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.
67487 Commits
32 Branches
866MB
Tree: 1a53c8dcfc
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '1a53c8dcfc'
${ noResults }
FFmpeg/libavcodec/vp9.c

4092 lines
160KB
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.     enum BlockLevel bl;

  92.     enum BlockPartition bp;

  93. } VP9Block;

  94. 

  95. typedef struct VP9Context {

  96.     VP9DSPContext dsp;

  97.     VideoDSPContext vdsp;

  98.     GetBitContext gb;

  99.     VP56RangeCoder c;

  100.     VP56RangeCoder *c_b;

  101.     unsigned c_b_size;

  102.     VP9Block *b_base, *b;

  103.     int pass, uses_2pass, last_uses_2pass;

  104.     int row, row7, col, col7;

  105.     uint8_t *dst[3];

  106.     ptrdiff_t y_stride, uv_stride;

  107. 

  108.     // bitstream header

  109.     uint8_t profile;

  110.     uint8_t keyframe, last_keyframe;

  111.     uint8_t invisible;

  112.     uint8_t use_last_frame_mvs;

  113.     uint8_t errorres;

  114.     uint8_t colorspace;

  115.     uint8_t fullrange;

  116.     uint8_t intraonly;

  117.     uint8_t resetctx;

  118.     uint8_t refreshrefmask;

  119.     uint8_t highprecisionmvs;

  120.     enum FilterMode filtermode;

  121.     uint8_t allowcompinter;

  122.     uint8_t fixcompref;

  123.     uint8_t refreshctx;

  124.     uint8_t parallelmode;

  125.     uint8_t framectxid;

  126.     uint8_t refidx[3];

  127.     uint8_t signbias[3];

  128.     uint8_t varcompref[2];

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

  130. #define CUR_FRAME 0

  131. #define LAST_FRAME 1

  132.     VP9Frame frames[2];

  133. 

  134.     struct {

  135.         uint8_t level;

  136.         int8_t sharpness;

  137.         uint8_t lim_lut[64];

  138.         uint8_t mblim_lut[64];

  139.     } filter;

  140.     struct {

  141.         uint8_t enabled;

  142.         int8_t mode[2];

  143.         int8_t ref[4];

  144.     } lf_delta;

  145.     uint8_t yac_qi;

  146.     int8_t ydc_qdelta, uvdc_qdelta, uvac_qdelta;

  147.     uint8_t lossless;

  148.     struct {

  149.         uint8_t enabled;

  150.         uint8_t temporal;

  151.         uint8_t absolute_vals;

  152.         uint8_t update_map;

  153.         struct {

  154.             uint8_t q_enabled;

  155.             uint8_t lf_enabled;

  156.             uint8_t ref_enabled;

  157.             uint8_t skip_enabled;

  158.             uint8_t ref_val;

  159.             int16_t q_val;

  160.             int8_t lf_val;

  161.             int16_t qmul[2][2];

  162.             uint8_t lflvl[4][2];

  163.         } feat[8];

  164.     } segmentation;

  165.     struct {

  166.         unsigned log2_tile_cols, log2_tile_rows;

  167.         unsigned tile_cols, tile_rows;

  168.         unsigned tile_row_start, tile_row_end, tile_col_start, tile_col_end;

  169.     } tiling;

  170.     unsigned sb_cols, sb_rows, rows, cols;

  171.     struct {

  172.         prob_context p;

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

  174.     } prob_ctx[4];

  175.     struct {

  176.         prob_context p;

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

  178.         uint8_t seg[7];

  179.         uint8_t segpred[3];

  180.     } prob;

  181.     struct {

  182.         unsigned y_mode[4][10];

  183.         unsigned uv_mode[10][10];

  184.         unsigned filter[4][3];

  185.         unsigned mv_mode[7][4];

  186.         unsigned intra[4][2];

  187.         unsigned comp[5][2];

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

  189.         unsigned comp_ref[5][2];

  190.         unsigned tx32p[2][4];

  191.         unsigned tx16p[2][3];

  192.         unsigned tx8p[2][2];

  193.         unsigned skip[3][2];

  194.         unsigned mv_joint[4];

  195.         struct {

  196.             unsigned sign[2];

  197.             unsigned classes[11];

  198.             unsigned class0[2];

  199.             unsigned bits[10][2];

  200.             unsigned class0_fp[2][4];

  201.             unsigned fp[4];

  202.             unsigned class0_hp[2];

  203.             unsigned hp[2];

  204.         } mv_comp[2];

  205.         unsigned partition[4][4][4];

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

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

  208.     } counts;

  209.     enum TxfmMode txfmmode;

  210.     enum CompPredMode comppredmode;

  211. 

  212.     // contextual (left/above) cache

  213.     DECLARE_ALIGNED(16, uint8_t, left_y_nnz_ctx)[16];

  214.     DECLARE_ALIGNED(16, uint8_t, left_mode_ctx)[16];

  215.     DECLARE_ALIGNED(16, VP56mv, left_mv_ctx)[16][2];

  216.     DECLARE_ALIGNED(8, uint8_t, left_uv_nnz_ctx)[2][8];

  217.     DECLARE_ALIGNED(8, uint8_t, left_partition_ctx)[8];

  218.     DECLARE_ALIGNED(8, uint8_t, left_skip_ctx)[8];

  219.     DECLARE_ALIGNED(8, uint8_t, left_txfm_ctx)[8];

  220.     DECLARE_ALIGNED(8, uint8_t, left_segpred_ctx)[8];

  221.     DECLARE_ALIGNED(8, uint8_t, left_intra_ctx)[8];

  222.     DECLARE_ALIGNED(8, uint8_t, left_comp_ctx)[8];

  223.     DECLARE_ALIGNED(8, uint8_t, left_ref_ctx)[8];

  224.     DECLARE_ALIGNED(8, uint8_t, left_filter_ctx)[8];

  225.     uint8_t *above_partition_ctx;

  226.     uint8_t *above_mode_ctx;

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

  228.     uint8_t *above_y_nnz_ctx;

  229.     uint8_t *above_uv_nnz_ctx[2];

  230.     uint8_t *above_skip_ctx; // 1bit

  231.     uint8_t *above_txfm_ctx; // 2bit

  232.     uint8_t *above_segpred_ctx; // 1bit

  233.     uint8_t *above_intra_ctx; // 1bit

  234.     uint8_t *above_comp_ctx; // 1bit

  235.     uint8_t *above_ref_ctx; // 2bit

  236.     uint8_t *above_filter_ctx;

  237.     VP56mv (*above_mv_ctx)[2];

  238. 

  239.     // whole-frame cache

  240.     uint8_t *intra_pred_data[3];

  241.     struct VP9Filter *lflvl;

  242.     DECLARE_ALIGNED(32, uint8_t, edge_emu_buffer)[135*144];

  243. 

  244.     // block reconstruction intermediates

  245.     int block_alloc_using_2pass;

  246.     int16_t *block_base, *block, *uvblock_base[2], *uvblock[2];

  247.     uint8_t *eob_base, *uveob_base[2], *eob, *uveob[2];

  248.     struct { int x, y; } min_mv, max_mv;

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

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

  251.     uint16_t mvscale[3][2];

  252.     uint8_t mvstep[3][2];

  253. } VP9Context;

  254. 

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

  256.     {

  257.         { 16, 16 }, { 16, 8 }, { 8, 16 }, { 8, 8 }, { 8, 4 }, { 4, 8 },

  258.         { 4, 4 }, { 4, 2 }, { 2, 4 }, { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 },

  259.     }, {

  260.         { 8, 8 }, { 8, 4 }, { 4, 8 }, { 4, 4 }, { 4, 2 }, { 2, 4 },

  261.         { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 }, { 1, 1 }, { 1, 1 }, { 1, 1 },

  262.     }

  263. };

  264. 

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

  266. {

  267.     VP9Context *s = ctx->priv_data;

  268.     int ret, sz;

  269. 

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

  271.         return ret;

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

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

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

  275.         return AVERROR(ENOMEM);

  276.     }

  277. 

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

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

  280. 

  281.     // retain segmentation map if it doesn't update

  282.     if (s->segmentation.enabled && !s->segmentation.update_map &&

  283.         !s->intraonly && !s->keyframe && !s->errorres &&

  284.         ctx->active_thread_type != FF_THREAD_FRAME) {

  285.         memcpy(f->segmentation_map, s->frames[LAST_FRAME].segmentation_map, sz);

  286.     }

  287. 

  288.     return 0;

  289. }

  290. 

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

  292. {

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

  294.     av_buffer_unref(&f->extradata);

  295. }

  296. 

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

  298. {

  299.     int res;

  300. 

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

  302.         return res;

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

  304.         vp9_unref_frame(ctx, dst);

  305.         return AVERROR(ENOMEM);

  306.     }

  307. 

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

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

  310. 

  311.     return 0;

  312. }

  313. 

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

  315. {

  316.     VP9Context *s = ctx->priv_data;

  317.     uint8_t *p;

  318. 

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

  320. 

  321.     if (s->intra_pred_data[0] && w == ctx->width && h == ctx->height)

  322.         return 0;

  323. 

  324.     ctx->width  = w;

  325.     ctx->height = h;

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

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

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

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

  330. 

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

  332.     av_freep(&s->intra_pred_data[0]);

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

  334.     if (!p)

  335.         return AVERROR(ENOMEM);

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  353. #undef assign

  354. 

  355.     // these will be re-allocated a little later

  356.     av_freep(&s->b_base);

  357.     av_freep(&s->block_base);

  358. 

  359.     return 0;

  360. }

  361. 

  362. static int update_block_buffers(AVCodecContext *ctx)

  363. {

  364.     VP9Context *s = ctx->priv_data;

  365. 

  366.     if (s->b_base && s->block_base && s->block_alloc_using_2pass == s->uses_2pass)

  367.         return 0;

  368. 

  369.     av_free(s->b_base);

  370.     av_free(s->block_base);

  371.     if (s->uses_2pass) {

  372.         int sbs = s->sb_cols * s->sb_rows;

  373. 

  374.         s->b_base = av_malloc_array(s->cols * s->rows, sizeof(VP9Block));

  375.         s->block_base = av_mallocz((64 * 64 + 128) * sbs * 3);

  376.         if (!s->b_base || !s->block_base)

  377.             return AVERROR(ENOMEM);

  378.         s->uvblock_base[0] = s->block_base + sbs * 64 * 64;

  379.         s->uvblock_base[1] = s->uvblock_base[0] + sbs * 32 * 32;

  380.         s->eob_base = (uint8_t *) (s->uvblock_base[1] + sbs * 32 * 32);

  381.         s->uveob_base[0] = s->eob_base + 256 * sbs;

  382.         s->uveob_base[1] = s->uveob_base[0] + 64 * sbs;

  383.     } else {

  384.         s->b_base = av_malloc(sizeof(VP9Block));

  385.         s->block_base = av_mallocz((64 * 64 + 128) * 3);

  386.         if (!s->b_base || !s->block_base)

  387.             return AVERROR(ENOMEM);

  388.         s->uvblock_base[0] = s->block_base + 64 * 64;

  389.         s->uvblock_base[1] = s->uvblock_base[0] + 32 * 32;

  390.         s->eob_base = (uint8_t *) (s->uvblock_base[1] + 32 * 32);

  391.         s->uveob_base[0] = s->eob_base + 256;

  392.         s->uveob_base[1] = s->uveob_base[0] + 64;

  393.     }

  394.     s->block_alloc_using_2pass = s->uses_2pass;

  395. 

  396.     return 0;

  397. }

  398. 

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

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

  401. {

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

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

  404. }

  405. 

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

  407. {

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

  409. }

  410. 

  411. // differential forward probability updates

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

  413. {

  414.     static const int inv_map_table[255] = {

  415.           7,  20,  33,  46,  59,  72,  85,  98, 111, 124, 137, 150, 163, 176,

  416.         189, 202, 215, 228, 241, 254,   1,   2,   3,   4,   5,   6,   8,   9,

  417.          10,  11,  12,  13,  14,  15,  16,  17,  18,  19,  21,  22,  23,  24,

  418.          25,  26,  27,  28,  29,  30,  31,  32,  34,  35,  36,  37,  38,  39,

  419.          40,  41,  42,  43,  44,  45,  47,  48,  49,  50,  51,  52,  53,  54,

  420.          55,  56,  57,  58,  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,

  421.          70,  71,  73,  74,  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,

  422.          86,  87,  88,  89,  90,  91,  92,  93,  94,  95,  96,  97,  99, 100,

  423.         101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 112, 113, 114, 115,

  424.         116, 117, 118, 119, 120, 121, 122, 123, 125, 126, 127, 128, 129, 130,

  425.         131, 132, 133, 134, 135, 136, 138, 139, 140, 141, 142, 143, 144, 145,

  426.         146, 147, 148, 149, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160,

  427.         161, 162, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175,

  428.         177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 190, 191,

  429.         192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 204, 205, 206,

  430.         207, 208, 209, 210, 211, 212, 213, 214, 216, 217, 218, 219, 220, 221,

  431.         222, 223, 224, 225, 226, 227, 229, 230, 231, 232, 233, 234, 235, 236,

  432.         237, 238, 239, 240, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251,

  433.         252, 253, 253,

  434.     };

  435.     int d;

  436. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  451. 

  452.     if (!vp8_rac_get(c)) {

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

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

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

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

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

  458.     } else {

  459.         d = vp8_rac_get_uint(c, 7);

  460.         if (d >= 65)

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

  462.         d += 64;

  463.         av_assert2(d < FF_ARRAY_ELEMS(inv_map_table));

  464.     }

  465. 

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

  467.                     255 - inv_recenter_nonneg(inv_map_table[d], 255 - p);

  468. }

  469. 

  470. static int decode_frame_header(AVCodecContext *ctx,

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

  472. {

  473.     VP9Context *s = ctx->priv_data;

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

  475.     int last_invisible;

  476.     const uint8_t *data2;

  477. 

  478.     /* general header */

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

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

  481.         return res;

  482.     }

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

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

  485.         return AVERROR_INVALIDDATA;

  486.     }

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

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

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

  490.         return AVERROR_INVALIDDATA;

  491.     }

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

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

  494.         return 0;

  495.     }

  496.     s->last_uses_2pass = s->uses_2pass;

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

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

  499.     last_invisible    = s->invisible;

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

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

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

  503.     if (s->keyframe) {

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

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

  506.             return AVERROR_INVALIDDATA;

  507.         }

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

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

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

  511.             return AVERROR_INVALIDDATA;

  512.         }

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

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

  515.         s->refreshrefmask = 0xff;

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

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

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

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

  520.     } else {

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

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

  523.         if (s->intraonly) {

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

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

  526.                 return AVERROR_INVALIDDATA;

  527.             }

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

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

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

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

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

  533.         } else {

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

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

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

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

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

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

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

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

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

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

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

  545.                 return AVERROR_INVALIDDATA;

  546.             }

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

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

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

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

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

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

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

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

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

  556.             } else {

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

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

  559.             }

  560.             // Note that in this code, "CUR_FRAME" is actually before we

  561.             // have formally allocated a frame, and thus actually represents

  562.             // the _last_ frame

  563.             s->use_last_frame_mvs &= s->frames[CUR_FRAME].tf.f->width == w &&

  564.                                      s->frames[CUR_FRAME].tf.f->height == h;

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

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

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

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

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

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

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

  572.             if (s->allowcompinter) {

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

  574.                     s->fixcompref    = 2;

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

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

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

  578.                     s->fixcompref    = 1;

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

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

  581.                 } else {

  582.                     s->fixcompref    = 0;

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

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

  585.                 }

  586.             }

  587. 

  588.             for (i = 0; i < 3; i++) {

  589.                 AVFrame *ref = s->refs[s->refidx[i]].f;

  590.                 int refw = ref->width, refh = ref->height;

  591. 

  592.                 if (refw == w && refh == h) {

  593.                     s->mvscale[i][0] = s->mvscale[i][1] = 0;

  594.                 } else {

  595.                     if (w * 2 < refw || h * 2 < refh || w > 16 * refw || h > 16 * refh) {

  596.                         av_log(ctx, AV_LOG_ERROR,

  597.                                "Invalid ref frame dimensions %dx%d for frame size %dx%d\n",

  598.                                refw, refh, w, h);

  599.                         return AVERROR_INVALIDDATA;

  600.                     }

  601.                     s->mvscale[i][0] = (refw << 14) / w;

  602.                     s->mvscale[i][1] = (refh << 14) / h;

  603.                     s->mvstep[i][0] = 16 * s->mvscale[i][0] >> 14;

  604.                     s->mvstep[i][1] = 16 * s->mvscale[i][1] >> 14;

  605.                 }

  606.             }

  607.         }

  608.     }

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

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

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

  612. 

  613.     /* loopfilter header data */

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

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

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

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

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

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

  620.     s->filter.sharpness = sharp;

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

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

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

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

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

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

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

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

  629.         }

  630.     } else {

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

  632.     }

  633. 

  634.     /* quantization header data */

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

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

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

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

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

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

  641. 

  642.     /* segmentation header info */

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

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

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

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

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

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

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

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

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

  652.             }

  653.         }

  654.         if ((!s->segmentation.update_map || s->segmentation.temporal) &&

  655.             (w != s->frames[CUR_FRAME].tf.f->width ||

  656.              h != s->frames[CUR_FRAME].tf.f->height)) {

  657.             av_log(ctx, AV_LOG_ERROR,

  658.                    "Reference segmap (temp=%d,update=%d) enabled on size-change!\n",

  659.                    s->segmentation.temporal, s->segmentation.update_map);

  660.             return AVERROR_INVALIDDATA;

  661.         }

  662. 

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

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

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

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

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

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

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

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

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

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

  673.             }

  674.         }

  675.     } else {

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

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

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

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

  680.     }

  681. 

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

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

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

  685. 

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

  687.             if (s->segmentation.absolute_vals)

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

  689.             else

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

  691.         } else {

  692.             qyac  = s->yac_qi;

  693.         }

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

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

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

  697.         qyac  = av_clip_uintp2(qyac, 8);

  698. 

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

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

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

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

  703. 

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

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

  706.             if (s->segmentation.absolute_vals)

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

  708.             else

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

  710.         } else {

  711.             lflvl  = s->filter.level;

  712.         }

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

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

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

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

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

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

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

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

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

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

  723.         }

  724.     }

  725. 

  726.     /* tiling info */

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

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

  729.         return res;

  730.     }

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

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

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

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

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

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

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

  738.             s->tiling.log2_tile_cols++;

  739.         else

  740.             break;

  741.     }

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

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

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

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

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

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

  748.         if (!s->c_b) {

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

  750.             return AVERROR(ENOMEM);

  751.         }

  752.     }

  753. 

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

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

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

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

  758.                sizeof(vp9_default_coef_probs));

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

  760.                sizeof(vp9_default_coef_probs));

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

  762.                sizeof(vp9_default_coef_probs));

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

  764.                sizeof(vp9_default_coef_probs));

  765.     }

  766. 

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

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

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

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

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

  772.         return AVERROR_INVALIDDATA;

  773.     }

  774.     res = ff_vp56_init_range_decoder(&s->c, data2, size2);

  775.     if (res < 0)

  776.         return res;

  777. 

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

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

  780.         return AVERROR_INVALIDDATA;

  781.     }

  782. 

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

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

  785.     } else {

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

  787.     }

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

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

  790.     // fw update)?

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

  792. 

  793.     // txfm updates

  794.     if (s->lossless) {

  795.         s->txfmmode = TX_4X4;

  796.     } else {

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

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

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

  800. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  815.         }

  816.     }

  817. 

  818.     // coef updates

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

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

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

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

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

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

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

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

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

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

  829.                                 break;

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

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

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

  833.                                 } else {

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

  835.                                 }

  836.                             }

  837.                             p[3] = 0;

  838.                         }

  839.         } else {

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

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

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

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

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

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

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

  847.                                 break;

  848.                             memcpy(p, r, 3);

  849.                             p[3] = 0;

  850.                         }

  851.         }

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

  853.             break;

  854.     }

  855. 

  856.     // mode updates

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

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

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

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

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

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

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

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

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

  866. 

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

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

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

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

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

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

  873. 

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

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

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

  877. 

  878.         if (s->allowcompinter) {

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

  880.             if (s->comppredmode)

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

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

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

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

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

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

  887.         } else {

  888.             s->comppredmode = PRED_SINGLEREF;

  889.         }

  890. 

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

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

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

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

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

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

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

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

  899.             }

  900.         }

  901. 

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

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

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

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

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

  907.         }

  908. 

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

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

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

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

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

  914. 

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

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

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

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

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

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

  921. 

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

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

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

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

  926. 

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

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

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

  930. 

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

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

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

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

  935. 

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

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

  938. 

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

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

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

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

  943.         }

  944. 

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

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

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

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

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

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

  951. 

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

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

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

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

  956.         }

  957. 

  958.         if (s->highprecisionmvs) {

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

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

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

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

  963. 

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

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

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

  967.             }

  968.         }

  969.     }

  970. 

  971.     return (data2 - data) + size2;

  972. }

  973. 

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

  975.                                       VP9Context *s)

  976. {

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

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

  979. }

  980. 

  981. static void find_ref_mvs(VP9Context *s,

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

  983. {

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

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

  986.                       { -1, -1 }, {  0, -1 }, { -1,  0 }, {  6, -1 }},

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

  988.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, {  2, -1 }},

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

  990.                       { -1, -1 }, { -3,  0 }, {  0, -3 }, { -1,  2 }},

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

  992.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  994.                       { -1,  1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  996.                       {  1, -1 }, { -3,  0 }, {  0, -3 }, { -3, -3 }},

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

  998.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  1000.                       {  0, -2 }, { -2,  0 }, { -2, -1 }, { -1, -2 }},

  1001.         [BS_8x16]  = {{ -1,  0 }, {  0, -1 }, { -1,  1 }, { -1, -1 },

  1002.                       { -2,  0 }, {  0, -2 }, { -1, -2 }, { -2, -1 }},

  1003.         [BS_8x8]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1004.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1005.         [BS_8x4]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1006.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1007.         [BS_4x8]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1008.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1009.         [BS_4x4]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1010.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1011.     };

  1012.     VP9Block *b = s->b;

  1013.     int row = s->row, col = s->col, row7 = s->row7;

  1014.     const int8_t (*p)[2] = mv_ref_blk_off[b->bs];

  1015. #define INVALID_MV 0x80008000U

  1016.     uint32_t mem = INVALID_MV;

  1017.     int i;

  1018. 

  1019. #define RETURN_DIRECT_MV(mv) \

  1020.     do { \

  1021.         uint32_t m = AV_RN32A(&mv); \

  1022.         if (!idx) { \

  1023.             AV_WN32A(pmv, m); \

  1024.             return; \

  1025.         } else if (mem == INVALID_MV) { \

  1026.             mem = m; \

  1027.         } else if (m != mem) { \

  1028.             AV_WN32A(pmv, m); \

  1029.             return; \

  1030.         } \

  1031.     } while (0)

  1032. 

  1033.     if (sb >= 0) {

  1034.         if (sb == 2 || sb == 1) {

  1035.             RETURN_DIRECT_MV(b->mv[0][z]);

  1036.         } else if (sb == 3) {

  1037.             RETURN_DIRECT_MV(b->mv[2][z]);

  1038.             RETURN_DIRECT_MV(b->mv[1][z]);

  1039.             RETURN_DIRECT_MV(b->mv[0][z]);

  1040.         }

  1041. 

  1042. #define RETURN_MV(mv) \

  1043.     do { \

  1044.         if (sb > 0) { \

  1045.             VP56mv tmp; \

  1046.             uint32_t m; \

  1047.             clamp_mv(&tmp, &mv, s); \

  1048.             m = AV_RN32A(&tmp); \

  1049.             if (!idx) { \

  1050.                 AV_WN32A(pmv, m); \

  1051.                 return; \

  1052.             } else if (mem == INVALID_MV) { \

  1053.                 mem = m; \

  1054.             } else if (m != mem) { \

  1055.                 AV_WN32A(pmv, m); \

  1056.                 return; \

  1057.             } \

  1058.         } else { \

  1059.             uint32_t m = AV_RN32A(&mv); \

  1060.             if (!idx) { \

  1061.                 clamp_mv(pmv, &mv, s); \

  1062.                 return; \

  1063.             } else if (mem == INVALID_MV) { \

  1064.                 mem = m; \

  1065.             } else if (m != mem) { \

  1066.                 clamp_mv(pmv, &mv, s); \

  1067.                 return; \

  1068.             } \

  1069.         } \

  1070.     } while (0)

  1071. 

  1072.         if (row > 0) {

  1073.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[(row - 1) * s->sb_cols * 8 + col];

  1074.             if (mv->ref[0] == ref) {

  1075.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][0]);

  1076.             } else if (mv->ref[1] == ref) {

  1077.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][1]);

  1078.             }

  1079.         }

  1080.         if (col > s->tiling.tile_col_start) {

  1081.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[row * s->sb_cols * 8 + col - 1];

  1082.             if (mv->ref[0] == ref) {

  1083.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][0]);

  1084.             } else if (mv->ref[1] == ref) {

  1085.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][1]);

  1086.             }

  1087.         }

  1088.         i = 2;

  1089.     } else {

  1090.         i = 0;

  1091.     }

  1092. 

  1093.     // previously coded MVs in this neighbourhood, using same reference frame

  1094.     for (; i < 8; i++) {

  1095.         int c = p[i][0] + col, r = p[i][1] + row;

  1096. 

  1097.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1098.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1099. 

  1100.             if (mv->ref[0] == ref) {

  1101.                 RETURN_MV(mv->mv[0]);

  1102.             } else if (mv->ref[1] == ref) {

  1103.                 RETURN_MV(mv->mv[1]);

  1104.             }

  1105.         }

  1106.     }

  1107. 

  1108.     // MV at this position in previous frame, using same reference frame

  1109.     if (s->use_last_frame_mvs) {

  1110.         struct VP9mvrefPair *mv = &s->frames[LAST_FRAME].mv[row * s->sb_cols * 8 + col];

  1111. 

  1112.         if (!s->last_uses_2pass)

  1113.             ff_thread_await_progress(&s->frames[LAST_FRAME].tf, row >> 3, 0);

  1114.         if (mv->ref[0] == ref) {

  1115.             RETURN_MV(mv->mv[0]);

  1116.         } else if (mv->ref[1] == ref) {

  1117.             RETURN_MV(mv->mv[1]);

  1118.         }

  1119.     }

  1120. 

  1121. #define RETURN_SCALE_MV(mv, scale) \

  1122.     do { \

  1123.         if (scale) { \

  1124.             VP56mv mv_temp = { -mv.x, -mv.y }; \

  1125.             RETURN_MV(mv_temp); \

  1126.         } else { \

  1127.             RETURN_MV(mv); \

  1128.         } \

  1129.     } while (0)

  1130. 

  1131.     // previously coded MVs in this neighbourhood, using different reference frame

  1132.     for (i = 0; i < 8; i++) {

  1133.         int c = p[i][0] + col, r = p[i][1] + row;

  1134. 

  1135.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1136.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1137. 

  1138.             if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1139.                 RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1140.             }

  1141.             if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1142.                 // BUG - libvpx has this condition regardless of whether

  1143.                 // we used the first ref MV and pre-scaling

  1144.                 AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1145.                 RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1146.             }

  1147.         }

  1148.     }

  1149. 

  1150.     // MV at this position in previous frame, using different reference frame

  1151.     if (s->use_last_frame_mvs) {

  1152.         struct VP9mvrefPair *mv = &s->frames[LAST_FRAME].mv[row * s->sb_cols * 8 + col];

  1153. 

  1154.         // no need to await_progress, because we already did that above

  1155.         if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1156.             RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1157.         }

  1158.         if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1159.             // BUG - libvpx has this condition regardless of whether

  1160.             // we used the first ref MV and pre-scaling

  1161.             AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1162.             RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1163.         }

  1164.     }

  1165. 

  1166.     AV_ZERO32(pmv);

  1167. #undef INVALID_MV

  1168. #undef RETURN_MV

  1169. #undef RETURN_SCALE_MV

  1170. }

  1171. 

  1172. static av_always_inline int read_mv_component(VP9Context *s, int idx, int hp)

  1173. {

  1174.     int bit, sign = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].sign);

  1175.     int n, c = vp8_rac_get_tree(&s->c, vp9_mv_class_tree,

  1176.                                 s->prob.p.mv_comp[idx].classes);

  1177. 

  1178.     s->counts.mv_comp[idx].sign[sign]++;

  1179.     s->counts.mv_comp[idx].classes[c]++;

  1180.     if (c) {

  1181.         int m;

  1182. 

  1183.         for (n = 0, m = 0; m < c; m++) {

  1184.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].bits[m]);

  1185.             n |= bit << m;

  1186.             s->counts.mv_comp[idx].bits[m][bit]++;

  1187.         }

  1188.         n <<= 3;

  1189.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree, s->prob.p.mv_comp[idx].fp);

  1190.         n |= bit << 1;

  1191.         s->counts.mv_comp[idx].fp[bit]++;

  1192.         if (hp) {

  1193.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].hp);

  1194.             s->counts.mv_comp[idx].hp[bit]++;

  1195.             n |= bit;

  1196.         } else {

  1197.             n |= 1;

  1198.             // bug in libvpx - we count for bw entropy purposes even if the

  1199.             // bit wasn't coded

  1200.             s->counts.mv_comp[idx].hp[1]++;

  1201.         }

  1202.         n += 8 << c;

  1203.     } else {

  1204.         n = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0);

  1205.         s->counts.mv_comp[idx].class0[n]++;

  1206.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree,

  1207.                                s->prob.p.mv_comp[idx].class0_fp[n]);

  1208.         s->counts.mv_comp[idx].class0_fp[n][bit]++;

  1209.         n = (n << 3) | (bit << 1);

  1210.         if (hp) {

  1211.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0_hp);

  1212.             s->counts.mv_comp[idx].class0_hp[bit]++;

  1213.             n |= bit;

  1214.         } else {

  1215.             n |= 1;

  1216.             // bug in libvpx - we count for bw entropy purposes even if the

  1217.             // bit wasn't coded

  1218.             s->counts.mv_comp[idx].class0_hp[1]++;

  1219.         }

  1220.     }

  1221. 

  1222.     return sign ? -(n + 1) : (n + 1);

  1223. }

  1224. 

  1225. static void fill_mv(VP9Context *s,

  1226.                     VP56mv *mv, int mode, int sb)

  1227. {

  1228.     VP9Block *b = s->b;

  1229. 

  1230.     if (mode == ZEROMV) {

  1231.         AV_ZERO64(mv);

  1232.     } else {

  1233.         int hp;

  1234. 

  1235.         // FIXME cache this value and reuse for other subblocks

  1236.         find_ref_mvs(s, &mv[0], b->ref[0], 0, mode == NEARMV,

  1237.                      mode == NEWMV ? -1 : sb);

  1238.         // FIXME maybe move this code into find_ref_mvs()

  1239.         if ((mode == NEWMV || sb == -1) &&

  1240.             !(hp = s->highprecisionmvs && abs(mv[0].x) < 64 && abs(mv[0].y) < 64)) {

  1241.             if (mv[0].y & 1) {

  1242.                 if (mv[0].y < 0)

  1243.                     mv[0].y++;

  1244.                 else

  1245.                     mv[0].y--;

  1246.             }

  1247.             if (mv[0].x & 1) {

  1248.                 if (mv[0].x < 0)

  1249.                     mv[0].x++;

  1250.                 else

  1251.                     mv[0].x--;

  1252.             }

  1253.         }

  1254.         if (mode == NEWMV) {

  1255.             enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1256.                                               s->prob.p.mv_joint);

  1257. 

  1258.             s->counts.mv_joint[j]++;

  1259.             if (j >= MV_JOINT_V)

  1260.                 mv[0].y += read_mv_component(s, 0, hp);

  1261.             if (j & 1)

  1262.                 mv[0].x += read_mv_component(s, 1, hp);

  1263.         }

  1264. 

  1265.         if (b->comp) {

  1266.             // FIXME cache this value and reuse for other subblocks

  1267.             find_ref_mvs(s, &mv[1], b->ref[1], 1, mode == NEARMV,

  1268.                          mode == NEWMV ? -1 : sb);

  1269.             if ((mode == NEWMV || sb == -1) &&

  1270.                 !(hp = s->highprecisionmvs && abs(mv[1].x) < 64 && abs(mv[1].y) < 64)) {

  1271.                 if (mv[1].y & 1) {

  1272.                     if (mv[1].y < 0)

  1273.                         mv[1].y++;

  1274.                     else

  1275.                         mv[1].y--;

  1276.                 }

  1277.                 if (mv[1].x & 1) {

  1278.                     if (mv[1].x < 0)

  1279.                         mv[1].x++;

  1280.                     else

  1281.                         mv[1].x--;

  1282.                 }

  1283.             }

  1284.             if (mode == NEWMV) {

  1285.                 enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1286.                                                   s->prob.p.mv_joint);

  1287. 

  1288.                 s->counts.mv_joint[j]++;

  1289.                 if (j >= MV_JOINT_V)

  1290.                     mv[1].y += read_mv_component(s, 0, hp);

  1291.                 if (j & 1)

  1292.                     mv[1].x += read_mv_component(s, 1, hp);

  1293.             }

  1294.         }

  1295.     }

  1296. }

  1297. 

  1298. static av_always_inline void setctx_2d(uint8_t *ptr, int w, int h,

  1299.                                        ptrdiff_t stride, int v)

  1300. {

  1301.     switch (w) {

  1302.     case 1:

  1303.         do {

  1304.             *ptr = v;

  1305.             ptr += stride;

  1306.         } while (--h);

  1307.         break;

  1308.     case 2: {

  1309.         int v16 = v * 0x0101;

  1310.         do {

  1311.             AV_WN16A(ptr, v16);

  1312.             ptr += stride;

  1313.         } while (--h);

  1314.         break;

  1315.     }

  1316.     case 4: {

  1317.         uint32_t v32 = v * 0x01010101;

  1318.         do {

  1319.             AV_WN32A(ptr, v32);

  1320.             ptr += stride;

  1321.         } while (--h);

  1322.         break;

  1323.     }

  1324.     case 8: {

  1325. #if HAVE_FAST_64BIT

  1326.         uint64_t v64 = v * 0x0101010101010101ULL;

  1327.         do {

  1328.             AV_WN64A(ptr, v64);

  1329.             ptr += stride;

  1330.         } while (--h);

  1331. #else

  1332.         uint32_t v32 = v * 0x01010101;

  1333.         do {

  1334.             AV_WN32A(ptr,     v32);

  1335.             AV_WN32A(ptr + 4, v32);

  1336.             ptr += stride;

  1337.         } while (--h);

  1338. #endif

  1339.         break;

  1340.     }

  1341.     }

  1342. }

  1343. 

  1344. static void decode_mode(AVCodecContext *ctx)

  1345. {

  1346.     static const uint8_t left_ctx[N_BS_SIZES] = {

  1347.         0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf

  1348.     };

  1349.     static const uint8_t above_ctx[N_BS_SIZES] = {

  1350.         0x0, 0x0, 0x8, 0x8, 0x8, 0xc, 0xc, 0xc, 0xe, 0xe, 0xe, 0xf, 0xf

  1351.     };

  1352.     static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {

  1353.         TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,

  1354.         TX_16X16, TX_8X8, TX_8X8, TX_8X8, TX_4X4, TX_4X4, TX_4X4

  1355.     };

  1356.     VP9Context *s = ctx->priv_data;

  1357.     VP9Block *b = s->b;

  1358.     int row = s->row, col = s->col, row7 = s->row7;

  1359.     enum TxfmMode max_tx = max_tx_for_bl_bp[b->bs];

  1360.     int w4 = FFMIN(s->cols - col, bwh_tab[1][b->bs][0]);

  1361.     int h4 = FFMIN(s->rows - row, bwh_tab[1][b->bs][1]), y;

  1362.     int have_a = row > 0, have_l = col > s->tiling.tile_col_start;

  1363.     int vref, filter_id;

  1364. 

  1365.     if (!s->segmentation.enabled) {

  1366.         b->seg_id = 0;

  1367.     } else if (s->keyframe || s->intraonly) {

  1368.         b->seg_id = vp8_rac_get_tree(&s->c, vp9_segmentation_tree, s->prob.seg);

  1369.     } else if (!s->segmentation.update_map ||

  1370.                (s->segmentation.temporal &&

  1371.                 vp56_rac_get_prob_branchy(&s->c,

  1372.                     s->prob.segpred[s->above_segpred_ctx[col] +

  1373.                                     s->left_segpred_ctx[row7]]))) {

  1374.         if (!s->errorres) {

  1375.             int pred = 8, x;

  1376.             uint8_t *refsegmap = s->frames[LAST_FRAME].segmentation_map;

  1377. 

  1378.             if (!s->last_uses_2pass)

  1379.                 ff_thread_await_progress(&s->frames[LAST_FRAME].tf, row >> 3, 0);

  1380.             for (y = 0; y < h4; y++) {

  1381.                 int idx_base = (y + row) * 8 * s->sb_cols + col;

  1382.                 for (x = 0; x < w4; x++)

  1383.                     pred = FFMIN(pred, refsegmap[idx_base + x]);

  1384.                 if (!s->segmentation.update_map && ctx->active_thread_type == FF_THREAD_FRAME) {

  1385.                     // FIXME maybe retain reference to previous frame as

  1386.                     // segmap reference instead of copying the whole map

  1387.                     // into a new buffer

  1388.                     memcpy(&s->frames[CUR_FRAME].segmentation_map[idx_base],

  1389.                            &refsegmap[idx_base], w4);

  1390.                 }

  1391.             }

  1392.             av_assert1(pred < 8);

  1393.             b->seg_id = pred;

  1394.         } else {

  1395.             b->seg_id = 0;

  1396.         }

  1397. 

  1398.         memset(&s->above_segpred_ctx[col], 1, w4);

  1399.         memset(&s->left_segpred_ctx[row7], 1, h4);

  1400.     } else {

  1401.         b->seg_id = vp8_rac_get_tree(&s->c, vp9_segmentation_tree,

  1402.                                      s->prob.seg);

  1403. 

  1404.         memset(&s->above_segpred_ctx[col], 0, w4);

  1405.         memset(&s->left_segpred_ctx[row7], 0, h4);

  1406.     }

  1407.     if (s->segmentation.enabled &&

  1408.         (s->segmentation.update_map || s->keyframe || s->intraonly)) {

  1409.         setctx_2d(&s->frames[CUR_FRAME].segmentation_map[row * 8 * s->sb_cols + col],

  1410.                   w4, h4, 8 * s->sb_cols, b->seg_id);

  1411.     }

  1412. 

  1413.     b->skip = s->segmentation.enabled &&

  1414.         s->segmentation.feat[b->seg_id].skip_enabled;

  1415.     if (!b->skip) {

  1416.         int c = s->left_skip_ctx[row7] + s->above_skip_ctx[col];

  1417.         b->skip = vp56_rac_get_prob(&s->c, s->prob.p.skip[c]);

  1418.         s->counts.skip[c][b->skip]++;

  1419.     }

  1420. 

  1421.     if (s->keyframe || s->intraonly) {

  1422.         b->intra = 1;

  1423.     } else if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1424.         b->intra = !s->segmentation.feat[b->seg_id].ref_val;

  1425.     } else {

  1426.         int c, bit;

  1427. 

  1428.         if (have_a && have_l) {

  1429.             c = s->above_intra_ctx[col] + s->left_intra_ctx[row7];

  1430.             c += (c == 2);

  1431.         } else {

  1432.             c = have_a ? 2 * s->above_intra_ctx[col] :

  1433.                 have_l ? 2 * s->left_intra_ctx[row7] : 0;

  1434.         }

  1435.         bit = vp56_rac_get_prob(&s->c, s->prob.p.intra[c]);

  1436.         s->counts.intra[c][bit]++;

  1437.         b->intra = !bit;

  1438.     }

  1439. 

  1440.     if ((b->intra || !b->skip) && s->txfmmode == TX_SWITCHABLE) {

  1441.         int c;

  1442.         if (have_a) {

  1443.             if (have_l) {

  1444.                 c = (s->above_skip_ctx[col] ? max_tx :

  1445.                      s->above_txfm_ctx[col]) +

  1446.                     (s->left_skip_ctx[row7] ? max_tx :

  1447.                      s->left_txfm_ctx[row7]) > max_tx;

  1448.             } else {

  1449.                 c = s->above_skip_ctx[col] ? 1 :

  1450.                     (s->above_txfm_ctx[col] * 2 > max_tx);

  1451.             }

  1452.         } else if (have_l) {

  1453.             c = s->left_skip_ctx[row7] ? 1 :

  1454.                 (s->left_txfm_ctx[row7] * 2 > max_tx);

  1455.         } else {

  1456.             c = 1;

  1457.         }

  1458.         switch (max_tx) {

  1459.         case TX_32X32:

  1460.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][0]);

  1461.             if (b->tx) {

  1462.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][1]);

  1463.                 if (b->tx == 2)

  1464.                     b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][2]);

  1465.             }

  1466.             s->counts.tx32p[c][b->tx]++;

  1467.             break;

  1468.         case TX_16X16:

  1469.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][0]);

  1470.             if (b->tx)

  1471.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][1]);

  1472.             s->counts.tx16p[c][b->tx]++;

  1473.             break;

  1474.         case TX_8X8:

  1475.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx8p[c]);

  1476.             s->counts.tx8p[c][b->tx]++;

  1477.             break;

  1478.         case TX_4X4:

  1479.             b->tx = TX_4X4;

  1480.             break;

  1481.         }

  1482.     } else {

  1483.         b->tx = FFMIN(max_tx, s->txfmmode);

  1484.     }

  1485. 

  1486.     if (s->keyframe || s->intraonly) {

  1487.         uint8_t *a = &s->above_mode_ctx[col * 2];

  1488.         uint8_t *l = &s->left_mode_ctx[(row7) << 1];

  1489. 

  1490.         b->comp = 0;

  1491.         if (b->bs > BS_8x8) {

  1492.             // FIXME the memory storage intermediates here aren't really

  1493.             // necessary, they're just there to make the code slightly

  1494.             // simpler for now

  1495.             b->mode[0] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1496.                                     vp9_default_kf_ymode_probs[a[0]][l[0]]);

  1497.             if (b->bs != BS_8x4) {

  1498.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1499.                                  vp9_default_kf_ymode_probs[a[1]][b->mode[0]]);

  1500.                 l[0] = a[1] = b->mode[1];

  1501.             } else {

  1502.                 l[0] = a[1] = b->mode[1] = b->mode[0];

  1503.             }

  1504.             if (b->bs != BS_4x8) {

  1505.                 b->mode[2] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1506.                                         vp9_default_kf_ymode_probs[a[0]][l[1]]);

  1507.                 if (b->bs != BS_8x4) {

  1508.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1509.                                   vp9_default_kf_ymode_probs[a[1]][b->mode[2]]);

  1510.                     l[1] = a[1] = b->mode[3];

  1511.                 } else {

  1512.                     l[1] = a[1] = b->mode[3] = b->mode[2];

  1513.                 }

  1514.             } else {

  1515.                 b->mode[2] = b->mode[0];

  1516.                 l[1] = a[1] = b->mode[3] = b->mode[1];

  1517.             }

  1518.         } else {

  1519.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1520.                                           vp9_default_kf_ymode_probs[*a][*l]);

  1521.             b->mode[3] = b->mode[2] = b->mode[1] = b->mode[0];

  1522.             // FIXME this can probably be optimized

  1523.             memset(a, b->mode[0], bwh_tab[0][b->bs][0]);

  1524.             memset(l, b->mode[0], bwh_tab[0][b->bs][1]);

  1525.         }

  1526.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1527.                                      vp9_default_kf_uvmode_probs[b->mode[3]]);

  1528.     } else if (b->intra) {

  1529.         b->comp = 0;

  1530.         if (b->bs > BS_8x8) {

  1531.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1532.                                           s->prob.p.y_mode[0]);

  1533.             s->counts.y_mode[0][b->mode[0]]++;

  1534.             if (b->bs != BS_8x4) {

  1535.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1536.                                               s->prob.p.y_mode[0]);

  1537.                 s->counts.y_mode[0][b->mode[1]]++;

  1538.             } else {

  1539.                 b->mode[1] = b->mode[0];

  1540.             }

  1541.             if (b->bs != BS_4x8) {

  1542.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1543.                                               s->prob.p.y_mode[0]);

  1544.                 s->counts.y_mode[0][b->mode[2]]++;

  1545.                 if (b->bs != BS_8x4) {

  1546.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1547.                                                   s->prob.p.y_mode[0]);

  1548.                     s->counts.y_mode[0][b->mode[3]]++;

  1549.                 } else {

  1550.                     b->mode[3] = b->mode[2];

  1551.                 }

  1552.             } else {

  1553.                 b->mode[2] = b->mode[0];

  1554.                 b->mode[3] = b->mode[1];

  1555.             }

  1556.         } else {

  1557.             static const uint8_t size_group[10] = {

  1558.                 3, 3, 3, 3, 2, 2, 2, 1, 1, 1

  1559.             };

  1560.             int sz = size_group[b->bs];

  1561. 

  1562.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1563.                                           s->prob.p.y_mode[sz]);

  1564.             b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1565.             s->counts.y_mode[sz][b->mode[3]]++;

  1566.         }

  1567.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1568.                                      s->prob.p.uv_mode[b->mode[3]]);

  1569.         s->counts.uv_mode[b->mode[3]][b->uvmode]++;

  1570.     } else {

  1571.         static const uint8_t inter_mode_ctx_lut[14][14] = {

  1572.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1573.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1574.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1575.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1576.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1577.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1578.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1579.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1580.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1581.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1582.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1583.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1584.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },

  1585.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },

  1586.         };

  1587. 

  1588.         if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1589.             av_assert2(s->segmentation.feat[b->seg_id].ref_val != 0);

  1590.             b->comp = 0;

  1591.             b->ref[0] = s->segmentation.feat[b->seg_id].ref_val - 1;

  1592.         } else {

  1593.             // read comp_pred flag

  1594.             if (s->comppredmode != PRED_SWITCHABLE) {

  1595.                 b->comp = s->comppredmode == PRED_COMPREF;

  1596.             } else {

  1597.                 int c;

  1598. 

  1599.                 // FIXME add intra as ref=0xff (or -1) to make these easier?

  1600.                 if (have_a) {

  1601.                     if (have_l) {

  1602.                         if (s->above_comp_ctx[col] && s->left_comp_ctx[row7]) {

  1603.                             c = 4;

  1604.                         } else if (s->above_comp_ctx[col]) {

  1605.                             c = 2 + (s->left_intra_ctx[row7] ||

  1606.                                      s->left_ref_ctx[row7] == s->fixcompref);

  1607.                         } else if (s->left_comp_ctx[row7]) {

  1608.                             c = 2 + (s->above_intra_ctx[col] ||

  1609.                                      s->above_ref_ctx[col] == s->fixcompref);

  1610.                         } else {

  1611.                             c = (!s->above_intra_ctx[col] &&

  1612.                                  s->above_ref_ctx[col] == s->fixcompref) ^

  1613.                             (!s->left_intra_ctx[row7] &&

  1614.                              s->left_ref_ctx[row & 7] == s->fixcompref);

  1615.                         }

  1616.                     } else {

  1617.                         c = s->above_comp_ctx[col] ? 3 :

  1618.                         (!s->above_intra_ctx[col] && s->above_ref_ctx[col] == s->fixcompref);

  1619.                     }

  1620.                 } else if (have_l) {

  1621.                     c = s->left_comp_ctx[row7] ? 3 :

  1622.                     (!s->left_intra_ctx[row7] && s->left_ref_ctx[row7] == s->fixcompref);

  1623.                 } else {

  1624.                     c = 1;

  1625.                 }

  1626.                 b->comp = vp56_rac_get_prob(&s->c, s->prob.p.comp[c]);

  1627.                 s->counts.comp[c][b->comp]++;

  1628.             }

  1629. 

  1630.             // read actual references

  1631.             // FIXME probably cache a few variables here to prevent repetitive

  1632.             // memory accesses below

  1633.             if (b->comp) /* two references */ {

  1634.                 int fix_idx = s->signbias[s->fixcompref], var_idx = !fix_idx, c, bit;

  1635. 

  1636.                 b->ref[fix_idx] = s->fixcompref;

  1637.                 // FIXME can this codeblob be replaced by some sort of LUT?

  1638.                 if (have_a) {

  1639.                     if (have_l) {

  1640.                         if (s->above_intra_ctx[col]) {

  1641.                             if (s->left_intra_ctx[row7]) {

  1642.                                 c = 2;

  1643.                             } else {

  1644.                                 c = 1 + 2 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1645.                             }

  1646.                         } else if (s->left_intra_ctx[row7]) {

  1647.                             c = 1 + 2 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1648.                         } else {

  1649.                             int refl = s->left_ref_ctx[row7], refa = s->above_ref_ctx[col];

  1650. 

  1651.                             if (refl == refa && refa == s->varcompref[1]) {

  1652.                                 c = 0;

  1653.                             } else if (!s->left_comp_ctx[row7] && !s->above_comp_ctx[col]) {

  1654.                                 if ((refa == s->fixcompref && refl == s->varcompref[0]) ||

  1655.                                     (refl == s->fixcompref && refa == s->varcompref[0])) {

  1656.                                     c = 4;

  1657.                                 } else {

  1658.                                     c = (refa == refl) ? 3 : 1;

  1659.                                 }

  1660.                             } else if (!s->left_comp_ctx[row7]) {

  1661.                                 if (refa == s->varcompref[1] && refl != s->varcompref[1]) {

  1662.                                     c = 1;

  1663.                                 } else {

  1664.                                     c = (refl == s->varcompref[1] &&

  1665.                                          refa != s->varcompref[1]) ? 2 : 4;

  1666.                                 }

  1667.                             } else if (!s->above_comp_ctx[col]) {

  1668.                                 if (refl == s->varcompref[1] && refa != s->varcompref[1]) {

  1669.                                     c = 1;

  1670.                                 } else {

  1671.                                     c = (refa == s->varcompref[1] &&

  1672.                                          refl != s->varcompref[1]) ? 2 : 4;

  1673.                                 }

  1674.                             } else {

  1675.                                 c = (refl == refa) ? 4 : 2;

  1676.                             }

  1677.                         }

  1678.                     } else {

  1679.                         if (s->above_intra_ctx[col]) {

  1680.                             c = 2;

  1681.                         } else if (s->above_comp_ctx[col]) {

  1682.                             c = 4 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1683.                         } else {

  1684.                             c = 3 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1685.                         }

  1686.                     }

  1687.                 } else if (have_l) {

  1688.                     if (s->left_intra_ctx[row7]) {

  1689.                         c = 2;

  1690.                     } else if (s->left_comp_ctx[row7]) {

  1691.                         c = 4 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1692.                     } else {

  1693.                         c = 3 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1694.                     }

  1695.                 } else {

  1696.                     c = 2;

  1697.                 }

  1698.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.comp_ref[c]);

  1699.                 b->ref[var_idx] = s->varcompref[bit];

  1700.                 s->counts.comp_ref[c][bit]++;

  1701.             } else /* single reference */ {

  1702.                 int bit, c;

  1703. 

  1704.                 if (have_a && !s->above_intra_ctx[col]) {

  1705.                     if (have_l && !s->left_intra_ctx[row7]) {

  1706.                         if (s->left_comp_ctx[row7]) {

  1707.                             if (s->above_comp_ctx[col]) {

  1708.                                 c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7] ||

  1709.                                          !s->above_ref_ctx[col]);

  1710.                             } else {

  1711.                                 c = (3 * !s->above_ref_ctx[col]) +

  1712.                                     (!s->fixcompref || !s->left_ref_ctx[row7]);

  1713.                             }

  1714.                         } else if (s->above_comp_ctx[col]) {

  1715.                             c = (3 * !s->left_ref_ctx[row7]) +

  1716.                                 (!s->fixcompref || !s->above_ref_ctx[col]);

  1717.                         } else {

  1718.                             c = 2 * !s->left_ref_ctx[row7] + 2 * !s->above_ref_ctx[col];

  1719.                         }

  1720.                     } else if (s->above_intra_ctx[col]) {

  1721.                         c = 2;

  1722.                     } else if (s->above_comp_ctx[col]) {

  1723.                         c = 1 + (!s->fixcompref || !s->above_ref_ctx[col]);

  1724.                     } else {

  1725.                         c = 4 * (!s->above_ref_ctx[col]);

  1726.                     }

  1727.                 } else if (have_l && !s->left_intra_ctx[row7]) {

  1728.                     if (s->left_intra_ctx[row7]) {

  1729.                         c = 2;

  1730.                     } else if (s->left_comp_ctx[row7]) {

  1731.                         c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7]);

  1732.                     } else {

  1733.                         c = 4 * (!s->left_ref_ctx[row7]);

  1734.                     }

  1735.                 } else {

  1736.                     c = 2;

  1737.                 }

  1738.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][0]);

  1739.                 s->counts.single_ref[c][0][bit]++;

  1740.                 if (!bit) {

  1741.                     b->ref[0] = 0;

  1742.                 } else {

  1743.                     // FIXME can this codeblob be replaced by some sort of LUT?

  1744.                     if (have_a) {

  1745.                         if (have_l) {

  1746.                             if (s->left_intra_ctx[row7]) {

  1747.                                 if (s->above_intra_ctx[col]) {

  1748.                                     c = 2;

  1749.                                 } else if (s->above_comp_ctx[col]) {

  1750.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1751.                                                  s->above_ref_ctx[col] == 1);

  1752.                                 } else if (!s->above_ref_ctx[col]) {

  1753.                                     c = 3;

  1754.                                 } else {

  1755.                                     c = 4 * (s->above_ref_ctx[col] == 1);

  1756.                                 }

  1757.                             } else if (s->above_intra_ctx[col]) {

  1758.                                 if (s->left_intra_ctx[row7]) {

  1759.                                     c = 2;

  1760.                                 } else if (s->left_comp_ctx[row7]) {

  1761.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1762.                                                  s->left_ref_ctx[row7] == 1);

  1763.                                 } else if (!s->left_ref_ctx[row7]) {

  1764.                                     c = 3;

  1765.                                 } else {

  1766.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1767.                                 }

  1768.                             } else if (s->above_comp_ctx[col]) {

  1769.                                 if (s->left_comp_ctx[row7]) {

  1770.                                     if (s->left_ref_ctx[row7] == s->above_ref_ctx[col]) {

  1771.                                         c = 3 * (s->fixcompref == 1 ||

  1772.                                                  s->left_ref_ctx[row7] == 1);

  1773.                                     } else {

  1774.                                         c = 2;

  1775.                                     }

  1776.                                 } else if (!s->left_ref_ctx[row7]) {

  1777.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1778.                                                  s->above_ref_ctx[col] == 1);

  1779.                                 } else {

  1780.                                     c = 3 * (s->left_ref_ctx[row7] == 1) +

  1781.                                     (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1782.                                 }

  1783.                             } else if (s->left_comp_ctx[row7]) {

  1784.                                 if (!s->above_ref_ctx[col]) {

  1785.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1786.                                                  s->left_ref_ctx[row7] == 1);

  1787.                                 } else {

  1788.                                     c = 3 * (s->above_ref_ctx[col] == 1) +

  1789.                                     (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1790.                                 }

  1791.                             } else if (!s->above_ref_ctx[col]) {

  1792.                                 if (!s->left_ref_ctx[row7]) {

  1793.                                     c = 3;

  1794.                                 } else {

  1795.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1796.                                 }

  1797.                             } else if (!s->left_ref_ctx[row7]) {

  1798.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1799.                             } else {

  1800.                                 c = 2 * (s->left_ref_ctx[row7] == 1) +

  1801.                                 2 * (s->above_ref_ctx[col] == 1);

  1802.                             }

  1803.                         } else {

  1804.                             if (s->above_intra_ctx[col] ||

  1805.                                 (!s->above_comp_ctx[col] && !s->above_ref_ctx[col])) {

  1806.                                 c = 2;

  1807.                             } else if (s->above_comp_ctx[col]) {

  1808.                                 c = 3 * (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1809.                             } else {

  1810.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1811.                             }

  1812.                         }

  1813.                     } else if (have_l) {

  1814.                         if (s->left_intra_ctx[row7] ||

  1815.                             (!s->left_comp_ctx[row7] && !s->left_ref_ctx[row7])) {

  1816.                             c = 2;

  1817.                         } else if (s->left_comp_ctx[row7]) {

  1818.                             c = 3 * (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1819.                         } else {

  1820.                             c = 4 * (s->left_ref_ctx[row7] == 1);

  1821.                         }

  1822.                     } else {

  1823.                         c = 2;

  1824.                     }

  1825.                     bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][1]);

  1826.                     s->counts.single_ref[c][1][bit]++;

  1827.                     b->ref[0] = 1 + bit;

  1828.                 }

  1829.             }

  1830.         }

  1831. 

  1832.         if (b->bs <= BS_8x8) {

  1833.             if (s->segmentation.feat[b->seg_id].skip_enabled) {

  1834.                 b->mode[0] = b->mode[1] = b->mode[2] = b->mode[3] = ZEROMV;

  1835.             } else {

  1836.                 static const uint8_t off[10] = {

  1837.                     3, 0, 0, 1, 0, 0, 0, 0, 0, 0

  1838.                 };

  1839. 

  1840.                 // FIXME this needs to use the LUT tables from find_ref_mvs

  1841.                 // because not all are -1,0/0,-1

  1842.                 int c = inter_mode_ctx_lut[s->above_mode_ctx[col + off[b->bs]]]

  1843.                                           [s->left_mode_ctx[row7 + off[b->bs]]];

  1844. 

  1845.                 b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1846.                                               s->prob.p.mv_mode[c]);

  1847.                 b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1848.                 s->counts.mv_mode[c][b->mode[0] - 10]++;

  1849.             }

  1850.         }

  1851. 

  1852.         if (s->filtermode == FILTER_SWITCHABLE) {

  1853.             int c;

  1854. 

  1855.             if (have_a && s->above_mode_ctx[col] >= NEARESTMV) {

  1856.                 if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1857.                     c = s->above_filter_ctx[col] == s->left_filter_ctx[row7] ?

  1858.                         s->left_filter_ctx[row7] : 3;

  1859.                 } else {

  1860.                     c = s->above_filter_ctx[col];

  1861.                 }

  1862.             } else if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1863.                 c = s->left_filter_ctx[row7];

  1864.             } else {

  1865.                 c = 3;

  1866.             }

  1867. 

  1868.             filter_id = vp8_rac_get_tree(&s->c, vp9_filter_tree,

  1869.                                          s->prob.p.filter[c]);

  1870.             s->counts.filter[c][filter_id]++;

  1871.             b->filter = vp9_filter_lut[filter_id];

  1872.         } else {

  1873.             b->filter = s->filtermode;

  1874.         }

  1875. 

  1876.         if (b->bs > BS_8x8) {

  1877.             int c = inter_mode_ctx_lut[s->above_mode_ctx[col]][s->left_mode_ctx[row7]];

  1878. 

  1879.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1880.                                           s->prob.p.mv_mode[c]);

  1881.             s->counts.mv_mode[c][b->mode[0] - 10]++;

  1882.             fill_mv(s, b->mv[0], b->mode[0], 0);

  1883. 

  1884.             if (b->bs != BS_8x4) {

  1885.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1886.                                               s->prob.p.mv_mode[c]);

  1887.                 s->counts.mv_mode[c][b->mode[1] - 10]++;

  1888.                 fill_mv(s, b->mv[1], b->mode[1], 1);

  1889.             } else {

  1890.                 b->mode[1] = b->mode[0];

  1891.                 AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1892.                 AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1893.             }

  1894. 

  1895.             if (b->bs != BS_4x8) {

  1896.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1897.                                               s->prob.p.mv_mode[c]);

  1898.                 s->counts.mv_mode[c][b->mode[2] - 10]++;

  1899.                 fill_mv(s, b->mv[2], b->mode[2], 2);

  1900. 

  1901.                 if (b->bs != BS_8x4) {

  1902.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1903.                                                   s->prob.p.mv_mode[c]);

  1904.                     s->counts.mv_mode[c][b->mode[3] - 10]++;

  1905.                     fill_mv(s, b->mv[3], b->mode[3], 3);

  1906.                 } else {

  1907.                     b->mode[3] = b->mode[2];

  1908.                     AV_COPY32(&b->mv[3][0], &b->mv[2][0]);

  1909.                     AV_COPY32(&b->mv[3][1], &b->mv[2][1]);

  1910.                 }

  1911.             } else {

  1912.                 b->mode[2] = b->mode[0];

  1913.                 AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1914.                 AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1915.                 b->mode[3] = b->mode[1];

  1916.                 AV_COPY32(&b->mv[3][0], &b->mv[1][0]);

  1917.                 AV_COPY32(&b->mv[3][1], &b->mv[1][1]);

  1918.             }

  1919.         } else {

  1920.             fill_mv(s, b->mv[0], b->mode[0], -1);

  1921.             AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1922.             AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1923.             AV_COPY32(&b->mv[3][0], &b->mv[0][0]);

  1924.             AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1925.             AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1926.             AV_COPY32(&b->mv[3][1], &b->mv[0][1]);

  1927.         }

  1928. 

  1929.         vref = b->ref[b->comp ? s->signbias[s->varcompref[0]] : 0];

  1930.     }

  1931. 

  1932. #if HAVE_FAST_64BIT

  1933. #define SPLAT_CTX(var, val, n) \

  1934.     switch (n) { \

  1935.     case 1:  var = val;                                    break; \

  1936.     case 2:  AV_WN16A(&var, val *             0x0101);     break; \

  1937.     case 4:  AV_WN32A(&var, val *         0x01010101);     break; \

  1938.     case 8:  AV_WN64A(&var, val * 0x0101010101010101ULL);  break; \

  1939.     case 16: { \

  1940.         uint64_t v64 = val * 0x0101010101010101ULL; \

  1941.         AV_WN64A(              &var,     v64); \

  1942.         AV_WN64A(&((uint8_t *) &var)[8], v64); \

  1943.         break; \

  1944.     } \

  1945.     }

  1946. #else

  1947. #define SPLAT_CTX(var, val, n) \

  1948.     switch (n) { \

  1949.     case 1:  var = val;                         break; \

  1950.     case 2:  AV_WN16A(&var, val *     0x0101);  break; \

  1951.     case 4:  AV_WN32A(&var, val * 0x01010101);  break; \

  1952.     case 8: { \

  1953.         uint32_t v32 = val * 0x01010101; \

  1954.         AV_WN32A(              &var,     v32); \

  1955.         AV_WN32A(&((uint8_t *) &var)[4], v32); \

  1956.         break; \

  1957.     } \

  1958.     case 16: { \

  1959.         uint32_t v32 = val * 0x01010101; \

  1960.         AV_WN32A(              &var,      v32); \

  1961.         AV_WN32A(&((uint8_t *) &var)[4],  v32); \

  1962.         AV_WN32A(&((uint8_t *) &var)[8],  v32); \

  1963.         AV_WN32A(&((uint8_t *) &var)[12], v32); \

  1964.         break; \

  1965.     } \

  1966.     }

  1967. #endif

  1968. 

  1969.     switch (bwh_tab[1][b->bs][0]) {

  1970. #define SET_CTXS(dir, off, n) \

  1971.     do { \

  1972.         SPLAT_CTX(s->dir##_skip_ctx[off],      b->skip,          n); \

  1973.         SPLAT_CTX(s->dir##_txfm_ctx[off],      b->tx,            n); \

  1974.         SPLAT_CTX(s->dir##_partition_ctx[off], dir##_ctx[b->bs], n); \

  1975.         if (!s->keyframe && !s->intraonly) { \

  1976.             SPLAT_CTX(s->dir##_intra_ctx[off], b->intra,   n); \

  1977.             SPLAT_CTX(s->dir##_comp_ctx[off],  b->comp,    n); \

  1978.             SPLAT_CTX(s->dir##_mode_ctx[off],  b->mode[3], n); \

  1979.             if (!b->intra) { \

  1980.                 SPLAT_CTX(s->dir##_ref_ctx[off], vref, n); \

  1981.                 if (s->filtermode == FILTER_SWITCHABLE) { \

  1982.                     SPLAT_CTX(s->dir##_filter_ctx[off], filter_id, n); \

  1983.                 } \

  1984.             } \

  1985.         } \

  1986.     } while (0)

  1987.     case 1: SET_CTXS(above, col, 1); break;

  1988.     case 2: SET_CTXS(above, col, 2); break;

  1989.     case 4: SET_CTXS(above, col, 4); break;

  1990.     case 8: SET_CTXS(above, col, 8); break;

  1991.     }

  1992.     switch (bwh_tab[1][b->bs][1]) {

  1993.     case 1: SET_CTXS(left, row7, 1); break;

  1994.     case 2: SET_CTXS(left, row7, 2); break;

  1995.     case 4: SET_CTXS(left, row7, 4); break;

  1996.     case 8: SET_CTXS(left, row7, 8); break;

  1997.     }

  1998. #undef SPLAT_CTX

  1999. #undef SET_CTXS

  2000. 

  2001.     if (!s->keyframe && !s->intraonly) {

  2002.         if (b->bs > BS_8x8) {

  2003.             int mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  2004. 

  2005.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][0], &b->mv[1][0]);

  2006.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][1], &b->mv[1][1]);

  2007.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][0], mv0);

  2008.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][1], mv1);

  2009.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][0], &b->mv[2][0]);

  2010.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][1], &b->mv[2][1]);

  2011.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][0], mv0);

  2012.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][1], mv1);

  2013.         } else {

  2014.             int n, mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  2015. 

  2016.             for (n = 0; n < w4 * 2; n++) {

  2017.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][0], mv0);

  2018.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][1], mv1);

  2019.             }

  2020.             for (n = 0; n < h4 * 2; n++) {

  2021.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][0], mv0);

  2022.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][1], mv1);

  2023.             }

  2024.         }

  2025.     }

  2026. 

  2027.     // FIXME kinda ugly

  2028.     for (y = 0; y < h4; y++) {

  2029.         int x, o = (row + y) * s->sb_cols * 8 + col;

  2030.         struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[o];

  2031. 

  2032.         if (b->intra) {

  2033.             for (x = 0; x < w4; x++) {

  2034.                 mv[x].ref[0] =

  2035.                 mv[x].ref[1] = -1;

  2036.             }

  2037.         } else if (b->comp) {

  2038.             for (x = 0; x < w4; x++) {

  2039.                 mv[x].ref[0] = b->ref[0];

  2040.                 mv[x].ref[1] = b->ref[1];

  2041.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  2042.                 AV_COPY32(&mv[x].mv[1], &b->mv[3][1]);

  2043.             }

  2044.         } else {

  2045.             for (x = 0; x < w4; x++) {

  2046.                 mv[x].ref[0] = b->ref[0];

  2047.                 mv[x].ref[1] = -1;

  2048.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  2049.             }

  2050.         }

  2051.     }

  2052. }

  2053. 

  2054. // FIXME merge cnt/eob arguments?

  2055. static av_always_inline int

  2056. decode_coeffs_b_generic(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  2057.                         int is_tx32x32, unsigned (*cnt)[6][3],

  2058.                         unsigned (*eob)[6][2], uint8_t (*p)[6][11],

  2059.                         int nnz, const int16_t *scan, const int16_t (*nb)[2],

  2060.                         const int16_t *band_counts, const int16_t *qmul)

  2061. {

  2062.     int i = 0, band = 0, band_left = band_counts[band];

  2063.     uint8_t *tp = p[0][nnz];

  2064.     uint8_t cache[1024];

  2065. 

  2066.     do {

  2067.         int val, rc;

  2068. 

  2069.         val = vp56_rac_get_prob_branchy(c, tp[0]); // eob

  2070.         eob[band][nnz][val]++;

  2071.         if (!val)

  2072.             break;

  2073. 

  2074.     skip_eob:

  2075.         if (!vp56_rac_get_prob_branchy(c, tp[1])) { // zero

  2076.             cnt[band][nnz][0]++;

  2077.             if (!--band_left)

  2078.                 band_left = band_counts[++band];

  2079.             cache[scan[i]] = 0;

  2080.             nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  2081.             tp = p[band][nnz];

  2082.             if (++i == n_coeffs)

  2083.                 break; //invalid input; blocks should end with EOB

  2084.             goto skip_eob;

  2085.         }

  2086. 

  2087.         rc = scan[i];

  2088.         if (!vp56_rac_get_prob_branchy(c, tp[2])) { // one

  2089.             cnt[band][nnz][1]++;

  2090.             val = 1;

  2091.             cache[rc] = 1;

  2092.         } else {

  2093.             // fill in p[3-10] (model fill) - only once per frame for each pos

  2094.             if (!tp[3])

  2095.                 memcpy(&tp[3], vp9_model_pareto8[tp[2]], 8);

  2096. 

  2097.             cnt[band][nnz][2]++;

  2098.             if (!vp56_rac_get_prob_branchy(c, tp[3])) { // 2, 3, 4

  2099.                 if (!vp56_rac_get_prob_branchy(c, tp[4])) {

  2100.                     cache[rc] = val = 2;

  2101.                 } else {

  2102.                     val = 3 + vp56_rac_get_prob(c, tp[5]);

  2103.                     cache[rc] = 3;

  2104.                 }

  2105.             } else if (!vp56_rac_get_prob_branchy(c, tp[6])) { // cat1/2

  2106.                 cache[rc] = 4;

  2107.                 if (!vp56_rac_get_prob_branchy(c, tp[7])) {

  2108.                     val = 5 + vp56_rac_get_prob(c, 159);

  2109.                 } else {

  2110.                     val  = 7 + (vp56_rac_get_prob(c, 165) << 1);

  2111.                     val +=      vp56_rac_get_prob(c, 145);

  2112.                 }

  2113.             } else { // cat 3-6

  2114.                 cache[rc] = 5;

  2115.                 if (!vp56_rac_get_prob_branchy(c, tp[8])) {

  2116.                     if (!vp56_rac_get_prob_branchy(c, tp[9])) {

  2117.                         val  = 11 + (vp56_rac_get_prob(c, 173) << 2);

  2118.                         val +=      (vp56_rac_get_prob(c, 148) << 1);

  2119.                         val +=       vp56_rac_get_prob(c, 140);

  2120.                     } else {

  2121.                         val  = 19 + (vp56_rac_get_prob(c, 176) << 3);

  2122.                         val +=      (vp56_rac_get_prob(c, 155) << 2);

  2123.                         val +=      (vp56_rac_get_prob(c, 140) << 1);

  2124.                         val +=       vp56_rac_get_prob(c, 135);

  2125.                     }

  2126.                 } else if (!vp56_rac_get_prob_branchy(c, tp[10])) {

  2127.                     val  = 35 + (vp56_rac_get_prob(c, 180) << 4);

  2128.                     val +=      (vp56_rac_get_prob(c, 157) << 3);

  2129.                     val +=      (vp56_rac_get_prob(c, 141) << 2);

  2130.                     val +=      (vp56_rac_get_prob(c, 134) << 1);

  2131.                     val +=       vp56_rac_get_prob(c, 130);

  2132.                 } else {

  2133.                     val  = 67 + (vp56_rac_get_prob(c, 254) << 13);

  2134.                     val +=      (vp56_rac_get_prob(c, 254) << 12);

  2135.                     val +=      (vp56_rac_get_prob(c, 254) << 11);

  2136.                     val +=      (vp56_rac_get_prob(c, 252) << 10);

  2137.                     val +=      (vp56_rac_get_prob(c, 249) << 9);

  2138.                     val +=      (vp56_rac_get_prob(c, 243) << 8);

  2139.                     val +=      (vp56_rac_get_prob(c, 230) << 7);

  2140.                     val +=      (vp56_rac_get_prob(c, 196) << 6);

  2141.                     val +=      (vp56_rac_get_prob(c, 177) << 5);

  2142.                     val +=      (vp56_rac_get_prob(c, 153) << 4);

  2143.                     val +=      (vp56_rac_get_prob(c, 140) << 3);

  2144.                     val +=      (vp56_rac_get_prob(c, 133) << 2);

  2145.                     val +=      (vp56_rac_get_prob(c, 130) << 1);

  2146.                     val +=       vp56_rac_get_prob(c, 129);

  2147.                 }

  2148.             }

  2149.         }

  2150.         if (!--band_left)

  2151.             band_left = band_counts[++band];

  2152.         if (is_tx32x32)

  2153.             coef[rc] = ((vp8_rac_get(c) ? -val : val) * qmul[!!i]) / 2;

  2154.         else

  2155.             coef[rc] = (vp8_rac_get(c) ? -val : val) * qmul[!!i];

  2156.         nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  2157.         tp = p[band][nnz];

  2158.     } while (++i < n_coeffs);

  2159. 

  2160.     return i;

  2161. }

  2162. 

  2163. static int decode_coeffs_b(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  2164.                            unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  2165.                            uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  2166.                            const int16_t (*nb)[2], const int16_t *band_counts,

  2167.                            const int16_t *qmul)

  2168. {

  2169.     return decode_coeffs_b_generic(c, coef, n_coeffs, 0, cnt, eob, p,

  2170.                                    nnz, scan, nb, band_counts, qmul);

  2171. }

  2172. 

  2173. static int decode_coeffs_b32(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  2174.                              unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  2175.                              uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  2176.                              const int16_t (*nb)[2], const int16_t *band_counts,

  2177.                              const int16_t *qmul)

  2178. {

  2179.     return decode_coeffs_b_generic(c, coef, n_coeffs, 1, cnt, eob, p,

  2180.                                    nnz, scan, nb, band_counts, qmul);

  2181. }

  2182. 

  2183. static void decode_coeffs(AVCodecContext *ctx)

  2184. {

  2185.     VP9Context *s = ctx->priv_data;

  2186.     VP9Block *b = s->b;

  2187.     int row = s->row, col = s->col;

  2188.     uint8_t (*p)[6][11] = s->prob.coef[b->tx][0 /* y */][!b->intra];

  2189.     unsigned (*c)[6][3] = s->counts.coef[b->tx][0 /* y */][!b->intra];

  2190.     unsigned (*e)[6][2] = s->counts.eob[b->tx][0 /* y */][!b->intra];

  2191.     int w4 = bwh_tab[1][b->bs][0] << 1, h4 = bwh_tab[1][b->bs][1] << 1;

  2192.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2193.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2194.     int n, pl, x, y, res;

  2195.     int16_t (*qmul)[2] = s->segmentation.feat[b->seg_id].qmul;

  2196.     int tx = 4 * s->lossless + b->tx;

  2197.     const int16_t * const *yscans = vp9_scans[tx];

  2198.     const int16_t (* const *ynbs)[2] = vp9_scans_nb[tx];

  2199.     const int16_t *uvscan = vp9_scans[b->uvtx][DCT_DCT];

  2200.     const int16_t (*uvnb)[2] = vp9_scans_nb[b->uvtx][DCT_DCT];

  2201.     uint8_t *a = &s->above_y_nnz_ctx[col * 2];

  2202.     uint8_t *l = &s->left_y_nnz_ctx[(row & 7) << 1];

  2203.     static const int16_t band_counts[4][8] = {

  2204.         { 1, 2, 3, 4,  3,   16 - 13 },

  2205.         { 1, 2, 3, 4, 11,   64 - 21 },

  2206.         { 1, 2, 3, 4, 11,  256 - 21 },

  2207.         { 1, 2, 3, 4, 11, 1024 - 21 },

  2208.     };

  2209.     const int16_t *y_band_counts = band_counts[b->tx];

  2210.     const int16_t *uv_band_counts = band_counts[b->uvtx];

  2211. 

  2212. #define MERGE(la, end, step, rd) \

  2213.     for (n = 0; n < end; n += step) \

  2214.         la[n] = !!rd(&la[n])

  2215. #define MERGE_CTX(step, rd) \

  2216.     do { \

  2217.         MERGE(l, end_y, step, rd); \

  2218.         MERGE(a, end_x, step, rd); \

  2219.     } while (0)

  2220. 

  2221. #define DECODE_Y_COEF_LOOP(step, mode_index, v) \

  2222.     for (n = 0, y = 0; y < end_y; y += step) { \

  2223.         for (x = 0; x < end_x; x += step, n += step * step) { \

  2224.             enum TxfmType txtp = vp9_intra_txfm_type[b->mode[mode_index]]; \

  2225.             res = decode_coeffs_b##v(&s->c, s->block + 16 * n, 16 * step * step, \

  2226.                                      c, e, p, a[x] + l[y], yscans[txtp], \

  2227.                                      ynbs[txtp], y_band_counts, qmul[0]); \

  2228.             a[x] = l[y] = !!res; \

  2229.             if (step >= 4) { \

  2230.                 AV_WN16A(&s->eob[n], res); \

  2231.             } else { \

  2232.                 s->eob[n] = res; \

  2233.             } \

  2234.         } \

  2235.     }

  2236. 

  2237. #define SPLAT(la, end, step, cond) \

  2238.     if (step == 2) { \

  2239.         for (n = 1; n < end; n += step) \

  2240.             la[n] = la[n - 1]; \

  2241.     } else if (step == 4) { \

  2242.         if (cond) { \

  2243.             for (n = 0; n < end; n += step) \

  2244.                 AV_WN32A(&la[n], la[n] * 0x01010101); \

  2245.         } else { \

  2246.             for (n = 0; n < end; n += step) \

  2247.                 memset(&la[n + 1], la[n], FFMIN(end - n - 1, 3)); \

  2248.         } \

  2249.     } else /* step == 8 */ { \

  2250.         if (cond) { \

  2251.             if (HAVE_FAST_64BIT) { \

  2252.                 for (n = 0; n < end; n += step) \

  2253.                     AV_WN64A(&la[n], la[n] * 0x0101010101010101ULL); \

  2254.             } else { \

  2255.                 for (n = 0; n < end; n += step) { \

  2256.                     uint32_t v32 = la[n] * 0x01010101; \

  2257.                     AV_WN32A(&la[n],     v32); \

  2258.                     AV_WN32A(&la[n + 4], v32); \

  2259.                 } \

  2260.             } \

  2261.         } else { \

  2262.             for (n = 0; n < end; n += step) \

  2263.                 memset(&la[n + 1], la[n], FFMIN(end - n - 1, 7)); \

  2264.         } \

  2265.     }

  2266. #define SPLAT_CTX(step) \

  2267.     do { \

  2268.         SPLAT(a, end_x, step, end_x == w4); \

  2269.         SPLAT(l, end_y, step, end_y == h4); \

  2270.     } while (0)

  2271. 

  2272.     /* y tokens */

  2273.     switch (b->tx) {

  2274.     case TX_4X4:

  2275.         DECODE_Y_COEF_LOOP(1, b->bs > BS_8x8 ? n : 0,);

  2276.         break;

  2277.     case TX_8X8:

  2278.         MERGE_CTX(2, AV_RN16A);

  2279.         DECODE_Y_COEF_LOOP(2, 0,);

  2280.         SPLAT_CTX(2);

  2281.         break;

  2282.     case TX_16X16:

  2283.         MERGE_CTX(4, AV_RN32A);

  2284.         DECODE_Y_COEF_LOOP(4, 0,);

  2285.         SPLAT_CTX(4);

  2286.         break;

  2287.     case TX_32X32:

  2288.         MERGE_CTX(8, AV_RN64A);

  2289.         DECODE_Y_COEF_LOOP(8, 0, 32);

  2290.         SPLAT_CTX(8);

  2291.         break;

  2292.     }

  2293. 

  2294. #define DECODE_UV_COEF_LOOP(step) \

  2295.     for (n = 0, y = 0; y < end_y; y += step) { \

  2296.         for (x = 0; x < end_x; x += step, n += step * step) { \

  2297.             res = decode_coeffs_b(&s->c, s->uvblock[pl] + 16 * n, \

  2298.                                   16 * step * step, c, e, p, a[x] + l[y], \

  2299.                                   uvscan, uvnb, uv_band_counts, qmul[1]); \

  2300.             a[x] = l[y] = !!res; \

  2301.             if (step >= 4) { \

  2302.                 AV_WN16A(&s->uveob[pl][n], res); \

  2303.             } else { \

  2304.                 s->uveob[pl][n] = res; \

  2305.             } \

  2306.         } \

  2307.     }

  2308. 

  2309.     p = s->prob.coef[b->uvtx][1 /* uv */][!b->intra];

  2310.     c = s->counts.coef[b->uvtx][1 /* uv */][!b->intra];

  2311.     e = s->counts.eob[b->uvtx][1 /* uv */][!b->intra];

  2312.     w4 >>= 1;

  2313.     h4 >>= 1;

  2314.     end_x >>= 1;

  2315.     end_y >>= 1;

  2316.     for (pl = 0; pl < 2; pl++) {

  2317.         a = &s->above_uv_nnz_ctx[pl][col];

  2318.         l = &s->left_uv_nnz_ctx[pl][row & 7];

  2319.         switch (b->uvtx) {

  2320.         case TX_4X4:

  2321.             DECODE_UV_COEF_LOOP(1);

  2322.             break;

  2323.         case TX_8X8:

  2324.             MERGE_CTX(2, AV_RN16A);

  2325.             DECODE_UV_COEF_LOOP(2);

  2326.             SPLAT_CTX(2);

  2327.             break;

  2328.         case TX_16X16:

  2329.             MERGE_CTX(4, AV_RN32A);

  2330.             DECODE_UV_COEF_LOOP(4);

  2331.             SPLAT_CTX(4);

  2332.             break;

  2333.         case TX_32X32:

  2334.             MERGE_CTX(8, AV_RN64A);

  2335.             // a 64x64 (max) uv block can ever only contain 1 tx32x32 block

  2336.             // so there is no need to loop

  2337.             res = decode_coeffs_b32(&s->c, s->uvblock[pl],

  2338.                                     1024, c, e, p, a[0] + l[0],

  2339.                                     uvscan, uvnb, uv_band_counts, qmul[1]);

  2340.             a[0] = l[0] = !!res;

  2341.             AV_WN16A(&s->uveob[pl][0], res);

  2342.             SPLAT_CTX(8);

  2343.             break;

  2344.         }

  2345.     }

  2346. }

  2347. 

  2348. static av_always_inline int check_intra_mode(VP9Context *s, int mode, uint8_t **a,

  2349.                                              uint8_t *dst_edge, ptrdiff_t stride_edge,

  2350.                                              uint8_t *dst_inner, ptrdiff_t stride_inner,

  2351.                                              uint8_t *l, int col, int x, int w,

  2352.                                              int row, int y, enum TxfmMode tx,

  2353.                                              int p)

  2354. {

  2355.     int have_top = row > 0 || y > 0;

  2356.     int have_left = col > s->tiling.tile_col_start || x > 0;

  2357.     int have_right = x < w - 1;

  2358.     static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {

  2359.         [VERT_PRED]            = { { DC_127_PRED,          VERT_PRED },

  2360.                                    { DC_127_PRED,          VERT_PRED } },

  2361.         [HOR_PRED]             = { { DC_129_PRED,          DC_129_PRED },

  2362.                                    { HOR_PRED,             HOR_PRED } },

  2363.         [DC_PRED]              = { { DC_128_PRED,          TOP_DC_PRED },

  2364.                                    { LEFT_DC_PRED,         DC_PRED } },

  2365.         [DIAG_DOWN_LEFT_PRED]  = { { DC_127_PRED,          DIAG_DOWN_LEFT_PRED },

  2366.                                    { DC_127_PRED,          DIAG_DOWN_LEFT_PRED } },

  2367.         [DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },

  2368.                                    { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },

  2369.         [VERT_RIGHT_PRED]      = { { VERT_RIGHT_PRED,      VERT_RIGHT_PRED },

  2370.                                    { VERT_RIGHT_PRED,      VERT_RIGHT_PRED } },

  2371.         [HOR_DOWN_PRED]        = { { HOR_DOWN_PRED,        HOR_DOWN_PRED },

  2372.                                    { HOR_DOWN_PRED,        HOR_DOWN_PRED } },

  2373.         [VERT_LEFT_PRED]       = { { DC_127_PRED,          VERT_LEFT_PRED },

  2374.                                    { DC_127_PRED,          VERT_LEFT_PRED } },

  2375.         [HOR_UP_PRED]          = { { DC_129_PRED,          DC_129_PRED },

  2376.                                    { HOR_UP_PRED,          HOR_UP_PRED } },

  2377.         [TM_VP8_PRED]          = { { DC_129_PRED,          VERT_PRED },

  2378.                                    { HOR_PRED,             TM_VP8_PRED } },

  2379.     };

  2380.     static const struct {

  2381.         uint8_t needs_left:1;

  2382.         uint8_t needs_top:1;

  2383.         uint8_t needs_topleft:1;

  2384.         uint8_t needs_topright:1;

  2385.     } edges[N_INTRA_PRED_MODES] = {

  2386.         [VERT_PRED]            = { .needs_top  = 1 },

  2387.         [HOR_PRED]             = { .needs_left = 1 },

  2388.         [DC_PRED]              = { .needs_top  = 1, .needs_left = 1 },

  2389.         [DIAG_DOWN_LEFT_PRED]  = { .needs_top  = 1, .needs_topright = 1 },

  2390.         [DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2391.         [VERT_RIGHT_PRED]      = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2392.         [HOR_DOWN_PRED]        = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2393.         [VERT_LEFT_PRED]       = { .needs_top  = 1, .needs_topright = 1 },

  2394.         [HOR_UP_PRED]          = { .needs_left = 1 },

  2395.         [TM_VP8_PRED]          = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2396.         [LEFT_DC_PRED]         = { .needs_left = 1 },

  2397.         [TOP_DC_PRED]          = { .needs_top  = 1 },

  2398.         [DC_128_PRED]          = { 0 },

  2399.         [DC_127_PRED]          = { 0 },

  2400.         [DC_129_PRED]          = { 0 }

  2401.     };

  2402. 

  2403.     av_assert2(mode >= 0 && mode < 10);

  2404.     mode = mode_conv[mode][have_left][have_top];

  2405.     if (edges[mode].needs_top) {

  2406.         uint8_t *top, *topleft;

  2407.         int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !p) - x) * 4;

  2408.         int n_px_need_tr = 0;

  2409. 

  2410.         if (tx == TX_4X4 && edges[mode].needs_topright && have_right)

  2411.             n_px_need_tr = 4;

  2412. 

  2413.         // if top of sb64-row, use s->intra_pred_data[] instead of

  2414.         // dst[-stride] for intra prediction (it contains pre- instead of

  2415.         // post-loopfilter data)

  2416.         if (have_top) {

  2417.             top = !(row & 7) && !y ?

  2418.                 s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  2419.                 y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];

  2420.             if (have_left)

  2421.                 topleft = !(row & 7) && !y ?

  2422.                     s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  2423.                     y == 0 || x == 0 ? &dst_edge[-stride_edge] :

  2424.                     &dst_inner[-stride_inner];

  2425.         }

  2426. 

  2427.         if (have_top &&

  2428.             (!edges[mode].needs_topleft || (have_left && top == topleft)) &&

  2429.             (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&

  2430.             n_px_need + n_px_need_tr <= n_px_have) {

  2431.             *a = top;

  2432.         } else {

  2433.             if (have_top) {

  2434.                 if (n_px_need <= n_px_have) {

  2435.                     memcpy(*a, top, n_px_need);

  2436.                 } else {

  2437.                     memcpy(*a, top, n_px_have);

  2438.                     memset(&(*a)[n_px_have], (*a)[n_px_have - 1],

  2439.                            n_px_need - n_px_have);

  2440.                 }

  2441.             } else {

  2442.                 memset(*a, 127, n_px_need);

  2443.             }

  2444.             if (edges[mode].needs_topleft) {

  2445.                 if (have_left && have_top) {

  2446.                     (*a)[-1] = topleft[-1];

  2447.                 } else {

  2448.                     (*a)[-1] = have_top ? 129 : 127;

  2449.                 }

  2450.             }

  2451.             if (tx == TX_4X4 && edges[mode].needs_topright) {

  2452.                 if (have_top && have_right &&

  2453.                     n_px_need + n_px_need_tr <= n_px_have) {

  2454.                     memcpy(&(*a)[4], &top[4], 4);

  2455.                 } else {

  2456.                     memset(&(*a)[4], (*a)[3], 4);

  2457.                 }

  2458.             }

  2459.         }

  2460.     }

  2461.     if (edges[mode].needs_left) {

  2462.         if (have_left) {

  2463.             int n_px_need = 4 << tx, i, n_px_have = (((s->rows - row) << !p) - y) * 4;

  2464.             uint8_t *dst = x == 0 ? dst_edge : dst_inner;

  2465.             ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;

  2466. 

  2467.             if (n_px_need <= n_px_have) {

  2468.                 for (i = 0; i < n_px_need; i++)

  2469.                     l[n_px_need - 1 - i] = dst[i * stride - 1];

  2470.             } else {

  2471.                 for (i = 0; i < n_px_have; i++)

  2472.                     l[n_px_need - 1 - i] = dst[i * stride - 1];

  2473.                 memset(l, l[n_px_need - n_px_have], n_px_need - n_px_have);

  2474.             }

  2475.         } else {

  2476.             memset(l, 129, 4 << tx);

  2477.         }

  2478.     }

  2479. 

  2480.     return mode;

  2481. }

  2482. 

  2483. static void intra_recon(AVCodecContext *ctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  2484. {

  2485.     VP9Context *s = ctx->priv_data;

  2486.     VP9Block *b = s->b;

  2487.     int row = s->row, col = s->col;

  2488.     int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2489.     int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2490.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2491.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2492.     int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2493.     int uvstep1d = 1 << b->uvtx, p;

  2494.     uint8_t *dst = s->dst[0], *dst_r = s->frames[CUR_FRAME].tf.f->data[0] + y_off;

  2495.     LOCAL_ALIGNED_32(uint8_t, a_buf, [64]);

  2496.     LOCAL_ALIGNED_32(uint8_t, l, [32]);

  2497. 

  2498.     for (n = 0, y = 0; y < end_y; y += step1d) {

  2499.         uint8_t *ptr = dst, *ptr_r = dst_r;

  2500.         for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d,

  2501.                                ptr_r += 4 * step1d, n += step) {

  2502.             int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?

  2503.                                y * 2 + x : 0];

  2504.             uint8_t *a = &a_buf[32];

  2505.             enum TxfmType txtp = vp9_intra_txfm_type[mode];

  2506.             int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2507. 

  2508.             mode = check_intra_mode(s, mode, &a, ptr_r,

  2509.                                     s->frames[CUR_FRAME].tf.f->linesize[0],

  2510.                                     ptr, s->y_stride, l,

  2511.                                     col, x, w4, row, y, b->tx, 0);

  2512.             s->dsp.intra_pred[b->tx][mode](ptr, s->y_stride, l, a);

  2513.             if (eob)

  2514.                 s->dsp.itxfm_add[tx][txtp](ptr, s->y_stride,

  2515.                                            s->block + 16 * n, eob);

  2516.         }

  2517.         dst_r += 4 * step1d * s->frames[CUR_FRAME].tf.f->linesize[0];

  2518.         dst   += 4 * step1d * s->y_stride;

  2519.     }

  2520. 

  2521.     // U/V

  2522.     w4 >>= 1;

  2523.     end_x >>= 1;

  2524.     end_y >>= 1;

  2525.     step = 1 << (b->uvtx * 2);

  2526.     for (p = 0; p < 2; p++) {

  2527.         dst   = s->dst[1 + p];

  2528.         dst_r = s->frames[CUR_FRAME].tf.f->data[1 + p] + uv_off;

  2529.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2530.             uint8_t *ptr = dst, *ptr_r = dst_r;

  2531.             for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d,

  2532.                                    ptr_r += 4 * uvstep1d, n += step) {

  2533.                 int mode = b->uvmode;

  2534.                 uint8_t *a = &a_buf[32];

  2535.                 int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2536. 

  2537.                 mode = check_intra_mode(s, mode, &a, ptr_r,

  2538.                                         s->frames[CUR_FRAME].tf.f->linesize[1],

  2539.                                         ptr, s->uv_stride, l,

  2540.                                         col, x, w4, row, y, b->uvtx, p + 1);

  2541.                 s->dsp.intra_pred[b->uvtx][mode](ptr, s->uv_stride, l, a);

  2542.                 if (eob)

  2543.                     s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2544.                                                     s->uvblock[p] + 16 * n, eob);

  2545.             }

  2546.             dst_r += 4 * uvstep1d * s->frames[CUR_FRAME].tf.f->linesize[1];

  2547.             dst   += 4 * uvstep1d * s->uv_stride;

  2548.         }

  2549.     }

  2550. }

  2551. 

  2552. static av_always_inline void mc_luma_scaled(VP9Context *s, vp9_scaled_mc_func smc,

  2553.                                             uint8_t *dst, ptrdiff_t dst_stride,

  2554.                                             const uint8_t *ref, ptrdiff_t ref_stride,

  2555.                                             ThreadFrame *ref_frame,

  2556.                                             ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2557.                                             int bw, int bh, int w, int h,

  2558.                                             const uint16_t *scale, const uint8_t *step)

  2559. {

  2560. #define scale_mv(n, dim) (((int64_t)n * scale[dim]) >> 14)

  2561.     // BUG libvpx seems to scale the two components separately. This introduces

  2562.     // rounding errors but we have to reproduce them to be exactly compatible

  2563.     // with the output from libvpx...

  2564.     int mx = scale_mv(mv->x * 2, 0) + scale_mv(x * 16, 0);

  2565.     int my = scale_mv(mv->y * 2, 1) + scale_mv(y * 16, 1);

  2566.     int refbw_m1, refbh_m1;

  2567.     int th;

  2568. 

  2569.     y = my >> 4;

  2570.     x = mx >> 4;

  2571.     ref += y * ref_stride + x;

  2572.     mx &= 15;

  2573.     my &= 15;

  2574.     refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;

  2575.     refbh_m1 = ((bh - 1) * step[1] + my) >> 4;

  2576.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2577.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2578.     // the longest loopfilter of the next sbrow

  2579.     th = (y + refbh_m1 + 4 + 7) >> 6;

  2580.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2581.     if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 4 >= h - refbh_m1) {

  2582.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2583.                                  ref - 3 * ref_stride - 3,

  2584.                                  144, ref_stride,

  2585.                                  refbw_m1 + 8, refbh_m1 + 8,

  2586.                                  x - 3, y - 3, w, h);

  2587.         ref = s->edge_emu_buffer + 3 * 144 + 3;

  2588.         ref_stride = 144;

  2589.     }

  2590.     smc(dst, dst_stride, ref, ref_stride, bh, mx, my, step[0], step[1]);

  2591. }

  2592. 

  2593. static av_always_inline void mc_chroma_scaled(VP9Context *s, vp9_scaled_mc_func smc,

  2594.                                               uint8_t *dst_u, uint8_t *dst_v,

  2595.                                               ptrdiff_t dst_stride,

  2596.                                               const uint8_t *ref_u, ptrdiff_t src_stride_u,

  2597.                                               const uint8_t *ref_v, ptrdiff_t src_stride_v,

  2598.                                               ThreadFrame *ref_frame,

  2599.                                               ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2600.                                               int bw, int bh, int w, int h,

  2601.                                               const uint16_t *scale, const uint8_t *step)

  2602. {

  2603.     // BUG https://code.google.com/p/webm/issues/detail?id=820

  2604.     int mx = scale_mv(mv->x, 0) + (scale_mv(x * 16, 0) & ~15) + (scale_mv(x * 32, 0) & 15);

  2605.     int my = scale_mv(mv->y, 1) + (scale_mv(y * 16, 1) & ~15) + (scale_mv(y * 32, 1) & 15);

  2606. #undef scale_mv

  2607.     int refbw_m1, refbh_m1;

  2608.     int th;

  2609. 

  2610.     y = my >> 4;

  2611.     x = mx >> 4;

  2612.     ref_u += y * src_stride_u + x;

  2613.     ref_v += y * src_stride_v + x;

  2614.     mx &= 15;

  2615.     my &= 15;

  2616.     refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;

  2617.     refbh_m1 = ((bh - 1) * step[1] + my) >> 4;

  2618.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2619.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2620.     // the longest loopfilter of the next sbrow

  2621.     th = (y + refbh_m1 + 4 + 7) >> 5;

  2622.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2623.     if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 4 >= h - refbh_m1) {

  2624.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2625.                                  ref_u - 3 * src_stride_u - 3,

  2626.                                  144, src_stride_u,

  2627.                                  refbw_m1 + 8, refbh_m1 + 8,

  2628.                                  x - 3, y - 3, w, h);

  2629.         ref_u = s->edge_emu_buffer + 3 * 144 + 3;

  2630.         smc(dst_u, dst_stride, ref_u, 144, bh, mx, my, step[0], step[1]);

  2631. 

  2632.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2633.                                  ref_v - 3 * src_stride_v - 3,

  2634.                                  144, src_stride_v,

  2635.                                  refbw_m1 + 8, refbh_m1 + 8,

  2636.                                  x - 3, y - 3, w, h);

  2637.         ref_v = s->edge_emu_buffer + 3 * 144 + 3;

  2638.         smc(dst_v, dst_stride, ref_v, 144, bh, mx, my, step[0], step[1]);

  2639.     } else {

  2640.         smc(dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my, step[0], step[1]);

  2641.         smc(dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my, step[0], step[1]);

  2642.     }

  2643. }

  2644. 

  2645. #define FN(x) x##_scaled

  2646. #define mc_luma_dir(s, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, bw, bh, w, h, i) \

  2647.     mc_luma_scaled(s, s->dsp.s##mc, dst, dst_ls, src, src_ls, tref, row, col, \

  2648.                    mv, bw, bh, w, h, s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])

  2649. #define mc_chroma_dir(s, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2650.                       row, col, mv, bw, bh, w, h, i) \

  2651.     mc_chroma_scaled(s, s->dsp.s##mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2652.                      row, col, mv, bw, bh, w, h, s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])

  2653. #include "vp9_mc_template.c"

  2654. #undef mc_luma_dir

  2655. #undef mc_chroma_dir

  2656. #undef FN

  2657. 

  2658. static av_always_inline void mc_luma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2],

  2659.                                               uint8_t *dst, ptrdiff_t dst_stride,

  2660.                                               const uint8_t *ref, ptrdiff_t ref_stride,

  2661.                                               ThreadFrame *ref_frame,

  2662.                                               ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2663.                                               int bw, int bh, int w, int h)

  2664. {

  2665.     int mx = mv->x, my = mv->y, th;

  2666. 

  2667.     y += my >> 3;

  2668.     x += mx >> 3;

  2669.     ref += y * ref_stride + x;

  2670.     mx &= 7;

  2671.     my &= 7;

  2672.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2673.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2674.     // the longest loopfilter of the next sbrow

  2675.     th = (y + bh + 4 * !!my + 7) >> 6;

  2676.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2677.     if (x < !!mx * 3 || y < !!my * 3 ||

  2678.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2679.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2680.                                  ref - !!my * 3 * ref_stride - !!mx * 3,

  2681.                                  80, ref_stride,

  2682.                                  bw + !!mx * 7, bh + !!my * 7,

  2683.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2684.         ref = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2685.         ref_stride = 80;

  2686.     }

  2687.     mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);

  2688. }

  2689. 

  2690. static av_always_inline void mc_chroma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2],

  2691.                                                 uint8_t *dst_u, uint8_t *dst_v,

  2692.                                                 ptrdiff_t dst_stride,

  2693.                                                 const uint8_t *ref_u, ptrdiff_t src_stride_u,

  2694.                                                 const uint8_t *ref_v, ptrdiff_t src_stride_v,

  2695.                                                 ThreadFrame *ref_frame,

  2696.                                                 ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2697.                                                 int bw, int bh, int w, int h)

  2698. {

  2699.     int mx = mv->x, my = mv->y, th;

  2700. 

  2701.     y += my >> 4;

  2702.     x += mx >> 4;

  2703.     ref_u += y * src_stride_u + x;

  2704.     ref_v += y * src_stride_v + x;

  2705.     mx &= 15;

  2706.     my &= 15;

  2707.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2708.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2709.     // the longest loopfilter of the next sbrow

  2710.     th = (y + bh + 4 * !!my + 7) >> 5;

  2711.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2712.     if (x < !!mx * 3 || y < !!my * 3 ||

  2713.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2714.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2715.                                  ref_u - !!my * 3 * src_stride_u - !!mx * 3,

  2716.                                  80, src_stride_u,

  2717.                                  bw + !!mx * 7, bh + !!my * 7,

  2718.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2719.         ref_u = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2720.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, 80, bh, mx, my);

  2721. 

  2722.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2723.                                  ref_v - !!my * 3 * src_stride_v - !!mx * 3,

  2724.                                  80, src_stride_v,

  2725.                                  bw + !!mx * 7, bh + !!my * 7,

  2726.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2727.         ref_v = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2728.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, 80, bh, mx, my);

  2729.     } else {

  2730.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);

  2731.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);

  2732.     }

  2733. }

  2734. 

  2735. #define FN(x) x

  2736. #define mc_luma_dir(s, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, bw, bh, w, h, i) \

  2737.     mc_luma_unscaled(s, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \

  2738.                      mv, bw, bh, w, h)

  2739. #define mc_chroma_dir(s, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2740.                       row, col, mv, bw, bh, w, h, i) \

  2741.     mc_chroma_unscaled(s, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2742.                        row, col, mv, bw, bh, w, h)

  2743. #include "vp9_mc_template.c"

  2744. #undef mc_luma_dir_dir

  2745. #undef mc_chroma_dir_dir

  2746. #undef FN

  2747. 

  2748. static void inter_recon(AVCodecContext *ctx)

  2749. {

  2750.     VP9Context *s = ctx->priv_data;

  2751.     VP9Block *b = s->b;

  2752.     int row = s->row, col = s->col;

  2753. 

  2754.     if (s->mvscale[b->ref[0]][0] || (b->comp && s->mvscale[b->ref[1]][0])) {

  2755.         inter_pred_scaled(ctx);

  2756.     } else {

  2757.         inter_pred(ctx);

  2758.     }

  2759.     if (!b->skip) {

  2760.         /* mostly copied intra_recon() */

  2761. 

  2762.         int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2763.         int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2764.         int end_x = FFMIN(2 * (s->cols - col), w4);

  2765.         int end_y = FFMIN(2 * (s->rows - row), h4);

  2766.         int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2767.         int uvstep1d = 1 << b->uvtx, p;

  2768.         uint8_t *dst = s->dst[0];

  2769. 

  2770.         // y itxfm add

  2771.         for (n = 0, y = 0; y < end_y; y += step1d) {

  2772.             uint8_t *ptr = dst;

  2773.             for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d, n += step) {

  2774.                 int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2775. 

  2776.                 if (eob)

  2777.                     s->dsp.itxfm_add[tx][DCT_DCT](ptr, s->y_stride,

  2778.                                                   s->block + 16 * n, eob);

  2779.             }

  2780.             dst += 4 * s->y_stride * step1d;

  2781.         }

  2782. 

  2783.         // uv itxfm add

  2784.         end_x >>= 1;

  2785.         end_y >>= 1;

  2786.         step = 1 << (b->uvtx * 2);

  2787.         for (p = 0; p < 2; p++) {

  2788.             dst = s->dst[p + 1];

  2789.             for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2790.                 uint8_t *ptr = dst;

  2791.                 for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d, n += step) {

  2792.                     int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2793. 

  2794.                     if (eob)

  2795.                         s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2796.                                                         s->uvblock[p] + 16 * n, eob);

  2797.                 }

  2798.                 dst += 4 * uvstep1d * s->uv_stride;

  2799.             }

  2800.         }

  2801.     }

  2802. }

  2803. 

  2804. static av_always_inline void mask_edges(struct VP9Filter *lflvl, int is_uv,

  2805.                                         int row_and_7, int col_and_7,

  2806.                                         int w, int h, int col_end, int row_end,

  2807.                                         enum TxfmMode tx, int skip_inter)

  2808. {

  2809.     // FIXME I'm pretty sure all loops can be replaced by a single LUT if

  2810.     // we make VP9Filter.mask uint64_t (i.e. row/col all single variable)

  2811.     // and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then

  2812.     // use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)

  2813. 

  2814.     // the intended behaviour of the vp9 loopfilter is to work on 8-pixel

  2815.     // edges. This means that for UV, we work on two subsampled blocks at

  2816.     // a time, and we only use the topleft block's mode information to set

  2817.     // things like block strength. Thus, for any block size smaller than

  2818.     // 16x16, ignore the odd portion of the block.

  2819.     if (tx == TX_4X4 && is_uv) {

  2820.         if (h == 1) {

  2821.             if (row_and_7 & 1)

  2822.                 return;

  2823.             if (!row_end)

  2824.                 h += 1;

  2825.         }

  2826.         if (w == 1) {

  2827.             if (col_and_7 & 1)

  2828.                 return;

  2829.             if (!col_end)

  2830.                 w += 1;

  2831.         }

  2832.     }

  2833. 

  2834.     if (tx == TX_4X4 && !skip_inter) {

  2835.         int t = 1 << col_and_7, m_col = (t << w) - t, y;

  2836.         int m_col_odd = (t << (w - 1)) - t;

  2837. 

  2838.         // on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide

  2839.         if (is_uv) {

  2840.             int m_row_8 = m_col & 0x01, m_row_4 = m_col - m_row_8;

  2841. 

  2842.             for (y = row_and_7; y < h + row_and_7; y++) {

  2843.                 int col_mask_id = 2 - !(y & 7);

  2844. 

  2845.                 lflvl->mask[is_uv][0][y][1] |= m_row_8;

  2846.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  2847.                 // for odd lines, if the odd col is not being filtered,

  2848.                 // skip odd row also:

  2849.                 // .---. <-- a

  2850.                 // |   |

  2851.                 // |___| <-- b

  2852.                 // ^   ^

  2853.                 // c   d

  2854.                 //

  2855.                 // if a/c are even row/col and b/d are odd, and d is skipped,

  2856.                 // e.g. right edge of size-66x66.webm, then skip b also (bug)

  2857.                 if ((col_end & 1) && (y & 1)) {

  2858.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col_odd;

  2859.                 } else {

  2860.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col;

  2861.                 }

  2862.             }

  2863.         } else {

  2864.             int m_row_8 = m_col & 0x11, m_row_4 = m_col - m_row_8;

  2865. 

  2866.             for (y = row_and_7; y < h + row_and_7; y++) {

  2867.                 int col_mask_id = 2 - !(y & 3);

  2868. 

  2869.                 lflvl->mask[is_uv][0][y][1] |= m_row_8; // row edge

  2870.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  2871.                 lflvl->mask[is_uv][1][y][col_mask_id] |= m_col; // col edge

  2872.                 lflvl->mask[is_uv][0][y][3] |= m_col;

  2873.                 lflvl->mask[is_uv][1][y][3] |= m_col;

  2874.             }

  2875.         }

  2876.     } else {

  2877.         int y, t = 1 << col_and_7, m_col = (t << w) - t;

  2878. 

  2879.         if (!skip_inter) {

  2880.             int mask_id = (tx == TX_8X8);

  2881.             int l2 = tx + is_uv - 1, step1d = 1 << l2;

  2882.             static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };

  2883.             int m_row = m_col & masks[l2];

  2884. 

  2885.             // at odd UV col/row edges tx16/tx32 loopfilter edges, force

  2886.             // 8wd loopfilter to prevent going off the visible edge.

  2887.             if (is_uv && tx > TX_8X8 && (w ^ (w - 1)) == 1) {

  2888.                 int m_row_16 = ((t << (w - 1)) - t) & masks[l2];

  2889.                 int m_row_8 = m_row - m_row_16;

  2890. 

  2891.                 for (y = row_and_7; y < h + row_and_7; y++) {

  2892.                     lflvl->mask[is_uv][0][y][0] |= m_row_16;

  2893.                     lflvl->mask[is_uv][0][y][1] |= m_row_8;

  2894.                 }

  2895.             } else {

  2896.                 for (y = row_and_7; y < h + row_and_7; y++)

  2897.                     lflvl->mask[is_uv][0][y][mask_id] |= m_row;

  2898.             }

  2899. 

  2900.             if (is_uv && tx > TX_8X8 && (h ^ (h - 1)) == 1) {

  2901.                 for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)

  2902.                     lflvl->mask[is_uv][1][y][0] |= m_col;

  2903.                 if (y - row_and_7 == h - 1)

  2904.                     lflvl->mask[is_uv][1][y][1] |= m_col;

  2905.             } else {

  2906.                 for (y = row_and_7; y < h + row_and_7; y += step1d)

  2907.                     lflvl->mask[is_uv][1][y][mask_id] |= m_col;

  2908.             }

  2909.         } else if (tx != TX_4X4) {

  2910.             int mask_id;

  2911. 

  2912.             mask_id = (tx == TX_8X8) || (is_uv && h == 1);

  2913.             lflvl->mask[is_uv][1][row_and_7][mask_id] |= m_col;

  2914.             mask_id = (tx == TX_8X8) || (is_uv && w == 1);

  2915.             for (y = row_and_7; y < h + row_and_7; y++)

  2916.                 lflvl->mask[is_uv][0][y][mask_id] |= t;

  2917.         } else if (is_uv) {

  2918.             int t8 = t & 0x01, t4 = t - t8;

  2919. 

  2920.             for (y = row_and_7; y < h + row_and_7; y++) {

  2921.                 lflvl->mask[is_uv][0][y][2] |= t4;

  2922.                 lflvl->mask[is_uv][0][y][1] |= t8;

  2923.             }

  2924.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 7)] |= m_col;

  2925.         } else {

  2926.             int t8 = t & 0x11, t4 = t - t8;

  2927. 

  2928.             for (y = row_and_7; y < h + row_and_7; y++) {

  2929.                 lflvl->mask[is_uv][0][y][2] |= t4;

  2930.                 lflvl->mask[is_uv][0][y][1] |= t8;

  2931.             }

  2932.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 3)] |= m_col;

  2933.         }

  2934.     }

  2935. }

  2936. 

  2937. static void decode_b(AVCodecContext *ctx, int row, int col,

  2938.                      struct VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,

  2939.                      enum BlockLevel bl, enum BlockPartition bp)

  2940. {

  2941.     VP9Context *s = ctx->priv_data;

  2942.     VP9Block *b = s->b;

  2943.     enum BlockSize bs = bl * 3 + bp;

  2944.     int w4 = bwh_tab[1][bs][0], h4 = bwh_tab[1][bs][1], lvl;

  2945.     int emu[2];

  2946.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2947. 

  2948.     s->row = row;

  2949.     s->row7 = row & 7;

  2950.     s->col = col;

  2951.     s->col7 = col & 7;

  2952.     s->min_mv.x = -(128 + col * 64);

  2953.     s->min_mv.y = -(128 + row * 64);

  2954.     s->max_mv.x = 128 + (s->cols - col - w4) * 64;

  2955.     s->max_mv.y = 128 + (s->rows - row - h4) * 64;

  2956.     if (s->pass < 2) {

  2957.         b->bs = bs;

  2958.         b->bl = bl;

  2959.         b->bp = bp;

  2960.         decode_mode(ctx);

  2961.         b->uvtx = b->tx - (w4 * 2 == (1 << b->tx) || h4 * 2 == (1 << b->tx));

  2962. 

  2963.         if (!b->skip) {

  2964.             decode_coeffs(ctx);

  2965.         } else {

  2966.             int row7 = s->row7;

  2967. 

  2968. #define SPLAT_ZERO_CTX(v, n) \

  2969.     switch (n) { \

  2970.     case 1:  v = 0;          break; \

  2971.     case 2:  AV_ZERO16(&v);  break; \

  2972.     case 4:  AV_ZERO32(&v);  break; \

  2973.     case 8:  AV_ZERO64(&v);  break; \

  2974.     case 16: AV_ZERO128(&v); break; \

  2975.     }

  2976. #define SPLAT_ZERO_YUV(dir, var, off, n) \

  2977.     do { \

  2978.         SPLAT_ZERO_CTX(s->dir##_y_##var[off * 2], n * 2); \

  2979.         SPLAT_ZERO_CTX(s->dir##_uv_##var[0][off], n); \

  2980.         SPLAT_ZERO_CTX(s->dir##_uv_##var[1][off], n); \

  2981.     } while (0)

  2982. 

  2983.             switch (w4) {

  2984.             case 1: SPLAT_ZERO_YUV(above, nnz_ctx, col, 1); break;

  2985.             case 2: SPLAT_ZERO_YUV(above, nnz_ctx, col, 2); break;

  2986.             case 4: SPLAT_ZERO_YUV(above, nnz_ctx, col, 4); break;

  2987.             case 8: SPLAT_ZERO_YUV(above, nnz_ctx, col, 8); break;

  2988.             }

  2989.             switch (h4) {

  2990.             case 1: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 1); break;

  2991.             case 2: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 2); break;

  2992.             case 4: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 4); break;

  2993.             case 8: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 8); break;

  2994.             }

  2995.         }

  2996.         if (s->pass == 1) {

  2997.             s->b++;

  2998.             s->block += w4 * h4 * 64;

  2999.             s->uvblock[0] += w4 * h4 * 16;

  3000.             s->uvblock[1] += w4 * h4 * 16;

  3001.             s->eob += 4 * w4 * h4;

  3002.             s->uveob[0] += w4 * h4;

  3003.             s->uveob[1] += w4 * h4;

  3004. 

  3005.             return;

  3006.         }

  3007.     }

  3008. 

  3009.     // emulated overhangs if the stride of the target buffer can't hold. This

  3010.     // allows to support emu-edge and so on even if we have large block

  3011.     // overhangs

  3012.     emu[0] = (col + w4) * 8 > f->linesize[0] ||

  3013.              (row + h4) > s->rows;

  3014.     emu[1] = (col + w4) * 4 > f->linesize[1] ||

  3015.              (row + h4) > s->rows;

  3016.     if (emu[0]) {

  3017.         s->dst[0] = s->tmp_y;

  3018.         s->y_stride = 64;

  3019.     } else {

  3020.         s->dst[0] = f->data[0] + yoff;

  3021.         s->y_stride = f->linesize[0];

  3022.     }

  3023.     if (emu[1]) {

  3024.         s->dst[1] = s->tmp_uv[0];

  3025.         s->dst[2] = s->tmp_uv[1];

  3026.         s->uv_stride = 32;

  3027.     } else {

  3028.         s->dst[1] = f->data[1] + uvoff;

  3029.         s->dst[2] = f->data[2] + uvoff;

  3030.         s->uv_stride = f->linesize[1];

  3031.     }

  3032.     if (b->intra) {

  3033.         intra_recon(ctx, yoff, uvoff);

  3034.     } else {

  3035.         inter_recon(ctx);

  3036.     }

  3037.     if (emu[0]) {

  3038.         int w = FFMIN(s->cols - col, w4) * 8, h = FFMIN(s->rows - row, h4) * 8, n, o = 0;

  3039. 

  3040.         for (n = 0; o < w; n++) {

  3041.             int bw = 64 >> n;

  3042. 

  3043.             av_assert2(n <= 4);

  3044.             if (w & bw) {

  3045.                 s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o, f->linesize[0],

  3046.                                          s->tmp_y + o, 64, h, 0, 0);

  3047.                 o += bw;

  3048.             }

  3049.         }

  3050.     }

  3051.     if (emu[1]) {

  3052.         int w = FFMIN(s->cols - col, w4) * 4, h = FFMIN(s->rows - row, h4) * 4, n, o = 0;

  3053. 

  3054.         for (n = 1; o < w; n++) {

  3055.             int bw = 64 >> n;

  3056. 

  3057.             av_assert2(n <= 4);

  3058.             if (w & bw) {

  3059.                 s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o, f->linesize[1],

  3060.                                          s->tmp_uv[0] + o, 32, h, 0, 0);

  3061.                 s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o, f->linesize[2],

  3062.                                          s->tmp_uv[1] + o, 32, h, 0, 0);

  3063.                 o += bw;

  3064.             }

  3065.         }

  3066.     }

  3067. 

  3068.     // pick filter level and find edges to apply filter to

  3069.     if (s->filter.level &&

  3070.         (lvl = s->segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]

  3071.                                                     [b->mode[3] != ZEROMV]) > 0) {

  3072.         int x_end = FFMIN(s->cols - col, w4), y_end = FFMIN(s->rows - row, h4);

  3073.         int skip_inter = !b->intra && b->skip, col7 = s->col7, row7 = s->row7;

  3074. 

  3075.         setctx_2d(&lflvl->level[row7 * 8 + col7], w4, h4, 8, lvl);

  3076.         mask_edges(lflvl, 0, row7, col7, x_end, y_end, 0, 0, b->tx, skip_inter);

  3077.         mask_edges(lflvl, 1, row7, col7, x_end, y_end,

  3078.                    s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,

  3079.                    s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,

  3080.                    b->uvtx, skip_inter);

  3081. 

  3082.         if (!s->filter.lim_lut[lvl]) {

  3083.             int sharp = s->filter.sharpness;

  3084.             int limit = lvl;

  3085. 

  3086.             if (sharp > 0) {

  3087.                 limit >>= (sharp + 3) >> 2;

  3088.                 limit = FFMIN(limit, 9 - sharp);

  3089.             }

  3090.             limit = FFMAX(limit, 1);

  3091. 

  3092.             s->filter.lim_lut[lvl] = limit;

  3093.             s->filter.mblim_lut[lvl] = 2 * (lvl + 2) + limit;

  3094.         }

  3095.     }

  3096. 

  3097.     if (s->pass == 2) {

  3098.         s->b++;

  3099.         s->block += w4 * h4 * 64;

  3100.         s->uvblock[0] += w4 * h4 * 16;

  3101.         s->uvblock[1] += w4 * h4 * 16;

  3102.         s->eob += 4 * w4 * h4;

  3103.         s->uveob[0] += w4 * h4;

  3104.         s->uveob[1] += w4 * h4;

  3105.     }

  3106. }

  3107. 

  3108. static void decode_sb(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  3109.                       ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  3110. {

  3111.     VP9Context *s = ctx->priv_data;

  3112.     int c = ((s->above_partition_ctx[col] >> (3 - bl)) & 1) |

  3113.             (((s->left_partition_ctx[row & 0x7] >> (3 - bl)) & 1) << 1);

  3114.     const uint8_t *p = s->keyframe ? vp9_default_kf_partition_probs[bl][c] :

  3115.                                      s->prob.p.partition[bl][c];

  3116.     enum BlockPartition bp;

  3117.     ptrdiff_t hbs = 4 >> bl;

  3118.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  3119.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  3120. 

  3121.     if (bl == BL_8X8) {

  3122.         bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  3123.         decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3124.     } else if (col + hbs < s->cols) { // FIXME why not <=?

  3125.         if (row + hbs < s->rows) { // FIXME why not <=?

  3126.             bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  3127.             switch (bp) {

  3128.             case PARTITION_NONE:

  3129.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3130.                 break;

  3131.             case PARTITION_H:

  3132.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3133.                 yoff  += hbs * 8 * y_stride;

  3134.                 uvoff += hbs * 4 * uv_stride;

  3135.                 decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, bl, bp);

  3136.                 break;

  3137.             case PARTITION_V:

  3138.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3139.                 yoff  += hbs * 8;

  3140.                 uvoff += hbs * 4;

  3141.                 decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, bl, bp);

  3142.                 break;

  3143.             case PARTITION_SPLIT:

  3144.                 decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3145.                 decode_sb(ctx, row, col + hbs, lflvl,

  3146.                           yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  3147.                 yoff  += hbs * 8 * y_stride;

  3148.                 uvoff += hbs * 4 * uv_stride;

  3149.                 decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3150.                 decode_sb(ctx, row + hbs, col + hbs, lflvl,

  3151.                           yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  3152.                 break;

  3153.             default:

  3154.                 av_assert0(0);

  3155.             }

  3156.         } else if (vp56_rac_get_prob_branchy(&s->c, p[1])) {

  3157.             bp = PARTITION_SPLIT;

  3158.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3159.             decode_sb(ctx, row, col + hbs, lflvl,

  3160.                       yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  3161.         } else {

  3162.             bp = PARTITION_H;

  3163.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3164.         }

  3165.     } else if (row + hbs < s->rows) { // FIXME why not <=?

  3166.         if (vp56_rac_get_prob_branchy(&s->c, p[2])) {

  3167.             bp = PARTITION_SPLIT;

  3168.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3169.             yoff  += hbs * 8 * y_stride;

  3170.             uvoff += hbs * 4 * uv_stride;

  3171.             decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3172.         } else {

  3173.             bp = PARTITION_V;

  3174.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3175.         }

  3176.     } else {

  3177.         bp = PARTITION_SPLIT;

  3178.         decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3179.     }

  3180.     s->counts.partition[bl][c][bp]++;

  3181. }

  3182. 

  3183. static void decode_sb_mem(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  3184.                           ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  3185. {

  3186.     VP9Context *s = ctx->priv_data;

  3187.     VP9Block *b = s->b;

  3188.     ptrdiff_t hbs = 4 >> bl;

  3189.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  3190.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  3191. 

  3192.     if (bl == BL_8X8) {

  3193.         av_assert2(b->bl == BL_8X8);

  3194.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  3195.     } else if (s->b->bl == bl) {

  3196.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  3197.         if (b->bp == PARTITION_H && row + hbs < s->rows) {

  3198.             yoff  += hbs * 8 * y_stride;

  3199.             uvoff += hbs * 4 * uv_stride;

  3200.             decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, b->bl, b->bp);

  3201.         } else if (b->bp == PARTITION_V && col + hbs < s->cols) {

  3202.             yoff  += hbs * 8;

  3203.             uvoff += hbs * 4;

  3204.             decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, b->bl, b->bp);

  3205.         }

  3206.     } else {

  3207.         decode_sb_mem(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3208.         if (col + hbs < s->cols) { // FIXME why not <=?

  3209.             if (row + hbs < s->rows) {

  3210.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff + 8 * hbs,

  3211.                               uvoff + 4 * hbs, bl + 1);

  3212.                 yoff  += hbs * 8 * y_stride;

  3213.                 uvoff += hbs * 4 * uv_stride;

  3214.                 decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3215.                 decode_sb_mem(ctx, row + hbs, col + hbs, lflvl,

  3216.                                     yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  3217.             } else {

  3218.                 yoff  += hbs * 8;

  3219.                 uvoff += hbs * 4;

  3220.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff, uvoff, bl + 1);

  3221.             }

  3222.         } else if (row + hbs < s->rows) {

  3223.             yoff  += hbs * 8 * y_stride;

  3224.             uvoff += hbs * 4 * uv_stride;

  3225.             decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3226.         }

  3227.     }

  3228. }

  3229. 

  3230. static void loopfilter_sb(AVCodecContext *ctx, struct VP9Filter *lflvl,

  3231.                           int row, int col, ptrdiff_t yoff, ptrdiff_t uvoff)

  3232. {

  3233.     VP9Context *s = ctx->priv_data;

  3234.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  3235.     uint8_t *dst = f->data[0] + yoff, *lvl = lflvl->level;

  3236.     ptrdiff_t ls_y = f->linesize[0], ls_uv = f->linesize[1];

  3237.     int y, x, p;

  3238. 

  3239.     // FIXME in how far can we interleave the v/h loopfilter calls? E.g.

  3240.     // if you think of them as acting on a 8x8 block max, we can interleave

  3241.     // each v/h within the single x loop, but that only works if we work on

  3242.     // 8 pixel blocks, and we won't always do that (we want at least 16px

  3243.     // to use SSE2 optimizations, perhaps 32 for AVX2)

  3244. 

  3245.     // filter edges between columns, Y plane (e.g. block1 | block2)

  3246.     for (y = 0; y < 8; y += 2, dst += 16 * ls_y, lvl += 16) {

  3247.         uint8_t *ptr = dst, *l = lvl, *hmask1 = lflvl->mask[0][0][y];

  3248.         uint8_t *hmask2 = lflvl->mask[0][0][y + 1];

  3249.         unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2], hm13 = hmask1[3];

  3250.         unsigned hm2 = hmask2[1] | hmask2[2], hm23 = hmask2[3];

  3251.         unsigned hm = hm1 | hm2 | hm13 | hm23;

  3252. 

  3253.         for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 8, l++) {

  3254.             if (hm1 & x) {

  3255.                 int L = *l, H = L >> 4;

  3256.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3257. 

  3258.                 if (col || x > 1) {

  3259.                     if (hmask1[0] & x) {

  3260.                         if (hmask2[0] & x) {

  3261.                             av_assert2(l[8] == L);

  3262.                             s->dsp.loop_filter_16[0](ptr, ls_y, E, I, H);

  3263.                         } else {

  3264.                             s->dsp.loop_filter_8[2][0](ptr, ls_y, E, I, H);

  3265.                         }

  3266.                     } else if (hm2 & x) {

  3267.                         L = l[8];

  3268.                         H |= (L >> 4) << 8;

  3269.                         E |= s->filter.mblim_lut[L] << 8;

  3270.                         I |= s->filter.lim_lut[L] << 8;

  3271.                         s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3272.                                                [!!(hmask2[1] & x)]

  3273.                                                [0](ptr, ls_y, E, I, H);

  3274.                     } else {

  3275.                         s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3276.                                             [0](ptr, ls_y, E, I, H);

  3277.                     }

  3278.                 }

  3279.             } else if (hm2 & x) {

  3280.                 int L = l[8], H = L >> 4;

  3281.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3282. 

  3283.                 if (col || x > 1) {

  3284.                     s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3285.                                         [0](ptr + 8 * ls_y, ls_y, E, I, H);

  3286.                 }

  3287.             }

  3288.             if (hm13 & x) {

  3289.                 int L = *l, H = L >> 4;

  3290.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3291. 

  3292.                 if (hm23 & x) {

  3293.                     L = l[8];

  3294.                     H |= (L >> 4) << 8;

  3295.                     E |= s->filter.mblim_lut[L] << 8;

  3296.                     I |= s->filter.lim_lut[L] << 8;

  3297.                     s->dsp.loop_filter_mix2[0][0][0](ptr + 4, ls_y, E, I, H);

  3298.                 } else {

  3299.                     s->dsp.loop_filter_8[0][0](ptr + 4, ls_y, E, I, H);

  3300.                 }

  3301.             } else if (hm23 & x) {

  3302.                 int L = l[8], H = L >> 4;

  3303.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3304. 

  3305.                 s->dsp.loop_filter_8[0][0](ptr + 8 * ls_y + 4, ls_y, E, I, H);

  3306.             }

  3307.         }

  3308.     }

  3309. 

  3310.     //                                          block1

  3311.     // filter edges between rows, Y plane (e.g. ------)

  3312.     //                                          block2

  3313.     dst = f->data[0] + yoff;

  3314.     lvl = lflvl->level;

  3315.     for (y = 0; y < 8; y++, dst += 8 * ls_y, lvl += 8) {

  3316.         uint8_t *ptr = dst, *l = lvl, *vmask = lflvl->mask[0][1][y];

  3317.         unsigned vm = vmask[0] | vmask[1] | vmask[2], vm3 = vmask[3];

  3318. 

  3319.         for (x = 1; vm & ~(x - 1); x <<= 2, ptr += 16, l += 2) {

  3320.             if (row || y) {

  3321.                 if (vm & x) {

  3322.                     int L = *l, H = L >> 4;

  3323.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3324. 

  3325.                     if (vmask[0] & x) {

  3326.                         if (vmask[0] & (x << 1)) {

  3327.                             av_assert2(l[1] == L);

  3328.                             s->dsp.loop_filter_16[1](ptr, ls_y, E, I, H);

  3329.                         } else {

  3330.                             s->dsp.loop_filter_8[2][1](ptr, ls_y, E, I, H);

  3331.                         }

  3332.                     } else if (vm & (x << 1)) {

  3333.                         L = l[1];

  3334.                         H |= (L >> 4) << 8;

  3335.                         E |= s->filter.mblim_lut[L] << 8;

  3336.                         I |= s->filter.lim_lut[L] << 8;

  3337.                         s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3338.                                                [!!(vmask[1] & (x << 1))]

  3339.                                                [1](ptr, ls_y, E, I, H);

  3340.                     } else {

  3341.                         s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3342.                                             [1](ptr, ls_y, E, I, H);

  3343.                     }

  3344.                 } else if (vm & (x << 1)) {

  3345.                     int L = l[1], H = L >> 4;

  3346.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3347. 

  3348.                     s->dsp.loop_filter_8[!!(vmask[1] & (x << 1))]

  3349.                                         [1](ptr + 8, ls_y, E, I, H);

  3350.                 }

  3351.             }

  3352.             if (vm3 & x) {

  3353.                 int L = *l, H = L >> 4;

  3354.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3355. 

  3356.                 if (vm3 & (x << 1)) {

  3357.                     L = l[1];

  3358.                     H |= (L >> 4) << 8;

  3359.                     E |= s->filter.mblim_lut[L] << 8;

  3360.                     I |= s->filter.lim_lut[L] << 8;

  3361.                     s->dsp.loop_filter_mix2[0][0][1](ptr + ls_y * 4, ls_y, E, I, H);

  3362.                 } else {

  3363.                     s->dsp.loop_filter_8[0][1](ptr + ls_y * 4, ls_y, E, I, H);

  3364.                 }

  3365.             } else if (vm3 & (x << 1)) {

  3366.                 int L = l[1], H = L >> 4;

  3367.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3368. 

  3369.                 s->dsp.loop_filter_8[0][1](ptr + ls_y * 4 + 8, ls_y, E, I, H);

  3370.             }

  3371.         }

  3372.     }

  3373. 

  3374.     // same principle but for U/V planes

  3375.     for (p = 0; p < 2; p++) {

  3376.         lvl = lflvl->level;

  3377.         dst = f->data[1 + p] + uvoff;

  3378.         for (y = 0; y < 8; y += 4, dst += 16 * ls_uv, lvl += 32) {

  3379.             uint8_t *ptr = dst, *l = lvl, *hmask1 = lflvl->mask[1][0][y];

  3380.             uint8_t *hmask2 = lflvl->mask[1][0][y + 2];

  3381.             unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2];

  3382.             unsigned hm2 = hmask2[1] | hmask2[2], hm = hm1 | hm2;

  3383. 

  3384.             for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 4) {

  3385.                 if (col || x > 1) {

  3386.                     if (hm1 & x) {

  3387.                         int L = *l, H = L >> 4;

  3388.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3389. 

  3390.                         if (hmask1[0] & x) {

  3391.                             if (hmask2[0] & x) {

  3392.                                 av_assert2(l[16] == L);

  3393.                                 s->dsp.loop_filter_16[0](ptr, ls_uv, E, I, H);

  3394.                             } else {

  3395.                                 s->dsp.loop_filter_8[2][0](ptr, ls_uv, E, I, H);

  3396.                             }

  3397.                         } else if (hm2 & x) {

  3398.                             L = l[16];

  3399.                             H |= (L >> 4) << 8;

  3400.                             E |= s->filter.mblim_lut[L] << 8;

  3401.                             I |= s->filter.lim_lut[L] << 8;

  3402.                             s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3403.                                                    [!!(hmask2[1] & x)]

  3404.                                                    [0](ptr, ls_uv, E, I, H);

  3405.                         } else {

  3406.                             s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3407.                                                 [0](ptr, ls_uv, E, I, H);

  3408.                         }

  3409.                     } else if (hm2 & x) {

  3410.                         int L = l[16], H = L >> 4;

  3411.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3412. 

  3413.                         s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3414.                                             [0](ptr + 8 * ls_uv, ls_uv, E, I, H);

  3415.                     }

  3416.                 }

  3417.                 if (x & 0xAA)

  3418.                     l += 2;

  3419.             }

  3420.         }

  3421.         lvl = lflvl->level;

  3422.         dst = f->data[1 + p] + uvoff;

  3423.         for (y = 0; y < 8; y++, dst += 4 * ls_uv) {

  3424.             uint8_t *ptr = dst, *l = lvl, *vmask = lflvl->mask[1][1][y];

  3425.             unsigned vm = vmask[0] | vmask[1] | vmask[2];

  3426. 

  3427.             for (x = 1; vm & ~(x - 1); x <<= 4, ptr += 16, l += 4) {

  3428.                 if (row || y) {

  3429.                     if (vm & x) {

  3430.                         int L = *l, H = L >> 4;

  3431.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3432. 

  3433.                         if (vmask[0] & x) {

  3434.                             if (vmask[0] & (x << 2)) {

  3435.                                 av_assert2(l[2] == L);

  3436.                                 s->dsp.loop_filter_16[1](ptr, ls_uv, E, I, H);

  3437.                             } else {

  3438.                                 s->dsp.loop_filter_8[2][1](ptr, ls_uv, E, I, H);

  3439.                             }

  3440.                         } else if (vm & (x << 2)) {

  3441.                             L = l[2];

  3442.                             H |= (L >> 4) << 8;

  3443.                             E |= s->filter.mblim_lut[L] << 8;

  3444.                             I |= s->filter.lim_lut[L] << 8;

  3445.                             s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3446.                                                    [!!(vmask[1] & (x << 2))]

  3447.                                                    [1](ptr, ls_uv, E, I, H);

  3448.                         } else {

  3449.                             s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3450.                                                 [1](ptr, ls_uv, E, I, H);

  3451.                         }

  3452.                     } else if (vm & (x << 2)) {

  3453.                         int L = l[2], H = L >> 4;

  3454.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3455. 

  3456.                         s->dsp.loop_filter_8[!!(vmask[1] & (x << 2))]

  3457.                                             [1](ptr + 8, ls_uv, E, I, H);

  3458.                     }

  3459.                 }

  3460.             }

  3461.             if (y & 1)

  3462.                 lvl += 16;

  3463.         }

  3464.     }

  3465. }

  3466. 

  3467. static void set_tile_offset(int *start, int *end, int idx, int log2_n, int n)

  3468. {

  3469.     int sb_start = ( idx      * n) >> log2_n;

  3470.     int sb_end   = ((idx + 1) * n) >> log2_n;

  3471.     *start = FFMIN(sb_start, n) << 3;

  3472.     *end   = FFMIN(sb_end,   n) << 3;

  3473. }

  3474. 

  3475. static av_always_inline void adapt_prob(uint8_t *p, unsigned ct0, unsigned ct1,

  3476.                                         int max_count, int update_factor)

  3477. {

  3478.     unsigned ct = ct0 + ct1, p2, p1;

  3479. 

  3480.     if (!ct)

  3481.         return;

  3482. 

  3483.     update_factor = FASTDIV(update_factor * FFMIN(ct, max_count), max_count);

  3484.     p1 = *p;

  3485.     p2 = ((((int64_t) ct0) << 8) + (ct >> 1)) / ct;

  3486.     p2 = av_clip(p2, 1, 255);

  3487. 

  3488.     // (p1 * (256 - update_factor) + p2 * update_factor + 128) >> 8

  3489.     *p = p1 + (((p2 - p1) * update_factor + 128) >> 8);

  3490. }

  3491. 

  3492. static void adapt_probs(VP9Context *s)

  3493. {

  3494.     int i, j, k, l, m;

  3495.     prob_context *p = &s->prob_ctx[s->framectxid].p;

  3496.     int uf = (s->keyframe || s->intraonly || !s->last_keyframe) ? 112 : 128;

  3497. 

  3498.     // coefficients

  3499.     for (i = 0; i < 4; i++)

  3500.         for (j = 0; j < 2; j++)

  3501.             for (k = 0; k < 2; k++)

  3502.                 for (l = 0; l < 6; l++)

  3503.                     for (m = 0; m < 6; m++) {

  3504.                         uint8_t *pp = s->prob_ctx[s->framectxid].coef[i][j][k][l][m];

  3505.                         unsigned *e = s->counts.eob[i][j][k][l][m];

  3506.                         unsigned *c = s->counts.coef[i][j][k][l][m];

  3507. 

  3508.                         if (l == 0 && m >= 3) // dc only has 3 pt

  3509.                             break;

  3510. 

  3511.                         adapt_prob(&pp[0], e[0], e[1], 24, uf);

  3512.                         adapt_prob(&pp[1], c[0], c[1] + c[2], 24, uf);

  3513.                         adapt_prob(&pp[2], c[1], c[2], 24, uf);

  3514.                     }

  3515. 

  3516.     if (s->keyframe || s->intraonly) {

  3517.         memcpy(p->skip,  s->prob.p.skip,  sizeof(p->skip));

  3518.         memcpy(p->tx32p, s->prob.p.tx32p, sizeof(p->tx32p));

  3519.         memcpy(p->tx16p, s->prob.p.tx16p, sizeof(p->tx16p));

  3520.         memcpy(p->tx8p,  s->prob.p.tx8p,  sizeof(p->tx8p));

  3521.         return;

  3522.     }

  3523. 

  3524.     // skip flag

  3525.     for (i = 0; i < 3; i++)

  3526.         adapt_prob(&p->skip[i], s->counts.skip[i][0], s->counts.skip[i][1], 20, 128);

  3527. 

  3528.     // intra/inter flag

  3529.     for (i = 0; i < 4; i++)

  3530.         adapt_prob(&p->intra[i], s->counts.intra[i][0], s->counts.intra[i][1], 20, 128);

  3531. 

  3532.     // comppred flag

  3533.     if (s->comppredmode == PRED_SWITCHABLE) {

  3534.       for (i = 0; i < 5; i++)

  3535.           adapt_prob(&p->comp[i], s->counts.comp[i][0], s->counts.comp[i][1], 20, 128);

  3536.     }

  3537. 

  3538.     // reference frames

  3539.     if (s->comppredmode != PRED_SINGLEREF) {

  3540.       for (i = 0; i < 5; i++)

  3541.           adapt_prob(&p->comp_ref[i], s->counts.comp_ref[i][0],

  3542.                      s->counts.comp_ref[i][1], 20, 128);

  3543.     }

  3544. 

  3545.     if (s->comppredmode != PRED_COMPREF) {

  3546.       for (i = 0; i < 5; i++) {

  3547.           uint8_t *pp = p->single_ref[i];

  3548.           unsigned (*c)[2] = s->counts.single_ref[i];

  3549. 

  3550.           adapt_prob(&pp[0], c[0][0], c[0][1], 20, 128);

  3551.           adapt_prob(&pp[1], c[1][0], c[1][1], 20, 128);

  3552.       }

  3553.     }

  3554. 

  3555.     // block partitioning

  3556.     for (i = 0; i < 4; i++)

  3557.         for (j = 0; j < 4; j++) {

  3558.             uint8_t *pp = p->partition[i][j];

  3559.             unsigned *c = s->counts.partition[i][j];

  3560. 

  3561.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3562.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3563.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3564.         }

  3565. 

  3566.     // tx size

  3567.     if (s->txfmmode == TX_SWITCHABLE) {

  3568.       for (i = 0; i < 2; i++) {

  3569.           unsigned *c16 = s->counts.tx16p[i], *c32 = s->counts.tx32p[i];

  3570. 

  3571.           adapt_prob(&p->tx8p[i], s->counts.tx8p[i][0], s->counts.tx8p[i][1], 20, 128);

  3572.           adapt_prob(&p->tx16p[i][0], c16[0], c16[1] + c16[2], 20, 128);

  3573.           adapt_prob(&p->tx16p[i][1], c16[1], c16[2], 20, 128);

  3574.           adapt_prob(&p->tx32p[i][0], c32[0], c32[1] + c32[2] + c32[3], 20, 128);

  3575.           adapt_prob(&p->tx32p[i][1], c32[1], c32[2] + c32[3], 20, 128);

  3576.           adapt_prob(&p->tx32p[i][2], c32[2], c32[3], 20, 128);

  3577.       }

  3578.     }

  3579. 

  3580.     // interpolation filter

  3581.     if (s->filtermode == FILTER_SWITCHABLE) {

  3582.         for (i = 0; i < 4; i++) {

  3583.             uint8_t *pp = p->filter[i];

  3584.             unsigned *c = s->counts.filter[i];

  3585. 

  3586.             adapt_prob(&pp[0], c[0], c[1] + c[2], 20, 128);

  3587.             adapt_prob(&pp[1], c[1], c[2], 20, 128);

  3588.         }

  3589.     }

  3590. 

  3591.     // inter modes

  3592.     for (i = 0; i < 7; i++) {

  3593.         uint8_t *pp = p->mv_mode[i];

  3594.         unsigned *c = s->counts.mv_mode[i];

  3595. 

  3596.         adapt_prob(&pp[0], c[2], c[1] + c[0] + c[3], 20, 128);

  3597.         adapt_prob(&pp[1], c[0], c[1] + c[3], 20, 128);

  3598.         adapt_prob(&pp[2], c[1], c[3], 20, 128);

  3599.     }

  3600. 

  3601.     // mv joints

  3602.     {

  3603.         uint8_t *pp = p->mv_joint;

  3604.         unsigned *c = s->counts.mv_joint;

  3605. 

  3606.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3607.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3608.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3609.     }

  3610. 

  3611.     // mv components

  3612.     for (i = 0; i < 2; i++) {

  3613.         uint8_t *pp;

  3614.         unsigned *c, (*c2)[2], sum;

  3615. 

  3616.         adapt_prob(&p->mv_comp[i].sign, s->counts.mv_comp[i].sign[0],

  3617.                    s->counts.mv_comp[i].sign[1], 20, 128);

  3618. 

  3619.         pp = p->mv_comp[i].classes;

  3620.         c = s->counts.mv_comp[i].classes;

  3621.         sum = c[1] + c[2] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9] + c[10];

  3622.         adapt_prob(&pp[0], c[0], sum, 20, 128);

  3623.         sum -= c[1];

  3624.         adapt_prob(&pp[1], c[1], sum, 20, 128);

  3625.         sum -= c[2] + c[3];

  3626.         adapt_prob(&pp[2], c[2] + c[3], sum, 20, 128);

  3627.         adapt_prob(&pp[3], c[2], c[3], 20, 128);

  3628.         sum -= c[4] + c[5];

  3629.         adapt_prob(&pp[4], c[4] + c[5], sum, 20, 128);

  3630.         adapt_prob(&pp[5], c[4], c[5], 20, 128);

  3631.         sum -= c[6];

  3632.         adapt_prob(&pp[6], c[6], sum, 20, 128);

  3633.         adapt_prob(&pp[7], c[7] + c[8], c[9] + c[10], 20, 128);

  3634.         adapt_prob(&pp[8], c[7], c[8], 20, 128);

  3635.         adapt_prob(&pp[9], c[9], c[10], 20, 128);

  3636. 

  3637.         adapt_prob(&p->mv_comp[i].class0, s->counts.mv_comp[i].class0[0],

  3638.                    s->counts.mv_comp[i].class0[1], 20, 128);

  3639.         pp = p->mv_comp[i].bits;

  3640.         c2 = s->counts.mv_comp[i].bits;

  3641.         for (j = 0; j < 10; j++)

  3642.             adapt_prob(&pp[j], c2[j][0], c2[j][1], 20, 128);

  3643. 

  3644.         for (j = 0; j < 2; j++) {

  3645.             pp = p->mv_comp[i].class0_fp[j];

  3646.             c = s->counts.mv_comp[i].class0_fp[j];

  3647.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3648.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3649.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3650.         }

  3651.         pp = p->mv_comp[i].fp;

  3652.         c = s->counts.mv_comp[i].fp;

  3653.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3654.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3655.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3656. 

  3657.         if (s->highprecisionmvs) {

  3658.             adapt_prob(&p->mv_comp[i].class0_hp, s->counts.mv_comp[i].class0_hp[0],

  3659.                        s->counts.mv_comp[i].class0_hp[1], 20, 128);

  3660.             adapt_prob(&p->mv_comp[i].hp, s->counts.mv_comp[i].hp[0],

  3661.                        s->counts.mv_comp[i].hp[1], 20, 128);

  3662.         }

  3663.     }

  3664. 

  3665.     // y intra modes

  3666.     for (i = 0; i < 4; i++) {

  3667.         uint8_t *pp = p->y_mode[i];

  3668.         unsigned *c = s->counts.y_mode[i], sum, s2;

  3669. 

  3670.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3671.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3672.         sum -= c[TM_VP8_PRED];

  3673.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3674.         sum -= c[VERT_PRED];

  3675.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3676.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3677.         sum -= s2;

  3678.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3679.         s2 -= c[HOR_PRED];

  3680.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3681.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3682.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3683.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3684.         sum -= c[VERT_LEFT_PRED];

  3685.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3686.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3687.     }

  3688. 

  3689.     // uv intra modes

  3690.     for (i = 0; i < 10; i++) {

  3691.         uint8_t *pp = p->uv_mode[i];

  3692.         unsigned *c = s->counts.uv_mode[i], sum, s2;

  3693. 

  3694.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3695.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3696.         sum -= c[TM_VP8_PRED];

  3697.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3698.         sum -= c[VERT_PRED];

  3699.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3700.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3701.         sum -= s2;

  3702.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3703.         s2 -= c[HOR_PRED];

  3704.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3705.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3706.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3707.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3708.         sum -= c[VERT_LEFT_PRED];

  3709.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3710.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3711.     }

  3712. }

  3713. 

  3714. static void free_buffers(VP9Context *s)

  3715. {

  3716.     av_freep(&s->intra_pred_data[0]);

  3717.     av_freep(&s->b_base);

  3718.     av_freep(&s->block_base);

  3719. }

  3720. 

  3721. static av_cold int vp9_decode_free(AVCodecContext *ctx)

  3722. {

  3723.     VP9Context *s = ctx->priv_data;

  3724.     int i;

  3725. 

  3726.     for (i = 0; i < 2; i++) {

  3727.         if (s->frames[i].tf.f->data[0])

  3728.             vp9_unref_frame(ctx, &s->frames[i]);

  3729.         av_frame_free(&s->frames[i].tf.f);

  3730.     }

  3731.     for (i = 0; i < 8; i++) {

  3732.         if (s->refs[i].f->data[0])

  3733.             ff_thread_release_buffer(ctx, &s->refs[i]);

  3734.         av_frame_free(&s->refs[i].f);

  3735.         if (s->next_refs[i].f->data[0])

  3736.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3737.         av_frame_free(&s->next_refs[i].f);

  3738.     }

  3739.     free_buffers(s);

  3740.     av_freep(&s->c_b);

  3741.     s->c_b_size = 0;

  3742. 

  3743.     return 0;

  3744. }

  3745. 

  3746. 

  3747. static int vp9_decode_frame(AVCodecContext *ctx, void *frame,

  3748.                             int *got_frame, AVPacket *pkt)

  3749. {

  3750.     const uint8_t *data = pkt->data;

  3751.     int size = pkt->size;

  3752.     VP9Context *s = ctx->priv_data;

  3753.     int res, tile_row, tile_col, i, ref, row, col;

  3754.     ptrdiff_t yoff, uvoff, ls_y, ls_uv;

  3755.     AVFrame *f;

  3756. 

  3757.     if ((res = decode_frame_header(ctx, data, size, &ref)) < 0) {

  3758.         return res;

  3759.     } else if (res == 0) {

  3760.         if (!s->refs[ref].f->data[0]) {

  3761.             av_log(ctx, AV_LOG_ERROR, "Requested reference %d not available\n", ref);

  3762.             return AVERROR_INVALIDDATA;

  3763.         }

  3764.         if ((res = av_frame_ref(frame, s->refs[ref].f)) < 0)

  3765.             return res;

  3766.         *got_frame = 1;

  3767.         return pkt->size;

  3768.     }

  3769.     data += res;

  3770.     size -= res;

  3771. 

  3772.     if (s->frames[LAST_FRAME].tf.f->data[0])

  3773.         vp9_unref_frame(ctx, &s->frames[LAST_FRAME]);

  3774.     if (!s->keyframe && s->frames[CUR_FRAME].tf.f->data[0] &&

  3775.         (res = vp9_ref_frame(ctx, &s->frames[LAST_FRAME], &s->frames[CUR_FRAME])) < 0)

  3776.         return res;

  3777.     if (s->frames[CUR_FRAME].tf.f->data[0])

  3778.         vp9_unref_frame(ctx, &s->frames[CUR_FRAME]);

  3779.     if ((res = vp9_alloc_frame(ctx, &s->frames[CUR_FRAME])) < 0)

  3780.         return res;

  3781.     f = s->frames[CUR_FRAME].tf.f;

  3782.     f->key_frame = s->keyframe;

  3783.     f->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;

  3784.     ls_y = f->linesize[0];

  3785.     ls_uv =f->linesize[1];

  3786. 

  3787.     // ref frame setup

  3788.     for (i = 0; i < 8; i++) {

  3789.         if (s->next_refs[i].f->data[0])

  3790.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3791.         if (s->refreshrefmask & (1 << i)) {

  3792.             res = ff_thread_ref_frame(&s->next_refs[i], &s->frames[CUR_FRAME].tf);

  3793.         } else {

  3794.             res = ff_thread_ref_frame(&s->next_refs[i], &s->refs[i]);

  3795.         }

  3796.         if (res < 0)

  3797.             return res;

  3798.     }

  3799. 

  3800.     // main tile decode loop

  3801.     memset(s->above_partition_ctx, 0, s->cols);

  3802.     memset(s->above_skip_ctx, 0, s->cols);

  3803.     if (s->keyframe || s->intraonly) {

  3804.         memset(s->above_mode_ctx, DC_PRED, s->cols * 2);

  3805.     } else {

  3806.         memset(s->above_mode_ctx, NEARESTMV, s->cols);

  3807.     }

  3808.     memset(s->above_y_nnz_ctx, 0, s->sb_cols * 16);

  3809.     memset(s->above_uv_nnz_ctx[0], 0, s->sb_cols * 8);

  3810.     memset(s->above_uv_nnz_ctx[1], 0, s->sb_cols * 8);

  3811.     memset(s->above_segpred_ctx, 0, s->cols);

  3812.     s->pass = s->uses_2pass =

  3813.         ctx->active_thread_type == FF_THREAD_FRAME && s->refreshctx && !s->parallelmode;

  3814.     if ((res = update_block_buffers(ctx)) < 0) {

  3815.         av_log(ctx, AV_LOG_ERROR,

  3816.                "Failed to allocate block buffers\n");

  3817.         return res;

  3818.     }

  3819.     if (s->refreshctx && s->parallelmode) {

  3820.         int j, k, l, m;

  3821. 

  3822.         for (i = 0; i < 4; i++) {

  3823.             for (j = 0; j < 2; j++)

  3824.                 for (k = 0; k < 2; k++)

  3825.                     for (l = 0; l < 6; l++)

  3826.                         for (m = 0; m < 6; m++)

  3827.                             memcpy(s->prob_ctx[s->framectxid].coef[i][j][k][l][m],

  3828.                                    s->prob.coef[i][j][k][l][m], 3);

  3829.             if (s->txfmmode == i)

  3830.                 break;

  3831.         }

  3832.         s->prob_ctx[s->framectxid].p = s->prob.p;

  3833.         ff_thread_finish_setup(ctx);

  3834.     }

  3835. 

  3836.     do {

  3837.         yoff = uvoff = 0;

  3838.         s->b = s->b_base;

  3839.         s->block = s->block_base;

  3840.         s->uvblock[0] = s->uvblock_base[0];

  3841.         s->uvblock[1] = s->uvblock_base[1];

  3842.         s->eob = s->eob_base;

  3843.         s->uveob[0] = s->uveob_base[0];

  3844.         s->uveob[1] = s->uveob_base[1];

  3845. 

  3846.         for (tile_row = 0; tile_row < s->tiling.tile_rows; tile_row++) {

  3847.             set_tile_offset(&s->tiling.tile_row_start, &s->tiling.tile_row_end,

  3848.                             tile_row, s->tiling.log2_tile_rows, s->sb_rows);

  3849.             if (s->pass != 2) {

  3850.                 for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  3851.                     int64_t tile_size;

  3852. 

  3853.                     if (tile_col == s->tiling.tile_cols - 1 &&

  3854.                         tile_row == s->tiling.tile_rows - 1) {

  3855.                         tile_size = size;

  3856.                     } else {

  3857.                         tile_size = AV_RB32(data);

  3858.                         data += 4;

  3859.                         size -= 4;

  3860.                     }

  3861.                     if (tile_size > size) {

  3862.                         ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  3863.                         return AVERROR_INVALIDDATA;

  3864.                     }

  3865.                     res = ff_vp56_init_range_decoder(&s->c_b[tile_col], data, tile_size);

  3866.                     if (res < 0)

  3867.                         return res;

  3868.                     if (vp56_rac_get_prob_branchy(&s->c_b[tile_col], 128)) { // marker bit

  3869.                         ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  3870.                         return AVERROR_INVALIDDATA;

  3871.                     }

  3872.                     data += tile_size;

  3873.                     size -= tile_size;

  3874.                 }

  3875.             }

  3876. 

  3877.             for (row = s->tiling.tile_row_start; row < s->tiling.tile_row_end;

  3878.                  row += 8, yoff += ls_y * 64, uvoff += ls_uv * 32) {

  3879.                 struct VP9Filter *lflvl_ptr = s->lflvl;

  3880.                 ptrdiff_t yoff2 = yoff, uvoff2 = uvoff;

  3881. 

  3882.                 for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  3883.                     set_tile_offset(&s->tiling.tile_col_start, &s->tiling.tile_col_end,

  3884.                                     tile_col, s->tiling.log2_tile_cols, s->sb_cols);

  3885. 

  3886.                     if (s->pass != 2) {

  3887.                         memset(s->left_partition_ctx, 0, 8);

  3888.                         memset(s->left_skip_ctx, 0, 8);

  3889.                         if (s->keyframe || s->intraonly) {

  3890.                             memset(s->left_mode_ctx, DC_PRED, 16);

  3891.                         } else {

  3892.                             memset(s->left_mode_ctx, NEARESTMV, 8);

  3893.                         }

  3894.                         memset(s->left_y_nnz_ctx, 0, 16);

  3895.                         memset(s->left_uv_nnz_ctx, 0, 16);

  3896.                         memset(s->left_segpred_ctx, 0, 8);

  3897. 

  3898.                         memcpy(&s->c, &s->c_b[tile_col], sizeof(s->c));

  3899.                     }

  3900. 

  3901.                     for (col = s->tiling.tile_col_start;

  3902.                          col < s->tiling.tile_col_end;

  3903.                          col += 8, yoff2 += 64, uvoff2 += 32, lflvl_ptr++) {

  3904.                         // FIXME integrate with lf code (i.e. zero after each

  3905.                         // use, similar to invtxfm coefficients, or similar)

  3906.                         if (s->pass != 1) {

  3907.                             memset(lflvl_ptr->mask, 0, sizeof(lflvl_ptr->mask));

  3908.                         }

  3909. 

  3910.                         if (s->pass == 2) {

  3911.                             decode_sb_mem(ctx, row, col, lflvl_ptr,

  3912.                                           yoff2, uvoff2, BL_64X64);

  3913.                         } else {

  3914.                             decode_sb(ctx, row, col, lflvl_ptr,

  3915.                                       yoff2, uvoff2, BL_64X64);

  3916.                         }

  3917.                     }

  3918.                     if (s->pass != 2) {

  3919.                         memcpy(&s->c_b[tile_col], &s->c, sizeof(s->c));

  3920.                     }

  3921.                 }

  3922. 

  3923.                 if (s->pass == 1) {

  3924.                     continue;

  3925.                 }

  3926. 

  3927.                 // backup pre-loopfilter reconstruction data for intra

  3928.                 // prediction of next row of sb64s

  3929.                 if (row + 8 < s->rows) {

  3930.                     memcpy(s->intra_pred_data[0],

  3931.                            f->data[0] + yoff + 63 * ls_y,

  3932.                            8 * s->cols);

  3933.                     memcpy(s->intra_pred_data[1],

  3934.                            f->data[1] + uvoff + 31 * ls_uv,

  3935.                            4 * s->cols);

  3936.                     memcpy(s->intra_pred_data[2],

  3937.                            f->data[2] + uvoff + 31 * ls_uv,

  3938.                            4 * s->cols);

  3939.                 }

  3940. 

  3941.                 // loopfilter one row

  3942.                 if (s->filter.level) {

  3943.                     yoff2 = yoff;

  3944.                     uvoff2 = uvoff;

  3945.                     lflvl_ptr = s->lflvl;

  3946.                     for (col = 0; col < s->cols;

  3947.                          col += 8, yoff2 += 64, uvoff2 += 32, lflvl_ptr++) {

  3948.                         loopfilter_sb(ctx, lflvl_ptr, row, col, yoff2, uvoff2);

  3949.                     }

  3950.                 }

  3951. 

  3952.                 // FIXME maybe we can make this more finegrained by running the

  3953.                 // loopfilter per-block instead of after each sbrow

  3954.                 // In fact that would also make intra pred left preparation easier?

  3955.                 ff_thread_report_progress(&s->frames[CUR_FRAME].tf, row >> 3, 0);

  3956.             }

  3957.         }

  3958. 

  3959.         if (s->pass < 2 && s->refreshctx && !s->parallelmode) {

  3960.             adapt_probs(s);

  3961.             ff_thread_finish_setup(ctx);

  3962.         }

  3963.     } while (s->pass++ == 1);

  3964.     ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  3965. 

  3966.     // ref frame setup

  3967.     for (i = 0; i < 8; i++) {

  3968.         if (s->refs[i].f->data[0])

  3969.             ff_thread_release_buffer(ctx, &s->refs[i]);

  3970.         ff_thread_ref_frame(&s->refs[i], &s->next_refs[i]);

  3971.     }

  3972. 

  3973.     if (!s->invisible) {

  3974.         if ((res = av_frame_ref(frame, s->frames[CUR_FRAME].tf.f)) < 0)

  3975.             return res;

  3976.         *got_frame = 1;

  3977.     }

  3978. 

  3979.     return pkt->size;

  3980. }

  3981. 

  3982. static void vp9_decode_flush(AVCodecContext *ctx)

  3983. {

  3984.     VP9Context *s = ctx->priv_data;

  3985.     int i;

  3986. 

  3987.     for (i = 0; i < 2; i++)

  3988.         vp9_unref_frame(ctx, &s->frames[i]);

  3989.     for (i = 0; i < 8; i++)

  3990.         ff_thread_release_buffer(ctx, &s->refs[i]);

  3991. }

  3992. 

  3993. static int init_frames(AVCodecContext *ctx)

  3994. {

  3995.     VP9Context *s = ctx->priv_data;

  3996.     int i;

  3997. 

  3998.     for (i = 0; i < 2; i++) {

  3999.         s->frames[i].tf.f = av_frame_alloc();

  4000.         if (!s->frames[i].tf.f) {

  4001.             vp9_decode_free(ctx);

  4002.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  4003.             return AVERROR(ENOMEM);

  4004.         }

  4005.     }

  4006.     for (i = 0; i < 8; i++) {

  4007.         s->refs[i].f = av_frame_alloc();

  4008.         s->next_refs[i].f = av_frame_alloc();

  4009.         if (!s->refs[i].f || !s->next_refs[i].f) {

  4010.             vp9_decode_free(ctx);

  4011.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  4012.             return AVERROR(ENOMEM);

  4013.         }

  4014.     }

  4015. 

  4016.     return 0;

  4017. }

  4018. 

  4019. static av_cold int vp9_decode_init(AVCodecContext *ctx)

  4020. {

  4021.     VP9Context *s = ctx->priv_data;

  4022. 

  4023.     ctx->internal->allocate_progress = 1;

  4024.     ctx->pix_fmt = AV_PIX_FMT_YUV420P;

  4025.     ff_vp9dsp_init(&s->dsp);

  4026.     ff_videodsp_init(&s->vdsp, 8);

  4027.     s->filter.sharpness = -1;

  4028. 

  4029.     return init_frames(ctx);

  4030. }

  4031. 

  4032. static av_cold int vp9_decode_init_thread_copy(AVCodecContext *avctx)

  4033. {

  4034.     return init_frames(avctx);

  4035. }

  4036. 

  4037. static int vp9_decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src)

  4038. {

  4039.     int i, res;

  4040.     VP9Context *s = dst->priv_data, *ssrc = src->priv_data;

  4041. 

  4042.     // detect size changes in other threads

  4043.     if (s->intra_pred_data[0] &&

  4044.         (!ssrc->intra_pred_data[0] || s->cols != ssrc->cols || s->rows != ssrc->rows)) {

  4045.         free_buffers(s);

  4046.     }

  4047. 

  4048.     for (i = 0; i < 2; i++) {

  4049.         if (s->frames[i].tf.f->data[0])

  4050.             vp9_unref_frame(dst, &s->frames[i]);

  4051.         if (ssrc->frames[i].tf.f->data[0]) {

  4052.             if ((res = vp9_ref_frame(dst, &s->frames[i], &ssrc->frames[i])) < 0)

  4053.                 return res;

  4054.         }

  4055.     }

  4056.     for (i = 0; i < 8; i++) {

  4057.         if (s->refs[i].f->data[0])

  4058.             ff_thread_release_buffer(dst, &s->refs[i]);

  4059.         if (ssrc->next_refs[i].f->data[0]) {

  4060.             if ((res = ff_thread_ref_frame(&s->refs[i], &ssrc->next_refs[i])) < 0)

  4061.                 return res;

  4062.         }

  4063.     }

  4064. 

  4065.     s->invisible = ssrc->invisible;

  4066.     s->keyframe = ssrc->keyframe;

  4067.     s->uses_2pass = ssrc->uses_2pass;

  4068.     memcpy(&s->prob_ctx, &ssrc->prob_ctx, sizeof(s->prob_ctx));

  4069.     memcpy(&s->lf_delta, &ssrc->lf_delta, sizeof(s->lf_delta));

  4070.     if (ssrc->segmentation.enabled) {

  4071.         memcpy(&s->segmentation.feat, &ssrc->segmentation.feat,

  4072.                sizeof(s->segmentation.feat));

  4073.     }

  4074. 

  4075.     return 0;

  4076. }

  4077. 

  4078. AVCodec ff_vp9_decoder = {

  4079.     .name                  = "vp9",

  4080.     .long_name             = NULL_IF_CONFIG_SMALL("Google VP9"),

  4081.     .type                  = AVMEDIA_TYPE_VIDEO,

  4082.     .id                    = AV_CODEC_ID_VP9,

  4083.     .priv_data_size        = sizeof(VP9Context),

  4084.     .init                  = vp9_decode_init,

  4085.     .close                 = vp9_decode_free,

  4086.     .decode                = vp9_decode_frame,

  4087.     .capabilities          = CODEC_CAP_DR1 | CODEC_CAP_FRAME_THREADS,

  4088.     .flush                 = vp9_decode_flush,

  4089.     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(vp9_decode_init_thread_copy),

  4090.     .update_thread_context = ONLY_IF_THREADS_ENABLED(vp9_decode_update_thread_context),

  4091. };