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.
58679 Commits
32 Branches
283MB
Tree: fc7d910b2e
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'fc7d910b2e'
${ noResults }
FFmpeg/libavcodec/vp9.c

3617 lines
144KB
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 "avcodec.h"

  25. #include "get_bits.h"

  26. #include "internal.h"

  27. #include "thread.h"

  28. #include "videodsp.h"

  29. #include "vp56.h"

  30. #include "vp9.h"

  31. #include "vp9data.h"

  32. #include "vp9dsp.h"

  33. #include "libavutil/avassert.h"

  34. 

  35. #define VP9_SYNCCODE 0x498342

  36. 

  37. enum CompPredMode {

  38.     PRED_SINGLEREF,

  39.     PRED_COMPREF,

  40.     PRED_SWITCHABLE,

  41. };

  42. 

  43. enum BlockLevel {

  44.     BL_64X64,

  45.     BL_32X32,

  46.     BL_16X16,

  47.     BL_8X8,

  48. };

  49. 

  50. enum BlockSize {

  51.     BS_64x64,

  52.     BS_64x32,

  53.     BS_32x64,

  54.     BS_32x32,

  55.     BS_32x16,

  56.     BS_16x32,

  57.     BS_16x16,

  58.     BS_16x8,

  59.     BS_8x16,

  60.     BS_8x8,

  61.     BS_8x4,

  62.     BS_4x8,

  63.     BS_4x4,

  64.     N_BS_SIZES,

  65. };

  66. 

  67. struct VP9mvrefPair {

  68.     VP56mv mv[2];

  69.     int8_t ref[2];

  70. };

  71. 

  72. typedef struct VP9Frame {

  73.     ThreadFrame tf;

  74.     AVBufferRef *extradata;

  75.     uint8_t *segmentation_map;

  76.     struct VP9mvrefPair *mv;

  77. } VP9Frame;

  78. 

  79. struct VP9Filter {

  80.     uint8_t level[8 * 8];

  81.     uint8_t /* bit=col */ mask[2 /* 0=y, 1=uv */][2 /* 0=col, 1=row */]

  82.                               [8 /* rows */][4 /* 0=16, 1=8, 2=4, 3=inner4 */];

  83. };

  84. 

  85. typedef struct VP9Block {

  86.     uint8_t seg_id, intra, comp, ref[2], mode[4], uvmode, skip;

  87.     enum FilterMode filter;

  88.     VP56mv mv[4 /* b_idx */][2 /* ref */];

  89.     enum BlockSize bs;

  90.     enum TxfmMode tx, uvtx;

  91. 

  92. } VP9Block;

  93. 

  94. typedef struct VP9Context {

  95.     VP9DSPContext dsp;

  96.     VideoDSPContext vdsp;

  97.     GetBitContext gb;

  98.     VP56RangeCoder c;

  99.     VP56RangeCoder *c_b;

  100.     unsigned c_b_size;

  101.     VP9Block b;

  102.     int row, row7, col, col7;

  103.     uint8_t *dst[3];

  104.     ptrdiff_t y_stride, uv_stride;

  105. 

  106.     // bitstream header

  107.     uint8_t profile;

  108.     uint8_t keyframe, last_keyframe;

  109.     uint8_t invisible;

  110.     uint8_t use_last_frame_mvs;

  111.     uint8_t errorres;

  112.     uint8_t colorspace;

  113.     uint8_t fullrange;

  114.     uint8_t intraonly;

  115.     uint8_t resetctx;

  116.     uint8_t refreshrefmask;

  117.     uint8_t highprecisionmvs;

  118.     enum FilterMode filtermode;

  119.     uint8_t allowcompinter;

  120.     uint8_t fixcompref;

  121.     uint8_t refreshctx;

  122.     uint8_t parallelmode;

  123.     uint8_t framectxid;

  124.     uint8_t refidx[3];

  125.     uint8_t signbias[3];

  126.     uint8_t varcompref[2];

  127.     ThreadFrame refs[8], next_refs[8];

  128. #define CUR_FRAME 0

  129. #define LAST_FRAME 1

  130.     VP9Frame frames[2];

  131. 

  132.     struct {

  133.         uint8_t level;

  134.         int8_t sharpness;

  135.         uint8_t lim_lut[64];

  136.         uint8_t mblim_lut[64];

  137.     } filter;

  138.     struct {

  139.         uint8_t enabled;

  140.         int8_t mode[2];

  141.         int8_t ref[4];

  142.     } lf_delta;

  143.     uint8_t yac_qi;

  144.     int8_t ydc_qdelta, uvdc_qdelta, uvac_qdelta;

  145.     uint8_t lossless;

  146.     struct {

  147.         uint8_t enabled;

  148.         uint8_t temporal;

  149.         uint8_t absolute_vals;

  150.         uint8_t update_map;

  151.         struct {

  152.             uint8_t q_enabled;

  153.             uint8_t lf_enabled;

  154.             uint8_t ref_enabled;

  155.             uint8_t skip_enabled;

  156.             uint8_t ref_val;

  157.             int16_t q_val;

  158.             int8_t lf_val;

  159.             int16_t qmul[2][2];

  160.             uint8_t lflvl[4][2];

  161.         } feat[8];

  162.     } segmentation;

  163.     struct {

  164.         unsigned log2_tile_cols, log2_tile_rows;

  165.         unsigned tile_cols, tile_rows;

  166.         unsigned tile_row_start, tile_row_end, tile_col_start, tile_col_end;

  167.     } tiling;

  168.     unsigned sb_cols, sb_rows, rows, cols;

  169.     struct {

  170.         prob_context p;

  171.         uint8_t coef[4][2][2][6][6][3];

  172.     } prob_ctx[4];

  173.     struct {

  174.         prob_context p;

  175.         uint8_t coef[4][2][2][6][6][11];

  176.         uint8_t seg[7];

  177.         uint8_t segpred[3];

  178.     } prob;

  179.     struct {

  180.         unsigned y_mode[4][10];

  181.         unsigned uv_mode[10][10];

  182.         unsigned filter[4][3];

  183.         unsigned mv_mode[7][4];

  184.         unsigned intra[4][2];

  185.         unsigned comp[5][2];

  186.         unsigned single_ref[5][2][2];

  187.         unsigned comp_ref[5][2];

  188.         unsigned tx32p[2][4];

  189.         unsigned tx16p[2][3];

  190.         unsigned tx8p[2][2];

  191.         unsigned skip[3][2];

  192.         unsigned mv_joint[4];

  193.         struct {

  194.             unsigned sign[2];

  195.             unsigned classes[11];

  196.             unsigned class0[2];

  197.             unsigned bits[10][2];

  198.             unsigned class0_fp[2][4];

  199.             unsigned fp[4];

  200.             unsigned class0_hp[2];

  201.             unsigned hp[2];

  202.         } mv_comp[2];

  203.         unsigned partition[4][4][4];

  204.         unsigned coef[4][2][2][6][6][3];

  205.         unsigned eob[4][2][2][6][6][2];

  206.     } counts;

  207.     enum TxfmMode txfmmode;

  208.     enum CompPredMode comppredmode;

  209. 

  210.     // contextual (left/above) cache

  211.     uint8_t left_partition_ctx[8], *above_partition_ctx;

  212.     uint8_t left_mode_ctx[16], *above_mode_ctx;

  213.     // FIXME maybe merge some of the below in a flags field?

  214.     uint8_t left_y_nnz_ctx[16], *above_y_nnz_ctx;

  215.     uint8_t left_uv_nnz_ctx[2][8], *above_uv_nnz_ctx[2];

  216.     uint8_t left_skip_ctx[8], *above_skip_ctx; // 1bit

  217.     uint8_t left_txfm_ctx[8], *above_txfm_ctx; // 2bit

  218.     uint8_t left_segpred_ctx[8], *above_segpred_ctx; // 1bit

  219.     uint8_t left_intra_ctx[8], *above_intra_ctx; // 1bit

  220.     uint8_t left_comp_ctx[8], *above_comp_ctx; // 1bit

  221.     uint8_t left_ref_ctx[8], *above_ref_ctx; // 2bit

  222.     uint8_t left_filter_ctx[8], *above_filter_ctx;

  223.     VP56mv left_mv_ctx[16][2], (*above_mv_ctx)[2];

  224. 

  225.     // whole-frame cache

  226.     uint8_t *intra_pred_data[3];

  227.     struct VP9Filter *lflvl;

  228.     DECLARE_ALIGNED(32, uint8_t, edge_emu_buffer)[71*80];

  229. 

  230.     // block reconstruction intermediates

  231.     DECLARE_ALIGNED(32, int16_t, block)[4096];

  232.     DECLARE_ALIGNED(32, int16_t, uvblock)[2][1024];

  233.     uint8_t eob[256];

  234.     uint8_t uveob[2][64];

  235.     VP56mv min_mv, max_mv;

  236.     DECLARE_ALIGNED(32, uint8_t, tmp_y)[64*64];

  237.     DECLARE_ALIGNED(32, uint8_t, tmp_uv)[2][32*32];

  238. } VP9Context;

  239. 

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

  241.     {

  242.         { 16, 16 }, { 16, 8 }, { 8, 16 }, { 8, 8 }, { 8, 4 }, { 4, 8 },

  243.         { 4, 4 }, { 4, 2 }, { 2, 4 }, { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 },

  244.     }, {

  245.         { 8, 8 }, { 8, 4 }, { 4, 8 }, { 4, 4 }, { 4, 2 }, { 2, 4 },

  246.         { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 }, { 1, 1 }, { 1, 1 }, { 1, 1 },

  247.     }

  248. };

  249. 

  250. static int vp9_alloc_frame(AVCodecContext *ctx, VP9Frame *f)

  251. {

  252.     VP9Context *s = ctx->priv_data;

  253.     int ret, sz;

  254. 

  255.     if ((ret = ff_thread_get_buffer(ctx, &f->tf, AV_GET_BUFFER_FLAG_REF)) < 0)

  256.         return ret;

  257.     sz = 64 * s->sb_cols * s->sb_rows;

  258.     if (!(f->extradata = av_buffer_allocz(sz * (1 + sizeof(struct VP9mvrefPair))))) {

  259.         ff_thread_release_buffer(ctx, &f->tf);

  260.         return AVERROR(ENOMEM);

  261.     }

  262. 

  263.     f->segmentation_map = f->extradata->data;

  264.     f->mv = (struct VP9mvrefPair *) (f->extradata->data + sz);

  265. 

  266.     return 0;

  267. }

  268. 

  269. static void vp9_unref_frame(AVCodecContext *ctx, VP9Frame *f)

  270. {

  271.     ff_thread_release_buffer(ctx, &f->tf);

  272.     av_buffer_unref(&f->extradata);

  273. }

  274. 

  275. static int vp9_ref_frame(AVCodecContext *ctx, VP9Frame *dst, VP9Frame *src)

  276. {

  277.     int res;

  278. 

  279.     if ((res = ff_thread_ref_frame(&dst->tf, &src->tf)) < 0) {

  280.         return res;

  281.     } else if (!(dst->extradata = av_buffer_ref(src->extradata))) {

  282.         vp9_unref_frame(ctx, dst);

  283.         return AVERROR(ENOMEM);

  284.     }

  285. 

  286.     dst->segmentation_map = src->segmentation_map;

  287.     dst->mv = src->mv;

  288. 

  289.     return 0;

  290. }

  291. 

  292. static int update_size(AVCodecContext *ctx, int w, int h)

  293. {

  294.     VP9Context *s = ctx->priv_data;

  295.     uint8_t *p;

  296. 

  297.     av_assert0(w > 0 && h > 0);

  298. 

  299.     if (s->above_partition_ctx && w == ctx->width && h == ctx->height)

  300.         return 0;

  301. 

  302.     ctx->width  = w;

  303.     ctx->height = h;

  304.     s->sb_cols  = (w + 63) >> 6;

  305.     s->sb_rows  = (h + 63) >> 6;

  306.     s->cols     = (w + 7) >> 3;

  307.     s->rows     = (h + 7) >> 3;

  308. 

  309. #define assign(var, type, n) var = (type) p; p += s->sb_cols * n * sizeof(*var)

  310.     av_freep(&s->above_partition_ctx);

  311.     p = av_malloc(s->sb_cols * (240 + sizeof(*s->lflvl) + 16 * sizeof(*s->above_mv_ctx)));

  312.     if (!p)

  313.         return AVERROR(ENOMEM);

  314.     assign(s->above_partition_ctx, uint8_t *,              8);

  315.     assign(s->above_skip_ctx,      uint8_t *,              8);

  316.     assign(s->above_txfm_ctx,      uint8_t *,              8);

  317.     assign(s->above_mode_ctx,      uint8_t *,             16);

  318.     assign(s->above_y_nnz_ctx,     uint8_t *,             16);

  319.     assign(s->above_uv_nnz_ctx[0], uint8_t *,              8);

  320.     assign(s->above_uv_nnz_ctx[1], uint8_t *,              8);

  321.     assign(s->intra_pred_data[0],  uint8_t *,             64);

  322.     assign(s->intra_pred_data[1],  uint8_t *,             32);

  323.     assign(s->intra_pred_data[2],  uint8_t *,             32);

  324.     assign(s->above_segpred_ctx,   uint8_t *,              8);

  325.     assign(s->above_intra_ctx,     uint8_t *,              8);

  326.     assign(s->above_comp_ctx,      uint8_t *,              8);

  327.     assign(s->above_ref_ctx,       uint8_t *,              8);

  328.     assign(s->above_filter_ctx,    uint8_t *,              8);

  329.     assign(s->lflvl,               struct VP9Filter *,     1);

  330.     assign(s->above_mv_ctx,        VP56mv(*)[2],          16);

  331. #undef assign

  332. 

  333.     return 0;

  334. }

  335. 

  336. // for some reason the sign bit is at the end, not the start, of a bit sequence

  337. static av_always_inline int get_sbits_inv(GetBitContext *gb, int n)

  338. {

  339.     int v = get_bits(gb, n);

  340.     return get_bits1(gb) ? -v : v;

  341. }

  342. 

  343. static av_always_inline int inv_recenter_nonneg(int v, int m)

  344. {

  345.     return v > 2 * m ? v : v & 1 ? m - ((v + 1) >> 1) : m + (v >> 1);

  346. }

  347. 

  348. // differential forward probability updates

  349. static int update_prob(VP56RangeCoder *c, int p)

  350. {

  351.     static const int inv_map_table[254] = {

  352.           7,  20,  33,  46,  59,  72,  85,  98, 111, 124, 137, 150, 163, 176,

  353.         189, 202, 215, 228, 241, 254,   1,   2,   3,   4,   5,   6,   8,   9,

  354.          10,  11,  12,  13,  14,  15,  16,  17,  18,  19,  21,  22,  23,  24,

  355.          25,  26,  27,  28,  29,  30,  31,  32,  34,  35,  36,  37,  38,  39,

  356.          40,  41,  42,  43,  44,  45,  47,  48,  49,  50,  51,  52,  53,  54,

  357.          55,  56,  57,  58,  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,

  358.          70,  71,  73,  74,  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,

  359.          86,  87,  88,  89,  90,  91,  92,  93,  94,  95,  96,  97,  99, 100,

  360.         101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 112, 113, 114, 115,

  361.         116, 117, 118, 119, 120, 121, 122, 123, 125, 126, 127, 128, 129, 130,

  362.         131, 132, 133, 134, 135, 136, 138, 139, 140, 141, 142, 143, 144, 145,

  363.         146, 147, 148, 149, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160,

  364.         161, 162, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175,

  365.         177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 190, 191,

  366.         192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 204, 205, 206,

  367.         207, 208, 209, 210, 211, 212, 213, 214, 216, 217, 218, 219, 220, 221,

  368.         222, 223, 224, 225, 226, 227, 229, 230, 231, 232, 233, 234, 235, 236,

  369.         237, 238, 239, 240, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251,

  370.         252, 253,

  371.     };

  372.     int d;

  373. 

  374.     /* This code is trying to do a differential probability update. For a

  375.      * current probability A in the range [1, 255], the difference to a new

  376.      * probability of any value can be expressed differentially as 1-A,255-A

  377.      * where some part of this (absolute range) exists both in positive as

  378.      * well as the negative part, whereas another part only exists in one

  379.      * half. We're trying to code this shared part differentially, i.e.

  380.      * times two where the value of the lowest bit specifies the sign, and

  381.      * the single part is then coded on top of this. This absolute difference

  382.      * then again has a value of [0,254], but a bigger value in this range

  383.      * indicates that we're further away from the original value A, so we

  384.      * can code this as a VLC code, since higher values are increasingly

  385.      * unlikely. The first 20 values in inv_map_table[] allow 'cheap, rough'

  386.      * updates vs. the 'fine, exact' updates further down the range, which

  387.      * adds one extra dimension to this differential update model. */

  388. 

  389.     if (!vp8_rac_get(c)) {

  390.         d = vp8_rac_get_uint(c, 4) + 0;

  391.     } else if (!vp8_rac_get(c)) {

  392.         d = vp8_rac_get_uint(c, 4) + 16;

  393.     } else if (!vp8_rac_get(c)) {

  394.         d = vp8_rac_get_uint(c, 5) + 32;

  395.     } else {

  396.         d = vp8_rac_get_uint(c, 7);

  397.         if (d >= 65)

  398.             d = (d << 1) - 65 + vp8_rac_get(c);

  399.         d += 64;

  400.     }

  401. 

  402.     return p <= 128 ? 1 + inv_recenter_nonneg(inv_map_table[d], p - 1) :

  403.                     255 - inv_recenter_nonneg(inv_map_table[d], 255 - p);

  404. }

  405. 

  406. static int decode_frame_header(AVCodecContext *ctx,

  407.                                const uint8_t *data, int size, int *ref)

  408. {

  409.     VP9Context *s = ctx->priv_data;

  410.     int c, i, j, k, l, m, n, w, h, max, size2, res, sharp;

  411.     int last_invisible;

  412.     const uint8_t *data2;

  413. 

  414.     /* general header */

  415.     if ((res = init_get_bits8(&s->gb, data, size)) < 0) {

  416.         av_log(ctx, AV_LOG_ERROR, "Failed to initialize bitstream reader\n");

  417.         return res;

  418.     }

  419.     if (get_bits(&s->gb, 2) != 0x2) { // frame marker

  420.         av_log(ctx, AV_LOG_ERROR, "Invalid frame marker\n");

  421.         return AVERROR_INVALIDDATA;

  422.     }

  423.     s->profile = get_bits1(&s->gb);

  424.     if (get_bits1(&s->gb)) { // reserved bit

  425.         av_log(ctx, AV_LOG_ERROR, "Reserved bit should be zero\n");

  426.         return AVERROR_INVALIDDATA;

  427.     }

  428.     if (get_bits1(&s->gb)) {

  429.         *ref = get_bits(&s->gb, 3);

  430.         return 0;

  431.     }

  432.     s->last_keyframe  = s->keyframe;

  433.     s->keyframe       = !get_bits1(&s->gb);

  434.     last_invisible    = s->invisible;

  435.     s->invisible      = !get_bits1(&s->gb);

  436.     s->errorres       = get_bits1(&s->gb);

  437.     // FIXME disable this upon resolution change

  438.     s->use_last_frame_mvs = !s->errorres && !last_invisible;

  439.     if (s->keyframe) {

  440.         if (get_bits_long(&s->gb, 24) != VP9_SYNCCODE) { // synccode

  441.             av_log(ctx, AV_LOG_ERROR, "Invalid sync code\n");

  442.             return AVERROR_INVALIDDATA;

  443.         }

  444.         s->colorspace = get_bits(&s->gb, 3);

  445.         if (s->colorspace == 7) { // RGB = profile 1

  446.             av_log(ctx, AV_LOG_ERROR, "RGB not supported in profile 0\n");

  447.             return AVERROR_INVALIDDATA;

  448.         }

  449.         s->fullrange  = get_bits1(&s->gb);

  450.         // for profile 1, here follows the subsampling bits

  451.         s->refreshrefmask = 0xff;

  452.         w = get_bits(&s->gb, 16) + 1;

  453.         h = get_bits(&s->gb, 16) + 1;

  454.         if (get_bits1(&s->gb)) // display size

  455.             skip_bits(&s->gb, 32);

  456.     } else {

  457.         s->intraonly  = s->invisible ? get_bits1(&s->gb) : 0;

  458.         s->resetctx   = s->errorres ? 0 : get_bits(&s->gb, 2);

  459.         if (s->intraonly) {

  460.             if (get_bits_long(&s->gb, 24) != VP9_SYNCCODE) { // synccode

  461.                 av_log(ctx, AV_LOG_ERROR, "Invalid sync code\n");

  462.                 return AVERROR_INVALIDDATA;

  463.             }

  464.             s->refreshrefmask = get_bits(&s->gb, 8);

  465.             w = get_bits(&s->gb, 16) + 1;

  466.             h = get_bits(&s->gb, 16) + 1;

  467.             if (get_bits1(&s->gb)) // display size

  468.                 skip_bits(&s->gb, 32);

  469.         } else {

  470.             s->refreshrefmask = get_bits(&s->gb, 8);

  471.             s->refidx[0]      = get_bits(&s->gb, 3);

  472.             s->signbias[0]    = get_bits1(&s->gb);

  473.             s->refidx[1]      = get_bits(&s->gb, 3);

  474.             s->signbias[1]    = get_bits1(&s->gb);

  475.             s->refidx[2]      = get_bits(&s->gb, 3);

  476.             s->signbias[2]    = get_bits1(&s->gb);

  477.             if (!s->refs[s->refidx[0]].f->data[0] ||

  478.                 !s->refs[s->refidx[1]].f->data[0] ||

  479.                 !s->refs[s->refidx[2]].f->data[0]) {

  480.                 av_log(ctx, AV_LOG_ERROR, "Not all references are available\n");

  481.                 return AVERROR_INVALIDDATA;

  482.             }

  483.             if (get_bits1(&s->gb)) {

  484.                 w = s->refs[s->refidx[0]].f->width;

  485.                 h = s->refs[s->refidx[0]].f->height;

  486.             } else if (get_bits1(&s->gb)) {

  487.                 w = s->refs[s->refidx[1]].f->width;

  488.                 h = s->refs[s->refidx[1]].f->height;

  489.             } else if (get_bits1(&s->gb)) {

  490.                 w = s->refs[s->refidx[2]].f->width;

  491.                 h = s->refs[s->refidx[2]].f->height;

  492.             } else {

  493.                 w = get_bits(&s->gb, 16) + 1;

  494.                 h = get_bits(&s->gb, 16) + 1;

  495.             }

  496.             if (get_bits1(&s->gb)) // display size

  497.                 skip_bits(&s->gb, 32);

  498.             s->highprecisionmvs = get_bits1(&s->gb);

  499.             s->filtermode = get_bits1(&s->gb) ? FILTER_SWITCHABLE :

  500.                                                 get_bits(&s->gb, 2);

  501.             s->allowcompinter = s->signbias[0] != s->signbias[1] ||

  502.                                 s->signbias[0] != s->signbias[2];

  503.             if (s->allowcompinter) {

  504.                 if (s->signbias[0] == s->signbias[1]) {

  505.                     s->fixcompref    = 2;

  506.                     s->varcompref[0] = 0;

  507.                     s->varcompref[1] = 1;

  508.                 } else if (s->signbias[0] == s->signbias[2]) {

  509.                     s->fixcompref    = 1;

  510.                     s->varcompref[0] = 0;

  511.                     s->varcompref[1] = 2;

  512.                 } else {

  513.                     s->fixcompref    = 0;

  514.                     s->varcompref[0] = 1;

  515.                     s->varcompref[1] = 2;

  516.                 }

  517.             }

  518.         }

  519.     }

  520.     s->refreshctx   = s->errorres ? 0 : get_bits1(&s->gb);

  521.     s->parallelmode = s->errorres ? 1 : get_bits1(&s->gb);

  522.     s->framectxid   = c = get_bits(&s->gb, 2);

  523. 

  524.     /* loopfilter header data */

  525.     s->filter.level = get_bits(&s->gb, 6);

  526.     sharp = get_bits(&s->gb, 3);

  527.     // if sharpness changed, reinit lim/mblim LUTs. if it didn't change, keep

  528.     // the old cache values since they are still valid

  529.     if (s->filter.sharpness != sharp)

  530.         memset(s->filter.lim_lut, 0, sizeof(s->filter.lim_lut));

  531.     s->filter.sharpness = sharp;

  532.     if ((s->lf_delta.enabled = get_bits1(&s->gb))) {

  533.         if (get_bits1(&s->gb)) {

  534.             for (i = 0; i < 4; i++)

  535.                 if (get_bits1(&s->gb))

  536.                     s->lf_delta.ref[i] = get_sbits_inv(&s->gb, 6);

  537.             for (i = 0; i < 2; i++)

  538.                 if (get_bits1(&s->gb))

  539.                     s->lf_delta.mode[i] = get_sbits_inv(&s->gb, 6);

  540.         }

  541.     } else {

  542.         memset(&s->lf_delta, 0, sizeof(s->lf_delta));

  543.     }

  544. 

  545.     /* quantization header data */

  546.     s->yac_qi      = get_bits(&s->gb, 8);

  547.     s->ydc_qdelta  = get_bits1(&s->gb) ? get_sbits_inv(&s->gb, 4) : 0;

  548.     s->uvdc_qdelta = get_bits1(&s->gb) ? get_sbits_inv(&s->gb, 4) : 0;

  549.     s->uvac_qdelta = get_bits1(&s->gb) ? get_sbits_inv(&s->gb, 4) : 0;

  550.     s->lossless    = s->yac_qi == 0 && s->ydc_qdelta == 0 &&

  551.                      s->uvdc_qdelta == 0 && s->uvac_qdelta == 0;

  552. 

  553.     /* segmentation header info */

  554.     if ((s->segmentation.enabled = get_bits1(&s->gb))) {

  555.         if ((s->segmentation.update_map = get_bits1(&s->gb))) {

  556.             for (i = 0; i < 7; i++)

  557.                 s->prob.seg[i] = get_bits1(&s->gb) ?

  558.                                  get_bits(&s->gb, 8) : 255;

  559.             if ((s->segmentation.temporal = get_bits1(&s->gb)))

  560.                 for (i = 0; i < 3; i++)

  561.                     s->prob.segpred[i] = get_bits1(&s->gb) ?

  562.                                          get_bits(&s->gb, 8) : 255;

  563.         }

  564. 

  565.         if (get_bits1(&s->gb)) {

  566.             s->segmentation.absolute_vals = get_bits1(&s->gb);

  567.             for (i = 0; i < 8; i++) {

  568.                 if ((s->segmentation.feat[i].q_enabled = get_bits1(&s->gb)))

  569.                     s->segmentation.feat[i].q_val = get_sbits_inv(&s->gb, 8);

  570.                 if ((s->segmentation.feat[i].lf_enabled = get_bits1(&s->gb)))

  571.                     s->segmentation.feat[i].lf_val = get_sbits_inv(&s->gb, 6);

  572.                 if ((s->segmentation.feat[i].ref_enabled = get_bits1(&s->gb)))

  573.                     s->segmentation.feat[i].ref_val = get_bits(&s->gb, 2);

  574.                 s->segmentation.feat[i].skip_enabled = get_bits1(&s->gb);

  575.             }

  576.         }

  577.     } else {

  578.         s->segmentation.feat[0].q_enabled    = 0;

  579.         s->segmentation.feat[0].lf_enabled   = 0;

  580.         s->segmentation.feat[0].skip_enabled = 0;

  581.         s->segmentation.feat[0].ref_enabled  = 0;

  582.     }

  583. 

  584.     // set qmul[] based on Y/UV, AC/DC and segmentation Q idx deltas

  585.     for (i = 0; i < (s->segmentation.enabled ? 8 : 1); i++) {

  586.         int qyac, qydc, quvac, quvdc, lflvl, sh;

  587. 

  588.         if (s->segmentation.feat[i].q_enabled) {

  589.             if (s->segmentation.absolute_vals)

  590.                 qyac = s->segmentation.feat[i].q_val;

  591.             else

  592.                 qyac = s->yac_qi + s->segmentation.feat[i].q_val;

  593.         } else {

  594.             qyac  = s->yac_qi;

  595.         }

  596.         qydc  = av_clip_uintp2(qyac + s->ydc_qdelta, 8);

  597.         quvdc = av_clip_uintp2(qyac + s->uvdc_qdelta, 8);

  598.         quvac = av_clip_uintp2(qyac + s->uvac_qdelta, 8);

  599.         qyac  = av_clip_uintp2(qyac, 8);

  600. 

  601.         s->segmentation.feat[i].qmul[0][0] = vp9_dc_qlookup[qydc];

  602.         s->segmentation.feat[i].qmul[0][1] = vp9_ac_qlookup[qyac];

  603.         s->segmentation.feat[i].qmul[1][0] = vp9_dc_qlookup[quvdc];

  604.         s->segmentation.feat[i].qmul[1][1] = vp9_ac_qlookup[quvac];

  605. 

  606.         sh = s->filter.level >= 32;

  607.         if (s->segmentation.feat[i].lf_enabled) {

  608.             if (s->segmentation.absolute_vals)

  609.                 lflvl = s->segmentation.feat[i].lf_val;

  610.             else

  611.                 lflvl = s->filter.level + s->segmentation.feat[i].lf_val;

  612.         } else {

  613.             lflvl  = s->filter.level;

  614.         }

  615.         s->segmentation.feat[i].lflvl[0][0] =

  616.         s->segmentation.feat[i].lflvl[0][1] =

  617.             av_clip_uintp2(lflvl + (s->lf_delta.ref[0] << sh), 6);

  618.         for (j = 1; j < 4; j++) {

  619.             s->segmentation.feat[i].lflvl[j][0] =

  620.                 av_clip_uintp2(lflvl + ((s->lf_delta.ref[j] +

  621.                                          s->lf_delta.mode[0]) << sh), 6);

  622.             s->segmentation.feat[i].lflvl[j][1] =

  623.                 av_clip_uintp2(lflvl + ((s->lf_delta.ref[j] +

  624.                                          s->lf_delta.mode[1]) << sh), 6);

  625.         }

  626.     }

  627. 

  628.     /* tiling info */

  629.     if ((res = update_size(ctx, w, h)) < 0) {

  630.         av_log(ctx, AV_LOG_ERROR, "Failed to initialize decoder for %dx%d\n", w, h);

  631.         return res;

  632.     }

  633.     for (s->tiling.log2_tile_cols = 0;

  634.          (s->sb_cols >> s->tiling.log2_tile_cols) > 64;

  635.          s->tiling.log2_tile_cols++) ;

  636.     for (max = 0; (s->sb_cols >> max) >= 4; max++) ;

  637.     max = FFMAX(0, max - 1);

  638.     while (max > s->tiling.log2_tile_cols) {

  639.         if (get_bits1(&s->gb))

  640.             s->tiling.log2_tile_cols++;

  641.         else

  642.             break;

  643.     }

  644.     s->tiling.log2_tile_rows = decode012(&s->gb);

  645.     s->tiling.tile_rows = 1 << s->tiling.log2_tile_rows;

  646.     if (s->tiling.tile_cols != (1 << s->tiling.log2_tile_cols)) {

  647.         s->tiling.tile_cols = 1 << s->tiling.log2_tile_cols;

  648.         s->c_b = av_fast_realloc(s->c_b, &s->c_b_size,

  649.                                  sizeof(VP56RangeCoder) * s->tiling.tile_cols);

  650.         if (!s->c_b) {

  651.             av_log(ctx, AV_LOG_ERROR, "Ran out of memory during range coder init\n");

  652.             return AVERROR(ENOMEM);

  653.         }

  654.     }

  655. 

  656.     if (s->keyframe || s->errorres || s->intraonly) {

  657.         s->prob_ctx[0].p = s->prob_ctx[1].p = s->prob_ctx[2].p =

  658.                            s->prob_ctx[3].p = vp9_default_probs;

  659.         memcpy(s->prob_ctx[0].coef, vp9_default_coef_probs,

  660.                sizeof(vp9_default_coef_probs));

  661.         memcpy(s->prob_ctx[1].coef, vp9_default_coef_probs,

  662.                sizeof(vp9_default_coef_probs));

  663.         memcpy(s->prob_ctx[2].coef, vp9_default_coef_probs,

  664.                sizeof(vp9_default_coef_probs));

  665.         memcpy(s->prob_ctx[3].coef, vp9_default_coef_probs,

  666.                sizeof(vp9_default_coef_probs));

  667.     }

  668. 

  669.     // next 16 bits is size of the rest of the header (arith-coded)

  670.     size2 = get_bits(&s->gb, 16);

  671.     data2 = align_get_bits(&s->gb);

  672.     if (size2 > size - (data2 - data)) {

  673.         av_log(ctx, AV_LOG_ERROR, "Invalid compressed header size\n");

  674.         return AVERROR_INVALIDDATA;

  675.     }

  676.     ff_vp56_init_range_decoder(&s->c, data2, size2);

  677.     if (vp56_rac_get_prob_branchy(&s->c, 128)) { // marker bit

  678.         av_log(ctx, AV_LOG_ERROR, "Marker bit was set\n");

  679.         return AVERROR_INVALIDDATA;

  680.     }

  681. 

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

  683.         memset(s->counts.coef, 0, sizeof(s->counts.coef) + sizeof(s->counts.eob));

  684.     } else {

  685.         memset(&s->counts, 0, sizeof(s->counts));

  686.     }

  687.     // FIXME is it faster to not copy here, but do it down in the fw updates

  688.     // as explicit copies if the fw update is missing (and skip the copy upon

  689.     // fw update)?

  690.     s->prob.p = s->prob_ctx[c].p;

  691. 

  692.     // txfm updates

  693.     if (s->lossless) {

  694.         s->txfmmode = TX_4X4;

  695.     } else {

  696.         s->txfmmode = vp8_rac_get_uint(&s->c, 2);

  697.         if (s->txfmmode == 3)

  698.             s->txfmmode += vp8_rac_get(&s->c);

  699. 

  700.         if (s->txfmmode == TX_SWITCHABLE) {

  701.             for (i = 0; i < 2; i++)

  702.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  703.                     s->prob.p.tx8p[i] = update_prob(&s->c, s->prob.p.tx8p[i]);

  704.             for (i = 0; i < 2; i++)

  705.                 for (j = 0; j < 2; j++)

  706.                     if (vp56_rac_get_prob_branchy(&s->c, 252))

  707.                         s->prob.p.tx16p[i][j] =

  708.                             update_prob(&s->c, s->prob.p.tx16p[i][j]);

  709.             for (i = 0; i < 2; i++)

  710.                 for (j = 0; j < 3; j++)

  711.                     if (vp56_rac_get_prob_branchy(&s->c, 252))

  712.                         s->prob.p.tx32p[i][j] =

  713.                             update_prob(&s->c, s->prob.p.tx32p[i][j]);

  714.         }

  715.     }

  716. 

  717.     // coef updates

  718.     for (i = 0; i < 4; i++) {

  719.         uint8_t (*ref)[2][6][6][3] = s->prob_ctx[c].coef[i];

  720.         if (vp8_rac_get(&s->c)) {

  721.             for (j = 0; j < 2; j++)

  722.                 for (k = 0; k < 2; k++)

  723.                     for (l = 0; l < 6; l++)

  724.                         for (m = 0; m < 6; m++) {

  725.                             uint8_t *p = s->prob.coef[i][j][k][l][m];

  726.                             uint8_t *r = ref[j][k][l][m];

  727.                             if (m >= 3 && l == 0) // dc only has 3 pt

  728.                                 break;

  729.                             for (n = 0; n < 3; n++) {

  730.                                 if (vp56_rac_get_prob_branchy(&s->c, 252)) {

  731.                                     p[n] = update_prob(&s->c, r[n]);

  732.                                 } else {

  733.                                     p[n] = r[n];

  734.                                 }

  735.                             }

  736.                             p[3] = 0;

  737.                         }

  738.         } else {

  739.             for (j = 0; j < 2; j++)

  740.                 for (k = 0; k < 2; k++)

  741.                     for (l = 0; l < 6; l++)

  742.                         for (m = 0; m < 6; m++) {

  743.                             uint8_t *p = s->prob.coef[i][j][k][l][m];

  744.                             uint8_t *r = ref[j][k][l][m];

  745.                             if (m > 3 && l == 0) // dc only has 3 pt

  746.                                 break;

  747.                             memcpy(p, r, 3);

  748.                             p[3] = 0;

  749.                         }

  750.         }

  751.         if (s->txfmmode == i)

  752.             break;

  753.     }

  754. 

  755.     // mode updates

  756.     for (i = 0; i < 3; i++)

  757.         if (vp56_rac_get_prob_branchy(&s->c, 252))

  758.             s->prob.p.skip[i] = update_prob(&s->c, s->prob.p.skip[i]);

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

  760.         for (i = 0; i < 7; i++)

  761.             for (j = 0; j < 3; j++)

  762.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  763.                     s->prob.p.mv_mode[i][j] =

  764.                         update_prob(&s->c, s->prob.p.mv_mode[i][j]);

  765. 

  766.         if (s->filtermode == FILTER_SWITCHABLE)

  767.             for (i = 0; i < 4; i++)

  768.                 for (j = 0; j < 2; j++)

  769.                     if (vp56_rac_get_prob_branchy(&s->c, 252))

  770.                         s->prob.p.filter[i][j] =

  771.                             update_prob(&s->c, s->prob.p.filter[i][j]);

  772. 

  773.         for (i = 0; i < 4; i++)

  774.             if (vp56_rac_get_prob_branchy(&s->c, 252))

  775.                 s->prob.p.intra[i] = update_prob(&s->c, s->prob.p.intra[i]);

  776. 

  777.         if (s->allowcompinter) {

  778.             s->comppredmode = vp8_rac_get(&s->c);

  779.             if (s->comppredmode)

  780.                 s->comppredmode += vp8_rac_get(&s->c);

  781.             if (s->comppredmode == PRED_SWITCHABLE)

  782.                 for (i = 0; i < 5; i++)

  783.                     if (vp56_rac_get_prob_branchy(&s->c, 252))

  784.                         s->prob.p.comp[i] =

  785.                             update_prob(&s->c, s->prob.p.comp[i]);

  786.         } else {

  787.             s->comppredmode = PRED_SINGLEREF;

  788.         }

  789. 

  790.         if (s->comppredmode != PRED_COMPREF) {

  791.             for (i = 0; i < 5; i++) {

  792.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  793.                     s->prob.p.single_ref[i][0] =

  794.                         update_prob(&s->c, s->prob.p.single_ref[i][0]);

  795.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  796.                     s->prob.p.single_ref[i][1] =

  797.                         update_prob(&s->c, s->prob.p.single_ref[i][1]);

  798.             }

  799.         }

  800. 

  801.         if (s->comppredmode != PRED_SINGLEREF) {

  802.             for (i = 0; i < 5; i++)

  803.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  804.                     s->prob.p.comp_ref[i] =

  805.                         update_prob(&s->c, s->prob.p.comp_ref[i]);

  806.         }

  807. 

  808.         for (i = 0; i < 4; i++)

  809.             for (j = 0; j < 9; j++)

  810.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  811.                     s->prob.p.y_mode[i][j] =

  812.                         update_prob(&s->c, s->prob.p.y_mode[i][j]);

  813. 

  814.         for (i = 0; i < 4; i++)

  815.             for (j = 0; j < 4; j++)

  816.                 for (k = 0; k < 3; k++)

  817.                     if (vp56_rac_get_prob_branchy(&s->c, 252))

  818.                         s->prob.p.partition[3 - i][j][k] =

  819.                             update_prob(&s->c, s->prob.p.partition[3 - i][j][k]);

  820. 

  821.         // mv fields don't use the update_prob subexp model for some reason

  822.         for (i = 0; i < 3; i++)

  823.             if (vp56_rac_get_prob_branchy(&s->c, 252))

  824.                 s->prob.p.mv_joint[i] = (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  825. 

  826.         for (i = 0; i < 2; i++) {

  827.             if (vp56_rac_get_prob_branchy(&s->c, 252))

  828.                 s->prob.p.mv_comp[i].sign = (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  829. 

  830.             for (j = 0; j < 10; j++)

  831.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  832.                     s->prob.p.mv_comp[i].classes[j] =

  833.                         (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  834. 

  835.             if (vp56_rac_get_prob_branchy(&s->c, 252))

  836.                 s->prob.p.mv_comp[i].class0 = (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  837. 

  838.             for (j = 0; j < 10; j++)

  839.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  840.                     s->prob.p.mv_comp[i].bits[j] =

  841.                         (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  842.         }

  843. 

  844.         for (i = 0; i < 2; i++) {

  845.             for (j = 0; j < 2; j++)

  846.                 for (k = 0; k < 3; k++)

  847.                     if (vp56_rac_get_prob_branchy(&s->c, 252))

  848.                         s->prob.p.mv_comp[i].class0_fp[j][k] =

  849.                             (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  850. 

  851.             for (j = 0; j < 3; j++)

  852.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  853.                     s->prob.p.mv_comp[i].fp[j] =

  854.                         (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  855.         }

  856. 

  857.         if (s->highprecisionmvs) {

  858.             for (i = 0; i < 2; i++) {

  859.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  860.                     s->prob.p.mv_comp[i].class0_hp =

  861.                         (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  862. 

  863.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  864.                     s->prob.p.mv_comp[i].hp =

  865.                         (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  866.             }

  867.         }

  868.     }

  869. 

  870.     return (data2 - data) + size2;

  871. }

  872. 

  873. static av_always_inline void clamp_mv(VP56mv *dst, const VP56mv *src,

  874.                                       VP9Context *s)

  875. {

  876.     dst->x = av_clip(src->x, s->min_mv.x, s->max_mv.x);

  877.     dst->y = av_clip(src->y, s->min_mv.y, s->max_mv.y);

  878. }

  879. 

  880. static void find_ref_mvs(VP9Context *s,

  881.                          VP56mv *pmv, int ref, int z, int idx, int sb)

  882. {

  883.     static const int8_t mv_ref_blk_off[N_BS_SIZES][8][2] = {

  884.         [BS_64x64] = {{  3, -1 }, { -1,  3 }, {  4, -1 }, { -1,  4 },

  885.                       { -1, -1 }, {  0, -1 }, { -1,  0 }, {  6, -1 }},

  886.         [BS_64x32] = {{  0, -1 }, { -1,  0 }, {  4, -1 }, { -1,  2 },

  887.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, {  2, -1 }},

  888.         [BS_32x64] = {{ -1,  0 }, {  0, -1 }, { -1,  4 }, {  2, -1 },

  889.                       { -1, -1 }, { -3,  0 }, {  0, -3 }, { -1,  2 }},

  890.         [BS_32x32] = {{  1, -1 }, { -1,  1 }, {  2, -1 }, { -1,  2 },

  891.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

  892.         [BS_32x16] = {{  0, -1 }, { -1,  0 }, {  2, -1 }, { -1, -1 },

  893.                       { -1,  1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

  894.         [BS_16x32] = {{ -1,  0 }, {  0, -1 }, { -1,  2 }, { -1, -1 },

  895.                       {  1, -1 }, { -3,  0 }, {  0, -3 }, { -3, -3 }},

  896.         [BS_16x16] = {{  0, -1 }, { -1,  0 }, {  1, -1 }, { -1,  1 },

  897.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

  898.         [BS_16x8]  = {{  0, -1 }, { -1,  0 }, {  1, -1 }, { -1, -1 },

  899.                       {  0, -2 }, { -2,  0 }, { -2, -1 }, { -1, -2 }},

  900.         [BS_8x16]  = {{ -1,  0 }, {  0, -1 }, { -1,  1 }, { -1, -1 },

  901.                       { -2,  0 }, {  0, -2 }, { -1, -2 }, { -2, -1 }},

  902.         [BS_8x8]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  903.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  904.         [BS_8x4]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  905.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  906.         [BS_4x8]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  907.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  908.         [BS_4x4]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  909.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  910.     };

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

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

  913.     const int8_t (*p)[2] = mv_ref_blk_off[b->bs];

  914. #define INVALID_MV 0x80008000U

  915.     uint32_t mem = INVALID_MV;

  916.     int i;

  917. 

  918. #define RETURN_DIRECT_MV(mv) \

  919.     do { \

  920.         uint32_t m = AV_RN32A(&mv); \

  921.         if (!idx) { \

  922.             AV_WN32A(pmv, m); \

  923.             return; \

  924.         } else if (mem == INVALID_MV) { \

  925.             mem = m; \

  926.         } else if (m != mem) { \

  927.             AV_WN32A(pmv, m); \

  928.             return; \

  929.         } \

  930.     } while (0)

  931. 

  932.     if (sb >= 0) {

  933.         if (sb == 2 || sb == 1) {

  934.             RETURN_DIRECT_MV(b->mv[0][z]);

  935.         } else if (sb == 3) {

  936.             RETURN_DIRECT_MV(b->mv[2][z]);

  937.             RETURN_DIRECT_MV(b->mv[1][z]);

  938.             RETURN_DIRECT_MV(b->mv[0][z]);

  939.         }

  940. 

  941. #define RETURN_MV(mv) \

  942.     do { \

  943.         if (sb > 0) { \

  944.             VP56mv tmp; \

  945.             uint32_t m; \

  946.             clamp_mv(&tmp, &mv, s); \

  947.             m = AV_RN32A(&tmp); \

  948.             if (!idx) { \

  949.                 AV_WN32A(pmv, m); \

  950.                 return; \

  951.             } else if (mem == INVALID_MV) { \

  952.                 mem = m; \

  953.             } else if (m != mem) { \

  954.                 AV_WN32A(pmv, m); \

  955.                 return; \

  956.             } \

  957.         } else { \

  958.             uint32_t m = AV_RN32A(&mv); \

  959.             if (!idx) { \

  960.                 clamp_mv(pmv, &mv, s); \

  961.                 return; \

  962.             } else if (mem == INVALID_MV) { \

  963.                 mem = m; \

  964.             } else if (m != mem) { \

  965.                 clamp_mv(pmv, &mv, s); \

  966.                 return; \

  967.             } \

  968.         } \

  969.     } while (0)

  970. 

  971.         if (row > 0) {

  972.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[(row - 1) * s->sb_cols * 8 + col];

  973.             if (mv->ref[0] == ref) {

  974.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][0]);

  975.             } else if (mv->ref[1] == ref) {

  976.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][1]);

  977.             }

  978.         }

  979.         if (col > s->tiling.tile_col_start) {

  980.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[row * s->sb_cols * 8 + col - 1];

  981.             if (mv->ref[0] == ref) {

  982.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][0]);

  983.             } else if (mv->ref[1] == ref) {

  984.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][1]);

  985.             }

  986.         }

  987.         i = 2;

  988.     } else {

  989.         i = 0;

  990.     }

  991. 

  992.     // previously coded MVs in this neighbourhood, using same reference frame

  993.     for (; i < 8; i++) {

  994.         int c = p[i][0] + col, r = p[i][1] + row;

  995. 

  996.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  997.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  998. 

  999.             if (mv->ref[0] == ref) {

  1000.                 RETURN_MV(mv->mv[0]);

  1001.             } else if (mv->ref[1] == ref) {

  1002.                 RETURN_MV(mv->mv[1]);

  1003.             }

  1004.         }

  1005.     }

  1006. 

  1007.     // MV at this position in previous frame, using same reference frame

  1008.     if (s->use_last_frame_mvs) {

  1009.         struct VP9mvrefPair *mv = &s->frames[LAST_FRAME].mv[row * s->sb_cols * 8 + col];

  1010. 

  1011.         if (mv->ref[0] == ref) {

  1012.             RETURN_MV(mv->mv[0]);

  1013.         } else if (mv->ref[1] == ref) {

  1014.             RETURN_MV(mv->mv[1]);

  1015.         }

  1016.     }

  1017. 

  1018. #define RETURN_SCALE_MV(mv, scale) \

  1019.     do { \

  1020.         if (scale) { \

  1021.             VP56mv mv_temp = { -mv.x, -mv.y }; \

  1022.             RETURN_MV(mv_temp); \

  1023.         } else { \

  1024.             RETURN_MV(mv); \

  1025.         } \

  1026.     } while (0)

  1027. 

  1028.     // previously coded MVs in this neighbourhood, using different reference frame

  1029.     for (i = 0; i < 8; i++) {

  1030.         int c = p[i][0] + col, r = p[i][1] + row;

  1031. 

  1032.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1033.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1034. 

  1035.             if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1036.                 RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1037.             }

  1038.             if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1039.                 // BUG - libvpx has this condition regardless of whether

  1040.                 // we used the first ref MV and pre-scaling

  1041.                 AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1042.                 RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1043.             }

  1044.         }

  1045.     }

  1046. 

  1047.     // MV at this position in previous frame, using different reference frame

  1048.     if (s->use_last_frame_mvs) {

  1049.         struct VP9mvrefPair *mv = &s->frames[LAST_FRAME].mv[row * s->sb_cols * 8 + col];

  1050. 

  1051.         if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1052.             RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1053.         }

  1054.         if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1055.             // BUG - libvpx has this condition regardless of whether

  1056.             // we used the first ref MV and pre-scaling

  1057.             AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1058.             RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1059.         }

  1060.     }

  1061. 

  1062.     AV_ZERO32(pmv);

  1063. #undef INVALID_MV

  1064. #undef RETURN_MV

  1065. #undef RETURN_SCALE_MV

  1066. }

  1067. 

  1068. static av_always_inline int read_mv_component(VP9Context *s, int idx, int hp)

  1069. {

  1070.     int bit, sign = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].sign);

  1071.     int n, c = vp8_rac_get_tree(&s->c, vp9_mv_class_tree,

  1072.                                 s->prob.p.mv_comp[idx].classes);

  1073. 

  1074.     s->counts.mv_comp[idx].sign[sign]++;

  1075.     s->counts.mv_comp[idx].classes[c]++;

  1076.     if (c) {

  1077.         int m;

  1078. 

  1079.         for (n = 0, m = 0; m < c; m++) {

  1080.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].bits[m]);

  1081.             n |= bit << m;

  1082.             s->counts.mv_comp[idx].bits[m][bit]++;

  1083.         }

  1084.         n <<= 3;

  1085.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree, s->prob.p.mv_comp[idx].fp);

  1086.         n |= bit << 1;

  1087.         s->counts.mv_comp[idx].fp[bit]++;

  1088.         if (hp) {

  1089.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].hp);

  1090.             s->counts.mv_comp[idx].hp[bit]++;

  1091.             n |= bit;

  1092.         } else {

  1093.             n |= 1;

  1094.             // bug in libvpx - we count for bw entropy purposes even if the

  1095.             // bit wasn't coded

  1096.             s->counts.mv_comp[idx].hp[1]++;

  1097.         }

  1098.         n += 8 << c;

  1099.     } else {

  1100.         n = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0);

  1101.         s->counts.mv_comp[idx].class0[n]++;

  1102.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree,

  1103.                                s->prob.p.mv_comp[idx].class0_fp[n]);

  1104.         s->counts.mv_comp[idx].class0_fp[n][bit]++;

  1105.         n = (n << 3) | (bit << 1);

  1106.         if (hp) {

  1107.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0_hp);

  1108.             s->counts.mv_comp[idx].class0_hp[bit]++;

  1109.             n |= bit;

  1110.         } else {

  1111.             n |= 1;

  1112.             // bug in libvpx - we count for bw entropy purposes even if the

  1113.             // bit wasn't coded

  1114.             s->counts.mv_comp[idx].class0_hp[1]++;

  1115.         }

  1116.     }

  1117. 

  1118.     return sign ? -(n + 1) : (n + 1);

  1119. }

  1120. 

  1121. static void fill_mv(VP9Context *s,

  1122.                     VP56mv *mv, int mode, int sb)

  1123. {

  1124.     VP9Block *const b = &s->b;

  1125. 

  1126.     if (mode == ZEROMV) {

  1127.         memset(mv, 0, sizeof(*mv) * 2);

  1128.     } else {

  1129.         int hp;

  1130. 

  1131.         // FIXME cache this value and reuse for other subblocks

  1132.         find_ref_mvs(s, &mv[0], b->ref[0], 0, mode == NEARMV,

  1133.                      mode == NEWMV ? -1 : sb);

  1134.         // FIXME maybe move this code into find_ref_mvs()

  1135.         if ((mode == NEWMV || sb == -1) &&

  1136.             !(hp = s->highprecisionmvs && abs(mv[0].x) < 64 && abs(mv[0].y) < 64)) {

  1137.             if (mv[0].y & 1) {

  1138.                 if (mv[0].y < 0)

  1139.                     mv[0].y++;

  1140.                 else

  1141.                     mv[0].y--;

  1142.             }

  1143.             if (mv[0].x & 1) {

  1144.                 if (mv[0].x < 0)

  1145.                     mv[0].x++;

  1146.                 else

  1147.                     mv[0].x--;

  1148.             }

  1149.         }

  1150.         if (mode == NEWMV) {

  1151.             enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1152.                                               s->prob.p.mv_joint);

  1153. 

  1154.             s->counts.mv_joint[j]++;

  1155.             if (j >= MV_JOINT_V)

  1156.                 mv[0].y += read_mv_component(s, 0, hp);

  1157.             if (j & 1)

  1158.                 mv[0].x += read_mv_component(s, 1, hp);

  1159.         }

  1160. 

  1161.         if (b->comp) {

  1162.             // FIXME cache this value and reuse for other subblocks

  1163.             find_ref_mvs(s, &mv[1], b->ref[1], 1, mode == NEARMV,

  1164.                          mode == NEWMV ? -1 : sb);

  1165.             if ((mode == NEWMV || sb == -1) &&

  1166.                 !(hp = s->highprecisionmvs && abs(mv[1].x) < 64 && abs(mv[1].y) < 64)) {

  1167.                 if (mv[1].y & 1) {

  1168.                     if (mv[1].y < 0)

  1169.                         mv[1].y++;

  1170.                     else

  1171.                         mv[1].y--;

  1172.                 }

  1173.                 if (mv[1].x & 1) {

  1174.                     if (mv[1].x < 0)

  1175.                         mv[1].x++;

  1176.                     else

  1177.                         mv[1].x--;

  1178.                 }

  1179.             }

  1180.             if (mode == NEWMV) {

  1181.                 enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1182.                                                   s->prob.p.mv_joint);

  1183. 

  1184.                 s->counts.mv_joint[j]++;

  1185.                 if (j >= MV_JOINT_V)

  1186.                     mv[1].y += read_mv_component(s, 0, hp);

  1187.                 if (j & 1)

  1188.                     mv[1].x += read_mv_component(s, 1, hp);

  1189.             }

  1190.         }

  1191.     }

  1192. }

  1193. 

  1194. static void decode_mode(AVCodecContext *ctx)

  1195. {

  1196.     static const uint8_t left_ctx[N_BS_SIZES] = {

  1197.         0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf

  1198.     };

  1199.     static const uint8_t above_ctx[N_BS_SIZES] = {

  1200.         0x0, 0x0, 0x8, 0x8, 0x8, 0xc, 0xc, 0xc, 0xe, 0xe, 0xe, 0xf, 0xf

  1201.     };

  1202.     static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {

  1203.         TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,

  1204.         TX_16X16, TX_8X8, TX_8X8, TX_8X8, TX_4X4, TX_4X4, TX_4X4

  1205.     };

  1206.     VP9Context *s = ctx->priv_data;

  1207.     VP9Block *const b = &s->b;

  1208.     int row = s->row, col = s->col, row7 = s->row7;

  1209.     enum TxfmMode max_tx = max_tx_for_bl_bp[b->bs];

  1210.     int w4 = FFMIN(s->cols - col, bwh_tab[1][b->bs][0]);

  1211.     int h4 = FFMIN(s->rows - row, bwh_tab[1][b->bs][1]), y;

  1212.     int have_a = row > 0, have_l = col > s->tiling.tile_col_start;

  1213. 

  1214.     if (!s->segmentation.enabled) {

  1215.         b->seg_id = 0;

  1216.     } else if (s->keyframe || s->intraonly) {

  1217.         b->seg_id = s->segmentation.update_map ?

  1218.             vp8_rac_get_tree(&s->c, vp9_segmentation_tree, s->prob.seg) : 0;

  1219.     } else if (!s->segmentation.update_map ||

  1220.                (s->segmentation.temporal &&

  1221.                 vp56_rac_get_prob_branchy(&s->c,

  1222.                     s->prob.segpred[s->above_segpred_ctx[col] +

  1223.                                     s->left_segpred_ctx[row7]]))) {

  1224.         int pred = 8, x;

  1225.         uint8_t *refsegmap = s->frames[LAST_FRAME].segmentation_map;

  1226. 

  1227.         for (y = 0; y < h4; y++)

  1228.             for (x = 0; x < w4; x++)

  1229.                 pred = FFMIN(pred, refsegmap[(y + row) * 8 * s->sb_cols + x + col]);

  1230.         av_assert1(pred < 8);

  1231.         b->seg_id = pred;

  1232. 

  1233.         memset(&s->above_segpred_ctx[col], 1, w4);

  1234.         memset(&s->left_segpred_ctx[row7], 1, h4);

  1235.     } else {

  1236.         b->seg_id = vp8_rac_get_tree(&s->c, vp9_segmentation_tree,

  1237.                                      s->prob.seg);

  1238. 

  1239.         memset(&s->above_segpred_ctx[col], 0, w4);

  1240.         memset(&s->left_segpred_ctx[row7], 0, h4);

  1241.     }

  1242.     if ((s->segmentation.enabled && s->segmentation.update_map) || s->keyframe) {

  1243.         uint8_t *segmap = s->frames[CUR_FRAME].segmentation_map;

  1244. 

  1245.         for (y = 0; y < h4; y++)

  1246.             memset(&segmap[(y + row) * 8 * s->sb_cols + col], b->seg_id, w4);

  1247.     }

  1248. 

  1249.     b->skip = s->segmentation.enabled &&

  1250.         s->segmentation.feat[b->seg_id].skip_enabled;

  1251.     if (!b->skip) {

  1252.         int c = s->left_skip_ctx[row7] + s->above_skip_ctx[col];

  1253.         b->skip = vp56_rac_get_prob(&s->c, s->prob.p.skip[c]);

  1254.         s->counts.skip[c][b->skip]++;

  1255.     }

  1256. 

  1257.     if (s->keyframe || s->intraonly) {

  1258.         b->intra = 1;

  1259.     } else if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1260.         b->intra = !s->segmentation.feat[b->seg_id].ref_val;

  1261.     } else {

  1262.         int c, bit;

  1263. 

  1264.         if (have_a && have_l) {

  1265.             c = s->above_intra_ctx[col] + s->left_intra_ctx[row7];

  1266.             c += (c == 2);

  1267.         } else {

  1268.             c = have_a ? 2 * s->above_intra_ctx[col] :

  1269.                 have_l ? 2 * s->left_intra_ctx[row7] : 0;

  1270.         }

  1271.         bit = vp56_rac_get_prob(&s->c, s->prob.p.intra[c]);

  1272.         s->counts.intra[c][bit]++;

  1273.         b->intra = !bit;

  1274.     }

  1275. 

  1276.     if ((b->intra || !b->skip) && s->txfmmode == TX_SWITCHABLE) {

  1277.         int c;

  1278.         if (have_a) {

  1279.             if (have_l) {

  1280.                 c = (s->above_skip_ctx[col] ? max_tx :

  1281.                      s->above_txfm_ctx[col]) +

  1282.                     (s->left_skip_ctx[row7] ? max_tx :

  1283.                      s->left_txfm_ctx[row7]) > max_tx;

  1284.             } else {

  1285.                 c = s->above_skip_ctx[col] ? 1 :

  1286.                     (s->above_txfm_ctx[col] * 2 > max_tx);

  1287.             }

  1288.         } else if (have_l) {

  1289.             c = s->left_skip_ctx[row7] ? 1 :

  1290.                 (s->left_txfm_ctx[row7] * 2 > max_tx);

  1291.         } else {

  1292.             c = 1;

  1293.         }

  1294.         switch (max_tx) {

  1295.         case TX_32X32:

  1296.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][0]);

  1297.             if (b->tx) {

  1298.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][1]);

  1299.                 if (b->tx == 2)

  1300.                     b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][2]);

  1301.             }

  1302.             s->counts.tx32p[c][b->tx]++;

  1303.             break;

  1304.         case TX_16X16:

  1305.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][0]);

  1306.             if (b->tx)

  1307.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][1]);

  1308.             s->counts.tx16p[c][b->tx]++;

  1309.             break;

  1310.         case TX_8X8:

  1311.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx8p[c]);

  1312.             s->counts.tx8p[c][b->tx]++;

  1313.             break;

  1314.         case TX_4X4:

  1315.             b->tx = TX_4X4;

  1316.             break;

  1317.         }

  1318.     } else {

  1319.         b->tx = FFMIN(max_tx, s->txfmmode);

  1320.     }

  1321. 

  1322.     if (s->keyframe || s->intraonly) {

  1323.         uint8_t *a = &s->above_mode_ctx[col * 2];

  1324.         uint8_t *l = &s->left_mode_ctx[(row7) << 1];

  1325. 

  1326.         b->comp = 0;

  1327.         if (b->bs > BS_8x8) {

  1328.             // FIXME the memory storage intermediates here aren't really

  1329.             // necessary, they're just there to make the code slightly

  1330.             // simpler for now

  1331.             b->mode[0] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1332.                                     vp9_default_kf_ymode_probs[a[0]][l[0]]);

  1333.             if (b->bs != BS_8x4) {

  1334.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1335.                                  vp9_default_kf_ymode_probs[a[1]][b->mode[0]]);

  1336.                 l[0] = a[1] = b->mode[1];

  1337.             } else {

  1338.                 l[0] = a[1] = b->mode[1] = b->mode[0];

  1339.             }

  1340.             if (b->bs != BS_4x8) {

  1341.                 b->mode[2] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1342.                                         vp9_default_kf_ymode_probs[a[0]][l[1]]);

  1343.                 if (b->bs != BS_8x4) {

  1344.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1345.                                   vp9_default_kf_ymode_probs[a[1]][b->mode[2]]);

  1346.                     l[1] = a[1] = b->mode[3];

  1347.                 } else {

  1348.                     l[1] = a[1] = b->mode[3] = b->mode[2];

  1349.                 }

  1350.             } else {

  1351.                 b->mode[2] = b->mode[0];

  1352.                 l[1] = a[1] = b->mode[3] = b->mode[1];

  1353.             }

  1354.         } else {

  1355.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1356.                                           vp9_default_kf_ymode_probs[*a][*l]);

  1357.             b->mode[3] = b->mode[2] = b->mode[1] = b->mode[0];

  1358.             // FIXME this can probably be optimized

  1359.             memset(a, b->mode[0], bwh_tab[0][b->bs][0]);

  1360.             memset(l, b->mode[0], bwh_tab[0][b->bs][1]);

  1361.         }

  1362.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1363.                                      vp9_default_kf_uvmode_probs[b->mode[3]]);

  1364.     } else if (b->intra) {

  1365.         b->comp = 0;

  1366.         if (b->bs > BS_8x8) {

  1367.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1368.                                           s->prob.p.y_mode[0]);

  1369.             s->counts.y_mode[0][b->mode[0]]++;

  1370.             if (b->bs != BS_8x4) {

  1371.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1372.                                               s->prob.p.y_mode[0]);

  1373.                 s->counts.y_mode[0][b->mode[1]]++;

  1374.             } else {

  1375.                 b->mode[1] = b->mode[0];

  1376.             }

  1377.             if (b->bs != BS_4x8) {

  1378.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1379.                                               s->prob.p.y_mode[0]);

  1380.                 s->counts.y_mode[0][b->mode[2]]++;

  1381.                 if (b->bs != BS_8x4) {

  1382.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1383.                                                   s->prob.p.y_mode[0]);

  1384.                     s->counts.y_mode[0][b->mode[3]]++;

  1385.                 } else {

  1386.                     b->mode[3] = b->mode[2];

  1387.                 }

  1388.             } else {

  1389.                 b->mode[2] = b->mode[0];

  1390.                 b->mode[3] = b->mode[1];

  1391.             }

  1392.         } else {

  1393.             static const uint8_t size_group[10] = {

  1394.                 3, 3, 3, 3, 2, 2, 2, 1, 1, 1

  1395.             };

  1396.             int sz = size_group[b->bs];

  1397. 

  1398.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1399.                                           s->prob.p.y_mode[sz]);

  1400.             b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1401.             s->counts.y_mode[sz][b->mode[3]]++;

  1402.         }

  1403.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1404.                                      s->prob.p.uv_mode[b->mode[3]]);

  1405.         s->counts.uv_mode[b->mode[3]][b->uvmode]++;

  1406.     } else {

  1407.         static const uint8_t inter_mode_ctx_lut[14][14] = {

  1408.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1409.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1410.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1411.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1412.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1413.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1414.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1415.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1416.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1417.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1418.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1419.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1420.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },

  1421.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },

  1422.         };

  1423. 

  1424.         if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1425.             av_assert2(s->segmentation.feat[b->seg_id].ref_val != 0);

  1426.             b->comp = 0;

  1427.             b->ref[0] = s->segmentation.feat[b->seg_id].ref_val - 1;

  1428.         } else {

  1429.             // read comp_pred flag

  1430.             if (s->comppredmode != PRED_SWITCHABLE) {

  1431.                 b->comp = s->comppredmode == PRED_COMPREF;

  1432.             } else {

  1433.                 int c;

  1434. 

  1435.                 // FIXME add intra as ref=0xff (or -1) to make these easier?

  1436.                 if (have_a) {

  1437.                     if (have_l) {

  1438.                         if (s->above_comp_ctx[col] && s->left_comp_ctx[row7]) {

  1439.                             c = 4;

  1440.                         } else if (s->above_comp_ctx[col]) {

  1441.                             c = 2 + (s->left_intra_ctx[row7] ||

  1442.                                      s->left_ref_ctx[row7] == s->fixcompref);

  1443.                         } else if (s->left_comp_ctx[row7]) {

  1444.                             c = 2 + (s->above_intra_ctx[col] ||

  1445.                                      s->above_ref_ctx[col] == s->fixcompref);

  1446.                         } else {

  1447.                             c = (!s->above_intra_ctx[col] &&

  1448.                                  s->above_ref_ctx[col] == s->fixcompref) ^

  1449.                             (!s->left_intra_ctx[row7] &&

  1450.                              s->left_ref_ctx[row & 7] == s->fixcompref);

  1451.                         }

  1452.                     } else {

  1453.                         c = s->above_comp_ctx[col] ? 3 :

  1454.                         (!s->above_intra_ctx[col] && s->above_ref_ctx[col] == s->fixcompref);

  1455.                     }

  1456.                 } else if (have_l) {

  1457.                     c = s->left_comp_ctx[row7] ? 3 :

  1458.                     (!s->left_intra_ctx[row7] && s->left_ref_ctx[row7] == s->fixcompref);

  1459.                 } else {

  1460.                     c = 1;

  1461.                 }

  1462.                 b->comp = vp56_rac_get_prob(&s->c, s->prob.p.comp[c]);

  1463.                 s->counts.comp[c][b->comp]++;

  1464.             }

  1465. 

  1466.             // read actual references

  1467.             // FIXME probably cache a few variables here to prevent repetitive

  1468.             // memory accesses below

  1469.             if (b->comp) /* two references */ {

  1470.                 int fix_idx = s->signbias[s->fixcompref], var_idx = !fix_idx, c, bit;

  1471. 

  1472.                 b->ref[fix_idx] = s->fixcompref;

  1473.                 // FIXME can this codeblob be replaced by some sort of LUT?

  1474.                 if (have_a) {

  1475.                     if (have_l) {

  1476.                         if (s->above_intra_ctx[col]) {

  1477.                             if (s->left_intra_ctx[row7]) {

  1478.                                 c = 2;

  1479.                             } else {

  1480.                                 c = 1 + 2 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1481.                             }

  1482.                         } else if (s->left_intra_ctx[row7]) {

  1483.                             c = 1 + 2 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1484.                         } else {

  1485.                             int refl = s->left_ref_ctx[row7], refa = s->above_ref_ctx[col];

  1486. 

  1487.                             if (refl == refa && refa == s->varcompref[1]) {

  1488.                                 c = 0;

  1489.                             } else if (!s->left_comp_ctx[row7] && !s->above_comp_ctx[col]) {

  1490.                                 if ((refa == s->fixcompref && refl == s->varcompref[0]) ||

  1491.                                     (refl == s->fixcompref && refa == s->varcompref[0])) {

  1492.                                     c = 4;

  1493.                                 } else {

  1494.                                     c = (refa == refl) ? 3 : 1;

  1495.                                 }

  1496.                             } else if (!s->left_comp_ctx[row7]) {

  1497.                                 if (refa == s->varcompref[1] && refl != s->varcompref[1]) {

  1498.                                     c = 1;

  1499.                                 } else {

  1500.                                     c = (refl == s->varcompref[1] &&

  1501.                                          refa != s->varcompref[1]) ? 2 : 4;

  1502.                                 }

  1503.                             } else if (!s->above_comp_ctx[col]) {

  1504.                                 if (refl == s->varcompref[1] && refa != s->varcompref[1]) {

  1505.                                     c = 1;

  1506.                                 } else {

  1507.                                     c = (refa == s->varcompref[1] &&

  1508.                                          refl != s->varcompref[1]) ? 2 : 4;

  1509.                                 }

  1510.                             } else {

  1511.                                 c = (refl == refa) ? 4 : 2;

  1512.                             }

  1513.                         }

  1514.                     } else {

  1515.                         if (s->above_intra_ctx[col]) {

  1516.                             c = 2;

  1517.                         } else if (s->above_comp_ctx[col]) {

  1518.                             c = 4 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1519.                         } else {

  1520.                             c = 3 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1521.                         }

  1522.                     }

  1523.                 } else if (have_l) {

  1524.                     if (s->left_intra_ctx[row7]) {

  1525.                         c = 2;

  1526.                     } else if (s->left_comp_ctx[row7]) {

  1527.                         c = 4 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1528.                     } else {

  1529.                         c = 3 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1530.                     }

  1531.                 } else {

  1532.                     c = 2;

  1533.                 }

  1534.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.comp_ref[c]);

  1535.                 b->ref[var_idx] = s->varcompref[bit];

  1536.                 s->counts.comp_ref[c][bit]++;

  1537.             } else /* single reference */ {

  1538.                 int bit, c;

  1539. 

  1540.                 if (have_a && !s->above_intra_ctx[col]) {

  1541.                     if (have_l && !s->left_intra_ctx[row7]) {

  1542.                         if (s->left_comp_ctx[row7]) {

  1543.                             if (s->above_comp_ctx[col]) {

  1544.                                 c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7] ||

  1545.                                          !s->above_ref_ctx[col]);

  1546.                             } else {

  1547.                                 c = (3 * !s->above_ref_ctx[col]) +

  1548.                                     (!s->fixcompref || !s->left_ref_ctx[row7]);

  1549.                             }

  1550.                         } else if (s->above_comp_ctx[col]) {

  1551.                             c = (3 * !s->left_ref_ctx[row7]) +

  1552.                                 (!s->fixcompref || !s->above_ref_ctx[col]);

  1553.                         } else {

  1554.                             c = 2 * !s->left_ref_ctx[row7] + 2 * !s->above_ref_ctx[col];

  1555.                         }

  1556.                     } else if (s->above_intra_ctx[col]) {

  1557.                         c = 2;

  1558.                     } else if (s->above_comp_ctx[col]) {

  1559.                         c = 1 + (!s->fixcompref || !s->above_ref_ctx[col]);

  1560.                     } else {

  1561.                         c = 4 * (!s->above_ref_ctx[col]);

  1562.                     }

  1563.                 } else if (have_l && !s->left_intra_ctx[row7]) {

  1564.                     if (s->left_intra_ctx[row7]) {

  1565.                         c = 2;

  1566.                     } else if (s->left_comp_ctx[row7]) {

  1567.                         c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7]);

  1568.                     } else {

  1569.                         c = 4 * (!s->left_ref_ctx[row7]);

  1570.                     }

  1571.                 } else {

  1572.                     c = 2;

  1573.                 }

  1574.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][0]);

  1575.                 s->counts.single_ref[c][0][bit]++;

  1576.                 if (!bit) {

  1577.                     b->ref[0] = 0;

  1578.                 } else {

  1579.                     // FIXME can this codeblob be replaced by some sort of LUT?

  1580.                     if (have_a) {

  1581.                         if (have_l) {

  1582.                             if (s->left_intra_ctx[row7]) {

  1583.                                 if (s->above_intra_ctx[col]) {

  1584.                                     c = 2;

  1585.                                 } else if (s->above_comp_ctx[col]) {

  1586.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1587.                                                  s->above_ref_ctx[col] == 1);

  1588.                                 } else if (!s->above_ref_ctx[col]) {

  1589.                                     c = 3;

  1590.                                 } else {

  1591.                                     c = 4 * (s->above_ref_ctx[col] == 1);

  1592.                                 }

  1593.                             } else if (s->above_intra_ctx[col]) {

  1594.                                 if (s->left_intra_ctx[row7]) {

  1595.                                     c = 2;

  1596.                                 } else if (s->left_comp_ctx[row7]) {

  1597.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1598.                                                  s->left_ref_ctx[row7] == 1);

  1599.                                 } else if (!s->left_ref_ctx[row7]) {

  1600.                                     c = 3;

  1601.                                 } else {

  1602.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1603.                                 }

  1604.                             } else if (s->above_comp_ctx[col]) {

  1605.                                 if (s->left_comp_ctx[row7]) {

  1606.                                     if (s->left_ref_ctx[row7] == s->above_ref_ctx[col]) {

  1607.                                         c = 3 * (s->fixcompref == 1 ||

  1608.                                                  s->left_ref_ctx[row7] == 1);

  1609.                                     } else {

  1610.                                         c = 2;

  1611.                                     }

  1612.                                 } else if (!s->left_ref_ctx[row7]) {

  1613.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1614.                                                  s->above_ref_ctx[col] == 1);

  1615.                                 } else {

  1616.                                     c = 3 * (s->left_ref_ctx[row7] == 1) +

  1617.                                     (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1618.                                 }

  1619.                             } else if (s->left_comp_ctx[row7]) {

  1620.                                 if (!s->above_ref_ctx[col]) {

  1621.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1622.                                                  s->left_ref_ctx[row7] == 1);

  1623.                                 } else {

  1624.                                     c = 3 * (s->above_ref_ctx[col] == 1) +

  1625.                                     (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1626.                                 }

  1627.                             } else if (!s->above_ref_ctx[col]) {

  1628.                                 if (!s->left_ref_ctx[row7]) {

  1629.                                     c = 3;

  1630.                                 } else {

  1631.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1632.                                 }

  1633.                             } else if (!s->left_ref_ctx[row7]) {

  1634.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1635.                             } else {

  1636.                                 c = 2 * (s->left_ref_ctx[row7] == 1) +

  1637.                                 2 * (s->above_ref_ctx[col] == 1);

  1638.                             }

  1639.                         } else {

  1640.                             if (s->above_intra_ctx[col] ||

  1641.                                 (!s->above_comp_ctx[col] && !s->above_ref_ctx[col])) {

  1642.                                 c = 2;

  1643.                             } else if (s->above_comp_ctx[col]) {

  1644.                                 c = 3 * (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1645.                             } else {

  1646.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1647.                             }

  1648.                         }

  1649.                     } else if (have_l) {

  1650.                         if (s->left_intra_ctx[row7] ||

  1651.                             (!s->left_comp_ctx[row7] && !s->left_ref_ctx[row7])) {

  1652.                             c = 2;

  1653.                         } else if (s->left_comp_ctx[row7]) {

  1654.                             c = 3 * (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1655.                         } else {

  1656.                             c = 4 * (s->left_ref_ctx[row7] == 1);

  1657.                         }

  1658.                     } else {

  1659.                         c = 2;

  1660.                     }

  1661.                     bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][1]);

  1662.                     s->counts.single_ref[c][1][bit]++;

  1663.                     b->ref[0] = 1 + bit;

  1664.                 }

  1665.             }

  1666.         }

  1667. 

  1668.         if (b->bs <= BS_8x8) {

  1669.             if (s->segmentation.feat[b->seg_id].skip_enabled) {

  1670.                 b->mode[0] = b->mode[1] = b->mode[2] = b->mode[3] = ZEROMV;

  1671.             } else {

  1672.                 static const uint8_t off[10] = {

  1673.                     3, 0, 0, 1, 0, 0, 0, 0, 0, 0

  1674.                 };

  1675. 

  1676.                 // FIXME this needs to use the LUT tables from find_ref_mvs

  1677.                 // because not all are -1,0/0,-1

  1678.                 int c = inter_mode_ctx_lut[s->above_mode_ctx[col + off[b->bs]]]

  1679.                                           [s->left_mode_ctx[row7 + off[b->bs]]];

  1680. 

  1681.                 b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1682.                                               s->prob.p.mv_mode[c]);

  1683.                 b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1684.                 s->counts.mv_mode[c][b->mode[0] - 10]++;

  1685.             }

  1686.         }

  1687. 

  1688.         if (s->filtermode == FILTER_SWITCHABLE) {

  1689.             int c;

  1690. 

  1691.             if (have_a && s->above_mode_ctx[col] >= NEARESTMV) {

  1692.                 if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1693.                     c = s->above_filter_ctx[col] == s->left_filter_ctx[row7] ?

  1694.                         s->left_filter_ctx[row7] : 3;

  1695.                 } else {

  1696.                     c = s->above_filter_ctx[col];

  1697.                 }

  1698.             } else if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1699.                 c = s->left_filter_ctx[row7];

  1700.             } else {

  1701.                 c = 3;

  1702.             }

  1703. 

  1704.             b->filter = vp8_rac_get_tree(&s->c, vp9_filter_tree,

  1705.                                          s->prob.p.filter[c]);

  1706.             s->counts.filter[c][b->filter]++;

  1707.         } else {

  1708.             b->filter = s->filtermode;

  1709.         }

  1710. 

  1711.         if (b->bs > BS_8x8) {

  1712.             int c = inter_mode_ctx_lut[s->above_mode_ctx[col]][s->left_mode_ctx[row7]];

  1713. 

  1714.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1715.                                           s->prob.p.mv_mode[c]);

  1716.             s->counts.mv_mode[c][b->mode[0] - 10]++;

  1717.             fill_mv(s, b->mv[0], b->mode[0], 0);

  1718. 

  1719.             if (b->bs != BS_8x4) {

  1720.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1721.                                               s->prob.p.mv_mode[c]);

  1722.                 s->counts.mv_mode[c][b->mode[1] - 10]++;

  1723.                 fill_mv(s, b->mv[1], b->mode[1], 1);

  1724.             } else {

  1725.                 b->mode[1] = b->mode[0];

  1726.                 AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1727.                 AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1728.             }

  1729. 

  1730.             if (b->bs != BS_4x8) {

  1731.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1732.                                               s->prob.p.mv_mode[c]);

  1733.                 s->counts.mv_mode[c][b->mode[2] - 10]++;

  1734.                 fill_mv(s, b->mv[2], b->mode[2], 2);

  1735. 

  1736.                 if (b->bs != BS_8x4) {

  1737.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1738.                                                   s->prob.p.mv_mode[c]);

  1739.                     s->counts.mv_mode[c][b->mode[3] - 10]++;

  1740.                     fill_mv(s, b->mv[3], b->mode[3], 3);

  1741.                 } else {

  1742.                     b->mode[3] = b->mode[2];

  1743.                     AV_COPY32(&b->mv[3][0], &b->mv[2][0]);

  1744.                     AV_COPY32(&b->mv[3][1], &b->mv[2][1]);

  1745.                 }

  1746.             } else {

  1747.                 b->mode[2] = b->mode[0];

  1748.                 AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1749.                 AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1750.                 b->mode[3] = b->mode[1];

  1751.                 AV_COPY32(&b->mv[3][0], &b->mv[1][0]);

  1752.                 AV_COPY32(&b->mv[3][1], &b->mv[1][1]);

  1753.             }

  1754.         } else {

  1755.             fill_mv(s, b->mv[0], b->mode[0], -1);

  1756.             AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1757.             AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1758.             AV_COPY32(&b->mv[3][0], &b->mv[0][0]);

  1759.             AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1760.             AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1761.             AV_COPY32(&b->mv[3][1], &b->mv[0][1]);

  1762.         }

  1763.     }

  1764. 

  1765.     // FIXME this can probably be optimized

  1766.     memset(&s->above_skip_ctx[col], b->skip, w4);

  1767.     memset(&s->left_skip_ctx[row7], b->skip, h4);

  1768.     memset(&s->above_txfm_ctx[col], b->tx, w4);

  1769.     memset(&s->left_txfm_ctx[row7], b->tx, h4);

  1770.     memset(&s->above_partition_ctx[col], above_ctx[b->bs], w4);

  1771.     memset(&s->left_partition_ctx[row7], left_ctx[b->bs], h4);

  1772.     if (!s->keyframe && !s->intraonly) {

  1773.         memset(&s->above_intra_ctx[col], b->intra, w4);

  1774.         memset(&s->left_intra_ctx[row7], b->intra, h4);

  1775.         memset(&s->above_comp_ctx[col], b->comp, w4);

  1776.         memset(&s->left_comp_ctx[row7], b->comp, h4);

  1777.         memset(&s->above_mode_ctx[col], b->mode[3], w4);

  1778.         memset(&s->left_mode_ctx[row7], b->mode[3], h4);

  1779.         if (s->filtermode == FILTER_SWITCHABLE && !b->intra ) {

  1780.             memset(&s->above_filter_ctx[col], b->filter, w4);

  1781.             memset(&s->left_filter_ctx[row7], b->filter, h4);

  1782.             b->filter = vp9_filter_lut[b->filter];

  1783.         }

  1784.         if (b->bs > BS_8x8) {

  1785.             int mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  1786. 

  1787.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][0], &b->mv[1][0]);

  1788.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][1], &b->mv[1][1]);

  1789.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][0], mv0);

  1790.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][1], mv1);

  1791.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][0], &b->mv[2][0]);

  1792.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][1], &b->mv[2][1]);

  1793.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][0], mv0);

  1794.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][1], mv1);

  1795.         } else {

  1796.             int n, mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  1797. 

  1798.             for (n = 0; n < w4 * 2; n++) {

  1799.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][0], mv0);

  1800.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][1], mv1);

  1801.             }

  1802.             for (n = 0; n < h4 * 2; n++) {

  1803.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][0], mv0);

  1804.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][1], mv1);

  1805.             }

  1806.         }

  1807. 

  1808.         if (!b->intra) { // FIXME write 0xff or -1 if intra, so we can use this

  1809.                          // as a direct check in above branches

  1810.             int vref = b->ref[b->comp ? s->signbias[s->varcompref[0]] : 0];

  1811. 

  1812.             memset(&s->above_ref_ctx[col], vref, w4);

  1813.             memset(&s->left_ref_ctx[row7], vref, h4);

  1814.         }

  1815.     }

  1816. 

  1817.     // FIXME kinda ugly

  1818.     for (y = 0; y < h4; y++) {

  1819.         int x, o = (row + y) * s->sb_cols * 8 + col;

  1820.         struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[o];

  1821. 

  1822.         if (b->intra) {

  1823.             for (x = 0; x < w4; x++) {

  1824.                 mv[x].ref[0] =

  1825.                 mv[x].ref[1] = -1;

  1826.             }

  1827.         } else if (b->comp) {

  1828.             for (x = 0; x < w4; x++) {

  1829.                 mv[x].ref[0] = b->ref[0];

  1830.                 mv[x].ref[1] = b->ref[1];

  1831.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  1832.                 AV_COPY32(&mv[x].mv[1], &b->mv[3][1]);

  1833.             }

  1834.         } else {

  1835.             for (x = 0; x < w4; x++) {

  1836.                 mv[x].ref[0] = b->ref[0];

  1837.                 mv[x].ref[1] = -1;

  1838.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  1839.             }

  1840.         }

  1841.     }

  1842. }

  1843. 

  1844. // FIXME remove tx argument, and merge cnt/eob arguments?

  1845. static int decode_coeffs_b(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  1846.                            enum TxfmMode tx, unsigned (*cnt)[6][3],

  1847.                            unsigned (*eob)[6][2], uint8_t (*p)[6][11],

  1848.                            int nnz, const int16_t *scan, const int16_t (*nb)[2],

  1849.                            const int16_t *band_counts, const int16_t *qmul)

  1850. {

  1851.     int i = 0, band = 0, band_left = band_counts[band];

  1852.     uint8_t *tp = p[0][nnz];

  1853.     uint8_t cache[1024];

  1854. 

  1855.     do {

  1856.         int val, rc;

  1857. 

  1858.         val = vp56_rac_get_prob_branchy(c, tp[0]); // eob

  1859.         eob[band][nnz][val]++;

  1860.         if (!val)

  1861.             break;

  1862. 

  1863.     skip_eob:

  1864.         if (!vp56_rac_get_prob_branchy(c, tp[1])) { // zero

  1865.             cnt[band][nnz][0]++;

  1866.             if (!--band_left)

  1867.                 band_left = band_counts[++band];

  1868.             cache[scan[i]] = 0;

  1869.             nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  1870.             tp = p[band][nnz];

  1871.             if (++i == n_coeffs)

  1872.                 break; //invalid input; blocks should end with EOB

  1873.             goto skip_eob;

  1874.         }

  1875. 

  1876.         rc = scan[i];

  1877.         if (!vp56_rac_get_prob_branchy(c, tp[2])) { // one

  1878.             cnt[band][nnz][1]++;

  1879.             val = 1;

  1880.             cache[rc] = 1;

  1881.         } else {

  1882.             // fill in p[3-10] (model fill) - only once per frame for each pos

  1883.             if (!tp[3])

  1884.                 memcpy(&tp[3], vp9_model_pareto8[tp[2]], 8);

  1885. 

  1886.             cnt[band][nnz][2]++;

  1887.             if (!vp56_rac_get_prob_branchy(c, tp[3])) { // 2, 3, 4

  1888.                 if (!vp56_rac_get_prob_branchy(c, tp[4])) {

  1889.                     cache[rc] = val = 2;

  1890.                 } else {

  1891.                     val = 3 + vp56_rac_get_prob(c, tp[5]);

  1892.                     cache[rc] = 3;

  1893.                 }

  1894.             } else if (!vp56_rac_get_prob_branchy(c, tp[6])) { // cat1/2

  1895.                 cache[rc] = 4;

  1896.                 if (!vp56_rac_get_prob_branchy(c, tp[7])) {

  1897.                     val = 5 + vp56_rac_get_prob(c, 159);

  1898.                 } else {

  1899.                     val  = 7 + (vp56_rac_get_prob(c, 165) << 1);

  1900.                     val +=      vp56_rac_get_prob(c, 145);

  1901.                 }

  1902.             } else { // cat 3-6

  1903.                 cache[rc] = 5;

  1904.                 if (!vp56_rac_get_prob_branchy(c, tp[8])) {

  1905.                     if (!vp56_rac_get_prob_branchy(c, tp[9])) {

  1906.                         val  = 11 + (vp56_rac_get_prob(c, 173) << 2);

  1907.                         val +=      (vp56_rac_get_prob(c, 148) << 1);

  1908.                         val +=       vp56_rac_get_prob(c, 140);

  1909.                     } else {

  1910.                         val  = 19 + (vp56_rac_get_prob(c, 176) << 3);

  1911.                         val +=      (vp56_rac_get_prob(c, 155) << 2);

  1912.                         val +=      (vp56_rac_get_prob(c, 140) << 1);

  1913.                         val +=       vp56_rac_get_prob(c, 135);

  1914.                     }

  1915.                 } else if (!vp56_rac_get_prob_branchy(c, tp[10])) {

  1916.                     val  = 35 + (vp56_rac_get_prob(c, 180) << 4);

  1917.                     val +=      (vp56_rac_get_prob(c, 157) << 3);

  1918.                     val +=      (vp56_rac_get_prob(c, 141) << 2);

  1919.                     val +=      (vp56_rac_get_prob(c, 134) << 1);

  1920.                     val +=       vp56_rac_get_prob(c, 130);

  1921.                 } else {

  1922.                     val  = 67 + (vp56_rac_get_prob(c, 254) << 13);

  1923.                     val +=      (vp56_rac_get_prob(c, 254) << 12);

  1924.                     val +=      (vp56_rac_get_prob(c, 254) << 11);

  1925.                     val +=      (vp56_rac_get_prob(c, 252) << 10);

  1926.                     val +=      (vp56_rac_get_prob(c, 249) << 9);

  1927.                     val +=      (vp56_rac_get_prob(c, 243) << 8);

  1928.                     val +=      (vp56_rac_get_prob(c, 230) << 7);

  1929.                     val +=      (vp56_rac_get_prob(c, 196) << 6);

  1930.                     val +=      (vp56_rac_get_prob(c, 177) << 5);

  1931.                     val +=      (vp56_rac_get_prob(c, 153) << 4);

  1932.                     val +=      (vp56_rac_get_prob(c, 140) << 3);

  1933.                     val +=      (vp56_rac_get_prob(c, 133) << 2);

  1934.                     val +=      (vp56_rac_get_prob(c, 130) << 1);

  1935.                     val +=       vp56_rac_get_prob(c, 129);

  1936.                 }

  1937.             }

  1938.         }

  1939.         if (!--band_left)

  1940.             band_left = band_counts[++band];

  1941.         if (tx == TX_32X32) // FIXME slow

  1942.             coef[rc] = ((vp8_rac_get(c) ? -val : val) * qmul[!!i]) / 2;

  1943.         else

  1944.             coef[rc] = (vp8_rac_get(c) ? -val : val) * qmul[!!i];

  1945.         nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  1946.         tp = p[band][nnz];

  1947.     } while (++i < n_coeffs);

  1948. 

  1949.     return i;

  1950. }

  1951. 

  1952. static int decode_coeffs(AVCodecContext *ctx)

  1953. {

  1954.     VP9Context *s = ctx->priv_data;

  1955.     VP9Block *const b = &s->b;

  1956.     int row = s->row, col = s->col;

  1957.     uint8_t (*p)[6][11] = s->prob.coef[b->tx][0 /* y */][!b->intra];

  1958.     unsigned (*c)[6][3] = s->counts.coef[b->tx][0 /* y */][!b->intra];

  1959.     unsigned (*e)[6][2] = s->counts.eob[b->tx][0 /* y */][!b->intra];

  1960.     int w4 = bwh_tab[1][b->bs][0] << 1, h4 = bwh_tab[1][b->bs][1] << 1;

  1961.     int end_x = FFMIN(2 * (s->cols - col), w4);

  1962.     int end_y = FFMIN(2 * (s->rows - row), h4);

  1963.     int n, pl, x, y, step1d = 1 << b->tx, step = 1 << (b->tx * 2);

  1964.     int uvstep1d = 1 << b->uvtx, uvstep = 1 << (b->uvtx * 2), res;

  1965.     int16_t (*qmul)[2] = s->segmentation.feat[b->seg_id].qmul;

  1966.     int tx = 4 * s->lossless + b->tx;

  1967.     const int16_t * const *yscans = vp9_scans[tx];

  1968.     const int16_t (* const *ynbs)[2] = vp9_scans_nb[tx];

  1969.     const int16_t *uvscan = vp9_scans[b->uvtx][DCT_DCT];

  1970.     const int16_t (*uvnb)[2] = vp9_scans_nb[b->uvtx][DCT_DCT];

  1971.     uint8_t *a = &s->above_y_nnz_ctx[col * 2];

  1972.     uint8_t *l = &s->left_y_nnz_ctx[(row & 7) << 1];

  1973.     static const int16_t band_counts[4][8] = {

  1974.         { 1, 2, 3, 4,  3,   16 - 13 },

  1975.         { 1, 2, 3, 4, 11,   64 - 21 },

  1976.         { 1, 2, 3, 4, 11,  256 - 21 },

  1977.         { 1, 2, 3, 4, 11, 1024 - 21 },

  1978.     };

  1979.     const int16_t *y_band_counts = band_counts[b->tx];

  1980.     const int16_t *uv_band_counts = band_counts[b->uvtx];

  1981. 

  1982.     /* y tokens */

  1983.     if (b->tx > TX_4X4) { // FIXME slow

  1984.         for (y = 0; y < end_y; y += step1d)

  1985.             for (x = 1; x < step1d; x++)

  1986.                 l[y] |= l[y + x];

  1987.         for (x = 0; x < end_x; x += step1d)

  1988.             for (y = 1; y < step1d; y++)

  1989.                 a[x] |= a[x + y];

  1990.     }

  1991.     for (n = 0, y = 0; y < end_y; y += step1d) {

  1992.         for (x = 0; x < end_x; x += step1d, n += step) {

  1993.             enum TxfmType txtp = vp9_intra_txfm_type[b->mode[b->tx == TX_4X4 &&

  1994.                                                              b->bs > BS_8x8 ?

  1995.                                                              n : 0]];

  1996.             int nnz = a[x] + l[y];

  1997.             if ((res = decode_coeffs_b(&s->c, s->block + 16 * n, 16 * step,

  1998.                                        b->tx, c, e, p, nnz, yscans[txtp],

  1999.                                        ynbs[txtp], y_band_counts, qmul[0])) < 0)

  2000.                 return res;

  2001.             a[x] = l[y] = !!res;

  2002.             if (b->tx > TX_8X8) {

  2003.                 AV_WN16A(&s->eob[n], res);

  2004.             } else {

  2005.                 s->eob[n] = res;

  2006.             }

  2007.         }

  2008.     }

  2009.     if (b->tx > TX_4X4) { // FIXME slow

  2010.         for (y = 0; y < end_y; y += step1d)

  2011.             memset(&l[y + 1], l[y], FFMIN(end_y - y - 1, step1d - 1));

  2012.         for (x = 0; x < end_x; x += step1d)

  2013.             memset(&a[x + 1], a[x], FFMIN(end_x - x - 1, step1d - 1));

  2014.     }

  2015. 

  2016.     p = s->prob.coef[b->uvtx][1 /* uv */][!b->intra];

  2017.     c = s->counts.coef[b->uvtx][1 /* uv */][!b->intra];

  2018.     e = s->counts.eob[b->uvtx][1 /* uv */][!b->intra];

  2019.     w4 >>= 1;

  2020.     h4 >>= 1;

  2021.     end_x >>= 1;

  2022.     end_y >>= 1;

  2023.     for (pl = 0; pl < 2; pl++) {

  2024.         a = &s->above_uv_nnz_ctx[pl][col];

  2025.         l = &s->left_uv_nnz_ctx[pl][row & 7];

  2026.         if (b->uvtx > TX_4X4) { // FIXME slow

  2027.             for (y = 0; y < end_y; y += uvstep1d)

  2028.                 for (x = 1; x < uvstep1d; x++)

  2029.                     l[y] |= l[y + x];

  2030.             for (x = 0; x < end_x; x += uvstep1d)

  2031.                 for (y = 1; y < uvstep1d; y++)

  2032.                     a[x] |= a[x + y];

  2033.         }

  2034.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2035.             for (x = 0; x < end_x; x += uvstep1d, n += uvstep) {

  2036.                 int nnz = a[x] + l[y];

  2037.                 if ((res = decode_coeffs_b(&s->c, s->uvblock[pl] + 16 * n,

  2038.                                            16 * uvstep, b->uvtx, c, e, p, nnz,

  2039.                                            uvscan, uvnb, uv_band_counts,

  2040.                                            qmul[1])) < 0)

  2041.                     return res;

  2042.                 a[x] = l[y] = !!res;

  2043.                 if (b->uvtx > TX_8X8) {

  2044.                     AV_WN16A(&s->uveob[pl][n], res);

  2045.                 } else {

  2046.                     s->uveob[pl][n] = res;

  2047.                 }

  2048.             }

  2049.         }

  2050.         if (b->uvtx > TX_4X4) { // FIXME slow

  2051.             for (y = 0; y < end_y; y += uvstep1d)

  2052.                 memset(&l[y + 1], l[y], FFMIN(end_y - y - 1, uvstep1d - 1));

  2053.             for (x = 0; x < end_x; x += uvstep1d)

  2054.                 memset(&a[x + 1], a[x], FFMIN(end_x - x - 1, uvstep1d - 1));

  2055.         }

  2056.     }

  2057. 

  2058.     return 0;

  2059. }

  2060. 

  2061. static av_always_inline int check_intra_mode(VP9Context *s, int mode, uint8_t **a,

  2062.                                              uint8_t *dst_edge, ptrdiff_t stride_edge,

  2063.                                              uint8_t *dst_inner, ptrdiff_t stride_inner,

  2064.                                              uint8_t *l, int col, int x, int w,

  2065.                                              int row, int y, enum TxfmMode tx,

  2066.                                              int p)

  2067. {

  2068.     int have_top = row > 0 || y > 0;

  2069.     int have_left = col > s->tiling.tile_col_start || x > 0;

  2070.     int have_right = x < w - 1;

  2071.     static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {

  2072.         [VERT_PRED]            = { { DC_127_PRED,          VERT_PRED },

  2073.                                    { DC_127_PRED,          VERT_PRED } },

  2074.         [HOR_PRED]             = { { DC_129_PRED,          DC_129_PRED },

  2075.                                    { HOR_PRED,             HOR_PRED } },

  2076.         [DC_PRED]              = { { DC_128_PRED,          TOP_DC_PRED },

  2077.                                    { LEFT_DC_PRED,         DC_PRED } },

  2078.         [DIAG_DOWN_LEFT_PRED]  = { { DC_127_PRED,          DIAG_DOWN_LEFT_PRED },

  2079.                                    { DC_127_PRED,          DIAG_DOWN_LEFT_PRED } },

  2080.         [DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },

  2081.                                    { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },

  2082.         [VERT_RIGHT_PRED]      = { { VERT_RIGHT_PRED,      VERT_RIGHT_PRED },

  2083.                                    { VERT_RIGHT_PRED,      VERT_RIGHT_PRED } },

  2084.         [HOR_DOWN_PRED]        = { { HOR_DOWN_PRED,        HOR_DOWN_PRED },

  2085.                                    { HOR_DOWN_PRED,        HOR_DOWN_PRED } },

  2086.         [VERT_LEFT_PRED]       = { { DC_127_PRED,          VERT_LEFT_PRED },

  2087.                                    { DC_127_PRED,          VERT_LEFT_PRED } },

  2088.         [HOR_UP_PRED]          = { { DC_129_PRED,          DC_129_PRED },

  2089.                                    { HOR_UP_PRED,          HOR_UP_PRED } },

  2090.         [TM_VP8_PRED]          = { { DC_129_PRED,          VERT_PRED },

  2091.                                    { HOR_PRED,             TM_VP8_PRED } },

  2092.     };

  2093.     static const struct {

  2094.         uint8_t needs_left:1;

  2095.         uint8_t needs_top:1;

  2096.         uint8_t needs_topleft:1;

  2097.         uint8_t needs_topright:1;

  2098.     } edges[N_INTRA_PRED_MODES] = {

  2099.         [VERT_PRED]            = { .needs_top  = 1 },

  2100.         [HOR_PRED]             = { .needs_left = 1 },

  2101.         [DC_PRED]              = { .needs_top  = 1, .needs_left = 1 },

  2102.         [DIAG_DOWN_LEFT_PRED]  = { .needs_top  = 1, .needs_topright = 1 },

  2103.         [DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2104.         [VERT_RIGHT_PRED]      = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2105.         [HOR_DOWN_PRED]        = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2106.         [VERT_LEFT_PRED]       = { .needs_top  = 1, .needs_topright = 1 },

  2107.         [HOR_UP_PRED]          = { .needs_left = 1 },

  2108.         [TM_VP8_PRED]          = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2109.         [LEFT_DC_PRED]         = { .needs_left = 1 },

  2110.         [TOP_DC_PRED]          = { .needs_top  = 1 },

  2111.         [DC_128_PRED]          = { 0 },

  2112.         [DC_127_PRED]          = { 0 },

  2113.         [DC_129_PRED]          = { 0 }

  2114.     };

  2115. 

  2116.     av_assert2(mode >= 0 && mode < 10);

  2117.     mode = mode_conv[mode][have_left][have_top];

  2118.     if (edges[mode].needs_top) {

  2119.         uint8_t *top, *topleft;

  2120.         int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !p) - x) * 4;

  2121.         int n_px_need_tr = 0;

  2122. 

  2123.         if (tx == TX_4X4 && edges[mode].needs_topright && have_right)

  2124.             n_px_need_tr = 4;

  2125. 

  2126.         // if top of sb64-row, use s->intra_pred_data[] instead of

  2127.         // dst[-stride] for intra prediction (it contains pre- instead of

  2128.         // post-loopfilter data)

  2129.         if (have_top) {

  2130.             top = !(row & 7) && !y ?

  2131.                 s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  2132.                 y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];

  2133.             if (have_left)

  2134.                 topleft = !(row & 7) && !y ?

  2135.                     s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  2136.                     y == 0 || x == 0 ? &dst_edge[-stride_edge] :

  2137.                     &dst_inner[-stride_inner];

  2138.         }

  2139. 

  2140.         if (have_top &&

  2141.             (!edges[mode].needs_topleft || (have_left && top == topleft)) &&

  2142.             (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&

  2143.             n_px_need + n_px_need_tr <= n_px_have) {

  2144.             *a = top;

  2145.         } else {

  2146.             if (have_top) {

  2147.                 if (n_px_need <= n_px_have) {

  2148.                     memcpy(*a, top, n_px_need);

  2149.                 } else {

  2150.                     memcpy(*a, top, n_px_have);

  2151.                     memset(&(*a)[n_px_have], (*a)[n_px_have - 1],

  2152.                            n_px_need - n_px_have);

  2153.                 }

  2154.             } else {

  2155.                 memset(*a, 127, n_px_need);

  2156.             }

  2157.             if (edges[mode].needs_topleft) {

  2158.                 if (have_left && have_top) {

  2159.                     (*a)[-1] = topleft[-1];

  2160.                 } else {

  2161.                     (*a)[-1] = have_top ? 129 : 127;

  2162.                 }

  2163.             }

  2164.             if (tx == TX_4X4 && edges[mode].needs_topright) {

  2165.                 if (have_top && have_right &&

  2166.                     n_px_need + n_px_need_tr <= n_px_have) {

  2167.                     memcpy(&(*a)[4], &top[4], 4);

  2168.                 } else {

  2169.                     memset(&(*a)[4], (*a)[3], 4);

  2170.                 }

  2171.             }

  2172.         }

  2173.     }

  2174.     if (edges[mode].needs_left) {

  2175.         if (have_left) {

  2176.             int n_px_need = 4 << tx, i, n_px_have = (((s->rows - row) << !p) - y) * 4;

  2177.             uint8_t *dst = x == 0 ? dst_edge : dst_inner;

  2178.             ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;

  2179. 

  2180.             if (n_px_need <= n_px_have) {

  2181.                 for (i = 0; i < n_px_need; i++)

  2182.                     l[i] = dst[i * stride - 1];

  2183.             } else {

  2184.                 for (i = 0; i < n_px_have; i++)

  2185.                     l[i] = dst[i * stride - 1];

  2186.                 memset(&l[i], l[i - 1], n_px_need - n_px_have);

  2187.             }

  2188.         } else {

  2189.             memset(l, 129, 4 << tx);

  2190.         }

  2191.     }

  2192. 

  2193.     return mode;

  2194. }

  2195. 

  2196. static void intra_recon(AVCodecContext *ctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  2197. {

  2198.     VP9Context *s = ctx->priv_data;

  2199.     VP9Block *const b = &s->b;

  2200.     int row = s->row, col = s->col;

  2201.     int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2202.     int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2203.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2204.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2205.     int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2206.     int uvstep1d = 1 << b->uvtx, p;

  2207.     uint8_t *dst = s->dst[0], *dst_r = s->frames[CUR_FRAME].tf.f->data[0] + y_off;

  2208. 

  2209.     for (n = 0, y = 0; y < end_y; y += step1d) {

  2210.         uint8_t *ptr = dst, *ptr_r = dst_r;

  2211.         for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d,

  2212.                                ptr_r += 4 * step1d, n += step) {

  2213.             int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?

  2214.                                y * 2 + x : 0];

  2215.             LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);

  2216.             uint8_t *a = &a_buf[16], l[32];

  2217.             enum TxfmType txtp = vp9_intra_txfm_type[mode];

  2218.             int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2219. 

  2220.             mode = check_intra_mode(s, mode, &a, ptr_r,

  2221.                                     s->frames[CUR_FRAME].tf.f->linesize[0],

  2222.                                     ptr, s->y_stride, l,

  2223.                                     col, x, w4, row, y, b->tx, 0);

  2224.             s->dsp.intra_pred[b->tx][mode](ptr, s->y_stride, l, a);

  2225.             if (eob)

  2226.                 s->dsp.itxfm_add[tx][txtp](ptr, s->y_stride,

  2227.                                            s->block + 16 * n, eob);

  2228.         }

  2229.         dst_r += 4 * step1d * s->frames[CUR_FRAME].tf.f->linesize[0];

  2230.         dst   += 4 * step1d * s->y_stride;

  2231.     }

  2232. 

  2233.     // U/V

  2234.     h4 >>= 1;

  2235.     w4 >>= 1;

  2236.     end_x >>= 1;

  2237.     end_y >>= 1;

  2238.     step = 1 << (b->uvtx * 2);

  2239.     for (p = 0; p < 2; p++) {

  2240.         dst   = s->dst[1 + p];

  2241.         dst_r = s->frames[CUR_FRAME].tf.f->data[1 + p] + uv_off;

  2242.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2243.             uint8_t *ptr = dst, *ptr_r = dst_r;

  2244.             for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d,

  2245.                                    ptr_r += 4 * uvstep1d, n += step) {

  2246.                 int mode = b->uvmode;

  2247.                 LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);

  2248.                 uint8_t *a = &a_buf[16], l[32];

  2249.                 int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2250. 

  2251.                 mode = check_intra_mode(s, mode, &a, ptr_r,

  2252.                                         s->frames[CUR_FRAME].tf.f->linesize[1],

  2253.                                         ptr, s->uv_stride, l,

  2254.                                         col, x, w4, row, y, b->uvtx, p + 1);

  2255.                 s->dsp.intra_pred[b->uvtx][mode](ptr, s->uv_stride, l, a);

  2256.                 if (eob)

  2257.                     s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2258.                                                     s->uvblock[p] + 16 * n, eob);

  2259.             }

  2260.             dst_r += 4 * uvstep1d * s->frames[CUR_FRAME].tf.f->linesize[1];

  2261.             dst   += 4 * uvstep1d * s->uv_stride;

  2262.         }

  2263.     }

  2264. }

  2265. 

  2266. static av_always_inline void mc_luma_dir(VP9Context *s, vp9_mc_func (*mc)[2],

  2267.                                          uint8_t *dst, ptrdiff_t dst_stride,

  2268.                                          const uint8_t *ref, ptrdiff_t ref_stride,

  2269.                                          ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2270.                                          int bw, int bh, int w, int h)

  2271. {

  2272.     int mx = mv->x, my = mv->y;

  2273. 

  2274.     y += my >> 3;

  2275.     x += mx >> 3;

  2276.     ref += y * ref_stride + x;

  2277.     mx &= 7;

  2278.     my &= 7;

  2279.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2280.     if (x < !!mx * 3 || y < !!my * 3 ||

  2281.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2282.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2283.                                  ref - !!my * 3 * ref_stride - !!mx * 3,

  2284.                                  80, ref_stride,

  2285.                                  bw + !!mx * 7, bh + !!my * 7,

  2286.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2287.         ref = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2288.         ref_stride = 80;

  2289.     }

  2290.     mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);

  2291. }

  2292. 

  2293. static av_always_inline void mc_chroma_dir(VP9Context *s, vp9_mc_func (*mc)[2],

  2294.                                            uint8_t *dst_u, uint8_t *dst_v,

  2295.                                            ptrdiff_t dst_stride,

  2296.                                            const uint8_t *ref_u, ptrdiff_t src_stride_u,

  2297.                                            const uint8_t *ref_v, ptrdiff_t src_stride_v,

  2298.                                            ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2299.                                            int bw, int bh, int w, int h)

  2300. {

  2301.     int mx = mv->x, my = mv->y;

  2302. 

  2303.     y += my >> 4;

  2304.     x += mx >> 4;

  2305.     ref_u += y * src_stride_u + x;

  2306.     ref_v += y * src_stride_v + x;

  2307.     mx &= 15;

  2308.     my &= 15;

  2309.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2310.     if (x < !!mx * 3 || y < !!my * 3 ||

  2311.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2312.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2313.                                  ref_u - !!my * 3 * src_stride_u - !!mx * 3,

  2314.                                  80, src_stride_u,

  2315.                                  bw + !!mx * 7, bh + !!my * 7,

  2316.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2317.         ref_u = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2318.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, 80, bh, mx, my);

  2319. 

  2320.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2321.                                  ref_v - !!my * 3 * src_stride_v - !!mx * 3,

  2322.                                  80, src_stride_v,

  2323.                                  bw + !!mx * 7, bh + !!my * 7,

  2324.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2325.         ref_v = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2326.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, 80, bh, mx, my);

  2327.     } else {

  2328.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);

  2329.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);

  2330.     }

  2331. }

  2332. 

  2333. static void inter_recon(AVCodecContext *ctx)

  2334. {

  2335.     static const uint8_t bwlog_tab[2][N_BS_SIZES] = {

  2336.         { 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4 },

  2337.         { 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 4, 4 },

  2338.     };

  2339.     VP9Context *s = ctx->priv_data;

  2340.     VP9Block *const b = &s->b;

  2341.     int row = s->row, col = s->col;

  2342.     ThreadFrame *tref1 = &s->refs[s->refidx[b->ref[0]]];

  2343.     AVFrame *ref1 = tref1->f;

  2344.     ThreadFrame *tref2 = b->comp ? &s->refs[s->refidx[b->ref[1]]] : NULL;

  2345.     AVFrame *ref2 = b->comp ? tref2->f : NULL;

  2346.     int w = ctx->width, h = ctx->height;

  2347.     ptrdiff_t ls_y = s->y_stride, ls_uv = s->uv_stride;

  2348. 

  2349.     // y inter pred

  2350.     if (b->bs > BS_8x8) {

  2351.         if (b->bs == BS_8x4) {

  2352.             mc_luma_dir(s, s->dsp.mc[3][b->filter][0], s->dst[0], ls_y,

  2353.                         ref1->data[0], ref1->linesize[0],

  2354.                         row << 3, col << 3, &b->mv[0][0], 8, 4, w, h);

  2355.             mc_luma_dir(s, s->dsp.mc[3][b->filter][0],

  2356.                         s->dst[0] + 4 * ls_y, ls_y,

  2357.                         ref1->data[0], ref1->linesize[0],

  2358.                         (row << 3) + 4, col << 3, &b->mv[2][0], 8, 4, w, h);

  2359. 

  2360.             if (b->comp) {

  2361.                 mc_luma_dir(s, s->dsp.mc[3][b->filter][1], s->dst[0], ls_y,

  2362.                             ref2->data[0], ref2->linesize[0],

  2363.                             row << 3, col << 3, &b->mv[0][1], 8, 4, w, h);

  2364.                 mc_luma_dir(s, s->dsp.mc[3][b->filter][1],

  2365.                             s->dst[0] + 4 * ls_y, ls_y,

  2366.                             ref2->data[0], ref2->linesize[0],

  2367.                             (row << 3) + 4, col << 3, &b->mv[2][1], 8, 4, w, h);

  2368.             }

  2369.         } else if (b->bs == BS_4x8) {

  2370.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0], ls_y,

  2371.                         ref1->data[0], ref1->linesize[0],

  2372.                         row << 3, col << 3, &b->mv[0][0], 4, 8, w, h);

  2373.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0] + 4, ls_y,

  2374.                         ref1->data[0], ref1->linesize[0],

  2375.                         row << 3, (col << 3) + 4, &b->mv[1][0], 4, 8, w, h);

  2376. 

  2377.             if (b->comp) {

  2378.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0], ls_y,

  2379.                             ref2->data[0], ref2->linesize[0],

  2380.                             row << 3, col << 3, &b->mv[0][1], 4, 8, w, h);

  2381.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0] + 4, ls_y,

  2382.                             ref2->data[0], ref2->linesize[0],

  2383.                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 8, w, h);

  2384.             }

  2385.         } else {

  2386.             av_assert2(b->bs == BS_4x4);

  2387. 

  2388.             // FIXME if two horizontally adjacent blocks have the same MV,

  2389.             // do a w8 instead of a w4 call

  2390.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0], ls_y,

  2391.                         ref1->data[0], ref1->linesize[0],

  2392.                         row << 3, col << 3, &b->mv[0][0], 4, 4, w, h);

  2393.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0] + 4, ls_y,

  2394.                         ref1->data[0], ref1->linesize[0],

  2395.                         row << 3, (col << 3) + 4, &b->mv[1][0], 4, 4, w, h);

  2396.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0],

  2397.                         s->dst[0] + 4 * ls_y, ls_y,

  2398.                         ref1->data[0], ref1->linesize[0],

  2399.                         (row << 3) + 4, col << 3, &b->mv[2][0], 4, 4, w, h);

  2400.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0],

  2401.                         s->dst[0] + 4 * ls_y + 4, ls_y,

  2402.                         ref1->data[0], ref1->linesize[0],

  2403.                         (row << 3) + 4, (col << 3) + 4, &b->mv[3][0], 4, 4, w, h);

  2404. 

  2405.             if (b->comp) {

  2406.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0], ls_y,

  2407.                             ref2->data[0], ref2->linesize[0],

  2408.                             row << 3, col << 3, &b->mv[0][1], 4, 4, w, h);

  2409.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0] + 4, ls_y,

  2410.                             ref2->data[0], ref2->linesize[0],

  2411.                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 4, w, h);

  2412.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1],

  2413.                             s->dst[0] + 4 * ls_y, ls_y,

  2414.                             ref2->data[0], ref2->linesize[0],

  2415.                             (row << 3) + 4, col << 3, &b->mv[2][1], 4, 4, w, h);

  2416.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1],

  2417.                             s->dst[0] + 4 * ls_y + 4, ls_y,

  2418.                             ref2->data[0], ref2->linesize[0],

  2419.                             (row << 3) + 4, (col << 3) + 4, &b->mv[3][1], 4, 4, w, h);

  2420.             }

  2421.         }

  2422.     } else {

  2423.         int bwl = bwlog_tab[0][b->bs];

  2424.         int bw = bwh_tab[0][b->bs][0] * 4, bh = bwh_tab[0][b->bs][1] * 4;

  2425. 

  2426.         mc_luma_dir(s, s->dsp.mc[bwl][b->filter][0], s->dst[0], ls_y,

  2427.                     ref1->data[0], ref1->linesize[0],

  2428.                     row << 3, col << 3, &b->mv[0][0],bw, bh, w, h);

  2429. 

  2430.         if (b->comp)

  2431.             mc_luma_dir(s, s->dsp.mc[bwl][b->filter][1], s->dst[0], ls_y,

  2432.                         ref2->data[0], ref2->linesize[0],

  2433.                         row << 3, col << 3, &b->mv[0][1], bw, bh, w, h);

  2434.     }

  2435. 

  2436.     // uv inter pred

  2437.     {

  2438.         int bwl = bwlog_tab[1][b->bs];

  2439.         int bw = bwh_tab[1][b->bs][0] * 4, bh = bwh_tab[1][b->bs][1] * 4;

  2440.         VP56mv mvuv;

  2441. 

  2442.         w = (w + 1) >> 1;

  2443.         h = (h + 1) >> 1;

  2444.         if (b->bs > BS_8x8) {

  2445.             mvuv.x = ROUNDED_DIV(b->mv[0][0].x + b->mv[1][0].x + b->mv[2][0].x + b->mv[3][0].x, 4);

  2446.             mvuv.y = ROUNDED_DIV(b->mv[0][0].y + b->mv[1][0].y + b->mv[2][0].y + b->mv[3][0].y, 4);

  2447.         } else {

  2448.             mvuv = b->mv[0][0];

  2449.         }

  2450. 

  2451.         mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][0],

  2452.                       s->dst[1], s->dst[2], ls_uv,

  2453.                       ref1->data[1], ref1->linesize[1],

  2454.                       ref1->data[2], ref1->linesize[2],

  2455.                       row << 2, col << 2, &mvuv, bw, bh, w, h);

  2456. 

  2457.         if (b->comp) {

  2458.             if (b->bs > BS_8x8) {

  2459.                 mvuv.x = ROUNDED_DIV(b->mv[0][1].x + b->mv[1][1].x + b->mv[2][1].x + b->mv[3][1].x, 4);

  2460.                 mvuv.y = ROUNDED_DIV(b->mv[0][1].y + b->mv[1][1].y + b->mv[2][1].y + b->mv[3][1].y, 4);

  2461.             } else {

  2462.                 mvuv = b->mv[0][1];

  2463.             }

  2464.             mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][1],

  2465.                           s->dst[1], s->dst[2], ls_uv,

  2466.                           ref2->data[1], ref2->linesize[1],

  2467.                           ref2->data[2], ref2->linesize[2],

  2468.                           row << 2, col << 2, &mvuv, bw, bh, w, h);

  2469.         }

  2470.     }

  2471. 

  2472.     if (!b->skip) {

  2473.         /* mostly copied intra_reconn() */

  2474. 

  2475.         int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2476.         int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2477.         int end_x = FFMIN(2 * (s->cols - col), w4);

  2478.         int end_y = FFMIN(2 * (s->rows - row), h4);

  2479.         int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2480.         int uvstep1d = 1 << b->uvtx, p;

  2481.         uint8_t *dst = s->dst[0];

  2482. 

  2483.         // y itxfm add

  2484.         for (n = 0, y = 0; y < end_y; y += step1d) {

  2485.             uint8_t *ptr = dst;

  2486.             for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d, n += step) {

  2487.                 int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2488. 

  2489.                 if (eob)

  2490.                     s->dsp.itxfm_add[tx][DCT_DCT](ptr, s->y_stride,

  2491.                                                   s->block + 16 * n, eob);

  2492.             }

  2493.             dst += 4 * s->y_stride * step1d;

  2494.         }

  2495. 

  2496.         // uv itxfm add

  2497.         h4 >>= 1;

  2498.         w4 >>= 1;

  2499.         end_x >>= 1;

  2500.         end_y >>= 1;

  2501.         step = 1 << (b->uvtx * 2);

  2502.         for (p = 0; p < 2; p++) {

  2503.             dst = s->dst[p + 1];

  2504.             for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2505.                 uint8_t *ptr = dst;

  2506.                 for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d, n += step) {

  2507.                     int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2508. 

  2509.                     if (eob)

  2510.                         s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2511.                                                         s->uvblock[p] + 16 * n, eob);

  2512.                 }

  2513.                 dst += 4 * uvstep1d * s->uv_stride;

  2514.             }

  2515.         }

  2516.     }

  2517. }

  2518. 

  2519. static av_always_inline void mask_edges(struct VP9Filter *lflvl, int is_uv,

  2520.                                         int row_and_7, int col_and_7,

  2521.                                         int w, int h, int col_end, int row_end,

  2522.                                         enum TxfmMode tx, int skip_inter)

  2523. {

  2524.     // FIXME I'm pretty sure all loops can be replaced by a single LUT if

  2525.     // we make VP9Filter.mask uint64_t (i.e. row/col all single variable)

  2526.     // and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then

  2527.     // use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)

  2528. 

  2529.     // the intended behaviour of the vp9 loopfilter is to work on 8-pixel

  2530.     // edges. This means that for UV, we work on two subsampled blocks at

  2531.     // a time, and we only use the topleft block's mode information to set

  2532.     // things like block strength. Thus, for any block size smaller than

  2533.     // 16x16, ignore the odd portion of the block.

  2534.     if (tx == TX_4X4 && is_uv) {

  2535.         if (h == 1) {

  2536.             if (row_and_7 & 1)

  2537.                 return;

  2538.             if (!row_end)

  2539.                 h += 1;

  2540.         }

  2541.         if (w == 1) {

  2542.             if (col_and_7 & 1)

  2543.                 return;

  2544.             if (!col_end)

  2545.                 w += 1;

  2546.         }

  2547.     }

  2548. 

  2549.     if (tx == TX_4X4 && !skip_inter) {

  2550.         int t = 1 << col_and_7, m_col = (t << w) - t, y;

  2551.         int m_col_odd = (t << (w - 1)) - t;

  2552. 

  2553.         // on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide

  2554.         if (is_uv) {

  2555.             int m_row_8 = m_col & 0x01, m_row_4 = m_col - m_row_8;

  2556. 

  2557.             for (y = row_and_7; y < h + row_and_7; y++) {

  2558.                 int col_mask_id = 2 - !(y & 7);

  2559. 

  2560.                 lflvl->mask[is_uv][0][y][1] |= m_row_8;

  2561.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  2562.                 // for odd lines, if the odd col is not being filtered,

  2563.                 // skip odd row also:

  2564.                 // .---. <-- a

  2565.                 // |   |

  2566.                 // |___| <-- b

  2567.                 // ^   ^

  2568.                 // c   d

  2569.                 //

  2570.                 // if a/c are even row/col and b/d are odd, and d is skipped,

  2571.                 // e.g. right edge of size-66x66.webm, then skip b also (bug)

  2572.                 if ((col_end & 1) && (y & 1)) {

  2573.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col_odd;

  2574.                 } else {

  2575.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col;

  2576.                 }

  2577.             }

  2578.         } else {

  2579.             int m_row_8 = m_col & 0x11, m_row_4 = m_col - m_row_8;

  2580. 

  2581.             for (y = row_and_7; y < h + row_and_7; y++) {

  2582.                 int col_mask_id = 2 - !(y & 3);

  2583. 

  2584.                 lflvl->mask[is_uv][0][y][1] |= m_row_8; // row edge

  2585.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  2586.                 lflvl->mask[is_uv][1][y][col_mask_id] |= m_col; // col edge

  2587.                 lflvl->mask[is_uv][0][y][3] |= m_col;

  2588.                 lflvl->mask[is_uv][1][y][3] |= m_col;

  2589.             }

  2590.         }

  2591.     } else {

  2592.         int y, t = 1 << col_and_7, m_col = (t << w) - t;

  2593. 

  2594.         if (!skip_inter) {

  2595.             int mask_id = (tx == TX_8X8);

  2596.             int l2 = tx + is_uv - 1, step1d = 1 << l2;

  2597.             static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };

  2598.             int m_row = m_col & masks[l2];

  2599. 

  2600.             // at odd UV col/row edges tx16/tx32 loopfilter edges, force

  2601.             // 8wd loopfilter to prevent going off the visible edge.

  2602.             if (is_uv && tx > TX_8X8 && (w ^ (w - 1)) == 1) {

  2603.                 int m_row_16 = ((t << (w - 1)) - t) & masks[l2];

  2604.                 int m_row_8 = m_row - m_row_16;

  2605. 

  2606.                 for (y = row_and_7; y < h + row_and_7; y++) {

  2607.                     lflvl->mask[is_uv][0][y][0] |= m_row_16;

  2608.                     lflvl->mask[is_uv][0][y][1] |= m_row_8;

  2609.                 }

  2610.             } else {

  2611.                 for (y = row_and_7; y < h + row_and_7; y++)

  2612.                     lflvl->mask[is_uv][0][y][mask_id] |= m_row;

  2613.             }

  2614. 

  2615.             if (is_uv && tx > TX_8X8 && (h ^ (h - 1)) == 1) {

  2616.                 for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)

  2617.                     lflvl->mask[is_uv][1][y][0] |= m_col;

  2618.                 if (y - row_and_7 == h - 1)

  2619.                     lflvl->mask[is_uv][1][y][1] |= m_col;

  2620.             } else {

  2621.                 for (y = row_and_7; y < h + row_and_7; y += step1d)

  2622.                     lflvl->mask[is_uv][1][y][mask_id] |= m_col;

  2623.             }

  2624.         } else if (tx != TX_4X4) {

  2625.             int mask_id;

  2626. 

  2627.             mask_id = (tx == TX_8X8) || (is_uv && h == 1);

  2628.             lflvl->mask[is_uv][1][row_and_7][mask_id] |= m_col;

  2629.             mask_id = (tx == TX_8X8) || (is_uv && w == 1);

  2630.             for (y = row_and_7; y < h + row_and_7; y++)

  2631.                 lflvl->mask[is_uv][0][y][mask_id] |= t;

  2632.         } else if (is_uv) {

  2633.             int t8 = t & 0x01, t4 = t - t8;

  2634. 

  2635.             for (y = row_and_7; y < h + row_and_7; y++) {

  2636.                 lflvl->mask[is_uv][0][y][2] |= t4;

  2637.                 lflvl->mask[is_uv][0][y][1] |= t8;

  2638.             }

  2639.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 7)] |= m_col;

  2640.         } else {

  2641.             int t8 = t & 0x11, t4 = t - t8;

  2642. 

  2643.             for (y = row_and_7; y < h + row_and_7; y++) {

  2644.                 lflvl->mask[is_uv][0][y][2] |= t4;

  2645.                 lflvl->mask[is_uv][0][y][1] |= t8;

  2646.             }

  2647.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 3)] |= m_col;

  2648.         }

  2649.     }

  2650. }

  2651. 

  2652. static int decode_b(AVCodecContext *ctx, int row, int col,

  2653.                     struct VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,

  2654.                     enum BlockLevel bl, enum BlockPartition bp)

  2655. {

  2656.     VP9Context *s = ctx->priv_data;

  2657.     VP9Block *const b = &s->b;

  2658.     enum BlockSize bs = bl * 3 + bp;

  2659.     int res, y, w4 = bwh_tab[1][bs][0], h4 = bwh_tab[1][bs][1], lvl;

  2660.     int emu[2];

  2661.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2662. 

  2663.     s->row = row;

  2664.     s->row7 = row & 7;

  2665.     s->col = col;

  2666.     s->col7 = col & 7;

  2667.     s->min_mv.x = -(128 + col * 64);

  2668.     s->min_mv.y = -(128 + row * 64);

  2669.     s->max_mv.x = 128 + (s->cols - col - w4) * 64;

  2670.     s->max_mv.y = 128 + (s->rows - row - h4) * 64;

  2671.     b->bs = bs;

  2672.     decode_mode(ctx);

  2673.     b->uvtx = b->tx - (w4 * 2 == (1 << b->tx) || h4 * 2 == (1 << b->tx));

  2674. 

  2675.     if (!b->skip) {

  2676.         if ((res = decode_coeffs(ctx)) < 0)

  2677.             return res;

  2678.     } else {

  2679.         int pl;

  2680. 

  2681.         memset(&s->above_y_nnz_ctx[col * 2], 0, w4 * 2);

  2682.         memset(&s->left_y_nnz_ctx[(row & 7) << 1], 0, h4 * 2);

  2683.         for (pl = 0; pl < 2; pl++) {

  2684.             memset(&s->above_uv_nnz_ctx[pl][col], 0, w4);

  2685.             memset(&s->left_uv_nnz_ctx[pl][row & 7], 0, h4);

  2686.         }

  2687.     }

  2688. 

  2689.     // emulated overhangs if the stride of the target buffer can't hold. This

  2690.     // allows to support emu-edge and so on even if we have large block

  2691.     // overhangs

  2692.     emu[0] = (col + w4) * 8 > f->linesize[0] ||

  2693.              (row + h4) > s->rows + 2 * !(ctx->flags & CODEC_FLAG_EMU_EDGE);

  2694.     emu[1] = (col + w4) * 4 > f->linesize[1] ||

  2695.              (row + h4) > s->rows + 2 * !(ctx->flags & CODEC_FLAG_EMU_EDGE);

  2696.     if (emu[0]) {

  2697.         s->dst[0] = s->tmp_y;

  2698.         s->y_stride = 64;

  2699.     } else {

  2700.         s->dst[0] = f->data[0] + yoff;

  2701.         s->y_stride = f->linesize[0];

  2702.     }

  2703.     if (emu[1]) {

  2704.         s->dst[1] = s->tmp_uv[0];

  2705.         s->dst[2] = s->tmp_uv[1];

  2706.         s->uv_stride = 32;

  2707.     } else {

  2708.         s->dst[1] = f->data[1] + uvoff;

  2709.         s->dst[2] = f->data[2] + uvoff;

  2710.         s->uv_stride = f->linesize[1];

  2711.     }

  2712.     if (b->intra) {

  2713.         intra_recon(ctx, yoff, uvoff);

  2714.     } else {

  2715.         inter_recon(ctx);

  2716.     }

  2717.     if (emu[0]) {

  2718.         int w = FFMIN(s->cols - col, w4) * 8, h = FFMIN(s->rows - row, h4) * 8, n, o = 0;

  2719. 

  2720.         for (n = 0; o < w; n++) {

  2721.             int bw = 64 >> n;

  2722. 

  2723.             av_assert2(n <= 4);

  2724.             if (w & bw) {

  2725.                 s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o, f->linesize[0],

  2726.                                          s->tmp_y + o, 64, h, 0, 0);

  2727.                 o += bw;

  2728.             }

  2729.         }

  2730.     }

  2731.     if (emu[1]) {

  2732.         int w = FFMIN(s->cols - col, w4) * 4, h = FFMIN(s->rows - row, h4) * 4, n, o = 0;

  2733. 

  2734.         for (n = 1; o < w; n++) {

  2735.             int bw = 64 >> n;

  2736. 

  2737.             av_assert2(n <= 4);

  2738.             if (w & bw) {

  2739.                 s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o, f->linesize[1],

  2740.                                          s->tmp_uv[0] + o, 32, h, 0, 0);

  2741.                 s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o, f->linesize[2],

  2742.                                          s->tmp_uv[1] + o, 32, h, 0, 0);

  2743.                 o += bw;

  2744.             }

  2745.         }

  2746.     }

  2747. 

  2748.     // pick filter level and find edges to apply filter to

  2749.     if (s->filter.level &&

  2750.         (lvl = s->segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]

  2751.                                                     [b->mode[3] != ZEROMV]) > 0) {

  2752.         int x_end = FFMIN(s->cols - col, w4), y_end = FFMIN(s->rows - row, h4);

  2753.         int skip_inter = !b->intra && b->skip;

  2754. 

  2755.         for (y = 0; y < h4; y++)

  2756.             memset(&lflvl->level[((row & 7) + y) * 8 + (col & 7)], lvl, w4);

  2757.         mask_edges(lflvl, 0, row & 7, col & 7, x_end, y_end, 0, 0, b->tx, skip_inter);

  2758.         mask_edges(lflvl, 1, row & 7, col & 7, x_end, y_end,

  2759.                    s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,

  2760.                    s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,

  2761.                    b->uvtx, skip_inter);

  2762. 

  2763.         if (!s->filter.lim_lut[lvl]) {

  2764.             int sharp = s->filter.sharpness;

  2765.             int limit = lvl;

  2766. 

  2767.             if (sharp > 0) {

  2768.                 limit >>= (sharp + 3) >> 2;

  2769.                 limit = FFMIN(limit, 9 - sharp);

  2770.             }

  2771.             limit = FFMAX(limit, 1);

  2772. 

  2773.             s->filter.lim_lut[lvl] = limit;

  2774.             s->filter.mblim_lut[lvl] = 2 * (lvl + 2) + limit;

  2775.         }

  2776.     }

  2777. 

  2778.     return 0;

  2779. }

  2780. 

  2781. static int decode_sb(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  2782.                      ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  2783. {

  2784.     VP9Context *s = ctx->priv_data;

  2785.     int c = ((s->above_partition_ctx[col] >> (3 - bl)) & 1) |

  2786.             (((s->left_partition_ctx[row & 0x7] >> (3 - bl)) & 1) << 1), res;

  2787.     const uint8_t *p = s->keyframe ? vp9_default_kf_partition_probs[bl][c] :

  2788.                                      s->prob.p.partition[bl][c];

  2789.     enum BlockPartition bp;

  2790.     ptrdiff_t hbs = 4 >> bl;

  2791.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2792.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  2793. 

  2794.     if (bl == BL_8X8) {

  2795.         bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  2796.         res = decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2797.     } else if (col + hbs < s->cols) {

  2798.         if (row + hbs < s->rows) {

  2799.             bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  2800.             switch (bp) {

  2801.             case PARTITION_NONE:

  2802.                 res = decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2803.                 break;

  2804.             case PARTITION_H:

  2805.                 if (!(res = decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp))) {

  2806.                     yoff  += hbs * 8 * y_stride;

  2807.                     uvoff += hbs * 4 * uv_stride;

  2808.                     res = decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, bl, bp);

  2809.                 }

  2810.                 break;

  2811.             case PARTITION_V:

  2812.                 if (!(res = decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp))) {

  2813.                     yoff  += hbs * 8;

  2814.                     uvoff += hbs * 4;

  2815.                     res = decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, bl, bp);

  2816.                 }

  2817.                 break;

  2818.             case PARTITION_SPLIT:

  2819.                 if (!(res = decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1))) {

  2820.                     if (!(res = decode_sb(ctx, row, col + hbs, lflvl,

  2821.                                           yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1))) {

  2822.                         yoff  += hbs * 8 * y_stride;

  2823.                         uvoff += hbs * 4 * uv_stride;

  2824.                         if (!(res = decode_sb(ctx, row + hbs, col, lflvl,

  2825.                                               yoff, uvoff, bl + 1)))

  2826.                             res = decode_sb(ctx, row + hbs, col + hbs, lflvl,

  2827.                                             yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2828.                     }

  2829.                 }

  2830.                 break;

  2831.             default:

  2832.                 av_assert0(0);

  2833.             }

  2834.         } else if (vp56_rac_get_prob_branchy(&s->c, p[1])) {

  2835.             bp = PARTITION_SPLIT;

  2836.             if (!(res = decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1)))

  2837.                 res = decode_sb(ctx, row, col + hbs, lflvl,

  2838.                                 yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2839.         } else {

  2840.             bp = PARTITION_H;

  2841.             res = decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2842.         }

  2843.     } else if (row + hbs < s->rows) {

  2844.         if (vp56_rac_get_prob_branchy(&s->c, p[2])) {

  2845.             bp = PARTITION_SPLIT;

  2846.             if (!(res = decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1))) {

  2847.                 yoff  += hbs * 8 * y_stride;

  2848.                 uvoff += hbs * 4 * uv_stride;

  2849.                 res = decode_sb(ctx, row + hbs, col, lflvl,

  2850.                                 yoff, uvoff, bl + 1);

  2851.             }

  2852.         } else {

  2853.             bp = PARTITION_V;

  2854.             res = decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2855.         }

  2856.     } else {

  2857.         bp = PARTITION_SPLIT;

  2858.         res = decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2859.     }

  2860.     s->counts.partition[bl][c][bp]++;

  2861. 

  2862.     return res;

  2863. }

  2864. 

  2865. static void loopfilter_sb(AVCodecContext *ctx, struct VP9Filter *lflvl,

  2866.                           int row, int col, ptrdiff_t yoff, ptrdiff_t uvoff)

  2867. {

  2868.     VP9Context *s = ctx->priv_data;

  2869.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2870.     uint8_t *dst = f->data[0] + yoff, *lvl = lflvl->level;

  2871.     ptrdiff_t ls_y = f->linesize[0], ls_uv = f->linesize[1];

  2872.     int y, x, p;

  2873. 

  2874.     // FIXME in how far can we interleave the v/h loopfilter calls? E.g.

  2875.     // if you think of them as acting on a 8x8 block max, we can interleave

  2876.     // each v/h within the single x loop, but that only works if we work on

  2877.     // 8 pixel blocks, and we won't always do that (we want at least 16px

  2878.     // to use SSE2 optimizations, perhaps 32 for AVX2)

  2879. 

  2880.     // filter edges between columns, Y plane (e.g. block1 | block2)

  2881.     for (y = 0; y < 8; y += 2, dst += 16 * ls_y, lvl += 16) {

  2882.         uint8_t *ptr = dst, *l = lvl, *hmask1 = lflvl->mask[0][0][y];

  2883.         uint8_t *hmask2 = lflvl->mask[0][0][y + 1];

  2884.         unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2], hm13 = hmask1[3];

  2885.         unsigned hm2 = hmask2[1] | hmask2[2], hm23 = hmask2[3];

  2886.         unsigned hm = hm1 | hm2 | hm13 | hm23;

  2887. 

  2888.         for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 8, l++) {

  2889.             if (hm1 & x) {

  2890.                 int L = *l, H = L >> 4;

  2891.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  2892. 

  2893.                 if (col || x > 1) {

  2894.                     if (hmask1[0] & x) {

  2895.                         if (hmask2[0] & x) {

  2896.                             av_assert2(l[8] == L);

  2897.                             s->dsp.loop_filter_16[0](ptr, ls_y, E, I, H);

  2898.                         } else {

  2899.                             s->dsp.loop_filter_8[2][0](ptr, ls_y, E, I, H);

  2900.                         }

  2901.                     } else if (hm2 & x) {

  2902.                         L = l[8];

  2903.                         H |= (L >> 4) << 8;

  2904.                         E |= s->filter.mblim_lut[L] << 8;

  2905.                         I |= s->filter.lim_lut[L] << 8;

  2906.                         s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  2907.                                                [!!(hmask2[1] & x)]

  2908.                                                [0](ptr, ls_y, E, I, H);

  2909.                     } else {

  2910.                         s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  2911.                                             [0](ptr, ls_y, E, I, H);

  2912.                     }

  2913.                 }

  2914.             } else if (hm2 & x) {

  2915.                 int L = l[8], H = L >> 4;

  2916.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  2917. 

  2918.                 if (col || x > 1) {

  2919.                     s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  2920.                                         [0](ptr + 8 * ls_y, ls_y, E, I, H);

  2921.                 }

  2922.             }

  2923.             if (hm13 & x) {

  2924.                 int L = *l, H = L >> 4;

  2925.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  2926. 

  2927.                 if (hm23 & x) {

  2928.                     L = l[8];

  2929.                     H |= (L >> 4) << 8;

  2930.                     E |= s->filter.mblim_lut[L] << 8;

  2931.                     I |= s->filter.lim_lut[L] << 8;

  2932.                     s->dsp.loop_filter_mix2[0][0][0](ptr + 4, ls_y, E, I, H);

  2933.                 } else {

  2934.                     s->dsp.loop_filter_8[0][0](ptr + 4, ls_y, E, I, H);

  2935.                 }

  2936.             } else if (hm23 & x) {

  2937.                 int L = l[8], H = L >> 4;

  2938.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  2939. 

  2940.                 s->dsp.loop_filter_8[0][0](ptr + 8 * ls_y + 4, ls_y, E, I, H);

  2941.             }

  2942.         }

  2943.     }

  2944. 

  2945.     //                                          block1

  2946.     // filter edges between rows, Y plane (e.g. ------)

  2947.     //                                          block2

  2948.     dst = f->data[0] + yoff;

  2949.     lvl = lflvl->level;

  2950.     for (y = 0; y < 8; y++, dst += 8 * ls_y, lvl += 8) {

  2951.         uint8_t *ptr = dst, *l = lvl, *vmask = lflvl->mask[0][1][y];

  2952.         unsigned vm = vmask[0] | vmask[1] | vmask[2], vm3 = vmask[3];

  2953. 

  2954.         for (x = 1; vm & ~(x - 1); x <<= 2, ptr += 16, l += 2) {

  2955.             if (row || y) {

  2956.                 if (vm & x) {

  2957.                     int L = *l, H = L >> 4;

  2958.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  2959. 

  2960.                     if (vmask[0] & x) {

  2961.                         if (vmask[0] & (x << 1)) {

  2962.                             av_assert2(l[1] == L);

  2963.                             s->dsp.loop_filter_16[1](ptr, ls_y, E, I, H);

  2964.                         } else {

  2965.                             s->dsp.loop_filter_8[2][1](ptr, ls_y, E, I, H);

  2966.                         }

  2967.                     } else if (vm & (x << 1)) {

  2968.                         L = l[1];

  2969.                         H |= (L >> 4) << 8;

  2970.                         E |= s->filter.mblim_lut[L] << 8;

  2971.                         I |= s->filter.lim_lut[L] << 8;

  2972.                         s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  2973.                                                [!!(vmask[1] & (x << 1))]

  2974.                                                [1](ptr, ls_y, E, I, H);

  2975.                     } else {

  2976.                         s->dsp.loop_filter_8[!!(vmask[1] & x)]

  2977.                                             [1](ptr, ls_y, E, I, H);

  2978.                     }

  2979.                 } else if (vm & (x << 1)) {

  2980.                     int L = l[1], H = L >> 4;

  2981.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  2982. 

  2983.                     s->dsp.loop_filter_8[!!(vmask[1] & (x << 1))]

  2984.                                         [1](ptr + 8, ls_y, E, I, H);

  2985.                 }

  2986.             }

  2987.             if (vm3 & x) {

  2988.                 int L = *l, H = L >> 4;

  2989.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  2990. 

  2991.                 if (vm3 & (x << 1)) {

  2992.                     L = l[1];

  2993.                     H |= (L >> 4) << 8;

  2994.                     E |= s->filter.mblim_lut[L] << 8;

  2995.                     I |= s->filter.lim_lut[L] << 8;

  2996.                     s->dsp.loop_filter_mix2[0][0][1](ptr + ls_y * 4, ls_y, E, I, H);

  2997.                 } else {

  2998.                     s->dsp.loop_filter_8[0][1](ptr + ls_y * 4, ls_y, E, I, H);

  2999.                 }

  3000.             } else if (vm3 & (x << 1)) {

  3001.                 int L = l[1], H = L >> 4;

  3002.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3003. 

  3004.                 s->dsp.loop_filter_8[0][1](ptr + ls_y * 4 + 8, ls_y, E, I, H);

  3005.             }

  3006.         }

  3007.     }

  3008. 

  3009.     // same principle but for U/V planes

  3010.     for (p = 0; p < 2; p++) {

  3011.         lvl = lflvl->level;

  3012.         dst = f->data[1 + p] + uvoff;

  3013.         for (y = 0; y < 8; y += 4, dst += 16 * ls_uv, lvl += 32) {

  3014.             uint8_t *ptr = dst, *l = lvl, *hmask1 = lflvl->mask[1][0][y];

  3015.             uint8_t *hmask2 = lflvl->mask[1][0][y + 2];

  3016.             unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2];

  3017.             unsigned hm2 = hmask2[1] | hmask2[2], hm = hm1 | hm2;

  3018. 

  3019.             for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 4) {

  3020.                 if (col || x > 1) {

  3021.                     if (hm1 & x) {

  3022.                         int L = *l, H = L >> 4;

  3023.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3024. 

  3025.                         if (hmask1[0] & x) {

  3026.                             if (hmask2[0] & x) {

  3027.                                 av_assert2(l[16] == L);

  3028.                                 s->dsp.loop_filter_16[0](ptr, ls_uv, E, I, H);

  3029.                             } else {

  3030.                                 s->dsp.loop_filter_8[2][0](ptr, ls_uv, E, I, H);

  3031.                             }

  3032.                         } else if (hm2 & x) {

  3033.                             L = l[16];

  3034.                             H |= (L >> 4) << 8;

  3035.                             E |= s->filter.mblim_lut[L] << 8;

  3036.                             I |= s->filter.lim_lut[L] << 8;

  3037.                             s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3038.                                                    [!!(hmask2[1] & x)]

  3039.                                                    [0](ptr, ls_uv, E, I, H);

  3040.                         } else {

  3041.                             s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3042.                                                 [0](ptr, ls_uv, E, I, H);

  3043.                         }

  3044.                     } else if (hm2 & x) {

  3045.                         int L = l[16], H = L >> 4;

  3046.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3047. 

  3048.                         s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3049.                                             [0](ptr + 8 * ls_uv, ls_uv, E, I, H);

  3050.                     }

  3051.                 }

  3052.                 if (x & 0xAA)

  3053.                     l += 2;

  3054.             }

  3055.         }

  3056.         lvl = lflvl->level;

  3057.         dst = f->data[1 + p] + uvoff;

  3058.         for (y = 0; y < 8; y++, dst += 4 * ls_uv) {

  3059.             uint8_t *ptr = dst, *l = lvl, *vmask = lflvl->mask[1][1][y];

  3060.             unsigned vm = vmask[0] | vmask[1] | vmask[2];

  3061. 

  3062.             for (x = 1; vm & ~(x - 1); x <<= 4, ptr += 16, l += 4) {

  3063.                 if (row || y) {

  3064.                     if (vm & x) {

  3065.                         int L = *l, H = L >> 4;

  3066.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3067. 

  3068.                         if (vmask[0] & x) {

  3069.                             if (vmask[0] & (x << 2)) {

  3070.                                 av_assert2(l[2] == L);

  3071.                                 s->dsp.loop_filter_16[1](ptr, ls_uv, E, I, H);

  3072.                             } else {

  3073.                                 s->dsp.loop_filter_8[2][1](ptr, ls_uv, E, I, H);

  3074.                             }

  3075.                         } else if (vm & (x << 2)) {

  3076.                             L = l[2];

  3077.                             H |= (L >> 4) << 8;

  3078.                             E |= s->filter.mblim_lut[L] << 8;

  3079.                             I |= s->filter.lim_lut[L] << 8;

  3080.                             s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3081.                                                    [!!(vmask[1] & (x << 2))]

  3082.                                                    [1](ptr, ls_uv, E, I, H);

  3083.                         } else {

  3084.                             s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3085.                                                 [1](ptr, ls_uv, E, I, H);

  3086.                         }

  3087.                     } else if (vm & (x << 2)) {

  3088.                         int L = l[2], H = L >> 4;

  3089.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3090. 

  3091.                         s->dsp.loop_filter_8[!!(vmask[1] & (x << 2))]

  3092.                                             [1](ptr + 8, ls_uv, E, I, H);

  3093.                     }

  3094.                 }

  3095.             }

  3096.             if (y & 1)

  3097.                 lvl += 16;

  3098.         }

  3099.     }

  3100. }

  3101. 

  3102. static void set_tile_offset(int *start, int *end, int idx, int log2_n, int n)

  3103. {

  3104.     int sb_start = ( idx      * n) >> log2_n;

  3105.     int sb_end   = ((idx + 1) * n) >> log2_n;

  3106.     *start = FFMIN(sb_start, n) << 3;

  3107.     *end   = FFMIN(sb_end,   n) << 3;

  3108. }

  3109. 

  3110. static av_always_inline void adapt_prob(uint8_t *p, unsigned ct0, unsigned ct1,

  3111.                                         int max_count, int update_factor)

  3112. {

  3113.     unsigned ct = ct0 + ct1, p2, p1;

  3114. 

  3115.     if (!ct)

  3116.         return;

  3117. 

  3118.     p1 = *p;

  3119.     p2 = ((ct0 << 8) + (ct >> 1)) / ct;

  3120.     p2 = av_clip(p2, 1, 255);

  3121.     ct = FFMIN(ct, max_count);

  3122.     update_factor = FASTDIV(update_factor * ct, max_count);

  3123. 

  3124.     // (p1 * (256 - update_factor) + p2 * update_factor + 128) >> 8

  3125.     *p = p1 + (((p2 - p1) * update_factor + 128) >> 8);

  3126. }

  3127. 

  3128. static void adapt_probs(VP9Context *s)

  3129. {

  3130.     int i, j, k, l, m;

  3131.     prob_context *p = &s->prob_ctx[s->framectxid].p;

  3132.     int uf = (s->keyframe || s->intraonly || !s->last_keyframe) ? 112 : 128;

  3133. 

  3134.     // coefficients

  3135.     for (i = 0; i < 4; i++)

  3136.         for (j = 0; j < 2; j++)

  3137.             for (k = 0; k < 2; k++)

  3138.                 for (l = 0; l < 6; l++)

  3139.                     for (m = 0; m < 6; m++) {

  3140.                         uint8_t *pp = s->prob_ctx[s->framectxid].coef[i][j][k][l][m];

  3141.                         unsigned *e = s->counts.eob[i][j][k][l][m];

  3142.                         unsigned *c = s->counts.coef[i][j][k][l][m];

  3143. 

  3144.                         if (l == 0 && m >= 3) // dc only has 3 pt

  3145.                             break;

  3146. 

  3147.                         adapt_prob(&pp[0], e[0], e[1], 24, uf);

  3148.                         adapt_prob(&pp[1], c[0], c[1] + c[2], 24, uf);

  3149.                         adapt_prob(&pp[2], c[1], c[2], 24, uf);

  3150.                     }

  3151. 

  3152.     if (s->keyframe || s->intraonly) {

  3153.         memcpy(p->skip,  s->prob.p.skip,  sizeof(p->skip));

  3154.         memcpy(p->tx32p, s->prob.p.tx32p, sizeof(p->tx32p));

  3155.         memcpy(p->tx16p, s->prob.p.tx16p, sizeof(p->tx16p));

  3156.         memcpy(p->tx8p,  s->prob.p.tx8p,  sizeof(p->tx8p));

  3157.         return;

  3158.     }

  3159. 

  3160.     // skip flag

  3161.     for (i = 0; i < 3; i++)

  3162.         adapt_prob(&p->skip[i], s->counts.skip[i][0], s->counts.skip[i][1], 20, 128);

  3163. 

  3164.     // intra/inter flag

  3165.     for (i = 0; i < 4; i++)

  3166.         adapt_prob(&p->intra[i], s->counts.intra[i][0], s->counts.intra[i][1], 20, 128);

  3167. 

  3168.     // comppred flag

  3169.     if (s->comppredmode == PRED_SWITCHABLE) {

  3170.       for (i = 0; i < 5; i++)

  3171.           adapt_prob(&p->comp[i], s->counts.comp[i][0], s->counts.comp[i][1], 20, 128);

  3172.     }

  3173. 

  3174.     // reference frames

  3175.     if (s->comppredmode != PRED_SINGLEREF) {

  3176.       for (i = 0; i < 5; i++)

  3177.           adapt_prob(&p->comp_ref[i], s->counts.comp_ref[i][0],

  3178.                      s->counts.comp_ref[i][1], 20, 128);

  3179.     }

  3180. 

  3181.     if (s->comppredmode != PRED_COMPREF) {

  3182.       for (i = 0; i < 5; i++) {

  3183.           uint8_t *pp = p->single_ref[i];

  3184.           unsigned (*c)[2] = s->counts.single_ref[i];

  3185. 

  3186.           adapt_prob(&pp[0], c[0][0], c[0][1], 20, 128);

  3187.           adapt_prob(&pp[1], c[1][0], c[1][1], 20, 128);

  3188.       }

  3189.     }

  3190. 

  3191.     // block partitioning

  3192.     for (i = 0; i < 4; i++)

  3193.         for (j = 0; j < 4; j++) {

  3194.             uint8_t *pp = p->partition[i][j];

  3195.             unsigned *c = s->counts.partition[i][j];

  3196. 

  3197.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3198.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3199.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3200.         }

  3201. 

  3202.     // tx size

  3203.     if (s->txfmmode == TX_SWITCHABLE) {

  3204.       for (i = 0; i < 2; i++) {

  3205.           unsigned *c16 = s->counts.tx16p[i], *c32 = s->counts.tx32p[i];

  3206. 

  3207.           adapt_prob(&p->tx8p[i], s->counts.tx8p[i][0], s->counts.tx8p[i][1], 20, 128);

  3208.           adapt_prob(&p->tx16p[i][0], c16[0], c16[1] + c16[2], 20, 128);

  3209.           adapt_prob(&p->tx16p[i][1], c16[1], c16[2], 20, 128);

  3210.           adapt_prob(&p->tx32p[i][0], c32[0], c32[1] + c32[2] + c32[3], 20, 128);

  3211.           adapt_prob(&p->tx32p[i][1], c32[1], c32[2] + c32[3], 20, 128);

  3212.           adapt_prob(&p->tx32p[i][2], c32[2], c32[3], 20, 128);

  3213.       }

  3214.     }

  3215. 

  3216.     // interpolation filter

  3217.     if (s->filtermode == FILTER_SWITCHABLE) {

  3218.         for (i = 0; i < 4; i++) {

  3219.             uint8_t *pp = p->filter[i];

  3220.             unsigned *c = s->counts.filter[i];

  3221. 

  3222.             adapt_prob(&pp[0], c[0], c[1] + c[2], 20, 128);

  3223.             adapt_prob(&pp[1], c[1], c[2], 20, 128);

  3224.         }

  3225.     }

  3226. 

  3227.     // inter modes

  3228.     for (i = 0; i < 7; i++) {

  3229.         uint8_t *pp = p->mv_mode[i];

  3230.         unsigned *c = s->counts.mv_mode[i];

  3231. 

  3232.         adapt_prob(&pp[0], c[2], c[1] + c[0] + c[3], 20, 128);

  3233.         adapt_prob(&pp[1], c[0], c[1] + c[3], 20, 128);

  3234.         adapt_prob(&pp[2], c[1], c[3], 20, 128);

  3235.     }

  3236. 

  3237.     // mv joints

  3238.     {

  3239.         uint8_t *pp = p->mv_joint;

  3240.         unsigned *c = s->counts.mv_joint;

  3241. 

  3242.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3243.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3244.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3245.     }

  3246. 

  3247.     // mv components

  3248.     for (i = 0; i < 2; i++) {

  3249.         uint8_t *pp;

  3250.         unsigned *c, (*c2)[2], sum;

  3251. 

  3252.         adapt_prob(&p->mv_comp[i].sign, s->counts.mv_comp[i].sign[0],

  3253.                    s->counts.mv_comp[i].sign[1], 20, 128);

  3254. 

  3255.         pp = p->mv_comp[i].classes;

  3256.         c = s->counts.mv_comp[i].classes;

  3257.         sum = c[1] + c[2] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9] + c[10];

  3258.         adapt_prob(&pp[0], c[0], sum, 20, 128);

  3259.         sum -= c[1];

  3260.         adapt_prob(&pp[1], c[1], sum, 20, 128);

  3261.         sum -= c[2] + c[3];

  3262.         adapt_prob(&pp[2], c[2] + c[3], sum, 20, 128);

  3263.         adapt_prob(&pp[3], c[2], c[3], 20, 128);

  3264.         sum -= c[4] + c[5];

  3265.         adapt_prob(&pp[4], c[4] + c[5], sum, 20, 128);

  3266.         adapt_prob(&pp[5], c[4], c[5], 20, 128);

  3267.         sum -= c[6];

  3268.         adapt_prob(&pp[6], c[6], sum, 20, 128);

  3269.         adapt_prob(&pp[7], c[7] + c[8], c[9] + c[10], 20, 128);

  3270.         adapt_prob(&pp[8], c[7], c[8], 20, 128);

  3271.         adapt_prob(&pp[9], c[9], c[10], 20, 128);

  3272. 

  3273.         adapt_prob(&p->mv_comp[i].class0, s->counts.mv_comp[i].class0[0],

  3274.                    s->counts.mv_comp[i].class0[1], 20, 128);

  3275.         pp = p->mv_comp[i].bits;

  3276.         c2 = s->counts.mv_comp[i].bits;

  3277.         for (j = 0; j < 10; j++)

  3278.             adapt_prob(&pp[j], c2[j][0], c2[j][1], 20, 128);

  3279. 

  3280.         for (j = 0; j < 2; j++) {

  3281.             pp = p->mv_comp[i].class0_fp[j];

  3282.             c = s->counts.mv_comp[i].class0_fp[j];

  3283.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3284.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3285.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3286.         }

  3287.         pp = p->mv_comp[i].fp;

  3288.         c = s->counts.mv_comp[i].fp;

  3289.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3290.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3291.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3292. 

  3293.         if (s->highprecisionmvs) {

  3294.             adapt_prob(&p->mv_comp[i].class0_hp, s->counts.mv_comp[i].class0_hp[0],

  3295.                        s->counts.mv_comp[i].class0_hp[1], 20, 128);

  3296.             adapt_prob(&p->mv_comp[i].hp, s->counts.mv_comp[i].hp[0],

  3297.                        s->counts.mv_comp[i].hp[1], 20, 128);

  3298.         }

  3299.     }

  3300. 

  3301.     // y intra modes

  3302.     for (i = 0; i < 4; i++) {

  3303.         uint8_t *pp = p->y_mode[i];

  3304.         unsigned *c = s->counts.y_mode[i], sum, s2;

  3305. 

  3306.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3307.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3308.         sum -= c[TM_VP8_PRED];

  3309.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3310.         sum -= c[VERT_PRED];

  3311.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3312.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3313.         sum -= s2;

  3314.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3315.         s2 -= c[HOR_PRED];

  3316.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3317.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3318.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3319.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3320.         sum -= c[VERT_LEFT_PRED];

  3321.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3322.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3323.     }

  3324. 

  3325.     // uv intra modes

  3326.     for (i = 0; i < 10; i++) {

  3327.         uint8_t *pp = p->uv_mode[i];

  3328.         unsigned *c = s->counts.uv_mode[i], sum, s2;

  3329. 

  3330.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3331.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3332.         sum -= c[TM_VP8_PRED];

  3333.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3334.         sum -= c[VERT_PRED];

  3335.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3336.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3337.         sum -= s2;

  3338.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3339.         s2 -= c[HOR_PRED];

  3340.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3341.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3342.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3343.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3344.         sum -= c[VERT_LEFT_PRED];

  3345.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3346.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3347.     }

  3348. }

  3349. 

  3350. static av_cold int vp9_decode_free(AVCodecContext *ctx)

  3351. {

  3352.     VP9Context *s = ctx->priv_data;

  3353.     int i;

  3354. 

  3355.     for (i = 0; i < 2; i++) {

  3356.         if (s->frames[i].tf.f->data[0])

  3357.             vp9_unref_frame(ctx, &s->frames[i]);

  3358.         av_frame_free(&s->frames[i].tf.f);

  3359.     }

  3360.     for (i = 0; i < 8; i++) {

  3361.         if (s->refs[i].f->data[0])

  3362.             ff_thread_release_buffer(ctx, &s->refs[i]);

  3363.         av_frame_free(&s->refs[i].f);

  3364.         if (s->next_refs[i].f->data[0])

  3365.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3366.         av_frame_free(&s->next_refs[i].f);

  3367.     }

  3368.     av_freep(&s->above_partition_ctx);

  3369.     av_freep(&s->c_b);

  3370. 

  3371.     return 0;

  3372. }

  3373. 

  3374. 

  3375. static int vp9_decode_frame(AVCodecContext *ctx, AVFrame *frame,

  3376.                             int *got_frame, AVPacket *pkt)

  3377. {

  3378.     const uint8_t *data = pkt->data;

  3379.     int size = pkt->size;

  3380.     VP9Context *s = ctx->priv_data;

  3381.     int res, tile_row, tile_col, i, ref, row, col;

  3382.     ptrdiff_t yoff = 0, uvoff = 0, ls_y, ls_uv;

  3383.     AVFrame *f;

  3384. 

  3385.     if ((res = decode_frame_header(ctx, data, size, &ref)) < 0) {

  3386.         return res;

  3387.     } else if (res == 0) {

  3388.         if (!s->refs[ref].f->data[0]) {

  3389.             av_log(ctx, AV_LOG_ERROR, "Requested reference %d not available\n", ref);

  3390.             return AVERROR_INVALIDDATA;

  3391.         }

  3392.         if ((res = av_frame_ref(frame, s->refs[ref].f)) < 0)

  3393.             return res;

  3394.         *got_frame = 1;

  3395.         return 0;

  3396.     }

  3397.     data += res;

  3398.     size -= res;

  3399. 

  3400.     if (s->frames[LAST_FRAME].tf.f->data[0])

  3401.         vp9_unref_frame(ctx, &s->frames[LAST_FRAME]);

  3402.     if (!s->keyframe && s->frames[CUR_FRAME].tf.f->data[0] &&

  3403.         (res = vp9_ref_frame(ctx, &s->frames[LAST_FRAME], &s->frames[CUR_FRAME])) < 0)

  3404.         return res;

  3405.     if (s->frames[CUR_FRAME].tf.f->data[0])

  3406.         vp9_unref_frame(ctx, &s->frames[CUR_FRAME]);

  3407.     if ((res = vp9_alloc_frame(ctx, &s->frames[CUR_FRAME])) < 0)

  3408.         return res;

  3409.     f = s->frames[CUR_FRAME].tf.f;

  3410.     f->key_frame = s->keyframe;

  3411.     f->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;

  3412.     ls_y = f->linesize[0];

  3413.     ls_uv =f->linesize[1];

  3414. 

  3415.     // ref frame setup

  3416.     for (i = 0; i < 8; i++)

  3417.         if (s->refreshrefmask & (1 << i)) {

  3418.             if (s->next_refs[i].f->data[0])

  3419.                 ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3420.             if ((res = ff_thread_ref_frame(&s->next_refs[i], &s->frames[CUR_FRAME].tf)) < 0)

  3421.                 return res;

  3422.         }

  3423. 

  3424.     // main tile decode loop

  3425.     memset(s->above_partition_ctx, 0, s->cols);

  3426.     memset(s->above_skip_ctx, 0, s->cols);

  3427.     if (s->keyframe || s->intraonly) {

  3428.         memset(s->above_mode_ctx, DC_PRED, s->cols * 2);

  3429.     } else {

  3430.         memset(s->above_mode_ctx, NEARESTMV, s->cols);

  3431.     }

  3432.     memset(s->above_y_nnz_ctx, 0, s->sb_cols * 16);

  3433.     memset(s->above_uv_nnz_ctx[0], 0, s->sb_cols * 8);

  3434.     memset(s->above_uv_nnz_ctx[1], 0, s->sb_cols * 8);

  3435.     memset(s->above_segpred_ctx, 0, s->cols);

  3436.     for (tile_row = 0; tile_row < s->tiling.tile_rows; tile_row++) {

  3437.         set_tile_offset(&s->tiling.tile_row_start, &s->tiling.tile_row_end,

  3438.                         tile_row, s->tiling.log2_tile_rows, s->sb_rows);

  3439.         for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  3440.             unsigned tile_size;

  3441. 

  3442.             if (tile_col == s->tiling.tile_cols - 1 &&

  3443.                 tile_row == s->tiling.tile_rows - 1) {

  3444.                 tile_size = size;

  3445.             } else {

  3446.                 tile_size = AV_RB32(data);

  3447.                 data += 4;

  3448.                 size -= 4;

  3449.             }

  3450.             if (tile_size > size)

  3451.                 return AVERROR_INVALIDDATA;

  3452.             ff_vp56_init_range_decoder(&s->c_b[tile_col], data, tile_size);

  3453.             if (vp56_rac_get_prob_branchy(&s->c_b[tile_col], 128)) // marker bit

  3454.                 return AVERROR_INVALIDDATA;

  3455.             data += tile_size;

  3456.             size -= tile_size;

  3457.         }

  3458. 

  3459.         for (row = s->tiling.tile_row_start; row < s->tiling.tile_row_end;

  3460.              row += 8, yoff += ls_y * 64, uvoff += ls_uv * 32) {

  3461.             struct VP9Filter *lflvl_ptr = s->lflvl;

  3462.             ptrdiff_t yoff2 = yoff, uvoff2 = uvoff;

  3463. 

  3464.             for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  3465.                 set_tile_offset(&s->tiling.tile_col_start, &s->tiling.tile_col_end,

  3466.                                 tile_col, s->tiling.log2_tile_cols, s->sb_cols);

  3467. 

  3468.                 memset(s->left_partition_ctx, 0, 8);

  3469.                 memset(s->left_skip_ctx, 0, 8);

  3470.                 if (s->keyframe || s->intraonly) {

  3471.                     memset(s->left_mode_ctx, DC_PRED, 16);

  3472.                 } else {

  3473.                     memset(s->left_mode_ctx, NEARESTMV, 8);

  3474.                 }

  3475.                 memset(s->left_y_nnz_ctx, 0, 16);

  3476.                 memset(s->left_uv_nnz_ctx, 0, 16);

  3477.                 memset(s->left_segpred_ctx, 0, 8);

  3478. 

  3479.                 memcpy(&s->c, &s->c_b[tile_col], sizeof(s->c));

  3480.                 for (col = s->tiling.tile_col_start;

  3481.                      col < s->tiling.tile_col_end;

  3482.                      col += 8, yoff2 += 64, uvoff2 += 32, lflvl_ptr++) {

  3483.                     // FIXME integrate with lf code (i.e. zero after each

  3484.                     // use, similar to invtxfm coefficients, or similar)

  3485.                     memset(lflvl_ptr->mask, 0, sizeof(lflvl_ptr->mask));

  3486. 

  3487.                     if ((res = decode_sb(ctx, row, col, lflvl_ptr,

  3488.                                          yoff2, uvoff2, BL_64X64)) < 0)

  3489.                         return res;

  3490.                 }

  3491.                 memcpy(&s->c_b[tile_col], &s->c, sizeof(s->c));

  3492.             }

  3493. 

  3494.             // backup pre-loopfilter reconstruction data for intra

  3495.             // prediction of next row of sb64s

  3496.             if (row + 8 < s->rows) {

  3497.                 memcpy(s->intra_pred_data[0],

  3498.                        f->data[0] + yoff + 63 * ls_y,

  3499.                        8 * s->cols);

  3500.                 memcpy(s->intra_pred_data[1],

  3501.                        f->data[1] + uvoff + 31 * ls_uv,

  3502.                        4 * s->cols);

  3503.                 memcpy(s->intra_pred_data[2],

  3504.                        f->data[2] + uvoff + 31 * ls_uv,

  3505.                        4 * s->cols);

  3506.             }

  3507. 

  3508.             // loopfilter one row

  3509.             if (s->filter.level) {

  3510.                 yoff2 = yoff;

  3511.                 uvoff2 = uvoff;

  3512.                 lflvl_ptr = s->lflvl;

  3513.                 for (col = 0; col < s->cols;

  3514.                      col += 8, yoff2 += 64, uvoff2 += 32, lflvl_ptr++) {

  3515.                     loopfilter_sb(ctx, lflvl_ptr, row, col, yoff2, uvoff2);

  3516.                 }

  3517.             }

  3518.         }

  3519.     }

  3520. 

  3521.     // bw adaptivity (or in case of parallel decoding mode, fw adaptivity

  3522.     // probability maintenance between frames)

  3523.     if (s->refreshctx) {

  3524.         if (s->parallelmode) {

  3525.             int j, k, l, m;

  3526. 

  3527.             for (i = 0; i < 4; i++)

  3528.                 for (j = 0; j < 2; j++)

  3529.                     for (k = 0; k < 2; k++)

  3530.                         for (l = 0; l < 6; l++)

  3531.                             for (m = 0; m < 6; m++)

  3532.                                 memcpy(s->prob_ctx[s->framectxid].coef[i][j][k][l][m],

  3533.                                        s->prob.coef[i][j][k][l][m], 3);

  3534.             s->prob_ctx[s->framectxid].p = s->prob.p;

  3535.         } else {

  3536.             adapt_probs(s);

  3537.         }

  3538.     }

  3539. 

  3540.     // ref frame setup

  3541.     for (i = 0; i < 8; i++) {

  3542.         if (s->refs[i].f->data[0])

  3543.             ff_thread_release_buffer(ctx, &s->refs[i]);

  3544.         ff_thread_ref_frame(&s->refs[i], &s->next_refs[i]);

  3545.     }

  3546. 

  3547.     if (!s->invisible) {

  3548.         if ((res = av_frame_ref(frame, s->frames[CUR_FRAME].tf.f)) < 0)

  3549.             return res;

  3550.         *got_frame = 1;

  3551.     }

  3552. 

  3553.     return 0;

  3554. }

  3555. 

  3556. static void vp9_decode_flush(AVCodecContext *ctx)

  3557. {

  3558.     VP9Context *s = ctx->priv_data;

  3559.     int i;

  3560. 

  3561.     for (i = 0; i < 2; i++)

  3562.         vp9_unref_frame(ctx, &s->frames[i]);

  3563.     for (i = 0; i < 8; i++)

  3564.         ff_thread_release_buffer(ctx, &s->refs[i]);

  3565. }

  3566. 

  3567. static int init_frames(AVCodecContext *ctx)

  3568. {

  3569.     VP9Context *s = ctx->priv_data;

  3570.     int i;

  3571. 

  3572.     for (i = 0; i < 2; i++) {

  3573.         s->frames[i].tf.f = av_frame_alloc();

  3574.         if (!s->frames[i].tf.f) {

  3575.             vp9_decode_free(ctx);

  3576.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  3577.             return AVERROR(ENOMEM);

  3578.         }

  3579.     }

  3580.     for (i = 0; i < 8; i++) {

  3581.         s->refs[i].f = av_frame_alloc();

  3582.         s->next_refs[i].f = av_frame_alloc();

  3583.         if (!s->refs[i].f || !s->next_refs[i].f) {

  3584.             vp9_decode_free(ctx);

  3585.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  3586.             return AVERROR(ENOMEM);

  3587.         }

  3588.     }

  3589. 

  3590.     return 0;

  3591. }

  3592. 

  3593. static av_cold int vp9_decode_init(AVCodecContext *ctx)

  3594. {

  3595.     VP9Context *s = ctx->priv_data;

  3596. 

  3597.     ctx->pix_fmt = AV_PIX_FMT_YUV420P;

  3598.     ff_vp9dsp_init(&s->dsp);

  3599.     ff_videodsp_init(&s->vdsp, 8);

  3600.     s->filter.sharpness = -1;

  3601. 

  3602.     return init_frames(ctx);

  3603. }

  3604. 

  3605. AVCodec ff_vp9_decoder = {

  3606.   .name                  = "vp9",

  3607.   .long_name             = NULL_IF_CONFIG_SMALL("Google VP9"),

  3608.   .type                  = AVMEDIA_TYPE_VIDEO,

  3609.   .id                    = AV_CODEC_ID_VP9,

  3610.   .priv_data_size        = sizeof(VP9Context),

  3611.   .init                  = vp9_decode_init,

  3612.   .close                 = vp9_decode_free,

  3613.   .capabilities          = CODEC_CAP_DR1,

  3614.   .flush                 = vp9_decode_flush,

  3615.     .decode                = vp9_decode_frame,

  3616. };