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.
77088 Commits
32 Branches
291MB
Tree: 585083dd1f
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '585083dd1f'
${ noResults }
FFmpeg/libavcodec/vp9.c

4350 lines
175KB
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. #include "libavutil/pixdesc.h"

  35. 

  36. #define VP9_SYNCCODE 0x498342

  37. 

  38. struct VP9Filter {

  39.     uint8_t level[8 * 8];

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

  41.                               [8 /* rows */][4 /* 0=16, 1=8, 2=4, 3=inner4 */];

  42. };

  43. 

  44. typedef struct VP9Block {

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

  46.     enum FilterMode filter;

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

  48.     enum BlockSize bs;

  49.     enum TxfmMode tx, uvtx;

  50.     enum BlockLevel bl;

  51.     enum BlockPartition bp;

  52. } VP9Block;

  53. 

  54. typedef struct VP9Context {

  55.     VP9SharedContext s;

  56. 

  57.     VP9DSPContext dsp;

  58.     VideoDSPContext vdsp;

  59.     GetBitContext gb;

  60.     VP56RangeCoder c;

  61.     VP56RangeCoder *c_b;

  62.     unsigned c_b_size;

  63.     VP9Block *b_base, *b;

  64.     int pass;

  65.     int row, row7, col, col7;

  66.     uint8_t *dst[3];

  67.     ptrdiff_t y_stride, uv_stride;

  68. 

  69.     uint8_t ss_h, ss_v;

  70.     uint8_t last_bpp, bpp, bpp_index, bytesperpixel;

  71.     uint8_t last_keyframe;

  72.     enum AVPixelFormat pix_fmt, last_fmt;

  73.     ThreadFrame next_refs[8];

  74. 

  75.     struct {

  76.         uint8_t lim_lut[64];

  77.         uint8_t mblim_lut[64];

  78.     } filter_lut;

  79.     unsigned tile_row_start, tile_row_end, tile_col_start, tile_col_end;

  80.     unsigned sb_cols, sb_rows, rows, cols;

  81.     struct {

  82.         prob_context p;

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

  84.     } prob_ctx[4];

  85.     struct {

  86.         prob_context p;

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

  88.     } prob;

  89.     struct {

  90.         unsigned y_mode[4][10];

  91.         unsigned uv_mode[10][10];

  92.         unsigned filter[4][3];

  93.         unsigned mv_mode[7][4];

  94.         unsigned intra[4][2];

  95.         unsigned comp[5][2];

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

  97.         unsigned comp_ref[5][2];

  98.         unsigned tx32p[2][4];

  99.         unsigned tx16p[2][3];

  100.         unsigned tx8p[2][2];

  101.         unsigned skip[3][2];

  102.         unsigned mv_joint[4];

  103.         struct {

  104.             unsigned sign[2];

  105.             unsigned classes[11];

  106.             unsigned class0[2];

  107.             unsigned bits[10][2];

  108.             unsigned class0_fp[2][4];

  109.             unsigned fp[4];

  110.             unsigned class0_hp[2];

  111.             unsigned hp[2];

  112.         } mv_comp[2];

  113.         unsigned partition[4][4][4];

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

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

  116.     } counts;

  117. 

  118.     // contextual (left/above) cache

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

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

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

  122.     DECLARE_ALIGNED(16, uint8_t, left_uv_nnz_ctx)[2][16];

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

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

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

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

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

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

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

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

  131.     uint8_t *above_partition_ctx;

  132.     uint8_t *above_mode_ctx;

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

  134.     uint8_t *above_y_nnz_ctx;

  135.     uint8_t *above_uv_nnz_ctx[2];

  136.     uint8_t *above_skip_ctx; // 1bit

  137.     uint8_t *above_txfm_ctx; // 2bit

  138.     uint8_t *above_segpred_ctx; // 1bit

  139.     uint8_t *above_intra_ctx; // 1bit

  140.     uint8_t *above_comp_ctx; // 1bit

  141.     uint8_t *above_ref_ctx; // 2bit

  142.     uint8_t *above_filter_ctx;

  143.     VP56mv (*above_mv_ctx)[2];

  144. 

  145.     // whole-frame cache

  146.     uint8_t *intra_pred_data[3];

  147.     struct VP9Filter *lflvl;

  148.     DECLARE_ALIGNED(32, uint8_t, edge_emu_buffer)[135 * 144 * 2];

  149. 

  150.     // block reconstruction intermediates

  151.     int block_alloc_using_2pass;

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

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

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

  155.     DECLARE_ALIGNED(32, uint8_t, tmp_y)[64 * 64 * 2];

  156.     DECLARE_ALIGNED(32, uint8_t, tmp_uv)[2][64 * 64 * 2];

  157.     uint16_t mvscale[3][2];

  158.     uint8_t mvstep[3][2];

  159. } VP9Context;

  160. 

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

  162.     {

  163.         { 16, 16 }, { 16, 8 }, { 8, 16 }, { 8, 8 }, { 8, 4 }, { 4, 8 },

  164.         { 4, 4 }, { 4, 2 }, { 2, 4 }, { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 },

  165.     }, {

  166.         { 8, 8 }, { 8, 4 }, { 4, 8 }, { 4, 4 }, { 4, 2 }, { 2, 4 },

  167.         { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 }, { 1, 1 }, { 1, 1 }, { 1, 1 },

  168.     }

  169. };

  170. 

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

  172. {

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

  174.     av_buffer_unref(&f->extradata);

  175.     av_buffer_unref(&f->hwaccel_priv_buf);

  176.     f->segmentation_map = NULL;

  177.     f->hwaccel_picture_private = NULL;

  178. }

  179. 

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

  181. {

  182.     VP9Context *s = ctx->priv_data;

  183.     int ret, sz;

  184. 

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

  186.         return ret;

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

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

  189.         goto fail;

  190.     }

  191. 

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

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

  194. 

  195.     if (ctx->hwaccel) {

  196.         const AVHWAccel *hwaccel = ctx->hwaccel;

  197.         av_assert0(!f->hwaccel_picture_private);

  198.         if (hwaccel->frame_priv_data_size) {

  199.             f->hwaccel_priv_buf = av_buffer_allocz(hwaccel->frame_priv_data_size);

  200.             if (!f->hwaccel_priv_buf)

  201.                 goto fail;

  202.             f->hwaccel_picture_private = f->hwaccel_priv_buf->data;

  203.         }

  204.     }

  205. 

  206.     return 0;

  207. 

  208. fail:

  209.     vp9_unref_frame(ctx, f);

  210.     return AVERROR(ENOMEM);

  211. }

  212. 

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

  214. {

  215.     int res;

  216. 

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

  218.         return res;

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

  220.         goto fail;

  221.     }

  222. 

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

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

  225.     dst->uses_2pass = src->uses_2pass;

  226. 

  227.     if (src->hwaccel_picture_private) {

  228.         dst->hwaccel_priv_buf = av_buffer_ref(src->hwaccel_priv_buf);

  229.         if (!dst->hwaccel_priv_buf)

  230.             goto fail;

  231.         dst->hwaccel_picture_private = dst->hwaccel_priv_buf->data;

  232.     }

  233. 

  234.     return 0;

  235. 

  236. fail:

  237.     vp9_unref_frame(ctx, dst);

  238.     return AVERROR(ENOMEM);

  239. }

  240. 

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

  242. {

  243. #define HWACCEL_MAX (0)

  244.     enum AVPixelFormat pix_fmts[HWACCEL_MAX + 2], *fmtp = pix_fmts;

  245.     VP9Context *s = ctx->priv_data;

  246.     uint8_t *p;

  247.     int bytesperpixel = s->bytesperpixel, res;

  248. 

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

  250. 

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

  252.         return 0;

  253. 

  254.     if ((res = ff_set_dimensions(ctx, w, h)) < 0)

  255.         return res;

  256. 

  257.     *fmtp++ = s->pix_fmt;

  258.     *fmtp = AV_PIX_FMT_NONE;

  259. 

  260.     res = ff_thread_get_format(ctx, pix_fmts);

  261.     if (res < 0)

  262.         return res;

  263. 

  264.     ctx->pix_fmt = res;

  265.     s->last_fmt  = s->pix_fmt;

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

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

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

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

  270. 

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

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

  273.     // FIXME we slightly over-allocate here for subsampled chroma, but a little

  274.     // bit of padding shouldn't affect performance...

  275.     p = av_malloc(s->sb_cols * (128 + 192 * bytesperpixel +

  276.                                 sizeof(*s->lflvl) + 16 * sizeof(*s->above_mv_ctx)));

  277.     if (!p)

  278.         return AVERROR(ENOMEM);

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

  280.     assign(s->intra_pred_data[1],  uint8_t *,             64 * bytesperpixel);

  281.     assign(s->intra_pred_data[2],  uint8_t *,             64 * bytesperpixel);

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

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

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

  285.     assign(s->above_uv_nnz_ctx[0], uint8_t *,             16);

  286.     assign(s->above_uv_nnz_ctx[1], uint8_t *,             16);

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

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

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

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

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

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

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

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

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

  296. #undef assign

  297. 

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

  299.     av_freep(&s->b_base);

  300.     av_freep(&s->block_base);

  301. 

  302.     if (s->bpp != s->last_bpp) {

  303.         ff_vp9dsp_init(&s->dsp, s->bpp, ctx->flags & AV_CODEC_FLAG_BITEXACT);

  304.         ff_videodsp_init(&s->vdsp, s->bpp);

  305.         s->last_bpp = s->bpp;

  306.     }

  307. 

  308.     return 0;

  309. }

  310. 

  311. static int update_block_buffers(AVCodecContext *ctx)

  312. {

  313.     VP9Context *s = ctx->priv_data;

  314.     int chroma_blocks, chroma_eobs, bytesperpixel = s->bytesperpixel;

  315. 

  316.     if (s->b_base && s->block_base && s->block_alloc_using_2pass == s->s.frames[CUR_FRAME].uses_2pass)

  317.         return 0;

  318. 

  319.     av_free(s->b_base);

  320.     av_free(s->block_base);

  321.     chroma_blocks = 64 * 64 >> (s->ss_h + s->ss_v);

  322.     chroma_eobs   = 16 * 16 >> (s->ss_h + s->ss_v);

  323.     if (s->s.frames[CUR_FRAME].uses_2pass) {

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

  325. 

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

  327.         s->block_base = av_mallocz(((64 * 64 + 2 * chroma_blocks) * bytesperpixel * sizeof(int16_t) +

  328.                                     16 * 16 + 2 * chroma_eobs) * sbs);

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

  330.             return AVERROR(ENOMEM);

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

  332.         s->uvblock_base[1] = s->uvblock_base[0] + sbs * chroma_blocks * bytesperpixel;

  333.         s->eob_base = (uint8_t *) (s->uvblock_base[1] + sbs * chroma_blocks * bytesperpixel);

  334.         s->uveob_base[0] = s->eob_base + 16 * 16 * sbs;

  335.         s->uveob_base[1] = s->uveob_base[0] + chroma_eobs * sbs;

  336.     } else {

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

  338.         s->block_base = av_mallocz((64 * 64 + 2 * chroma_blocks) * bytesperpixel * sizeof(int16_t) +

  339.                                    16 * 16 + 2 * chroma_eobs);

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

  341.             return AVERROR(ENOMEM);

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

  343.         s->uvblock_base[1] = s->uvblock_base[0] + chroma_blocks * bytesperpixel;

  344.         s->eob_base = (uint8_t *) (s->uvblock_base[1] + chroma_blocks * bytesperpixel);

  345.         s->uveob_base[0] = s->eob_base + 16 * 16;

  346.         s->uveob_base[1] = s->uveob_base[0] + chroma_eobs;

  347.     }

  348.     s->block_alloc_using_2pass = s->s.frames[CUR_FRAME].uses_2pass;

  349. 

  350.     return 0;

  351. }

  352. 

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

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

  355. {

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

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

  358. }

  359. 

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

  361. {

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

  363. }

  364. 

  365. // differential forward probability updates

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

  367. {

  368.     static const int inv_map_table[255] = {

  369.           7,  20,  33,  46,  59,  72,  85,  98, 111, 124, 137, 150, 163, 176,

  370.         189, 202, 215, 228, 241, 254,   1,   2,   3,   4,   5,   6,   8,   9,

  371.          10,  11,  12,  13,  14,  15,  16,  17,  18,  19,  21,  22,  23,  24,

  372.          25,  26,  27,  28,  29,  30,  31,  32,  34,  35,  36,  37,  38,  39,

  373.          40,  41,  42,  43,  44,  45,  47,  48,  49,  50,  51,  52,  53,  54,

  374.          55,  56,  57,  58,  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,

  375.          70,  71,  73,  74,  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,

  376.          86,  87,  88,  89,  90,  91,  92,  93,  94,  95,  96,  97,  99, 100,

  377.         101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 112, 113, 114, 115,

  378.         116, 117, 118, 119, 120, 121, 122, 123, 125, 126, 127, 128, 129, 130,

  379.         131, 132, 133, 134, 135, 136, 138, 139, 140, 141, 142, 143, 144, 145,

  380.         146, 147, 148, 149, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160,

  381.         161, 162, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175,

  382.         177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 190, 191,

  383.         192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 204, 205, 206,

  384.         207, 208, 209, 210, 211, 212, 213, 214, 216, 217, 218, 219, 220, 221,

  385.         222, 223, 224, 225, 226, 227, 229, 230, 231, 232, 233, 234, 235, 236,

  386.         237, 238, 239, 240, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251,

  387.         252, 253, 253,

  388.     };

  389.     int d;

  390. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  405. 

  406.     if (!vp8_rac_get(c)) {

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

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

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

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

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

  412.     } else {

  413.         d = vp8_rac_get_uint(c, 7);

  414.         if (d >= 65)

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

  416.         d += 64;

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

  418.     }

  419. 

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

  421.                     255 - inv_recenter_nonneg(inv_map_table[d], 255 - p);

  422. }

  423. 

  424. static int read_colorspace_details(AVCodecContext *ctx)

  425. {

  426.     static const enum AVColorSpace colorspaces[8] = {

  427.         AVCOL_SPC_UNSPECIFIED, AVCOL_SPC_BT470BG, AVCOL_SPC_BT709, AVCOL_SPC_SMPTE170M,

  428.         AVCOL_SPC_SMPTE240M, AVCOL_SPC_BT2020_NCL, AVCOL_SPC_RESERVED, AVCOL_SPC_RGB,

  429.     };

  430.     VP9Context *s = ctx->priv_data;

  431.     int bits = ctx->profile <= 1 ? 0 : 1 + get_bits1(&s->gb); // 0:8, 1:10, 2:12

  432. 

  433.     s->bpp_index = bits;

  434.     s->bpp = 8 + bits * 2;

  435.     s->bytesperpixel = (7 + s->bpp) >> 3;

  436.     ctx->colorspace = colorspaces[get_bits(&s->gb, 3)];

  437.     if (ctx->colorspace == AVCOL_SPC_RGB) { // RGB = profile 1

  438.         static const enum AVPixelFormat pix_fmt_rgb[3] = {

  439.             AV_PIX_FMT_GBRP, AV_PIX_FMT_GBRP10, AV_PIX_FMT_GBRP12

  440.         };

  441.         s->ss_h = s->ss_v = 0;

  442.         ctx->color_range = AVCOL_RANGE_JPEG;

  443.         s->pix_fmt = pix_fmt_rgb[bits];

  444.         if (ctx->profile & 1) {

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

  446.                 av_log(ctx, AV_LOG_ERROR, "Reserved bit set in RGB\n");

  447.                 return AVERROR_INVALIDDATA;

  448.             }

  449.         } else {

  450.             av_log(ctx, AV_LOG_ERROR, "RGB not supported in profile %d\n",

  451.                    ctx->profile);

  452.             return AVERROR_INVALIDDATA;

  453.         }

  454.     } else {

  455.         static const enum AVPixelFormat pix_fmt_for_ss[3][2 /* v */][2 /* h */] = {

  456.             { { AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV422P },

  457.               { AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV420P } },

  458.             { { AV_PIX_FMT_YUV444P10, AV_PIX_FMT_YUV422P10 },

  459.               { AV_PIX_FMT_YUV440P10, AV_PIX_FMT_YUV420P10 } },

  460.             { { AV_PIX_FMT_YUV444P12, AV_PIX_FMT_YUV422P12 },

  461.               { AV_PIX_FMT_YUV440P12, AV_PIX_FMT_YUV420P12 } }

  462.         };

  463.         ctx->color_range = get_bits1(&s->gb) ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;

  464.         if (ctx->profile & 1) {

  465.             s->ss_h = get_bits1(&s->gb);

  466.             s->ss_v = get_bits1(&s->gb);

  467.             s->pix_fmt = pix_fmt_for_ss[bits][s->ss_v][s->ss_h];

  468.             if (s->pix_fmt == AV_PIX_FMT_YUV420P) {

  469.                 av_log(ctx, AV_LOG_ERROR, "YUV 4:2:0 not supported in profile %d\n",

  470.                        ctx->profile);

  471.                 return AVERROR_INVALIDDATA;

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

  473.                 av_log(ctx, AV_LOG_ERROR, "Profile %d color details reserved bit set\n",

  474.                        ctx->profile);

  475.                 return AVERROR_INVALIDDATA;

  476.             }

  477.         } else {

  478.             s->ss_h = s->ss_v = 1;

  479.             s->pix_fmt = pix_fmt_for_ss[bits][1][1];

  480.         }

  481.     }

  482. 

  483.     return 0;

  484. }

  485. 

  486. static int decode_frame_header(AVCodecContext *ctx,

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

  488. {

  489.     VP9Context *s = ctx->priv_data;

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

  491.     int last_invisible;

  492.     const uint8_t *data2;

  493. 

  494.     /* general header */

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

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

  497.         return res;

  498.     }

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

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

  501.         return AVERROR_INVALIDDATA;

  502.     }

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

  504.     ctx->profile |= get_bits1(&s->gb) << 1;

  505.     if (ctx->profile == 3) ctx->profile += get_bits1(&s->gb);

  506.     if (ctx->profile > 3) {

  507.         av_log(ctx, AV_LOG_ERROR, "Profile %d is not yet supported\n", ctx->profile);

  508.         return AVERROR_INVALIDDATA;

  509.     }

  510.     s->s.h.profile = ctx->profile;

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

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

  513.         return 0;

  514.     }

  515.     s->last_keyframe  = s->s.h.keyframe;

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

  517.     last_invisible    = s->s.h.invisible;

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

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

  520.     s->s.h.use_last_frame_mvs = !s->s.h.errorres && !last_invisible;

  521.     if (s->s.h.keyframe) {

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

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

  524.             return AVERROR_INVALIDDATA;

  525.         }

  526.         if ((res = read_colorspace_details(ctx)) < 0)

  527.             return res;

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

  529.         s->s.h.refreshrefmask = 0xff;

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

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

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

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

  534.     } else {

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

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

  537.         if (s->s.h.intraonly) {

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

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

  540.                 return AVERROR_INVALIDDATA;

  541.             }

  542.             if (ctx->profile >= 1) {

  543.                 if ((res = read_colorspace_details(ctx)) < 0)

  544.                     return res;

  545.             } else {

  546.                 s->ss_h = s->ss_v = 1;

  547.                 s->bpp = 8;

  548.                 s->bpp_index = 0;

  549.                 s->bytesperpixel = 1;

  550.                 s->pix_fmt = AV_PIX_FMT_YUV420P;

  551.                 ctx->colorspace = AVCOL_SPC_BT470BG;

  552.                 ctx->color_range = AVCOL_RANGE_JPEG;

  553.             }

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

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

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

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

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

  559.         } else {

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

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

  562.             s->s.h.signbias[0]    = get_bits1(&s->gb) && !s->s.h.errorres;

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

  564.             s->s.h.signbias[1]    = get_bits1(&s->gb) && !s->s.h.errorres;

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

  566.             s->s.h.signbias[2]    = get_bits1(&s->gb) && !s->s.h.errorres;

  567.             if (!s->s.refs[s->s.h.refidx[0]].f->buf[0] ||

  568.                 !s->s.refs[s->s.h.refidx[1]].f->buf[0] ||

  569.                 !s->s.refs[s->s.h.refidx[2]].f->buf[0]) {

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

  571.                 return AVERROR_INVALIDDATA;

  572.             }

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

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

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

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

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

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

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

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

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

  582.             } else {

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

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

  585.             }

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

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

  588.             // the _last_ frame

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

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

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

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

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

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

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

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

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

  598.             if (s->s.h.allowcompinter) {

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

  600.                     s->s.h.fixcompref    = 2;

  601.                     s->s.h.varcompref[0] = 0;

  602.                     s->s.h.varcompref[1] = 1;

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

  604.                     s->s.h.fixcompref    = 1;

  605.                     s->s.h.varcompref[0] = 0;

  606.                     s->s.h.varcompref[1] = 2;

  607.                 } else {

  608.                     s->s.h.fixcompref    = 0;

  609.                     s->s.h.varcompref[0] = 1;

  610.                     s->s.h.varcompref[1] = 2;

  611.                 }

  612.             }

  613.         }

  614.     }

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

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

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

  618. 

  619.     /* loopfilter header data */

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

  621.         // reset loopfilter defaults

  622.         s->s.h.lf_delta.ref[0] = 1;

  623.         s->s.h.lf_delta.ref[1] = 0;

  624.         s->s.h.lf_delta.ref[2] = -1;

  625.         s->s.h.lf_delta.ref[3] = -1;

  626.         s->s.h.lf_delta.mode[0] = 0;

  627.         s->s.h.lf_delta.mode[1] = 0;

  628.         memset(s->s.h.segmentation.feat, 0, sizeof(s->s.h.segmentation.feat));

  629.     }

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

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

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

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

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

  635.         memset(s->filter_lut.lim_lut, 0, sizeof(s->filter_lut.lim_lut));

  636.     s->s.h.filter.sharpness = sharp;

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

  638.         if ((s->s.h.lf_delta.updated = get_bits1(&s->gb))) {

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

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

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

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

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

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

  645.         }

  646.     }

  647. 

  648.     /* quantization header data */

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

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

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

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

  653.     s->s.h.lossless    = s->s.h.yac_qi == 0 && s->s.h.ydc_qdelta == 0 &&

  654.                        s->s.h.uvdc_qdelta == 0 && s->s.h.uvac_qdelta == 0;

  655.     if (s->s.h.lossless)

  656.         ctx->properties |= FF_CODEC_PROPERTY_LOSSLESS;

  657. 

  658.     /* segmentation header info */

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

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

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

  662.                 s->s.h.segmentation.prob[i] = get_bits1(&s->gb) ?

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

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

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

  666.                     s->s.h.segmentation.pred_prob[i] = get_bits1(&s->gb) ?

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

  668.             }

  669.         }

  670. 

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

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

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

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

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

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

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

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

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

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

  681.             }

  682.         }

  683.     }

  684. 

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

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

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

  688. 

  689.         if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[i].q_enabled) {

  690.             if (s->s.h.segmentation.absolute_vals)

  691.                 qyac = av_clip_uintp2(s->s.h.segmentation.feat[i].q_val, 8);

  692.             else

  693.                 qyac = av_clip_uintp2(s->s.h.yac_qi + s->s.h.segmentation.feat[i].q_val, 8);

  694.         } else {

  695.             qyac  = s->s.h.yac_qi;

  696.         }

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

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

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

  700.         qyac  = av_clip_uintp2(qyac, 8);

  701. 

  702.         s->s.h.segmentation.feat[i].qmul[0][0] = vp9_dc_qlookup[s->bpp_index][qydc];

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

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

  705.         s->s.h.segmentation.feat[i].qmul[1][1] = vp9_ac_qlookup[s->bpp_index][quvac];

  706. 

  707.         sh = s->s.h.filter.level >= 32;

  708.         if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[i].lf_enabled) {

  709.             if (s->s.h.segmentation.absolute_vals)

  710.                 lflvl = av_clip_uintp2(s->s.h.segmentation.feat[i].lf_val, 6);

  711.             else

  712.                 lflvl = av_clip_uintp2(s->s.h.filter.level + s->s.h.segmentation.feat[i].lf_val, 6);

  713.         } else {

  714.             lflvl  = s->s.h.filter.level;

  715.         }

  716.         if (s->s.h.lf_delta.enabled) {

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

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

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

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

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

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

  723.                                              s->s.h.lf_delta.mode[0]) * (1 << sh)), 6);

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

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

  726.                                              s->s.h.lf_delta.mode[1]) * (1 << sh)), 6);

  727.             }

  728.         } else {

  729.             memset(s->s.h.segmentation.feat[i].lflvl, lflvl,

  730.                    sizeof(s->s.h.segmentation.feat[i].lflvl));

  731.         }

  732.     }

  733. 

  734.     /* tiling info */

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

  736.         av_log(ctx, AV_LOG_ERROR, "Failed to initialize decoder for %dx%d @ %d\n",

  737.                w, h, s->pix_fmt);

  738.         return res;

  739.     }

  740.     for (s->s.h.tiling.log2_tile_cols = 0;

  741.          s->sb_cols > (64 << s->s.h.tiling.log2_tile_cols);

  742.          s->s.h.tiling.log2_tile_cols++) ;

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

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

  745.     while (max > s->s.h.tiling.log2_tile_cols) {

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

  747.             s->s.h.tiling.log2_tile_cols++;

  748.         else

  749.             break;

  750.     }

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

  752.     s->s.h.tiling.tile_rows = 1 << s->s.h.tiling.log2_tile_rows;

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

  754.         s->s.h.tiling.tile_cols = 1 << s->s.h.tiling.log2_tile_cols;

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

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

  757.         if (!s->c_b) {

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

  759.             return AVERROR(ENOMEM);

  760.         }

  761.     }

  762. 

  763.     /* check reference frames */

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

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

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

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

  768. 

  769.             if (ref->format != ctx->pix_fmt) {

  770.                 av_log(ctx, AV_LOG_ERROR,

  771.                        "Ref pixfmt (%s) did not match current frame (%s)",

  772.                        av_get_pix_fmt_name(ref->format),

  773.                        av_get_pix_fmt_name(ctx->pix_fmt));

  774.                 return AVERROR_INVALIDDATA;

  775.             } else if (refw == w && refh == h) {

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

  777.             } else {

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

  779.                     av_log(ctx, AV_LOG_ERROR,

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

  781.                            refw, refh, w, h);

  782.                     return AVERROR_INVALIDDATA;

  783.                 }

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

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

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

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

  788.             }

  789.         }

  790.     }

  791. 

  792.     if (s->s.h.keyframe || s->s.h.errorres || (s->s.h.intraonly && s->s.h.resetctx == 3)) {

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

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

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

  796.                sizeof(vp9_default_coef_probs));

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

  798.                sizeof(vp9_default_coef_probs));

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

  800.                sizeof(vp9_default_coef_probs));

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

  802.                sizeof(vp9_default_coef_probs));

  803.     } else if (s->s.h.intraonly && s->s.h.resetctx == 2) {

  804.         s->prob_ctx[c].p = vp9_default_probs;

  805.         memcpy(s->prob_ctx[c].coef, vp9_default_coef_probs,

  806.                sizeof(vp9_default_coef_probs));

  807.     }

  808. 

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

  810.     s->s.h.compressed_header_size = size2 = get_bits(&s->gb, 16);

  811.     s->s.h.uncompressed_header_size = (get_bits_count(&s->gb) + 7) / 8;

  812. 

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

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

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

  816.         return AVERROR_INVALIDDATA;

  817.     }

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

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

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

  821.         return AVERROR_INVALIDDATA;

  822.     }

  823. 

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

  825.         memset(s->counts.coef, 0, sizeof(s->counts.coef));

  826.         memset(s->counts.eob,  0, sizeof(s->counts.eob));

  827.     } else {

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

  829.     }

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

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

  832.     // fw update)?

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

  834. 

  835.     // txfm updates

  836.     if (s->s.h.lossless) {

  837.         s->s.h.txfmmode = TX_4X4;

  838.     } else {

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

  840.         if (s->s.h.txfmmode == 3)

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

  842. 

  843.         if (s->s.h.txfmmode == TX_SWITCHABLE) {

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

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

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

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

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

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

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

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

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

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

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

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

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

  857.         }

  858.     }

  859. 

  860.     // coef updates

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

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

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

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

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

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

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

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

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

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

  871.                                 break;

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

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

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

  875.                                 } else {

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

  877.                                 }

  878.                             }

  879.                             p[3] = 0;

  880.                         }

  881.         } else {

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

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

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

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

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

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

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

  889.                                 break;

  890.                             memcpy(p, r, 3);

  891.                             p[3] = 0;

  892.                         }

  893.         }

  894.         if (s->s.h.txfmmode == i)

  895.             break;

  896.     }

  897. 

  898.     // mode updates

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

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

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

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

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

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

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

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

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

  908. 

  909.         if (s->s.h.filtermode == FILTER_SWITCHABLE)

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

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

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

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

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

  915. 

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

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

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

  919. 

  920.         if (s->s.h.allowcompinter) {

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

  922.             if (s->s.h.comppredmode)

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

  924.             if (s->s.h.comppredmode == PRED_SWITCHABLE)

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

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

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

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

  929.         } else {

  930.             s->s.h.comppredmode = PRED_SINGLEREF;

  931.         }

  932. 

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

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

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

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

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

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

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

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

  941.             }

  942.         }

  943. 

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

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

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

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

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

  949.         }

  950. 

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

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

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

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

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

  956. 

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

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

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

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

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

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

  963. 

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

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

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

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

  968. 

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

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

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

  972. 

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

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

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

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

  977. 

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

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

  980. 

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

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

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

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

  985.         }

  986. 

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

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

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

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

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

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

  993. 

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

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

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

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

  998.         }

  999. 

  1000.         if (s->s.h.highprecisionmvs) {

  1001.             for (i = 0; i < 2; i++) {

  1002.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  1003.                     s->prob.p.mv_comp[i].class0_hp =

  1004.                         (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  1005. 

  1006.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  1007.                     s->prob.p.mv_comp[i].hp =

  1008.                         (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  1009.             }

  1010.         }

  1011.     }

  1012. 

  1013.     return (data2 - data) + size2;

  1014. }

  1015. 

  1016. static av_always_inline void clamp_mv(VP56mv *dst, const VP56mv *src,

  1017.                                       VP9Context *s)

  1018. {

  1019.     dst->x = av_clip(src->x, s->min_mv.x, s->max_mv.x);

  1020.     dst->y = av_clip(src->y, s->min_mv.y, s->max_mv.y);

  1021. }

  1022. 

  1023. static void find_ref_mvs(VP9Context *s,

  1024.                          VP56mv *pmv, int ref, int z, int idx, int sb)

  1025. {

  1026.     static const int8_t mv_ref_blk_off[N_BS_SIZES][8][2] = {

  1027.         [BS_64x64] = {{  3, -1 }, { -1,  3 }, {  4, -1 }, { -1,  4 },

  1028.                       { -1, -1 }, {  0, -1 }, { -1,  0 }, {  6, -1 }},

  1029.         [BS_64x32] = {{  0, -1 }, { -1,  0 }, {  4, -1 }, { -1,  2 },

  1030.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, {  2, -1 }},

  1031.         [BS_32x64] = {{ -1,  0 }, {  0, -1 }, { -1,  4 }, {  2, -1 },

  1032.                       { -1, -1 }, { -3,  0 }, {  0, -3 }, { -1,  2 }},

  1033.         [BS_32x32] = {{  1, -1 }, { -1,  1 }, {  2, -1 }, { -1,  2 },

  1034.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

  1035.         [BS_32x16] = {{  0, -1 }, { -1,  0 }, {  2, -1 }, { -1, -1 },

  1036.                       { -1,  1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

  1037.         [BS_16x32] = {{ -1,  0 }, {  0, -1 }, { -1,  2 }, { -1, -1 },

  1038.                       {  1, -1 }, { -3,  0 }, {  0, -3 }, { -3, -3 }},

  1039.         [BS_16x16] = {{  0, -1 }, { -1,  0 }, {  1, -1 }, { -1,  1 },

  1040.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

  1041.         [BS_16x8]  = {{  0, -1 }, { -1,  0 }, {  1, -1 }, { -1, -1 },

  1042.                       {  0, -2 }, { -2,  0 }, { -2, -1 }, { -1, -2 }},

  1043.         [BS_8x16]  = {{ -1,  0 }, {  0, -1 }, { -1,  1 }, { -1, -1 },

  1044.                       { -2,  0 }, {  0, -2 }, { -1, -2 }, { -2, -1 }},

  1045.         [BS_8x8]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1046.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1047.         [BS_8x4]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1048.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1049.         [BS_4x8]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1050.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1051.         [BS_4x4]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1052.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1053.     };

  1054.     VP9Block *b = s->b;

  1055.     int row = s->row, col = s->col, row7 = s->row7;

  1056.     const int8_t (*p)[2] = mv_ref_blk_off[b->bs];

  1057. #define INVALID_MV 0x80008000U

  1058.     uint32_t mem = INVALID_MV, mem_sub8x8 = INVALID_MV;

  1059.     int i;

  1060. 

  1061. #define RETURN_DIRECT_MV(mv) \

  1062.     do { \

  1063.         uint32_t m = AV_RN32A(&mv); \

  1064.         if (!idx) { \

  1065.             AV_WN32A(pmv, m); \

  1066.             return; \

  1067.         } else if (mem == INVALID_MV) { \

  1068.             mem = m; \

  1069.         } else if (m != mem) { \

  1070.             AV_WN32A(pmv, m); \

  1071.             return; \

  1072.         } \

  1073.     } while (0)

  1074. 

  1075.     if (sb >= 0) {

  1076.         if (sb == 2 || sb == 1) {

  1077.             RETURN_DIRECT_MV(b->mv[0][z]);

  1078.         } else if (sb == 3) {

  1079.             RETURN_DIRECT_MV(b->mv[2][z]);

  1080.             RETURN_DIRECT_MV(b->mv[1][z]);

  1081.             RETURN_DIRECT_MV(b->mv[0][z]);

  1082.         }

  1083. 

  1084. #define RETURN_MV(mv) \

  1085.     do { \

  1086.         if (sb > 0) { \

  1087.             VP56mv tmp; \

  1088.             uint32_t m; \

  1089.             av_assert2(idx == 1); \

  1090.             av_assert2(mem != INVALID_MV); \

  1091.             if (mem_sub8x8 == INVALID_MV) { \

  1092.                 clamp_mv(&tmp, &mv, s); \

  1093.                 m = AV_RN32A(&tmp); \

  1094.                 if (m != mem) { \

  1095.                     AV_WN32A(pmv, m); \

  1096.                     return; \

  1097.                 } \

  1098.                 mem_sub8x8 = AV_RN32A(&mv); \

  1099.             } else if (mem_sub8x8 != AV_RN32A(&mv)) { \

  1100.                 clamp_mv(&tmp, &mv, s); \

  1101.                 m = AV_RN32A(&tmp); \

  1102.                 if (m != mem) { \

  1103.                     AV_WN32A(pmv, m); \

  1104.                 } else { \

  1105.                     /* BUG I'm pretty sure this isn't the intention */ \

  1106.                     AV_WN32A(pmv, 0); \

  1107.                 } \

  1108.                 return; \

  1109.             } \

  1110.         } else { \

  1111.             uint32_t m = AV_RN32A(&mv); \

  1112.             if (!idx) { \

  1113.                 clamp_mv(pmv, &mv, s); \

  1114.                 return; \

  1115.             } else if (mem == INVALID_MV) { \

  1116.                 mem = m; \

  1117.             } else if (m != mem) { \

  1118.                 clamp_mv(pmv, &mv, s); \

  1119.                 return; \

  1120.             } \

  1121.         } \

  1122.     } while (0)

  1123. 

  1124.         if (row > 0) {

  1125.             struct VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[(row - 1) * s->sb_cols * 8 + col];

  1126.             if (mv->ref[0] == ref) {

  1127.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][0]);

  1128.             } else if (mv->ref[1] == ref) {

  1129.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][1]);

  1130.             }

  1131.         }

  1132.         if (col > s->tile_col_start) {

  1133.             struct VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[row * s->sb_cols * 8 + col - 1];

  1134.             if (mv->ref[0] == ref) {

  1135.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][0]);

  1136.             } else if (mv->ref[1] == ref) {

  1137.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][1]);

  1138.             }

  1139.         }

  1140.         i = 2;

  1141.     } else {

  1142.         i = 0;

  1143.     }

  1144. 

  1145.     // previously coded MVs in this neighbourhood, using same reference frame

  1146.     for (; i < 8; i++) {

  1147.         int c = p[i][0] + col, r = p[i][1] + row;

  1148. 

  1149.         if (c >= s->tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1150.             struct VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1151. 

  1152.             if (mv->ref[0] == ref) {

  1153.                 RETURN_MV(mv->mv[0]);

  1154.             } else if (mv->ref[1] == ref) {

  1155.                 RETURN_MV(mv->mv[1]);

  1156.             }

  1157.         }

  1158.     }

  1159. 

  1160.     // MV at this position in previous frame, using same reference frame

  1161.     if (s->s.h.use_last_frame_mvs) {

  1162.         struct VP9mvrefPair *mv = &s->s.frames[REF_FRAME_MVPAIR].mv[row * s->sb_cols * 8 + col];

  1163. 

  1164.         if (!s->s.frames[REF_FRAME_MVPAIR].uses_2pass)

  1165.             ff_thread_await_progress(&s->s.frames[REF_FRAME_MVPAIR].tf, row >> 3, 0);

  1166.         if (mv->ref[0] == ref) {

  1167.             RETURN_MV(mv->mv[0]);

  1168.         } else if (mv->ref[1] == ref) {

  1169.             RETURN_MV(mv->mv[1]);

  1170.         }

  1171.     }

  1172. 

  1173. #define RETURN_SCALE_MV(mv, scale) \

  1174.     do { \

  1175.         if (scale) { \

  1176.             VP56mv mv_temp = { -mv.x, -mv.y }; \

  1177.             RETURN_MV(mv_temp); \

  1178.         } else { \

  1179.             RETURN_MV(mv); \

  1180.         } \

  1181.     } while (0)

  1182. 

  1183.     // previously coded MVs in this neighbourhood, using different reference frame

  1184.     for (i = 0; i < 8; i++) {

  1185.         int c = p[i][0] + col, r = p[i][1] + row;

  1186. 

  1187.         if (c >= s->tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1188.             struct VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1189. 

  1190.             if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1191.                 RETURN_SCALE_MV(mv->mv[0], s->s.h.signbias[mv->ref[0]] != s->s.h.signbias[ref]);

  1192.             }

  1193.             if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1194.                 // BUG - libvpx has this condition regardless of whether

  1195.                 // we used the first ref MV and pre-scaling

  1196.                 AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1197.                 RETURN_SCALE_MV(mv->mv[1], s->s.h.signbias[mv->ref[1]] != s->s.h.signbias[ref]);

  1198.             }

  1199.         }

  1200.     }

  1201. 

  1202.     // MV at this position in previous frame, using different reference frame

  1203.     if (s->s.h.use_last_frame_mvs) {

  1204.         struct VP9mvrefPair *mv = &s->s.frames[REF_FRAME_MVPAIR].mv[row * s->sb_cols * 8 + col];

  1205. 

  1206.         // no need to await_progress, because we already did that above

  1207.         if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1208.             RETURN_SCALE_MV(mv->mv[0], s->s.h.signbias[mv->ref[0]] != s->s.h.signbias[ref]);

  1209.         }

  1210.         if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1211.             // BUG - libvpx has this condition regardless of whether

  1212.             // we used the first ref MV and pre-scaling

  1213.             AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1214.             RETURN_SCALE_MV(mv->mv[1], s->s.h.signbias[mv->ref[1]] != s->s.h.signbias[ref]);

  1215.         }

  1216.     }

  1217. 

  1218.     AV_ZERO32(pmv);

  1219.     clamp_mv(pmv, pmv, s);

  1220. #undef INVALID_MV

  1221. #undef RETURN_MV

  1222. #undef RETURN_SCALE_MV

  1223. }

  1224. 

  1225. static av_always_inline int read_mv_component(VP9Context *s, int idx, int hp)

  1226. {

  1227.     int bit, sign = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].sign);

  1228.     int n, c = vp8_rac_get_tree(&s->c, vp9_mv_class_tree,

  1229.                                 s->prob.p.mv_comp[idx].classes);

  1230. 

  1231.     s->counts.mv_comp[idx].sign[sign]++;

  1232.     s->counts.mv_comp[idx].classes[c]++;

  1233.     if (c) {

  1234.         int m;

  1235. 

  1236.         for (n = 0, m = 0; m < c; m++) {

  1237.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].bits[m]);

  1238.             n |= bit << m;

  1239.             s->counts.mv_comp[idx].bits[m][bit]++;

  1240.         }

  1241.         n <<= 3;

  1242.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree, s->prob.p.mv_comp[idx].fp);

  1243.         n |= bit << 1;

  1244.         s->counts.mv_comp[idx].fp[bit]++;

  1245.         if (hp) {

  1246.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].hp);

  1247.             s->counts.mv_comp[idx].hp[bit]++;

  1248.             n |= bit;

  1249.         } else {

  1250.             n |= 1;

  1251.             // bug in libvpx - we count for bw entropy purposes even if the

  1252.             // bit wasn't coded

  1253.             s->counts.mv_comp[idx].hp[1]++;

  1254.         }

  1255.         n += 8 << c;

  1256.     } else {

  1257.         n = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0);

  1258.         s->counts.mv_comp[idx].class0[n]++;

  1259.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree,

  1260.                                s->prob.p.mv_comp[idx].class0_fp[n]);

  1261.         s->counts.mv_comp[idx].class0_fp[n][bit]++;

  1262.         n = (n << 3) | (bit << 1);

  1263.         if (hp) {

  1264.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0_hp);

  1265.             s->counts.mv_comp[idx].class0_hp[bit]++;

  1266.             n |= bit;

  1267.         } else {

  1268.             n |= 1;

  1269.             // bug in libvpx - we count for bw entropy purposes even if the

  1270.             // bit wasn't coded

  1271.             s->counts.mv_comp[idx].class0_hp[1]++;

  1272.         }

  1273.     }

  1274. 

  1275.     return sign ? -(n + 1) : (n + 1);

  1276. }

  1277. 

  1278. static void fill_mv(VP9Context *s,

  1279.                     VP56mv *mv, int mode, int sb)

  1280. {

  1281.     VP9Block *b = s->b;

  1282. 

  1283.     if (mode == ZEROMV) {

  1284.         AV_ZERO64(mv);

  1285.     } else {

  1286.         int hp;

  1287. 

  1288.         // FIXME cache this value and reuse for other subblocks

  1289.         find_ref_mvs(s, &mv[0], b->ref[0], 0, mode == NEARMV,

  1290.                      mode == NEWMV ? -1 : sb);

  1291.         // FIXME maybe move this code into find_ref_mvs()

  1292.         if ((mode == NEWMV || sb == -1) &&

  1293.             !(hp = s->s.h.highprecisionmvs && abs(mv[0].x) < 64 && abs(mv[0].y) < 64)) {

  1294.             if (mv[0].y & 1) {

  1295.                 if (mv[0].y < 0)

  1296.                     mv[0].y++;

  1297.                 else

  1298.                     mv[0].y--;

  1299.             }

  1300.             if (mv[0].x & 1) {

  1301.                 if (mv[0].x < 0)

  1302.                     mv[0].x++;

  1303.                 else

  1304.                     mv[0].x--;

  1305.             }

  1306.         }

  1307.         if (mode == NEWMV) {

  1308.             enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1309.                                               s->prob.p.mv_joint);

  1310. 

  1311.             s->counts.mv_joint[j]++;

  1312.             if (j >= MV_JOINT_V)

  1313.                 mv[0].y += read_mv_component(s, 0, hp);

  1314.             if (j & 1)

  1315.                 mv[0].x += read_mv_component(s, 1, hp);

  1316.         }

  1317. 

  1318.         if (b->comp) {

  1319.             // FIXME cache this value and reuse for other subblocks

  1320.             find_ref_mvs(s, &mv[1], b->ref[1], 1, mode == NEARMV,

  1321.                          mode == NEWMV ? -1 : sb);

  1322.             if ((mode == NEWMV || sb == -1) &&

  1323.                 !(hp = s->s.h.highprecisionmvs && abs(mv[1].x) < 64 && abs(mv[1].y) < 64)) {

  1324.                 if (mv[1].y & 1) {

  1325.                     if (mv[1].y < 0)

  1326.                         mv[1].y++;

  1327.                     else

  1328.                         mv[1].y--;

  1329.                 }

  1330.                 if (mv[1].x & 1) {

  1331.                     if (mv[1].x < 0)

  1332.                         mv[1].x++;

  1333.                     else

  1334.                         mv[1].x--;

  1335.                 }

  1336.             }

  1337.             if (mode == NEWMV) {

  1338.                 enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1339.                                                   s->prob.p.mv_joint);

  1340. 

  1341.                 s->counts.mv_joint[j]++;

  1342.                 if (j >= MV_JOINT_V)

  1343.                     mv[1].y += read_mv_component(s, 0, hp);

  1344.                 if (j & 1)

  1345.                     mv[1].x += read_mv_component(s, 1, hp);

  1346.             }

  1347.         }

  1348.     }

  1349. }

  1350. 

  1351. static av_always_inline void setctx_2d(uint8_t *ptr, int w, int h,

  1352.                                        ptrdiff_t stride, int v)

  1353. {

  1354.     switch (w) {

  1355.     case 1:

  1356.         do {

  1357.             *ptr = v;

  1358.             ptr += stride;

  1359.         } while (--h);

  1360.         break;

  1361.     case 2: {

  1362.         int v16 = v * 0x0101;

  1363.         do {

  1364.             AV_WN16A(ptr, v16);

  1365.             ptr += stride;

  1366.         } while (--h);

  1367.         break;

  1368.     }

  1369.     case 4: {

  1370.         uint32_t v32 = v * 0x01010101;

  1371.         do {

  1372.             AV_WN32A(ptr, v32);

  1373.             ptr += stride;

  1374.         } while (--h);

  1375.         break;

  1376.     }

  1377.     case 8: {

  1378. #if HAVE_FAST_64BIT

  1379.         uint64_t v64 = v * 0x0101010101010101ULL;

  1380.         do {

  1381.             AV_WN64A(ptr, v64);

  1382.             ptr += stride;

  1383.         } while (--h);

  1384. #else

  1385.         uint32_t v32 = v * 0x01010101;

  1386.         do {

  1387.             AV_WN32A(ptr,     v32);

  1388.             AV_WN32A(ptr + 4, v32);

  1389.             ptr += stride;

  1390.         } while (--h);

  1391. #endif

  1392.         break;

  1393.     }

  1394.     }

  1395. }

  1396. 

  1397. static void decode_mode(AVCodecContext *ctx)

  1398. {

  1399.     static const uint8_t left_ctx[N_BS_SIZES] = {

  1400.         0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf

  1401.     };

  1402.     static const uint8_t above_ctx[N_BS_SIZES] = {

  1403.         0x0, 0x0, 0x8, 0x8, 0x8, 0xc, 0xc, 0xc, 0xe, 0xe, 0xe, 0xf, 0xf

  1404.     };

  1405.     static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {

  1406.         TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,

  1407.         TX_16X16, TX_8X8, TX_8X8, TX_8X8, TX_4X4, TX_4X4, TX_4X4

  1408.     };

  1409.     VP9Context *s = ctx->priv_data;

  1410.     VP9Block *b = s->b;

  1411.     int row = s->row, col = s->col, row7 = s->row7;

  1412.     enum TxfmMode max_tx = max_tx_for_bl_bp[b->bs];

  1413.     int bw4 = bwh_tab[1][b->bs][0], w4 = FFMIN(s->cols - col, bw4);

  1414.     int bh4 = bwh_tab[1][b->bs][1], h4 = FFMIN(s->rows - row, bh4), y;

  1415.     int have_a = row > 0, have_l = col > s->tile_col_start;

  1416.     int vref, filter_id;

  1417. 

  1418.     if (!s->s.h.segmentation.enabled) {

  1419.         b->seg_id = 0;

  1420.     } else if (s->s.h.keyframe || s->s.h.intraonly) {

  1421.         b->seg_id = !s->s.h.segmentation.update_map ? 0 :

  1422.                     vp8_rac_get_tree(&s->c, vp9_segmentation_tree, s->s.h.segmentation.prob);

  1423.     } else if (!s->s.h.segmentation.update_map ||

  1424.                (s->s.h.segmentation.temporal &&

  1425.                 vp56_rac_get_prob_branchy(&s->c,

  1426.                     s->s.h.segmentation.pred_prob[s->above_segpred_ctx[col] +

  1427.                                     s->left_segpred_ctx[row7]]))) {

  1428.         if (!s->s.h.errorres && s->s.frames[REF_FRAME_SEGMAP].segmentation_map) {

  1429.             int pred = 8, x;

  1430.             uint8_t *refsegmap = s->s.frames[REF_FRAME_SEGMAP].segmentation_map;

  1431. 

  1432.             if (!s->s.frames[REF_FRAME_SEGMAP].uses_2pass)

  1433.                 ff_thread_await_progress(&s->s.frames[REF_FRAME_SEGMAP].tf, row >> 3, 0);

  1434.             for (y = 0; y < h4; y++) {

  1435.                 int idx_base = (y + row) * 8 * s->sb_cols + col;

  1436.                 for (x = 0; x < w4; x++)

  1437.                     pred = FFMIN(pred, refsegmap[idx_base + x]);

  1438.             }

  1439.             av_assert1(pred < 8);

  1440.             b->seg_id = pred;

  1441.         } else {

  1442.             b->seg_id = 0;

  1443.         }

  1444. 

  1445.         memset(&s->above_segpred_ctx[col], 1, w4);

  1446.         memset(&s->left_segpred_ctx[row7], 1, h4);

  1447.     } else {

  1448.         b->seg_id = vp8_rac_get_tree(&s->c, vp9_segmentation_tree,

  1449.                                      s->s.h.segmentation.prob);

  1450. 

  1451.         memset(&s->above_segpred_ctx[col], 0, w4);

  1452.         memset(&s->left_segpred_ctx[row7], 0, h4);

  1453.     }

  1454.     if (s->s.h.segmentation.enabled &&

  1455.         (s->s.h.segmentation.update_map || s->s.h.keyframe || s->s.h.intraonly)) {

  1456.         setctx_2d(&s->s.frames[CUR_FRAME].segmentation_map[row * 8 * s->sb_cols + col],

  1457.                   bw4, bh4, 8 * s->sb_cols, b->seg_id);

  1458.     }

  1459. 

  1460.     b->skip = s->s.h.segmentation.enabled &&

  1461.         s->s.h.segmentation.feat[b->seg_id].skip_enabled;

  1462.     if (!b->skip) {

  1463.         int c = s->left_skip_ctx[row7] + s->above_skip_ctx[col];

  1464.         b->skip = vp56_rac_get_prob(&s->c, s->prob.p.skip[c]);

  1465.         s->counts.skip[c][b->skip]++;

  1466.     }

  1467. 

  1468.     if (s->s.h.keyframe || s->s.h.intraonly) {

  1469.         b->intra = 1;

  1470.     } else if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[b->seg_id].ref_enabled) {

  1471.         b->intra = !s->s.h.segmentation.feat[b->seg_id].ref_val;

  1472.     } else {

  1473.         int c, bit;

  1474. 

  1475.         if (have_a && have_l) {

  1476.             c = s->above_intra_ctx[col] + s->left_intra_ctx[row7];

  1477.             c += (c == 2);

  1478.         } else {

  1479.             c = have_a ? 2 * s->above_intra_ctx[col] :

  1480.                 have_l ? 2 * s->left_intra_ctx[row7] : 0;

  1481.         }

  1482.         bit = vp56_rac_get_prob(&s->c, s->prob.p.intra[c]);

  1483.         s->counts.intra[c][bit]++;

  1484.         b->intra = !bit;

  1485.     }

  1486. 

  1487.     if ((b->intra || !b->skip) && s->s.h.txfmmode == TX_SWITCHABLE) {

  1488.         int c;

  1489.         if (have_a) {

  1490.             if (have_l) {

  1491.                 c = (s->above_skip_ctx[col] ? max_tx :

  1492.                      s->above_txfm_ctx[col]) +

  1493.                     (s->left_skip_ctx[row7] ? max_tx :

  1494.                      s->left_txfm_ctx[row7]) > max_tx;

  1495.             } else {

  1496.                 c = s->above_skip_ctx[col] ? 1 :

  1497.                     (s->above_txfm_ctx[col] * 2 > max_tx);

  1498.             }

  1499.         } else if (have_l) {

  1500.             c = s->left_skip_ctx[row7] ? 1 :

  1501.                 (s->left_txfm_ctx[row7] * 2 > max_tx);

  1502.         } else {

  1503.             c = 1;

  1504.         }

  1505.         switch (max_tx) {

  1506.         case TX_32X32:

  1507.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][0]);

  1508.             if (b->tx) {

  1509.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][1]);

  1510.                 if (b->tx == 2)

  1511.                     b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][2]);

  1512.             }

  1513.             s->counts.tx32p[c][b->tx]++;

  1514.             break;

  1515.         case TX_16X16:

  1516.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][0]);

  1517.             if (b->tx)

  1518.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][1]);

  1519.             s->counts.tx16p[c][b->tx]++;

  1520.             break;

  1521.         case TX_8X8:

  1522.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx8p[c]);

  1523.             s->counts.tx8p[c][b->tx]++;

  1524.             break;

  1525.         case TX_4X4:

  1526.             b->tx = TX_4X4;

  1527.             break;

  1528.         }

  1529.     } else {

  1530.         b->tx = FFMIN(max_tx, s->s.h.txfmmode);

  1531.     }

  1532. 

  1533.     if (s->s.h.keyframe || s->s.h.intraonly) {

  1534.         uint8_t *a = &s->above_mode_ctx[col * 2];

  1535.         uint8_t *l = &s->left_mode_ctx[(row7) << 1];

  1536. 

  1537.         b->comp = 0;

  1538.         if (b->bs > BS_8x8) {

  1539.             // FIXME the memory storage intermediates here aren't really

  1540.             // necessary, they're just there to make the code slightly

  1541.             // simpler for now

  1542.             b->mode[0] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1543.                                     vp9_default_kf_ymode_probs[a[0]][l[0]]);

  1544.             if (b->bs != BS_8x4) {

  1545.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1546.                                  vp9_default_kf_ymode_probs[a[1]][b->mode[0]]);

  1547.                 l[0] = a[1] = b->mode[1];

  1548.             } else {

  1549.                 l[0] = a[1] = b->mode[1] = b->mode[0];

  1550.             }

  1551.             if (b->bs != BS_4x8) {

  1552.                 b->mode[2] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1553.                                         vp9_default_kf_ymode_probs[a[0]][l[1]]);

  1554.                 if (b->bs != BS_8x4) {

  1555.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1556.                                   vp9_default_kf_ymode_probs[a[1]][b->mode[2]]);

  1557.                     l[1] = a[1] = b->mode[3];

  1558.                 } else {

  1559.                     l[1] = a[1] = b->mode[3] = b->mode[2];

  1560.                 }

  1561.             } else {

  1562.                 b->mode[2] = b->mode[0];

  1563.                 l[1] = a[1] = b->mode[3] = b->mode[1];

  1564.             }

  1565.         } else {

  1566.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1567.                                           vp9_default_kf_ymode_probs[*a][*l]);

  1568.             b->mode[3] = b->mode[2] = b->mode[1] = b->mode[0];

  1569.             // FIXME this can probably be optimized

  1570.             memset(a, b->mode[0], bwh_tab[0][b->bs][0]);

  1571.             memset(l, b->mode[0], bwh_tab[0][b->bs][1]);

  1572.         }

  1573.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1574.                                      vp9_default_kf_uvmode_probs[b->mode[3]]);

  1575.     } else if (b->intra) {

  1576.         b->comp = 0;

  1577.         if (b->bs > BS_8x8) {

  1578.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1579.                                           s->prob.p.y_mode[0]);

  1580.             s->counts.y_mode[0][b->mode[0]]++;

  1581.             if (b->bs != BS_8x4) {

  1582.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1583.                                               s->prob.p.y_mode[0]);

  1584.                 s->counts.y_mode[0][b->mode[1]]++;

  1585.             } else {

  1586.                 b->mode[1] = b->mode[0];

  1587.             }

  1588.             if (b->bs != BS_4x8) {

  1589.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1590.                                               s->prob.p.y_mode[0]);

  1591.                 s->counts.y_mode[0][b->mode[2]]++;

  1592.                 if (b->bs != BS_8x4) {

  1593.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1594.                                                   s->prob.p.y_mode[0]);

  1595.                     s->counts.y_mode[0][b->mode[3]]++;

  1596.                 } else {

  1597.                     b->mode[3] = b->mode[2];

  1598.                 }

  1599.             } else {

  1600.                 b->mode[2] = b->mode[0];

  1601.                 b->mode[3] = b->mode[1];

  1602.             }

  1603.         } else {

  1604.             static const uint8_t size_group[10] = {

  1605.                 3, 3, 3, 3, 2, 2, 2, 1, 1, 1

  1606.             };

  1607.             int sz = size_group[b->bs];

  1608. 

  1609.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1610.                                           s->prob.p.y_mode[sz]);

  1611.             b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1612.             s->counts.y_mode[sz][b->mode[3]]++;

  1613.         }

  1614.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1615.                                      s->prob.p.uv_mode[b->mode[3]]);

  1616.         s->counts.uv_mode[b->mode[3]][b->uvmode]++;

  1617.     } else {

  1618.         static const uint8_t inter_mode_ctx_lut[14][14] = {

  1619.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1620.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1621.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1622.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1623.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1624.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1625.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1626.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1627.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1628.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1629.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1630.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1631.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },

  1632.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },

  1633.         };

  1634. 

  1635.         if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[b->seg_id].ref_enabled) {

  1636.             av_assert2(s->s.h.segmentation.feat[b->seg_id].ref_val != 0);

  1637.             b->comp = 0;

  1638.             b->ref[0] = s->s.h.segmentation.feat[b->seg_id].ref_val - 1;

  1639.         } else {

  1640.             // read comp_pred flag

  1641.             if (s->s.h.comppredmode != PRED_SWITCHABLE) {

  1642.                 b->comp = s->s.h.comppredmode == PRED_COMPREF;

  1643.             } else {

  1644.                 int c;

  1645. 

  1646.                 // FIXME add intra as ref=0xff (or -1) to make these easier?

  1647.                 if (have_a) {

  1648.                     if (have_l) {

  1649.                         if (s->above_comp_ctx[col] && s->left_comp_ctx[row7]) {

  1650.                             c = 4;

  1651.                         } else if (s->above_comp_ctx[col]) {

  1652.                             c = 2 + (s->left_intra_ctx[row7] ||

  1653.                                      s->left_ref_ctx[row7] == s->s.h.fixcompref);

  1654.                         } else if (s->left_comp_ctx[row7]) {

  1655.                             c = 2 + (s->above_intra_ctx[col] ||

  1656.                                      s->above_ref_ctx[col] == s->s.h.fixcompref);

  1657.                         } else {

  1658.                             c = (!s->above_intra_ctx[col] &&

  1659.                                  s->above_ref_ctx[col] == s->s.h.fixcompref) ^

  1660.                             (!s->left_intra_ctx[row7] &&

  1661.                              s->left_ref_ctx[row & 7] == s->s.h.fixcompref);

  1662.                         }

  1663.                     } else {

  1664.                         c = s->above_comp_ctx[col] ? 3 :

  1665.                         (!s->above_intra_ctx[col] && s->above_ref_ctx[col] == s->s.h.fixcompref);

  1666.                     }

  1667.                 } else if (have_l) {

  1668.                     c = s->left_comp_ctx[row7] ? 3 :

  1669.                     (!s->left_intra_ctx[row7] && s->left_ref_ctx[row7] == s->s.h.fixcompref);

  1670.                 } else {

  1671.                     c = 1;

  1672.                 }

  1673.                 b->comp = vp56_rac_get_prob(&s->c, s->prob.p.comp[c]);

  1674.                 s->counts.comp[c][b->comp]++;

  1675.             }

  1676. 

  1677.             // read actual references

  1678.             // FIXME probably cache a few variables here to prevent repetitive

  1679.             // memory accesses below

  1680.             if (b->comp) /* two references */ {

  1681.                 int fix_idx = s->s.h.signbias[s->s.h.fixcompref], var_idx = !fix_idx, c, bit;

  1682. 

  1683.                 b->ref[fix_idx] = s->s.h.fixcompref;

  1684.                 // FIXME can this codeblob be replaced by some sort of LUT?

  1685.                 if (have_a) {

  1686.                     if (have_l) {

  1687.                         if (s->above_intra_ctx[col]) {

  1688.                             if (s->left_intra_ctx[row7]) {

  1689.                                 c = 2;

  1690.                             } else {

  1691.                                 c = 1 + 2 * (s->left_ref_ctx[row7] != s->s.h.varcompref[1]);

  1692.                             }

  1693.                         } else if (s->left_intra_ctx[row7]) {

  1694.                             c = 1 + 2 * (s->above_ref_ctx[col] != s->s.h.varcompref[1]);

  1695.                         } else {

  1696.                             int refl = s->left_ref_ctx[row7], refa = s->above_ref_ctx[col];

  1697. 

  1698.                             if (refl == refa && refa == s->s.h.varcompref[1]) {

  1699.                                 c = 0;

  1700.                             } else if (!s->left_comp_ctx[row7] && !s->above_comp_ctx[col]) {

  1701.                                 if ((refa == s->s.h.fixcompref && refl == s->s.h.varcompref[0]) ||

  1702.                                     (refl == s->s.h.fixcompref && refa == s->s.h.varcompref[0])) {

  1703.                                     c = 4;

  1704.                                 } else {

  1705.                                     c = (refa == refl) ? 3 : 1;

  1706.                                 }

  1707.                             } else if (!s->left_comp_ctx[row7]) {

  1708.                                 if (refa == s->s.h.varcompref[1] && refl != s->s.h.varcompref[1]) {

  1709.                                     c = 1;

  1710.                                 } else {

  1711.                                     c = (refl == s->s.h.varcompref[1] &&

  1712.                                          refa != s->s.h.varcompref[1]) ? 2 : 4;

  1713.                                 }

  1714.                             } else if (!s->above_comp_ctx[col]) {

  1715.                                 if (refl == s->s.h.varcompref[1] && refa != s->s.h.varcompref[1]) {

  1716.                                     c = 1;

  1717.                                 } else {

  1718.                                     c = (refa == s->s.h.varcompref[1] &&

  1719.                                          refl != s->s.h.varcompref[1]) ? 2 : 4;

  1720.                                 }

  1721.                             } else {

  1722.                                 c = (refl == refa) ? 4 : 2;

  1723.                             }

  1724.                         }

  1725.                     } else {

  1726.                         if (s->above_intra_ctx[col]) {

  1727.                             c = 2;

  1728.                         } else if (s->above_comp_ctx[col]) {

  1729.                             c = 4 * (s->above_ref_ctx[col] != s->s.h.varcompref[1]);

  1730.                         } else {

  1731.                             c = 3 * (s->above_ref_ctx[col] != s->s.h.varcompref[1]);

  1732.                         }

  1733.                     }

  1734.                 } else if (have_l) {

  1735.                     if (s->left_intra_ctx[row7]) {

  1736.                         c = 2;

  1737.                     } else if (s->left_comp_ctx[row7]) {

  1738.                         c = 4 * (s->left_ref_ctx[row7] != s->s.h.varcompref[1]);

  1739.                     } else {

  1740.                         c = 3 * (s->left_ref_ctx[row7] != s->s.h.varcompref[1]);

  1741.                     }

  1742.                 } else {

  1743.                     c = 2;

  1744.                 }

  1745.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.comp_ref[c]);

  1746.                 b->ref[var_idx] = s->s.h.varcompref[bit];

  1747.                 s->counts.comp_ref[c][bit]++;

  1748.             } else /* single reference */ {

  1749.                 int bit, c;

  1750. 

  1751.                 if (have_a && !s->above_intra_ctx[col]) {

  1752.                     if (have_l && !s->left_intra_ctx[row7]) {

  1753.                         if (s->left_comp_ctx[row7]) {

  1754.                             if (s->above_comp_ctx[col]) {

  1755.                                 c = 1 + (!s->s.h.fixcompref || !s->left_ref_ctx[row7] ||

  1756.                                          !s->above_ref_ctx[col]);

  1757.                             } else {

  1758.                                 c = (3 * !s->above_ref_ctx[col]) +

  1759.                                     (!s->s.h.fixcompref || !s->left_ref_ctx[row7]);

  1760.                             }

  1761.                         } else if (s->above_comp_ctx[col]) {

  1762.                             c = (3 * !s->left_ref_ctx[row7]) +

  1763.                                 (!s->s.h.fixcompref || !s->above_ref_ctx[col]);

  1764.                         } else {

  1765.                             c = 2 * !s->left_ref_ctx[row7] + 2 * !s->above_ref_ctx[col];

  1766.                         }

  1767.                     } else if (s->above_intra_ctx[col]) {

  1768.                         c = 2;

  1769.                     } else if (s->above_comp_ctx[col]) {

  1770.                         c = 1 + (!s->s.h.fixcompref || !s->above_ref_ctx[col]);

  1771.                     } else {

  1772.                         c = 4 * (!s->above_ref_ctx[col]);

  1773.                     }

  1774.                 } else if (have_l && !s->left_intra_ctx[row7]) {

  1775.                     if (s->left_intra_ctx[row7]) {

  1776.                         c = 2;

  1777.                     } else if (s->left_comp_ctx[row7]) {

  1778.                         c = 1 + (!s->s.h.fixcompref || !s->left_ref_ctx[row7]);

  1779.                     } else {

  1780.                         c = 4 * (!s->left_ref_ctx[row7]);

  1781.                     }

  1782.                 } else {

  1783.                     c = 2;

  1784.                 }

  1785.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][0]);

  1786.                 s->counts.single_ref[c][0][bit]++;

  1787.                 if (!bit) {

  1788.                     b->ref[0] = 0;

  1789.                 } else {

  1790.                     // FIXME can this codeblob be replaced by some sort of LUT?

  1791.                     if (have_a) {

  1792.                         if (have_l) {

  1793.                             if (s->left_intra_ctx[row7]) {

  1794.                                 if (s->above_intra_ctx[col]) {

  1795.                                     c = 2;

  1796.                                 } else if (s->above_comp_ctx[col]) {

  1797.                                     c = 1 + 2 * (s->s.h.fixcompref == 1 ||

  1798.                                                  s->above_ref_ctx[col] == 1);

  1799.                                 } else if (!s->above_ref_ctx[col]) {

  1800.                                     c = 3;

  1801.                                 } else {

  1802.                                     c = 4 * (s->above_ref_ctx[col] == 1);

  1803.                                 }

  1804.                             } else if (s->above_intra_ctx[col]) {

  1805.                                 if (s->left_intra_ctx[row7]) {

  1806.                                     c = 2;

  1807.                                 } else if (s->left_comp_ctx[row7]) {

  1808.                                     c = 1 + 2 * (s->s.h.fixcompref == 1 ||

  1809.                                                  s->left_ref_ctx[row7] == 1);

  1810.                                 } else if (!s->left_ref_ctx[row7]) {

  1811.                                     c = 3;

  1812.                                 } else {

  1813.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1814.                                 }

  1815.                             } else if (s->above_comp_ctx[col]) {

  1816.                                 if (s->left_comp_ctx[row7]) {

  1817.                                     if (s->left_ref_ctx[row7] == s->above_ref_ctx[col]) {

  1818.                                         c = 3 * (s->s.h.fixcompref == 1 ||

  1819.                                                  s->left_ref_ctx[row7] == 1);

  1820.                                     } else {

  1821.                                         c = 2;

  1822.                                     }

  1823.                                 } else if (!s->left_ref_ctx[row7]) {

  1824.                                     c = 1 + 2 * (s->s.h.fixcompref == 1 ||

  1825.                                                  s->above_ref_ctx[col] == 1);

  1826.                                 } else {

  1827.                                     c = 3 * (s->left_ref_ctx[row7] == 1) +

  1828.                                     (s->s.h.fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1829.                                 }

  1830.                             } else if (s->left_comp_ctx[row7]) {

  1831.                                 if (!s->above_ref_ctx[col]) {

  1832.                                     c = 1 + 2 * (s->s.h.fixcompref == 1 ||

  1833.                                                  s->left_ref_ctx[row7] == 1);

  1834.                                 } else {

  1835.                                     c = 3 * (s->above_ref_ctx[col] == 1) +

  1836.                                     (s->s.h.fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1837.                                 }

  1838.                             } else if (!s->above_ref_ctx[col]) {

  1839.                                 if (!s->left_ref_ctx[row7]) {

  1840.                                     c = 3;

  1841.                                 } else {

  1842.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1843.                                 }

  1844.                             } else if (!s->left_ref_ctx[row7]) {

  1845.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1846.                             } else {

  1847.                                 c = 2 * (s->left_ref_ctx[row7] == 1) +

  1848.                                 2 * (s->above_ref_ctx[col] == 1);

  1849.                             }

  1850.                         } else {

  1851.                             if (s->above_intra_ctx[col] ||

  1852.                                 (!s->above_comp_ctx[col] && !s->above_ref_ctx[col])) {

  1853.                                 c = 2;

  1854.                             } else if (s->above_comp_ctx[col]) {

  1855.                                 c = 3 * (s->s.h.fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1856.                             } else {

  1857.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1858.                             }

  1859.                         }

  1860.                     } else if (have_l) {

  1861.                         if (s->left_intra_ctx[row7] ||

  1862.                             (!s->left_comp_ctx[row7] && !s->left_ref_ctx[row7])) {

  1863.                             c = 2;

  1864.                         } else if (s->left_comp_ctx[row7]) {

  1865.                             c = 3 * (s->s.h.fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1866.                         } else {

  1867.                             c = 4 * (s->left_ref_ctx[row7] == 1);

  1868.                         }

  1869.                     } else {

  1870.                         c = 2;

  1871.                     }

  1872.                     bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][1]);

  1873.                     s->counts.single_ref[c][1][bit]++;

  1874.                     b->ref[0] = 1 + bit;

  1875.                 }

  1876.             }

  1877.         }

  1878. 

  1879.         if (b->bs <= BS_8x8) {

  1880.             if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[b->seg_id].skip_enabled) {

  1881.                 b->mode[0] = b->mode[1] = b->mode[2] = b->mode[3] = ZEROMV;

  1882.             } else {

  1883.                 static const uint8_t off[10] = {

  1884.                     3, 0, 0, 1, 0, 0, 0, 0, 0, 0

  1885.                 };

  1886. 

  1887.                 // FIXME this needs to use the LUT tables from find_ref_mvs

  1888.                 // because not all are -1,0/0,-1

  1889.                 int c = inter_mode_ctx_lut[s->above_mode_ctx[col + off[b->bs]]]

  1890.                                           [s->left_mode_ctx[row7 + off[b->bs]]];

  1891. 

  1892.                 b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1893.                                               s->prob.p.mv_mode[c]);

  1894.                 b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1895.                 s->counts.mv_mode[c][b->mode[0] - 10]++;

  1896.             }

  1897.         }

  1898. 

  1899.         if (s->s.h.filtermode == FILTER_SWITCHABLE) {

  1900.             int c;

  1901. 

  1902.             if (have_a && s->above_mode_ctx[col] >= NEARESTMV) {

  1903.                 if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1904.                     c = s->above_filter_ctx[col] == s->left_filter_ctx[row7] ?

  1905.                         s->left_filter_ctx[row7] : 3;

  1906.                 } else {

  1907.                     c = s->above_filter_ctx[col];

  1908.                 }

  1909.             } else if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1910.                 c = s->left_filter_ctx[row7];

  1911.             } else {

  1912.                 c = 3;

  1913.             }

  1914. 

  1915.             filter_id = vp8_rac_get_tree(&s->c, vp9_filter_tree,

  1916.                                          s->prob.p.filter[c]);

  1917.             s->counts.filter[c][filter_id]++;

  1918.             b->filter = vp9_filter_lut[filter_id];

  1919.         } else {

  1920.             b->filter = s->s.h.filtermode;

  1921.         }

  1922. 

  1923.         if (b->bs > BS_8x8) {

  1924.             int c = inter_mode_ctx_lut[s->above_mode_ctx[col]][s->left_mode_ctx[row7]];

  1925. 

  1926.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1927.                                           s->prob.p.mv_mode[c]);

  1928.             s->counts.mv_mode[c][b->mode[0] - 10]++;

  1929.             fill_mv(s, b->mv[0], b->mode[0], 0);

  1930. 

  1931.             if (b->bs != BS_8x4) {

  1932.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1933.                                               s->prob.p.mv_mode[c]);

  1934.                 s->counts.mv_mode[c][b->mode[1] - 10]++;

  1935.                 fill_mv(s, b->mv[1], b->mode[1], 1);

  1936.             } else {

  1937.                 b->mode[1] = b->mode[0];

  1938.                 AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1939.                 AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1940.             }

  1941. 

  1942.             if (b->bs != BS_4x8) {

  1943.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1944.                                               s->prob.p.mv_mode[c]);

  1945.                 s->counts.mv_mode[c][b->mode[2] - 10]++;

  1946.                 fill_mv(s, b->mv[2], b->mode[2], 2);

  1947. 

  1948.                 if (b->bs != BS_8x4) {

  1949.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1950.                                                   s->prob.p.mv_mode[c]);

  1951.                     s->counts.mv_mode[c][b->mode[3] - 10]++;

  1952.                     fill_mv(s, b->mv[3], b->mode[3], 3);

  1953.                 } else {

  1954.                     b->mode[3] = b->mode[2];

  1955.                     AV_COPY32(&b->mv[3][0], &b->mv[2][0]);

  1956.                     AV_COPY32(&b->mv[3][1], &b->mv[2][1]);

  1957.                 }

  1958.             } else {

  1959.                 b->mode[2] = b->mode[0];

  1960.                 AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1961.                 AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1962.                 b->mode[3] = b->mode[1];

  1963.                 AV_COPY32(&b->mv[3][0], &b->mv[1][0]);

  1964.                 AV_COPY32(&b->mv[3][1], &b->mv[1][1]);

  1965.             }

  1966.         } else {

  1967.             fill_mv(s, b->mv[0], b->mode[0], -1);

  1968.             AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1969.             AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1970.             AV_COPY32(&b->mv[3][0], &b->mv[0][0]);

  1971.             AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1972.             AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1973.             AV_COPY32(&b->mv[3][1], &b->mv[0][1]);

  1974.         }

  1975. 

  1976.         vref = b->ref[b->comp ? s->s.h.signbias[s->s.h.varcompref[0]] : 0];

  1977.     }

  1978. 

  1979. #if HAVE_FAST_64BIT

  1980. #define SPLAT_CTX(var, val, n) \

  1981.     switch (n) { \

  1982.     case 1:  var = val;                                    break; \

  1983.     case 2:  AV_WN16A(&var, val *             0x0101);     break; \

  1984.     case 4:  AV_WN32A(&var, val *         0x01010101);     break; \

  1985.     case 8:  AV_WN64A(&var, val * 0x0101010101010101ULL);  break; \

  1986.     case 16: { \

  1987.         uint64_t v64 = val * 0x0101010101010101ULL; \

  1988.         AV_WN64A(              &var,     v64); \

  1989.         AV_WN64A(&((uint8_t *) &var)[8], v64); \

  1990.         break; \

  1991.     } \

  1992.     }

  1993. #else

  1994. #define SPLAT_CTX(var, val, n) \

  1995.     switch (n) { \

  1996.     case 1:  var = val;                         break; \

  1997.     case 2:  AV_WN16A(&var, val *     0x0101);  break; \

  1998.     case 4:  AV_WN32A(&var, val * 0x01010101);  break; \

  1999.     case 8: { \

  2000.         uint32_t v32 = val * 0x01010101; \

  2001.         AV_WN32A(              &var,     v32); \

  2002.         AV_WN32A(&((uint8_t *) &var)[4], v32); \

  2003.         break; \

  2004.     } \

  2005.     case 16: { \

  2006.         uint32_t v32 = val * 0x01010101; \

  2007.         AV_WN32A(              &var,      v32); \

  2008.         AV_WN32A(&((uint8_t *) &var)[4],  v32); \

  2009.         AV_WN32A(&((uint8_t *) &var)[8],  v32); \

  2010.         AV_WN32A(&((uint8_t *) &var)[12], v32); \

  2011.         break; \

  2012.     } \

  2013.     }

  2014. #endif

  2015. 

  2016.     switch (bwh_tab[1][b->bs][0]) {

  2017. #define SET_CTXS(dir, off, n) \

  2018.     do { \

  2019.         SPLAT_CTX(s->dir##_skip_ctx[off],      b->skip,          n); \

  2020.         SPLAT_CTX(s->dir##_txfm_ctx[off],      b->tx,            n); \

  2021.         SPLAT_CTX(s->dir##_partition_ctx[off], dir##_ctx[b->bs], n); \

  2022.         if (!s->s.h.keyframe && !s->s.h.intraonly) { \

  2023.             SPLAT_CTX(s->dir##_intra_ctx[off], b->intra,   n); \

  2024.             SPLAT_CTX(s->dir##_comp_ctx[off],  b->comp,    n); \

  2025.             SPLAT_CTX(s->dir##_mode_ctx[off],  b->mode[3], n); \

  2026.             if (!b->intra) { \

  2027.                 SPLAT_CTX(s->dir##_ref_ctx[off], vref, n); \

  2028.                 if (s->s.h.filtermode == FILTER_SWITCHABLE) { \

  2029.                     SPLAT_CTX(s->dir##_filter_ctx[off], filter_id, n); \

  2030.                 } \

  2031.             } \

  2032.         } \

  2033.     } while (0)

  2034.     case 1: SET_CTXS(above, col, 1); break;

  2035.     case 2: SET_CTXS(above, col, 2); break;

  2036.     case 4: SET_CTXS(above, col, 4); break;

  2037.     case 8: SET_CTXS(above, col, 8); break;

  2038.     }

  2039.     switch (bwh_tab[1][b->bs][1]) {

  2040.     case 1: SET_CTXS(left, row7, 1); break;

  2041.     case 2: SET_CTXS(left, row7, 2); break;

  2042.     case 4: SET_CTXS(left, row7, 4); break;

  2043.     case 8: SET_CTXS(left, row7, 8); break;

  2044.     }

  2045. #undef SPLAT_CTX

  2046. #undef SET_CTXS

  2047. 

  2048.     if (!s->s.h.keyframe && !s->s.h.intraonly) {

  2049.         if (b->bs > BS_8x8) {

  2050.             int mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  2051. 

  2052.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][0], &b->mv[1][0]);

  2053.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][1], &b->mv[1][1]);

  2054.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][0], mv0);

  2055.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][1], mv1);

  2056.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][0], &b->mv[2][0]);

  2057.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][1], &b->mv[2][1]);

  2058.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][0], mv0);

  2059.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][1], mv1);

  2060.         } else {

  2061.             int n, mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  2062. 

  2063.             for (n = 0; n < w4 * 2; n++) {

  2064.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][0], mv0);

  2065.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][1], mv1);

  2066.             }

  2067.             for (n = 0; n < h4 * 2; n++) {

  2068.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][0], mv0);

  2069.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][1], mv1);

  2070.             }

  2071.         }

  2072.     }

  2073. 

  2074.     // FIXME kinda ugly

  2075.     for (y = 0; y < h4; y++) {

  2076.         int x, o = (row + y) * s->sb_cols * 8 + col;

  2077.         struct VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[o];

  2078. 

  2079.         if (b->intra) {

  2080.             for (x = 0; x < w4; x++) {

  2081.                 mv[x].ref[0] =

  2082.                 mv[x].ref[1] = -1;

  2083.             }

  2084.         } else if (b->comp) {

  2085.             for (x = 0; x < w4; x++) {

  2086.                 mv[x].ref[0] = b->ref[0];

  2087.                 mv[x].ref[1] = b->ref[1];

  2088.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  2089.                 AV_COPY32(&mv[x].mv[1], &b->mv[3][1]);

  2090.             }

  2091.         } else {

  2092.             for (x = 0; x < w4; x++) {

  2093.                 mv[x].ref[0] = b->ref[0];

  2094.                 mv[x].ref[1] = -1;

  2095.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  2096.             }

  2097.         }

  2098.     }

  2099. }

  2100. 

  2101. // FIXME merge cnt/eob arguments?

  2102. static av_always_inline int

  2103. decode_coeffs_b_generic(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  2104.                         int is_tx32x32, int is8bitsperpixel, int bpp, unsigned (*cnt)[6][3],

  2105.                         unsigned (*eob)[6][2], uint8_t (*p)[6][11],

  2106.                         int nnz, const int16_t *scan, const int16_t (*nb)[2],

  2107.                         const int16_t *band_counts, const int16_t *qmul)

  2108. {

  2109.     int i = 0, band = 0, band_left = band_counts[band];

  2110.     uint8_t *tp = p[0][nnz];

  2111.     uint8_t cache[1024];

  2112. 

  2113.     do {

  2114.         int val, rc;

  2115. 

  2116.         val = vp56_rac_get_prob_branchy(c, tp[0]); // eob

  2117.         eob[band][nnz][val]++;

  2118.         if (!val)

  2119.             break;

  2120. 

  2121.     skip_eob:

  2122.         if (!vp56_rac_get_prob_branchy(c, tp[1])) { // zero

  2123.             cnt[band][nnz][0]++;

  2124.             if (!--band_left)

  2125.                 band_left = band_counts[++band];

  2126.             cache[scan[i]] = 0;

  2127.             nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  2128.             tp = p[band][nnz];

  2129.             if (++i == n_coeffs)

  2130.                 break; //invalid input; blocks should end with EOB

  2131.             goto skip_eob;

  2132.         }

  2133. 

  2134.         rc = scan[i];

  2135.         if (!vp56_rac_get_prob_branchy(c, tp[2])) { // one

  2136.             cnt[band][nnz][1]++;

  2137.             val = 1;

  2138.             cache[rc] = 1;

  2139.         } else {

  2140.             // fill in p[3-10] (model fill) - only once per frame for each pos

  2141.             if (!tp[3])

  2142.                 memcpy(&tp[3], vp9_model_pareto8[tp[2]], 8);

  2143. 

  2144.             cnt[band][nnz][2]++;

  2145.             if (!vp56_rac_get_prob_branchy(c, tp[3])) { // 2, 3, 4

  2146.                 if (!vp56_rac_get_prob_branchy(c, tp[4])) {

  2147.                     cache[rc] = val = 2;

  2148.                 } else {

  2149.                     val = 3 + vp56_rac_get_prob(c, tp[5]);

  2150.                     cache[rc] = 3;

  2151.                 }

  2152.             } else if (!vp56_rac_get_prob_branchy(c, tp[6])) { // cat1/2

  2153.                 cache[rc] = 4;

  2154.                 if (!vp56_rac_get_prob_branchy(c, tp[7])) {

  2155.                     val = 5 + vp56_rac_get_prob(c, 159);

  2156.                 } else {

  2157.                     val  = 7 + (vp56_rac_get_prob(c, 165) << 1);

  2158.                     val +=      vp56_rac_get_prob(c, 145);

  2159.                 }

  2160.             } else { // cat 3-6

  2161.                 cache[rc] = 5;

  2162.                 if (!vp56_rac_get_prob_branchy(c, tp[8])) {

  2163.                     if (!vp56_rac_get_prob_branchy(c, tp[9])) {

  2164.                         val  = 11 + (vp56_rac_get_prob(c, 173) << 2);

  2165.                         val +=      (vp56_rac_get_prob(c, 148) << 1);

  2166.                         val +=       vp56_rac_get_prob(c, 140);

  2167.                     } else {

  2168.                         val  = 19 + (vp56_rac_get_prob(c, 176) << 3);

  2169.                         val +=      (vp56_rac_get_prob(c, 155) << 2);

  2170.                         val +=      (vp56_rac_get_prob(c, 140) << 1);

  2171.                         val +=       vp56_rac_get_prob(c, 135);

  2172.                     }

  2173.                 } else if (!vp56_rac_get_prob_branchy(c, tp[10])) {

  2174.                     val  = 35 + (vp56_rac_get_prob(c, 180) << 4);

  2175.                     val +=      (vp56_rac_get_prob(c, 157) << 3);

  2176.                     val +=      (vp56_rac_get_prob(c, 141) << 2);

  2177.                     val +=      (vp56_rac_get_prob(c, 134) << 1);

  2178.                     val +=       vp56_rac_get_prob(c, 130);

  2179.                 } else {

  2180.                     val = 67;

  2181.                     if (!is8bitsperpixel) {

  2182.                         if (bpp == 12) {

  2183.                             val += vp56_rac_get_prob(c, 255) << 17;

  2184.                             val += vp56_rac_get_prob(c, 255) << 16;

  2185.                         }

  2186.                         val +=  (vp56_rac_get_prob(c, 255) << 15);

  2187.                         val +=  (vp56_rac_get_prob(c, 255) << 14);

  2188.                     }

  2189.                     val +=      (vp56_rac_get_prob(c, 254) << 13);

  2190.                     val +=      (vp56_rac_get_prob(c, 254) << 12);

  2191.                     val +=      (vp56_rac_get_prob(c, 254) << 11);

  2192.                     val +=      (vp56_rac_get_prob(c, 252) << 10);

  2193.                     val +=      (vp56_rac_get_prob(c, 249) << 9);

  2194.                     val +=      (vp56_rac_get_prob(c, 243) << 8);

  2195.                     val +=      (vp56_rac_get_prob(c, 230) << 7);

  2196.                     val +=      (vp56_rac_get_prob(c, 196) << 6);

  2197.                     val +=      (vp56_rac_get_prob(c, 177) << 5);

  2198.                     val +=      (vp56_rac_get_prob(c, 153) << 4);

  2199.                     val +=      (vp56_rac_get_prob(c, 140) << 3);

  2200.                     val +=      (vp56_rac_get_prob(c, 133) << 2);

  2201.                     val +=      (vp56_rac_get_prob(c, 130) << 1);

  2202.                     val +=       vp56_rac_get_prob(c, 129);

  2203.                 }

  2204.             }

  2205.         }

  2206. #define STORE_COEF(c, i, v) do { \

  2207.     if (is8bitsperpixel) { \

  2208.         c[i] = v; \

  2209.     } else { \

  2210.         AV_WN32A(&c[i * 2], v); \

  2211.     } \

  2212. } while (0)

  2213.         if (!--band_left)

  2214.             band_left = band_counts[++band];

  2215.         if (is_tx32x32)

  2216.             STORE_COEF(coef, rc, ((vp8_rac_get(c) ? -val : val) * qmul[!!i]) / 2);

  2217.         else

  2218.             STORE_COEF(coef, rc, (vp8_rac_get(c) ? -val : val) * qmul[!!i]);

  2219.         nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  2220.         tp = p[band][nnz];

  2221.     } while (++i < n_coeffs);

  2222. 

  2223.     return i;

  2224. }

  2225. 

  2226. static int decode_coeffs_b_8bpp(VP9Context *s, int16_t *coef, int n_coeffs,

  2227.                                 unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  2228.                                 uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  2229.                                 const int16_t (*nb)[2], const int16_t *band_counts,

  2230.                                 const int16_t *qmul)

  2231. {

  2232.     return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 0, 1, 8, cnt, eob, p,

  2233.                                    nnz, scan, nb, band_counts, qmul);

  2234. }

  2235. 

  2236. static int decode_coeffs_b32_8bpp(VP9Context *s, int16_t *coef, int n_coeffs,

  2237.                                   unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  2238.                                   uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  2239.                                   const int16_t (*nb)[2], const int16_t *band_counts,

  2240.                                   const int16_t *qmul)

  2241. {

  2242.     return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 1, 1, 8, cnt, eob, p,

  2243.                                    nnz, scan, nb, band_counts, qmul);

  2244. }

  2245. 

  2246. static int decode_coeffs_b_16bpp(VP9Context *s, int16_t *coef, int n_coeffs,

  2247.                                  unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  2248.                                  uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  2249.                                  const int16_t (*nb)[2], const int16_t *band_counts,

  2250.                                  const int16_t *qmul)

  2251. {

  2252.     return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 0, 0, s->bpp, cnt, eob, p,

  2253.                                    nnz, scan, nb, band_counts, qmul);

  2254. }

  2255. 

  2256. static int decode_coeffs_b32_16bpp(VP9Context *s, int16_t *coef, int n_coeffs,

  2257.                                    unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  2258.                                    uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  2259.                                    const int16_t (*nb)[2], const int16_t *band_counts,

  2260.                                    const int16_t *qmul)

  2261. {

  2262.     return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 1, 0, s->bpp, cnt, eob, p,

  2263.                                    nnz, scan, nb, band_counts, qmul);

  2264. }

  2265. 

  2266. static av_always_inline int decode_coeffs(AVCodecContext *ctx, int is8bitsperpixel)

  2267. {

  2268.     VP9Context *s = ctx->priv_data;

  2269.     VP9Block *b = s->b;

  2270.     int row = s->row, col = s->col;

  2271.     uint8_t (*p)[6][11] = s->prob.coef[b->tx][0 /* y */][!b->intra];

  2272.     unsigned (*c)[6][3] = s->counts.coef[b->tx][0 /* y */][!b->intra];

  2273.     unsigned (*e)[6][2] = s->counts.eob[b->tx][0 /* y */][!b->intra];

  2274.     int w4 = bwh_tab[1][b->bs][0] << 1, h4 = bwh_tab[1][b->bs][1] << 1;

  2275.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2276.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2277.     int n, pl, x, y, res;

  2278.     int16_t (*qmul)[2] = s->s.h.segmentation.feat[b->seg_id].qmul;

  2279.     int tx = 4 * s->s.h.lossless + b->tx;

  2280.     const int16_t * const *yscans = vp9_scans[tx];

  2281.     const int16_t (* const *ynbs)[2] = vp9_scans_nb[tx];

  2282.     const int16_t *uvscan = vp9_scans[b->uvtx][DCT_DCT];

  2283.     const int16_t (*uvnb)[2] = vp9_scans_nb[b->uvtx][DCT_DCT];

  2284.     uint8_t *a = &s->above_y_nnz_ctx[col * 2];

  2285.     uint8_t *l = &s->left_y_nnz_ctx[(row & 7) << 1];

  2286.     static const int16_t band_counts[4][8] = {

  2287.         { 1, 2, 3, 4,  3,   16 - 13 },

  2288.         { 1, 2, 3, 4, 11,   64 - 21 },

  2289.         { 1, 2, 3, 4, 11,  256 - 21 },

  2290.         { 1, 2, 3, 4, 11, 1024 - 21 },

  2291.     };

  2292.     const int16_t *y_band_counts = band_counts[b->tx];

  2293.     const int16_t *uv_band_counts = band_counts[b->uvtx];

  2294.     int bytesperpixel = is8bitsperpixel ? 1 : 2;

  2295.     int total_coeff = 0;

  2296. 

  2297. #define MERGE(la, end, step, rd) \

  2298.     for (n = 0; n < end; n += step) \

  2299.         la[n] = !!rd(&la[n])

  2300. #define MERGE_CTX(step, rd) \

  2301.     do { \

  2302.         MERGE(l, end_y, step, rd); \

  2303.         MERGE(a, end_x, step, rd); \

  2304.     } while (0)

  2305. 

  2306. #define DECODE_Y_COEF_LOOP(step, mode_index, v) \

  2307.     for (n = 0, y = 0; y < end_y; y += step) { \

  2308.         for (x = 0; x < end_x; x += step, n += step * step) { \

  2309.             enum TxfmType txtp = vp9_intra_txfm_type[b->mode[mode_index]]; \

  2310.             res = (is8bitsperpixel ? decode_coeffs_b##v##_8bpp : decode_coeffs_b##v##_16bpp) \

  2311.                                     (s, s->block + 16 * n * bytesperpixel, 16 * step * step, \

  2312.                                      c, e, p, a[x] + l[y], yscans[txtp], \

  2313.                                      ynbs[txtp], y_band_counts, qmul[0]); \

  2314.             a[x] = l[y] = !!res; \

  2315.             total_coeff |= !!res; \

  2316.             if (step >= 4) { \

  2317.                 AV_WN16A(&s->eob[n], res); \

  2318.             } else { \

  2319.                 s->eob[n] = res; \

  2320.             } \

  2321.         } \

  2322.     }

  2323. 

  2324. #define SPLAT(la, end, step, cond) \

  2325.     if (step == 2) { \

  2326.         for (n = 1; n < end; n += step) \

  2327.             la[n] = la[n - 1]; \

  2328.     } else if (step == 4) { \

  2329.         if (cond) { \

  2330.             for (n = 0; n < end; n += step) \

  2331.                 AV_WN32A(&la[n], la[n] * 0x01010101); \

  2332.         } else { \

  2333.             for (n = 0; n < end; n += step) \

  2334.                 memset(&la[n + 1], la[n], FFMIN(end - n - 1, 3)); \

  2335.         } \

  2336.     } else /* step == 8 */ { \

  2337.         if (cond) { \

  2338.             if (HAVE_FAST_64BIT) { \

  2339.                 for (n = 0; n < end; n += step) \

  2340.                     AV_WN64A(&la[n], la[n] * 0x0101010101010101ULL); \

  2341.             } else { \

  2342.                 for (n = 0; n < end; n += step) { \

  2343.                     uint32_t v32 = la[n] * 0x01010101; \

  2344.                     AV_WN32A(&la[n],     v32); \

  2345.                     AV_WN32A(&la[n + 4], v32); \

  2346.                 } \

  2347.             } \

  2348.         } else { \

  2349.             for (n = 0; n < end; n += step) \

  2350.                 memset(&la[n + 1], la[n], FFMIN(end - n - 1, 7)); \

  2351.         } \

  2352.     }

  2353. #define SPLAT_CTX(step) \

  2354.     do { \

  2355.         SPLAT(a, end_x, step, end_x == w4); \

  2356.         SPLAT(l, end_y, step, end_y == h4); \

  2357.     } while (0)

  2358. 

  2359.     /* y tokens */

  2360.     switch (b->tx) {

  2361.     case TX_4X4:

  2362.         DECODE_Y_COEF_LOOP(1, b->bs > BS_8x8 ? n : 0,);

  2363.         break;

  2364.     case TX_8X8:

  2365.         MERGE_CTX(2, AV_RN16A);

  2366.         DECODE_Y_COEF_LOOP(2, 0,);

  2367.         SPLAT_CTX(2);

  2368.         break;

  2369.     case TX_16X16:

  2370.         MERGE_CTX(4, AV_RN32A);

  2371.         DECODE_Y_COEF_LOOP(4, 0,);

  2372.         SPLAT_CTX(4);

  2373.         break;

  2374.     case TX_32X32:

  2375.         MERGE_CTX(8, AV_RN64A);

  2376.         DECODE_Y_COEF_LOOP(8, 0, 32);

  2377.         SPLAT_CTX(8);

  2378.         break;

  2379.     }

  2380. 

  2381. #define DECODE_UV_COEF_LOOP(step, v) \

  2382.     for (n = 0, y = 0; y < end_y; y += step) { \

  2383.         for (x = 0; x < end_x; x += step, n += step * step) { \

  2384.             res = (is8bitsperpixel ? decode_coeffs_b##v##_8bpp : decode_coeffs_b##v##_16bpp) \

  2385.                                     (s, s->uvblock[pl] + 16 * n * bytesperpixel, \

  2386.                                      16 * step * step, c, e, p, a[x] + l[y], \

  2387.                                      uvscan, uvnb, uv_band_counts, qmul[1]); \

  2388.             a[x] = l[y] = !!res; \

  2389.             total_coeff |= !!res; \

  2390.             if (step >= 4) { \

  2391.                 AV_WN16A(&s->uveob[pl][n], res); \

  2392.             } else { \

  2393.                 s->uveob[pl][n] = res; \

  2394.             } \

  2395.         } \

  2396.     }

  2397. 

  2398.     p = s->prob.coef[b->uvtx][1 /* uv */][!b->intra];

  2399.     c = s->counts.coef[b->uvtx][1 /* uv */][!b->intra];

  2400.     e = s->counts.eob[b->uvtx][1 /* uv */][!b->intra];

  2401.     w4 >>= s->ss_h;

  2402.     end_x >>= s->ss_h;

  2403.     h4 >>= s->ss_v;

  2404.     end_y >>= s->ss_v;

  2405.     for (pl = 0; pl < 2; pl++) {

  2406.         a = &s->above_uv_nnz_ctx[pl][col << !s->ss_h];

  2407.         l = &s->left_uv_nnz_ctx[pl][(row & 7) << !s->ss_v];

  2408.         switch (b->uvtx) {

  2409.         case TX_4X4:

  2410.             DECODE_UV_COEF_LOOP(1,);

  2411.             break;

  2412.         case TX_8X8:

  2413.             MERGE_CTX(2, AV_RN16A);

  2414.             DECODE_UV_COEF_LOOP(2,);

  2415.             SPLAT_CTX(2);

  2416.             break;

  2417.         case TX_16X16:

  2418.             MERGE_CTX(4, AV_RN32A);

  2419.             DECODE_UV_COEF_LOOP(4,);

  2420.             SPLAT_CTX(4);

  2421.             break;

  2422.         case TX_32X32:

  2423.             MERGE_CTX(8, AV_RN64A);

  2424.             DECODE_UV_COEF_LOOP(8, 32);

  2425.             SPLAT_CTX(8);

  2426.             break;

  2427.         }

  2428.     }

  2429. 

  2430.     return total_coeff;

  2431. }

  2432. 

  2433. static int decode_coeffs_8bpp(AVCodecContext *ctx)

  2434. {

  2435.     return decode_coeffs(ctx, 1);

  2436. }

  2437. 

  2438. static int decode_coeffs_16bpp(AVCodecContext *ctx)

  2439. {

  2440.     return decode_coeffs(ctx, 0);

  2441. }

  2442. 

  2443. static av_always_inline int check_intra_mode(VP9Context *s, int mode, uint8_t **a,

  2444.                                              uint8_t *dst_edge, ptrdiff_t stride_edge,

  2445.                                              uint8_t *dst_inner, ptrdiff_t stride_inner,

  2446.                                              uint8_t *l, int col, int x, int w,

  2447.                                              int row, int y, enum TxfmMode tx,

  2448.                                              int p, int ss_h, int ss_v, int bytesperpixel)

  2449. {

  2450.     int have_top = row > 0 || y > 0;

  2451.     int have_left = col > s->tile_col_start || x > 0;

  2452.     int have_right = x < w - 1;

  2453.     int bpp = s->bpp;

  2454.     static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {

  2455.         [VERT_PRED]            = { { DC_127_PRED,          VERT_PRED },

  2456.                                    { DC_127_PRED,          VERT_PRED } },

  2457.         [HOR_PRED]             = { { DC_129_PRED,          DC_129_PRED },

  2458.                                    { HOR_PRED,             HOR_PRED } },

  2459.         [DC_PRED]              = { { DC_128_PRED,          TOP_DC_PRED },

  2460.                                    { LEFT_DC_PRED,         DC_PRED } },

  2461.         [DIAG_DOWN_LEFT_PRED]  = { { DC_127_PRED,          DIAG_DOWN_LEFT_PRED },

  2462.                                    { DC_127_PRED,          DIAG_DOWN_LEFT_PRED } },

  2463.         [DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },

  2464.                                    { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },

  2465.         [VERT_RIGHT_PRED]      = { { VERT_RIGHT_PRED,      VERT_RIGHT_PRED },

  2466.                                    { VERT_RIGHT_PRED,      VERT_RIGHT_PRED } },

  2467.         [HOR_DOWN_PRED]        = { { HOR_DOWN_PRED,        HOR_DOWN_PRED },

  2468.                                    { HOR_DOWN_PRED,        HOR_DOWN_PRED } },

  2469.         [VERT_LEFT_PRED]       = { { DC_127_PRED,          VERT_LEFT_PRED },

  2470.                                    { DC_127_PRED,          VERT_LEFT_PRED } },

  2471.         [HOR_UP_PRED]          = { { DC_129_PRED,          DC_129_PRED },

  2472.                                    { HOR_UP_PRED,          HOR_UP_PRED } },

  2473.         [TM_VP8_PRED]          = { { DC_129_PRED,          VERT_PRED },

  2474.                                    { HOR_PRED,             TM_VP8_PRED } },

  2475.     };

  2476.     static const struct {

  2477.         uint8_t needs_left:1;

  2478.         uint8_t needs_top:1;

  2479.         uint8_t needs_topleft:1;

  2480.         uint8_t needs_topright:1;

  2481.         uint8_t invert_left:1;

  2482.     } edges[N_INTRA_PRED_MODES] = {

  2483.         [VERT_PRED]            = { .needs_top  = 1 },

  2484.         [HOR_PRED]             = { .needs_left = 1 },

  2485.         [DC_PRED]              = { .needs_top  = 1, .needs_left = 1 },

  2486.         [DIAG_DOWN_LEFT_PRED]  = { .needs_top  = 1, .needs_topright = 1 },

  2487.         [DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2488.         [VERT_RIGHT_PRED]      = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2489.         [HOR_DOWN_PRED]        = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2490.         [VERT_LEFT_PRED]       = { .needs_top  = 1, .needs_topright = 1 },

  2491.         [HOR_UP_PRED]          = { .needs_left = 1, .invert_left = 1 },

  2492.         [TM_VP8_PRED]          = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2493.         [LEFT_DC_PRED]         = { .needs_left = 1 },

  2494.         [TOP_DC_PRED]          = { .needs_top  = 1 },

  2495.         [DC_128_PRED]          = { 0 },

  2496.         [DC_127_PRED]          = { 0 },

  2497.         [DC_129_PRED]          = { 0 }

  2498.     };

  2499. 

  2500.     av_assert2(mode >= 0 && mode < 10);

  2501.     mode = mode_conv[mode][have_left][have_top];

  2502.     if (edges[mode].needs_top) {

  2503.         uint8_t *top, *topleft;

  2504.         int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !ss_h) - x) * 4;

  2505.         int n_px_need_tr = 0;

  2506. 

  2507.         if (tx == TX_4X4 && edges[mode].needs_topright && have_right)

  2508.             n_px_need_tr = 4;

  2509. 

  2510.         // if top of sb64-row, use s->intra_pred_data[] instead of

  2511.         // dst[-stride] for intra prediction (it contains pre- instead of

  2512.         // post-loopfilter data)

  2513.         if (have_top) {

  2514.             top = !(row & 7) && !y ?

  2515.                 s->intra_pred_data[p] + (col * (8 >> ss_h) + x * 4) * bytesperpixel :

  2516.                 y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];

  2517.             if (have_left)

  2518.                 topleft = !(row & 7) && !y ?

  2519.                     s->intra_pred_data[p] + (col * (8 >> ss_h) + x * 4) * bytesperpixel :

  2520.                     y == 0 || x == 0 ? &dst_edge[-stride_edge] :

  2521.                     &dst_inner[-stride_inner];

  2522.         }

  2523. 

  2524.         if (have_top &&

  2525.             (!edges[mode].needs_topleft || (have_left && top == topleft)) &&

  2526.             (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&

  2527.             n_px_need + n_px_need_tr <= n_px_have) {

  2528.             *a = top;

  2529.         } else {

  2530.             if (have_top) {

  2531.                 if (n_px_need <= n_px_have) {

  2532.                     memcpy(*a, top, n_px_need * bytesperpixel);

  2533.                 } else {

  2534. #define memset_bpp(c, i1, v, i2, num) do { \

  2535.     if (bytesperpixel == 1) { \

  2536.         memset(&(c)[(i1)], (v)[(i2)], (num)); \

  2537.     } else { \

  2538.         int n, val = AV_RN16A(&(v)[(i2) * 2]); \

  2539.         for (n = 0; n < (num); n++) { \

  2540.             AV_WN16A(&(c)[((i1) + n) * 2], val); \

  2541.         } \

  2542.     } \

  2543. } while (0)

  2544.                     memcpy(*a, top, n_px_have * bytesperpixel);

  2545.                     memset_bpp(*a, n_px_have, (*a), n_px_have - 1, n_px_need - n_px_have);

  2546.                 }

  2547.             } else {

  2548. #define memset_val(c, val, num) do { \

  2549.     if (bytesperpixel == 1) { \

  2550.         memset((c), (val), (num)); \

  2551.     } else { \

  2552.         int n; \

  2553.         for (n = 0; n < (num); n++) { \

  2554.             AV_WN16A(&(c)[n * 2], (val)); \

  2555.         } \

  2556.     } \

  2557. } while (0)

  2558.                 memset_val(*a, (128 << (bpp - 8)) - 1, n_px_need);

  2559.             }

  2560.             if (edges[mode].needs_topleft) {

  2561.                 if (have_left && have_top) {

  2562. #define assign_bpp(c, i1, v, i2) do { \

  2563.     if (bytesperpixel == 1) { \

  2564.         (c)[(i1)] = (v)[(i2)]; \

  2565.     } else { \

  2566.         AV_COPY16(&(c)[(i1) * 2], &(v)[(i2) * 2]); \

  2567.     } \

  2568. } while (0)

  2569.                     assign_bpp(*a, -1, topleft, -1);

  2570.                 } else {

  2571. #define assign_val(c, i, v) do { \

  2572.     if (bytesperpixel == 1) { \

  2573.         (c)[(i)] = (v); \

  2574.     } else { \

  2575.         AV_WN16A(&(c)[(i) * 2], (v)); \

  2576.     } \

  2577. } while (0)

  2578.                     assign_val((*a), -1, (128 << (bpp - 8)) + (have_top ? +1 : -1));

  2579.                 }

  2580.             }

  2581.             if (tx == TX_4X4 && edges[mode].needs_topright) {

  2582.                 if (have_top && have_right &&

  2583.                     n_px_need + n_px_need_tr <= n_px_have) {

  2584.                     memcpy(&(*a)[4 * bytesperpixel], &top[4 * bytesperpixel], 4 * bytesperpixel);

  2585.                 } else {

  2586.                     memset_bpp(*a, 4, *a, 3, 4);

  2587.                 }

  2588.             }

  2589.         }

  2590.     }

  2591.     if (edges[mode].needs_left) {

  2592.         if (have_left) {

  2593.             int n_px_need = 4 << tx, i, n_px_have = (((s->rows - row) << !ss_v) - y) * 4;

  2594.             uint8_t *dst = x == 0 ? dst_edge : dst_inner;

  2595.             ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;

  2596. 

  2597.             if (edges[mode].invert_left) {

  2598.                 if (n_px_need <= n_px_have) {

  2599.                     for (i = 0; i < n_px_need; i++)

  2600.                         assign_bpp(l, i, &dst[i * stride], -1);

  2601.                 } else {

  2602.                     for (i = 0; i < n_px_have; i++)

  2603.                         assign_bpp(l, i, &dst[i * stride], -1);

  2604.                     memset_bpp(l, n_px_have, l, n_px_have - 1, n_px_need - n_px_have);

  2605.                 }

  2606.             } else {

  2607.                 if (n_px_need <= n_px_have) {

  2608.                     for (i = 0; i < n_px_need; i++)

  2609.                         assign_bpp(l, n_px_need - 1 - i, &dst[i * stride], -1);

  2610.                 } else {

  2611.                     for (i = 0; i < n_px_have; i++)

  2612.                         assign_bpp(l, n_px_need - 1 - i, &dst[i * stride], -1);

  2613.                     memset_bpp(l, 0, l, n_px_need - n_px_have, n_px_need - n_px_have);

  2614.                 }

  2615.             }

  2616.         } else {

  2617.             memset_val(l, (128 << (bpp - 8)) + 1, 4 << tx);

  2618.         }

  2619.     }

  2620. 

  2621.     return mode;

  2622. }

  2623. 

  2624. static av_always_inline void intra_recon(AVCodecContext *ctx, ptrdiff_t y_off,

  2625.                                          ptrdiff_t uv_off, int bytesperpixel)

  2626. {

  2627.     VP9Context *s = ctx->priv_data;

  2628.     VP9Block *b = s->b;

  2629.     int row = s->row, col = s->col;

  2630.     int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2631.     int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2632.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2633.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2634.     int tx = 4 * s->s.h.lossless + b->tx, uvtx = b->uvtx + 4 * s->s.h.lossless;

  2635.     int uvstep1d = 1 << b->uvtx, p;

  2636.     uint8_t *dst = s->dst[0], *dst_r = s->s.frames[CUR_FRAME].tf.f->data[0] + y_off;

  2637.     LOCAL_ALIGNED_32(uint8_t, a_buf, [96]);

  2638.     LOCAL_ALIGNED_32(uint8_t, l, [64]);

  2639. 

  2640.     for (n = 0, y = 0; y < end_y; y += step1d) {

  2641.         uint8_t *ptr = dst, *ptr_r = dst_r;

  2642.         for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d * bytesperpixel,

  2643.                                ptr_r += 4 * step1d * bytesperpixel, n += step) {

  2644.             int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?

  2645.                                y * 2 + x : 0];

  2646.             uint8_t *a = &a_buf[32];

  2647.             enum TxfmType txtp = vp9_intra_txfm_type[mode];

  2648.             int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2649. 

  2650.             mode = check_intra_mode(s, mode, &a, ptr_r,

  2651.                                     s->s.frames[CUR_FRAME].tf.f->linesize[0],

  2652.                                     ptr, s->y_stride, l,

  2653.                                     col, x, w4, row, y, b->tx, 0, 0, 0, bytesperpixel);

  2654.             s->dsp.intra_pred[b->tx][mode](ptr, s->y_stride, l, a);

  2655.             if (eob)

  2656.                 s->dsp.itxfm_add[tx][txtp](ptr, s->y_stride,

  2657.                                            s->block + 16 * n * bytesperpixel, eob);

  2658.         }

  2659.         dst_r += 4 * step1d * s->s.frames[CUR_FRAME].tf.f->linesize[0];

  2660.         dst   += 4 * step1d * s->y_stride;

  2661.     }

  2662. 

  2663.     // U/V

  2664.     w4 >>= s->ss_h;

  2665.     end_x >>= s->ss_h;

  2666.     end_y >>= s->ss_v;

  2667.     step = 1 << (b->uvtx * 2);

  2668.     for (p = 0; p < 2; p++) {

  2669.         dst   = s->dst[1 + p];

  2670.         dst_r = s->s.frames[CUR_FRAME].tf.f->data[1 + p] + uv_off;

  2671.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2672.             uint8_t *ptr = dst, *ptr_r = dst_r;

  2673.             for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d * bytesperpixel,

  2674.                                    ptr_r += 4 * uvstep1d * bytesperpixel, n += step) {

  2675.                 int mode = b->uvmode;

  2676.                 uint8_t *a = &a_buf[32];

  2677.                 int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2678. 

  2679.                 mode = check_intra_mode(s, mode, &a, ptr_r,

  2680.                                         s->s.frames[CUR_FRAME].tf.f->linesize[1],

  2681.                                         ptr, s->uv_stride, l, col, x, w4, row, y,

  2682.                                         b->uvtx, p + 1, s->ss_h, s->ss_v, bytesperpixel);

  2683.                 s->dsp.intra_pred[b->uvtx][mode](ptr, s->uv_stride, l, a);

  2684.                 if (eob)

  2685.                     s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2686.                                                     s->uvblock[p] + 16 * n * bytesperpixel, eob);

  2687.             }

  2688.             dst_r += 4 * uvstep1d * s->s.frames[CUR_FRAME].tf.f->linesize[1];

  2689.             dst   += 4 * uvstep1d * s->uv_stride;

  2690.         }

  2691.     }

  2692. }

  2693. 

  2694. static void intra_recon_8bpp(AVCodecContext *ctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  2695. {

  2696.     intra_recon(ctx, y_off, uv_off, 1);

  2697. }

  2698. 

  2699. static void intra_recon_16bpp(AVCodecContext *ctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  2700. {

  2701.     intra_recon(ctx, y_off, uv_off, 2);

  2702. }

  2703. 

  2704. static av_always_inline void mc_luma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2],

  2705.                                               uint8_t *dst, ptrdiff_t dst_stride,

  2706.                                               const uint8_t *ref, ptrdiff_t ref_stride,

  2707.                                               ThreadFrame *ref_frame,

  2708.                                               ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2709.                                               int bw, int bh, int w, int h, int bytesperpixel)

  2710. {

  2711.     int mx = mv->x, my = mv->y, th;

  2712. 

  2713.     y += my >> 3;

  2714.     x += mx >> 3;

  2715.     ref += y * ref_stride + x * bytesperpixel;

  2716.     mx &= 7;

  2717.     my &= 7;

  2718.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2719.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2720.     // the longest loopfilter of the next sbrow

  2721.     th = (y + bh + 4 * !!my + 7) >> 6;

  2722.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2723.     if (x < !!mx * 3 || y < !!my * 3 ||

  2724.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2725.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2726.                                  ref - !!my * 3 * ref_stride - !!mx * 3 * bytesperpixel,

  2727.                                  160, ref_stride,

  2728.                                  bw + !!mx * 7, bh + !!my * 7,

  2729.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2730.         ref = s->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;

  2731.         ref_stride = 160;

  2732.     }

  2733.     mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);

  2734. }

  2735. 

  2736. static av_always_inline void mc_chroma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2],

  2737.                                                 uint8_t *dst_u, uint8_t *dst_v,

  2738.                                                 ptrdiff_t dst_stride,

  2739.                                                 const uint8_t *ref_u, ptrdiff_t src_stride_u,

  2740.                                                 const uint8_t *ref_v, ptrdiff_t src_stride_v,

  2741.                                                 ThreadFrame *ref_frame,

  2742.                                                 ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2743.                                                 int bw, int bh, int w, int h, int bytesperpixel)

  2744. {

  2745.     int mx = mv->x << !s->ss_h, my = mv->y << !s->ss_v, th;

  2746. 

  2747.     y += my >> 4;

  2748.     x += mx >> 4;

  2749.     ref_u += y * src_stride_u + x * bytesperpixel;

  2750.     ref_v += y * src_stride_v + x * bytesperpixel;

  2751.     mx &= 15;

  2752.     my &= 15;

  2753.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2754.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2755.     // the longest loopfilter of the next sbrow

  2756.     th = (y + bh + 4 * !!my + 7) >> (6 - s->ss_v);

  2757.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2758.     if (x < !!mx * 3 || y < !!my * 3 ||

  2759.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2760.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2761.                                  ref_u - !!my * 3 * src_stride_u - !!mx * 3 * bytesperpixel,

  2762.                                  160, src_stride_u,

  2763.                                  bw + !!mx * 7, bh + !!my * 7,

  2764.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2765.         ref_u = s->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;

  2766.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, 160, bh, mx, my);

  2767. 

  2768.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2769.                                  ref_v - !!my * 3 * src_stride_v - !!mx * 3 * bytesperpixel,

  2770.                                  160, src_stride_v,

  2771.                                  bw + !!mx * 7, bh + !!my * 7,

  2772.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2773.         ref_v = s->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;

  2774.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, 160, bh, mx, my);

  2775.     } else {

  2776.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);

  2777.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);

  2778.     }

  2779. }

  2780. 

  2781. #define mc_luma_dir(s, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, \

  2782.                     px, py, pw, ph, bw, bh, w, h, i) \

  2783.     mc_luma_unscaled(s, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \

  2784.                      mv, bw, bh, w, h, bytesperpixel)

  2785. #define mc_chroma_dir(s, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2786.                       row, col, mv, px, py, pw, ph, bw, bh, w, h, i) \

  2787.     mc_chroma_unscaled(s, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2788.                        row, col, mv, bw, bh, w, h, bytesperpixel)

  2789. #define SCALED 0

  2790. #define FN(x) x##_8bpp

  2791. #define BYTES_PER_PIXEL 1

  2792. #include "vp9_mc_template.c"

  2793. #undef FN

  2794. #undef BYTES_PER_PIXEL

  2795. #define FN(x) x##_16bpp

  2796. #define BYTES_PER_PIXEL 2

  2797. #include "vp9_mc_template.c"

  2798. #undef mc_luma_dir

  2799. #undef mc_chroma_dir

  2800. #undef FN

  2801. #undef BYTES_PER_PIXEL

  2802. #undef SCALED

  2803. 

  2804. static av_always_inline void mc_luma_scaled(VP9Context *s, vp9_scaled_mc_func smc,

  2805.                                             vp9_mc_func (*mc)[2],

  2806.                                             uint8_t *dst, ptrdiff_t dst_stride,

  2807.                                             const uint8_t *ref, ptrdiff_t ref_stride,

  2808.                                             ThreadFrame *ref_frame,

  2809.                                             ptrdiff_t y, ptrdiff_t x, const VP56mv *in_mv,

  2810.                                             int px, int py, int pw, int ph,

  2811.                                             int bw, int bh, int w, int h, int bytesperpixel,

  2812.                                             const uint16_t *scale, const uint8_t *step)

  2813. {

  2814.     if (s->s.frames[CUR_FRAME].tf.f->width == ref_frame->f->width &&

  2815.         s->s.frames[CUR_FRAME].tf.f->height == ref_frame->f->height) {

  2816.         mc_luma_unscaled(s, mc, dst, dst_stride, ref, ref_stride, ref_frame,

  2817.                          y, x, in_mv, bw, bh, w, h, bytesperpixel);

  2818.     } else {

  2819. #define scale_mv(n, dim) (((int64_t)(n) * scale[dim]) >> 14)

  2820.     int mx, my;

  2821.     int refbw_m1, refbh_m1;

  2822.     int th;

  2823.     VP56mv mv;

  2824. 

  2825.     mv.x = av_clip(in_mv->x, -(x + pw - px + 4) << 3, (s->cols * 8 - x + px + 3) << 3);

  2826.     mv.y = av_clip(in_mv->y, -(y + ph - py + 4) << 3, (s->rows * 8 - y + py + 3) << 3);

  2827.     // BUG libvpx seems to scale the two components separately. This introduces

  2828.     // rounding errors but we have to reproduce them to be exactly compatible

  2829.     // with the output from libvpx...

  2830.     mx = scale_mv(mv.x * 2, 0) + scale_mv(x * 16, 0);

  2831.     my = scale_mv(mv.y * 2, 1) + scale_mv(y * 16, 1);

  2832. 

  2833.     y = my >> 4;

  2834.     x = mx >> 4;

  2835.     ref += y * ref_stride + x * bytesperpixel;

  2836.     mx &= 15;

  2837.     my &= 15;

  2838.     refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;

  2839.     refbh_m1 = ((bh - 1) * step[1] + my) >> 4;

  2840.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2841.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2842.     // the longest loopfilter of the next sbrow

  2843.     th = (y + refbh_m1 + 4 + 7) >> 6;

  2844.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2845.     if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 4 >= h - refbh_m1) {

  2846.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2847.                                  ref - 3 * ref_stride - 3 * bytesperpixel,

  2848.                                  288, ref_stride,

  2849.                                  refbw_m1 + 8, refbh_m1 + 8,

  2850.                                  x - 3, y - 3, w, h);

  2851.         ref = s->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;

  2852.         ref_stride = 288;

  2853.     }

  2854.     smc(dst, dst_stride, ref, ref_stride, bh, mx, my, step[0], step[1]);

  2855.     }

  2856. }

  2857. 

  2858. static av_always_inline void mc_chroma_scaled(VP9Context *s, vp9_scaled_mc_func smc,

  2859.                                               vp9_mc_func (*mc)[2],

  2860.                                               uint8_t *dst_u, uint8_t *dst_v,

  2861.                                               ptrdiff_t dst_stride,

  2862.                                               const uint8_t *ref_u, ptrdiff_t src_stride_u,

  2863.                                               const uint8_t *ref_v, ptrdiff_t src_stride_v,

  2864.                                               ThreadFrame *ref_frame,

  2865.                                               ptrdiff_t y, ptrdiff_t x, const VP56mv *in_mv,

  2866.                                               int px, int py, int pw, int ph,

  2867.                                               int bw, int bh, int w, int h, int bytesperpixel,

  2868.                                               const uint16_t *scale, const uint8_t *step)

  2869. {

  2870.     if (s->s.frames[CUR_FRAME].tf.f->width == ref_frame->f->width &&

  2871.         s->s.frames[CUR_FRAME].tf.f->height == ref_frame->f->height) {

  2872.         mc_chroma_unscaled(s, mc, dst_u, dst_v, dst_stride, ref_u, src_stride_u,

  2873.                            ref_v, src_stride_v, ref_frame,

  2874.                            y, x, in_mv, bw, bh, w, h, bytesperpixel);

  2875.     } else {

  2876.     int mx, my;

  2877.     int refbw_m1, refbh_m1;

  2878.     int th;

  2879.     VP56mv mv;

  2880. 

  2881.     if (s->ss_h) {

  2882.         // BUG https://code.google.com/p/webm/issues/detail?id=820

  2883.         mv.x = av_clip(in_mv->x, -(x + pw - px + 4) << 4, (s->cols * 4 - x + px + 3) << 4);

  2884.         mx = scale_mv(mv.x, 0) + (scale_mv(x * 16, 0) & ~15) + (scale_mv(x * 32, 0) & 15);

  2885.     } else {

  2886.         mv.x = av_clip(in_mv->x, -(x + pw - px + 4) << 3, (s->cols * 8 - x + px + 3) << 3);

  2887.         mx = scale_mv(mv.x << 1, 0) + scale_mv(x * 16, 0);

  2888.     }

  2889.     if (s->ss_v) {

  2890.         // BUG https://code.google.com/p/webm/issues/detail?id=820

  2891.         mv.y = av_clip(in_mv->y, -(y + ph - py + 4) << 4, (s->rows * 4 - y + py + 3) << 4);

  2892.         my = scale_mv(mv.y, 1) + (scale_mv(y * 16, 1) & ~15) + (scale_mv(y * 32, 1) & 15);

  2893.     } else {

  2894.         mv.y = av_clip(in_mv->y, -(y + ph - py + 4) << 3, (s->rows * 8 - y + py + 3) << 3);

  2895.         my = scale_mv(mv.y << 1, 1) + scale_mv(y * 16, 1);

  2896.     }

  2897. #undef scale_mv

  2898.     y = my >> 4;

  2899.     x = mx >> 4;

  2900.     ref_u += y * src_stride_u + x * bytesperpixel;

  2901.     ref_v += y * src_stride_v + x * bytesperpixel;

  2902.     mx &= 15;

  2903.     my &= 15;

  2904.     refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;

  2905.     refbh_m1 = ((bh - 1) * step[1] + my) >> 4;

  2906.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2907.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2908.     // the longest loopfilter of the next sbrow

  2909.     th = (y + refbh_m1 + 4 + 7) >> (6 - s->ss_v);

  2910.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2911.     if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 4 >= h - refbh_m1) {

  2912.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2913.                                  ref_u - 3 * src_stride_u - 3 * bytesperpixel,

  2914.                                  288, src_stride_u,

  2915.                                  refbw_m1 + 8, refbh_m1 + 8,

  2916.                                  x - 3, y - 3, w, h);

  2917.         ref_u = s->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;

  2918.         smc(dst_u, dst_stride, ref_u, 288, bh, mx, my, step[0], step[1]);

  2919. 

  2920.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2921.                                  ref_v - 3 * src_stride_v - 3 * bytesperpixel,

  2922.                                  288, src_stride_v,

  2923.                                  refbw_m1 + 8, refbh_m1 + 8,

  2924.                                  x - 3, y - 3, w, h);

  2925.         ref_v = s->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;

  2926.         smc(dst_v, dst_stride, ref_v, 288, bh, mx, my, step[0], step[1]);

  2927.     } else {

  2928.         smc(dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my, step[0], step[1]);

  2929.         smc(dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my, step[0], step[1]);

  2930.     }

  2931.     }

  2932. }

  2933. 

  2934. #define mc_luma_dir(s, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, \

  2935.                     px, py, pw, ph, bw, bh, w, h, i) \

  2936.     mc_luma_scaled(s, s->dsp.s##mc, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \

  2937.                    mv, px, py, pw, ph, bw, bh, w, h, bytesperpixel, \

  2938.                    s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])

  2939. #define mc_chroma_dir(s, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2940.                       row, col, mv, px, py, pw, ph, bw, bh, w, h, i) \

  2941.     mc_chroma_scaled(s, s->dsp.s##mc, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2942.                      row, col, mv, px, py, pw, ph, bw, bh, w, h, bytesperpixel, \

  2943.                      s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])

  2944. #define SCALED 1

  2945. #define FN(x) x##_scaled_8bpp

  2946. #define BYTES_PER_PIXEL 1

  2947. #include "vp9_mc_template.c"

  2948. #undef FN

  2949. #undef BYTES_PER_PIXEL

  2950. #define FN(x) x##_scaled_16bpp

  2951. #define BYTES_PER_PIXEL 2

  2952. #include "vp9_mc_template.c"

  2953. #undef mc_luma_dir

  2954. #undef mc_chroma_dir

  2955. #undef FN

  2956. #undef BYTES_PER_PIXEL

  2957. #undef SCALED

  2958. 

  2959. static av_always_inline void inter_recon(AVCodecContext *ctx, int bytesperpixel)

  2960. {

  2961.     VP9Context *s = ctx->priv_data;

  2962.     VP9Block *b = s->b;

  2963.     int row = s->row, col = s->col;

  2964. 

  2965.     if (s->mvscale[b->ref[0]][0] || (b->comp && s->mvscale[b->ref[1]][0])) {

  2966.         if (bytesperpixel == 1) {

  2967.             inter_pred_scaled_8bpp(ctx);

  2968.         } else {

  2969.             inter_pred_scaled_16bpp(ctx);

  2970.         }

  2971.     } else {

  2972.         if (bytesperpixel == 1) {

  2973.             inter_pred_8bpp(ctx);

  2974.         } else {

  2975.             inter_pred_16bpp(ctx);

  2976.         }

  2977.     }

  2978.     if (!b->skip) {

  2979.         /* mostly copied intra_recon() */

  2980. 

  2981.         int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2982.         int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2983.         int end_x = FFMIN(2 * (s->cols - col), w4);

  2984.         int end_y = FFMIN(2 * (s->rows - row), h4);

  2985.         int tx = 4 * s->s.h.lossless + b->tx, uvtx = b->uvtx + 4 * s->s.h.lossless;

  2986.         int uvstep1d = 1 << b->uvtx, p;

  2987.         uint8_t *dst = s->dst[0];

  2988. 

  2989.         // y itxfm add

  2990.         for (n = 0, y = 0; y < end_y; y += step1d) {

  2991.             uint8_t *ptr = dst;

  2992.             for (x = 0; x < end_x; x += step1d,

  2993.                  ptr += 4 * step1d * bytesperpixel, n += step) {

  2994.                 int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2995. 

  2996.                 if (eob)

  2997.                     s->dsp.itxfm_add[tx][DCT_DCT](ptr, s->y_stride,

  2998.                                                   s->block + 16 * n * bytesperpixel, eob);

  2999.             }

  3000.             dst += 4 * s->y_stride * step1d;

  3001.         }

  3002. 

  3003.         // uv itxfm add

  3004.         end_x >>= s->ss_h;

  3005.         end_y >>= s->ss_v;

  3006.         step = 1 << (b->uvtx * 2);

  3007.         for (p = 0; p < 2; p++) {

  3008.             dst = s->dst[p + 1];

  3009.             for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  3010.                 uint8_t *ptr = dst;

  3011.                 for (x = 0; x < end_x; x += uvstep1d,

  3012.                      ptr += 4 * uvstep1d * bytesperpixel, n += step) {

  3013.                     int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  3014. 

  3015.                     if (eob)

  3016.                         s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  3017.                                                         s->uvblock[p] + 16 * n * bytesperpixel, eob);

  3018.                 }

  3019.                 dst += 4 * uvstep1d * s->uv_stride;

  3020.             }

  3021.         }

  3022.     }

  3023. }

  3024. 

  3025. static void inter_recon_8bpp(AVCodecContext *ctx)

  3026. {

  3027.     inter_recon(ctx, 1);

  3028. }

  3029. 

  3030. static void inter_recon_16bpp(AVCodecContext *ctx)

  3031. {

  3032.     inter_recon(ctx, 2);

  3033. }

  3034. 

  3035. static av_always_inline void mask_edges(uint8_t (*mask)[8][4], int ss_h, int ss_v,

  3036.                                         int row_and_7, int col_and_7,

  3037.                                         int w, int h, int col_end, int row_end,

  3038.                                         enum TxfmMode tx, int skip_inter)

  3039. {

  3040.     static const unsigned wide_filter_col_mask[2] = { 0x11, 0x01 };

  3041.     static const unsigned wide_filter_row_mask[2] = { 0x03, 0x07 };

  3042. 

  3043.     // FIXME I'm pretty sure all loops can be replaced by a single LUT if

  3044.     // we make VP9Filter.mask uint64_t (i.e. row/col all single variable)

  3045.     // and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then

  3046.     // use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)

  3047. 

  3048.     // the intended behaviour of the vp9 loopfilter is to work on 8-pixel

  3049.     // edges. This means that for UV, we work on two subsampled blocks at

  3050.     // a time, and we only use the topleft block's mode information to set

  3051.     // things like block strength. Thus, for any block size smaller than

  3052.     // 16x16, ignore the odd portion of the block.

  3053.     if (tx == TX_4X4 && (ss_v | ss_h)) {

  3054.         if (h == ss_v) {

  3055.             if (row_and_7 & 1)

  3056.                 return;

  3057.             if (!row_end)

  3058.                 h += 1;

  3059.         }

  3060.         if (w == ss_h) {

  3061.             if (col_and_7 & 1)

  3062.                 return;

  3063.             if (!col_end)

  3064.                 w += 1;

  3065.         }

  3066.     }

  3067. 

  3068.     if (tx == TX_4X4 && !skip_inter) {

  3069.         int t = 1 << col_and_7, m_col = (t << w) - t, y;

  3070.         // on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide

  3071.         int m_row_8 = m_col & wide_filter_col_mask[ss_h], m_row_4 = m_col - m_row_8;

  3072. 

  3073.         for (y = row_and_7; y < h + row_and_7; y++) {

  3074.             int col_mask_id = 2 - !(y & wide_filter_row_mask[ss_v]);

  3075. 

  3076.             mask[0][y][1] |= m_row_8;

  3077.             mask[0][y][2] |= m_row_4;

  3078.             // for odd lines, if the odd col is not being filtered,

  3079.             // skip odd row also:

  3080.             // .---. <-- a

  3081.             // |   |

  3082.             // |___| <-- b

  3083.             // ^   ^

  3084.             // c   d

  3085.             //

  3086.             // if a/c are even row/col and b/d are odd, and d is skipped,

  3087.             // e.g. right edge of size-66x66.webm, then skip b also (bug)

  3088.             if ((ss_h & ss_v) && (col_end & 1) && (y & 1)) {

  3089.                 mask[1][y][col_mask_id] |= (t << (w - 1)) - t;

  3090.             } else {

  3091.                 mask[1][y][col_mask_id] |= m_col;

  3092.             }

  3093.             if (!ss_h)

  3094.                 mask[0][y][3] |= m_col;

  3095.             if (!ss_v) {

  3096.                 if (ss_h && (col_end & 1))

  3097.                     mask[1][y][3] |= (t << (w - 1)) - t;

  3098.                 else

  3099.                     mask[1][y][3] |= m_col;

  3100.             }

  3101.         }

  3102.     } else {

  3103.         int y, t = 1 << col_and_7, m_col = (t << w) - t;

  3104. 

  3105.         if (!skip_inter) {

  3106.             int mask_id = (tx == TX_8X8);

  3107.             static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };

  3108.             int l2 = tx + ss_h - 1, step1d;

  3109.             int m_row = m_col & masks[l2];

  3110. 

  3111.             // at odd UV col/row edges tx16/tx32 loopfilter edges, force

  3112.             // 8wd loopfilter to prevent going off the visible edge.

  3113.             if (ss_h && tx > TX_8X8 && (w ^ (w - 1)) == 1) {

  3114.                 int m_row_16 = ((t << (w - 1)) - t) & masks[l2];

  3115.                 int m_row_8 = m_row - m_row_16;

  3116. 

  3117.                 for (y = row_and_7; y < h + row_and_7; y++) {

  3118.                     mask[0][y][0] |= m_row_16;

  3119.                     mask[0][y][1] |= m_row_8;

  3120.                 }

  3121.             } else {

  3122.                 for (y = row_and_7; y < h + row_and_7; y++)

  3123.                     mask[0][y][mask_id] |= m_row;

  3124.             }

  3125. 

  3126.             l2 = tx + ss_v - 1;

  3127.             step1d = 1 << l2;

  3128.             if (ss_v && tx > TX_8X8 && (h ^ (h - 1)) == 1) {

  3129.                 for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)

  3130.                     mask[1][y][0] |= m_col;

  3131.                 if (y - row_and_7 == h - 1)

  3132.                     mask[1][y][1] |= m_col;

  3133.             } else {

  3134.                 for (y = row_and_7; y < h + row_and_7; y += step1d)

  3135.                     mask[1][y][mask_id] |= m_col;

  3136.             }

  3137.         } else if (tx != TX_4X4) {

  3138.             int mask_id;

  3139. 

  3140.             mask_id = (tx == TX_8X8) || (h == ss_v);

  3141.             mask[1][row_and_7][mask_id] |= m_col;

  3142.             mask_id = (tx == TX_8X8) || (w == ss_h);

  3143.             for (y = row_and_7; y < h + row_and_7; y++)

  3144.                 mask[0][y][mask_id] |= t;

  3145.         } else {

  3146.             int t8 = t & wide_filter_col_mask[ss_h], t4 = t - t8;

  3147. 

  3148.             for (y = row_and_7; y < h + row_and_7; y++) {

  3149.                 mask[0][y][2] |= t4;

  3150.                 mask[0][y][1] |= t8;

  3151.             }

  3152.             mask[1][row_and_7][2 - !(row_and_7 & wide_filter_row_mask[ss_v])] |= m_col;

  3153.         }

  3154.     }

  3155. }

  3156. 

  3157. static void decode_b(AVCodecContext *ctx, int row, int col,

  3158.                      struct VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,

  3159.                      enum BlockLevel bl, enum BlockPartition bp)

  3160. {

  3161.     VP9Context *s = ctx->priv_data;

  3162.     VP9Block *b = s->b;

  3163.     enum BlockSize bs = bl * 3 + bp;

  3164.     int bytesperpixel = s->bytesperpixel;

  3165.     int w4 = bwh_tab[1][bs][0], h4 = bwh_tab[1][bs][1], lvl;

  3166.     int emu[2];

  3167.     AVFrame *f = s->s.frames[CUR_FRAME].tf.f;

  3168. 

  3169.     s->row = row;

  3170.     s->row7 = row & 7;

  3171.     s->col = col;

  3172.     s->col7 = col & 7;

  3173.     s->min_mv.x = -(128 + col * 64);

  3174.     s->min_mv.y = -(128 + row * 64);

  3175.     s->max_mv.x = 128 + (s->cols - col - w4) * 64;

  3176.     s->max_mv.y = 128 + (s->rows - row - h4) * 64;

  3177.     if (s->pass < 2) {

  3178.         b->bs = bs;

  3179.         b->bl = bl;

  3180.         b->bp = bp;

  3181.         decode_mode(ctx);

  3182.         b->uvtx = b->tx - ((s->ss_h && w4 * 2 == (1 << b->tx)) ||

  3183.                            (s->ss_v && h4 * 2 == (1 << b->tx)));

  3184. 

  3185.         if (!b->skip) {

  3186.             int has_coeffs;

  3187. 

  3188.             if (bytesperpixel == 1) {

  3189.                 has_coeffs = decode_coeffs_8bpp(ctx);

  3190.             } else {

  3191.                 has_coeffs = decode_coeffs_16bpp(ctx);

  3192.             }

  3193.             if (!has_coeffs && b->bs <= BS_8x8 && !b->intra) {

  3194.                 b->skip = 1;

  3195.                 memset(&s->above_skip_ctx[col], 1, w4);

  3196.                 memset(&s->left_skip_ctx[s->row7], 1, h4);

  3197.             }

  3198.         } else {

  3199.             int row7 = s->row7;

  3200. 

  3201. #define SPLAT_ZERO_CTX(v, n) \

  3202.     switch (n) { \

  3203.     case 1:  v = 0;          break; \

  3204.     case 2:  AV_ZERO16(&v);  break; \

  3205.     case 4:  AV_ZERO32(&v);  break; \

  3206.     case 8:  AV_ZERO64(&v);  break; \

  3207.     case 16: AV_ZERO128(&v); break; \

  3208.     }

  3209. #define SPLAT_ZERO_YUV(dir, var, off, n, dir2) \

  3210.     do { \

  3211.         SPLAT_ZERO_CTX(s->dir##_y_##var[off * 2], n * 2); \

  3212.         if (s->ss_##dir2) { \

  3213.             SPLAT_ZERO_CTX(s->dir##_uv_##var[0][off], n); \

  3214.             SPLAT_ZERO_CTX(s->dir##_uv_##var[1][off], n); \

  3215.         } else { \

  3216.             SPLAT_ZERO_CTX(s->dir##_uv_##var[0][off * 2], n * 2); \

  3217.             SPLAT_ZERO_CTX(s->dir##_uv_##var[1][off * 2], n * 2); \

  3218.         } \

  3219.     } while (0)

  3220. 

  3221.             switch (w4) {

  3222.             case 1: SPLAT_ZERO_YUV(above, nnz_ctx, col, 1, h); break;

  3223.             case 2: SPLAT_ZERO_YUV(above, nnz_ctx, col, 2, h); break;

  3224.             case 4: SPLAT_ZERO_YUV(above, nnz_ctx, col, 4, h); break;

  3225.             case 8: SPLAT_ZERO_YUV(above, nnz_ctx, col, 8, h); break;

  3226.             }

  3227.             switch (h4) {

  3228.             case 1: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 1, v); break;

  3229.             case 2: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 2, v); break;

  3230.             case 4: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 4, v); break;

  3231.             case 8: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 8, v); break;

  3232.             }

  3233.         }

  3234. 

  3235.         if (s->pass == 1) {

  3236.             s->b++;

  3237.             s->block += w4 * h4 * 64 * bytesperpixel;

  3238.             s->uvblock[0] += w4 * h4 * 64 * bytesperpixel >> (s->ss_h + s->ss_v);

  3239.             s->uvblock[1] += w4 * h4 * 64 * bytesperpixel >> (s->ss_h + s->ss_v);

  3240.             s->eob += 4 * w4 * h4;

  3241.             s->uveob[0] += 4 * w4 * h4 >> (s->ss_h + s->ss_v);

  3242.             s->uveob[1] += 4 * w4 * h4 >> (s->ss_h + s->ss_v);

  3243. 

  3244.             return;

  3245.         }

  3246.     }

  3247. 

  3248.     // emulated overhangs if the stride of the target buffer can't hold. This

  3249.     // makes it possible to support emu-edge and so on even if we have large block

  3250.     // overhangs

  3251.     emu[0] = (col + w4) * 8 * bytesperpixel > f->linesize[0] ||

  3252.              (row + h4) > s->rows;

  3253.     emu[1] = ((col + w4) * 8 >> s->ss_h) * bytesperpixel > f->linesize[1] ||

  3254.              (row + h4) > s->rows;

  3255.     if (emu[0]) {

  3256.         s->dst[0] = s->tmp_y;

  3257.         s->y_stride = 128;

  3258.     } else {

  3259.         s->dst[0] = f->data[0] + yoff;

  3260.         s->y_stride = f->linesize[0];

  3261.     }

  3262.     if (emu[1]) {

  3263.         s->dst[1] = s->tmp_uv[0];

  3264.         s->dst[2] = s->tmp_uv[1];

  3265.         s->uv_stride = 128;

  3266.     } else {

  3267.         s->dst[1] = f->data[1] + uvoff;

  3268.         s->dst[2] = f->data[2] + uvoff;

  3269.         s->uv_stride = f->linesize[1];

  3270.     }

  3271.     if (b->intra) {

  3272.         if (s->bpp > 8) {

  3273.             intra_recon_16bpp(ctx, yoff, uvoff);

  3274.         } else {

  3275.             intra_recon_8bpp(ctx, yoff, uvoff);

  3276.         }

  3277.     } else {

  3278.         if (s->bpp > 8) {

  3279.             inter_recon_16bpp(ctx);

  3280.         } else {

  3281.             inter_recon_8bpp(ctx);

  3282.         }

  3283.     }

  3284.     if (emu[0]) {

  3285.         int w = FFMIN(s->cols - col, w4) * 8, h = FFMIN(s->rows - row, h4) * 8, n, o = 0;

  3286. 

  3287.         for (n = 0; o < w; n++) {

  3288.             int bw = 64 >> n;

  3289. 

  3290.             av_assert2(n <= 4);

  3291.             if (w & bw) {

  3292.                 s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o * bytesperpixel, f->linesize[0],

  3293.                                          s->tmp_y + o * bytesperpixel, 128, h, 0, 0);

  3294.                 o += bw;

  3295.             }

  3296.         }

  3297.     }

  3298.     if (emu[1]) {

  3299.         int w = FFMIN(s->cols - col, w4) * 8 >> s->ss_h;

  3300.         int h = FFMIN(s->rows - row, h4) * 8 >> s->ss_v, n, o = 0;

  3301. 

  3302.         for (n = s->ss_h; o < w; n++) {

  3303.             int bw = 64 >> n;

  3304. 

  3305.             av_assert2(n <= 4);

  3306.             if (w & bw) {

  3307.                 s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o * bytesperpixel, f->linesize[1],

  3308.                                          s->tmp_uv[0] + o * bytesperpixel, 128, h, 0, 0);

  3309.                 s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o * bytesperpixel, f->linesize[2],

  3310.                                          s->tmp_uv[1] + o * bytesperpixel, 128, h, 0, 0);

  3311.                 o += bw;

  3312.             }

  3313.         }

  3314.     }

  3315. 

  3316.     // pick filter level and find edges to apply filter to

  3317.     if (s->s.h.filter.level &&

  3318.         (lvl = s->s.h.segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]

  3319.                                                       [b->mode[3] != ZEROMV]) > 0) {

  3320.         int x_end = FFMIN(s->cols - col, w4), y_end = FFMIN(s->rows - row, h4);

  3321.         int skip_inter = !b->intra && b->skip, col7 = s->col7, row7 = s->row7;

  3322. 

  3323.         setctx_2d(&lflvl->level[row7 * 8 + col7], w4, h4, 8, lvl);

  3324.         mask_edges(lflvl->mask[0], 0, 0, row7, col7, x_end, y_end, 0, 0, b->tx, skip_inter);

  3325.         if (s->ss_h || s->ss_v)

  3326.             mask_edges(lflvl->mask[1], s->ss_h, s->ss_v, row7, col7, x_end, y_end,

  3327.                        s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,

  3328.                        s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,

  3329.                        b->uvtx, skip_inter);

  3330. 

  3331.         if (!s->filter_lut.lim_lut[lvl]) {

  3332.             int sharp = s->s.h.filter.sharpness;

  3333.             int limit = lvl;

  3334. 

  3335.             if (sharp > 0) {

  3336.                 limit >>= (sharp + 3) >> 2;

  3337.                 limit = FFMIN(limit, 9 - sharp);

  3338.             }

  3339.             limit = FFMAX(limit, 1);

  3340. 

  3341.             s->filter_lut.lim_lut[lvl] = limit;

  3342.             s->filter_lut.mblim_lut[lvl] = 2 * (lvl + 2) + limit;

  3343.         }

  3344.     }

  3345. 

  3346.     if (s->pass == 2) {

  3347.         s->b++;

  3348.         s->block += w4 * h4 * 64 * bytesperpixel;

  3349.         s->uvblock[0] += w4 * h4 * 64 * bytesperpixel >> (s->ss_v + s->ss_h);

  3350.         s->uvblock[1] += w4 * h4 * 64 * bytesperpixel >> (s->ss_v + s->ss_h);

  3351.         s->eob += 4 * w4 * h4;

  3352.         s->uveob[0] += 4 * w4 * h4 >> (s->ss_v + s->ss_h);

  3353.         s->uveob[1] += 4 * w4 * h4 >> (s->ss_v + s->ss_h);

  3354.     }

  3355. }

  3356. 

  3357. static void decode_sb(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  3358.                       ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  3359. {

  3360.     VP9Context *s = ctx->priv_data;

  3361.     int c = ((s->above_partition_ctx[col] >> (3 - bl)) & 1) |

  3362.             (((s->left_partition_ctx[row & 0x7] >> (3 - bl)) & 1) << 1);

  3363.     const uint8_t *p = s->s.h.keyframe || s->s.h.intraonly ? vp9_default_kf_partition_probs[bl][c] :

  3364.                                                      s->prob.p.partition[bl][c];

  3365.     enum BlockPartition bp;

  3366.     ptrdiff_t hbs = 4 >> bl;

  3367.     AVFrame *f = s->s.frames[CUR_FRAME].tf.f;

  3368.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  3369.     int bytesperpixel = s->bytesperpixel;

  3370. 

  3371.     if (bl == BL_8X8) {

  3372.         bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  3373.         decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3374.     } else if (col + hbs < s->cols) { // FIXME why not <=?

  3375.         if (row + hbs < s->rows) { // FIXME why not <=?

  3376.             bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  3377.             switch (bp) {

  3378.             case PARTITION_NONE:

  3379.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3380.                 break;

  3381.             case PARTITION_H:

  3382.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3383.                 yoff  += hbs * 8 * y_stride;

  3384.                 uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3385.                 decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, bl, bp);

  3386.                 break;

  3387.             case PARTITION_V:

  3388.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3389.                 yoff  += hbs * 8 * bytesperpixel;

  3390.                 uvoff += hbs * 8 * bytesperpixel >> s->ss_h;

  3391.                 decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, bl, bp);

  3392.                 break;

  3393.             case PARTITION_SPLIT:

  3394.                 decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3395.                 decode_sb(ctx, row, col + hbs, lflvl,

  3396.                           yoff + 8 * hbs * bytesperpixel,

  3397.                           uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);

  3398.                 yoff  += hbs * 8 * y_stride;

  3399.                 uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3400.                 decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3401.                 decode_sb(ctx, row + hbs, col + hbs, lflvl,

  3402.                           yoff + 8 * hbs * bytesperpixel,

  3403.                           uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);

  3404.                 break;

  3405.             default:

  3406.                 av_assert0(0);

  3407.             }

  3408.         } else if (vp56_rac_get_prob_branchy(&s->c, p[1])) {

  3409.             bp = PARTITION_SPLIT;

  3410.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3411.             decode_sb(ctx, row, col + hbs, lflvl,

  3412.                       yoff + 8 * hbs * bytesperpixel,

  3413.                       uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);

  3414.         } else {

  3415.             bp = PARTITION_H;

  3416.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3417.         }

  3418.     } else if (row + hbs < s->rows) { // FIXME why not <=?

  3419.         if (vp56_rac_get_prob_branchy(&s->c, p[2])) {

  3420.             bp = PARTITION_SPLIT;

  3421.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3422.             yoff  += hbs * 8 * y_stride;

  3423.             uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3424.             decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3425.         } else {

  3426.             bp = PARTITION_V;

  3427.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3428.         }

  3429.     } else {

  3430.         bp = PARTITION_SPLIT;

  3431.         decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3432.     }

  3433.     s->counts.partition[bl][c][bp]++;

  3434. }

  3435. 

  3436. static void decode_sb_mem(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  3437.                           ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  3438. {

  3439.     VP9Context *s = ctx->priv_data;

  3440.     VP9Block *b = s->b;

  3441.     ptrdiff_t hbs = 4 >> bl;

  3442.     AVFrame *f = s->s.frames[CUR_FRAME].tf.f;

  3443.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  3444.     int bytesperpixel = s->bytesperpixel;

  3445. 

  3446.     if (bl == BL_8X8) {

  3447.         av_assert2(b->bl == BL_8X8);

  3448.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  3449.     } else if (s->b->bl == bl) {

  3450.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  3451.         if (b->bp == PARTITION_H && row + hbs < s->rows) {

  3452.             yoff  += hbs * 8 * y_stride;

  3453.             uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3454.             decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, b->bl, b->bp);

  3455.         } else if (b->bp == PARTITION_V && col + hbs < s->cols) {

  3456.             yoff  += hbs * 8 * bytesperpixel;

  3457.             uvoff += hbs * 8 * bytesperpixel >> s->ss_h;

  3458.             decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, b->bl, b->bp);

  3459.         }

  3460.     } else {

  3461.         decode_sb_mem(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3462.         if (col + hbs < s->cols) { // FIXME why not <=?

  3463.             if (row + hbs < s->rows) {

  3464.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff + 8 * hbs * bytesperpixel,

  3465.                               uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);

  3466.                 yoff  += hbs * 8 * y_stride;

  3467.                 uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3468.                 decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3469.                 decode_sb_mem(ctx, row + hbs, col + hbs, lflvl,

  3470.                               yoff + 8 * hbs * bytesperpixel,

  3471.                               uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);

  3472.             } else {

  3473.                 yoff  += hbs * 8 * bytesperpixel;

  3474.                 uvoff += hbs * 8 * bytesperpixel >> s->ss_h;

  3475.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff, uvoff, bl + 1);

  3476.             }

  3477.         } else if (row + hbs < s->rows) {

  3478.             yoff  += hbs * 8 * y_stride;

  3479.             uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3480.             decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3481.         }

  3482.     }

  3483. }

  3484. 

  3485. static av_always_inline void filter_plane_cols(VP9Context *s, int col, int ss_h, int ss_v,

  3486.                                                uint8_t *lvl, uint8_t (*mask)[4],

  3487.                                                uint8_t *dst, ptrdiff_t ls)

  3488. {

  3489.     int y, x, bytesperpixel = s->bytesperpixel;

  3490. 

  3491.     // filter edges between columns (e.g. block1 | block2)

  3492.     for (y = 0; y < 8; y += 2 << ss_v, dst += 16 * ls, lvl += 16 << ss_v) {

  3493.         uint8_t *ptr = dst, *l = lvl, *hmask1 = mask[y], *hmask2 = mask[y + 1 + ss_v];

  3494.         unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2], hm13 = hmask1[3];

  3495.         unsigned hm2 = hmask2[1] | hmask2[2], hm23 = hmask2[3];

  3496.         unsigned hm = hm1 | hm2 | hm13 | hm23;

  3497. 

  3498.         for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 8 * bytesperpixel >> ss_h) {

  3499.             if (col || x > 1) {

  3500.                 if (hm1 & x) {

  3501.                     int L = *l, H = L >> 4;

  3502.                     int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

  3503. 

  3504.                     if (hmask1[0] & x) {

  3505.                         if (hmask2[0] & x) {

  3506.                             av_assert2(l[8 << ss_v] == L);

  3507.                             s->dsp.loop_filter_16[0](ptr, ls, E, I, H);

  3508.                         } else {

  3509.                             s->dsp.loop_filter_8[2][0](ptr, ls, E, I, H);

  3510.                         }

  3511.                     } else if (hm2 & x) {

  3512.                         L = l[8 << ss_v];

  3513.                         H |= (L >> 4) << 8;

  3514.                         E |= s->filter_lut.mblim_lut[L] << 8;

  3515.                         I |= s->filter_lut.lim_lut[L] << 8;

  3516.                         s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3517.                                                [!!(hmask2[1] & x)]

  3518.                                                [0](ptr, ls, E, I, H);

  3519.                     } else {

  3520.                         s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3521.                                             [0](ptr, ls, E, I, H);

  3522.                     }

  3523.                 } else if (hm2 & x) {

  3524.                     int L = l[8 << ss_v], H = L >> 4;

  3525.                     int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

  3526. 

  3527.                     s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3528.                                         [0](ptr + 8 * ls, ls, E, I, H);

  3529.                 }

  3530.             }

  3531.             if (ss_h) {

  3532.                 if (x & 0xAA)

  3533.                     l += 2;

  3534.             } else {

  3535.                 if (hm13 & x) {

  3536.                     int L = *l, H = L >> 4;

  3537.                     int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

  3538. 

  3539.                     if (hm23 & x) {

  3540.                         L = l[8 << ss_v];

  3541.                         H |= (L >> 4) << 8;

  3542.                         E |= s->filter_lut.mblim_lut[L] << 8;

  3543.                         I |= s->filter_lut.lim_lut[L] << 8;

  3544.                         s->dsp.loop_filter_mix2[0][0][0](ptr + 4 * bytesperpixel, ls, E, I, H);

  3545.                     } else {

  3546.                         s->dsp.loop_filter_8[0][0](ptr + 4 * bytesperpixel, ls, E, I, H);

  3547.                     }

  3548.                 } else if (hm23 & x) {

  3549.                     int L = l[8 << ss_v], H = L >> 4;

  3550.                     int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

  3551. 

  3552.                     s->dsp.loop_filter_8[0][0](ptr + 8 * ls + 4 * bytesperpixel, ls, E, I, H);

  3553.                 }

  3554.                 l++;

  3555.             }

  3556.         }

  3557.     }

  3558. }

  3559. 

  3560. static av_always_inline void filter_plane_rows(VP9Context *s, int row, int ss_h, int ss_v,

  3561.                                                uint8_t *lvl, uint8_t (*mask)[4],

  3562.                                                uint8_t *dst, ptrdiff_t ls)

  3563. {

  3564.     int y, x, bytesperpixel = s->bytesperpixel;

  3565. 

  3566.     //                                 block1

  3567.     // filter edges between rows (e.g. ------)

  3568.     //                                 block2

  3569.     for (y = 0; y < 8; y++, dst += 8 * ls >> ss_v) {

  3570.         uint8_t *ptr = dst, *l = lvl, *vmask = mask[y];

  3571.         unsigned vm = vmask[0] | vmask[1] | vmask[2], vm3 = vmask[3];

  3572. 

  3573.         for (x = 1; vm & ~(x - 1); x <<= (2 << ss_h), ptr += 16 * bytesperpixel, l += 2 << ss_h) {

  3574.             if (row || y) {

  3575.                 if (vm & x) {

  3576.                     int L = *l, H = L >> 4;

  3577.                     int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

  3578. 

  3579.                     if (vmask[0] & x) {

  3580.                         if (vmask[0] & (x << (1 + ss_h))) {

  3581.                             av_assert2(l[1 + ss_h] == L);

  3582.                             s->dsp.loop_filter_16[1](ptr, ls, E, I, H);

  3583.                         } else {

  3584.                             s->dsp.loop_filter_8[2][1](ptr, ls, E, I, H);

  3585.                         }

  3586.                     } else if (vm & (x << (1 + ss_h))) {

  3587.                         L = l[1 + ss_h];

  3588.                         H |= (L >> 4) << 8;

  3589.                         E |= s->filter_lut.mblim_lut[L] << 8;

  3590.                         I |= s->filter_lut.lim_lut[L] << 8;

  3591.                         s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3592.                                                [!!(vmask[1] & (x << (1 + ss_h)))]

  3593.                                                [1](ptr, ls, E, I, H);

  3594.                     } else {

  3595.                         s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3596.                                             [1](ptr, ls, E, I, H);

  3597.                     }

  3598.                 } else if (vm & (x << (1 + ss_h))) {

  3599.                     int L = l[1 + ss_h], H = L >> 4;

  3600.                     int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

  3601. 

  3602.                     s->dsp.loop_filter_8[!!(vmask[1] & (x << (1 + ss_h)))]

  3603.                                         [1](ptr + 8 * bytesperpixel, ls, E, I, H);

  3604.                 }

  3605.             }

  3606.             if (!ss_v) {

  3607.                 if (vm3 & x) {

  3608.                     int L = *l, H = L >> 4;

  3609.                     int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

  3610. 

  3611.                     if (vm3 & (x << (1 + ss_h))) {

  3612.                         L = l[1 + ss_h];

  3613.                         H |= (L >> 4) << 8;

  3614.                         E |= s->filter_lut.mblim_lut[L] << 8;

  3615.                         I |= s->filter_lut.lim_lut[L] << 8;

  3616.                         s->dsp.loop_filter_mix2[0][0][1](ptr + ls * 4, ls, E, I, H);

  3617.                     } else {

  3618.                         s->dsp.loop_filter_8[0][1](ptr + ls * 4, ls, E, I, H);

  3619.                     }

  3620.                 } else if (vm3 & (x << (1 + ss_h))) {

  3621.                     int L = l[1 + ss_h], H = L >> 4;

  3622.                     int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

  3623. 

  3624.                     s->dsp.loop_filter_8[0][1](ptr + ls * 4 + 8 * bytesperpixel, ls, E, I, H);

  3625.                 }

  3626.             }

  3627.         }

  3628.         if (ss_v) {

  3629.             if (y & 1)

  3630.                 lvl += 16;

  3631.         } else {

  3632.             lvl += 8;

  3633.         }

  3634.     }

  3635. }

  3636. 

  3637. static void loopfilter_sb(AVCodecContext *ctx, struct VP9Filter *lflvl,

  3638.                           int row, int col, ptrdiff_t yoff, ptrdiff_t uvoff)

  3639. {

  3640.     VP9Context *s = ctx->priv_data;

  3641.     AVFrame *f = s->s.frames[CUR_FRAME].tf.f;

  3642.     uint8_t *dst = f->data[0] + yoff;

  3643.     ptrdiff_t ls_y = f->linesize[0], ls_uv = f->linesize[1];

  3644.     uint8_t (*uv_masks)[8][4] = lflvl->mask[s->ss_h | s->ss_v];

  3645.     int p;

  3646. 

  3647.     // FIXME in how far can we interleave the v/h loopfilter calls? E.g.

  3648.     // if you think of them as acting on a 8x8 block max, we can interleave

  3649.     // each v/h within the single x loop, but that only works if we work on

  3650.     // 8 pixel blocks, and we won't always do that (we want at least 16px

  3651.     // to use SSE2 optimizations, perhaps 32 for AVX2)

  3652. 

  3653.     filter_plane_cols(s, col, 0, 0, lflvl->level, lflvl->mask[0][0], dst, ls_y);

  3654.     filter_plane_rows(s, row, 0, 0, lflvl->level, lflvl->mask[0][1], dst, ls_y);

  3655. 

  3656.     for (p = 0; p < 2; p++) {

  3657.         dst = f->data[1 + p] + uvoff;

  3658.         filter_plane_cols(s, col, s->ss_h, s->ss_v, lflvl->level, uv_masks[0], dst, ls_uv);

  3659.         filter_plane_rows(s, row, s->ss_h, s->ss_v, lflvl->level, uv_masks[1], dst, ls_uv);

  3660.     }

  3661. }

  3662. 

  3663. static void set_tile_offset(int *start, int *end, int idx, int log2_n, int n)

  3664. {

  3665.     int sb_start = ( idx      * n) >> log2_n;

  3666.     int sb_end   = ((idx + 1) * n) >> log2_n;

  3667.     *start = FFMIN(sb_start, n) << 3;

  3668.     *end   = FFMIN(sb_end,   n) << 3;

  3669. }

  3670. 

  3671. static av_always_inline void adapt_prob(uint8_t *p, unsigned ct0, unsigned ct1,

  3672.                                         int max_count, int update_factor)

  3673. {

  3674.     unsigned ct = ct0 + ct1, p2, p1;

  3675. 

  3676.     if (!ct)

  3677.         return;

  3678. 

  3679.     p1 = *p;

  3680.     p2 = ((ct0 << 8) + (ct >> 1)) / ct;

  3681.     p2 = av_clip(p2, 1, 255);

  3682.     ct = FFMIN(ct, max_count);

  3683.     update_factor = FASTDIV(update_factor * ct, max_count);

  3684. 

  3685.     // (p1 * (256 - update_factor) + p2 * update_factor + 128) >> 8

  3686.     *p = p1 + (((p2 - p1) * update_factor + 128) >> 8);

  3687. }

  3688. 

  3689. static void adapt_probs(VP9Context *s)

  3690. {

  3691.     int i, j, k, l, m;

  3692.     prob_context *p = &s->prob_ctx[s->s.h.framectxid].p;

  3693.     int uf = (s->s.h.keyframe || s->s.h.intraonly || !s->last_keyframe) ? 112 : 128;

  3694. 

  3695.     // coefficients

  3696.     for (i = 0; i < 4; i++)

  3697.         for (j = 0; j < 2; j++)

  3698.             for (k = 0; k < 2; k++)

  3699.                 for (l = 0; l < 6; l++)

  3700.                     for (m = 0; m < 6; m++) {

  3701.                         uint8_t *pp = s->prob_ctx[s->s.h.framectxid].coef[i][j][k][l][m];

  3702.                         unsigned *e = s->counts.eob[i][j][k][l][m];

  3703.                         unsigned *c = s->counts.coef[i][j][k][l][m];

  3704. 

  3705.                         if (l == 0 && m >= 3) // dc only has 3 pt

  3706.                             break;

  3707. 

  3708.                         adapt_prob(&pp[0], e[0], e[1], 24, uf);

  3709.                         adapt_prob(&pp[1], c[0], c[1] + c[2], 24, uf);

  3710.                         adapt_prob(&pp[2], c[1], c[2], 24, uf);

  3711.                     }

  3712. 

  3713.     if (s->s.h.keyframe || s->s.h.intraonly) {

  3714.         memcpy(p->skip,  s->prob.p.skip,  sizeof(p->skip));

  3715.         memcpy(p->tx32p, s->prob.p.tx32p, sizeof(p->tx32p));

  3716.         memcpy(p->tx16p, s->prob.p.tx16p, sizeof(p->tx16p));

  3717.         memcpy(p->tx8p,  s->prob.p.tx8p,  sizeof(p->tx8p));

  3718.         return;

  3719.     }

  3720. 

  3721.     // skip flag

  3722.     for (i = 0; i < 3; i++)

  3723.         adapt_prob(&p->skip[i], s->counts.skip[i][0], s->counts.skip[i][1], 20, 128);

  3724. 

  3725.     // intra/inter flag

  3726.     for (i = 0; i < 4; i++)

  3727.         adapt_prob(&p->intra[i], s->counts.intra[i][0], s->counts.intra[i][1], 20, 128);

  3728. 

  3729.     // comppred flag

  3730.     if (s->s.h.comppredmode == PRED_SWITCHABLE) {

  3731.       for (i = 0; i < 5; i++)

  3732.           adapt_prob(&p->comp[i], s->counts.comp[i][0], s->counts.comp[i][1], 20, 128);

  3733.     }

  3734. 

  3735.     // reference frames

  3736.     if (s->s.h.comppredmode != PRED_SINGLEREF) {

  3737.       for (i = 0; i < 5; i++)

  3738.           adapt_prob(&p->comp_ref[i], s->counts.comp_ref[i][0],

  3739.                      s->counts.comp_ref[i][1], 20, 128);

  3740.     }

  3741. 

  3742.     if (s->s.h.comppredmode != PRED_COMPREF) {

  3743.       for (i = 0; i < 5; i++) {

  3744.           uint8_t *pp = p->single_ref[i];

  3745.           unsigned (*c)[2] = s->counts.single_ref[i];

  3746. 

  3747.           adapt_prob(&pp[0], c[0][0], c[0][1], 20, 128);

  3748.           adapt_prob(&pp[1], c[1][0], c[1][1], 20, 128);

  3749.       }

  3750.     }

  3751. 

  3752.     // block partitioning

  3753.     for (i = 0; i < 4; i++)

  3754.         for (j = 0; j < 4; j++) {

  3755.             uint8_t *pp = p->partition[i][j];

  3756.             unsigned *c = s->counts.partition[i][j];

  3757. 

  3758.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3759.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3760.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3761.         }

  3762. 

  3763.     // tx size

  3764.     if (s->s.h.txfmmode == TX_SWITCHABLE) {

  3765.       for (i = 0; i < 2; i++) {

  3766.           unsigned *c16 = s->counts.tx16p[i], *c32 = s->counts.tx32p[i];

  3767. 

  3768.           adapt_prob(&p->tx8p[i], s->counts.tx8p[i][0], s->counts.tx8p[i][1], 20, 128);

  3769.           adapt_prob(&p->tx16p[i][0], c16[0], c16[1] + c16[2], 20, 128);

  3770.           adapt_prob(&p->tx16p[i][1], c16[1], c16[2], 20, 128);

  3771.           adapt_prob(&p->tx32p[i][0], c32[0], c32[1] + c32[2] + c32[3], 20, 128);

  3772.           adapt_prob(&p->tx32p[i][1], c32[1], c32[2] + c32[3], 20, 128);

  3773.           adapt_prob(&p->tx32p[i][2], c32[2], c32[3], 20, 128);

  3774.       }

  3775.     }

  3776. 

  3777.     // interpolation filter

  3778.     if (s->s.h.filtermode == FILTER_SWITCHABLE) {

  3779.         for (i = 0; i < 4; i++) {

  3780.             uint8_t *pp = p->filter[i];

  3781.             unsigned *c = s->counts.filter[i];

  3782. 

  3783.             adapt_prob(&pp[0], c[0], c[1] + c[2], 20, 128);

  3784.             adapt_prob(&pp[1], c[1], c[2], 20, 128);

  3785.         }

  3786.     }

  3787. 

  3788.     // inter modes

  3789.     for (i = 0; i < 7; i++) {

  3790.         uint8_t *pp = p->mv_mode[i];

  3791.         unsigned *c = s->counts.mv_mode[i];

  3792. 

  3793.         adapt_prob(&pp[0], c[2], c[1] + c[0] + c[3], 20, 128);

  3794.         adapt_prob(&pp[1], c[0], c[1] + c[3], 20, 128);

  3795.         adapt_prob(&pp[2], c[1], c[3], 20, 128);

  3796.     }

  3797. 

  3798.     // mv joints

  3799.     {

  3800.         uint8_t *pp = p->mv_joint;

  3801.         unsigned *c = s->counts.mv_joint;

  3802. 

  3803.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3804.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3805.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3806.     }

  3807. 

  3808.     // mv components

  3809.     for (i = 0; i < 2; i++) {

  3810.         uint8_t *pp;

  3811.         unsigned *c, (*c2)[2], sum;

  3812. 

  3813.         adapt_prob(&p->mv_comp[i].sign, s->counts.mv_comp[i].sign[0],

  3814.                    s->counts.mv_comp[i].sign[1], 20, 128);

  3815. 

  3816.         pp = p->mv_comp[i].classes;

  3817.         c = s->counts.mv_comp[i].classes;

  3818.         sum = c[1] + c[2] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9] + c[10];

  3819.         adapt_prob(&pp[0], c[0], sum, 20, 128);

  3820.         sum -= c[1];

  3821.         adapt_prob(&pp[1], c[1], sum, 20, 128);

  3822.         sum -= c[2] + c[3];

  3823.         adapt_prob(&pp[2], c[2] + c[3], sum, 20, 128);

  3824.         adapt_prob(&pp[3], c[2], c[3], 20, 128);

  3825.         sum -= c[4] + c[5];

  3826.         adapt_prob(&pp[4], c[4] + c[5], sum, 20, 128);

  3827.         adapt_prob(&pp[5], c[4], c[5], 20, 128);

  3828.         sum -= c[6];

  3829.         adapt_prob(&pp[6], c[6], sum, 20, 128);

  3830.         adapt_prob(&pp[7], c[7] + c[8], c[9] + c[10], 20, 128);

  3831.         adapt_prob(&pp[8], c[7], c[8], 20, 128);

  3832.         adapt_prob(&pp[9], c[9], c[10], 20, 128);

  3833. 

  3834.         adapt_prob(&p->mv_comp[i].class0, s->counts.mv_comp[i].class0[0],

  3835.                    s->counts.mv_comp[i].class0[1], 20, 128);

  3836.         pp = p->mv_comp[i].bits;

  3837.         c2 = s->counts.mv_comp[i].bits;

  3838.         for (j = 0; j < 10; j++)

  3839.             adapt_prob(&pp[j], c2[j][0], c2[j][1], 20, 128);

  3840. 

  3841.         for (j = 0; j < 2; j++) {

  3842.             pp = p->mv_comp[i].class0_fp[j];

  3843.             c = s->counts.mv_comp[i].class0_fp[j];

  3844.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3845.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3846.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3847.         }

  3848.         pp = p->mv_comp[i].fp;

  3849.         c = s->counts.mv_comp[i].fp;

  3850.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3851.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3852.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3853. 

  3854.         if (s->s.h.highprecisionmvs) {

  3855.             adapt_prob(&p->mv_comp[i].class0_hp, s->counts.mv_comp[i].class0_hp[0],

  3856.                        s->counts.mv_comp[i].class0_hp[1], 20, 128);

  3857.             adapt_prob(&p->mv_comp[i].hp, s->counts.mv_comp[i].hp[0],

  3858.                        s->counts.mv_comp[i].hp[1], 20, 128);

  3859.         }

  3860.     }

  3861. 

  3862.     // y intra modes

  3863.     for (i = 0; i < 4; i++) {

  3864.         uint8_t *pp = p->y_mode[i];

  3865.         unsigned *c = s->counts.y_mode[i], sum, s2;

  3866. 

  3867.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3868.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3869.         sum -= c[TM_VP8_PRED];

  3870.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3871.         sum -= c[VERT_PRED];

  3872.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3873.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3874.         sum -= s2;

  3875.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3876.         s2 -= c[HOR_PRED];

  3877.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3878.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3879.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3880.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3881.         sum -= c[VERT_LEFT_PRED];

  3882.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3883.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3884.     }

  3885. 

  3886.     // uv intra modes

  3887.     for (i = 0; i < 10; i++) {

  3888.         uint8_t *pp = p->uv_mode[i];

  3889.         unsigned *c = s->counts.uv_mode[i], sum, s2;

  3890. 

  3891.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3892.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3893.         sum -= c[TM_VP8_PRED];

  3894.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3895.         sum -= c[VERT_PRED];

  3896.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3897.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3898.         sum -= s2;

  3899.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3900.         s2 -= c[HOR_PRED];

  3901.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3902.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3903.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3904.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3905.         sum -= c[VERT_LEFT_PRED];

  3906.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3907.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3908.     }

  3909. }

  3910. 

  3911. static void free_buffers(VP9Context *s)

  3912. {

  3913.     av_freep(&s->intra_pred_data[0]);

  3914.     av_freep(&s->b_base);

  3915.     av_freep(&s->block_base);

  3916. }

  3917. 

  3918. static av_cold int vp9_decode_free(AVCodecContext *ctx)

  3919. {

  3920.     VP9Context *s = ctx->priv_data;

  3921.     int i;

  3922. 

  3923.     for (i = 0; i < 3; i++) {

  3924.         if (s->s.frames[i].tf.f->buf[0])

  3925.             vp9_unref_frame(ctx, &s->s.frames[i]);

  3926.         av_frame_free(&s->s.frames[i].tf.f);

  3927.     }

  3928.     for (i = 0; i < 8; i++) {

  3929.         if (s->s.refs[i].f->buf[0])

  3930.             ff_thread_release_buffer(ctx, &s->s.refs[i]);

  3931.         av_frame_free(&s->s.refs[i].f);

  3932.         if (s->next_refs[i].f->buf[0])

  3933.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3934.         av_frame_free(&s->next_refs[i].f);

  3935.     }

  3936.     free_buffers(s);

  3937.     av_freep(&s->c_b);

  3938.     s->c_b_size = 0;

  3939. 

  3940.     return 0;

  3941. }

  3942. 

  3943. 

  3944. static int vp9_decode_frame(AVCodecContext *ctx, void *frame,

  3945.                             int *got_frame, AVPacket *pkt)

  3946. {

  3947.     const uint8_t *data = pkt->data;

  3948.     int size = pkt->size;

  3949.     VP9Context *s = ctx->priv_data;

  3950.     int res, tile_row, tile_col, i, ref, row, col;

  3951.     int retain_segmap_ref = s->s.frames[REF_FRAME_SEGMAP].segmentation_map &&

  3952.                             (!s->s.h.segmentation.enabled || !s->s.h.segmentation.update_map);

  3953.     ptrdiff_t yoff, uvoff, ls_y, ls_uv;

  3954.     AVFrame *f;

  3955.     int bytesperpixel;

  3956. 

  3957.     if ((res = decode_frame_header(ctx, data, size, &ref)) < 0) {

  3958.         return res;

  3959.     } else if (res == 0) {

  3960.         if (!s->s.refs[ref].f->buf[0]) {

  3961.             av_log(ctx, AV_LOG_ERROR, "Requested reference %d not available\n", ref);

  3962.             return AVERROR_INVALIDDATA;

  3963.         }

  3964.         if ((res = av_frame_ref(frame, s->s.refs[ref].f)) < 0)

  3965.             return res;

  3966.         ((AVFrame *)frame)->pkt_pts = pkt->pts;

  3967.         ((AVFrame *)frame)->pkt_dts = pkt->dts;

  3968.         for (i = 0; i < 8; i++) {

  3969.             if (s->next_refs[i].f->buf[0])

  3970.                 ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3971.             if (s->s.refs[i].f->buf[0] &&

  3972.                 (res = ff_thread_ref_frame(&s->next_refs[i], &s->s.refs[i])) < 0)

  3973.                 return res;

  3974.         }

  3975.         *got_frame = 1;

  3976.         return pkt->size;

  3977.     }

  3978.     data += res;

  3979.     size -= res;

  3980. 

  3981.     if (!retain_segmap_ref || s->s.h.keyframe || s->s.h.intraonly) {

  3982.         if (s->s.frames[REF_FRAME_SEGMAP].tf.f->buf[0])

  3983.             vp9_unref_frame(ctx, &s->s.frames[REF_FRAME_SEGMAP]);

  3984.         if (!s->s.h.keyframe && !s->s.h.intraonly && !s->s.h.errorres && s->s.frames[CUR_FRAME].tf.f->buf[0] &&

  3985.             (res = vp9_ref_frame(ctx, &s->s.frames[REF_FRAME_SEGMAP], &s->s.frames[CUR_FRAME])) < 0)

  3986.             return res;

  3987.     }

  3988.     if (s->s.frames[REF_FRAME_MVPAIR].tf.f->buf[0])

  3989.         vp9_unref_frame(ctx, &s->s.frames[REF_FRAME_MVPAIR]);

  3990.     if (!s->s.h.intraonly && !s->s.h.keyframe && !s->s.h.errorres && s->s.frames[CUR_FRAME].tf.f->buf[0] &&

  3991.         (res = vp9_ref_frame(ctx, &s->s.frames[REF_FRAME_MVPAIR], &s->s.frames[CUR_FRAME])) < 0)

  3992.         return res;

  3993.     if (s->s.frames[CUR_FRAME].tf.f->buf[0])

  3994.         vp9_unref_frame(ctx, &s->s.frames[CUR_FRAME]);

  3995.     if ((res = vp9_alloc_frame(ctx, &s->s.frames[CUR_FRAME])) < 0)

  3996.         return res;

  3997.     f = s->s.frames[CUR_FRAME].tf.f;

  3998.     f->key_frame = s->s.h.keyframe;

  3999.     f->pict_type = (s->s.h.keyframe || s->s.h.intraonly) ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;

  4000.     ls_y = f->linesize[0];

  4001.     ls_uv =f->linesize[1];

  4002. 

  4003.     if (s->s.frames[REF_FRAME_SEGMAP].tf.f->buf[0] &&

  4004.         (s->s.frames[REF_FRAME_MVPAIR].tf.f->width  != s->s.frames[CUR_FRAME].tf.f->width ||

  4005.          s->s.frames[REF_FRAME_MVPAIR].tf.f->height != s->s.frames[CUR_FRAME].tf.f->height)) {

  4006.         vp9_unref_frame(ctx, &s->s.frames[REF_FRAME_SEGMAP]);

  4007.     }

  4008. 

  4009.     // ref frame setup

  4010.     for (i = 0; i < 8; i++) {

  4011.         if (s->next_refs[i].f->buf[0])

  4012.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  4013.         if (s->s.h.refreshrefmask & (1 << i)) {

  4014.             res = ff_thread_ref_frame(&s->next_refs[i], &s->s.frames[CUR_FRAME].tf);

  4015.         } else if (s->s.refs[i].f->buf[0]) {

  4016.             res = ff_thread_ref_frame(&s->next_refs[i], &s->s.refs[i]);

  4017.         }

  4018.         if (res < 0)

  4019.             return res;

  4020.     }

  4021. 

  4022.     if (ctx->hwaccel) {

  4023.         res = ctx->hwaccel->start_frame(ctx, NULL, 0);

  4024.         if (res < 0)

  4025.             return res;

  4026.         res = ctx->hwaccel->decode_slice(ctx, pkt->data, pkt->size);

  4027.         if (res < 0)

  4028.             return res;

  4029.         res = ctx->hwaccel->end_frame(ctx);

  4030.         if (res < 0)

  4031.             return res;

  4032.         goto finish;

  4033.     }

  4034. 

  4035.     // main tile decode loop

  4036.     bytesperpixel = s->bytesperpixel;

  4037.     memset(s->above_partition_ctx, 0, s->cols);

  4038.     memset(s->above_skip_ctx, 0, s->cols);

  4039.     if (s->s.h.keyframe || s->s.h.intraonly) {

  4040.         memset(s->above_mode_ctx, DC_PRED, s->cols * 2);

  4041.     } else {

  4042.         memset(s->above_mode_ctx, NEARESTMV, s->cols);

  4043.     }

  4044.     memset(s->above_y_nnz_ctx, 0, s->sb_cols * 16);

  4045.     memset(s->above_uv_nnz_ctx[0], 0, s->sb_cols * 16 >> s->ss_h);

  4046.     memset(s->above_uv_nnz_ctx[1], 0, s->sb_cols * 16 >> s->ss_h);

  4047.     memset(s->above_segpred_ctx, 0, s->cols);

  4048.     s->pass = s->s.frames[CUR_FRAME].uses_2pass =

  4049.         ctx->active_thread_type == FF_THREAD_FRAME && s->s.h.refreshctx && !s->s.h.parallelmode;

  4050.     if ((res = update_block_buffers(ctx)) < 0) {

  4051.         av_log(ctx, AV_LOG_ERROR,

  4052.                "Failed to allocate block buffers\n");

  4053.         return res;

  4054.     }

  4055.     if (s->s.h.refreshctx && s->s.h.parallelmode) {

  4056.         int j, k, l, m;

  4057. 

  4058.         for (i = 0; i < 4; i++) {

  4059.             for (j = 0; j < 2; j++)

  4060.                 for (k = 0; k < 2; k++)

  4061.                     for (l = 0; l < 6; l++)

  4062.                         for (m = 0; m < 6; m++)

  4063.                             memcpy(s->prob_ctx[s->s.h.framectxid].coef[i][j][k][l][m],

  4064.                                    s->prob.coef[i][j][k][l][m], 3);

  4065.             if (s->s.h.txfmmode == i)

  4066.                 break;

  4067.         }

  4068.         s->prob_ctx[s->s.h.framectxid].p = s->prob.p;

  4069.         ff_thread_finish_setup(ctx);

  4070.     } else if (!s->s.h.refreshctx) {

  4071.         ff_thread_finish_setup(ctx);

  4072.     }

  4073. 

  4074.     do {

  4075.         yoff = uvoff = 0;

  4076.         s->b = s->b_base;

  4077.         s->block = s->block_base;

  4078.         s->uvblock[0] = s->uvblock_base[0];

  4079.         s->uvblock[1] = s->uvblock_base[1];

  4080.         s->eob = s->eob_base;

  4081.         s->uveob[0] = s->uveob_base[0];

  4082.         s->uveob[1] = s->uveob_base[1];

  4083. 

  4084.         for (tile_row = 0; tile_row < s->s.h.tiling.tile_rows; tile_row++) {

  4085.             set_tile_offset(&s->tile_row_start, &s->tile_row_end,

  4086.                             tile_row, s->s.h.tiling.log2_tile_rows, s->sb_rows);

  4087.             if (s->pass != 2) {

  4088.                 for (tile_col = 0; tile_col < s->s.h.tiling.tile_cols; tile_col++) {

  4089.                     int64_t tile_size;

  4090. 

  4091.                     if (tile_col == s->s.h.tiling.tile_cols - 1 &&

  4092.                         tile_row == s->s.h.tiling.tile_rows - 1) {

  4093.                         tile_size = size;

  4094.                     } else {

  4095.                         tile_size = AV_RB32(data);

  4096.                         data += 4;

  4097.                         size -= 4;

  4098.                     }

  4099.                     if (tile_size > size) {

  4100.                         ff_thread_report_progress(&s->s.frames[CUR_FRAME].tf, INT_MAX, 0);

  4101.                         return AVERROR_INVALIDDATA;

  4102.                     }

  4103.                     ff_vp56_init_range_decoder(&s->c_b[tile_col], data, tile_size);

  4104.                     if (vp56_rac_get_prob_branchy(&s->c_b[tile_col], 128)) { // marker bit

  4105.                         ff_thread_report_progress(&s->s.frames[CUR_FRAME].tf, INT_MAX, 0);

  4106.                         return AVERROR_INVALIDDATA;

  4107.                     }

  4108.                     data += tile_size;

  4109.                     size -= tile_size;

  4110.                 }

  4111.             }

  4112. 

  4113.             for (row = s->tile_row_start; row < s->tile_row_end;

  4114.                  row += 8, yoff += ls_y * 64, uvoff += ls_uv * 64 >> s->ss_v) {

  4115.                 struct VP9Filter *lflvl_ptr = s->lflvl;

  4116.                 ptrdiff_t yoff2 = yoff, uvoff2 = uvoff;

  4117. 

  4118.                 for (tile_col = 0; tile_col < s->s.h.tiling.tile_cols; tile_col++) {

  4119.                     set_tile_offset(&s->tile_col_start, &s->tile_col_end,

  4120.                                     tile_col, s->s.h.tiling.log2_tile_cols, s->sb_cols);

  4121. 

  4122.                     if (s->pass != 2) {

  4123.                         memset(s->left_partition_ctx, 0, 8);

  4124.                         memset(s->left_skip_ctx, 0, 8);

  4125.                         if (s->s.h.keyframe || s->s.h.intraonly) {

  4126.                             memset(s->left_mode_ctx, DC_PRED, 16);

  4127.                         } else {

  4128.                             memset(s->left_mode_ctx, NEARESTMV, 8);

  4129.                         }

  4130.                         memset(s->left_y_nnz_ctx, 0, 16);

  4131.                         memset(s->left_uv_nnz_ctx, 0, 32);

  4132.                         memset(s->left_segpred_ctx, 0, 8);

  4133. 

  4134.                         memcpy(&s->c, &s->c_b[tile_col], sizeof(s->c));

  4135.                     }

  4136. 

  4137.                     for (col = s->tile_col_start;

  4138.                          col < s->tile_col_end;

  4139.                          col += 8, yoff2 += 64 * bytesperpixel,

  4140.                          uvoff2 += 64 * bytesperpixel >> s->ss_h, lflvl_ptr++) {

  4141.                         // FIXME integrate with lf code (i.e. zero after each

  4142.                         // use, similar to invtxfm coefficients, or similar)

  4143.                         if (s->pass != 1) {

  4144.                             memset(lflvl_ptr->mask, 0, sizeof(lflvl_ptr->mask));

  4145.                         }

  4146. 

  4147.                         if (s->pass == 2) {

  4148.                             decode_sb_mem(ctx, row, col, lflvl_ptr,

  4149.                                           yoff2, uvoff2, BL_64X64);

  4150.                         } else {

  4151.                             decode_sb(ctx, row, col, lflvl_ptr,

  4152.                                       yoff2, uvoff2, BL_64X64);

  4153.                         }

  4154.                     }

  4155.                     if (s->pass != 2) {

  4156.                         memcpy(&s->c_b[tile_col], &s->c, sizeof(s->c));

  4157.                     }

  4158.                 }

  4159. 

  4160.                 if (s->pass == 1) {

  4161.                     continue;

  4162.                 }

  4163. 

  4164.                 // backup pre-loopfilter reconstruction data for intra

  4165.                 // prediction of next row of sb64s

  4166.                 if (row + 8 < s->rows) {

  4167.                     memcpy(s->intra_pred_data[0],

  4168.                            f->data[0] + yoff + 63 * ls_y,

  4169.                            8 * s->cols * bytesperpixel);

  4170.                     memcpy(s->intra_pred_data[1],

  4171.                            f->data[1] + uvoff + ((64 >> s->ss_v) - 1) * ls_uv,

  4172.                            8 * s->cols * bytesperpixel >> s->ss_h);

  4173.                     memcpy(s->intra_pred_data[2],

  4174.                            f->data[2] + uvoff + ((64 >> s->ss_v) - 1) * ls_uv,

  4175.                            8 * s->cols * bytesperpixel >> s->ss_h);

  4176.                 }

  4177. 

  4178.                 // loopfilter one row

  4179.                 if (s->s.h.filter.level) {

  4180.                     yoff2 = yoff;

  4181.                     uvoff2 = uvoff;

  4182.                     lflvl_ptr = s->lflvl;

  4183.                     for (col = 0; col < s->cols;

  4184.                          col += 8, yoff2 += 64 * bytesperpixel,

  4185.                          uvoff2 += 64 * bytesperpixel >> s->ss_h, lflvl_ptr++) {

  4186.                         loopfilter_sb(ctx, lflvl_ptr, row, col, yoff2, uvoff2);

  4187.                     }

  4188.                 }

  4189. 

  4190.                 // FIXME maybe we can make this more finegrained by running the

  4191.                 // loopfilter per-block instead of after each sbrow

  4192.                 // In fact that would also make intra pred left preparation easier?

  4193.                 ff_thread_report_progress(&s->s.frames[CUR_FRAME].tf, row >> 3, 0);

  4194.             }

  4195.         }

  4196. 

  4197.         if (s->pass < 2 && s->s.h.refreshctx && !s->s.h.parallelmode) {

  4198.             adapt_probs(s);

  4199.             ff_thread_finish_setup(ctx);

  4200.         }

  4201.     } while (s->pass++ == 1);

  4202.     ff_thread_report_progress(&s->s.frames[CUR_FRAME].tf, INT_MAX, 0);

  4203. 

  4204. finish:

  4205.     // ref frame setup

  4206.     for (i = 0; i < 8; i++) {

  4207.         if (s->s.refs[i].f->buf[0])

  4208.             ff_thread_release_buffer(ctx, &s->s.refs[i]);

  4209.         if (s->next_refs[i].f->buf[0] &&

  4210.             (res = ff_thread_ref_frame(&s->s.refs[i], &s->next_refs[i])) < 0)

  4211.             return res;

  4212.     }

  4213. 

  4214.     if (!s->s.h.invisible) {

  4215.         if ((res = av_frame_ref(frame, s->s.frames[CUR_FRAME].tf.f)) < 0)

  4216.             return res;

  4217.         *got_frame = 1;

  4218.     }

  4219. 

  4220.     return pkt->size;

  4221. }

  4222. 

  4223. static void vp9_decode_flush(AVCodecContext *ctx)

  4224. {

  4225.     VP9Context *s = ctx->priv_data;

  4226.     int i;

  4227. 

  4228.     for (i = 0; i < 3; i++)

  4229.         vp9_unref_frame(ctx, &s->s.frames[i]);

  4230.     for (i = 0; i < 8; i++)

  4231.         ff_thread_release_buffer(ctx, &s->s.refs[i]);

  4232. }

  4233. 

  4234. static int init_frames(AVCodecContext *ctx)

  4235. {

  4236.     VP9Context *s = ctx->priv_data;

  4237.     int i;

  4238. 

  4239.     for (i = 0; i < 3; i++) {

  4240.         s->s.frames[i].tf.f = av_frame_alloc();

  4241.         if (!s->s.frames[i].tf.f) {

  4242.             vp9_decode_free(ctx);

  4243.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  4244.             return AVERROR(ENOMEM);

  4245.         }

  4246.     }

  4247.     for (i = 0; i < 8; i++) {

  4248.         s->s.refs[i].f = av_frame_alloc();

  4249.         s->next_refs[i].f = av_frame_alloc();

  4250.         if (!s->s.refs[i].f || !s->next_refs[i].f) {

  4251.             vp9_decode_free(ctx);

  4252.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  4253.             return AVERROR(ENOMEM);

  4254.         }

  4255.     }

  4256. 

  4257.     return 0;

  4258. }

  4259. 

  4260. static av_cold int vp9_decode_init(AVCodecContext *ctx)

  4261. {

  4262.     VP9Context *s = ctx->priv_data;

  4263. 

  4264.     ctx->internal->allocate_progress = 1;

  4265.     s->last_bpp = 0;

  4266.     s->s.h.filter.sharpness = -1;

  4267. 

  4268.     return init_frames(ctx);

  4269. }

  4270. 

  4271. #if HAVE_THREADS

  4272. static av_cold int vp9_decode_init_thread_copy(AVCodecContext *avctx)

  4273. {

  4274.     return init_frames(avctx);

  4275. }

  4276. 

  4277. static int vp9_decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src)

  4278. {

  4279.     int i, res;

  4280.     VP9Context *s = dst->priv_data, *ssrc = src->priv_data;

  4281. 

  4282.     // detect size changes in other threads

  4283.     if (s->intra_pred_data[0] &&

  4284.         (!ssrc->intra_pred_data[0] || s->cols != ssrc->cols ||

  4285.          s->rows != ssrc->rows || s->bpp != ssrc->bpp || s->pix_fmt != ssrc->pix_fmt)) {

  4286.         free_buffers(s);

  4287.     }

  4288. 

  4289.     for (i = 0; i < 3; i++) {

  4290.         if (s->s.frames[i].tf.f->buf[0])

  4291.             vp9_unref_frame(dst, &s->s.frames[i]);

  4292.         if (ssrc->s.frames[i].tf.f->buf[0]) {

  4293.             if ((res = vp9_ref_frame(dst, &s->s.frames[i], &ssrc->s.frames[i])) < 0)

  4294.                 return res;

  4295.         }

  4296.     }

  4297.     for (i = 0; i < 8; i++) {

  4298.         if (s->s.refs[i].f->buf[0])

  4299.             ff_thread_release_buffer(dst, &s->s.refs[i]);

  4300.         if (ssrc->next_refs[i].f->buf[0]) {

  4301.             if ((res = ff_thread_ref_frame(&s->s.refs[i], &ssrc->next_refs[i])) < 0)

  4302.                 return res;

  4303.         }

  4304.     }

  4305. 

  4306.     s->s.h.invisible = ssrc->s.h.invisible;

  4307.     s->s.h.keyframe = ssrc->s.h.keyframe;

  4308.     s->s.h.intraonly = ssrc->s.h.intraonly;

  4309.     s->ss_v = ssrc->ss_v;

  4310.     s->ss_h = ssrc->ss_h;

  4311.     s->s.h.segmentation.enabled = ssrc->s.h.segmentation.enabled;

  4312.     s->s.h.segmentation.update_map = ssrc->s.h.segmentation.update_map;

  4313.     s->s.h.segmentation.absolute_vals = ssrc->s.h.segmentation.absolute_vals;

  4314.     s->bytesperpixel = ssrc->bytesperpixel;

  4315.     s->bpp = ssrc->bpp;

  4316.     s->bpp_index = ssrc->bpp_index;

  4317.     s->pix_fmt = ssrc->pix_fmt;

  4318.     memcpy(&s->prob_ctx, &ssrc->prob_ctx, sizeof(s->prob_ctx));

  4319.     memcpy(&s->s.h.lf_delta, &ssrc->s.h.lf_delta, sizeof(s->s.h.lf_delta));

  4320.     memcpy(&s->s.h.segmentation.feat, &ssrc->s.h.segmentation.feat,

  4321.            sizeof(s->s.h.segmentation.feat));

  4322. 

  4323.     return 0;

  4324. }

  4325. #endif

  4326. 

  4327. static const AVProfile profiles[] = {

  4328.     { FF_PROFILE_VP9_0, "Profile 0" },

  4329.     { FF_PROFILE_VP9_1, "Profile 1" },

  4330.     { FF_PROFILE_VP9_2, "Profile 2" },

  4331.     { FF_PROFILE_VP9_3, "Profile 3" },

  4332.     { FF_PROFILE_UNKNOWN },

  4333. };

  4334. 

  4335. AVCodec ff_vp9_decoder = {

  4336.     .name                  = "vp9",

  4337.     .long_name             = NULL_IF_CONFIG_SMALL("Google VP9"),

  4338.     .type                  = AVMEDIA_TYPE_VIDEO,

  4339.     .id                    = AV_CODEC_ID_VP9,

  4340.     .priv_data_size        = sizeof(VP9Context),

  4341.     .init                  = vp9_decode_init,

  4342.     .close                 = vp9_decode_free,

  4343.     .decode                = vp9_decode_frame,

  4344.     .capabilities          = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_FRAME_THREADS,

  4345.     .flush                 = vp9_decode_flush,

  4346.     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(vp9_decode_init_thread_copy),

  4347.     .update_thread_context = ONLY_IF_THREADS_ENABLED(vp9_decode_update_thread_context),

  4348.     .profiles              = NULL_IF_CONFIG_SMALL(profiles),

  4349. };