_stride, l, col, x, w4, row, y,

  1397.                                         b->uvtx, p + 1, s->ss_h, s->ss_v, bytesperpixel);

  1398.                 s->dsp.intra_pred[b->uvtx][mode](ptr, s->uv_stride, l, a);

  1399.                 if (eob)

  1400.                     s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  1401.                                                     s->uvblock[p] + 16 * n * bytesperpixel, eob);

  1402.             }

  1403.             dst_r += 4 * uvstep1d * s->s.frames[CUR_FRAME].tf.f->linesize[1];

  1404.             dst   += 4 * uvstep1d * s->uv_stride;

  1405.         }

  1406.     }

  1407. }

  1408. 

  1409. static void intra_recon_8bpp(AVCodecContext *avctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  1410. {

  1411.     intra_recon(avctx, y_off, uv_off, 1);

  1412. }

  1413. 

  1414. static void intra_recon_16bpp(AVCodecContext *avctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  1415. {

  1416.     intra_recon(avctx, y_off, uv_off, 2);

  1417. }

  1418. 

  1419. static av_always_inline void mc_luma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2],

  1420.                                               uint8_t *dst, ptrdiff_t dst_stride,

  1421.                                               const uint8_t *ref, ptrdiff_t ref_stride,

  1422.                                               ThreadFrame *ref_frame,

  1423.                                               ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  1424.                                               int bw, int bh, int w, int h, int bytesperpixel)

  1425. {

  1426.     int mx = mv->x, my = mv->y, th;

  1427. 

  1428.     y += my >> 3;

  1429.     x += mx >> 3;

  1430.     ref += y * ref_stride + x * bytesperpixel;

  1431.     mx &= 7;

  1432.     my &= 7;

  1433.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  1434.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  1435.     // the longest loopfilter of the next sbrow

  1436.     th = (y + bh + 4 * !!my + 7) >> 6;

  1437.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  1438.     // The arm/aarch64 _hv filters read one more row than what actually is

  1439.     // needed, so switch to emulated edge one pixel sooner vertically

  1440.     // (!!my * 5) than horizontally (!!mx * 4).

  1441.     if (x < !!mx * 3 || y < !!my * 3 ||

  1442.         x + !!mx * 4 > w - bw || y + !!my * 5 > h - bh) {

  1443.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1444.                                  ref - !!my * 3 * ref_stride - !!mx * 3 * bytesperpixel,

  1445.                                  160, ref_stride,

  1446.                                  bw + !!mx * 7, bh + !!my * 7,

  1447.                                  x - !!mx * 3, y - !!my * 3, w, h);

  1448.         ref = s->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;

  1449.         ref_stride = 160;

  1450.     }

  1451.     mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);

  1452. }

  1453. 

  1454. static av_always_inline void mc_chroma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2],

  1455.                                                 uint8_t *dst_u, uint8_t *dst_v,

  1456.                                                 ptrdiff_t dst_stride,

  1457.                                                 const uint8_t *ref_u, ptrdiff_t src_stride_u,

  1458.                                                 const uint8_t *ref_v, ptrdiff_t src_stride_v,

  1459.                                                 ThreadFrame *ref_frame,

  1460.                                                 ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  1461.                                                 int bw, int bh, int w, int h, int bytesperpixel)

  1462. {

  1463.     int mx = mv->x * (1 << !s->ss_h), my = mv->y * (1 << !s->ss_v), th;

  1464. 

  1465.     y += my >> 4;

  1466.     x += mx >> 4;

  1467.     ref_u += y * src_stride_u + x * bytesperpixel;

  1468.     ref_v += y * src_stride_v + x * bytesperpixel;

  1469.     mx &= 15;

  1470.     my &= 15;

  1471.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  1472.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  1473.     // the longest loopfilter of the next sbrow

  1474.     th = (y + bh + 4 * !!my + 7) >> (6 - s->ss_v);

  1475.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  1476.     // The arm/aarch64 _hv filters read one more row than what actually is

  1477.     // needed, so switch to emulated edge one pixel sooner vertically

  1478.     // (!!my * 5) than horizontally (!!mx * 4).

  1479.     if (x < !!mx * 3 || y < !!my * 3 ||

  1480.         x + !!mx * 4 > w - bw || y + !!my * 5 > h - bh) {

  1481.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1482.                                  ref_u - !!my * 3 * src_stride_u - !!mx * 3 * bytesperpixel,

  1483.                                  160, src_stride_u,

  1484.                                  bw + !!mx * 7, bh + !!my * 7,

  1485.                                  x - !!mx * 3, y - !!my * 3, w, h);

  1486.         ref_u = s->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;

  1487.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, 160, bh, mx, my);

  1488. 

  1489.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1490.                                  ref_v - !!my * 3 * src_stride_v - !!mx * 3 * bytesperpixel,

  1491.                                  160, src_stride_v,

  1492.                                  bw + !!mx * 7, bh + !!my * 7,

  1493.                                  x - !!mx * 3, y - !!my * 3, w, h);

  1494.         ref_v = s->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;

  1495.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, 160, bh, mx, my);

  1496.     } else {

  1497.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);

  1498.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);

  1499.     }

  1500. }

  1501. 

  1502. #define mc_luma_dir(s, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, \

  1503.                     px, py, pw, ph, bw, bh, w, h, i) \

  1504.     mc_luma_unscaled(s, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \

  1505.                      mv, bw, bh, w, h, bytesperpixel)

  1506. #define mc_chroma_dir(s, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  1507.                       row, col, mv, px, py, pw, ph, bw, bh, w, h, i) \

  1508.     mc_chroma_unscaled(s, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  1509.                        row, col, mv, bw, bh, w, h, bytesperpixel)

  1510. #define SCALED 0

  1511. #define FN(x) x##_8bpp

  1512. #define BYTES_PER_PIXEL 1

  1513. #include "vp9_mc_template.c"

  1514. #undef FN

  1515. #undef BYTES_PER_PIXEL

  1516. #define FN(x) x##_16bpp

  1517. #define BYTES_PER_PIXEL 2

  1518. #include "vp9_mc_template.c"

  1519. #undef mc_luma_dir

  1520. #undef mc_chroma_dir

  1521. #undef FN

  1522. #undef BYTES_PER_PIXEL

  1523. #undef SCALED

  1524. 

  1525. static av_always_inline void mc_luma_scaled(VP9Context *s, vp9_scaled_mc_func smc,

  1526.                                             vp9_mc_func (*mc)[2],

  1527.                                             uint8_t *dst, ptrdiff_t dst_stride,

  1528.                                             const uint8_t *ref, ptrdiff_t ref_stride,

  1529.                                             ThreadFrame *ref_frame,

  1530.                                             ptrdiff_t y, ptrdiff_t x, const VP56mv *in_mv,

  1531.                                             int px, int py, int pw, int ph,

  1532.                                             int bw, int bh, int w, int h, int bytesperpixel,

  1533.                                             const uint16_t *scale, const uint8_t *step)

  1534. {

  1535.     if (s->s.frames[CUR_FRAME].tf.f->width == ref_frame->f->width &&

  1536.         s->s.frames[CUR_FRAME].tf.f->height == ref_frame->f->height) {

  1537.         mc_luma_unscaled(s, mc, dst, dst_stride, ref, ref_stride, ref_frame,

  1538.                          y, x, in_mv, bw, bh, w, h, bytesperpixel);

  1539.     } else {

  1540. #define scale_mv(n, dim) (((int64_t)(n) * scale[dim]) >> 14)

  1541.     int mx, my;

  1542.     int refbw_m1, refbh_m1;

  1543.     int th;

  1544.     VP56mv mv;

  1545. 

  1546.     mv.x = av_clip(in_mv->x, -(x + pw - px + 4) * 8, (s->cols * 8 - x + px + 3) * 8);

  1547.     mv.y = av_clip(in_mv->y, -(y + ph - py + 4) * 8, (s->rows * 8 - y + py + 3) * 8);

  1548.     // BUG libvpx seems to scale the two components separately. This introduces

  1549.     // rounding errors but we have to reproduce them to be exactly compatible

  1550.     // with the output from libvpx...

  1551.     mx = scale_mv(mv.x * 2, 0) + scale_mv(x * 16, 0);

  1552.     my = scale_mv(mv.y * 2, 1) + scale_mv(y * 16, 1);

  1553. 

  1554.     y = my >> 4;

  1555.     x = mx >> 4;

  1556.     ref += y * ref_stride + x * bytesperpixel;

  1557.     mx &= 15;

  1558.     my &= 15;

  1559.     refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;

  1560.     refbh_m1 = ((bh - 1) * step[1] + my) >> 4;

  1561.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  1562.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  1563.     // the longest loopfilter of the next sbrow

  1564.     th = (y + refbh_m1 + 4 + 7) >> 6;

  1565.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  1566.     // The arm/aarch64 _hv filters read one more row than what actually is

  1567.     // needed, so switch to emulated edge one pixel sooner vertically

  1568.     // (y + 5 >= h - refbh_m1) than horizontally (x + 4 >= w - refbw_m1).

  1569.     if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 5 >= h - refbh_m1) {

  1570.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1571.                                  ref - 3 * ref_stride - 3 * bytesperpixel,

  1572.                                  288, ref_stride,

  1573.                                  refbw_m1 + 8, refbh_m1 + 8,

  1574.                                  x - 3, y - 3, w, h);

  1575.         ref = s->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;

  1576.         ref_stride = 288;

  1577.     }

  1578.     smc(dst, dst_stride, ref, ref_stride, bh, mx, my, step[0], step[1]);

  1579.     }

  1580. }

  1581. 

  1582. static av_always_inline void mc_chroma_scaled(VP9Context *s, vp9_scaled_mc_func smc,

  1583.                                               vp9_mc_func (*mc)[2],

  1584.                                               uint8_t *dst_u, uint8_t *dst_v,

  1585.                                               ptrdiff_t dst_stride,

  1586.                                               const uint8_t *ref_u, ptrdiff_t src_stride_u,

  1587.                                               const uint8_t *ref_v, ptrdiff_t src_stride_v,

  1588.                                               ThreadFrame *ref_frame,

  1589.                                               ptrdiff_t y, ptrdiff_t x, const VP56mv *in_mv,

  1590.                                               int px, int py, int pw, int ph,

  1591.                                               int bw, int bh, int w, int h, int bytesperpixel,

  1592.                                               const uint16_t *scale, const uint8_t *step)

  1593. {

  1594.     if (s->s.frames[CUR_FRAME].tf.f->width == ref_frame->f->width &&

  1595.         s->s.frames[CUR_FRAME].tf.f->height == ref_frame->f->height) {

  1596.         mc_chroma_unscaled(s, mc, dst_u, dst_v, dst_stride, ref_u, src_stride_u,

  1597.                            ref_v, src_stride_v, ref_frame,

  1598.                            y, x, in_mv, bw, bh, w, h, bytesperpixel);

  1599.     } else {

  1600.     int mx, my;

  1601.     int refbw_m1, refbh_m1;

  1602.     int th;

  1603.     VP56mv mv;

  1604. 

  1605.     if (s->ss_h) {

  1606.         // BUG https://code.google.com/p/webm/issues/detail?id=820

  1607.         mv.x = av_clip(in_mv->x, -(x + pw - px + 4) * 16, (s->cols * 4 - x + px + 3) * 16);

  1608.         mx = scale_mv(mv.x, 0) + (scale_mv(x * 16, 0) & ~15) + (scale_mv(x * 32, 0) & 15);

  1609.     } else {

  1610.         mv.x = av_clip(in_mv->x, -(x + pw - px + 4) * 8, (s->cols * 8 - x + px + 3) * 8);

  1611.         mx = scale_mv(mv.x * 2, 0) + scale_mv(x * 16, 0);

  1612.     }

  1613.     if (s->ss_v) {

  1614.         // BUG https://code.google.com/p/webm/issues/detail?id=820

  1615.         mv.y = av_clip(in_mv->y, -(y + ph - py + 4) * 16, (s->rows * 4 - y + py + 3) * 16);

  1616.         my = scale_mv(mv.y, 1) + (scale_mv(y * 16, 1) & ~15) + (scale_mv(y * 32, 1) & 15);

  1617.     } else {

  1618.         mv.y = av_clip(in_mv->y, -(y + ph - py + 4) * 8, (s->rows * 8 - y + py + 3) * 8);

  1619.         my = scale_mv(mv.y * 2, 1) + scale_mv(y * 16, 1);

  1620.     }

  1621. #undef scale_mv

  1622.     y = my >> 4;

  1623.     x = mx >> 4;

  1624.     ref_u += y * src_stride_u + x * bytesperpixel;

  1625.     ref_v += y * src_stride_v + x * bytesperpixel;

  1626.     mx &= 15;

  1627.     my &= 15;

  1628.     refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;

  1629.     refbh_m1 = ((bh - 1) * step[1] + my) >> 4;

  1630.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  1631.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  1632.     // the longest loopfilter of the next sbrow

  1633.     th = (y + refbh_m1 + 4 + 7) >> (6 - s->ss_v);

  1634.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  1635.     // The arm/aarch64 _hv filters read one more row than what actually is

  1636.     // needed, so switch to emulated edge one pixel sooner vertically

  1637.     // (y + 5 >= h - refbh_m1) than horizontally (x + 4 >= w - refbw_m1).

  1638.     if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 5 >= h - refbh_m1) {

  1639.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1640.                                  ref_u - 3 * src_stride_u - 3 * bytesperpixel,

  1641.                                  288, src_stride_u,

  1642.                                  refbw_m1 + 8, refbh_m1 + 8,

  1643.                                  x - 3, y - 3, w, h);

  1644.         ref_u = s->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;

  1645.         smc(dst_u, dst_stride, ref_u, 288, bh, mx, my, step[0], step[1]);

  1646. 

  1647.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1648.                                  ref_v - 3 * src_stride_v - 3 * bytesperpixel,

  1649.                                  288, src_stride_v,

  1650.                                  refbw_m1 + 8, refbh_m1 + 8,

  1651.                                  x - 3, y - 3, w, h);

  1652.         ref_v = s->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;

  1653.         smc(dst_v, dst_stride, ref_v, 288, bh, mx, my, step[0], step[1]);

  1654.     } else {

  1655.         smc(dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my, step[0], step[1]);

  1656.         smc(dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my, step[0], step[1]);

  1657.     }

  1658.     }

  1659. }

  1660. 

  1661. #define mc_luma_dir(s, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, \

  1662.                     px, py, pw, ph, bw, bh, w, h, i) \

  1663.     mc_luma_scaled(s, s->dsp.s##mc, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \

  1664.                    mv, px, py, pw, ph, bw, bh, w, h, bytesperpixel, \

  1665.                    s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])

  1666. #define mc_chroma_dir(s, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  1667.                       row, col, mv, px, py, pw, ph, bw, bh, w, h, i) \

  1668.     mc_chroma_scaled(s, s->dsp.s##mc, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  1669.                      row, col, mv, px, py, pw, ph, bw, bh, w, h, bytesperpixel, \

  1670.                      s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])

  1671. #define SCALED 1

  1672. #define FN(x) x##_scaled_8bpp

  1673. #define BYTES_PER_PIXEL 1

  1674. #include "vp9_mc_template.c"

  1675. #undef FN

  1676. #undef BYTES_PER_PIXEL

  1677. #define FN(x) x##_scaled_16bpp

  1678. #define BYTES_PER_PIXEL 2

  1679. #include "vp9_mc_template.c"

  1680. #undef mc_luma_dir

  1681. #undef mc_chroma_dir

  1682. #undef FN

  1683. #undef BYTES_PER_PIXEL

  1684. #undef SCALED

  1685. 

  1686. static av_always_inline void inter_recon(AVCodecContext *avctx, int bytesperpixel)

  1687. {

  1688.     VP9Context *s = avctx->priv_data;

  1689.     VP9Block *b = s->b;

  1690.     int row = s->row, col = s->col;

  1691. 

  1692.     if (s->mvscale[b->ref[0]][0] || (b->comp && s->mvscale[b->ref[1]][0])) {

  1693.         if (bytesperpixel == 1) {

  1694.             inter_pred_scaled_8bpp(avctx);

  1695.         } else {

  1696.             inter_pred_scaled_16bpp(avctx);

  1697.         }

  1698.     } else {

  1699.         if (bytesperpixel == 1) {

  1700.             inter_pred_8bpp(avctx);

  1701.         } else {

  1702.             inter_pred_16bpp(avctx);

  1703.         }

  1704.     }

  1705. 

  1706.     if (!b->skip) {

  1707.         /* mostly copied intra_recon() */

  1708. 

  1709.         int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  1710.         int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  1711.         int end_x = FFMIN(2 * (s->cols - col), w4);

  1712.         int end_y = FFMIN(2 * (s->rows - row), h4);

  1713.         int tx = 4 * s->s.h.lossless + b->tx, uvtx = b->uvtx + 4 * s->s.h.lossless;

  1714.         int uvstep1d = 1 << b->uvtx, p;

  1715.         uint8_t *dst = s->dst[0];

  1716. 

  1717.         // y itxfm add

  1718.         for (n = 0, y = 0; y < end_y; y += step1d) {

  1719.             uint8_t *ptr = dst;

  1720.             for (x = 0; x < end_x; x += step1d,

  1721.                  ptr += 4 * step1d * bytesperpixel, n += step) {

  1722.                 int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  1723. 

  1724.                 if (eob)

  1725.                     s->dsp.itxfm_add[tx][DCT_DCT](ptr, s->y_stride,

  1726.                                                   s->block + 16 * n * bytesperpixel, eob);

  1727.             }

  1728.             dst += 4 * s->y_stride * step1d;

  1729.         }

  1730. 

  1731.         // uv itxfm add

  1732.         end_x >>= s->ss_h;

  1733.         end_y >>= s->ss_v;

  1734.         step = 1 << (b->uvtx * 2);

  1735.         for (p = 0; p < 2; p++) {

  1736.             dst = s->dst[p + 1];

  1737.             for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  1738.                 uint8_t *ptr = dst;

  1739.                 for (x = 0; x < end_x; x += uvstep1d,

  1740.                      ptr += 4 * uvstep1d * bytesperpixel, n += step) {

  1741.                     int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  1742. 

  1743.                     if (eob)

  1744.                         s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  1745.                                                         s->uvblock[p] + 16 * n * bytesperpixel, eob);

  1746.                 }

  1747.                 dst += 4 * uvstep1d * s->uv_stride;

  1748.             }

  1749.         }

  1750.     }

  1751. }

  1752. 

  1753. static void inter_recon_8bpp(AVCodecContext *avctx)

  1754. {

  1755.     inter_recon(avctx, 1);

  1756. }

  1757. 

  1758. static void inter_recon_16bpp(AVCodecContext *avctx)

  1759. {

  1760.     inter_recon(avctx, 2);

  1761. }

  1762. 

  1763. static av_always_inline void mask_edges(uint8_t (*mask)[8][4], int ss_h, int ss_v,

  1764.                                         int row_and_7, int col_and_7,

  1765.                                         int w, int h, int col_end, int row_end,

  1766.                                         enum TxfmMode tx, int skip_inter)

  1767. {

  1768.     static const unsigned wide_filter_col_mask[2] = { 0x11, 0x01 };

  1769.     static const unsigned wide_filter_row_mask[2] = { 0x03, 0x07 };

  1770. 

  1771.     // FIXME I'm pretty sure all loops can be replaced by a single LUT if

  1772.     // we make VP9Filter.mask uint64_t (i.e. row/col all single variable)

  1773.     // and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then

  1774.     // use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)

  1775. 

  1776.     // the intended behaviour of the vp9 loopfilter is to work on 8-pixel

  1777.     // edges. This means that for UV, we work on two subsampled blocks at

  1778.     // a time, and we only use the topleft block's mode information to set

  1779.     // things like block strength. Thus, for any block size smaller than

  1780.     // 16x16, ignore the odd portion of the block.

  1781.     if (tx == TX_4X4 && (ss_v | ss_h)) {

  1782.         if (h == ss_v) {

  1783.             if (row_and_7 & 1)

  1784.                 return;

  1785.             if (!row_end)

  1786.                 h += 1;

  1787.         }

  1788.         if (w == ss_h) {

  1789.             if (col_and_7 & 1)

  1790.                 return;

  1791.             if (!col_end)

  1792.                 w += 1;

  1793.         }

  1794.     }

  1795. 

  1796.     if (tx == TX_4X4 && !skip_inter) {

  1797.         int t = 1 << col_and_7, m_col = (t << w) - t, y;

  1798.         // on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide

  1799.         int m_row_8 = m_col & wide_filter_col_mask[ss_h], m_row_4 = m_col - m_row_8;

  1800. 

  1801.         for (y = row_and_7; y < h + row_and_7; y++) {

  1802.             int col_mask_id = 2 - !(y & wide_filter_row_mask[ss_v]);

  1803. 

  1804.             mask[0][y][1] |= m_row_8;

  1805.             mask[0][y][2] |= m_row_4;

  1806.             // for odd lines, if the odd col is not being filtered,

  1807.             // skip odd row also:

  1808.             // .---. <-- a

  1809.             // |   |

  1810.             // |___| <-- b

  1811.             // ^   ^

  1812.             // c   d

  1813.             //

  1814.             // if a/c are even row/col and b/d are odd, and d is skipped,

  1815.             // e.g. right edge of size-66x66.webm, then skip b also (bug)

  1816.             if ((ss_h & ss_v) && (col_end & 1) && (y & 1)) {

  1817.                 mask[1][y][col_mask_id] |= (t << (w - 1)) - t;

  1818.             } else {

  1819.                 mask[1][y][col_mask_id] |= m_col;

  1820.             }

  1821.             if (!ss_h)

  1822.                 mask[0][y][3] |= m_col;

  1823.             if (!ss_v) {

  1824.                 if (ss_h && (col_end & 1))

  1825.                     mask[1][y][3] |= (t << (w - 1)) - t;

  1826.                 else

  1827.                     mask[1][y][3] |= m_col;

  1828.             }

  1829.         }

  1830.     } else {

  1831.         int y, t = 1 << col_and_7, m_col = (t << w) - t;

  1832. 

  1833.         if (!skip_inter) {

  1834.             int mask_id = (tx == TX_8X8);

  1835.             int l2 = tx + ss_h - 1, step1d;

  1836.             static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };

  1837.             int m_row = m_col & masks[l2];

  1838. 

  1839.             // at odd UV col/row edges tx16/tx32 loopfilter edges, force

  1840.             // 8wd loopfilter to prevent going off the visible edge.

  1841.             if (ss_h && tx > TX_8X8 && (w ^ (w - 1)) == 1) {

  1842.                 int m_row_16 = ((t << (w - 1)) - t) & masks[l2];

  1843.                 int m_row_8 = m_row - m_row_16;

  1844. 

  1845.                 for (y = row_and_7; y < h + row_and_7; y++) {

  1846.                     mask[0][y][0] |= m_row_16;

  1847.                     mask[0][y][1] |= m_row_8;

  1848.                 }

  1849.             } else {

  1850.                 for (y = row_and_7; y < h + row_and_7; y++)

  1851.                     mask[0][y][mask_id] |= m_row;

  1852.             }

  1853. 

  1854.             l2 = tx + ss_v - 1;

  1855.             step1d = 1 << l2;

  1856.             if (ss_v && tx > TX_8X8 && (h ^ (h - 1)) == 1) {

  1857.                 for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)

  1858.                     mask[1][y][0] |= m_col;

  1859.                 if (y - row_and_7 == h - 1)

  1860.                     mask[1][y][1] |= m_col;

  1861.             } else {

  1862.                 for (y = row_and_7; y < h + row_and_7; y += step1d)

  1863.                     mask[1][y][mask_id] |= m_col;

  1864.             }

  1865.         } else if (tx != TX_4X4) {

  1866.             int mask_id;

  1867. 

  1868.             mask_id = (tx == TX_8X8) || (h == ss_v);

  1869.             mask[1][row_and_7][mask_id] |= m_col;

  1870.             mask_id = (tx == TX_8X8) || (w == ss_h);

  1871.             for (y = row_and_7; y < h + row_and_7; y++)

  1872.                 mask[0][y][mask_id] |= t;

  1873.         } else {

  1874.             int t8 = t & wide_filter_col_mask[ss_h], t4 = t - t8;

  1875. 

  1876.             for (y = row_and_7; y < h + row_and_7; y++) {

  1877.                 mask[0][y][2] |= t4;

  1878.                 mask[0][y][1] |= t8;

  1879.             }

  1880.             mask[1][row_and_7][2 - !(row_and_7 & wide_filter_row_mask[ss_v])] |= m_col;

  1881.         }

  1882.     }

  1883. }

  1884. 

  1885. void ff_vp9_decode_block(AVCodecContext *avctx, int row, int col,

  1886.                          VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,

  1887.                          enum BlockLevel bl, enum BlockPartition bp)

  1888. {

  1889.     VP9Context *s = avctx->priv_data;

  1890.     VP9Block *b = s->b;

  1891.     enum BlockSize bs = bl * 3 + bp;

  1892.     int bytesperpixel = s->bytesperpixel;

  1893.     int w4 = bwh_tab[1][bs][0], h4 = bwh_tab[1][bs][1], lvl;

  1894.     int emu[2];

  1895.     AVFrame *f = s->s.frames[CUR_FRAME].tf.f;

  1896. 

  1897.     s->row = row;

  1898.     s->row7 = row & 7;

  1899.     s->col = col;

  1900.     s->col7 = col & 7;

  1901. 

  1902.     s->min_mv.x = -(128 + col * 64);

  1903.     s->min_mv.y = -(128 + row * 64);

  1904.     s->max_mv.x = 128 + (s->cols - col - w4) * 64;

  1905.     s->max_mv.y = 128 + (s->rows - row - h4) * 64;

  1906. 

  1907.     if (s->pass < 2) {

  1908.         b->bs = bs;

  1909.         b->bl = bl;

  1910.         b->bp = bp;

  1911.         decode_mode(avctx);

  1912.         b->uvtx = b->tx - ((s->ss_h && w4 * 2 == (1 << b->tx)) ||

  1913.                            (s->ss_v && h4 * 2 == (1 << b->tx)));

  1914. 

  1915.         if (!b->skip) {

  1916.             int has_coeffs;

  1917. 

  1918.             if (bytesperpixel == 1) {

  1919.                 has_coeffs = decode_coeffs_8bpp(avctx);

  1920.             } else {

  1921.                 has_coeffs = decode_coeffs_16bpp(avctx);

  1922.             }

  1923.             if (!has_coeffs && b->bs <= BS_8x8 && !b->intra) {

  1924.                 b->skip = 1;

  1925.                 memset(&s->above_skip_ctx[col], 1, w4);

  1926.                 memset(&s->left_skip_ctx[s->row7], 1, h4);

  1927.             }

  1928.         } else {

  1929.             int row7 = s->row7;

  1930. 

  1931. #define SPLAT_ZERO_CTX(v, n) \

  1932.     switch (n) { \

  1933.     case 1:  v = 0;          break; \

  1934.     case 2:  AV_ZERO16(&v);  break; \

  1935.     case 4:  AV_ZERO32(&v);  break; \

  1936.     case 8:  AV_ZERO64(&v);  break; \

  1937.     case 16: AV_ZERO128(&v); break; \

  1938.     }

  1939. #define SPLAT_ZERO_YUV(dir, var, off, n, dir2) \

  1940.     do { \

  1941.         SPLAT_ZERO_CTX(s->dir##_y_##var[off * 2], n * 2); \

  1942.         if (s->ss_##dir2) { \

  1943.             SPLAT_ZERO_CTX(s->dir##_uv_##var[0][off], n); \

  1944.             SPLAT_ZERO_CTX(s->dir##_uv_##var[1][off], n); \

  1945.         } else { \

  1946.             SPLAT_ZERO_CTX(s->dir##_uv_##var[0][off * 2], n * 2); \

  1947.             SPLAT_ZERO_CTX(s->dir##_uv_##var[1][off * 2], n * 2); \

  1948.         } \

  1949.     } while (0)

  1950. 

  1951.             switch (w4) {

  1952.             case 1: SPLAT_ZERO_YUV(above, nnz_ctx, col, 1, h); break;

  1953.             case 2: SPLAT_ZERO_YUV(above, nnz_ctx, col, 2, h); break;

  1954.             case 4: SPLAT_ZERO_YUV(above, nnz_ctx, col, 4, h); break;

  1955.             case 8: SPLAT_ZERO_YUV(above, nnz_ctx, col, 8, h); break;

  1956.             }

  1957.             switch (h4) {

  1958.             case 1: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 1, v); break;

  1959.             case 2: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 2, v); break;

  1960.             case 4: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 4, v); break;

  1961.             case 8: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 8, v); break;

  1962.             }

  1963.         }

  1964. 

  1965.         if (s->pass == 1) {

  1966.             s->b++;

  1967.             s->block += w4 * h4 * 64 * bytesperpixel;

  1968.             s->uvblock[0] += w4 * h4 * 64 * bytesperpixel >> (s->ss_h + s->ss_v);

  1969.             s->uvblock[1] += w4 * h4 * 64 * bytesperpixel >> (s->ss_h + s->ss_v);

  1970.             s->eob += 4 * w4 * h4;

  1971.             s->uveob[0] += 4 * w4 * h4 >> (s->ss_h + s->ss_v);

  1972.             s->uveob[1] += 4 * w4 * h4 >> (s->ss_h + s->ss_v);

  1973. 

  1974.             return;

  1975.         }

  1976.     }

  1977. 

  1978.     // emulated overhangs if the stride of the target buffer can't hold. This

  1979.     // makes it possible to support emu-edge and so on even if we have large block

  1980.     // overhangs

  1981.     emu[0] = (col + w4) * 8 * bytesperpixel > f->linesize[0] ||

  1982.              (row + h4) > s->rows;

  1983.     emu[1] = ((col + w4) * 8 >> s->ss_h) * bytesperpixel > f->linesize[1] ||

  1984.              (row + h4) > s->rows;

  1985.     if (emu[0]) {

  1986.         s->dst[0] = s->tmp_y;

  1987.         s->y_stride = 128;

  1988.     } else {

  1989.         s->dst[0] = f->data[0] + yoff;

  1990.         s->y_stride = f->linesize[0];

  1991.     }

  1992.     if (emu[1]) {

  1993.         s->dst[1] = s->tmp_uv[0];

  1994.         s->dst[2] = s->tmp_uv[1];

  1995.         s->uv_stride = 128;

  1996.     } else {

  1997.         s->dst[1] = f->data[1] + uvoff;

  1998.         s->dst[2] = f->data[2] + uvoff;

  1999.         s->uv_stride = f->linesize[1];

  2000.     }

  2001.     if (b->intra) {

  2002.         if (s->s.h.bpp > 8) {

  2003.             intra_recon_16bpp(avctx, yoff, uvoff);

  2004.         } else {

  2005.             intra_recon_8bpp(avctx, yoff, uvoff);

  2006.         }

  2007.     } else {

  2008.         if (s->s.h.bpp > 8) {

  2009.             inter_recon_16bpp(avctx);

  2010.         } else {

  2011.             inter_recon_8bpp(avctx);

  2012.         }

  2013.     }

  2014.     if (emu[0]) {

  2015.         int w = FFMIN(s->cols - col, w4) * 8, h = FFMIN(s->rows - row, h4) * 8, n, o = 0;

  2016. 

  2017.         for (n = 0; o < w; n++) {

  2018.             int bw = 64 >> n;

  2019. 

  2020.             av_assert2(n <= 4);

  2021.             if (w & bw) {

  2022.                 s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o * bytesperpixel, f->linesize[0],

  2023.                                          s->tmp_y + o * bytesperpixel, 128, h, 0, 0);

  2024.                 o += bw;

  2025.             }

  2026.         }

  2027.     }

  2028.     if (emu[1]) {

  2029.         int w = FFMIN(s->cols - col, w4) * 8 >> s->ss_h;

  2030.         int h = FFMIN(s->rows - row, h4) * 8 >> s->ss_v, n, o = 0;

  2031. 

  2032.         for (n = s->ss_h; o < w; n++) {

  2033.             int bw = 64 >> n;

  2034. 

  2035.             av_assert2(n <= 4);

  2036.             if (w & bw) {

  2037.                 s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o * bytesperpixel, f->linesize[1],

  2038.                                          s->tmp_uv[0] + o * bytesperpixel, 128, h, 0, 0);

  2039.                 s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o * bytesperpixel, f->linesize[2],

  2040.                                          s->tmp_uv[1] + o * bytesperpixel, 128, h, 0, 0);

  2041.                 o += bw;

  2042.             }

  2043.         }

  2044.     }

  2045. 

  2046.     // pick filter level and find edges to apply filter to

  2047.     if (s->s.h.filter.level &&

  2048.         (lvl = s->s.h.segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]

  2049.                                                       [b->mode[3] != ZEROMV]) > 0) {

  2050.         int x_end = FFMIN(s->cols - col, w4), y_end = FFMIN(s->rows - row, h4);

  2051.         int skip_inter = !b->intra && b->skip, col7 = s->col7, row7 = s->row7;

  2052. 

  2053.         setctx_2d(&lflvl->level[row7 * 8 + col7], w4, h4, 8, lvl);

  2054.         mask_edges(lflvl->mask[0], 0, 0, row7, col7, x_end, y_end, 0, 0, b->tx, skip_inter);

  2055.         if (s->ss_h || s->ss_v)

  2056.             mask_edges(lflvl->mask[1], s->ss_h, s->ss_v, row7, col7, x_end, y_end,

  2057.                        s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,

  2058.                        s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,

  2059.                        b->uvtx, skip_inter);

  2060. 

  2061.         if (!s->filter_lut.lim_lut[lvl]) {

  2062.             int sharp = s->s.h.filter.sharpness;

  2063.             int limit = lvl;

  2064. 

  2065.             if (sharp > 0) {

  2066.                 limit >>= (sharp + 3) >> 2;

  2067.                 limit = FFMIN(limit, 9 - sharp);

  2068.             }

  2069.             limit = FFMAX(limit, 1);

  2070. 

  2071.             s->filter_lut.lim_lut[lvl] = limit;

  2072.             s->filter_lut.mblim_lut[lvl] = 2 * (lvl + 2) + limit;

  2073.         }

  2074.     }

  2075. 

  2076.     if (s->pass == 2) {

  2077.         s->b++;

  2078.         s->block += w4 * h4 * 64 * bytesperpixel;

  2079.         s->uvblock[0] += w4 * h4 * 64 * bytesperpixel >> (s->ss_v + s->ss_h);

  2080.         s->uvblock[1] += w4 * h4 * 64 * bytesperpixel >> (s->ss_v + s->ss_h);

  2081.         s->eob += 4 * w4 * h4;

  2082.         s->uveob[0] += 4 * w4 * h4 >> (s->ss_v + s->ss_h);

  2083.         s->uveob[1] += 4 * w4 * h4 >> (s->ss_v + s->ss_h);

  2084.     }

  2085. }