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.
84741 Commits
32 Branches
283MB
Tree: f8c019944d
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'f8c019944d'
${ noResults }
FFmpeg/libavcodec/vp9block.c

2087 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. #include "vp9dec.h"

  33. 

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

  35.     {

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

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

  38.     }, {

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

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

  41.     }

  42. };

  43. 

  44. static av_always_inline void setctx_2d(uint8_t *ptr, int w, int h,

  45.                                        ptrdiff_t stride, int v)

  46. {

  47.     switch (w) {

  48.     case 1:

  49.         do {

  50.             *ptr = v;

  51.             ptr += stride;

  52.         } while (--h);

  53.         break;

  54.     case 2: {

  55.         int v16 = v * 0x0101;

  56.         do {

  57.             AV_WN16A(ptr, v16);

  58.             ptr += stride;

  59.         } while (--h);

  60.         break;

  61.     }

  62.     case 4: {

  63.         uint32_t v32 = v * 0x01010101;

  64.         do {

  65.             AV_WN32A(ptr, v32);

  66.             ptr += stride;

  67.         } while (--h);

  68.         break;

  69.     }

  70.     case 8: {

  71. #if HAVE_FAST_64BIT

  72.         uint64_t v64 = v * 0x0101010101010101ULL;

  73.         do {

  74.             AV_WN64A(ptr, v64);

  75.             ptr += stride;

  76.         } while (--h);

  77. #else

  78.         uint32_t v32 = v * 0x01010101;

  79.         do {

  80.             AV_WN32A(ptr,     v32);

  81.             AV_WN32A(ptr + 4, v32);

  82.             ptr += stride;

  83.         } while (--h);

  84. #endif

  85.         break;

  86.     }

  87.     }

  88. }

  89. 

  90. static void decode_mode(AVCodecContext *avctx)

  91. {

  92.     static const uint8_t left_ctx[N_BS_SIZES] = {

  93.         0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf

  94.     };

  95.     static const uint8_t above_ctx[N_BS_SIZES] = {

  96.         0x0, 0x0, 0x8, 0x8, 0x8, 0xc, 0xc, 0xc, 0xe, 0xe, 0xe, 0xf, 0xf

  97.     };

  98.     static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {

  99.         TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,

  100.         TX_16X16, TX_8X8,   TX_8X8,   TX_8X8,   TX_4X4,   TX_4X4,  TX_4X4

  101.     };

  102.     VP9Context *s = avctx->priv_data;

  103.     VP9Block *b = s->b;

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

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

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

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

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

  109.     int vref, filter_id;

  110. 

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

  112.         b->seg_id = 0;

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

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

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

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

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

  118.                 vp56_rac_get_prob_branchy(&s->c,

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

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

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

  122.             int pred = 8, x;

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

  124. 

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

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

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

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

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

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

  131.             }

  132.             av_assert1(pred < 8);

  133.             b->seg_id = pred;

  134.         } else {

  135.             b->seg_id = 0;

  136.         }

  137. 

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

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

  140.     } else {

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

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

  143. 

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

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

  146.     }

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

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

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

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

  151.     }

  152. 

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

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

  155.     if (!b->skip) {

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

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

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

  159.     }

  160. 

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

  162.         b->intra = 1;

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

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

  165.     } else {

  166.         int c, bit;

  167. 

  168.         if (have_a && have_l) {

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

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

  171.         } else {

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

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

  174.         }

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

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

  177.         b->intra = !bit;

  178.     }

  179. 

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

  181.         int c;

  182.         if (have_a) {

  183.             if (have_l) {

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

  185.                      s->above_txfm_ctx[col]) +

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

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

  188.             } else {

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

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

  191.             }

  192.         } else if (have_l) {

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

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

  195.         } else {

  196.             c = 1;

  197.         }

  198.         switch (max_tx) {

  199.         case TX_32X32:

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

  201.             if (b->tx) {

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

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

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

  205.             }

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

  207.             break;

  208.         case TX_16X16:

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

  210.             if (b->tx)

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

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

  213.             break;

  214.         case TX_8X8:

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

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

  217.             break;

  218.         case TX_4X4:

  219.             b->tx = TX_4X4;

  220.             break;

  221.         }

  222.     } else {

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

  224.     }

  225. 

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

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

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

  229. 

  230.         b->comp = 0;

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

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

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

  234.             // simpler for now

  235.             b->mode[0] =

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

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

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

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

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

  241.                 l[0]       =

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

  243.             } else {

  244.                 l[0]       =

  245.                 a[1]       =

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

  247.             }

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

  249.                 b->mode[2] =

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

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

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

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

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

  255.                     l[1]       =

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

  257.                 } else {

  258.                     l[1]       =

  259.                     a[1]       =

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

  261.                 }

  262.             } else {

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

  264.                 l[1]       =

  265.                 a[1]       =

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

  267.             }

  268.         } else {

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

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

  271.             b->mode[3] =

  272.             b->mode[2] =

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

  274.             // FIXME this can probably be optimized

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

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

  277.         }

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

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

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

  281.         b->comp = 0;

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

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

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

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

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

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

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

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

  290.             } else {

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

  292.             }

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

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

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

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

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

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

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

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

  301.                 } else {

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

  303.                 }

  304.             } else {

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

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

  307.             }

  308.         } else {

  309.             static const uint8_t size_group[10] = {

  310.                 3, 3, 3, 3, 2, 2, 2, 1, 1, 1

  311.             };

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

  313. 

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

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

  316.             b->mode[1] =

  317.             b->mode[2] =

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

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

  320.         }

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

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

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

  324.     } else {

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

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

  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, 2, 2, 1, 3 },

  338.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },

  339.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },

  340.         };

  341. 

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

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

  344.             b->comp = 0;

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

  346.         } else {

  347.             // read comp_pred flag

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

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

  350.             } else {

  351.                 int c;

  352. 

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

  354.                 if (have_a) {

  355.                     if (have_l) {

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

  357.                             c = 4;

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

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

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

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

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

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

  364.                         } else {

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

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

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

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

  369.                         }

  370.                     } else {

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

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

  373.                     }

  374.                 } else if (have_l) {

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

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

  377.                 } else {

  378.                     c = 1;

  379.                 }

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

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

  382.             }

  383. 

  384.             // read actual references

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

  386.             // memory accesses below

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

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

  389. 

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

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

  392.                 if (have_a) {

  393.                     if (have_l) {

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

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

  396.                                 c = 2;

  397.                             } else {

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

  399.                             }

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

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

  402.                         } else {

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

  404. 

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

  406.                                 c = 0;

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

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

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

  410.                                     c = 4;

  411.                                 } else {

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

  413.                                 }

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

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

  416.                                     c = 1;

  417.                                 } else {

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

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

  420.                                 }

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

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

  423.                                     c = 1;

  424.                                 } else {

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

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

  427.                                 }

  428.                             } else {

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

  430.                             }

  431.                         }

  432.                     } else {

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

  434.                             c = 2;

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

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

  437.                         } else {

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

  439.                         }

  440.                     }

  441.                 } else if (have_l) {

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

  443.                         c = 2;

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

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

  446.                     } else {

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

  448.                     }

  449.                 } else {

  450.                     c = 2;

  451.                 }

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

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

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

  455.             } else /* single reference */ {

  456.                 int bit, c;

  457. 

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

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

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

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

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

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

  464.                             } else {

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

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

  467.                             }

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

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

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

  471.                         } else {

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

  473.                         }

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

  475.                         c = 2;

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

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

  478.                     } else {

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

  480.                     }

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

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

  483.                         c = 2;

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

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

  486.                     } else {

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

  488.                     }

  489.                 } else {

  490.                     c = 2;

  491.                 }

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

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

  494.                 if (!bit) {

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

  496.                 } else {

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

  498.                     if (have_a) {

  499.                         if (have_l) {

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

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

  502.                                     c = 2;

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

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

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

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

  507.                                     c = 3;

  508.                                 } else {

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

  510.                                 }

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

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

  513.                                     c = 2;

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

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

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

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

  518.                                     c = 3;

  519.                                 } else {

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

  521.                                 }

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

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

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

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

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

  527.                                     } else {

  528.                                         c = 2;

  529.                                     }

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

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

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

  533.                                 } else {

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

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

  536.                                 }

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

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

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

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

  541.                                 } else {

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

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

  544.                                 }

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

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

  547.                                     c = 3;

  548.                                 } else {

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

  550.                                 }

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

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

  553.                             } else {

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

  555.                                     2 * (s->above_ref_ctx[col] == 1);

  556.                             }

  557.                         } else {

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

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

  560.                                 c = 2;

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

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

  563.                             } else {

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

  565.                             }

  566.                         }

  567.                     } else if (have_l) {

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

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

  570.                             c = 2;

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

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

  573.                         } else {

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

  575.                         }

  576.                     } else {

  577.                         c = 2;

  578.                     }

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

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

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

  582.                 }

  583.             }

  584.         }

  585. 

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

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

  588.                 b->mode[0] =

  589.                 b->mode[1] =

  590.                 b->mode[2] =

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

  592.             } else {

  593.                 static const uint8_t off[10] = {

  594.                     3, 0, 0, 1, 0, 0, 0, 0, 0, 0

  595.                 };

  596. 

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

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

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

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

  601. 

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

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

  604.                 b->mode[1] =

  605.                 b->mode[2] =

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

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

  608.             }

  609.         }

  610. 

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

  612.             int c;

  613. 

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

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

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

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

  618.                 } else {

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

  620.                 }

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

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

  623.             } else {

  624.                 c = 3;

  625.             }

  626. 

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

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

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

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

  631.         } else {

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

  633.         }

  634. 

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

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

  637. 

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

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

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

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

  642. 

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

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

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

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

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

  648.             } else {

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

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

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

  652.             }

  653. 

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

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

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

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

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

  659. 

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

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

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

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

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

  665.                 } else {

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

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

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

  669.                 }

  670.             } else {

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

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

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

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

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

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

  677.             }

  678.         } else {

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

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

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

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

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

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

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

  686.         }

  687. 

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

  689.     }

  690. 

  691. #if HAVE_FAST_64BIT

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

  693.     switch (n) { \

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

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

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

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

  698.     case 16: { \

  699.         uint64_t v64 = val * 0x0101010101010101ULL; \

  700.         AV_WN64A(              &var,     v64); \

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

  702.         break; \

  703.     } \

  704.     }

  705. #else

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

  707.     switch (n) { \

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

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

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

  711.     case 8: { \

  712.         uint32_t v32 = val * 0x01010101; \

  713.         AV_WN32A(              &var,     v32); \

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

  715.         break; \

  716.     } \

  717.     case 16: { \

  718.         uint32_t v32 = val * 0x01010101; \

  719.         AV_WN32A(              &var,      v32); \

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

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

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

  723.         break; \

  724.     } \

  725.     }

  726. #endif

  727. 

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

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

  730.     do { \

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

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

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

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

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

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

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

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

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

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

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

  742.                 } \

  743.             } \

  744.         } \

  745.     } while (0)

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

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

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

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

  750.     }

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

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

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

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

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

  756.     }

  757. #undef SPLAT_CTX

  758. #undef SET_CTXS

  759. 

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

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

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

  763. 

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

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

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

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

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

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

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

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

  772.         } else {

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

  774. 

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

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

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

  778.             }

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

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

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

  782.             }

  783.         }

  784.     }

  785. 

  786.     // FIXME kinda ugly

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

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

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

  790. 

  791.         if (b->intra) {

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

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

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

  795.             }

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

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

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

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

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

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

  802.             }

  803.         } else {

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

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

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

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

  808.             }

  809.         }

  810.     }

  811. }

  812. 

  813. // FIXME merge cnt/eob arguments?

  814. static av_always_inline int

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

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

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

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

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

  820. {

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

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

  823.     uint8_t cache[1024];

  824. 

  825.     do {

  826.         int val, rc;

  827. 

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

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

  830.         if (!val)

  831.             break;

  832. 

  833. skip_eob:

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

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

  836.             if (!--band_left)

  837.                 band_left = band_counts[++band];

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

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

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

  841.             if (++i == n_coeffs)

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

  843.             goto skip_eob;

  844.         }

  845. 

  846.         rc = scan[i];

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

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

  849.             val       = 1;

  850.             cache[rc] = 1;

  851.         } else {

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

  853.             if (!tp[3])

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

  855. 

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

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

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

  859.                     cache[rc] = val = 2;

  860.                 } else {

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

  862.                     cache[rc] = 3;

  863.                 }

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

  865.                 cache[rc] = 4;

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

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

  868.                 } else {

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

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

  871.                 }

  872.             } else { // cat 3-6

  873.                 cache[rc] = 5;

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

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

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

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

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

  879.                     } else {

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

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

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

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

  884.                     }

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

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

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

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

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

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

  891.                 } else {

  892.                     val = 67;

  893.                     if (!is8bitsperpixel) {

  894.                         if (bpp == 12) {

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

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

  897.                         }

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

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

  900.                     }

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  915.                 }

  916.             }

  917.         }

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

  919.     if (is8bitsperpixel) { \

  920.         c[i] = v; \

  921.     } else { \

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

  923.     } \

  924. } while (0)

  925.         if (!--band_left)

  926.             band_left = band_counts[++band];

  927.         if (is_tx32x32)

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

  929.         else

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

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

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

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

  934. 

  935.     return i;

  936. }

  937. 

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

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

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

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

  942.                                 const int16_t *qmul)

  943. {

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

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

  946. }

  947. 

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

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

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

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

  952.                                   const int16_t *qmul)

  953. {

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

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

  956. }

  957. 

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

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

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

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

  962.                                  const int16_t *qmul)

  963. {

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

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

  966. }

  967. 

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

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

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

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

  972.                                    const int16_t *qmul)

  973. {

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

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

  976. }

  977. 

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

  979. {

  980.     VP9Context *s = avctx->priv_data;

  981.     VP9Block *b = s->b;

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  999.         { 1, 2, 3, 4,  3,   16 - 13 },

  1000.         { 1, 2, 3, 4, 11,   64 - 21 },

  1001.         { 1, 2, 3, 4, 11,  256 - 21 },

  1002.         { 1, 2, 3, 4, 11, 1024 - 21 },

  1003.     };

  1004.     const int16_t *y_band_counts = band_counts[b->tx];

  1005.     const int16_t *uv_band_counts = band_counts[b->uvtx];

  1006.     int bytesperpixel = is8bitsperpixel ? 1 : 2;

  1007.     int total_coeff = 0;

  1008. 

  1009. #define MERGE(la, end, step, rd) \

  1010.     for (n = 0; n < end; n += step) \

  1011.         la[n] = !!rd(&la[n])

  1012. #define MERGE_CTX(step, rd) \

  1013.     do { \

  1014.         MERGE(l, end_y, step, rd); \

  1015.         MERGE(a, end_x, step, rd); \

  1016.     } while (0)

  1017. 

  1018. #define DECODE_Y_COEF_LOOP(step, mode_index, v) \

  1019.     for (n = 0, y = 0; y < end_y; y += step) { \

  1020.         for (x = 0; x < end_x; x += step, n += step * step) { \

  1021.             enum TxfmType txtp = ff_vp9_intra_txfm_type[b->mode[mode_index]]; \

  1022.             ret = (is8bitsperpixel ? decode_coeffs_b##v##_8bpp : decode_coeffs_b##v##_16bpp) \

  1023.                                     (s, s->block + 16 * n * bytesperpixel, 16 * step * step, \

  1024.                                      c, e, p, a[x] + l[y], yscans[txtp], \

  1025.                                      ynbs[txtp], y_band_counts, qmul[0]); \

  1026.             a[x] = l[y] = !!ret; \

  1027.             total_coeff |= !!ret; \

  1028.             if (step >= 4) { \

  1029.                 AV_WN16A(&s->eob[n], ret); \

  1030.             } else { \

  1031.                 s->eob[n] = ret; \

  1032.             } \

  1033.         } \

  1034.     }

  1035. 

  1036. #define SPLAT(la, end, step, cond) \

  1037.     if (step == 2) { \

  1038.         for (n = 1; n < end; n += step) \

  1039.             la[n] = la[n - 1]; \

  1040.     } else if (step == 4) { \

  1041.         if (cond) { \

  1042.             for (n = 0; n < end; n += step) \

  1043.                 AV_WN32A(&la[n], la[n] * 0x01010101); \

  1044.         } else { \

  1045.             for (n = 0; n < end; n += step) \

  1046.                 memset(&la[n + 1], la[n], FFMIN(end - n - 1, 3)); \

  1047.         } \

  1048.     } else /* step == 8 */ { \

  1049.         if (cond) { \

  1050.             if (HAVE_FAST_64BIT) { \

  1051.                 for (n = 0; n < end; n += step) \

  1052.                     AV_WN64A(&la[n], la[n] * 0x0101010101010101ULL); \

  1053.             } else { \

  1054.                 for (n = 0; n < end; n += step) { \

  1055.                     uint32_t v32 = la[n] * 0x01010101; \

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

  1057.                     AV_WN32A(&la[n + 4], v32); \

  1058.                 } \

  1059.             } \

  1060.         } else { \

  1061.             for (n = 0; n < end; n += step) \

  1062.                 memset(&la[n + 1], la[n], FFMIN(end - n - 1, 7)); \

  1063.         } \

  1064.     }

  1065. #define SPLAT_CTX(step) \

  1066.     do { \

  1067.         SPLAT(a, end_x, step, end_x == w4); \

  1068.         SPLAT(l, end_y, step, end_y == h4); \

  1069.     } while (0)

  1070. 

  1071.     /* y tokens */

  1072.     switch (b->tx) {

  1073.     case TX_4X4:

  1074.         DECODE_Y_COEF_LOOP(1, b->bs > BS_8x8 ? n : 0,);

  1075.         break;

  1076.     case TX_8X8:

  1077.         MERGE_CTX(2, AV_RN16A);

  1078.         DECODE_Y_COEF_LOOP(2, 0,);

  1079.         SPLAT_CTX(2);

  1080.         break;

  1081.     case TX_16X16:

  1082.         MERGE_CTX(4, AV_RN32A);

  1083.         DECODE_Y_COEF_LOOP(4, 0,);

  1084.         SPLAT_CTX(4);

  1085.         break;

  1086.     case TX_32X32:

  1087.         MERGE_CTX(8, AV_RN64A);

  1088.         DECODE_Y_COEF_LOOP(8, 0, 32);

  1089.         SPLAT_CTX(8);

  1090.         break;

  1091.     }

  1092. 

  1093. #define DECODE_UV_COEF_LOOP(step, v) \

  1094.     for (n = 0, y = 0; y < end_y; y += step) { \

  1095.         for (x = 0; x < end_x; x += step, n += step * step) { \

  1096.             ret = (is8bitsperpixel ? decode_coeffs_b##v##_8bpp : decode_coeffs_b##v##_16bpp) \

  1097.                                     (s, s->uvblock[pl] + 16 * n * bytesperpixel, \

  1098.                                      16 * step * step, c, e, p, a[x] + l[y], \

  1099.                                      uvscan, uvnb, uv_band_counts, qmul[1]); \

  1100.             a[x] = l[y] = !!ret; \

  1101.             total_coeff |= !!ret; \

  1102.             if (step >= 4) { \

  1103.                 AV_WN16A(&s->uveob[pl][n], ret); \

  1104.             } else { \

  1105.                 s->uveob[pl][n] = ret; \

  1106.             } \

  1107.         } \

  1108.     }

  1109. 

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

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

  1112.     e = s->counts.eob[b->uvtx][1 /* uv */][!b->intra];

  1113.     w4 >>= s->ss_h;

  1114.     end_x >>= s->ss_h;

  1115.     h4 >>= s->ss_v;

  1116.     end_y >>= s->ss_v;

  1117.     for (pl = 0; pl < 2; pl++) {

  1118.         a = &s->above_uv_nnz_ctx[pl][col << !s->ss_h];

  1119.         l = &s->left_uv_nnz_ctx[pl][(row & 7) << !s->ss_v];

  1120.         switch (b->uvtx) {

  1121.         case TX_4X4:

  1122.             DECODE_UV_COEF_LOOP(1,);

  1123.             break;

  1124.         case TX_8X8:

  1125.             MERGE_CTX(2, AV_RN16A);

  1126.             DECODE_UV_COEF_LOOP(2,);

  1127.             SPLAT_CTX(2);

  1128.             break;

  1129.         case TX_16X16:

  1130.             MERGE_CTX(4, AV_RN32A);

  1131.             DECODE_UV_COEF_LOOP(4,);

  1132.             SPLAT_CTX(4);

  1133.             break;

  1134.         case TX_32X32:

  1135.             MERGE_CTX(8, AV_RN64A);

  1136.             DECODE_UV_COEF_LOOP(8, 32);

  1137.             SPLAT_CTX(8);

  1138.             break;

  1139.         }

  1140.     }

  1141. 

  1142.     return total_coeff;

  1143. }

  1144. 

  1145. static int decode_coeffs_8bpp(AVCodecContext *avctx)

  1146. {

  1147.     return decode_coeffs(avctx, 1);

  1148. }

  1149. 

  1150. static int decode_coeffs_16bpp(AVCodecContext *avctx)

  1151. {

  1152.     return decode_coeffs(avctx, 0);

  1153. }

  1154. 

  1155. static av_always_inline int check_intra_mode(VP9Context *s, int mode, uint8_t **a,

  1156.                                              uint8_t *dst_edge, ptrdiff_t stride_edge,

  1157.                                              uint8_t *dst_inner, ptrdiff_t stride_inner,

  1158.                                              uint8_t *l, int col, int x, int w,

  1159.                                              int row, int y, enum TxfmMode tx,

  1160.                                              int p, int ss_h, int ss_v, int bytesperpixel)

  1161. {

  1162.     int have_top = row > 0 || y > 0;

  1163.     int have_left = col > s->tile_col_start || x > 0;

  1164.     int have_right = x < w - 1;

  1165.     int bpp = s->s.h.bpp;

  1166.     static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {

  1167.         [VERT_PRED]            = { { DC_127_PRED,          VERT_PRED            },

  1168.                                    { DC_127_PRED,          VERT_PRED            } },

  1169.         [HOR_PRED]             = { { DC_129_PRED,          DC_129_PRED          },

  1170.                                    { HOR_PRED,             HOR_PRED             } },

  1171.         [DC_PRED]              = { { DC_128_PRED,          TOP_DC_PRED          },

  1172.                                    { LEFT_DC_PRED,         DC_PRED              } },

  1173.         [DIAG_DOWN_LEFT_PRED]  = { { DC_127_PRED,          DIAG_DOWN_LEFT_PRED  },

  1174.                                    { DC_127_PRED,          DIAG_DOWN_LEFT_PRED  } },

  1175.         [DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },

  1176.                                    { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },

  1177.         [VERT_RIGHT_PRED]      = { { VERT_RIGHT_PRED,      VERT_RIGHT_PRED      },

  1178.                                    { VERT_RIGHT_PRED,      VERT_RIGHT_PRED      } },

  1179.         [HOR_DOWN_PRED]        = { { HOR_DOWN_PRED,        HOR_DOWN_PRED        },

  1180.                                    { HOR_DOWN_PRED,        HOR_DOWN_PRED        } },

  1181.         [VERT_LEFT_PRED]       = { { DC_127_PRED,          VERT_LEFT_PRED       },

  1182.                                    { DC_127_PRED,          VERT_LEFT_PRED       } },

  1183.         [HOR_UP_PRED]          = { { DC_129_PRED,          DC_129_PRED          },

  1184.                                    { HOR_UP_PRED,          HOR_UP_PRED          } },

  1185.         [TM_VP8_PRED]          = { { DC_129_PRED,          VERT_PRED            },

  1186.                                    { HOR_PRED,             TM_VP8_PRED          } },

  1187.     };

  1188.     static const struct {

  1189.         uint8_t needs_left:1;

  1190.         uint8_t needs_top:1;

  1191.         uint8_t needs_topleft:1;

  1192.         uint8_t needs_topright:1;

  1193.         uint8_t invert_left:1;

  1194.     } edges[N_INTRA_PRED_MODES] = {

  1195.         [VERT_PRED]            = { .needs_top  = 1 },

  1196.         [HOR_PRED]             = { .needs_left = 1 },

  1197.         [DC_PRED]              = { .needs_top  = 1, .needs_left = 1 },

  1198.         [DIAG_DOWN_LEFT_PRED]  = { .needs_top  = 1, .needs_topright = 1 },

  1199.         [DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1,

  1200.                                    .needs_topleft = 1 },

  1201.         [VERT_RIGHT_PRED]      = { .needs_left = 1, .needs_top = 1,

  1202.                                    .needs_topleft = 1 },

  1203.         [HOR_DOWN_PRED]        = { .needs_left = 1, .needs_top = 1,

  1204.                                    .needs_topleft = 1 },

  1205.         [VERT_LEFT_PRED]       = { .needs_top  = 1, .needs_topright = 1 },

  1206.         [HOR_UP_PRED]          = { .needs_left = 1, .invert_left = 1 },

  1207.         [TM_VP8_PRED]          = { .needs_left = 1, .needs_top = 1,

  1208.                                    .needs_topleft = 1 },

  1209.         [LEFT_DC_PRED]         = { .needs_left = 1 },

  1210.         [TOP_DC_PRED]          = { .needs_top  = 1 },

  1211.         [DC_128_PRED]          = { 0 },

  1212.         [DC_127_PRED]          = { 0 },

  1213.         [DC_129_PRED]          = { 0 }

  1214.     };

  1215. 

  1216.     av_assert2(mode >= 0 && mode < 10);

  1217.     mode = mode_conv[mode][have_left][have_top];

  1218.     if (edges[mode].needs_top) {

  1219.         uint8_t *top, *topleft;

  1220.         int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !ss_h) - x) * 4;

  1221.         int n_px_need_tr = 0;

  1222. 

  1223.         if (tx == TX_4X4 && edges[mode].needs_topright && have_right)

  1224.             n_px_need_tr = 4;

  1225. 

  1226.         // if top of sb64-row, use s->intra_pred_data[] instead of

  1227.         // dst[-stride] for intra prediction (it contains pre- instead of

  1228.         // post-loopfilter data)

  1229.         if (have_top) {

  1230.             top = !(row & 7) && !y ?

  1231.                 s->intra_pred_data[p] + (col * (8 >> ss_h) + x * 4) * bytesperpixel :

  1232.                 y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];

  1233.             if (have_left)

  1234.                 topleft = !(row & 7) && !y ?

  1235.                     s->intra_pred_data[p] + (col * (8 >> ss_h) + x * 4) * bytesperpixel :

  1236.                     y == 0 || x == 0 ? &dst_edge[-stride_edge] :

  1237.                     &dst_inner[-stride_inner];

  1238.         }

  1239. 

  1240.         if (have_top &&

  1241.             (!edges[mode].needs_topleft || (have_left && top == topleft)) &&

  1242.             (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&

  1243.             n_px_need + n_px_need_tr <= n_px_have) {

  1244.             *a = top;

  1245.         } else {

  1246.             if (have_top) {

  1247.                 if (n_px_need <= n_px_have) {

  1248.                     memcpy(*a, top, n_px_need * bytesperpixel);

  1249.                 } else {

  1250. #define memset_bpp(c, i1, v, i2, num) do { \

  1251.     if (bytesperpixel == 1) { \

  1252.         memset(&(c)[(i1)], (v)[(i2)], (num)); \

  1253.     } else { \

  1254.         int n, val = AV_RN16A(&(v)[(i2) * 2]); \

  1255.         for (n = 0; n < (num); n++) { \

  1256.             AV_WN16A(&(c)[((i1) + n) * 2], val); \

  1257.         } \

  1258.     } \

  1259. } while (0)

  1260.                     memcpy(*a, top, n_px_have * bytesperpixel);

  1261.                     memset_bpp(*a, n_px_have, (*a), n_px_have - 1, n_px_need - n_px_have);

  1262.                 }

  1263.             } else {

  1264. #define memset_val(c, val, num) do { \

  1265.     if (bytesperpixel == 1) { \

  1266.         memset((c), (val), (num)); \

  1267.     } else { \

  1268.         int n; \

  1269.         for (n = 0; n < (num); n++) { \

  1270.             AV_WN16A(&(c)[n * 2], (val)); \

  1271.         } \

  1272.     } \

  1273. } while (0)

  1274.                 memset_val(*a, (128 << (bpp - 8)) - 1, n_px_need);

  1275.             }

  1276.             if (edges[mode].needs_topleft) {

  1277.                 if (have_left && have_top) {

  1278. #define assign_bpp(c, i1, v, i2) do { \

  1279.     if (bytesperpixel == 1) { \

  1280.         (c)[(i1)] = (v)[(i2)]; \

  1281.     } else { \

  1282.         AV_COPY16(&(c)[(i1) * 2], &(v)[(i2) * 2]); \

  1283.     } \

  1284. } while (0)

  1285.                     assign_bpp(*a, -1, topleft, -1);

  1286.                 } else {

  1287. #define assign_val(c, i, v) do { \

  1288.     if (bytesperpixel == 1) { \

  1289.         (c)[(i)] = (v); \

  1290.     } else { \

  1291.         AV_WN16A(&(c)[(i) * 2], (v)); \

  1292.     } \

  1293. } while (0)

  1294.                     assign_val((*a), -1, (128 << (bpp - 8)) + (have_top ? +1 : -1));

  1295.                 }

  1296.             }

  1297.             if (tx == TX_4X4 && edges[mode].needs_topright) {

  1298.                 if (have_top && have_right &&

  1299.                     n_px_need + n_px_need_tr <= n_px_have) {

  1300.                     memcpy(&(*a)[4 * bytesperpixel], &top[4 * bytesperpixel], 4 * bytesperpixel);

  1301.                 } else {

  1302.                     memset_bpp(*a, 4, *a, 3, 4);

  1303.                 }

  1304.             }

  1305.         }

  1306.     }

  1307.     if (edges[mode].needs_left) {

  1308.         if (have_left) {

  1309.             int n_px_need = 4 << tx, i, n_px_have = (((s->rows - row) << !ss_v) - y) * 4;

  1310.             uint8_t *dst = x == 0 ? dst_edge : dst_inner;

  1311.             ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;

  1312. 

  1313.             if (edges[mode].invert_left) {

  1314.                 if (n_px_need <= n_px_have) {

  1315.                     for (i = 0; i < n_px_need; i++)

  1316.                         assign_bpp(l, i, &dst[i * stride], -1);

  1317.                 } else {

  1318.                     for (i = 0; i < n_px_have; i++)

  1319.                         assign_bpp(l, i, &dst[i * stride], -1);

  1320.                     memset_bpp(l, n_px_have, l, n_px_have - 1, n_px_need - n_px_have);

  1321.                 }

  1322.             } else {

  1323.                 if (n_px_need <= n_px_have) {

  1324.                     for (i = 0; i < n_px_need; i++)

  1325.                         assign_bpp(l, n_px_need - 1 - i, &dst[i * stride], -1);

  1326.                 } else {

  1327.                     for (i = 0; i < n_px_have; i++)

  1328.                         assign_bpp(l, n_px_need - 1 - i, &dst[i * stride], -1);

  1329.                     memset_bpp(l, 0, l, n_px_need - n_px_have, n_px_need - n_px_have);

  1330.                 }

  1331.             }

  1332.         } else {

  1333.             memset_val(l, (128 << (bpp - 8)) + 1, 4 << tx);

  1334.         }

  1335.     }

  1336. 

  1337.     return mode;

  1338. }

  1339. 

  1340. static av_always_inline void intra_recon(AVCodecContext *avctx, ptrdiff_t y_off,

  1341.                                          ptrdiff_t uv_off, int bytesperpixel)

  1342. {

  1343.     VP9Context *s = avctx->priv_data;

  1344.     VP9Block *b = s->b;

  1345.     int row = s->row, col = s->col;

  1346.     int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  1347.     int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  1348.     int end_x = FFMIN(2 * (s->cols - col), w4);

  1349.     int end_y = FFMIN(2 * (s->rows - row), h4);

  1350.     int tx = 4 * s->s.h.lossless + b->tx, uvtx = b->uvtx + 4 * s->s.h.lossless;

  1351.     int uvstep1d = 1 << b->uvtx, p;

  1352.     uint8_t *dst = s->dst[0], *dst_r = s->s.frames[CUR_FRAME].tf.f->data[0] + y_off;

  1353.     LOCAL_ALIGNED_32(uint8_t, a_buf, [96]);

  1354.     LOCAL_ALIGNED_32(uint8_t, l, [64]);

  1355. 

  1356.     for (n = 0, y = 0; y < end_y; y += step1d) {

  1357.         uint8_t *ptr = dst, *ptr_r = dst_r;

  1358.         for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d * bytesperpixel,

  1359.                                ptr_r += 4 * step1d * bytesperpixel, n += step) {

  1360.             int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?

  1361.                                y * 2 + x : 0];

  1362.             uint8_t *a = &a_buf[32];

  1363.             enum TxfmType txtp = ff_vp9_intra_txfm_type[mode];

  1364.             int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  1365. 

  1366.             mode = check_intra_mode(s, mode, &a, ptr_r,

  1367.                                     s->s.frames[CUR_FRAME].tf.f->linesize[0],

  1368.                                     ptr, s->y_stride, l,

  1369.                                     col, x, w4, row, y, b->tx, 0, 0, 0, bytesperpixel);

  1370.             s->dsp.intra_pred[b->tx][mode](ptr, s->y_stride, l, a);

  1371.             if (eob)

  1372.                 s->dsp.itxfm_add[tx][txtp](ptr, s->y_stride,

  1373.                                            s->block + 16 * n * bytesperpixel, eob);

  1374.         }

  1375.         dst_r += 4 * step1d * s->s.frames[CUR_FRAME].tf.f->linesize[0];

  1376.         dst   += 4 * step1d * s->y_stride;

  1377.     }

  1378. 

  1379.     // U/V

  1380.     w4    >>= s->ss_h;

  1381.     end_x >>= s->ss_h;

  1382.     end_y >>= s->ss_v;

  1383.     step = 1 << (b->uvtx * 2);

  1384.     for (p = 0; p < 2; p++) {

  1385.         dst   = s->dst[1 + p];

  1386.         dst_r = s->s.frames[CUR_FRAME].tf.f->data[1 + p] + uv_off;

  1387.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  1388.             uint8_t *ptr = dst, *ptr_r = dst_r;

  1389.             for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d * bytesperpixel,

  1390.                                    ptr_r += 4 * uvstep1d * bytesperpixel, n += step) {

  1391.                 int mode = b->uvmode;

  1392.                 uint8_t *a = &a_buf[32];

  1393.                 int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  1394. 

  1395.                 mode = check_intra_mode(s, mode, &a, ptr_r,

  1396.                                         s->s.frames[CUR_FRAME].tf.f->linesize[1],

  1397.                                         ptr, s->uv_stride, l, col, x, w4, row, y,

  1398.                                         b->uvtx, p + 1, s->ss_h, s->ss_v, bytesperpixel);

  1399.                 s->dsp.intra_pred[b->uvtx][mode](ptr, s->uv_stride, l, a);

  1400.                 if (eob)

  1401.                     s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  1402.                                                     s->uvblock[p] + 16 * n * bytesperpixel, eob);

  1403.             }

  1404.             dst_r += 4 * uvstep1d * s->s.frames[CUR_FRAME].tf.f->linesize[1];

  1405.             dst   += 4 * uvstep1d * s->uv_stride;

  1406.         }

  1407.     }

  1408. }

  1409. 

  1410. static void intra_recon_8bpp(AVCodecContext *avctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  1411. {

  1412.     intra_recon(avctx, y_off, uv_off, 1);

  1413. }

  1414. 

  1415. static void intra_recon_16bpp(AVCodecContext *avctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  1416. {

  1417.     intra_recon(avctx, y_off, uv_off, 2);

  1418. }

  1419. 

  1420. static av_always_inline void mc_luma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2],

  1421.                                               uint8_t *dst, ptrdiff_t dst_stride,

  1422.                                               const uint8_t *ref, ptrdiff_t ref_stride,

  1423.                                               ThreadFrame *ref_frame,

  1424.                                               ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  1425.                                               int bw, int bh, int w, int h, int bytesperpixel)

  1426. {

  1427.     int mx = mv->x, my = mv->y, th;

  1428. 

  1429.     y += my >> 3;

  1430.     x += mx >> 3;

  1431.     ref += y * ref_stride + x * bytesperpixel;

  1432.     mx &= 7;

  1433.     my &= 7;

  1434.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  1435.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  1436.     // the longest loopfilter of the next sbrow

  1437.     th = (y + bh + 4 * !!my + 7) >> 6;

  1438.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  1439.     // The arm/aarch64 _hv filters read one more row than what actually is

  1440.     // needed, so switch to emulated edge one pixel sooner vertically

  1441.     // (!!my * 5) than horizontally (!!mx * 4).

  1442.     if (x < !!mx * 3 || y < !!my * 3 ||

  1443.         x + !!mx * 4 > w - bw || y + !!my * 5 > h - bh) {

  1444.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1445.                                  ref - !!my * 3 * ref_stride - !!mx * 3 * bytesperpixel,

  1446.                                  160, ref_stride,

  1447.                                  bw + !!mx * 7, bh + !!my * 7,

  1448.                                  x - !!mx * 3, y - !!my * 3, w, h);

  1449.         ref = s->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;

  1450.         ref_stride = 160;

  1451.     }

  1452.     mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);

  1453. }

  1454. 

  1455. static av_always_inline void mc_chroma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2],

  1456.                                                 uint8_t *dst_u, uint8_t *dst_v,

  1457.                                                 ptrdiff_t dst_stride,

  1458.                                                 const uint8_t *ref_u, ptrdiff_t src_stride_u,

  1459.                                                 const uint8_t *ref_v, ptrdiff_t src_stride_v,

  1460.                                                 ThreadFrame *ref_frame,

  1461.                                                 ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  1462.                                                 int bw, int bh, int w, int h, int bytesperpixel)

  1463. {

  1464.     int mx = mv->x * (1 << !s->ss_h), my = mv->y * (1 << !s->ss_v), th;

  1465. 

  1466.     y += my >> 4;

  1467.     x += mx >> 4;

  1468.     ref_u += y * src_stride_u + x * bytesperpixel;

  1469.     ref_v += y * src_stride_v + x * bytesperpixel;

  1470.     mx &= 15;

  1471.     my &= 15;

  1472.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  1473.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  1474.     // the longest loopfilter of the next sbrow

  1475.     th = (y + bh + 4 * !!my + 7) >> (6 - s->ss_v);

  1476.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  1477.     // The arm/aarch64 _hv filters read one more row than what actually is

  1478.     // needed, so switch to emulated edge one pixel sooner vertically

  1479.     // (!!my * 5) than horizontally (!!mx * 4).

  1480.     if (x < !!mx * 3 || y < !!my * 3 ||

  1481.         x + !!mx * 4 > w - bw || y + !!my * 5 > h - bh) {

  1482.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1483.                                  ref_u - !!my * 3 * src_stride_u - !!mx * 3 * bytesperpixel,

  1484.                                  160, src_stride_u,

  1485.                                  bw + !!mx * 7, bh + !!my * 7,

  1486.                                  x - !!mx * 3, y - !!my * 3, w, h);

  1487.         ref_u = s->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;

  1488.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, 160, bh, mx, my);

  1489. 

  1490.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1491.                                  ref_v - !!my * 3 * src_stride_v - !!mx * 3 * bytesperpixel,

  1492.                                  160, src_stride_v,

  1493.                                  bw + !!mx * 7, bh + !!my * 7,

  1494.                                  x - !!mx * 3, y - !!my * 3, w, h);

  1495.         ref_v = s->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;

  1496.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, 160, bh, mx, my);

  1497.     } else {

  1498.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);

  1499.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);

  1500.     }

  1501. }

  1502. 

  1503. #define mc_luma_dir(s, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, \

  1504.                     px, py, pw, ph, bw, bh, w, h, i) \

  1505.     mc_luma_unscaled(s, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \

  1506.                      mv, bw, bh, w, h, bytesperpixel)

  1507. #define mc_chroma_dir(s, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  1508.                       row, col, mv, px, py, pw, ph, bw, bh, w, h, i) \

  1509.     mc_chroma_unscaled(s, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  1510.                        row, col, mv, bw, bh, w, h, bytesperpixel)

  1511. #define SCALED 0

  1512. #define FN(x) x##_8bpp

  1513. #define BYTES_PER_PIXEL 1

  1514. #include "vp9_mc_template.c"

  1515. #undef FN

  1516. #undef BYTES_PER_PIXEL

  1517. #define FN(x) x##_16bpp

  1518. #define BYTES_PER_PIXEL 2

  1519. #include "vp9_mc_template.c"

  1520. #undef mc_luma_dir

  1521. #undef mc_chroma_dir

  1522. #undef FN

  1523. #undef BYTES_PER_PIXEL

  1524. #undef SCALED

  1525. 

  1526. static av_always_inline void mc_luma_scaled(VP9Context *s, vp9_scaled_mc_func smc,

  1527.                                             vp9_mc_func (*mc)[2],

  1528.                                             uint8_t *dst, ptrdiff_t dst_stride,

  1529.                                             const uint8_t *ref, ptrdiff_t ref_stride,

  1530.                                             ThreadFrame *ref_frame,

  1531.                                             ptrdiff_t y, ptrdiff_t x, const VP56mv *in_mv,

  1532.                                             int px, int py, int pw, int ph,

  1533.                                             int bw, int bh, int w, int h, int bytesperpixel,

  1534.                                             const uint16_t *scale, const uint8_t *step)

  1535. {

  1536.     if (s->s.frames[CUR_FRAME].tf.f->width == ref_frame->f->width &&

  1537.         s->s.frames[CUR_FRAME].tf.f->height == ref_frame->f->height) {

  1538.         mc_luma_unscaled(s, mc, dst, dst_stride, ref, ref_stride, ref_frame,

  1539.                          y, x, in_mv, bw, bh, w, h, bytesperpixel);

  1540.     } else {

  1541. #define scale_mv(n, dim) (((int64_t)(n) * scale[dim]) >> 14)

  1542.     int mx, my;

  1543.     int refbw_m1, refbh_m1;

  1544.     int th;

  1545.     VP56mv mv;

  1546. 

  1547.     mv.x = av_clip(in_mv->x, -(x + pw - px + 4) * 8, (s->cols * 8 - x + px + 3) * 8);

  1548.     mv.y = av_clip(in_mv->y, -(y + ph - py + 4) * 8, (s->rows * 8 - y + py + 3) * 8);

  1549.     // BUG libvpx seems to scale the two components separately. This introduces

  1550.     // rounding errors but we have to reproduce them to be exactly compatible

  1551.     // with the output from libvpx...

  1552.     mx = scale_mv(mv.x * 2, 0) + scale_mv(x * 16, 0);

  1553.     my = scale_mv(mv.y * 2, 1) + scale_mv(y * 16, 1);

  1554. 

  1555.     y = my >> 4;

  1556.     x = mx >> 4;

  1557.     ref += y * ref_stride + x * bytesperpixel;

  1558.     mx &= 15;

  1559.     my &= 15;

  1560.     refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;

  1561.     refbh_m1 = ((bh - 1) * step[1] + my) >> 4;

  1562.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  1563.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  1564.     // the longest loopfilter of the next sbrow

  1565.     th = (y + refbh_m1 + 4 + 7) >> 6;

  1566.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  1567.     // The arm/aarch64 _hv filters read one more row than what actually is

  1568.     // needed, so switch to emulated edge one pixel sooner vertically

  1569.     // (y + 5 >= h - refbh_m1) than horizontally (x + 4 >= w - refbw_m1).

  1570.     if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 5 >= h - refbh_m1) {

  1571.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1572.                                  ref - 3 * ref_stride - 3 * bytesperpixel,

  1573.                                  288, ref_stride,

  1574.                                  refbw_m1 + 8, refbh_m1 + 8,

  1575.                                  x - 3, y - 3, w, h);

  1576.         ref = s->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;

  1577.         ref_stride = 288;

  1578.     }

  1579.     smc(dst, dst_stride, ref, ref_stride, bh, mx, my, step[0], step[1]);

  1580.     }

  1581. }

  1582. 

  1583. static av_always_inline void mc_chroma_scaled(VP9Context *s, vp9_scaled_mc_func smc,

  1584.                                               vp9_mc_func (*mc)[2],

  1585.                                               uint8_t *dst_u, uint8_t *dst_v,

  1586.                                               ptrdiff_t dst_stride,

  1587.                                               const uint8_t *ref_u, ptrdiff_t src_stride_u,

  1588.                                               const uint8_t *ref_v, ptrdiff_t src_stride_v,

  1589.                                               ThreadFrame *ref_frame,

  1590.                                               ptrdiff_t y, ptrdiff_t x, const VP56mv *in_mv,

  1591.                                               int px, int py, int pw, int ph,

  1592.                                               int bw, int bh, int w, int h, int bytesperpixel,

  1593.                                               const uint16_t *scale, const uint8_t *step)

  1594. {

  1595.     if (s->s.frames[CUR_FRAME].tf.f->width == ref_frame->f->width &&

  1596.         s->s.frames[CUR_FRAME].tf.f->height == ref_frame->f->height) {

  1597.         mc_chroma_unscaled(s, mc, dst_u, dst_v, dst_stride, ref_u, src_stride_u,

  1598.                            ref_v, src_stride_v, ref_frame,

  1599.                            y, x, in_mv, bw, bh, w, h, bytesperpixel);

  1600.     } else {

  1601.     int mx, my;

  1602.     int refbw_m1, refbh_m1;

  1603.     int th;

  1604.     VP56mv mv;

  1605. 

  1606.     if (s->ss_h) {

  1607.         // BUG https://code.google.com/p/webm/issues/detail?id=820

  1608.         mv.x = av_clip(in_mv->x, -(x + pw - px + 4) * 16, (s->cols * 4 - x + px + 3) * 16);

  1609.         mx = scale_mv(mv.x, 0) + (scale_mv(x * 16, 0) & ~15) + (scale_mv(x * 32, 0) & 15);

  1610.     } else {

  1611.         mv.x = av_clip(in_mv->x, -(x + pw - px + 4) * 8, (s->cols * 8 - x + px + 3) * 8);

  1612.         mx = scale_mv(mv.x * 2, 0) + scale_mv(x * 16, 0);

  1613.     }

  1614.     if (s->ss_v) {

  1615.         // BUG https://code.google.com/p/webm/issues/detail?id=820

  1616.         mv.y = av_clip(in_mv->y, -(y + ph - py + 4) * 16, (s->rows * 4 - y + py + 3) * 16);

  1617.         my = scale_mv(mv.y, 1) + (scale_mv(y * 16, 1) & ~15) + (scale_mv(y * 32, 1) & 15);

  1618.     } else {

  1619.         mv.y = av_clip(in_mv->y, -(y + ph - py + 4) * 8, (s->rows * 8 - y + py + 3) * 8);

  1620.         my = scale_mv(mv.y * 2, 1) + scale_mv(y * 16, 1);

  1621.     }

  1622. #undef scale_mv

  1623.     y = my >> 4;

  1624.     x = mx >> 4;

  1625.     ref_u += y * src_stride_u + x * bytesperpixel;

  1626.     ref_v += y * src_stride_v + x * bytesperpixel;

  1627.     mx &= 15;

  1628.     my &= 15;

  1629.     refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;

  1630.     refbh_m1 = ((bh - 1) * step[1] + my) >> 4;

  1631.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  1632.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  1633.     // the longest loopfilter of the next sbrow

  1634.     th = (y + refbh_m1 + 4 + 7) >> (6 - s->ss_v);

  1635.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  1636.     // The arm/aarch64 _hv filters read one more row than what actually is

  1637.     // needed, so switch to emulated edge one pixel sooner vertically

  1638.     // (y + 5 >= h - refbh_m1) than horizontally (x + 4 >= w - refbw_m1).

  1639.     if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 5 >= h - refbh_m1) {

  1640.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1641.                                  ref_u - 3 * src_stride_u - 3 * bytesperpixel,

  1642.                                  288, src_stride_u,

  1643.                                  refbw_m1 + 8, refbh_m1 + 8,

  1644.                                  x - 3, y - 3, w, h);

  1645.         ref_u = s->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;

  1646.         smc(dst_u, dst_stride, ref_u, 288, bh, mx, my, step[0], step[1]);

  1647. 

  1648.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  1649.                                  ref_v - 3 * src_stride_v - 3 * bytesperpixel,

  1650.                                  288, src_stride_v,

  1651.                                  refbw_m1 + 8, refbh_m1 + 8,

  1652.                                  x - 3, y - 3, w, h);

  1653.         ref_v = s->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;

  1654.         smc(dst_v, dst_stride, ref_v, 288, bh, mx, my, step[0], step[1]);

  1655.     } else {

  1656.         smc(dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my, step[0], step[1]);

  1657.         smc(dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my, step[0], step[1]);

  1658.     }

  1659.     }

  1660. }

  1661. 

  1662. #define mc_luma_dir(s, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, \

  1663.                     px, py, pw, ph, bw, bh, w, h, i) \

  1664.     mc_luma_scaled(s, s->dsp.s##mc, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \

  1665.                    mv, px, py, pw, ph, bw, bh, w, h, bytesperpixel, \

  1666.                    s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])

  1667. #define mc_chroma_dir(s, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  1668.                       row, col, mv, px, py, pw, ph, bw, bh, w, h, i) \

  1669.     mc_chroma_scaled(s, s->dsp.s##mc, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  1670.                      row, col, mv, px, py, pw, ph, bw, bh, w, h, bytesperpixel, \

  1671.                      s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])

  1672. #define SCALED 1

  1673. #define FN(x) x##_scaled_8bpp

  1674. #define BYTES_PER_PIXEL 1

  1675. #include "vp9_mc_template.c"

  1676. #undef FN

  1677. #undef BYTES_PER_PIXEL

  1678. #define FN(x) x##_scaled_16bpp

  1679. #define BYTES_PER_PIXEL 2

  1680. #include "vp9_mc_template.c"

  1681. #undef mc_luma_dir

  1682. #undef mc_chroma_dir

  1683. #undef FN

  1684. #undef BYTES_PER_PIXEL

  1685. #undef SCALED

  1686. 

  1687. static av_always_inline void inter_recon(AVCodecContext *avctx, int bytesperpixel)

  1688. {

  1689.     VP9Context *s = avctx->priv_data;

  1690.     VP9Block *b = s->b;

  1691.     int row = s->row, col = s->col;

  1692. 

  1693.     if (s->mvscale[b->ref[0]][0] || (b->comp && s->mvscale[b->ref[1]][0])) {

  1694.         if (bytesperpixel == 1) {

  1695.             inter_pred_scaled_8bpp(avctx);

  1696.         } else {

  1697.             inter_pred_scaled_16bpp(avctx);

  1698.         }

  1699.     } else {

  1700.         if (bytesperpixel == 1) {

  1701.             inter_pred_8bpp(avctx);

  1702.         } else {

  1703.             inter_pred_16bpp(avctx);

  1704.         }

  1705.     }

  1706. 

  1707.     if (!b->skip) {

  1708.         /* mostly copied intra_recon() */

  1709. 

  1710.         int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  1711.         int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  1712.         int end_x = FFMIN(2 * (s->cols - col), w4);

  1713.         int end_y = FFMIN(2 * (s->rows - row), h4);

  1714.         int tx = 4 * s->s.h.lossless + b->tx, uvtx = b->uvtx + 4 * s->s.h.lossless;

  1715.         int uvstep1d = 1 << b->uvtx, p;

  1716.         uint8_t *dst = s->dst[0];

  1717. 

  1718.         // y itxfm add

  1719.         for (n = 0, y = 0; y < end_y; y += step1d) {

  1720.             uint8_t *ptr = dst;

  1721.             for (x = 0; x < end_x; x += step1d,

  1722.                  ptr += 4 * step1d * bytesperpixel, n += step) {

  1723.                 int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  1724. 

  1725.                 if (eob)

  1726.                     s->dsp.itxfm_add[tx][DCT_DCT](ptr, s->y_stride,

  1727.                                                   s->block + 16 * n * bytesperpixel, eob);

  1728.             }

  1729.             dst += 4 * s->y_stride * step1d;

  1730.         }

  1731. 

  1732.         // uv itxfm add

  1733.         end_x >>= s->ss_h;

  1734.         end_y >>= s->ss_v;

  1735.         step = 1 << (b->uvtx * 2);

  1736.         for (p = 0; p < 2; p++) {

  1737.             dst = s->dst[p + 1];

  1738.             for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  1739.                 uint8_t *ptr = dst;

  1740.                 for (x = 0; x < end_x; x += uvstep1d,

  1741.                      ptr += 4 * uvstep1d * bytesperpixel, n += step) {

  1742.                     int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  1743. 

  1744.                     if (eob)

  1745.                         s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  1746.                                                         s->uvblock[p] + 16 * n * bytesperpixel, eob);

  1747.                 }

  1748.                 dst += 4 * uvstep1d * s->uv_stride;

  1749.             }

  1750.         }

  1751.     }

  1752. }

  1753. 

  1754. static void inter_recon_8bpp(AVCodecContext *avctx)

  1755. {

  1756.     inter_recon(avctx, 1);

  1757. }

  1758. 

  1759. static void inter_recon_16bpp(AVCodecContext *avctx)

  1760. {

  1761.     inter_recon(avctx, 2);

  1762. }

  1763. 

  1764. static av_always_inline void mask_edges(uint8_t (*mask)[8][4], int ss_h, int ss_v,

  1765.                                         int row_and_7, int col_and_7,

  1766.                                         int w, int h, int col_end, int row_end,

  1767.                                         enum TxfmMode tx, int skip_inter)

  1768. {

  1769.     static const unsigned wide_filter_col_mask[2] = { 0x11, 0x01 };

  1770.     static const unsigned wide_filter_row_mask[2] = { 0x03, 0x07 };

  1771. 

  1772.     // FIXME I'm pretty sure all loops can be replaced by a single LUT if

  1773.     // we make VP9Filter.mask uint64_t (i.e. row/col all single variable)

  1774.     // and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then

  1775.     // use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)

  1776. 

  1777.     // the intended behaviour of the vp9 loopfilter is to work on 8-pixel

  1778.     // edges. This means that for UV, we work on two subsampled blocks at

  1779.     // a time, and we only use the topleft block's mode information to set

  1780.     // things like block strength. Thus, for any block size smaller than

  1781.     // 16x16, ignore the odd portion of the block.

  1782.     if (tx == TX_4X4 && (ss_v | ss_h)) {

  1783.         if (h == ss_v) {

  1784.             if (row_and_7 & 1)

  1785.                 return;

  1786.             if (!row_end)

  1787.                 h += 1;

  1788.         }

  1789.         if (w == ss_h) {

  1790.             if (col_and_7 & 1)

  1791.                 return;

  1792.             if (!col_end)

  1793.                 w += 1;

  1794.         }

  1795.     }

  1796. 

  1797.     if (tx == TX_4X4 && !skip_inter) {

  1798.         int t = 1 << col_and_7, m_col = (t << w) - t, y;

  1799.         // on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide

  1800.         int m_row_8 = m_col & wide_filter_col_mask[ss_h], m_row_4 = m_col - m_row_8;

  1801. 

  1802.         for (y = row_and_7; y < h + row_and_7; y++) {

  1803.             int col_mask_id = 2 - !(y & wide_filter_row_mask[ss_v]);

  1804. 

  1805.             mask[0][y][1] |= m_row_8;

  1806.             mask[0][y][2] |= m_row_4;

  1807.             // for odd lines, if the odd col is not being filtered,

  1808.             // skip odd row also:

  1809.             // .---. <-- a

  1810.             // |   |

  1811.             // |___| <-- b

  1812.             // ^   ^

  1813.             // c   d

  1814.             //

  1815.             // if a/c are even row/col and b/d are odd, and d is skipped,

  1816.             // e.g. right edge of size-66x66.webm, then skip b also (bug)

  1817.             if ((ss_h & ss_v) && (col_end & 1) && (y & 1)) {

  1818.                 mask[1][y][col_mask_id] |= (t << (w - 1)) - t;

  1819.             } else {

  1820.                 mask[1][y][col_mask_id] |= m_col;

  1821.             }

  1822.             if (!ss_h)

  1823.                 mask[0][y][3] |= m_col;

  1824.             if (!ss_v) {

  1825.                 if (ss_h && (col_end & 1))

  1826.                     mask[1][y][3] |= (t << (w - 1)) - t;

  1827.                 else

  1828.                     mask[1][y][3] |= m_col;

  1829.             }

  1830.         }

  1831.     } else {

  1832.         int y, t = 1 << col_and_7, m_col = (t << w) - t;

  1833. 

  1834.         if (!skip_inter) {

  1835.             int mask_id = (tx == TX_8X8);

  1836.             int l2 = tx + ss_h - 1, step1d;

  1837.             static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };

  1838.             int m_row = m_col & masks[l2];

  1839. 

  1840.             // at odd UV col/row edges tx16/tx32 loopfilter edges, force

  1841.             // 8wd loopfilter to prevent going off the visible edge.

  1842.             if (ss_h && tx > TX_8X8 && (w ^ (w - 1)) == 1) {

  1843.                 int m_row_16 = ((t << (w - 1)) - t) & masks[l2];

  1844.                 int m_row_8 = m_row - m_row_16;

  1845. 

  1846.                 for (y = row_and_7; y < h + row_and_7; y++) {

  1847.                     mask[0][y][0] |= m_row_16;

  1848.                     mask[0][y][1] |= m_row_8;

  1849.                 }

  1850.             } else {

  1851.                 for (y = row_and_7; y < h + row_and_7; y++)

  1852.                     mask[0][y][mask_id] |= m_row;

  1853.             }

  1854. 

  1855.             l2 = tx + ss_v - 1;

  1856.             step1d = 1 << l2;

  1857.             if (ss_v && tx > TX_8X8 && (h ^ (h - 1)) == 1) {

  1858.                 for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)

  1859.                     mask[1][y][0] |= m_col;

  1860.                 if (y - row_and_7 == h - 1)

  1861.                     mask[1][y][1] |= m_col;

  1862.             } else {

  1863.                 for (y = row_and_7; y < h + row_and_7; y += step1d)

  1864.                     mask[1][y][mask_id] |= m_col;

  1865.             }

  1866.         } else if (tx != TX_4X4) {

  1867.             int mask_id;

  1868. 

  1869.             mask_id = (tx == TX_8X8) || (h == ss_v);

  1870.             mask[1][row_and_7][mask_id] |= m_col;

  1871.             mask_id = (tx == TX_8X8) || (w == ss_h);

  1872.             for (y = row_and_7; y < h + row_and_7; y++)

  1873.                 mask[0][y][mask_id] |= t;

  1874.         } else {

  1875.             int t8 = t & wide_filter_col_mask[ss_h], t4 = t - t8;

  1876. 

  1877.             for (y = row_and_7; y < h + row_and_7; y++) {

  1878.                 mask[0][y][2] |= t4;

  1879.                 mask[0][y][1] |= t8;

  1880.             }

  1881.             mask[1][row_and_7][2 - !(row_and_7 & wide_filter_row_mask[ss_v])] |= m_col;

  1882.         }

  1883.     }

  1884. }

  1885. 

  1886. void ff_vp9_decode_block(AVCodecContext *avctx, int row, int col,

  1887.                          VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,

  1888.                          enum BlockLevel bl, enum BlockPartition bp)

  1889. {

  1890.     VP9Context *s = avctx->priv_data;

  1891.     VP9Block *b = s->b;

  1892.     enum BlockSize bs = bl * 3 + bp;

  1893.     int bytesperpixel = s->bytesperpixel;

  1894.     int w4 = bwh_tab[1][bs][0], h4 = bwh_tab[1][bs][1], lvl;

  1895.     int emu[2];

  1896.     AVFrame *f = s->s.frames[CUR_FRAME].tf.f;

  1897. 

  1898.     s->row = row;

  1899.     s->row7 = row & 7;

  1900.     s->col = col;

  1901.     s->col7 = col & 7;

  1902. 

  1903.     s->min_mv.x = -(128 + col * 64);

  1904.     s->min_mv.y = -(128 + row * 64);

  1905.     s->max_mv.x = 128 + (s->cols - col - w4) * 64;

  1906.     s->max_mv.y = 128 + (s->rows - row - h4) * 64;

  1907. 

  1908.     if (s->pass < 2) {

  1909.         b->bs = bs;

  1910.         b->bl = bl;

  1911.         b->bp = bp;

  1912.         decode_mode(avctx);

  1913.         b->uvtx = b->tx - ((s->ss_h && w4 * 2 == (1 << b->tx)) ||

  1914.                            (s->ss_v && h4 * 2 == (1 << b->tx)));

  1915. 

  1916.         if (!b->skip) {

  1917.             int has_coeffs;

  1918. 

  1919.             if (bytesperpixel == 1) {

  1920.                 has_coeffs = decode_coeffs_8bpp(avctx);

  1921.             } else {

  1922.                 has_coeffs = decode_coeffs_16bpp(avctx);

  1923.             }

  1924.             if (!has_coeffs && b->bs <= BS_8x8 && !b->intra) {

  1925.                 b->skip = 1;

  1926.                 memset(&s->above_skip_ctx[col], 1, w4);

  1927.                 memset(&s->left_skip_ctx[s->row7], 1, h4);

  1928.             }

  1929.         } else {

  1930.             int row7 = s->row7;

  1931. 

  1932. #define SPLAT_ZERO_CTX(v, n) \

  1933.     switch (n) { \

  1934.     case 1:  v = 0;          break; \

  1935.     case 2:  AV_ZERO16(&v);  break; \

  1936.     case 4:  AV_ZERO32(&v);  break; \

  1937.     case 8:  AV_ZERO64(&v);  break; \

  1938.     case 16: AV_ZERO128(&v); break; \

  1939.     }

  1940. #define SPLAT_ZERO_YUV(dir, var, off, n, dir2) \

  1941.     do { \

  1942.         SPLAT_ZERO_CTX(s->dir##_y_##var[off * 2], n * 2); \

  1943.         if (s->ss_##dir2) { \

  1944.             SPLAT_ZERO_CTX(s->dir##_uv_##var[0][off], n); \

  1945.             SPLAT_ZERO_CTX(s->dir##_uv_##var[1][off], n); \

  1946.         } else { \

  1947.             SPLAT_ZERO_CTX(s->dir##_uv_##var[0][off * 2], n * 2); \

  1948.             SPLAT_ZERO_CTX(s->dir##_uv_##var[1][off * 2], n * 2); \

  1949.         } \

  1950.     } while (0)

  1951. 

  1952.             switch (w4) {

  1953.             case 1: SPLAT_ZERO_YUV(above, nnz_ctx, col, 1, h); break;

  1954.             case 2: SPLAT_ZERO_YUV(above, nnz_ctx, col, 2, h); break;

  1955.             case 4: SPLAT_ZERO_YUV(above, nnz_ctx, col, 4, h); break;

  1956.             case 8: SPLAT_ZERO_YUV(above, nnz_ctx, col, 8, h); break;

  1957.             }

  1958.             switch (h4) {

  1959.             case 1: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 1, v); break;

  1960.             case 2: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 2, v); break;

  1961.             case 4: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 4, v); break;

  1962.             case 8: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 8, v); break;

  1963.             }

  1964.         }

  1965. 

  1966.         if (s->pass == 1) {

  1967.             s->b++;

  1968.             s->block += w4 * h4 * 64 * bytesperpixel;

  1969.             s->uvblock[0] += w4 * h4 * 64 * bytesperpixel >> (s->ss_h + s->ss_v);

  1970.             s->uvblock[1] += w4 * h4 * 64 * bytesperpixel >> (s->ss_h + s->ss_v);

  1971.             s->eob += 4 * w4 * h4;

  1972.             s->uveob[0] += 4 * w4 * h4 >> (s->ss_h + s->ss_v);

  1973.             s->uveob[1] += 4 * w4 * h4 >> (s->ss_h + s->ss_v);

  1974. 

  1975.             return;

  1976.         }

  1977.     }

  1978. 

  1979.     // emulated overhangs if the stride of the target buffer can't hold. This

  1980.     // makes it possible to support emu-edge and so on even if we have large block

  1981.     // overhangs

  1982.     emu[0] = (col + w4) * 8 * bytesperpixel > f->linesize[0] ||

  1983.              (row + h4) > s->rows;

  1984.     emu[1] = ((col + w4) * 8 >> s->ss_h) * bytesperpixel > f->linesize[1] ||

  1985.              (row + h4) > s->rows;

  1986.     if (emu[0]) {

  1987.         s->dst[0] = s->tmp_y;

  1988.         s->y_stride = 128;

  1989.     } else {

  1990.         s->dst[0] = f->data[0] + yoff;

  1991.         s->y_stride = f->linesize[0];

  1992.     }

  1993.     if (emu[1]) {

  1994.         s->dst[1] = s->tmp_uv[0];

  1995.         s->dst[2] = s->tmp_uv[1];

  1996.         s->uv_stride = 128;

  1997.     } else {

  1998.         s->dst[1] = f->data[1] + uvoff;

  1999.         s->dst[2] = f->data[2] + uvoff;

  2000.         s->uv_stride = f->linesize[1];

  2001.     }

  2002.     if (b->intra) {

  2003.         if (s->s.h.bpp > 8) {

  2004.             intra_recon_16bpp(avctx, yoff, uvoff);

  2005.         } else {

  2006.             intra_recon_8bpp(avctx, yoff, uvoff);

  2007.         }

  2008.     } else {

  2009.         if (s->s.h.bpp > 8) {

  2010.             inter_recon_16bpp(avctx);

  2011.         } else {

  2012.             inter_recon_8bpp(avctx);

  2013.         }

  2014.     }

  2015.     if (emu[0]) {

  2016.         int w = FFMIN(s->cols - col, w4) * 8, h = FFMIN(s->rows - row, h4) * 8, n, o = 0;

  2017. 

  2018.         for (n = 0; o < w; n++) {

  2019.             int bw = 64 >> n;

  2020. 

  2021.             av_assert2(n <= 4);

  2022.             if (w & bw) {

  2023.                 s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o * bytesperpixel, f->linesize[0],

  2024.                                          s->tmp_y + o * bytesperpixel, 128, h, 0, 0);

  2025.                 o += bw;

  2026.             }

  2027.         }

  2028.     }

  2029.     if (emu[1]) {

  2030.         int w = FFMIN(s->cols - col, w4) * 8 >> s->ss_h;

  2031.         int h = FFMIN(s->rows - row, h4) * 8 >> s->ss_v, n, o = 0;

  2032. 

  2033.         for (n = s->ss_h; o < w; n++) {

  2034.             int bw = 64 >> n;

  2035. 

  2036.             av_assert2(n <= 4);

  2037.             if (w & bw) {

  2038.                 s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o * bytesperpixel, f->linesize[1],

  2039.                                          s->tmp_uv[0] + o * bytesperpixel, 128, h, 0, 0);

  2040.                 s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o * bytesperpixel, f->linesize[2],

  2041.                                          s->tmp_uv[1] + o * bytesperpixel, 128, h, 0, 0);

  2042.                 o += bw;

  2043.             }

  2044.         }

  2045.     }

  2046. 

  2047.     // pick filter level and find edges to apply filter to

  2048.     if (s->s.h.filter.level &&

  2049.         (lvl = s->s.h.segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]

  2050.                                                       [b->mode[3] != ZEROMV]) > 0) {

  2051.         int x_end = FFMIN(s->cols - col, w4), y_end = FFMIN(s->rows - row, h4);

  2052.         int skip_inter = !b->intra && b->skip, col7 = s->col7, row7 = s->row7;

  2053. 

  2054.         setctx_2d(&lflvl->level[row7 * 8 + col7], w4, h4, 8, lvl);

  2055.         mask_edges(lflvl->mask[0], 0, 0, row7, col7, x_end, y_end, 0, 0, b->tx, skip_inter);

  2056.         if (s->ss_h || s->ss_v)

  2057.             mask_edges(lflvl->mask[1], s->ss_h, s->ss_v, row7, col7, x_end, y_end,

  2058.                        s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,

  2059.                        s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,

  2060.                        b->uvtx, skip_inter);

  2061. 

  2062.         if (!s->filter_lut.lim_lut[lvl]) {

  2063.             int sharp = s->s.h.filter.sharpness;

  2064.             int limit = lvl;

  2065. 

  2066.             if (sharp > 0) {

  2067.                 limit >>= (sharp + 3) >> 2;

  2068.                 limit = FFMIN(limit, 9 - sharp);

  2069.             }

  2070.             limit = FFMAX(limit, 1);

  2071. 

  2072.             s->filter_lut.lim_lut[lvl] = limit;

  2073.             s->filter_lut.mblim_lut[lvl] = 2 * (lvl + 2) + limit;

  2074.         }

  2075.     }

  2076. 

  2077.     if (s->pass == 2) {

  2078.         s->b++;

  2079.         s->block += w4 * h4 * 64 * bytesperpixel;

  2080.         s->uvblock[0] += w4 * h4 * 64 * bytesperpixel >> (s->ss_v + s->ss_h);

  2081.         s->uvblock[1] += w4 * h4 * 64 * bytesperpixel >> (s->ss_v + s->ss_h);

  2082.         s->eob += 4 * w4 * h4;

  2083.         s->uveob[0] += 4 * w4 * h4 >> (s->ss_v + s->ss_h);

  2084.         s->uveob[1] += 4 * w4 * h4 >> (s->ss_v + s->ss_h);

  2085.     }

  2086. }