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.
72008 Commits
32 Branches
313MB
Tree: 1e4a77d001
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '1e4a77d001'
${ noResults }
FFmpeg/libavcodec/vp9.c

4312 lines
170KB
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. enum CompPredMode {

  39.     PRED_SINGLEREF,

  40.     PRED_COMPREF,

  41.     PRED_SWITCHABLE,

  42. };

  43. 

  44. enum BlockLevel {

  45.     BL_64X64,

  46.     BL_32X32,

  47.     BL_16X16,

  48.     BL_8X8,

  49. };

  50. 

  51. enum BlockSize {

  52.     BS_64x64,

  53.     BS_64x32,

  54.     BS_32x64,

  55.     BS_32x32,

  56.     BS_32x16,

  57.     BS_16x32,

  58.     BS_16x16,

  59.     BS_16x8,

  60.     BS_8x16,

  61.     BS_8x8,

  62.     BS_8x4,

  63.     BS_4x8,

  64.     BS_4x4,

  65.     N_BS_SIZES,

  66. };

  67. 

  68. struct VP9mvrefPair {

  69.     VP56mv mv[2];

  70.     int8_t ref[2];

  71. };

  72. 

  73. typedef struct VP9Frame {

  74.     ThreadFrame tf;

  75.     AVBufferRef *extradata;

  76.     uint8_t *segmentation_map;

  77.     struct VP9mvrefPair *mv;

  78.     int uses_2pass;

  79. } VP9Frame;

  80. 

  81. struct VP9Filter {

  82.     uint8_t level[8 * 8];

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

  84.                               [8 /* rows */][4 /* 0=16, 1=8, 2=4, 3=inner4 */];

  85. };

  86. 

  87. typedef struct VP9Block {

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

  89.     enum FilterMode filter;

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

  91.     enum BlockSize bs;

  92.     enum TxfmMode tx, uvtx;

  93.     enum BlockLevel bl;

  94.     enum BlockPartition bp;

  95. } VP9Block;

  96. 

  97. typedef struct VP9Context {

  98.     VP9DSPContext dsp;

  99.     VideoDSPContext vdsp;

  100.     GetBitContext gb;

  101.     VP56RangeCoder c;

  102.     VP56RangeCoder *c_b;

  103.     unsigned c_b_size;

  104.     VP9Block *b_base, *b;

  105.     int pass;

  106.     int row, row7, col, col7;

  107.     uint8_t *dst[3];

  108.     ptrdiff_t y_stride, uv_stride;

  109. 

  110.     // bitstream header

  111.     uint8_t keyframe, last_keyframe;

  112.     uint8_t last_bpp, bpp, bpp_index, bytesperpixel;

  113.     uint8_t invisible;

  114.     uint8_t use_last_frame_mvs;

  115.     uint8_t errorres;

  116.     uint8_t ss_h, ss_v;

  117.     uint8_t intraonly;

  118.     uint8_t resetctx;

  119.     uint8_t refreshrefmask;

  120.     uint8_t highprecisionmvs;

  121.     enum FilterMode filtermode;

  122.     uint8_t allowcompinter;

  123.     uint8_t fixcompref;

  124.     uint8_t refreshctx;

  125.     uint8_t parallelmode;

  126.     uint8_t framectxid;

  127.     uint8_t refidx[3];

  128.     uint8_t signbias[3];

  129.     uint8_t varcompref[2];

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

  131. #define CUR_FRAME 0

  132. #define REF_FRAME_MVPAIR 1

  133. #define REF_FRAME_SEGMAP 2

  134.     VP9Frame frames[3];

  135. 

  136.     struct {

  137.         uint8_t level;

  138.         int8_t sharpness;

  139.         uint8_t lim_lut[64];

  140.         uint8_t mblim_lut[64];

  141.     } filter;

  142.     struct {

  143.         uint8_t enabled;

  144.         int8_t mode[2];

  145.         int8_t ref[4];

  146.     } lf_delta;

  147.     uint8_t yac_qi;

  148.     int8_t ydc_qdelta, uvdc_qdelta, uvac_qdelta;

  149.     uint8_t lossless;

  150. #define MAX_SEGMENT 8

  151.     struct {

  152.         uint8_t enabled;

  153.         uint8_t temporal;

  154.         uint8_t absolute_vals;

  155.         uint8_t update_map;

  156.         uint8_t ignore_refmap;

  157.         struct {

  158.             uint8_t q_enabled;

  159.             uint8_t lf_enabled;

  160.             uint8_t ref_enabled;

  161.             uint8_t skip_enabled;

  162.             uint8_t ref_val;

  163.             int16_t q_val;

  164.             int8_t lf_val;

  165.             int16_t qmul[2][2];

  166.             uint8_t lflvl[4][2];

  167.         } feat[MAX_SEGMENT];

  168.     } segmentation;

  169.     struct {

  170.         unsigned log2_tile_cols, log2_tile_rows;

  171.         unsigned tile_cols, tile_rows;

  172.         unsigned tile_row_start, tile_row_end, tile_col_start, tile_col_end;

  173.     } tiling;

  174.     unsigned sb_cols, sb_rows, rows, cols;

  175.     struct {

  176.         prob_context p;

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

  178.     } prob_ctx[4];

  179.     struct {

  180.         prob_context p;

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

  182.         uint8_t seg[7];

  183.         uint8_t segpred[3];

  184.     } prob;

  185.     struct {

  186.         unsigned y_mode[4][10];

  187.         unsigned uv_mode[10][10];

  188.         unsigned filter[4][3];

  189.         unsigned mv_mode[7][4];

  190.         unsigned intra[4][2];

  191.         unsigned comp[5][2];

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

  193.         unsigned comp_ref[5][2];

  194.         unsigned tx32p[2][4];

  195.         unsigned tx16p[2][3];

  196.         unsigned tx8p[2][2];

  197.         unsigned skip[3][2];

  198.         unsigned mv_joint[4];

  199.         struct {

  200.             unsigned sign[2];

  201.             unsigned classes[11];

  202.             unsigned class0[2];

  203.             unsigned bits[10][2];

  204.             unsigned class0_fp[2][4];

  205.             unsigned fp[4];

  206.             unsigned class0_hp[2];

  207.             unsigned hp[2];

  208.         } mv_comp[2];

  209.         unsigned partition[4][4][4];

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

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

  212.     } counts;

  213.     enum TxfmMode txfmmode;

  214.     enum CompPredMode comppredmode;

  215. 

  216.     // contextual (left/above) cache

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

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

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

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

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

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

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

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

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

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

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

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

  229.     uint8_t *above_partition_ctx;

  230.     uint8_t *above_mode_ctx;

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

  232.     uint8_t *above_y_nnz_ctx;

  233.     uint8_t *above_uv_nnz_ctx[2];

  234.     uint8_t *above_skip_ctx; // 1bit

  235.     uint8_t *above_txfm_ctx; // 2bit

  236.     uint8_t *above_segpred_ctx; // 1bit

  237.     uint8_t *above_intra_ctx; // 1bit

  238.     uint8_t *above_comp_ctx; // 1bit

  239.     uint8_t *above_ref_ctx; // 2bit

  240.     uint8_t *above_filter_ctx;

  241.     VP56mv (*above_mv_ctx)[2];

  242. 

  243.     // whole-frame cache

  244.     uint8_t *intra_pred_data[3];

  245.     struct VP9Filter *lflvl;

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

  247. 

  248.     // block reconstruction intermediates

  249.     int block_alloc_using_2pass;

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

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

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

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

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

  255.     uint16_t mvscale[3][2];

  256.     uint8_t mvstep[3][2];

  257. } VP9Context;

  258. 

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

  260.     {

  261.         { 16, 16 }, { 16, 8 }, { 8, 16 }, { 8, 8 }, { 8, 4 }, { 4, 8 },

  262.         { 4, 4 }, { 4, 2 }, { 2, 4 }, { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 },

  263.     }, {

  264.         { 8, 8 }, { 8, 4 }, { 4, 8 }, { 4, 4 }, { 4, 2 }, { 2, 4 },

  265.         { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 }, { 1, 1 }, { 1, 1 }, { 1, 1 },

  266.     }

  267. };

  268. 

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

  270. {

  271.     VP9Context *s = ctx->priv_data;

  272.     int ret, sz;

  273. 

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

  275.         return ret;

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

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

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

  279.         return AVERROR(ENOMEM);

  280.     }

  281. 

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

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

  284. 

  285.     return 0;

  286. }

  287. 

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

  289. {

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

  291.     av_buffer_unref(&f->extradata);

  292. }

  293. 

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

  295. {

  296.     int res;

  297. 

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

  299.         return res;

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

  301.         vp9_unref_frame(ctx, dst);

  302.         return AVERROR(ENOMEM);

  303.     }

  304. 

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

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

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

  308. 

  309.     return 0;

  310. }

  311. 

  312. static int update_size(AVCodecContext *ctx, int w, int h, enum AVPixelFormat fmt)

  313. {

  314.     VP9Context *s = ctx->priv_data;

  315.     uint8_t *p;

  316.     int bytesperpixel = s->bytesperpixel;

  317. 

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

  319. 

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

  321.         return 0;

  322. 

  323.     ctx->width   = w;

  324.     ctx->height  = h;

  325.     ctx->pix_fmt = fmt;

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

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

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

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

  330. 

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

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

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

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

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

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

  337.     if (!p)

  338.         return AVERROR(ENOMEM);

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  356. #undef assign

  357. 

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

  359.     av_freep(&s->b_base);

  360.     av_freep(&s->block_base);

  361. 

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

  363.         ff_vp9dsp_init(&s->dsp, s->bpp);

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

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

  366.     }

  367. 

  368.     return 0;

  369. }

  370. 

  371. static int update_block_buffers(AVCodecContext *ctx)

  372. {

  373.     VP9Context *s = ctx->priv_data;

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

  375. 

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

  377.         return 0;

  378. 

  379.     av_free(s->b_base);

  380.     av_free(s->block_base);

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

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

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

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

  385. 

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

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

  388.                                     16 * 16 + 2 * chroma_eobs) * sbs);

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

  390.             return AVERROR(ENOMEM);

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

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

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

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

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

  396.     } else {

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

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

  399.                                    16 * 16 + 2 * chroma_eobs);

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

  401.             return AVERROR(ENOMEM);

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

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

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

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

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

  407.     }

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

  409. 

  410.     return 0;

  411. }

  412. 

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

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

  415. {

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

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

  418. }

  419. 

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

  421. {

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

  423. }

  424. 

  425. // differential forward probability updates

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

  427. {

  428.     static const int inv_map_table[254] = {

  429.           7,  20,  33,  46,  59,  72,  85,  98, 111, 124, 137, 150, 163, 176,

  430.         189, 202, 215, 228, 241, 254,   1,   2,   3,   4,   5,   6,   8,   9,

  431.          10,  11,  12,  13,  14,  15,  16,  17,  18,  19,  21,  22,  23,  24,

  432.          25,  26,  27,  28,  29,  30,  31,  32,  34,  35,  36,  37,  38,  39,

  433.          40,  41,  42,  43,  44,  45,  47,  48,  49,  50,  51,  52,  53,  54,

  434.          55,  56,  57,  58,  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,

  435.          70,  71,  73,  74,  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,

  436.          86,  87,  88,  89,  90,  91,  92,  93,  94,  95,  96,  97,  99, 100,

  437.         101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 112, 113, 114, 115,

  438.         116, 117, 118, 119, 120, 121, 122, 123, 125, 126, 127, 128, 129, 130,

  439.         131, 132, 133, 134, 135, 136, 138, 139, 140, 141, 142, 143, 144, 145,

  440.         146, 147, 148, 149, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160,

  441.         161, 162, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175,

  442.         177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 190, 191,

  443.         192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 204, 205, 206,

  444.         207, 208, 209, 210, 211, 212, 213, 214, 216, 217, 218, 219, 220, 221,

  445.         222, 223, 224, 225, 226, 227, 229, 230, 231, 232, 233, 234, 235, 236,

  446.         237, 238, 239, 240, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251,

  447.         252, 253,

  448.     };

  449.     int d;

  450. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  465. 

  466.     if (!vp8_rac_get(c)) {

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

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

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

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

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

  472.     } else {

  473.         d = vp8_rac_get_uint(c, 7);

  474.         if (d >= 65)

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

  476.         d += 64;

  477.     }

  478. 

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

  480.                     255 - inv_recenter_nonneg(inv_map_table[d], 255 - p);

  481. }

  482. 

  483. static enum AVPixelFormat read_colorspace_details(AVCodecContext *ctx)

  484. {

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

  486.         AVCOL_SPC_UNSPECIFIED, AVCOL_SPC_BT470BG, AVCOL_SPC_BT709, AVCOL_SPC_SMPTE170M,

  487.         AVCOL_SPC_SMPTE240M, AVCOL_SPC_BT2020_NCL, AVCOL_SPC_RESERVED, AVCOL_SPC_RGB,

  488.     };

  489.     VP9Context *s = ctx->priv_data;

  490.     enum AVPixelFormat res;

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

  492. 

  493.     s->bpp_index = bits;

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

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

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

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

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

  499.             AV_PIX_FMT_GBRP, AV_PIX_FMT_GBRP10, AV_PIX_FMT_GBRP12

  500.         };

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

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

  503.             res = pix_fmt_rgb[bits];

  504.             ctx->color_range = AVCOL_RANGE_JPEG;

  505.         } else {

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

  507.                    ctx->profile);

  508.             return AVERROR_INVALIDDATA;

  509.         }

  510.     } else {

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

  512.             { { AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV422P },

  513.               { AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV420P } },

  514.             { { AV_PIX_FMT_YUV444P10, AV_PIX_FMT_YUV422P10 },

  515.               { AV_PIX_FMT_YUV440P10, AV_PIX_FMT_YUV420P10 } },

  516.             { { AV_PIX_FMT_YUV444P12, AV_PIX_FMT_YUV422P12 },

  517.               { AV_PIX_FMT_YUV440P12, AV_PIX_FMT_YUV420P12 } }

  518.         };

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

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

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

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

  523.             if ((res = pix_fmt_for_ss[bits][s->ss_v][s->ss_h]) == AV_PIX_FMT_YUV420P) {

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

  525.                        ctx->profile);

  526.                 return AVERROR_INVALIDDATA;

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

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

  529.                        ctx->profile);

  530.                 return AVERROR_INVALIDDATA;

  531.             }

  532.         } else {

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

  534.             res = pix_fmt_for_ss[bits][1][1];

  535.         }

  536.     }

  537. 

  538.     return res;

  539. }

  540. 

  541. static int decode_frame_header(AVCodecContext *ctx,

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

  543. {

  544.     VP9Context *s = ctx->priv_data;

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

  546.     enum AVPixelFormat fmt = ctx->pix_fmt;

  547.     int last_invisible;

  548.     const uint8_t *data2;

  549. 

  550.     /* general header */

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

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

  553.         return res;

  554.     }

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

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

  557.         return AVERROR_INVALIDDATA;

  558.     }

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

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

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

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

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

  564.         return AVERROR_INVALIDDATA;

  565.     }

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

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

  568.         return 0;

  569.     }

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

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

  572.     last_invisible    = s->invisible;

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

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

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

  576.     if (s->keyframe) {

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

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

  579.             return AVERROR_INVALIDDATA;

  580.         }

  581.         if ((fmt = read_colorspace_details(ctx)) < 0)

  582.             return fmt;

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

  584.         s->refreshrefmask = 0xff;

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

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

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

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

  589.     } else {

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

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

  592.         if (s->intraonly) {

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

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

  595.                 return AVERROR_INVALIDDATA;

  596.             }

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

  598.                 if ((fmt = read_colorspace_details(ctx)) < 0)

  599.                     return fmt;

  600.             } else {

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

  602.                 s->bpp = 8;

  603.                 s->bpp_index = 0;

  604.                 s->bytesperpixel = 1;

  605.                 fmt = AV_PIX_FMT_YUV420P;

  606.                 ctx->colorspace = AVCOL_SPC_BT470BG;

  607.                 ctx->color_range = AVCOL_RANGE_JPEG;

  608.             }

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

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

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

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

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

  614.         } else {

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

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

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

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

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

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

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

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

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

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

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

  626.                 return AVERROR_INVALIDDATA;

  627.             }

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

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

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

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

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

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

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

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

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

  637.             } else {

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

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

  640.             }

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

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

  643.             // the _last_ frame

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

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

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

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

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

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

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

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

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

  653.             if (s->allowcompinter) {

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

  655.                     s->fixcompref    = 2;

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

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

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

  659.                     s->fixcompref    = 1;

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

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

  662.                 } else {

  663.                     s->fixcompref    = 0;

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

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

  666.                 }

  667.             }

  668. 

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

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

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

  672. 

  673.                 if (ref->format != fmt) {

  674.                     av_log(ctx, AV_LOG_ERROR,

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

  676.                            av_get_pix_fmt_name(ref->format),

  677.                            av_get_pix_fmt_name(fmt));

  678.                     return AVERROR_INVALIDDATA;

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

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

  681.                 } else {

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

  683.                         av_log(ctx, AV_LOG_ERROR,

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

  685.                                refw, refh, w, h);

  686.                         return AVERROR_INVALIDDATA;

  687.                     }

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

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

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

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

  692.                 }

  693.             }

  694.         }

  695.     }

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

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

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

  699. 

  700.     /* loopfilter header data */

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

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

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

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

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

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

  707.     s->filter.sharpness = sharp;

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

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

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

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

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

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

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

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

  716.         }

  717.     }

  718. 

  719.     /* quantization header data */

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

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

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

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

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

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

  726. 

  727.     /* segmentation header info */

  728.     s->segmentation.ignore_refmap = 0;

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

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

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

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

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

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

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

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

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

  738.             }

  739.         }

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

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

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

  743.             av_log(ctx, AV_LOG_WARNING,

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

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

  746.                 s->segmentation.ignore_refmap = 1;

  747.             //return AVERROR_INVALIDDATA;

  748.         }

  749. 

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

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

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

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

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

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

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

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

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

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

  760.             }

  761.         }

  762.     } else {

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

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

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

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

  767.     }

  768. 

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

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

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

  772. 

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

  774.             if (s->segmentation.absolute_vals)

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

  776.             else

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

  778.         } else {

  779.             qyac  = s->yac_qi;

  780.         }

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

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

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

  784.         qyac  = av_clip_uintp2(qyac, 8);

  785. 

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

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

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

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

  790. 

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

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

  793.             if (s->segmentation.absolute_vals)

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

  795.             else

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

  797.         } else {

  798.             lflvl  = s->filter.level;

  799.         }

  800.         if (s->lf_delta.enabled) {

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

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

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

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

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

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

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

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

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

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

  811.             }

  812.         } else {

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

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

  815.         }

  816.     }

  817. 

  818.     /* tiling info */

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

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

  821.         return res;

  822.     }

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

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

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

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

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

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

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

  830.             s->tiling.log2_tile_cols++;

  831.         else

  832.             break;

  833.     }

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

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

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

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

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

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

  840.         if (!s->c_b) {

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

  842.             return AVERROR(ENOMEM);

  843.         }

  844.     }

  845. 

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

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

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

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

  850.                sizeof(vp9_default_coef_probs));

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

  852.                sizeof(vp9_default_coef_probs));

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

  854.                sizeof(vp9_default_coef_probs));

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

  856.                sizeof(vp9_default_coef_probs));

  857.     }

  858. 

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

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

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

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

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

  864.         return AVERROR_INVALIDDATA;

  865.     }

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

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

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

  869.         return AVERROR_INVALIDDATA;

  870.     }

  871. 

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

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

  874.     } else {

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

  876.     }

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

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

  879.     // fw update)?

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

  881. 

  882.     // txfm updates

  883.     if (s->lossless) {

  884.         s->txfmmode = TX_4X4;

  885.     } else {

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

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

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

  889. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  904.         }

  905.     }

  906. 

  907.     // coef updates

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

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

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

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

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

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

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

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

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

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

  918.                                 break;

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

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

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

  922.                                 } else {

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

  924.                                 }

  925.                             }

  926.                             p[3] = 0;

  927.                         }

  928.         } else {

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

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

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

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

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

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

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

  936.                                 break;

  937.                             memcpy(p, r, 3);

  938.                             p[3] = 0;

  939.                         }

  940.         }

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

  942.             break;

  943.     }

  944. 

  945.     // mode updates

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

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

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

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

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

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

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

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

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

  955. 

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

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

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

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

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

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

  962. 

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

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

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

  966. 

  967.         if (s->allowcompinter) {

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

  969.             if (s->comppredmode)

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

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

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

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

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

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

  976.         } else {

  977.             s->comppredmode = PRED_SINGLEREF;

  978.         }

  979. 

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

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

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

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

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

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

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

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

  988.             }

  989.         }

  990. 

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

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

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

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

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

  996.         }

  997. 

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

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

  1000.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  1001.                     s->prob.p.y_mode[i][j] =

  1002.                         update_prob(&s->c, s->prob.p.y_mode[i][j]);

  1003. 

  1004.         for (i = 0; i < 4; i++)

  1005.             for (j = 0; j < 4; j++)

  1006.                 for (k = 0; k < 3; k++)

  1007.                     if (vp56_rac_get_prob_branchy(&s->c, 252))

  1008.                         s->prob.p.partition[3 - i][j][k] =

  1009.                             update_prob(&s->c, s->prob.p.partition[3 - i][j][k]);

  1010. 

  1011.         // mv fields don't use the update_prob subexp model for some reason

  1012.         for (i = 0; i < 3; i++)

  1013.             if (vp56_rac_get_prob_branchy(&s->c, 252))

  1014.                 s->prob.p.mv_joint[i] = (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  1015. 

  1016.         for (i = 0; i < 2; i++) {

  1017.             if (vp56_rac_get_prob_branchy(&s->c, 252))

  1018.                 s->prob.p.mv_comp[i].sign = (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  1019. 

  1020.             for (j = 0; j < 10; j++)

  1021.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  1022.                     s->prob.p.mv_comp[i].classes[j] =

  1023.                         (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  1024. 

  1025.             if (vp56_rac_get_prob_branchy(&s->c, 252))

  1026.                 s->prob.p.mv_comp[i].class0 = (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  1027. 

  1028.             for (j = 0; j < 10; j++)

  1029.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  1030.                     s->prob.p.mv_comp[i].bits[j] =

  1031.                         (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  1032.         }

  1033. 

  1034.         for (i = 0; i < 2; i++) {

  1035.             for (j = 0; j < 2; j++)

  1036.                 for (k = 0; k < 3; k++)

  1037.                     if (vp56_rac_get_prob_branchy(&s->c, 252))

  1038.                         s->prob.p.mv_comp[i].class0_fp[j][k] =

  1039.                             (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  1040. 

  1041.             for (j = 0; j < 3; j++)

  1042.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  1043.                     s->prob.p.mv_comp[i].fp[j] =

  1044.                         (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  1045.         }

  1046. 

  1047.         if (s->highprecisionmvs) {

  1048.             for (i = 0; i < 2; i++) {

  1049.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  1050.                     s->prob.p.mv_comp[i].class0_hp =

  1051.                         (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  1052. 

  1053.                 if (vp56_rac_get_prob_branchy(&s->c, 252))

  1054.                     s->prob.p.mv_comp[i].hp =

  1055.                         (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  1056.             }

  1057.         }

  1058.     }

  1059. 

  1060.     return (data2 - data) + size2;

  1061. }

  1062. 

  1063. static av_always_inline void clamp_mv(VP56mv *dst, const VP56mv *src,

  1064.                                       VP9Context *s)

  1065. {

  1066.     dst->x = av_clip(src->x, s->min_mv.x, s->max_mv.x);

  1067.     dst->y = av_clip(src->y, s->min_mv.y, s->max_mv.y);

  1068. }

  1069. 

  1070. static void find_ref_mvs(VP9Context *s,

  1071.                          VP56mv *pmv, int ref, int z, int idx, int sb)

  1072. {

  1073.     static const int8_t mv_ref_blk_off[N_BS_SIZES][8][2] = {

  1074.         [BS_64x64] = {{  3, -1 }, { -1,  3 }, {  4, -1 }, { -1,  4 },

  1075.                       { -1, -1 }, {  0, -1 }, { -1,  0 }, {  6, -1 }},

  1076.         [BS_64x32] = {{  0, -1 }, { -1,  0 }, {  4, -1 }, { -1,  2 },

  1077.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, {  2, -1 }},

  1078.         [BS_32x64] = {{ -1,  0 }, {  0, -1 }, { -1,  4 }, {  2, -1 },

  1079.                       { -1, -1 }, { -3,  0 }, {  0, -3 }, { -1,  2 }},

  1080.         [BS_32x32] = {{  1, -1 }, { -1,  1 }, {  2, -1 }, { -1,  2 },

  1081.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

  1082.         [BS_32x16] = {{  0, -1 }, { -1,  0 }, {  2, -1 }, { -1, -1 },

  1083.                       { -1,  1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

  1084.         [BS_16x32] = {{ -1,  0 }, {  0, -1 }, { -1,  2 }, { -1, -1 },

  1085.                       {  1, -1 }, { -3,  0 }, {  0, -3 }, { -3, -3 }},

  1086.         [BS_16x16] = {{  0, -1 }, { -1,  0 }, {  1, -1 }, { -1,  1 },

  1087.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

  1088.         [BS_16x8]  = {{  0, -1 }, { -1,  0 }, {  1, -1 }, { -1, -1 },

  1089.                       {  0, -2 }, { -2,  0 }, { -2, -1 }, { -1, -2 }},

  1090.         [BS_8x16]  = {{ -1,  0 }, {  0, -1 }, { -1,  1 }, { -1, -1 },

  1091.                       { -2,  0 }, {  0, -2 }, { -1, -2 }, { -2, -1 }},

  1092.         [BS_8x8]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1093.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1094.         [BS_8x4]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1095.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1096.         [BS_4x8]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1097.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1098.         [BS_4x4]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1099.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1100.     };

  1101.     VP9Block *b = s->b;

  1102.     int row = s->row, col = s->col, row7 = s->row7;

  1103.     const int8_t (*p)[2] = mv_ref_blk_off[b->bs];

  1104. #define INVALID_MV 0x80008000U

  1105.     uint32_t mem = INVALID_MV;

  1106.     int i;

  1107. 

  1108. #define RETURN_DIRECT_MV(mv) \

  1109.     do { \

  1110.         uint32_t m = AV_RN32A(&mv); \

  1111.         if (!idx) { \

  1112.             AV_WN32A(pmv, m); \

  1113.             return; \

  1114.         } else if (mem == INVALID_MV) { \

  1115.             mem = m; \

  1116.         } else if (m != mem) { \

  1117.             AV_WN32A(pmv, m); \

  1118.             return; \

  1119.         } \

  1120.     } while (0)

  1121. 

  1122.     if (sb >= 0) {

  1123.         if (sb == 2 || sb == 1) {

  1124.             RETURN_DIRECT_MV(b->mv[0][z]);

  1125.         } else if (sb == 3) {

  1126.             RETURN_DIRECT_MV(b->mv[2][z]);

  1127.             RETURN_DIRECT_MV(b->mv[1][z]);

  1128.             RETURN_DIRECT_MV(b->mv[0][z]);

  1129.         }

  1130. 

  1131. #define RETURN_MV(mv) \

  1132.     do { \

  1133.         if (sb > 0) { \

  1134.             VP56mv tmp; \

  1135.             uint32_t m; \

  1136.             clamp_mv(&tmp, &mv, s); \

  1137.             m = AV_RN32A(&tmp); \

  1138.             if (!idx) { \

  1139.                 AV_WN32A(pmv, m); \

  1140.                 return; \

  1141.             } else if (mem == INVALID_MV) { \

  1142.                 mem = m; \

  1143.             } else if (m != mem) { \

  1144.                 AV_WN32A(pmv, m); \

  1145.                 return; \

  1146.             } \

  1147.         } else { \

  1148.             uint32_t m = AV_RN32A(&mv); \

  1149.             if (!idx) { \

  1150.                 clamp_mv(pmv, &mv, s); \

  1151.                 return; \

  1152.             } else if (mem == INVALID_MV) { \

  1153.                 mem = m; \

  1154.             } else if (m != mem) { \

  1155.                 clamp_mv(pmv, &mv, s); \

  1156.                 return; \

  1157.             } \

  1158.         } \

  1159.     } while (0)

  1160. 

  1161.         if (row > 0) {

  1162.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[(row - 1) * s->sb_cols * 8 + col];

  1163.             if (mv->ref[0] == ref) {

  1164.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][0]);

  1165.             } else if (mv->ref[1] == ref) {

  1166.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][1]);

  1167.             }

  1168.         }

  1169.         if (col > s->tiling.tile_col_start) {

  1170.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[row * s->sb_cols * 8 + col - 1];

  1171.             if (mv->ref[0] == ref) {

  1172.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][0]);

  1173.             } else if (mv->ref[1] == ref) {

  1174.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][1]);

  1175.             }

  1176.         }

  1177.         i = 2;

  1178.     } else {

  1179.         i = 0;

  1180.     }

  1181. 

  1182.     // previously coded MVs in this neighbourhood, using same reference frame

  1183.     for (; i < 8; i++) {

  1184.         int c = p[i][0] + col, r = p[i][1] + row;

  1185. 

  1186.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1187.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1188. 

  1189.             if (mv->ref[0] == ref) {

  1190.                 RETURN_MV(mv->mv[0]);

  1191.             } else if (mv->ref[1] == ref) {

  1192.                 RETURN_MV(mv->mv[1]);

  1193.             }

  1194.         }

  1195.     }

  1196. 

  1197.     // MV at this position in previous frame, using same reference frame

  1198.     if (s->use_last_frame_mvs) {

  1199.         struct VP9mvrefPair *mv = &s->frames[REF_FRAME_MVPAIR].mv[row * s->sb_cols * 8 + col];

  1200. 

  1201.         if (!s->frames[REF_FRAME_MVPAIR].uses_2pass)

  1202.             ff_thread_await_progress(&s->frames[REF_FRAME_MVPAIR].tf, row >> 3, 0);

  1203.         if (mv->ref[0] == ref) {

  1204.             RETURN_MV(mv->mv[0]);

  1205.         } else if (mv->ref[1] == ref) {

  1206.             RETURN_MV(mv->mv[1]);

  1207.         }

  1208.     }

  1209. 

  1210. #define RETURN_SCALE_MV(mv, scale) \

  1211.     do { \

  1212.         if (scale) { \

  1213.             VP56mv mv_temp = { -mv.x, -mv.y }; \

  1214.             RETURN_MV(mv_temp); \

  1215.         } else { \

  1216.             RETURN_MV(mv); \

  1217.         } \

  1218.     } while (0)

  1219. 

  1220.     // previously coded MVs in this neighbourhood, using different reference frame

  1221.     for (i = 0; i < 8; i++) {

  1222.         int c = p[i][0] + col, r = p[i][1] + row;

  1223. 

  1224.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1225.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1226. 

  1227.             if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1228.                 RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1229.             }

  1230.             if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1231.                 // BUG - libvpx has this condition regardless of whether

  1232.                 // we used the first ref MV and pre-scaling

  1233.                 AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1234.                 RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1235.             }

  1236.         }

  1237.     }

  1238. 

  1239.     // MV at this position in previous frame, using different reference frame

  1240.     if (s->use_last_frame_mvs) {

  1241.         struct VP9mvrefPair *mv = &s->frames[REF_FRAME_MVPAIR].mv[row * s->sb_cols * 8 + col];

  1242. 

  1243.         // no need to await_progress, because we already did that above

  1244.         if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1245.             RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1246.         }

  1247.         if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1248.             // BUG - libvpx has this condition regardless of whether

  1249.             // we used the first ref MV and pre-scaling

  1250.             AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1251.             RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1252.         }

  1253.     }

  1254. 

  1255.     AV_ZERO32(pmv);

  1256. #undef INVALID_MV

  1257. #undef RETURN_MV

  1258. #undef RETURN_SCALE_MV

  1259. }

  1260. 

  1261. static av_always_inline int read_mv_component(VP9Context *s, int idx, int hp)

  1262. {

  1263.     int bit, sign = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].sign);

  1264.     int n, c = vp8_rac_get_tree(&s->c, vp9_mv_class_tree,

  1265.                                 s->prob.p.mv_comp[idx].classes);

  1266. 

  1267.     s->counts.mv_comp[idx].sign[sign]++;

  1268.     s->counts.mv_comp[idx].classes[c]++;

  1269.     if (c) {

  1270.         int m;

  1271. 

  1272.         for (n = 0, m = 0; m < c; m++) {

  1273.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].bits[m]);

  1274.             n |= bit << m;

  1275.             s->counts.mv_comp[idx].bits[m][bit]++;

  1276.         }

  1277.         n <<= 3;

  1278.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree, s->prob.p.mv_comp[idx].fp);

  1279.         n |= bit << 1;

  1280.         s->counts.mv_comp[idx].fp[bit]++;

  1281.         if (hp) {

  1282.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].hp);

  1283.             s->counts.mv_comp[idx].hp[bit]++;

  1284.             n |= bit;

  1285.         } else {

  1286.             n |= 1;

  1287.             // bug in libvpx - we count for bw entropy purposes even if the

  1288.             // bit wasn't coded

  1289.             s->counts.mv_comp[idx].hp[1]++;

  1290.         }

  1291.         n += 8 << c;

  1292.     } else {

  1293.         n = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0);

  1294.         s->counts.mv_comp[idx].class0[n]++;

  1295.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree,

  1296.                                s->prob.p.mv_comp[idx].class0_fp[n]);

  1297.         s->counts.mv_comp[idx].class0_fp[n][bit]++;

  1298.         n = (n << 3) | (bit << 1);

  1299.         if (hp) {

  1300.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0_hp);

  1301.             s->counts.mv_comp[idx].class0_hp[bit]++;

  1302.             n |= bit;

  1303.         } else {

  1304.             n |= 1;

  1305.             // bug in libvpx - we count for bw entropy purposes even if the

  1306.             // bit wasn't coded

  1307.             s->counts.mv_comp[idx].class0_hp[1]++;

  1308.         }

  1309.     }

  1310. 

  1311.     return sign ? -(n + 1) : (n + 1);

  1312. }

  1313. 

  1314. static void fill_mv(VP9Context *s,

  1315.                     VP56mv *mv, int mode, int sb)

  1316. {

  1317.     VP9Block *b = s->b;

  1318. 

  1319.     if (mode == ZEROMV) {

  1320.         AV_ZERO64(mv);

  1321.     } else {

  1322.         int hp;

  1323. 

  1324.         // FIXME cache this value and reuse for other subblocks

  1325.         find_ref_mvs(s, &mv[0], b->ref[0], 0, mode == NEARMV,

  1326.                      mode == NEWMV ? -1 : sb);

  1327.         // FIXME maybe move this code into find_ref_mvs()

  1328.         if ((mode == NEWMV || sb == -1) &&

  1329.             !(hp = s->highprecisionmvs && abs(mv[0].x) < 64 && abs(mv[0].y) < 64)) {

  1330.             if (mv[0].y & 1) {

  1331.                 if (mv[0].y < 0)

  1332.                     mv[0].y++;

  1333.                 else

  1334.                     mv[0].y--;

  1335.             }

  1336.             if (mv[0].x & 1) {

  1337.                 if (mv[0].x < 0)

  1338.                     mv[0].x++;

  1339.                 else

  1340.                     mv[0].x--;

  1341.             }

  1342.         }

  1343.         if (mode == NEWMV) {

  1344.             enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1345.                                               s->prob.p.mv_joint);

  1346. 

  1347.             s->counts.mv_joint[j]++;

  1348.             if (j >= MV_JOINT_V)

  1349.                 mv[0].y += read_mv_component(s, 0, hp);

  1350.             if (j & 1)

  1351.                 mv[0].x += read_mv_component(s, 1, hp);

  1352.         }

  1353. 

  1354.         if (b->comp) {

  1355.             // FIXME cache this value and reuse for other subblocks

  1356.             find_ref_mvs(s, &mv[1], b->ref[1], 1, mode == NEARMV,

  1357.                          mode == NEWMV ? -1 : sb);

  1358.             if ((mode == NEWMV || sb == -1) &&

  1359.                 !(hp = s->highprecisionmvs && abs(mv[1].x) < 64 && abs(mv[1].y) < 64)) {

  1360.                 if (mv[1].y & 1) {

  1361.                     if (mv[1].y < 0)

  1362.                         mv[1].y++;

  1363.                     else

  1364.                         mv[1].y--;

  1365.                 }

  1366.                 if (mv[1].x & 1) {

  1367.                     if (mv[1].x < 0)

  1368.                         mv[1].x++;

  1369.                     else

  1370.                         mv[1].x--;

  1371.                 }

  1372.             }

  1373.             if (mode == NEWMV) {

  1374.                 enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1375.                                                   s->prob.p.mv_joint);

  1376. 

  1377.                 s->counts.mv_joint[j]++;

  1378.                 if (j >= MV_JOINT_V)

  1379.                     mv[1].y += read_mv_component(s, 0, hp);

  1380.                 if (j & 1)

  1381.                     mv[1].x += read_mv_component(s, 1, hp);

  1382.             }

  1383.         }

  1384.     }

  1385. }

  1386. 

  1387. static av_always_inline void setctx_2d(uint8_t *ptr, int w, int h,

  1388.                                        ptrdiff_t stride, int v)

  1389. {

  1390.     switch (w) {

  1391.     case 1:

  1392.         do {

  1393.             *ptr = v;

  1394.             ptr += stride;

  1395.         } while (--h);

  1396.         break;

  1397.     case 2: {

  1398.         int v16 = v * 0x0101;

  1399.         do {

  1400.             AV_WN16A(ptr, v16);

  1401.             ptr += stride;

  1402.         } while (--h);

  1403.         break;

  1404.     }

  1405.     case 4: {

  1406.         uint32_t v32 = v * 0x01010101;

  1407.         do {

  1408.             AV_WN32A(ptr, v32);

  1409.             ptr += stride;

  1410.         } while (--h);

  1411.         break;

  1412.     }

  1413.     case 8: {

  1414. #if HAVE_FAST_64BIT

  1415.         uint64_t v64 = v * 0x0101010101010101ULL;

  1416.         do {

  1417.             AV_WN64A(ptr, v64);

  1418.             ptr += stride;

  1419.         } while (--h);

  1420. #else

  1421.         uint32_t v32 = v * 0x01010101;

  1422.         do {

  1423.             AV_WN32A(ptr,     v32);

  1424.             AV_WN32A(ptr + 4, v32);

  1425.             ptr += stride;

  1426.         } while (--h);

  1427. #endif

  1428.         break;

  1429.     }

  1430.     }

  1431. }

  1432. 

  1433. static void decode_mode(AVCodecContext *ctx)

  1434. {

  1435.     static const uint8_t left_ctx[N_BS_SIZES] = {

  1436.         0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf

  1437.     };

  1438.     static const uint8_t above_ctx[N_BS_SIZES] = {

  1439.         0x0, 0x0, 0x8, 0x8, 0x8, 0xc, 0xc, 0xc, 0xe, 0xe, 0xe, 0xf, 0xf

  1440.     };

  1441.     static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {

  1442.         TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,

  1443.         TX_16X16, TX_8X8, TX_8X8, TX_8X8, TX_4X4, TX_4X4, TX_4X4

  1444.     };

  1445.     VP9Context *s = ctx->priv_data;

  1446.     VP9Block *b = s->b;

  1447.     int row = s->row, col = s->col, row7 = s->row7;

  1448.     enum TxfmMode max_tx = max_tx_for_bl_bp[b->bs];

  1449.     int bw4 = bwh_tab[1][b->bs][0], w4 = FFMIN(s->cols - col, bw4);

  1450.     int bh4 = bwh_tab[1][b->bs][1], h4 = FFMIN(s->rows - row, bh4), y;

  1451.     int have_a = row > 0, have_l = col > s->tiling.tile_col_start;

  1452.     int vref, filter_id;

  1453. 

  1454.     if (!s->segmentation.enabled) {

  1455.         b->seg_id = 0;

  1456.     } else if (s->keyframe || s->intraonly) {

  1457.         b->seg_id = vp8_rac_get_tree(&s->c, vp9_segmentation_tree, s->prob.seg);

  1458.     } else if (!s->segmentation.update_map ||

  1459.                (s->segmentation.temporal &&

  1460.                 vp56_rac_get_prob_branchy(&s->c,

  1461.                     s->prob.segpred[s->above_segpred_ctx[col] +

  1462.                                     s->left_segpred_ctx[row7]]))) {

  1463.         if (!s->errorres && !s->segmentation.ignore_refmap) {

  1464.             int pred = 8, x;

  1465.             uint8_t *refsegmap = s->frames[REF_FRAME_SEGMAP].segmentation_map;

  1466. 

  1467.             if (!s->frames[REF_FRAME_SEGMAP].uses_2pass)

  1468.                 ff_thread_await_progress(&s->frames[REF_FRAME_SEGMAP].tf, row >> 3, 0);

  1469.             for (y = 0; y < h4; y++) {

  1470.                 int idx_base = (y + row) * 8 * s->sb_cols + col;

  1471.                 for (x = 0; x < w4; x++)

  1472.                     pred = FFMIN(pred, refsegmap[idx_base + x]);

  1473.             }

  1474.             av_assert1(pred < 8);

  1475.             b->seg_id = pred;

  1476.         } else {

  1477.             b->seg_id = 0;

  1478.         }

  1479. 

  1480.         memset(&s->above_segpred_ctx[col], 1, w4);

  1481.         memset(&s->left_segpred_ctx[row7], 1, h4);

  1482.     } else {

  1483.         b->seg_id = vp8_rac_get_tree(&s->c, vp9_segmentation_tree,

  1484.                                      s->prob.seg);

  1485. 

  1486.         memset(&s->above_segpred_ctx[col], 0, w4);

  1487.         memset(&s->left_segpred_ctx[row7], 0, h4);

  1488.     }

  1489.     if (s->segmentation.enabled &&

  1490.         (s->segmentation.update_map || s->keyframe || s->intraonly)) {

  1491.         setctx_2d(&s->frames[CUR_FRAME].segmentation_map[row * 8 * s->sb_cols + col],

  1492.                   bw4, bh4, 8 * s->sb_cols, b->seg_id);

  1493.     }

  1494. 

  1495.     b->skip = s->segmentation.enabled &&

  1496.         s->segmentation.feat[b->seg_id].skip_enabled;

  1497.     if (!b->skip) {

  1498.         int c = s->left_skip_ctx[row7] + s->above_skip_ctx[col];

  1499.         b->skip = vp56_rac_get_prob(&s->c, s->prob.p.skip[c]);

  1500.         s->counts.skip[c][b->skip]++;

  1501.     }

  1502. 

  1503.     if (s->keyframe || s->intraonly) {

  1504.         b->intra = 1;

  1505.     } else if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1506.         b->intra = !s->segmentation.feat[b->seg_id].ref_val;

  1507.     } else {

  1508.         int c, bit;

  1509. 

  1510.         if (have_a && have_l) {

  1511.             c = s->above_intra_ctx[col] + s->left_intra_ctx[row7];

  1512.             c += (c == 2);

  1513.         } else {

  1514.             c = have_a ? 2 * s->above_intra_ctx[col] :

  1515.                 have_l ? 2 * s->left_intra_ctx[row7] : 0;

  1516.         }

  1517.         bit = vp56_rac_get_prob(&s->c, s->prob.p.intra[c]);

  1518.         s->counts.intra[c][bit]++;

  1519.         b->intra = !bit;

  1520.     }

  1521. 

  1522.     if ((b->intra || !b->skip) && s->txfmmode == TX_SWITCHABLE) {

  1523.         int c;

  1524.         if (have_a) {

  1525.             if (have_l) {

  1526.                 c = (s->above_skip_ctx[col] ? max_tx :

  1527.                      s->above_txfm_ctx[col]) +

  1528.                     (s->left_skip_ctx[row7] ? max_tx :

  1529.                      s->left_txfm_ctx[row7]) > max_tx;

  1530.             } else {

  1531.                 c = s->above_skip_ctx[col] ? 1 :

  1532.                     (s->above_txfm_ctx[col] * 2 > max_tx);

  1533.             }

  1534.         } else if (have_l) {

  1535.             c = s->left_skip_ctx[row7] ? 1 :

  1536.                 (s->left_txfm_ctx[row7] * 2 > max_tx);

  1537.         } else {

  1538.             c = 1;

  1539.         }

  1540.         switch (max_tx) {

  1541.         case TX_32X32:

  1542.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][0]);

  1543.             if (b->tx) {

  1544.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][1]);

  1545.                 if (b->tx == 2)

  1546.                     b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][2]);

  1547.             }

  1548.             s->counts.tx32p[c][b->tx]++;

  1549.             break;

  1550.         case TX_16X16:

  1551.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][0]);

  1552.             if (b->tx)

  1553.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][1]);

  1554.             s->counts.tx16p[c][b->tx]++;

  1555.             break;

  1556.         case TX_8X8:

  1557.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx8p[c]);

  1558.             s->counts.tx8p[c][b->tx]++;

  1559.             break;

  1560.         case TX_4X4:

  1561.             b->tx = TX_4X4;

  1562.             break;

  1563.         }

  1564.     } else {

  1565.         b->tx = FFMIN(max_tx, s->txfmmode);

  1566.     }

  1567. 

  1568.     if (s->keyframe || s->intraonly) {

  1569.         uint8_t *a = &s->above_mode_ctx[col * 2];

  1570.         uint8_t *l = &s->left_mode_ctx[(row7) << 1];

  1571. 

  1572.         b->comp = 0;

  1573.         if (b->bs > BS_8x8) {

  1574.             // FIXME the memory storage intermediates here aren't really

  1575.             // necessary, they're just there to make the code slightly

  1576.             // simpler for now

  1577.             b->mode[0] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1578.                                     vp9_default_kf_ymode_probs[a[0]][l[0]]);

  1579.             if (b->bs != BS_8x4) {

  1580.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1581.                                  vp9_default_kf_ymode_probs[a[1]][b->mode[0]]);

  1582.                 l[0] = a[1] = b->mode[1];

  1583.             } else {

  1584.                 l[0] = a[1] = b->mode[1] = b->mode[0];

  1585.             }

  1586.             if (b->bs != BS_4x8) {

  1587.                 b->mode[2] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1588.                                         vp9_default_kf_ymode_probs[a[0]][l[1]]);

  1589.                 if (b->bs != BS_8x4) {

  1590.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1591.                                   vp9_default_kf_ymode_probs[a[1]][b->mode[2]]);

  1592.                     l[1] = a[1] = b->mode[3];

  1593.                 } else {

  1594.                     l[1] = a[1] = b->mode[3] = b->mode[2];

  1595.                 }

  1596.             } else {

  1597.                 b->mode[2] = b->mode[0];

  1598.                 l[1] = a[1] = b->mode[3] = b->mode[1];

  1599.             }

  1600.         } else {

  1601.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1602.                                           vp9_default_kf_ymode_probs[*a][*l]);

  1603.             b->mode[3] = b->mode[2] = b->mode[1] = b->mode[0];

  1604.             // FIXME this can probably be optimized

  1605.             memset(a, b->mode[0], bwh_tab[0][b->bs][0]);

  1606.             memset(l, b->mode[0], bwh_tab[0][b->bs][1]);

  1607.         }

  1608.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1609.                                      vp9_default_kf_uvmode_probs[b->mode[3]]);

  1610.     } else if (b->intra) {

  1611.         b->comp = 0;

  1612.         if (b->bs > BS_8x8) {

  1613.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1614.                                           s->prob.p.y_mode[0]);

  1615.             s->counts.y_mode[0][b->mode[0]]++;

  1616.             if (b->bs != BS_8x4) {

  1617.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1618.                                               s->prob.p.y_mode[0]);

  1619.                 s->counts.y_mode[0][b->mode[1]]++;

  1620.             } else {

  1621.                 b->mode[1] = b->mode[0];

  1622.             }

  1623.             if (b->bs != BS_4x8) {

  1624.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1625.                                               s->prob.p.y_mode[0]);

  1626.                 s->counts.y_mode[0][b->mode[2]]++;

  1627.                 if (b->bs != BS_8x4) {

  1628.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1629.                                                   s->prob.p.y_mode[0]);

  1630.                     s->counts.y_mode[0][b->mode[3]]++;

  1631.                 } else {

  1632.                     b->mode[3] = b->mode[2];

  1633.                 }

  1634.             } else {

  1635.                 b->mode[2] = b->mode[0];

  1636.                 b->mode[3] = b->mode[1];

  1637.             }

  1638.         } else {

  1639.             static const uint8_t size_group[10] = {

  1640.                 3, 3, 3, 3, 2, 2, 2, 1, 1, 1

  1641.             };

  1642.             int sz = size_group[b->bs];

  1643. 

  1644.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1645.                                           s->prob.p.y_mode[sz]);

  1646.             b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1647.             s->counts.y_mode[sz][b->mode[3]]++;

  1648.         }

  1649.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1650.                                      s->prob.p.uv_mode[b->mode[3]]);

  1651.         s->counts.uv_mode[b->mode[3]][b->uvmode]++;

  1652.     } else {

  1653.         static const uint8_t inter_mode_ctx_lut[14][14] = {

  1654.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1655.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1656.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1657.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1658.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1659.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1660.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1661.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1662.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1663.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1664.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1665.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1666.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },

  1667.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },

  1668.         };

  1669. 

  1670.         if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1671.             av_assert2(s->segmentation.feat[b->seg_id].ref_val != 0);

  1672.             b->comp = 0;

  1673.             b->ref[0] = s->segmentation.feat[b->seg_id].ref_val - 1;

  1674.         } else {

  1675.             // read comp_pred flag

  1676.             if (s->comppredmode != PRED_SWITCHABLE) {

  1677.                 b->comp = s->comppredmode == PRED_COMPREF;

  1678.             } else {

  1679.                 int c;

  1680. 

  1681.                 // FIXME add intra as ref=0xff (or -1) to make these easier?

  1682.                 if (have_a) {

  1683.                     if (have_l) {

  1684.                         if (s->above_comp_ctx[col] && s->left_comp_ctx[row7]) {

  1685.                             c = 4;

  1686.                         } else if (s->above_comp_ctx[col]) {

  1687.                             c = 2 + (s->left_intra_ctx[row7] ||

  1688.                                      s->left_ref_ctx[row7] == s->fixcompref);

  1689.                         } else if (s->left_comp_ctx[row7]) {

  1690.                             c = 2 + (s->above_intra_ctx[col] ||

  1691.                                      s->above_ref_ctx[col] == s->fixcompref);

  1692.                         } else {

  1693.                             c = (!s->above_intra_ctx[col] &&

  1694.                                  s->above_ref_ctx[col] == s->fixcompref) ^

  1695.                             (!s->left_intra_ctx[row7] &&

  1696.                              s->left_ref_ctx[row & 7] == s->fixcompref);

  1697.                         }

  1698.                     } else {

  1699.                         c = s->above_comp_ctx[col] ? 3 :

  1700.                         (!s->above_intra_ctx[col] && s->above_ref_ctx[col] == s->fixcompref);

  1701.                     }

  1702.                 } else if (have_l) {

  1703.                     c = s->left_comp_ctx[row7] ? 3 :

  1704.                     (!s->left_intra_ctx[row7] && s->left_ref_ctx[row7] == s->fixcompref);

  1705.                 } else {

  1706.                     c = 1;

  1707.                 }

  1708.                 b->comp = vp56_rac_get_prob(&s->c, s->prob.p.comp[c]);

  1709.                 s->counts.comp[c][b->comp]++;

  1710.             }

  1711. 

  1712.             // read actual references

  1713.             // FIXME probably cache a few variables here to prevent repetitive

  1714.             // memory accesses below

  1715.             if (b->comp) /* two references */ {

  1716.                 int fix_idx = s->signbias[s->fixcompref], var_idx = !fix_idx, c, bit;

  1717. 

  1718.                 b->ref[fix_idx] = s->fixcompref;

  1719.                 // FIXME can this codeblob be replaced by some sort of LUT?

  1720.                 if (have_a) {

  1721.                     if (have_l) {

  1722.                         if (s->above_intra_ctx[col]) {

  1723.                             if (s->left_intra_ctx[row7]) {

  1724.                                 c = 2;

  1725.                             } else {

  1726.                                 c = 1 + 2 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1727.                             }

  1728.                         } else if (s->left_intra_ctx[row7]) {

  1729.                             c = 1 + 2 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1730.                         } else {

  1731.                             int refl = s->left_ref_ctx[row7], refa = s->above_ref_ctx[col];

  1732. 

  1733.                             if (refl == refa && refa == s->varcompref[1]) {

  1734.                                 c = 0;

  1735.                             } else if (!s->left_comp_ctx[row7] && !s->above_comp_ctx[col]) {

  1736.                                 if ((refa == s->fixcompref && refl == s->varcompref[0]) ||

  1737.                                     (refl == s->fixcompref && refa == s->varcompref[0])) {

  1738.                                     c = 4;

  1739.                                 } else {

  1740.                                     c = (refa == refl) ? 3 : 1;

  1741.                                 }

  1742.                             } else if (!s->left_comp_ctx[row7]) {

  1743.                                 if (refa == s->varcompref[1] && refl != s->varcompref[1]) {

  1744.                                     c = 1;

  1745.                                 } else {

  1746.                                     c = (refl == s->varcompref[1] &&

  1747.                                          refa != s->varcompref[1]) ? 2 : 4;

  1748.                                 }

  1749.                             } else if (!s->above_comp_ctx[col]) {

  1750.                                 if (refl == s->varcompref[1] && refa != s->varcompref[1]) {

  1751.                                     c = 1;

  1752.                                 } else {

  1753.                                     c = (refa == s->varcompref[1] &&

  1754.                                          refl != s->varcompref[1]) ? 2 : 4;

  1755.                                 }

  1756.                             } else {

  1757.                                 c = (refl == refa) ? 4 : 2;

  1758.                             }

  1759.                         }

  1760.                     } else {

  1761.                         if (s->above_intra_ctx[col]) {

  1762.                             c = 2;

  1763.                         } else if (s->above_comp_ctx[col]) {

  1764.                             c = 4 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1765.                         } else {

  1766.                             c = 3 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1767.                         }

  1768.                     }

  1769.                 } else if (have_l) {

  1770.                     if (s->left_intra_ctx[row7]) {

  1771.                         c = 2;

  1772.                     } else if (s->left_comp_ctx[row7]) {

  1773.                         c = 4 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1774.                     } else {

  1775.                         c = 3 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1776.                     }

  1777.                 } else {

  1778.                     c = 2;

  1779.                 }

  1780.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.comp_ref[c]);

  1781.                 b->ref[var_idx] = s->varcompref[bit];

  1782.                 s->counts.comp_ref[c][bit]++;

  1783.             } else /* single reference */ {

  1784.                 int bit, c;

  1785. 

  1786.                 if (have_a && !s->above_intra_ctx[col]) {

  1787.                     if (have_l && !s->left_intra_ctx[row7]) {

  1788.                         if (s->left_comp_ctx[row7]) {

  1789.                             if (s->above_comp_ctx[col]) {

  1790.                                 c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7] ||

  1791.                                          !s->above_ref_ctx[col]);

  1792.                             } else {

  1793.                                 c = (3 * !s->above_ref_ctx[col]) +

  1794.                                     (!s->fixcompref || !s->left_ref_ctx[row7]);

  1795.                             }

  1796.                         } else if (s->above_comp_ctx[col]) {

  1797.                             c = (3 * !s->left_ref_ctx[row7]) +

  1798.                                 (!s->fixcompref || !s->above_ref_ctx[col]);

  1799.                         } else {

  1800.                             c = 2 * !s->left_ref_ctx[row7] + 2 * !s->above_ref_ctx[col];

  1801.                         }

  1802.                     } else if (s->above_intra_ctx[col]) {

  1803.                         c = 2;

  1804.                     } else if (s->above_comp_ctx[col]) {

  1805.                         c = 1 + (!s->fixcompref || !s->above_ref_ctx[col]);

  1806.                     } else {

  1807.                         c = 4 * (!s->above_ref_ctx[col]);

  1808.                     }

  1809.                 } else if (have_l && !s->left_intra_ctx[row7]) {

  1810.                     if (s->left_intra_ctx[row7]) {

  1811.                         c = 2;

  1812.                     } else if (s->left_comp_ctx[row7]) {

  1813.                         c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7]);

  1814.                     } else {

  1815.                         c = 4 * (!s->left_ref_ctx[row7]);

  1816.                     }

  1817.                 } else {

  1818.                     c = 2;

  1819.                 }

  1820.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][0]);

  1821.                 s->counts.single_ref[c][0][bit]++;

  1822.                 if (!bit) {

  1823.                     b->ref[0] = 0;

  1824.                 } else {

  1825.                     // FIXME can this codeblob be replaced by some sort of LUT?

  1826.                     if (have_a) {

  1827.                         if (have_l) {

  1828.                             if (s->left_intra_ctx[row7]) {

  1829.                                 if (s->above_intra_ctx[col]) {

  1830.                                     c = 2;

  1831.                                 } else if (s->above_comp_ctx[col]) {

  1832.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1833.                                                  s->above_ref_ctx[col] == 1);

  1834.                                 } else if (!s->above_ref_ctx[col]) {

  1835.                                     c = 3;

  1836.                                 } else {

  1837.                                     c = 4 * (s->above_ref_ctx[col] == 1);

  1838.                                 }

  1839.                             } else if (s->above_intra_ctx[col]) {

  1840.                                 if (s->left_intra_ctx[row7]) {

  1841.                                     c = 2;

  1842.                                 } else if (s->left_comp_ctx[row7]) {

  1843.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1844.                                                  s->left_ref_ctx[row7] == 1);

  1845.                                 } else if (!s->left_ref_ctx[row7]) {

  1846.                                     c = 3;

  1847.                                 } else {

  1848.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1849.                                 }

  1850.                             } else if (s->above_comp_ctx[col]) {

  1851.                                 if (s->left_comp_ctx[row7]) {

  1852.                                     if (s->left_ref_ctx[row7] == s->above_ref_ctx[col]) {

  1853.                                         c = 3 * (s->fixcompref == 1 ||

  1854.                                                  s->left_ref_ctx[row7] == 1);

  1855.                                     } else {

  1856.                                         c = 2;

  1857.                                     }

  1858.                                 } else if (!s->left_ref_ctx[row7]) {

  1859.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1860.                                                  s->above_ref_ctx[col] == 1);

  1861.                                 } else {

  1862.                                     c = 3 * (s->left_ref_ctx[row7] == 1) +

  1863.                                     (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1864.                                 }

  1865.                             } else if (s->left_comp_ctx[row7]) {

  1866.                                 if (!s->above_ref_ctx[col]) {

  1867.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1868.                                                  s->left_ref_ctx[row7] == 1);

  1869.                                 } else {

  1870.                                     c = 3 * (s->above_ref_ctx[col] == 1) +

  1871.                                     (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1872.                                 }

  1873.                             } else if (!s->above_ref_ctx[col]) {

  1874.                                 if (!s->left_ref_ctx[row7]) {

  1875.                                     c = 3;

  1876.                                 } else {

  1877.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1878.                                 }

  1879.                             } else if (!s->left_ref_ctx[row7]) {

  1880.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1881.                             } else {

  1882.                                 c = 2 * (s->left_ref_ctx[row7] == 1) +

  1883.                                 2 * (s->above_ref_ctx[col] == 1);

  1884.                             }

  1885.                         } else {

  1886.                             if (s->above_intra_ctx[col] ||

  1887.                                 (!s->above_comp_ctx[col] && !s->above_ref_ctx[col])) {

  1888.                                 c = 2;

  1889.                             } else if (s->above_comp_ctx[col]) {

  1890.                                 c = 3 * (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1891.                             } else {

  1892.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1893.                             }

  1894.                         }

  1895.                     } else if (have_l) {

  1896.                         if (s->left_intra_ctx[row7] ||

  1897.                             (!s->left_comp_ctx[row7] && !s->left_ref_ctx[row7])) {

  1898.                             c = 2;

  1899.                         } else if (s->left_comp_ctx[row7]) {

  1900.                             c = 3 * (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1901.                         } else {

  1902.                             c = 4 * (s->left_ref_ctx[row7] == 1);

  1903.                         }

  1904.                     } else {

  1905.                         c = 2;

  1906.                     }

  1907.                     bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][1]);

  1908.                     s->counts.single_ref[c][1][bit]++;

  1909.                     b->ref[0] = 1 + bit;

  1910.                 }

  1911.             }

  1912.         }

  1913. 

  1914.         if (b->bs <= BS_8x8) {

  1915.             if (s->segmentation.feat[b->seg_id].skip_enabled) {

  1916.                 b->mode[0] = b->mode[1] = b->mode[2] = b->mode[3] = ZEROMV;

  1917.             } else {

  1918.                 static const uint8_t off[10] = {

  1919.                     3, 0, 0, 1, 0, 0, 0, 0, 0, 0

  1920.                 };

  1921. 

  1922.                 // FIXME this needs to use the LUT tables from find_ref_mvs

  1923.                 // because not all are -1,0/0,-1

  1924.                 int c = inter_mode_ctx_lut[s->above_mode_ctx[col + off[b->bs]]]

  1925.                                           [s->left_mode_ctx[row7 + off[b->bs]]];

  1926. 

  1927.                 b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1928.                                               s->prob.p.mv_mode[c]);

  1929.                 b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1930.                 s->counts.mv_mode[c][b->mode[0] - 10]++;

  1931.             }

  1932.         }

  1933. 

  1934.         if (s->filtermode == FILTER_SWITCHABLE) {

  1935.             int c;

  1936. 

  1937.             if (have_a && s->above_mode_ctx[col] >= NEARESTMV) {

  1938.                 if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1939.                     c = s->above_filter_ctx[col] == s->left_filter_ctx[row7] ?

  1940.                         s->left_filter_ctx[row7] : 3;

  1941.                 } else {

  1942.                     c = s->above_filter_ctx[col];

  1943.                 }

  1944.             } else if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1945.                 c = s->left_filter_ctx[row7];

  1946.             } else {

  1947.                 c = 3;

  1948.             }

  1949. 

  1950.             filter_id = vp8_rac_get_tree(&s->c, vp9_filter_tree,

  1951.                                          s->prob.p.filter[c]);

  1952.             s->counts.filter[c][filter_id]++;

  1953.             b->filter = vp9_filter_lut[filter_id];

  1954.         } else {

  1955.             b->filter = s->filtermode;

  1956.         }

  1957. 

  1958.         if (b->bs > BS_8x8) {

  1959.             int c = inter_mode_ctx_lut[s->above_mode_ctx[col]][s->left_mode_ctx[row7]];

  1960. 

  1961.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1962.                                           s->prob.p.mv_mode[c]);

  1963.             s->counts.mv_mode[c][b->mode[0] - 10]++;

  1964.             fill_mv(s, b->mv[0], b->mode[0], 0);

  1965. 

  1966.             if (b->bs != BS_8x4) {

  1967.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1968.                                               s->prob.p.mv_mode[c]);

  1969.                 s->counts.mv_mode[c][b->mode[1] - 10]++;

  1970.                 fill_mv(s, b->mv[1], b->mode[1], 1);

  1971.             } else {

  1972.                 b->mode[1] = b->mode[0];

  1973.                 AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1974.                 AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1975.             }

  1976. 

  1977.             if (b->bs != BS_4x8) {

  1978.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1979.                                               s->prob.p.mv_mode[c]);

  1980.                 s->counts.mv_mode[c][b->mode[2] - 10]++;

  1981.                 fill_mv(s, b->mv[2], b->mode[2], 2);

  1982. 

  1983.                 if (b->bs != BS_8x4) {

  1984.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1985.                                                   s->prob.p.mv_mode[c]);

  1986.                     s->counts.mv_mode[c][b->mode[3] - 10]++;

  1987.                     fill_mv(s, b->mv[3], b->mode[3], 3);

  1988.                 } else {

  1989.                     b->mode[3] = b->mode[2];

  1990.                     AV_COPY32(&b->mv[3][0], &b->mv[2][0]);

  1991.                     AV_COPY32(&b->mv[3][1], &b->mv[2][1]);

  1992.                 }

  1993.             } else {

  1994.                 b->mode[2] = b->mode[0];

  1995.                 AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1996.                 AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1997.                 b->mode[3] = b->mode[1];

  1998.                 AV_COPY32(&b->mv[3][0], &b->mv[1][0]);

  1999.                 AV_COPY32(&b->mv[3][1], &b->mv[1][1]);

  2000.             }

  2001.         } else {

  2002.             fill_mv(s, b->mv[0], b->mode[0], -1);

  2003.             AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  2004.             AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  2005.             AV_COPY32(&b->mv[3][0], &b->mv[0][0]);

  2006.             AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  2007.             AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  2008.             AV_COPY32(&b->mv[3][1], &b->mv[0][1]);

  2009.         }

  2010. 

  2011.         vref = b->ref[b->comp ? s->signbias[s->varcompref[0]] : 0];

  2012.     }

  2013. 

  2014. #if HAVE_FAST_64BIT

  2015. #define SPLAT_CTX(var, val, n) \

  2016.     switch (n) { \

  2017.     case 1:  var = val;                                    break; \

  2018.     case 2:  AV_WN16A(&var, val *             0x0101);     break; \

  2019.     case 4:  AV_WN32A(&var, val *         0x01010101);     break; \

  2020.     case 8:  AV_WN64A(&var, val * 0x0101010101010101ULL);  break; \

  2021.     case 16: { \

  2022.         uint64_t v64 = val * 0x0101010101010101ULL; \

  2023.         AV_WN64A(              &var,     v64); \

  2024.         AV_WN64A(&((uint8_t *) &var)[8], v64); \

  2025.         break; \

  2026.     } \

  2027.     }

  2028. #else

  2029. #define SPLAT_CTX(var, val, n) \

  2030.     switch (n) { \

  2031.     case 1:  var = val;                         break; \

  2032.     case 2:  AV_WN16A(&var, val *     0x0101);  break; \

  2033.     case 4:  AV_WN32A(&var, val * 0x01010101);  break; \

  2034.     case 8: { \

  2035.         uint32_t v32 = val * 0x01010101; \

  2036.         AV_WN32A(              &var,     v32); \

  2037.         AV_WN32A(&((uint8_t *) &var)[4], v32); \

  2038.         break; \

  2039.     } \

  2040.     case 16: { \

  2041.         uint32_t v32 = val * 0x01010101; \

  2042.         AV_WN32A(              &var,      v32); \

  2043.         AV_WN32A(&((uint8_t *) &var)[4],  v32); \

  2044.         AV_WN32A(&((uint8_t *) &var)[8],  v32); \

  2045.         AV_WN32A(&((uint8_t *) &var)[12], v32); \

  2046.         break; \

  2047.     } \

  2048.     }

  2049. #endif

  2050. 

  2051.     switch (bwh_tab[1][b->bs][0]) {

  2052. #define SET_CTXS(dir, off, n) \

  2053.     do { \

  2054.         SPLAT_CTX(s->dir##_skip_ctx[off],      b->skip,          n); \

  2055.         SPLAT_CTX(s->dir##_txfm_ctx[off],      b->tx,            n); \

  2056.         SPLAT_CTX(s->dir##_partition_ctx[off], dir##_ctx[b->bs], n); \

  2057.         if (!s->keyframe && !s->intraonly) { \

  2058.             SPLAT_CTX(s->dir##_intra_ctx[off], b->intra,   n); \

  2059.             SPLAT_CTX(s->dir##_comp_ctx[off],  b->comp,    n); \

  2060.             SPLAT_CTX(s->dir##_mode_ctx[off],  b->mode[3], n); \

  2061.             if (!b->intra) { \

  2062.                 SPLAT_CTX(s->dir##_ref_ctx[off], vref, n); \

  2063.                 if (s->filtermode == FILTER_SWITCHABLE) { \

  2064.                     SPLAT_CTX(s->dir##_filter_ctx[off], filter_id, n); \

  2065.                 } \

  2066.             } \

  2067.         } \

  2068.     } while (0)

  2069.     case 1: SET_CTXS(above, col, 1); break;

  2070.     case 2: SET_CTXS(above, col, 2); break;

  2071.     case 4: SET_CTXS(above, col, 4); break;

  2072.     case 8: SET_CTXS(above, col, 8); break;

  2073.     }

  2074.     switch (bwh_tab[1][b->bs][1]) {

  2075.     case 1: SET_CTXS(left, row7, 1); break;

  2076.     case 2: SET_CTXS(left, row7, 2); break;

  2077.     case 4: SET_CTXS(left, row7, 4); break;

  2078.     case 8: SET_CTXS(left, row7, 8); break;

  2079.     }

  2080. #undef SPLAT_CTX

  2081. #undef SET_CTXS

  2082. 

  2083.     if (!s->keyframe && !s->intraonly) {

  2084.         if (b->bs > BS_8x8) {

  2085.             int mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  2086. 

  2087.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][0], &b->mv[1][0]);

  2088.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][1], &b->mv[1][1]);

  2089.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][0], mv0);

  2090.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][1], mv1);

  2091.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][0], &b->mv[2][0]);

  2092.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][1], &b->mv[2][1]);

  2093.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][0], mv0);

  2094.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][1], mv1);

  2095.         } else {

  2096.             int n, mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  2097. 

  2098.             for (n = 0; n < w4 * 2; n++) {

  2099.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][0], mv0);

  2100.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][1], mv1);

  2101.             }

  2102.             for (n = 0; n < h4 * 2; n++) {

  2103.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][0], mv0);

  2104.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][1], mv1);

  2105.             }

  2106.         }

  2107.     }

  2108. 

  2109.     // FIXME kinda ugly

  2110.     for (y = 0; y < h4; y++) {

  2111.         int x, o = (row + y) * s->sb_cols * 8 + col;

  2112.         struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[o];

  2113. 

  2114.         if (b->intra) {

  2115.             for (x = 0; x < w4; x++) {

  2116.                 mv[x].ref[0] =

  2117.                 mv[x].ref[1] = -1;

  2118.             }

  2119.         } else if (b->comp) {

  2120.             for (x = 0; x < w4; x++) {

  2121.                 mv[x].ref[0] = b->ref[0];

  2122.                 mv[x].ref[1] = b->ref[1];

  2123.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  2124.                 AV_COPY32(&mv[x].mv[1], &b->mv[3][1]);

  2125.             }

  2126.         } else {

  2127.             for (x = 0; x < w4; x++) {

  2128.                 mv[x].ref[0] = b->ref[0];

  2129.                 mv[x].ref[1] = -1;

  2130.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  2131.             }

  2132.         }

  2133.     }

  2134. }

  2135. 

  2136. // FIXME merge cnt/eob arguments?

  2137. static av_always_inline int

  2138. decode_coeffs_b_generic(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  2139.                         int is_tx32x32, int is8bitsperpixel, int bpp, unsigned (*cnt)[6][3],

  2140.                         unsigned (*eob)[6][2], uint8_t (*p)[6][11],

  2141.                         int nnz, const int16_t *scan, const int16_t (*nb)[2],

  2142.                         const int16_t *band_counts, const int16_t *qmul)

  2143. {

  2144.     int i = 0, band = 0, band_left = band_counts[band];

  2145.     uint8_t *tp = p[0][nnz];

  2146.     uint8_t cache[1024];

  2147. 

  2148.     do {

  2149.         int val, rc;

  2150. 

  2151.         val = vp56_rac_get_prob_branchy(c, tp[0]); // eob

  2152.         eob[band][nnz][val]++;

  2153.         if (!val)

  2154.             break;

  2155. 

  2156.     skip_eob:

  2157.         if (!vp56_rac_get_prob_branchy(c, tp[1])) { // zero

  2158.             cnt[band][nnz][0]++;

  2159.             if (!--band_left)

  2160.                 band_left = band_counts[++band];

  2161.             cache[scan[i]] = 0;

  2162.             nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  2163.             tp = p[band][nnz];

  2164.             if (++i == n_coeffs)

  2165.                 break; //invalid input; blocks should end with EOB

  2166.             goto skip_eob;

  2167.         }

  2168. 

  2169.         rc = scan[i];

  2170.         if (!vp56_rac_get_prob_branchy(c, tp[2])) { // one

  2171.             cnt[band][nnz][1]++;

  2172.             val = 1;

  2173.             cache[rc] = 1;

  2174.         } else {

  2175.             // fill in p[3-10] (model fill) - only once per frame for each pos

  2176.             if (!tp[3])

  2177.                 memcpy(&tp[3], vp9_model_pareto8[tp[2]], 8);

  2178. 

  2179.             cnt[band][nnz][2]++;

  2180.             if (!vp56_rac_get_prob_branchy(c, tp[3])) { // 2, 3, 4

  2181.                 if (!vp56_rac_get_prob_branchy(c, tp[4])) {

  2182.                     cache[rc] = val = 2;

  2183.                 } else {

  2184.                     val = 3 + vp56_rac_get_prob(c, tp[5]);

  2185.                     cache[rc] = 3;

  2186.                 }

  2187.             } else if (!vp56_rac_get_prob_branchy(c, tp[6])) { // cat1/2

  2188.                 cache[rc] = 4;

  2189.                 if (!vp56_rac_get_prob_branchy(c, tp[7])) {

  2190.                     val = 5 + vp56_rac_get_prob(c, 159);

  2191.                 } else {

  2192.                     val  = 7 + (vp56_rac_get_prob(c, 165) << 1);

  2193.                     val +=      vp56_rac_get_prob(c, 145);

  2194.                 }

  2195.             } else { // cat 3-6

  2196.                 cache[rc] = 5;

  2197.                 if (!vp56_rac_get_prob_branchy(c, tp[8])) {

  2198.                     if (!vp56_rac_get_prob_branchy(c, tp[9])) {

  2199.                         val  = 11 + (vp56_rac_get_prob(c, 173) << 2);

  2200.                         val +=      (vp56_rac_get_prob(c, 148) << 1);

  2201.                         val +=       vp56_rac_get_prob(c, 140);

  2202.                     } else {

  2203.                         val  = 19 + (vp56_rac_get_prob(c, 176) << 3);

  2204.                         val +=      (vp56_rac_get_prob(c, 155) << 2);

  2205.                         val +=      (vp56_rac_get_prob(c, 140) << 1);

  2206.                         val +=       vp56_rac_get_prob(c, 135);

  2207.                     }

  2208.                 } else if (!vp56_rac_get_prob_branchy(c, tp[10])) {

  2209.                     val  = 35 + (vp56_rac_get_prob(c, 180) << 4);

  2210.                     val +=      (vp56_rac_get_prob(c, 157) << 3);

  2211.                     val +=      (vp56_rac_get_prob(c, 141) << 2);

  2212.                     val +=      (vp56_rac_get_prob(c, 134) << 1);

  2213.                     val +=       vp56_rac_get_prob(c, 130);

  2214.                 } else {

  2215.                     val = 67;

  2216.                     if (!is8bitsperpixel) {

  2217.                         if (bpp == 12) {

  2218.                             val += vp56_rac_get_prob(c, 255) << 17;

  2219.                             val += vp56_rac_get_prob(c, 255) << 16;

  2220.                         }

  2221.                         val +=  (vp56_rac_get_prob(c, 255) << 15);

  2222.                         val +=  (vp56_rac_get_prob(c, 255) << 14);

  2223.                     }

  2224.                     val +=      (vp56_rac_get_prob(c, 254) << 13);

  2225.                     val +=      (vp56_rac_get_prob(c, 254) << 12);

  2226.                     val +=      (vp56_rac_get_prob(c, 254) << 11);

  2227.                     val +=      (vp56_rac_get_prob(c, 252) << 10);

  2228.                     val +=      (vp56_rac_get_prob(c, 249) << 9);

  2229.                     val +=      (vp56_rac_get_prob(c, 243) << 8);

  2230.                     val +=      (vp56_rac_get_prob(c, 230) << 7);

  2231.                     val +=      (vp56_rac_get_prob(c, 196) << 6);

  2232.                     val +=      (vp56_rac_get_prob(c, 177) << 5);

  2233.                     val +=      (vp56_rac_get_prob(c, 153) << 4);

  2234.                     val +=      (vp56_rac_get_prob(c, 140) << 3);

  2235.                     val +=      (vp56_rac_get_prob(c, 133) << 2);

  2236.                     val +=      (vp56_rac_get_prob(c, 130) << 1);

  2237.                     val +=       vp56_rac_get_prob(c, 129);

  2238.                 }

  2239.             }

  2240.         }

  2241. #define STORE_COEF(c, i, v) do { \

  2242.     if (is8bitsperpixel) { \

  2243.         c[i] = v; \

  2244.     } else { \

  2245.         AV_WN32A(&c[i * 2], v); \

  2246.     } \

  2247. } while (0)

  2248.         if (!--band_left)

  2249.             band_left = band_counts[++band];

  2250.         if (is_tx32x32)

  2251.             STORE_COEF(coef, rc, ((vp8_rac_get(c) ? -val : val) * qmul[!!i]) / 2);

  2252.         else

  2253.             STORE_COEF(coef, rc, (vp8_rac_get(c) ? -val : val) * qmul[!!i]);

  2254.         nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  2255.         tp = p[band][nnz];

  2256.     } while (++i < n_coeffs);

  2257. 

  2258.     return i;

  2259. }

  2260. 

  2261. static int decode_coeffs_b_8bpp(VP9Context *s, int16_t *coef, int n_coeffs,

  2262.                                 unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  2263.                                 uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  2264.                                 const int16_t (*nb)[2], const int16_t *band_counts,

  2265.                                 const int16_t *qmul)

  2266. {

  2267.     return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 0, 1, 8, cnt, eob, p,

  2268.                                    nnz, scan, nb, band_counts, qmul);

  2269. }

  2270. 

  2271. static int decode_coeffs_b32_8bpp(VP9Context *s, int16_t *coef, int n_coeffs,

  2272.                                   unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  2273.                                   uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  2274.                                   const int16_t (*nb)[2], const int16_t *band_counts,

  2275.                                   const int16_t *qmul)

  2276. {

  2277.     return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 1, 1, 8, cnt, eob, p,

  2278.                                    nnz, scan, nb, band_counts, qmul);

  2279. }

  2280. 

  2281. static int decode_coeffs_b_16bpp(VP9Context *s, int16_t *coef, int n_coeffs,

  2282.                                  unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  2283.                                  uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  2284.                                  const int16_t (*nb)[2], const int16_t *band_counts,

  2285.                                  const int16_t *qmul)

  2286. {

  2287.     return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 0, 0, s->bpp, cnt, eob, p,

  2288.                                    nnz, scan, nb, band_counts, qmul);

  2289. }

  2290. 

  2291. static int decode_coeffs_b32_16bpp(VP9Context *s, int16_t *coef, int n_coeffs,

  2292.                                    unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  2293.                                    uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  2294.                                    const int16_t (*nb)[2], const int16_t *band_counts,

  2295.                                    const int16_t *qmul)

  2296. {

  2297.     return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 1, 0, s->bpp, cnt, eob, p,

  2298.                                    nnz, scan, nb, band_counts, qmul);

  2299. }

  2300. 

  2301. static av_always_inline void decode_coeffs(AVCodecContext *ctx, int is8bitsperpixel)

  2302. {

  2303.     VP9Context *s = ctx->priv_data;

  2304.     VP9Block *b = s->b;

  2305.     int row = s->row, col = s->col;

  2306.     uint8_t (*p)[6][11] = s->prob.coef[b->tx][0 /* y */][!b->intra];

  2307.     unsigned (*c)[6][3] = s->counts.coef[b->tx][0 /* y */][!b->intra];

  2308.     unsigned (*e)[6][2] = s->counts.eob[b->tx][0 /* y */][!b->intra];

  2309.     int w4 = bwh_tab[1][b->bs][0] << 1, h4 = bwh_tab[1][b->bs][1] << 1;

  2310.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2311.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2312.     int n, pl, x, y, res;

  2313.     int16_t (*qmul)[2] = s->segmentation.feat[b->seg_id].qmul;

  2314.     int tx = 4 * s->lossless + b->tx;

  2315.     const int16_t * const *yscans = vp9_scans[tx];

  2316.     const int16_t (* const *ynbs)[2] = vp9_scans_nb[tx];

  2317.     const int16_t *uvscan = vp9_scans[b->uvtx][DCT_DCT];

  2318.     const int16_t (*uvnb)[2] = vp9_scans_nb[b->uvtx][DCT_DCT];

  2319.     uint8_t *a = &s->above_y_nnz_ctx[col * 2];

  2320.     uint8_t *l = &s->left_y_nnz_ctx[(row & 7) << 1];

  2321.     static const int16_t band_counts[4][8] = {

  2322.         { 1, 2, 3, 4,  3,   16 - 13 },

  2323.         { 1, 2, 3, 4, 11,   64 - 21 },

  2324.         { 1, 2, 3, 4, 11,  256 - 21 },

  2325.         { 1, 2, 3, 4, 11, 1024 - 21 },

  2326.     };

  2327.     const int16_t *y_band_counts = band_counts[b->tx];

  2328.     const int16_t *uv_band_counts = band_counts[b->uvtx];

  2329.     int bytesperpixel = is8bitsperpixel ? 1 : 2;

  2330. 

  2331. #define MERGE(la, end, step, rd) \

  2332.     for (n = 0; n < end; n += step) \

  2333.         la[n] = !!rd(&la[n])

  2334. #define MERGE_CTX(step, rd) \

  2335.     do { \

  2336.         MERGE(l, end_y, step, rd); \

  2337.         MERGE(a, end_x, step, rd); \

  2338.     } while (0)

  2339. 

  2340. #define DECODE_Y_COEF_LOOP(step, mode_index, v) \

  2341.     for (n = 0, y = 0; y < end_y; y += step) { \

  2342.         for (x = 0; x < end_x; x += step, n += step * step) { \

  2343.             enum TxfmType txtp = vp9_intra_txfm_type[b->mode[mode_index]]; \

  2344.             res = (is8bitsperpixel ? decode_coeffs_b##v##_8bpp : decode_coeffs_b##v##_16bpp) \

  2345.                                     (s, s->block + 16 * n * bytesperpixel, 16 * step * step, \

  2346.                                      c, e, p, a[x] + l[y], yscans[txtp], \

  2347.                                      ynbs[txtp], y_band_counts, qmul[0]); \

  2348.             a[x] = l[y] = !!res; \

  2349.             if (step >= 4) { \

  2350.                 AV_WN16A(&s->eob[n], res); \

  2351.             } else { \

  2352.                 s->eob[n] = res; \

  2353.             } \

  2354.         } \

  2355.     }

  2356. 

  2357. #define SPLAT(la, end, step, cond) \

  2358.     if (step == 2) { \

  2359.         for (n = 1; n < end; n += step) \

  2360.             la[n] = la[n - 1]; \

  2361.     } else if (step == 4) { \

  2362.         if (cond) { \

  2363.             for (n = 0; n < end; n += step) \

  2364.                 AV_WN32A(&la[n], la[n] * 0x01010101); \

  2365.         } else { \

  2366.             for (n = 0; n < end; n += step) \

  2367.                 memset(&la[n + 1], la[n], FFMIN(end - n - 1, 3)); \

  2368.         } \

  2369.     } else /* step == 8 */ { \

  2370.         if (cond) { \

  2371.             if (HAVE_FAST_64BIT) { \

  2372.                 for (n = 0; n < end; n += step) \

  2373.                     AV_WN64A(&la[n], la[n] * 0x0101010101010101ULL); \

  2374.             } else { \

  2375.                 for (n = 0; n < end; n += step) { \

  2376.                     uint32_t v32 = la[n] * 0x01010101; \

  2377.                     AV_WN32A(&la[n],     v32); \

  2378.                     AV_WN32A(&la[n + 4], v32); \

  2379.                 } \

  2380.             } \

  2381.         } else { \

  2382.             for (n = 0; n < end; n += step) \

  2383.                 memset(&la[n + 1], la[n], FFMIN(end - n - 1, 7)); \

  2384.         } \

  2385.     }

  2386. #define SPLAT_CTX(step) \

  2387.     do { \

  2388.         SPLAT(a, end_x, step, end_x == w4); \

  2389.         SPLAT(l, end_y, step, end_y == h4); \

  2390.     } while (0)

  2391. 

  2392.     /* y tokens */

  2393.     switch (b->tx) {

  2394.     case TX_4X4:

  2395.         DECODE_Y_COEF_LOOP(1, b->bs > BS_8x8 ? n : 0,);

  2396.         break;

  2397.     case TX_8X8:

  2398.         MERGE_CTX(2, AV_RN16A);

  2399.         DECODE_Y_COEF_LOOP(2, 0,);

  2400.         SPLAT_CTX(2);

  2401.         break;

  2402.     case TX_16X16:

  2403.         MERGE_CTX(4, AV_RN32A);

  2404.         DECODE_Y_COEF_LOOP(4, 0,);

  2405.         SPLAT_CTX(4);

  2406.         break;

  2407.     case TX_32X32:

  2408.         MERGE_CTX(8, AV_RN64A);

  2409.         DECODE_Y_COEF_LOOP(8, 0, 32);

  2410.         SPLAT_CTX(8);

  2411.         break;

  2412.     }

  2413. 

  2414. #define DECODE_UV_COEF_LOOP(step, v) \

  2415.     for (n = 0, y = 0; y < end_y; y += step) { \

  2416.         for (x = 0; x < end_x; x += step, n += step * step) { \

  2417.             res = (is8bitsperpixel ? decode_coeffs_b##v##_8bpp : decode_coeffs_b##v##_16bpp) \

  2418.                                     (s, s->uvblock[pl] + 16 * n * bytesperpixel, \

  2419.                                      16 * step * step, c, e, p, a[x] + l[y], \

  2420.                                      uvscan, uvnb, uv_band_counts, qmul[1]); \

  2421.             a[x] = l[y] = !!res; \

  2422.             if (step >= 4) { \

  2423.                 AV_WN16A(&s->uveob[pl][n], res); \

  2424.             } else { \

  2425.                 s->uveob[pl][n] = res; \

  2426.             } \

  2427.         } \

  2428.     }

  2429. 

  2430.     p = s->prob.coef[b->uvtx][1 /* uv */][!b->intra];

  2431.     c = s->counts.coef[b->uvtx][1 /* uv */][!b->intra];

  2432.     e = s->counts.eob[b->uvtx][1 /* uv */][!b->intra];

  2433.     w4 >>= s->ss_h;

  2434.     end_x >>= s->ss_h;

  2435.     h4 >>= s->ss_v;

  2436.     end_y >>= s->ss_v;

  2437.     for (pl = 0; pl < 2; pl++) {

  2438.         a = &s->above_uv_nnz_ctx[pl][col << !s->ss_h];

  2439.         l = &s->left_uv_nnz_ctx[pl][(row & 7) << !s->ss_v];

  2440.         switch (b->uvtx) {

  2441.         case TX_4X4:

  2442.             DECODE_UV_COEF_LOOP(1,);

  2443.             break;

  2444.         case TX_8X8:

  2445.             MERGE_CTX(2, AV_RN16A);

  2446.             DECODE_UV_COEF_LOOP(2,);

  2447.             SPLAT_CTX(2);

  2448.             break;

  2449.         case TX_16X16:

  2450.             MERGE_CTX(4, AV_RN32A);

  2451.             DECODE_UV_COEF_LOOP(4,);

  2452.             SPLAT_CTX(4);

  2453.             break;

  2454.         case TX_32X32:

  2455.             MERGE_CTX(8, AV_RN64A);

  2456.             DECODE_UV_COEF_LOOP(8, 32);

  2457.             SPLAT_CTX(8);

  2458.             break;

  2459.         }

  2460.     }

  2461. }

  2462. 

  2463. static void decode_coeffs_8bpp(AVCodecContext *ctx)

  2464. {

  2465.     decode_coeffs(ctx, 1);

  2466. }

  2467. 

  2468. static void decode_coeffs_16bpp(AVCodecContext *ctx)

  2469. {

  2470.     decode_coeffs(ctx, 0);

  2471. }

  2472. 

  2473. static av_always_inline int check_intra_mode(VP9Context *s, int mode, uint8_t **a,

  2474.                                              uint8_t *dst_edge, ptrdiff_t stride_edge,

  2475.                                              uint8_t *dst_inner, ptrdiff_t stride_inner,

  2476.                                              uint8_t *l, int col, int x, int w,

  2477.                                              int row, int y, enum TxfmMode tx,

  2478.                                              int p, int ss_h, int ss_v, int bytesperpixel)

  2479. {

  2480.     int have_top = row > 0 || y > 0;

  2481.     int have_left = col > s->tiling.tile_col_start || x > 0;

  2482.     int have_right = x < w - 1;

  2483.     int bpp = s->bpp;

  2484.     static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {

  2485.         [VERT_PRED]            = { { DC_127_PRED,          VERT_PRED },

  2486.                                    { DC_127_PRED,          VERT_PRED } },

  2487.         [HOR_PRED]             = { { DC_129_PRED,          DC_129_PRED },

  2488.                                    { HOR_PRED,             HOR_PRED } },

  2489.         [DC_PRED]              = { { DC_128_PRED,          TOP_DC_PRED },

  2490.                                    { LEFT_DC_PRED,         DC_PRED } },

  2491.         [DIAG_DOWN_LEFT_PRED]  = { { DC_127_PRED,          DIAG_DOWN_LEFT_PRED },

  2492.                                    { DC_127_PRED,          DIAG_DOWN_LEFT_PRED } },

  2493.         [DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },

  2494.                                    { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },

  2495.         [VERT_RIGHT_PRED]      = { { VERT_RIGHT_PRED,      VERT_RIGHT_PRED },

  2496.                                    { VERT_RIGHT_PRED,      VERT_RIGHT_PRED } },

  2497.         [HOR_DOWN_PRED]        = { { HOR_DOWN_PRED,        HOR_DOWN_PRED },

  2498.                                    { HOR_DOWN_PRED,        HOR_DOWN_PRED } },

  2499.         [VERT_LEFT_PRED]       = { { DC_127_PRED,          VERT_LEFT_PRED },

  2500.                                    { DC_127_PRED,          VERT_LEFT_PRED } },

  2501.         [HOR_UP_PRED]          = { { DC_129_PRED,          DC_129_PRED },

  2502.                                    { HOR_UP_PRED,          HOR_UP_PRED } },

  2503.         [TM_VP8_PRED]          = { { DC_129_PRED,          VERT_PRED },

  2504.                                    { HOR_PRED,             TM_VP8_PRED } },

  2505.     };

  2506.     static const struct {

  2507.         uint8_t needs_left:1;

  2508.         uint8_t needs_top:1;

  2509.         uint8_t needs_topleft:1;

  2510.         uint8_t needs_topright:1;

  2511.         uint8_t invert_left:1;

  2512.     } edges[N_INTRA_PRED_MODES] = {

  2513.         [VERT_PRED]            = { .needs_top  = 1 },

  2514.         [HOR_PRED]             = { .needs_left = 1 },

  2515.         [DC_PRED]              = { .needs_top  = 1, .needs_left = 1 },

  2516.         [DIAG_DOWN_LEFT_PRED]  = { .needs_top  = 1, .needs_topright = 1 },

  2517.         [DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2518.         [VERT_RIGHT_PRED]      = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2519.         [HOR_DOWN_PRED]        = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2520.         [VERT_LEFT_PRED]       = { .needs_top  = 1, .needs_topright = 1 },

  2521.         [HOR_UP_PRED]          = { .needs_left = 1, .invert_left = 1 },

  2522.         [TM_VP8_PRED]          = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2523.         [LEFT_DC_PRED]         = { .needs_left = 1 },

  2524.         [TOP_DC_PRED]          = { .needs_top  = 1 },

  2525.         [DC_128_PRED]          = { 0 },

  2526.         [DC_127_PRED]          = { 0 },

  2527.         [DC_129_PRED]          = { 0 }

  2528.     };

  2529. 

  2530.     av_assert2(mode >= 0 && mode < 10);

  2531.     mode = mode_conv[mode][have_left][have_top];

  2532.     if (edges[mode].needs_top) {

  2533.         uint8_t *top, *topleft;

  2534.         int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !ss_h) - x) * 4;

  2535.         int n_px_need_tr = 0;

  2536. 

  2537.         if (tx == TX_4X4 && edges[mode].needs_topright && have_right)

  2538.             n_px_need_tr = 4;

  2539. 

  2540.         // if top of sb64-row, use s->intra_pred_data[] instead of

  2541.         // dst[-stride] for intra prediction (it contains pre- instead of

  2542.         // post-loopfilter data)

  2543.         if (have_top) {

  2544.             top = !(row & 7) && !y ?

  2545.                 s->intra_pred_data[p] + (col * (8 >> ss_h) + x * 4) * bytesperpixel :

  2546.                 y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];

  2547.             if (have_left)

  2548.                 topleft = !(row & 7) && !y ?

  2549.                     s->intra_pred_data[p] + (col * (8 >> ss_h) + x * 4) * bytesperpixel :

  2550.                     y == 0 || x == 0 ? &dst_edge[-stride_edge] :

  2551.                     &dst_inner[-stride_inner];

  2552.         }

  2553. 

  2554.         if (have_top &&

  2555.             (!edges[mode].needs_topleft || (have_left && top == topleft)) &&

  2556.             (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&

  2557.             n_px_need + n_px_need_tr <= n_px_have) {

  2558.             *a = top;

  2559.         } else {

  2560.             if (have_top) {

  2561.                 if (n_px_need <= n_px_have) {

  2562.                     memcpy(*a, top, n_px_need * bytesperpixel);

  2563.                 } else {

  2564. #define memset_bpp(c, i1, v, i2, num) do { \

  2565.     if (bytesperpixel == 1) { \

  2566.         memset(&(c)[(i1)], (v)[(i2)], (num)); \

  2567.     } else { \

  2568.         int n, val = AV_RN16A(&(v)[(i2) * 2]); \

  2569.         for (n = 0; n < (num); n++) { \

  2570.             AV_WN16A(&(c)[((i1) + n) * 2], val); \

  2571.         } \

  2572.     } \

  2573. } while (0)

  2574.                     memcpy(*a, top, n_px_have * bytesperpixel);

  2575.                     memset_bpp(*a, n_px_have, (*a), n_px_have - 1, n_px_need - n_px_have);

  2576.                 }

  2577.             } else {

  2578. #define memset_val(c, val, num) do { \

  2579.     if (bytesperpixel == 1) { \

  2580.         memset((c), (val), (num)); \

  2581.     } else { \

  2582.         int n; \

  2583.         for (n = 0; n < (num); n++) { \

  2584.             AV_WN16A(&(c)[n * 2], (val)); \

  2585.         } \

  2586.     } \

  2587. } while (0)

  2588.                 memset_val(*a, (128 << (bpp - 8)) - 1, n_px_need);

  2589.             }

  2590.             if (edges[mode].needs_topleft) {

  2591.                 if (have_left && have_top) {

  2592. #define assign_bpp(c, i1, v, i2) do { \

  2593.     if (bytesperpixel == 1) { \

  2594.         (c)[(i1)] = (v)[(i2)]; \

  2595.     } else { \

  2596.         AV_COPY16(&(c)[(i1) * 2], &(v)[(i2) * 2]); \

  2597.     } \

  2598. } while (0)

  2599.                     assign_bpp(*a, -1, topleft, -1);

  2600.                 } else {

  2601. #define assign_val(c, i, v) do { \

  2602.     if (bytesperpixel == 1) { \

  2603.         (c)[(i)] = (v); \

  2604.     } else { \

  2605.         AV_WN16A(&(c)[(i) * 2], (v)); \

  2606.     } \

  2607. } while (0)

  2608.                     assign_val((*a), -1, (128 << (bpp - 8)) + (have_top ? +1 : -1));

  2609.                 }

  2610.             }

  2611.             if (tx == TX_4X4 && edges[mode].needs_topright) {

  2612.                 if (have_top && have_right &&

  2613.                     n_px_need + n_px_need_tr <= n_px_have) {

  2614.                     memcpy(&(*a)[4 * bytesperpixel], &top[4 * bytesperpixel], 4 * bytesperpixel);

  2615.                 } else {

  2616.                     memset_bpp(*a, 4, *a, 3, 4);

  2617.                 }

  2618.             }

  2619.         }

  2620.     }

  2621.     if (edges[mode].needs_left) {

  2622.         if (have_left) {

  2623.             int n_px_need = 4 << tx, i, n_px_have = (((s->rows - row) << !ss_v) - y) * 4;

  2624.             uint8_t *dst = x == 0 ? dst_edge : dst_inner;

  2625.             ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;

  2626. 

  2627.             if (edges[mode].invert_left) {

  2628.                 if (n_px_need <= n_px_have) {

  2629.                     for (i = 0; i < n_px_need; i++)

  2630.                         assign_bpp(l, i, &dst[i * stride], -1);

  2631.                 } else {

  2632.                     for (i = 0; i < n_px_have; i++)

  2633.                         assign_bpp(l, i, &dst[i * stride], -1);

  2634.                     memset_bpp(l, n_px_have, l, n_px_have - 1, n_px_need - n_px_have);

  2635.                 }

  2636.             } else {

  2637.                 if (n_px_need <= n_px_have) {

  2638.                     for (i = 0; i < n_px_need; i++)

  2639.                         assign_bpp(l, n_px_need - 1 - i, &dst[i * stride], -1);

  2640.                 } else {

  2641.                     for (i = 0; i < n_px_have; i++)

  2642.                         assign_bpp(l, n_px_need - 1 - i, &dst[i * stride], -1);

  2643.                     memset_bpp(l, 0, l, n_px_need - n_px_have, n_px_need - n_px_have);

  2644.                 }

  2645.             }

  2646.         } else {

  2647.             memset_val(l, (128 << (bpp - 8)) + 1, 4 << tx);

  2648.         }

  2649.     }

  2650. 

  2651.     return mode;

  2652. }

  2653. 

  2654. static av_always_inline void intra_recon(AVCodecContext *ctx, ptrdiff_t y_off,

  2655.                                          ptrdiff_t uv_off, int bytesperpixel)

  2656. {

  2657.     VP9Context *s = ctx->priv_data;

  2658.     VP9Block *b = s->b;

  2659.     int row = s->row, col = s->col;

  2660.     int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2661.     int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2662.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2663.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2664.     int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2665.     int uvstep1d = 1 << b->uvtx, p;

  2666.     uint8_t *dst = s->dst[0], *dst_r = s->frames[CUR_FRAME].tf.f->data[0] + y_off;

  2667.     LOCAL_ALIGNED_32(uint8_t, a_buf, [96]);

  2668.     LOCAL_ALIGNED_32(uint8_t, l, [64]);

  2669. 

  2670.     for (n = 0, y = 0; y < end_y; y += step1d) {

  2671.         uint8_t *ptr = dst, *ptr_r = dst_r;

  2672.         for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d * bytesperpixel,

  2673.                                ptr_r += 4 * step1d * bytesperpixel, n += step) {

  2674.             int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?

  2675.                                y * 2 + x : 0];

  2676.             uint8_t *a = &a_buf[32];

  2677.             enum TxfmType txtp = vp9_intra_txfm_type[mode];

  2678.             int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2679. 

  2680.             mode = check_intra_mode(s, mode, &a, ptr_r,

  2681.                                     s->frames[CUR_FRAME].tf.f->linesize[0],

  2682.                                     ptr, s->y_stride, l,

  2683.                                     col, x, w4, row, y, b->tx, 0, 0, 0, bytesperpixel);

  2684.             s->dsp.intra_pred[b->tx][mode](ptr, s->y_stride, l, a);

  2685.             if (eob)

  2686.                 s->dsp.itxfm_add[tx][txtp](ptr, s->y_stride,

  2687.                                            s->block + 16 * n * bytesperpixel, eob);

  2688.         }

  2689.         dst_r += 4 * step1d * s->frames[CUR_FRAME].tf.f->linesize[0];

  2690.         dst   += 4 * step1d * s->y_stride;

  2691.     }

  2692. 

  2693.     // U/V

  2694.     w4 >>= s->ss_h;

  2695.     end_x >>= s->ss_h;

  2696.     end_y >>= s->ss_v;

  2697.     step = 1 << (b->uvtx * 2);

  2698.     for (p = 0; p < 2; p++) {

  2699.         dst   = s->dst[1 + p];

  2700.         dst_r = s->frames[CUR_FRAME].tf.f->data[1 + p] + uv_off;

  2701.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2702.             uint8_t *ptr = dst, *ptr_r = dst_r;

  2703.             for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d * bytesperpixel,

  2704.                                    ptr_r += 4 * uvstep1d * bytesperpixel, n += step) {

  2705.                 int mode = b->uvmode;

  2706.                 uint8_t *a = &a_buf[32];

  2707.                 int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2708. 

  2709.                 mode = check_intra_mode(s, mode, &a, ptr_r,

  2710.                                         s->frames[CUR_FRAME].tf.f->linesize[1],

  2711.                                         ptr, s->uv_stride, l, col, x, w4, row, y,

  2712.                                         b->uvtx, p + 1, s->ss_h, s->ss_v, bytesperpixel);

  2713.                 s->dsp.intra_pred[b->uvtx][mode](ptr, s->uv_stride, l, a);

  2714.                 if (eob)

  2715.                     s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2716.                                                     s->uvblock[p] + 16 * n * bytesperpixel, eob);

  2717.             }

  2718.             dst_r += 4 * uvstep1d * s->frames[CUR_FRAME].tf.f->linesize[1];

  2719.             dst   += 4 * uvstep1d * s->uv_stride;

  2720.         }

  2721.     }

  2722. }

  2723. 

  2724. static void intra_recon_8bpp(AVCodecContext *ctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  2725. {

  2726.     intra_recon(ctx, y_off, uv_off, 1);

  2727. }

  2728. 

  2729. static void intra_recon_16bpp(AVCodecContext *ctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  2730. {

  2731.     intra_recon(ctx, y_off, uv_off, 2);

  2732. }

  2733. 

  2734. static av_always_inline void mc_luma_scaled(VP9Context *s, vp9_scaled_mc_func smc,

  2735.                                             uint8_t *dst, ptrdiff_t dst_stride,

  2736.                                             const uint8_t *ref, ptrdiff_t ref_stride,

  2737.                                             ThreadFrame *ref_frame,

  2738.                                             ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2739.                                             int bw, int bh, int w, int h, int bytesperpixel,

  2740.                                             const uint16_t *scale, const uint8_t *step)

  2741. {

  2742. #define scale_mv(n, dim) (((int64_t)(n) * scale[dim]) >> 14)

  2743.     // BUG libvpx seems to scale the two components separately. This introduces

  2744.     // rounding errors but we have to reproduce them to be exactly compatible

  2745.     // with the output from libvpx...

  2746.     int mx = scale_mv(mv->x * 2, 0) + scale_mv(x * 16, 0);

  2747.     int my = scale_mv(mv->y * 2, 1) + scale_mv(y * 16, 1);

  2748.     int refbw_m1, refbh_m1;

  2749.     int th;

  2750. 

  2751.     y = my >> 4;

  2752.     x = mx >> 4;

  2753.     ref += y * ref_stride + x * bytesperpixel;

  2754.     mx &= 15;

  2755.     my &= 15;

  2756.     refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;

  2757.     refbh_m1 = ((bh - 1) * step[1] + my) >> 4;

  2758.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2759.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2760.     // the longest loopfilter of the next sbrow

  2761.     th = (y + refbh_m1 + 4 + 7) >> 6;

  2762.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2763.     if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 4 >= h - refbh_m1) {

  2764.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2765.                                  ref - 3 * ref_stride - 3 * bytesperpixel,

  2766.                                  288, ref_stride,

  2767.                                  refbw_m1 + 8, refbh_m1 + 8,

  2768.                                  x - 3, y - 3, w, h);

  2769.         ref = s->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;

  2770.         ref_stride = 288;

  2771.     }

  2772.     smc(dst, dst_stride, ref, ref_stride, bh, mx, my, step[0], step[1]);

  2773. }

  2774. 

  2775. static av_always_inline void mc_chroma_scaled(VP9Context *s, vp9_scaled_mc_func smc,

  2776.                                               uint8_t *dst_u, uint8_t *dst_v,

  2777.                                               ptrdiff_t dst_stride,

  2778.                                               const uint8_t *ref_u, ptrdiff_t src_stride_u,

  2779.                                               const uint8_t *ref_v, ptrdiff_t src_stride_v,

  2780.                                               ThreadFrame *ref_frame,

  2781.                                               ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2782.                                               int bw, int bh, int w, int h, int bytesperpixel,

  2783.                                               const uint16_t *scale, const uint8_t *step)

  2784. {

  2785.     int mx, my;

  2786.     int refbw_m1, refbh_m1;

  2787.     int th;

  2788. 

  2789.     if (s->ss_h) {

  2790.         // BUG https://code.google.com/p/webm/issues/detail?id=820

  2791.         mx = scale_mv(mv->x, 0) + (scale_mv(x * 16, 0) & ~15) + (scale_mv(x * 32, 0) & 15);

  2792.     } else {

  2793.         mx = scale_mv(mv->x << 1, 0) + scale_mv(x * 16, 0);

  2794.     }

  2795.     if (s->ss_v) {

  2796.         // BUG https://code.google.com/p/webm/issues/detail?id=820

  2797.         my = scale_mv(mv->y, 1) + (scale_mv(y * 16, 1) & ~15) + (scale_mv(y * 32, 1) & 15);

  2798.     } else {

  2799.         my = scale_mv(mv->y << 1, 1) + scale_mv(y * 16, 1);

  2800.     }

  2801. #undef scale_mv

  2802.     y = my >> 4;

  2803.     x = mx >> 4;

  2804.     ref_u += y * src_stride_u + x * bytesperpixel;

  2805.     ref_v += y * src_stride_v + x * bytesperpixel;

  2806.     mx &= 15;

  2807.     my &= 15;

  2808.     refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;

  2809.     refbh_m1 = ((bh - 1) * step[1] + my) >> 4;

  2810.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2811.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2812.     // the longest loopfilter of the next sbrow

  2813.     th = (y + refbh_m1 + 4 + 7) >> (6 - s->ss_v);

  2814.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2815.     if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 4 >= h - refbh_m1) {

  2816.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2817.                                  ref_u - 3 * src_stride_u - 3 * bytesperpixel,

  2818.                                  288, src_stride_u,

  2819.                                  refbw_m1 + 8, refbh_m1 + 8,

  2820.                                  x - 3, y - 3, w, h);

  2821.         ref_u = s->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;

  2822.         smc(dst_u, dst_stride, ref_u, 288, bh, mx, my, step[0], step[1]);

  2823. 

  2824.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2825.                                  ref_v - 3 * src_stride_v - 3 * bytesperpixel,

  2826.                                  288, src_stride_v,

  2827.                                  refbw_m1 + 8, refbh_m1 + 8,

  2828.                                  x - 3, y - 3, w, h);

  2829.         ref_v = s->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;

  2830.         smc(dst_v, dst_stride, ref_v, 288, bh, mx, my, step[0], step[1]);

  2831.     } else {

  2832.         smc(dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my, step[0], step[1]);

  2833.         smc(dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my, step[0], step[1]);

  2834.     }

  2835. }

  2836. 

  2837. #define mc_luma_dir(s, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, bw, bh, w, h, i) \

  2838.     mc_luma_scaled(s, s->dsp.s##mc, dst, dst_ls, src, src_ls, tref, row, col, \

  2839.                    mv, bw, bh, w, h, bytesperpixel, \

  2840.                    s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])

  2841. #define mc_chroma_dir(s, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2842.                       row, col, mv, bw, bh, w, h, i) \

  2843.     mc_chroma_scaled(s, s->dsp.s##mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2844.                      row, col, mv, bw, bh, w, h, bytesperpixel, \

  2845.                      s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])

  2846. #define SCALED 1

  2847. #define FN(x) x##_scaled_8bpp

  2848. #define BYTES_PER_PIXEL 1

  2849. #include "vp9_mc_template.c"

  2850. #undef FN

  2851. #undef BYTES_PER_PIXEL

  2852. #define FN(x) x##_scaled_16bpp

  2853. #define BYTES_PER_PIXEL 2

  2854. #include "vp9_mc_template.c"

  2855. #undef mc_luma_dir

  2856. #undef mc_chroma_dir

  2857. #undef FN

  2858. #undef BYTES_PER_PIXEL

  2859. #undef SCALED

  2860. 

  2861. static av_always_inline void mc_luma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2],

  2862.                                               uint8_t *dst, ptrdiff_t dst_stride,

  2863.                                               const uint8_t *ref, ptrdiff_t ref_stride,

  2864.                                               ThreadFrame *ref_frame,

  2865.                                               ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2866.                                               int bw, int bh, int w, int h, int bytesperpixel)

  2867. {

  2868.     int mx = mv->x, my = mv->y, th;

  2869. 

  2870.     y += my >> 3;

  2871.     x += mx >> 3;

  2872.     ref += y * ref_stride + x * bytesperpixel;

  2873.     mx &= 7;

  2874.     my &= 7;

  2875.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2876.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2877.     // the longest loopfilter of the next sbrow

  2878.     th = (y + bh + 4 * !!my + 7) >> 6;

  2879.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2880.     if (x < !!mx * 3 || y < !!my * 3 ||

  2881.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2882.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2883.                                  ref - !!my * 3 * ref_stride - !!mx * 3 * bytesperpixel,

  2884.                                  160, ref_stride,

  2885.                                  bw + !!mx * 7, bh + !!my * 7,

  2886.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2887.         ref = s->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;

  2888.         ref_stride = 160;

  2889.     }

  2890.     mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);

  2891. }

  2892. 

  2893. static av_always_inline void mc_chroma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2],

  2894.                                                 uint8_t *dst_u, uint8_t *dst_v,

  2895.                                                 ptrdiff_t dst_stride,

  2896.                                                 const uint8_t *ref_u, ptrdiff_t src_stride_u,

  2897.                                                 const uint8_t *ref_v, ptrdiff_t src_stride_v,

  2898.                                                 ThreadFrame *ref_frame,

  2899.                                                 ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2900.                                                 int bw, int bh, int w, int h, int bytesperpixel)

  2901. {

  2902.     int mx = mv->x << !s->ss_h, my = mv->y << !s->ss_v, th;

  2903. 

  2904.     y += my >> 4;

  2905.     x += mx >> 4;

  2906.     ref_u += y * src_stride_u + x * bytesperpixel;

  2907.     ref_v += y * src_stride_v + x * bytesperpixel;

  2908.     mx &= 15;

  2909.     my &= 15;

  2910.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2911.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2912.     // the longest loopfilter of the next sbrow

  2913.     th = (y + bh + 4 * !!my + 7) >> (6 - s->ss_v);

  2914.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2915.     if (x < !!mx * 3 || y < !!my * 3 ||

  2916.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2917.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2918.                                  ref_u - !!my * 3 * src_stride_u - !!mx * 3 * bytesperpixel,

  2919.                                  160, src_stride_u,

  2920.                                  bw + !!mx * 7, bh + !!my * 7,

  2921.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2922.         ref_u = s->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;

  2923.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, 160, bh, mx, my);

  2924. 

  2925.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2926.                                  ref_v - !!my * 3 * src_stride_v - !!mx * 3 * bytesperpixel,

  2927.                                  160, src_stride_v,

  2928.                                  bw + !!mx * 7, bh + !!my * 7,

  2929.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2930.         ref_v = s->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;

  2931.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, 160, bh, mx, my);

  2932.     } else {

  2933.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);

  2934.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);

  2935.     }

  2936. }

  2937. 

  2938. #define mc_luma_dir(s, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, bw, bh, w, h, i) \

  2939.     mc_luma_unscaled(s, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \

  2940.                      mv, bw, bh, w, h, bytesperpixel)

  2941. #define mc_chroma_dir(s, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2942.                       row, col, mv, bw, bh, w, h, i) \

  2943.     mc_chroma_unscaled(s, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2944.                        row, col, mv, bw, bh, w, h, bytesperpixel)

  2945. #define SCALED 0

  2946. #define FN(x) x##_8bpp

  2947. #define BYTES_PER_PIXEL 1

  2948. #include "vp9_mc_template.c"

  2949. #undef FN

  2950. #undef BYTES_PER_PIXEL

  2951. #define FN(x) x##_16bpp

  2952. #define BYTES_PER_PIXEL 2

  2953. #include "vp9_mc_template.c"

  2954. #undef mc_luma_dir_dir

  2955. #undef mc_chroma_dir_dir

  2956. #undef FN

  2957. #undef BYTES_PER_PIXEL

  2958. #undef SCALED

  2959. 

  2960. static av_always_inline void inter_recon(AVCodecContext *ctx, int bytesperpixel)

  2961. {

  2962.     VP9Context *s = ctx->priv_data;

  2963.     VP9Block *b = s->b;

  2964.     int row = s->row, col = s->col;

  2965. 

  2966.     if (s->mvscale[b->ref[0]][0] || (b->comp && s->mvscale[b->ref[1]][0])) {

  2967.         if (bytesperpixel == 1) {

  2968.             inter_pred_scaled_8bpp(ctx);

  2969.         } else {

  2970.             inter_pred_scaled_16bpp(ctx);

  2971.         }

  2972.     } else {

  2973.         if (bytesperpixel == 1) {

  2974.             inter_pred_8bpp(ctx);

  2975.         } else {

  2976.             inter_pred_16bpp(ctx);

  2977.         }

  2978.     }

  2979.     if (!b->skip) {

  2980.         /* mostly copied intra_recon() */

  2981. 

  2982.         int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2983.         int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2984.         int end_x = FFMIN(2 * (s->cols - col), w4);

  2985.         int end_y = FFMIN(2 * (s->rows - row), h4);

  2986.         int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2987.         int uvstep1d = 1 << b->uvtx, p;

  2988.         uint8_t *dst = s->dst[0];

  2989. 

  2990.         // y itxfm add

  2991.         for (n = 0, y = 0; y < end_y; y += step1d) {

  2992.             uint8_t *ptr = dst;

  2993.             for (x = 0; x < end_x; x += step1d,

  2994.                  ptr += 4 * step1d * bytesperpixel, n += step) {

  2995.                 int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2996. 

  2997.                 if (eob)

  2998.                     s->dsp.itxfm_add[tx][DCT_DCT](ptr, s->y_stride,

  2999.                                                   s->block + 16 * n * bytesperpixel, eob);

  3000.             }

  3001.             dst += 4 * s->y_stride * step1d;

  3002.         }

  3003. 

  3004.         // uv itxfm add

  3005.         end_x >>= s->ss_h;

  3006.         end_y >>= s->ss_v;

  3007.         step = 1 << (b->uvtx * 2);

  3008.         for (p = 0; p < 2; p++) {

  3009.             dst = s->dst[p + 1];

  3010.             for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  3011.                 uint8_t *ptr = dst;

  3012.                 for (x = 0; x < end_x; x += uvstep1d,

  3013.                      ptr += 4 * uvstep1d * bytesperpixel, n += step) {

  3014.                     int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  3015. 

  3016.                     if (eob)

  3017.                         s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  3018.                                                         s->uvblock[p] + 16 * n * bytesperpixel, eob);

  3019.                 }

  3020.                 dst += 4 * uvstep1d * s->uv_stride;

  3021.             }

  3022.         }

  3023.     }

  3024. }

  3025. 

  3026. static void inter_recon_8bpp(AVCodecContext *ctx)

  3027. {

  3028.     inter_recon(ctx, 1);

  3029. }

  3030. 

  3031. static void inter_recon_16bpp(AVCodecContext *ctx)

  3032. {

  3033.     inter_recon(ctx, 2);

  3034. }

  3035. 

  3036. static av_always_inline void mask_edges(uint8_t (*mask)[8][4], int ss_h, int ss_v,

  3037.                                         int row_and_7, int col_and_7,

  3038.                                         int w, int h, int col_end, int row_end,

  3039.                                         enum TxfmMode tx, int skip_inter)

  3040. {

  3041.     static const unsigned wide_filter_col_mask[2] = { 0x11, 0x01 };

  3042.     static const unsigned wide_filter_row_mask[2] = { 0x03, 0x07 };

  3043. 

  3044.     // FIXME I'm pretty sure all loops can be replaced by a single LUT if

  3045.     // we make VP9Filter.mask uint64_t (i.e. row/col all single variable)

  3046.     // and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then

  3047.     // use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)

  3048. 

  3049.     // the intended behaviour of the vp9 loopfilter is to work on 8-pixel

  3050.     // edges. This means that for UV, we work on two subsampled blocks at

  3051.     // a time, and we only use the topleft block's mode information to set

  3052.     // things like block strength. Thus, for any block size smaller than

  3053.     // 16x16, ignore the odd portion of the block.

  3054.     if (tx == TX_4X4 && (ss_v | ss_h)) {

  3055.         if (h == ss_v) {

  3056.             if (row_and_7 & 1)

  3057.                 return;

  3058.             if (!row_end)

  3059.                 h += 1;

  3060.         }

  3061.         if (w == ss_h) {

  3062.             if (col_and_7 & 1)

  3063.                 return;

  3064.             if (!col_end)

  3065.                 w += 1;

  3066.         }

  3067.     }

  3068. 

  3069.     if (tx == TX_4X4 && !skip_inter) {

  3070.         int t = 1 << col_and_7, m_col = (t << w) - t, y;

  3071.         // on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide

  3072.         int m_row_8 = m_col & wide_filter_col_mask[ss_h], m_row_4 = m_col - m_row_8;

  3073. 

  3074.         for (y = row_and_7; y < h + row_and_7; y++) {

  3075.             int col_mask_id = 2 - !(y & wide_filter_row_mask[ss_v]);

  3076. 

  3077.             mask[0][y][1] |= m_row_8;

  3078.             mask[0][y][2] |= m_row_4;

  3079.             // for odd lines, if the odd col is not being filtered,

  3080.             // skip odd row also:

  3081.             // .---. <-- a

  3082.             // |   |

  3083.             // |___| <-- b

  3084.             // ^   ^

  3085.             // c   d

  3086.             //

  3087.             // if a/c are even row/col and b/d are odd, and d is skipped,

  3088.             // e.g. right edge of size-66x66.webm, then skip b also (bug)

  3089.             if ((ss_h & ss_v) && (col_end & 1) && (y & 1)) {

  3090.                 mask[1][y][col_mask_id] |= (t << (w - 1)) - t;

  3091.             } else {

  3092.                 mask[1][y][col_mask_id] |= m_col;

  3093.             }

  3094.             if (!ss_h)

  3095.                 mask[0][y][3] |= m_col;

  3096.             if (!ss_v)

  3097.                 mask[1][y][3] |= m_col;

  3098.         }

  3099.     } else {

  3100.         int y, t = 1 << col_and_7, m_col = (t << w) - t;

  3101. 

  3102.         if (!skip_inter) {

  3103.             int mask_id = (tx == TX_8X8);

  3104.             static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };

  3105.             int l2 = tx + ss_h - 1, step1d;

  3106.             int m_row = m_col & masks[l2];

  3107. 

  3108.             // at odd UV col/row edges tx16/tx32 loopfilter edges, force

  3109.             // 8wd loopfilter to prevent going off the visible edge.

  3110.             if (ss_h && tx > TX_8X8 && (w ^ (w - 1)) == 1) {

  3111.                 int m_row_16 = ((t << (w - 1)) - t) & masks[l2];

  3112.                 int m_row_8 = m_row - m_row_16;

  3113. 

  3114.                 for (y = row_and_7; y < h + row_and_7; y++) {

  3115.                     mask[0][y][0] |= m_row_16;

  3116.                     mask[0][y][1] |= m_row_8;

  3117.                 }

  3118.             } else {

  3119.                 for (y = row_and_7; y < h + row_and_7; y++)

  3120.                     mask[0][y][mask_id] |= m_row;

  3121.             }

  3122. 

  3123.             l2 = tx + ss_v - 1;

  3124.             step1d = 1 << l2;

  3125.             if (ss_v && tx > TX_8X8 && (h ^ (h - 1)) == 1) {

  3126.                 for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)

  3127.                     mask[1][y][0] |= m_col;

  3128.                 if (y - row_and_7 == h - 1)

  3129.                     mask[1][y][1] |= m_col;

  3130.             } else {

  3131.                 for (y = row_and_7; y < h + row_and_7; y += step1d)

  3132.                     mask[1][y][mask_id] |= m_col;

  3133.             }

  3134.         } else if (tx != TX_4X4) {

  3135.             int mask_id;

  3136. 

  3137.             mask_id = (tx == TX_8X8) || (h == ss_v);

  3138.             mask[1][row_and_7][mask_id] |= m_col;

  3139.             mask_id = (tx == TX_8X8) || (w == ss_h);

  3140.             for (y = row_and_7; y < h + row_and_7; y++)

  3141.                 mask[0][y][mask_id] |= t;

  3142.         } else {

  3143.             int t8 = t & wide_filter_col_mask[ss_h], t4 = t - t8;

  3144. 

  3145.             for (y = row_and_7; y < h + row_and_7; y++) {

  3146.                 mask[0][y][2] |= t4;

  3147.                 mask[0][y][1] |= t8;

  3148.             }

  3149.             mask[1][row_and_7][2 - !(row_and_7 & wide_filter_row_mask[ss_v])] |= m_col;

  3150.         }

  3151.     }

  3152. }

  3153. 

  3154. static void decode_b(AVCodecContext *ctx, int row, int col,

  3155.                      struct VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,

  3156.                      enum BlockLevel bl, enum BlockPartition bp)

  3157. {

  3158.     VP9Context *s = ctx->priv_data;

  3159.     VP9Block *b = s->b;

  3160.     enum BlockSize bs = bl * 3 + bp;

  3161.     int bytesperpixel = s->bytesperpixel;

  3162.     int w4 = bwh_tab[1][bs][0], h4 = bwh_tab[1][bs][1], lvl;

  3163.     int emu[2];

  3164.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  3165. 

  3166.     s->row = row;

  3167.     s->row7 = row & 7;

  3168.     s->col = col;

  3169.     s->col7 = col & 7;

  3170.     s->min_mv.x = -(128 + col * 64);

  3171.     s->min_mv.y = -(128 + row * 64);

  3172.     s->max_mv.x = 128 + (s->cols - col - w4) * 64;

  3173.     s->max_mv.y = 128 + (s->rows - row - h4) * 64;

  3174.     if (s->pass < 2) {

  3175.         b->bs = bs;

  3176.         b->bl = bl;

  3177.         b->bp = bp;

  3178.         decode_mode(ctx);

  3179.         b->uvtx = b->tx - ((s->ss_h && w4 * 2 == (1 << b->tx)) ||

  3180.                            (s->ss_v && h4 * 2 == (1 << b->tx)));

  3181. 

  3182.         if (!b->skip) {

  3183.             if (bytesperpixel == 1) {

  3184.                 decode_coeffs_8bpp(ctx);

  3185.             } else {

  3186.                 decode_coeffs_16bpp(ctx);

  3187.             }

  3188.         } else {

  3189.             int row7 = s->row7;

  3190. 

  3191. #define SPLAT_ZERO_CTX(v, n) \

  3192.     switch (n) { \

  3193.     case 1:  v = 0;          break; \

  3194.     case 2:  AV_ZERO16(&v);  break; \

  3195.     case 4:  AV_ZERO32(&v);  break; \

  3196.     case 8:  AV_ZERO64(&v);  break; \

  3197.     case 16: AV_ZERO128(&v); break; \

  3198.     }

  3199. #define SPLAT_ZERO_YUV(dir, var, off, n, dir2) \

  3200.     do { \

  3201.         SPLAT_ZERO_CTX(s->dir##_y_##var[off * 2], n * 2); \

  3202.         if (s->ss_##dir2) { \

  3203.             SPLAT_ZERO_CTX(s->dir##_uv_##var[0][off], n); \

  3204.             SPLAT_ZERO_CTX(s->dir##_uv_##var[1][off], n); \

  3205.         } else { \

  3206.             SPLAT_ZERO_CTX(s->dir##_uv_##var[0][off * 2], n * 2); \

  3207.             SPLAT_ZERO_CTX(s->dir##_uv_##var[1][off * 2], n * 2); \

  3208.         } \

  3209.     } while (0)

  3210. 

  3211.             switch (w4) {

  3212.             case 1: SPLAT_ZERO_YUV(above, nnz_ctx, col, 1, h); break;

  3213.             case 2: SPLAT_ZERO_YUV(above, nnz_ctx, col, 2, h); break;

  3214.             case 4: SPLAT_ZERO_YUV(above, nnz_ctx, col, 4, h); break;

  3215.             case 8: SPLAT_ZERO_YUV(above, nnz_ctx, col, 8, h); break;

  3216.             }

  3217.             switch (h4) {

  3218.             case 1: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 1, v); break;

  3219.             case 2: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 2, v); break;

  3220.             case 4: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 4, v); break;

  3221.             case 8: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 8, v); break;

  3222.             }

  3223.         }

  3224.         if (s->pass == 1) {

  3225.             s->b++;

  3226.             s->block += w4 * h4 * 64 * bytesperpixel;

  3227.             s->uvblock[0] += w4 * h4 * 64 * bytesperpixel >> (s->ss_h + s->ss_v);

  3228.             s->uvblock[1] += w4 * h4 * 64 * bytesperpixel >> (s->ss_h + s->ss_v);

  3229.             s->eob += 4 * w4 * h4;

  3230.             s->uveob[0] += 4 * w4 * h4 >> (s->ss_h + s->ss_v);

  3231.             s->uveob[1] += 4 * w4 * h4 >> (s->ss_h + s->ss_v);

  3232. 

  3233.             return;

  3234.         }

  3235.     }

  3236. 

  3237.     // emulated overhangs if the stride of the target buffer can't hold. This

  3238.     // allows to support emu-edge and so on even if we have large block

  3239.     // overhangs

  3240.     emu[0] = (col + w4) * 8 > f->linesize[0] ||

  3241.              (row + h4) > s->rows;

  3242.     emu[1] = (col + w4) * 4 > f->linesize[1] ||

  3243.              (row + h4) > s->rows;

  3244.     if (emu[0]) {

  3245.         s->dst[0] = s->tmp_y;

  3246.         s->y_stride = 128;

  3247.     } else {

  3248.         s->dst[0] = f->data[0] + yoff;

  3249.         s->y_stride = f->linesize[0];

  3250.     }

  3251.     if (emu[1]) {

  3252.         s->dst[1] = s->tmp_uv[0];

  3253.         s->dst[2] = s->tmp_uv[1];

  3254.         s->uv_stride = 128;

  3255.     } else {

  3256.         s->dst[1] = f->data[1] + uvoff;

  3257.         s->dst[2] = f->data[2] + uvoff;

  3258.         s->uv_stride = f->linesize[1];

  3259.     }

  3260.     if (b->intra) {

  3261.         if (s->bpp > 8) {

  3262.             intra_recon_16bpp(ctx, yoff, uvoff);

  3263.         } else {

  3264.             intra_recon_8bpp(ctx, yoff, uvoff);

  3265.         }

  3266.     } else {

  3267.         if (s->bpp > 8) {

  3268.             inter_recon_16bpp(ctx);

  3269.         } else {

  3270.             inter_recon_8bpp(ctx);

  3271.         }

  3272.     }

  3273.     if (emu[0]) {

  3274.         int w = FFMIN(s->cols - col, w4) * 8, h = FFMIN(s->rows - row, h4) * 8, n, o = 0;

  3275. 

  3276.         for (n = 0; o < w; n++) {

  3277.             int bw = 64 >> n;

  3278. 

  3279.             av_assert2(n <= 4);

  3280.             if (w & bw) {

  3281.                 s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o, f->linesize[0],

  3282.                                          s->tmp_y + o, 128, h, 0, 0);

  3283.                 o += bw * bytesperpixel;

  3284.             }

  3285.         }

  3286.     }

  3287.     if (emu[1]) {

  3288.         int w = FFMIN(s->cols - col, w4) * 8 >> s->ss_h;

  3289.         int h = FFMIN(s->rows - row, h4) * 8 >> s->ss_v, n, o = 0;

  3290. 

  3291.         for (n = 1; o < w; n++) {

  3292.             int bw = 64 >> n;

  3293. 

  3294.             av_assert2(n <= 4);

  3295.             if (w & bw) {

  3296.                 s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o, f->linesize[1],

  3297.                                          s->tmp_uv[0] + o, 128, h, 0, 0);

  3298.                 s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o, f->linesize[2],

  3299.                                          s->tmp_uv[1] + o, 128, h, 0, 0);

  3300.                 o += bw * bytesperpixel;

  3301.             }

  3302.         }

  3303.     }

  3304. 

  3305.     // pick filter level and find edges to apply filter to

  3306.     if (s->filter.level &&

  3307.         (lvl = s->segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]

  3308.                                                     [b->mode[3] != ZEROMV]) > 0) {

  3309.         int x_end = FFMIN(s->cols - col, w4), y_end = FFMIN(s->rows - row, h4);

  3310.         int skip_inter = !b->intra && b->skip, col7 = s->col7, row7 = s->row7;

  3311. 

  3312.         setctx_2d(&lflvl->level[row7 * 8 + col7], w4, h4, 8, lvl);

  3313.         mask_edges(lflvl->mask[0], 0, 0, row7, col7, x_end, y_end, 0, 0, b->tx, skip_inter);

  3314.         if (s->ss_h || s->ss_v)

  3315.             mask_edges(lflvl->mask[1], s->ss_h, s->ss_v, row7, col7, x_end, y_end,

  3316.                        s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,

  3317.                        s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,

  3318.                        b->uvtx, skip_inter);

  3319. 

  3320.         if (!s->filter.lim_lut[lvl]) {

  3321.             int sharp = s->filter.sharpness;

  3322.             int limit = lvl;

  3323. 

  3324.             if (sharp > 0) {

  3325.                 limit >>= (sharp + 3) >> 2;

  3326.                 limit = FFMIN(limit, 9 - sharp);

  3327.             }

  3328.             limit = FFMAX(limit, 1);

  3329. 

  3330.             s->filter.lim_lut[lvl] = limit;

  3331.             s->filter.mblim_lut[lvl] = 2 * (lvl + 2) + limit;

  3332.         }

  3333.     }

  3334. 

  3335.     if (s->pass == 2) {

  3336.         s->b++;

  3337.         s->block += w4 * h4 * 64 * bytesperpixel;

  3338.         s->uvblock[0] += w4 * h4 * 64 * bytesperpixel >> (s->ss_v + s->ss_h);

  3339.         s->uvblock[1] += w4 * h4 * 64 * bytesperpixel >> (s->ss_v + s->ss_h);

  3340.         s->eob += 4 * w4 * h4;

  3341.         s->uveob[0] += 4 * w4 * h4 >> (s->ss_v + s->ss_h);

  3342.         s->uveob[1] += 4 * w4 * h4 >> (s->ss_v + s->ss_h);

  3343.     }

  3344. }

  3345. 

  3346. static void decode_sb(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  3347.                       ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  3348. {

  3349.     VP9Context *s = ctx->priv_data;

  3350.     int c = ((s->above_partition_ctx[col] >> (3 - bl)) & 1) |

  3351.             (((s->left_partition_ctx[row & 0x7] >> (3 - bl)) & 1) << 1);

  3352.     const uint8_t *p = s->keyframe || s->intraonly ? vp9_default_kf_partition_probs[bl][c] :

  3353.                                                      s->prob.p.partition[bl][c];

  3354.     enum BlockPartition bp;

  3355.     ptrdiff_t hbs = 4 >> bl;

  3356.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  3357.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  3358.     int bytesperpixel = s->bytesperpixel;

  3359. 

  3360.     if (bl == BL_8X8) {

  3361.         bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  3362.         decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3363.     } else if (col + hbs < s->cols) { // FIXME why not <=?

  3364.         if (row + hbs < s->rows) { // FIXME why not <=?

  3365.             bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  3366.             switch (bp) {

  3367.             case PARTITION_NONE:

  3368.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3369.                 break;

  3370.             case PARTITION_H:

  3371.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3372.                 yoff  += hbs * 8 * y_stride;

  3373.                 uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3374.                 decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, bl, bp);

  3375.                 break;

  3376.             case PARTITION_V:

  3377.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3378.                 yoff  += hbs * 8 * bytesperpixel;

  3379.                 uvoff += hbs * 8 * bytesperpixel >> s->ss_h;

  3380.                 decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, bl, bp);

  3381.                 break;

  3382.             case PARTITION_SPLIT:

  3383.                 decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3384.                 decode_sb(ctx, row, col + hbs, lflvl,

  3385.                           yoff + 8 * hbs * bytesperpixel,

  3386.                           uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);

  3387.                 yoff  += hbs * 8 * y_stride;

  3388.                 uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3389.                 decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3390.                 decode_sb(ctx, row + hbs, col + hbs, lflvl,

  3391.                           yoff + 8 * hbs * bytesperpixel,

  3392.                           uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);

  3393.                 break;

  3394.             default:

  3395.                 av_assert0(0);

  3396.             }

  3397.         } else if (vp56_rac_get_prob_branchy(&s->c, p[1])) {

  3398.             bp = PARTITION_SPLIT;

  3399.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3400.             decode_sb(ctx, row, col + hbs, lflvl,

  3401.                       yoff + 8 * hbs * bytesperpixel,

  3402.                       uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);

  3403.         } else {

  3404.             bp = PARTITION_H;

  3405.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3406.         }

  3407.     } else if (row + hbs < s->rows) { // FIXME why not <=?

  3408.         if (vp56_rac_get_prob_branchy(&s->c, p[2])) {

  3409.             bp = PARTITION_SPLIT;

  3410.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3411.             yoff  += hbs * 8 * y_stride;

  3412.             uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3413.             decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3414.         } else {

  3415.             bp = PARTITION_V;

  3416.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3417.         }

  3418.     } else {

  3419.         bp = PARTITION_SPLIT;

  3420.         decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3421.     }

  3422.     s->counts.partition[bl][c][bp]++;

  3423. }

  3424. 

  3425. static void decode_sb_mem(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  3426.                           ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  3427. {

  3428.     VP9Context *s = ctx->priv_data;

  3429.     VP9Block *b = s->b;

  3430.     ptrdiff_t hbs = 4 >> bl;

  3431.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  3432.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  3433.     int bytesperpixel = s->bytesperpixel;

  3434. 

  3435.     if (bl == BL_8X8) {

  3436.         av_assert2(b->bl == BL_8X8);

  3437.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  3438.     } else if (s->b->bl == bl) {

  3439.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  3440.         if (b->bp == PARTITION_H && row + hbs < s->rows) {

  3441.             yoff  += hbs * 8 * y_stride;

  3442.             uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3443.             decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, b->bl, b->bp);

  3444.         } else if (b->bp == PARTITION_V && col + hbs < s->cols) {

  3445.             yoff  += hbs * 8 * bytesperpixel;

  3446.             uvoff += hbs * 8 * bytesperpixel >> s->ss_h;

  3447.             decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, b->bl, b->bp);

  3448.         }

  3449.     } else {

  3450.         decode_sb_mem(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3451.         if (col + hbs < s->cols) { // FIXME why not <=?

  3452.             if (row + hbs < s->rows) {

  3453.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff + 8 * hbs * bytesperpixel,

  3454.                               uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);

  3455.                 yoff  += hbs * 8 * y_stride;

  3456.                 uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3457.                 decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3458.                 decode_sb_mem(ctx, row + hbs, col + hbs, lflvl,

  3459.                               yoff + 8 * hbs * bytesperpixel,

  3460.                               uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);

  3461.             } else {

  3462.                 yoff  += hbs * 8 * bytesperpixel;

  3463.                 uvoff += hbs * 8 * bytesperpixel >> s->ss_h;

  3464.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff, uvoff, bl + 1);

  3465.             }

  3466.         } else if (row + hbs < s->rows) {

  3467.             yoff  += hbs * 8 * y_stride;

  3468.             uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3469.             decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3470.         }

  3471.     }

  3472. }

  3473. 

  3474. static av_always_inline void filter_plane_cols(VP9Context *s, int col, int ss_h, int ss_v,

  3475.                                                uint8_t *lvl, uint8_t (*mask)[4],

  3476.                                                uint8_t *dst, ptrdiff_t ls)

  3477. {

  3478.     int y, x, bytesperpixel = s->bytesperpixel;

  3479. 

  3480.     // filter edges between columns (e.g. block1 | block2)

  3481.     for (y = 0; y < 8; y += 2 << ss_v, dst += 16 * ls, lvl += 16 << ss_v) {

  3482.         uint8_t *ptr = dst, *l = lvl, *hmask1 = mask[y], *hmask2 = mask[y + 1 + ss_v];

  3483.         unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2], hm13 = hmask1[3];

  3484.         unsigned hm2 = hmask2[1] | hmask2[2], hm23 = hmask2[3];

  3485.         unsigned hm = hm1 | hm2 | hm13 | hm23;

  3486. 

  3487.         for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 8 * bytesperpixel >> ss_h) {

  3488.             if (col || x > 1) {

  3489.                 if (hm1 & x) {

  3490.                     int L = *l, H = L >> 4;

  3491.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3492. 

  3493.                     if (hmask1[0] & x) {

  3494.                         if (hmask2[0] & x) {

  3495.                             av_assert2(l[8 << ss_v] == L);

  3496.                             s->dsp.loop_filter_16[0](ptr, ls, E, I, H);

  3497.                         } else {

  3498.                             s->dsp.loop_filter_8[2][0](ptr, ls, E, I, H);

  3499.                         }

  3500.                     } else if (hm2 & x) {

  3501.                         L = l[8 << ss_v];

  3502.                         H |= (L >> 4) << 8;

  3503.                         E |= s->filter.mblim_lut[L] << 8;

  3504.                         I |= s->filter.lim_lut[L] << 8;

  3505.                         s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3506.                                                [!!(hmask2[1] & x)]

  3507.                                                [0](ptr, ls, E, I, H);

  3508.                     } else {

  3509.                         s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3510.                                             [0](ptr, ls, E, I, H);

  3511.                     }

  3512.                 } else if (hm2 & x) {

  3513.                     int L = l[8 << ss_v], H = L >> 4;

  3514.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3515. 

  3516.                     s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3517.                                         [0](ptr + 8 * ls, ls, E, I, H);

  3518.                 }

  3519.             }

  3520.             if (ss_h) {

  3521.                 if (x & 0xAA)

  3522.                     l += 2;

  3523.             } else {

  3524.                 if (hm13 & x) {

  3525.                     int L = *l, H = L >> 4;

  3526.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3527. 

  3528.                     if (hm23 & x) {

  3529.                         L = l[8 << ss_v];

  3530.                         H |= (L >> 4) << 8;

  3531.                         E |= s->filter.mblim_lut[L] << 8;

  3532.                         I |= s->filter.lim_lut[L] << 8;

  3533.                         s->dsp.loop_filter_mix2[0][0][0](ptr + 4 * bytesperpixel, ls, E, I, H);

  3534.                     } else {

  3535.                         s->dsp.loop_filter_8[0][0](ptr + 4 * bytesperpixel, ls, E, I, H);

  3536.                     }

  3537.                 } else if (hm23 & x) {

  3538.                     int L = l[8 << ss_v], H = L >> 4;

  3539.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3540. 

  3541.                     s->dsp.loop_filter_8[0][0](ptr + 8 * ls + 4 * bytesperpixel, ls, E, I, H);

  3542.                 }

  3543.                 l++;

  3544.             }

  3545.         }

  3546.     }

  3547. }

  3548. 

  3549. static av_always_inline void filter_plane_rows(VP9Context *s, int row, int ss_h, int ss_v,

  3550.                                                uint8_t *lvl, uint8_t (*mask)[4],

  3551.                                                uint8_t *dst, ptrdiff_t ls)

  3552. {

  3553.     int y, x, bytesperpixel = s->bytesperpixel;

  3554. 

  3555.     //                                 block1

  3556.     // filter edges between rows (e.g. ------)

  3557.     //                                 block2

  3558.     for (y = 0; y < 8; y++, dst += 8 * ls >> ss_v) {

  3559.         uint8_t *ptr = dst, *l = lvl, *vmask = mask[y];

  3560.         unsigned vm = vmask[0] | vmask[1] | vmask[2], vm3 = vmask[3];

  3561. 

  3562.         for (x = 1; vm & ~(x - 1); x <<= (2 << ss_h), ptr += 16 * bytesperpixel, l += 2 << ss_h) {

  3563.             if (row || y) {

  3564.                 if (vm & x) {

  3565.                     int L = *l, H = L >> 4;

  3566.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3567. 

  3568.                     if (vmask[0] & x) {

  3569.                         if (vmask[0] & (x << (1 + ss_h))) {

  3570.                             av_assert2(l[1 + ss_h] == L);

  3571.                             s->dsp.loop_filter_16[1](ptr, ls, E, I, H);

  3572.                         } else {

  3573.                             s->dsp.loop_filter_8[2][1](ptr, ls, E, I, H);

  3574.                         }

  3575.                     } else if (vm & (x << (1 + ss_h))) {

  3576.                         L = l[1 + ss_h];

  3577.                         H |= (L >> 4) << 8;

  3578.                         E |= s->filter.mblim_lut[L] << 8;

  3579.                         I |= s->filter.lim_lut[L] << 8;

  3580.                         s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3581.                                                [!!(vmask[1] & (x << (1 + ss_h)))]

  3582.                                                [1](ptr, ls, E, I, H);

  3583.                     } else {

  3584.                         s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3585.                                             [1](ptr, ls, E, I, H);

  3586.                     }

  3587.                 } else if (vm & (x << (1 + ss_h))) {

  3588.                     int L = l[1 + ss_h], H = L >> 4;

  3589.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3590. 

  3591.                     s->dsp.loop_filter_8[!!(vmask[1] & (x << (1 + ss_h)))]

  3592.                                         [1](ptr + 8 * bytesperpixel, ls, E, I, H);

  3593.                 }

  3594.             }

  3595.             if (!ss_v) {

  3596.                 if (vm3 & x) {

  3597.                     int L = *l, H = L >> 4;

  3598.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3599. 

  3600.                     if (vm3 & (x << (1 + ss_h))) {

  3601.                         L = l[1 + ss_h];

  3602.                         H |= (L >> 4) << 8;

  3603.                         E |= s->filter.mblim_lut[L] << 8;

  3604.                         I |= s->filter.lim_lut[L] << 8;

  3605.                         s->dsp.loop_filter_mix2[0][0][1](ptr + ls * 4, ls, E, I, H);

  3606.                     } else {

  3607.                         s->dsp.loop_filter_8[0][1](ptr + ls * 4, ls, E, I, H);

  3608.                     }

  3609.                 } else if (vm3 & (x << (1 + ss_h))) {

  3610.                     int L = l[1 + ss_h], H = L >> 4;

  3611.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3612. 

  3613.                     s->dsp.loop_filter_8[0][1](ptr + ls * 4 + 8 * bytesperpixel, ls, E, I, H);

  3614.                 }

  3615.             }

  3616.         }

  3617.         if (ss_v) {

  3618.             if (y & 1)

  3619.                 lvl += 16;

  3620.         } else {

  3621.             lvl += 8;

  3622.         }

  3623.     }

  3624. }

  3625. 

  3626. static void loopfilter_sb(AVCodecContext *ctx, struct VP9Filter *lflvl,

  3627.                           int row, int col, ptrdiff_t yoff, ptrdiff_t uvoff)

  3628. {

  3629.     VP9Context *s = ctx->priv_data;

  3630.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  3631.     uint8_t *dst = f->data[0] + yoff;

  3632.     ptrdiff_t ls_y = f->linesize[0], ls_uv = f->linesize[1];

  3633.     uint8_t (*uv_masks)[8][4] = lflvl->mask[s->ss_h | s->ss_v];

  3634.     int p;

  3635. 

  3636.     // FIXME in how far can we interleave the v/h loopfilter calls? E.g.

  3637.     // if you think of them as acting on a 8x8 block max, we can interleave

  3638.     // each v/h within the single x loop, but that only works if we work on

  3639.     // 8 pixel blocks, and we won't always do that (we want at least 16px

  3640.     // to use SSE2 optimizations, perhaps 32 for AVX2)

  3641. 

  3642.     filter_plane_cols(s, col, 0, 0, lflvl->level, lflvl->mask[0][0], dst, ls_y);

  3643.     filter_plane_rows(s, row, 0, 0, lflvl->level, lflvl->mask[0][1], dst, ls_y);

  3644. 

  3645.     for (p = 0; p < 2; p++) {

  3646.         dst = f->data[1 + p] + uvoff;

  3647.         filter_plane_cols(s, col, s->ss_h, s->ss_v, lflvl->level, uv_masks[0], dst, ls_uv);

  3648.         filter_plane_rows(s, row, s->ss_h, s->ss_v, lflvl->level, uv_masks[1], dst, ls_uv);

  3649.     }

  3650. }

  3651. 

  3652. static void set_tile_offset(int *start, int *end, int idx, int log2_n, int n)

  3653. {

  3654.     int sb_start = ( idx      * n) >> log2_n;

  3655.     int sb_end   = ((idx + 1) * n) >> log2_n;

  3656.     *start = FFMIN(sb_start, n) << 3;

  3657.     *end   = FFMIN(sb_end,   n) << 3;

  3658. }

  3659. 

  3660. static av_always_inline void adapt_prob(uint8_t *p, unsigned ct0, unsigned ct1,

  3661.                                         int max_count, int update_factor)

  3662. {

  3663.     unsigned ct = ct0 + ct1, p2, p1;

  3664. 

  3665.     if (!ct)

  3666.         return;

  3667. 

  3668.     p1 = *p;

  3669.     p2 = ((ct0 << 8) + (ct >> 1)) / ct;

  3670.     p2 = av_clip(p2, 1, 255);

  3671.     ct = FFMIN(ct, max_count);

  3672.     update_factor = FASTDIV(update_factor * ct, max_count);

  3673. 

  3674.     // (p1 * (256 - update_factor) + p2 * update_factor + 128) >> 8

  3675.     *p = p1 + (((p2 - p1) * update_factor + 128) >> 8);

  3676. }

  3677. 

  3678. static void adapt_probs(VP9Context *s)

  3679. {

  3680.     int i, j, k, l, m;

  3681.     prob_context *p = &s->prob_ctx[s->framectxid].p;

  3682.     int uf = (s->keyframe || s->intraonly || !s->last_keyframe) ? 112 : 128;

  3683. 

  3684.     // coefficients

  3685.     for (i = 0; i < 4; i++)

  3686.         for (j = 0; j < 2; j++)

  3687.             for (k = 0; k < 2; k++)

  3688.                 for (l = 0; l < 6; l++)

  3689.                     for (m = 0; m < 6; m++) {

  3690.                         uint8_t *pp = s->prob_ctx[s->framectxid].coef[i][j][k][l][m];

  3691.                         unsigned *e = s->counts.eob[i][j][k][l][m];

  3692.                         unsigned *c = s->counts.coef[i][j][k][l][m];

  3693. 

  3694.                         if (l == 0 && m >= 3) // dc only has 3 pt

  3695.                             break;

  3696. 

  3697.                         adapt_prob(&pp[0], e[0], e[1], 24, uf);

  3698.                         adapt_prob(&pp[1], c[0], c[1] + c[2], 24, uf);

  3699.                         adapt_prob(&pp[2], c[1], c[2], 24, uf);

  3700.                     }

  3701. 

  3702.     if (s->keyframe || s->intraonly) {

  3703.         memcpy(p->skip,  s->prob.p.skip,  sizeof(p->skip));

  3704.         memcpy(p->tx32p, s->prob.p.tx32p, sizeof(p->tx32p));

  3705.         memcpy(p->tx16p, s->prob.p.tx16p, sizeof(p->tx16p));

  3706.         memcpy(p->tx8p,  s->prob.p.tx8p,  sizeof(p->tx8p));

  3707.         return;

  3708.     }

  3709. 

  3710.     // skip flag

  3711.     for (i = 0; i < 3; i++)

  3712.         adapt_prob(&p->skip[i], s->counts.skip[i][0], s->counts.skip[i][1], 20, 128);

  3713. 

  3714.     // intra/inter flag

  3715.     for (i = 0; i < 4; i++)

  3716.         adapt_prob(&p->intra[i], s->counts.intra[i][0], s->counts.intra[i][1], 20, 128);

  3717. 

  3718.     // comppred flag

  3719.     if (s->comppredmode == PRED_SWITCHABLE) {

  3720.       for (i = 0; i < 5; i++)

  3721.           adapt_prob(&p->comp[i], s->counts.comp[i][0], s->counts.comp[i][1], 20, 128);

  3722.     }

  3723. 

  3724.     // reference frames

  3725.     if (s->comppredmode != PRED_SINGLEREF) {

  3726.       for (i = 0; i < 5; i++)

  3727.           adapt_prob(&p->comp_ref[i], s->counts.comp_ref[i][0],

  3728.                      s->counts.comp_ref[i][1], 20, 128);

  3729.     }

  3730. 

  3731.     if (s->comppredmode != PRED_COMPREF) {

  3732.       for (i = 0; i < 5; i++) {

  3733.           uint8_t *pp = p->single_ref[i];

  3734.           unsigned (*c)[2] = s->counts.single_ref[i];

  3735. 

  3736.           adapt_prob(&pp[0], c[0][0], c[0][1], 20, 128);

  3737.           adapt_prob(&pp[1], c[1][0], c[1][1], 20, 128);

  3738.       }

  3739.     }

  3740. 

  3741.     // block partitioning

  3742.     for (i = 0; i < 4; i++)

  3743.         for (j = 0; j < 4; j++) {

  3744.             uint8_t *pp = p->partition[i][j];

  3745.             unsigned *c = s->counts.partition[i][j];

  3746. 

  3747.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3748.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3749.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3750.         }

  3751. 

  3752.     // tx size

  3753.     if (s->txfmmode == TX_SWITCHABLE) {

  3754.       for (i = 0; i < 2; i++) {

  3755.           unsigned *c16 = s->counts.tx16p[i], *c32 = s->counts.tx32p[i];

  3756. 

  3757.           adapt_prob(&p->tx8p[i], s->counts.tx8p[i][0], s->counts.tx8p[i][1], 20, 128);

  3758.           adapt_prob(&p->tx16p[i][0], c16[0], c16[1] + c16[2], 20, 128);

  3759.           adapt_prob(&p->tx16p[i][1], c16[1], c16[2], 20, 128);

  3760.           adapt_prob(&p->tx32p[i][0], c32[0], c32[1] + c32[2] + c32[3], 20, 128);

  3761.           adapt_prob(&p->tx32p[i][1], c32[1], c32[2] + c32[3], 20, 128);

  3762.           adapt_prob(&p->tx32p[i][2], c32[2], c32[3], 20, 128);

  3763.       }

  3764.     }

  3765. 

  3766.     // interpolation filter

  3767.     if (s->filtermode == FILTER_SWITCHABLE) {

  3768.         for (i = 0; i < 4; i++) {

  3769.             uint8_t *pp = p->filter[i];

  3770.             unsigned *c = s->counts.filter[i];

  3771. 

  3772.             adapt_prob(&pp[0], c[0], c[1] + c[2], 20, 128);

  3773.             adapt_prob(&pp[1], c[1], c[2], 20, 128);

  3774.         }

  3775.     }

  3776. 

  3777.     // inter modes

  3778.     for (i = 0; i < 7; i++) {

  3779.         uint8_t *pp = p->mv_mode[i];

  3780.         unsigned *c = s->counts.mv_mode[i];

  3781. 

  3782.         adapt_prob(&pp[0], c[2], c[1] + c[0] + c[3], 20, 128);

  3783.         adapt_prob(&pp[1], c[0], c[1] + c[3], 20, 128);

  3784.         adapt_prob(&pp[2], c[1], c[3], 20, 128);

  3785.     }

  3786. 

  3787.     // mv joints

  3788.     {

  3789.         uint8_t *pp = p->mv_joint;

  3790.         unsigned *c = s->counts.mv_joint;

  3791. 

  3792.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3793.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3794.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3795.     }

  3796. 

  3797.     // mv components

  3798.     for (i = 0; i < 2; i++) {

  3799.         uint8_t *pp;

  3800.         unsigned *c, (*c2)[2], sum;

  3801. 

  3802.         adapt_prob(&p->mv_comp[i].sign, s->counts.mv_comp[i].sign[0],

  3803.                    s->counts.mv_comp[i].sign[1], 20, 128);

  3804. 

  3805.         pp = p->mv_comp[i].classes;

  3806.         c = s->counts.mv_comp[i].classes;

  3807.         sum = c[1] + c[2] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9] + c[10];

  3808.         adapt_prob(&pp[0], c[0], sum, 20, 128);

  3809.         sum -= c[1];

  3810.         adapt_prob(&pp[1], c[1], sum, 20, 128);

  3811.         sum -= c[2] + c[3];

  3812.         adapt_prob(&pp[2], c[2] + c[3], sum, 20, 128);

  3813.         adapt_prob(&pp[3], c[2], c[3], 20, 128);

  3814.         sum -= c[4] + c[5];

  3815.         adapt_prob(&pp[4], c[4] + c[5], sum, 20, 128);

  3816.         adapt_prob(&pp[5], c[4], c[5], 20, 128);

  3817.         sum -= c[6];

  3818.         adapt_prob(&pp[6], c[6], sum, 20, 128);

  3819.         adapt_prob(&pp[7], c[7] + c[8], c[9] + c[10], 20, 128);

  3820.         adapt_prob(&pp[8], c[7], c[8], 20, 128);

  3821.         adapt_prob(&pp[9], c[9], c[10], 20, 128);

  3822. 

  3823.         adapt_prob(&p->mv_comp[i].class0, s->counts.mv_comp[i].class0[0],

  3824.                    s->counts.mv_comp[i].class0[1], 20, 128);

  3825.         pp = p->mv_comp[i].bits;

  3826.         c2 = s->counts.mv_comp[i].bits;

  3827.         for (j = 0; j < 10; j++)

  3828.             adapt_prob(&pp[j], c2[j][0], c2[j][1], 20, 128);

  3829. 

  3830.         for (j = 0; j < 2; j++) {

  3831.             pp = p->mv_comp[i].class0_fp[j];

  3832.             c = s->counts.mv_comp[i].class0_fp[j];

  3833.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3834.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3835.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3836.         }

  3837.         pp = p->mv_comp[i].fp;

  3838.         c = s->counts.mv_comp[i].fp;

  3839.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3840.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3841.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3842. 

  3843.         if (s->highprecisionmvs) {

  3844.             adapt_prob(&p->mv_comp[i].class0_hp, s->counts.mv_comp[i].class0_hp[0],

  3845.                        s->counts.mv_comp[i].class0_hp[1], 20, 128);

  3846.             adapt_prob(&p->mv_comp[i].hp, s->counts.mv_comp[i].hp[0],

  3847.                        s->counts.mv_comp[i].hp[1], 20, 128);

  3848.         }

  3849.     }

  3850. 

  3851.     // y intra modes

  3852.     for (i = 0; i < 4; i++) {

  3853.         uint8_t *pp = p->y_mode[i];

  3854.         unsigned *c = s->counts.y_mode[i], sum, s2;

  3855. 

  3856.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3857.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3858.         sum -= c[TM_VP8_PRED];

  3859.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3860.         sum -= c[VERT_PRED];

  3861.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3862.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3863.         sum -= s2;

  3864.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3865.         s2 -= c[HOR_PRED];

  3866.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3867.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3868.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3869.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3870.         sum -= c[VERT_LEFT_PRED];

  3871.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3872.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3873.     }

  3874. 

  3875.     // uv intra modes

  3876.     for (i = 0; i < 10; i++) {

  3877.         uint8_t *pp = p->uv_mode[i];

  3878.         unsigned *c = s->counts.uv_mode[i], sum, s2;

  3879. 

  3880.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3881.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3882.         sum -= c[TM_VP8_PRED];

  3883.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3884.         sum -= c[VERT_PRED];

  3885.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3886.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3887.         sum -= s2;

  3888.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3889.         s2 -= c[HOR_PRED];

  3890.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3891.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3892.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3893.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3894.         sum -= c[VERT_LEFT_PRED];

  3895.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3896.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3897.     }

  3898. }

  3899. 

  3900. static void free_buffers(VP9Context *s)

  3901. {

  3902.     av_freep(&s->intra_pred_data[0]);

  3903.     av_freep(&s->b_base);

  3904.     av_freep(&s->block_base);

  3905. }

  3906. 

  3907. static av_cold int vp9_decode_free(AVCodecContext *ctx)

  3908. {

  3909.     VP9Context *s = ctx->priv_data;

  3910.     int i;

  3911. 

  3912.     for (i = 0; i < 3; i++) {

  3913.         if (s->frames[i].tf.f->data[0])

  3914.             vp9_unref_frame(ctx, &s->frames[i]);

  3915.         av_frame_free(&s->frames[i].tf.f);

  3916.     }

  3917.     for (i = 0; i < 8; i++) {

  3918.         if (s->refs[i].f->data[0])

  3919.             ff_thread_release_buffer(ctx, &s->refs[i]);

  3920.         av_frame_free(&s->refs[i].f);

  3921.         if (s->next_refs[i].f->data[0])

  3922.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3923.         av_frame_free(&s->next_refs[i].f);

  3924.     }

  3925.     free_buffers(s);

  3926.     av_freep(&s->c_b);

  3927.     s->c_b_size = 0;

  3928. 

  3929.     return 0;

  3930. }

  3931. 

  3932. 

  3933. static int vp9_decode_frame(AVCodecContext *ctx, void *frame,

  3934.                             int *got_frame, AVPacket *pkt)

  3935. {

  3936.     const uint8_t *data = pkt->data;

  3937.     int size = pkt->size;

  3938.     VP9Context *s = ctx->priv_data;

  3939.     int res, tile_row, tile_col, i, ref, row, col;

  3940.     int retain_segmap_ref = s->segmentation.enabled && !s->segmentation.update_map;

  3941.     ptrdiff_t yoff, uvoff, ls_y, ls_uv;

  3942.     AVFrame *f;

  3943.     int bytesperpixel;

  3944. 

  3945.     if ((res = decode_frame_header(ctx, data, size, &ref)) < 0) {

  3946.         return res;

  3947.     } else if (res == 0) {

  3948.         if (!s->refs[ref].f->data[0]) {

  3949.             av_log(ctx, AV_LOG_ERROR, "Requested reference %d not available\n", ref);

  3950.             return AVERROR_INVALIDDATA;

  3951.         }

  3952.         if ((res = av_frame_ref(frame, s->refs[ref].f)) < 0)

  3953.             return res;

  3954.         ((AVFrame *)frame)->pkt_pts = pkt->pts;

  3955.         ((AVFrame *)frame)->pkt_dts = pkt->dts;

  3956.         for (i = 0; i < 8; i++) {

  3957.             if (s->next_refs[i].f->data[0])

  3958.                 ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3959.             if (s->refs[i].f->data[0] &&

  3960.                 (res = ff_thread_ref_frame(&s->next_refs[i], &s->refs[i])) < 0)

  3961.                 return res;

  3962.         }

  3963.         *got_frame = 1;

  3964.         return pkt->size;

  3965.     }

  3966.     data += res;

  3967.     size -= res;

  3968. 

  3969.     if (!retain_segmap_ref) {

  3970.         if (s->frames[REF_FRAME_SEGMAP].tf.f->data[0])

  3971.             vp9_unref_frame(ctx, &s->frames[REF_FRAME_SEGMAP]);

  3972.         if (!s->keyframe && !s->intraonly && !s->errorres && s->frames[CUR_FRAME].tf.f->data[0] &&

  3973.             (res = vp9_ref_frame(ctx, &s->frames[REF_FRAME_SEGMAP], &s->frames[CUR_FRAME])) < 0)

  3974.             return res;

  3975.     }

  3976.     if (s->frames[REF_FRAME_MVPAIR].tf.f->data[0])

  3977.         vp9_unref_frame(ctx, &s->frames[REF_FRAME_MVPAIR]);

  3978.     if (!s->intraonly && !s->keyframe && !s->errorres && s->frames[CUR_FRAME].tf.f->data[0] &&

  3979.         (res = vp9_ref_frame(ctx, &s->frames[REF_FRAME_MVPAIR], &s->frames[CUR_FRAME])) < 0)

  3980.         return res;

  3981.     if (s->frames[CUR_FRAME].tf.f->data[0])

  3982.         vp9_unref_frame(ctx, &s->frames[CUR_FRAME]);

  3983.     if ((res = vp9_alloc_frame(ctx, &s->frames[CUR_FRAME])) < 0)

  3984.         return res;

  3985.     f = s->frames[CUR_FRAME].tf.f;

  3986.     f->key_frame = s->keyframe;

  3987.     f->pict_type = (s->keyframe || s->intraonly) ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;

  3988.     ls_y = f->linesize[0];

  3989.     ls_uv =f->linesize[1];

  3990. 

  3991.     // ref frame setup

  3992.     for (i = 0; i < 8; i++) {

  3993.         if (s->next_refs[i].f->data[0])

  3994.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3995.         if (s->refreshrefmask & (1 << i)) {

  3996.             res = ff_thread_ref_frame(&s->next_refs[i], &s->frames[CUR_FRAME].tf);

  3997.         } else if (s->refs[i].f->data[0]) {

  3998.             res = ff_thread_ref_frame(&s->next_refs[i], &s->refs[i]);

  3999.         }

  4000.         if (res < 0)

  4001.             return res;

  4002.     }

  4003. 

  4004.     // main tile decode loop

  4005.     bytesperpixel = s->bytesperpixel;

  4006.     memset(s->above_partition_ctx, 0, s->cols);

  4007.     memset(s->above_skip_ctx, 0, s->cols);

  4008.     if (s->keyframe || s->intraonly) {

  4009.         memset(s->above_mode_ctx, DC_PRED, s->cols * 2);

  4010.     } else {

  4011.         memset(s->above_mode_ctx, NEARESTMV, s->cols);

  4012.     }

  4013.     memset(s->above_y_nnz_ctx, 0, s->sb_cols * 16);

  4014.     memset(s->above_uv_nnz_ctx[0], 0, s->sb_cols * 16 >> s->ss_h);

  4015.     memset(s->above_uv_nnz_ctx[1], 0, s->sb_cols * 16 >> s->ss_h);

  4016.     memset(s->above_segpred_ctx, 0, s->cols);

  4017.     s->pass = s->frames[CUR_FRAME].uses_2pass =

  4018.         ctx->active_thread_type == FF_THREAD_FRAME && s->refreshctx && !s->parallelmode;

  4019.     if ((res = update_block_buffers(ctx)) < 0) {

  4020.         av_log(ctx, AV_LOG_ERROR,

  4021.                "Failed to allocate block buffers\n");

  4022.         return res;

  4023.     }

  4024.     if (s->refreshctx && s->parallelmode) {

  4025.         int j, k, l, m;

  4026. 

  4027.         for (i = 0; i < 4; i++) {

  4028.             for (j = 0; j < 2; j++)

  4029.                 for (k = 0; k < 2; k++)

  4030.                     for (l = 0; l < 6; l++)

  4031.                         for (m = 0; m < 6; m++)

  4032.                             memcpy(s->prob_ctx[s->framectxid].coef[i][j][k][l][m],

  4033.                                    s->prob.coef[i][j][k][l][m], 3);

  4034.             if (s->txfmmode == i)

  4035.                 break;

  4036.         }

  4037.         s->prob_ctx[s->framectxid].p = s->prob.p;

  4038.         ff_thread_finish_setup(ctx);

  4039.     } else if (!s->refreshctx) {

  4040.         ff_thread_finish_setup(ctx);

  4041.     }

  4042. 

  4043.     do {

  4044.         yoff = uvoff = 0;

  4045.         s->b = s->b_base;

  4046.         s->block = s->block_base;

  4047.         s->uvblock[0] = s->uvblock_base[0];

  4048.         s->uvblock[1] = s->uvblock_base[1];

  4049.         s->eob = s->eob_base;

  4050.         s->uveob[0] = s->uveob_base[0];

  4051.         s->uveob[1] = s->uveob_base[1];

  4052. 

  4053.         for (tile_row = 0; tile_row < s->tiling.tile_rows; tile_row++) {

  4054.             set_tile_offset(&s->tiling.tile_row_start, &s->tiling.tile_row_end,

  4055.                             tile_row, s->tiling.log2_tile_rows, s->sb_rows);

  4056.             if (s->pass != 2) {

  4057.                 for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  4058.                     unsigned tile_size;

  4059. 

  4060.                     if (tile_col == s->tiling.tile_cols - 1 &&

  4061.                         tile_row == s->tiling.tile_rows - 1) {

  4062.                         tile_size = size;

  4063.                     } else {

  4064.                         tile_size = AV_RB32(data);

  4065.                         data += 4;

  4066.                         size -= 4;

  4067.                     }

  4068.                     if (tile_size > size) {

  4069.                         ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  4070.                         return AVERROR_INVALIDDATA;

  4071.                     }

  4072.                     ff_vp56_init_range_decoder(&s->c_b[tile_col], data, tile_size);

  4073.                     if (vp56_rac_get_prob_branchy(&s->c_b[tile_col], 128)) { // marker bit

  4074.                         ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  4075.                         return AVERROR_INVALIDDATA;

  4076.                     }

  4077.                     data += tile_size;

  4078.                     size -= tile_size;

  4079.                 }

  4080.             }

  4081. 

  4082.             for (row = s->tiling.tile_row_start; row < s->tiling.tile_row_end;

  4083.                  row += 8, yoff += ls_y * 64, uvoff += ls_uv * 64 >> s->ss_v) {

  4084.                 struct VP9Filter *lflvl_ptr = s->lflvl;

  4085.                 ptrdiff_t yoff2 = yoff, uvoff2 = uvoff;

  4086. 

  4087.                 for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  4088.                     set_tile_offset(&s->tiling.tile_col_start, &s->tiling.tile_col_end,

  4089.                                     tile_col, s->tiling.log2_tile_cols, s->sb_cols);

  4090. 

  4091.                     if (s->pass != 2) {

  4092.                         memset(s->left_partition_ctx, 0, 8);

  4093.                         memset(s->left_skip_ctx, 0, 8);

  4094.                         if (s->keyframe || s->intraonly) {

  4095.                             memset(s->left_mode_ctx, DC_PRED, 16);

  4096.                         } else {

  4097.                             memset(s->left_mode_ctx, NEARESTMV, 8);

  4098.                         }

  4099.                         memset(s->left_y_nnz_ctx, 0, 16);

  4100.                         memset(s->left_uv_nnz_ctx, 0, 32);

  4101.                         memset(s->left_segpred_ctx, 0, 8);

  4102. 

  4103.                         memcpy(&s->c, &s->c_b[tile_col], sizeof(s->c));

  4104.                     }

  4105. 

  4106.                     for (col = s->tiling.tile_col_start;

  4107.                          col < s->tiling.tile_col_end;

  4108.                          col += 8, yoff2 += 64 * bytesperpixel,

  4109.                          uvoff2 += 64 * bytesperpixel >> s->ss_h, lflvl_ptr++) {

  4110.                         // FIXME integrate with lf code (i.e. zero after each

  4111.                         // use, similar to invtxfm coefficients, or similar)

  4112.                         if (s->pass != 1) {

  4113.                             memset(lflvl_ptr->mask, 0, sizeof(lflvl_ptr->mask));

  4114.                         }

  4115. 

  4116.                         if (s->pass == 2) {

  4117.                             decode_sb_mem(ctx, row, col, lflvl_ptr,

  4118.                                           yoff2, uvoff2, BL_64X64);

  4119.                         } else {

  4120.                             decode_sb(ctx, row, col, lflvl_ptr,

  4121.                                       yoff2, uvoff2, BL_64X64);

  4122.                         }

  4123.                     }

  4124.                     if (s->pass != 2) {

  4125.                         memcpy(&s->c_b[tile_col], &s->c, sizeof(s->c));

  4126.                     }

  4127.                 }

  4128. 

  4129.                 if (s->pass == 1) {

  4130.                     continue;

  4131.                 }

  4132. 

  4133.                 // backup pre-loopfilter reconstruction data for intra

  4134.                 // prediction of next row of sb64s

  4135.                 if (row + 8 < s->rows) {

  4136.                     memcpy(s->intra_pred_data[0],

  4137.                            f->data[0] + yoff + 63 * ls_y,

  4138.                            8 * s->cols * bytesperpixel);

  4139.                     memcpy(s->intra_pred_data[1],

  4140.                            f->data[1] + uvoff + ((64 >> s->ss_v) - 1) * ls_uv,

  4141.                            8 * s->cols * bytesperpixel >> s->ss_h);

  4142.                     memcpy(s->intra_pred_data[2],

  4143.                            f->data[2] + uvoff + ((64 >> s->ss_v) - 1) * ls_uv,

  4144.                            8 * s->cols * bytesperpixel >> s->ss_h);

  4145.                 }

  4146. 

  4147.                 // loopfilter one row

  4148.                 if (s->filter.level) {

  4149.                     yoff2 = yoff;

  4150.                     uvoff2 = uvoff;

  4151.                     lflvl_ptr = s->lflvl;

  4152.                     for (col = 0; col < s->cols;

  4153.                          col += 8, yoff2 += 64 * bytesperpixel,

  4154.                          uvoff2 += 64 * bytesperpixel >> s->ss_h, lflvl_ptr++) {

  4155.                         loopfilter_sb(ctx, lflvl_ptr, row, col, yoff2, uvoff2);

  4156.                     }

  4157.                 }

  4158. 

  4159.                 // FIXME maybe we can make this more finegrained by running the

  4160.                 // loopfilter per-block instead of after each sbrow

  4161.                 // In fact that would also make intra pred left preparation easier?

  4162.                 ff_thread_report_progress(&s->frames[CUR_FRAME].tf, row >> 3, 0);

  4163.             }

  4164.         }

  4165. 

  4166.         if (s->pass < 2 && s->refreshctx && !s->parallelmode) {

  4167.             adapt_probs(s);

  4168.             ff_thread_finish_setup(ctx);

  4169.         }

  4170.     } while (s->pass++ == 1);

  4171.     ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  4172. 

  4173.     // ref frame setup

  4174.     for (i = 0; i < 8; i++) {

  4175.         if (s->refs[i].f->data[0])

  4176.             ff_thread_release_buffer(ctx, &s->refs[i]);

  4177.         ff_thread_ref_frame(&s->refs[i], &s->next_refs[i]);

  4178.     }

  4179. 

  4180.     if (!s->invisible) {

  4181.         if ((res = av_frame_ref(frame, s->frames[CUR_FRAME].tf.f)) < 0)

  4182.             return res;

  4183.         *got_frame = 1;

  4184.     }

  4185. 

  4186.     return pkt->size;

  4187. }

  4188. 

  4189. static void vp9_decode_flush(AVCodecContext *ctx)

  4190. {

  4191.     VP9Context *s = ctx->priv_data;

  4192.     int i;

  4193. 

  4194.     for (i = 0; i < 3; i++)

  4195.         vp9_unref_frame(ctx, &s->frames[i]);

  4196.     for (i = 0; i < 8; i++)

  4197.         ff_thread_release_buffer(ctx, &s->refs[i]);

  4198. }

  4199. 

  4200. static int init_frames(AVCodecContext *ctx)

  4201. {

  4202.     VP9Context *s = ctx->priv_data;

  4203.     int i;

  4204. 

  4205.     for (i = 0; i < 3; i++) {

  4206.         s->frames[i].tf.f = av_frame_alloc();

  4207.         if (!s->frames[i].tf.f) {

  4208.             vp9_decode_free(ctx);

  4209.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  4210.             return AVERROR(ENOMEM);

  4211.         }

  4212.     }

  4213.     for (i = 0; i < 8; i++) {

  4214.         s->refs[i].f = av_frame_alloc();

  4215.         s->next_refs[i].f = av_frame_alloc();

  4216.         if (!s->refs[i].f || !s->next_refs[i].f) {

  4217.             vp9_decode_free(ctx);

  4218.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  4219.             return AVERROR(ENOMEM);

  4220.         }

  4221.     }

  4222. 

  4223.     return 0;

  4224. }

  4225. 

  4226. static av_cold int vp9_decode_init(AVCodecContext *ctx)

  4227. {

  4228.     VP9Context *s = ctx->priv_data;

  4229. 

  4230.     ctx->internal->allocate_progress = 1;

  4231.     s->last_bpp = 0;

  4232.     s->filter.sharpness = -1;

  4233. 

  4234.     return init_frames(ctx);

  4235. }

  4236. 

  4237. static av_cold int vp9_decode_init_thread_copy(AVCodecContext *avctx)

  4238. {

  4239.     return init_frames(avctx);

  4240. }

  4241. 

  4242. static int vp9_decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src)

  4243. {

  4244.     int i, res;

  4245.     VP9Context *s = dst->priv_data, *ssrc = src->priv_data;

  4246. 

  4247.     // detect size changes in other threads

  4248.     if (s->intra_pred_data[0] &&

  4249.         (!ssrc->intra_pred_data[0] || s->cols != ssrc->cols || s->rows != ssrc->rows)) {

  4250.         free_buffers(s);

  4251.     }

  4252. 

  4253.     for (i = 0; i < 3; i++) {

  4254.         if (s->frames[i].tf.f->data[0])

  4255.             vp9_unref_frame(dst, &s->frames[i]);

  4256.         if (ssrc->frames[i].tf.f->data[0]) {

  4257.             if ((res = vp9_ref_frame(dst, &s->frames[i], &ssrc->frames[i])) < 0)

  4258.                 return res;

  4259.         }

  4260.     }

  4261.     for (i = 0; i < 8; i++) {

  4262.         if (s->refs[i].f->data[0])

  4263.             ff_thread_release_buffer(dst, &s->refs[i]);

  4264.         if (ssrc->next_refs[i].f->data[0]) {

  4265.             if ((res = ff_thread_ref_frame(&s->refs[i], &ssrc->next_refs[i])) < 0)

  4266.                 return res;

  4267.         }

  4268.     }

  4269. 

  4270.     s->invisible = ssrc->invisible;

  4271.     s->keyframe = ssrc->keyframe;

  4272.     s->ss_v = ssrc->ss_v;

  4273.     s->ss_h = ssrc->ss_h;

  4274.     s->segmentation.enabled = ssrc->segmentation.enabled;

  4275.     s->segmentation.update_map = ssrc->segmentation.update_map;

  4276.     s->bytesperpixel = ssrc->bytesperpixel;

  4277.     s->bpp = ssrc->bpp;

  4278.     s->bpp_index = ssrc->bpp_index;

  4279.     memcpy(&s->prob_ctx, &ssrc->prob_ctx, sizeof(s->prob_ctx));

  4280.     memcpy(&s->lf_delta, &ssrc->lf_delta, sizeof(s->lf_delta));

  4281.     if (ssrc->segmentation.enabled) {

  4282.         memcpy(&s->segmentation.feat, &ssrc->segmentation.feat,

  4283.                sizeof(s->segmentation.feat));

  4284.     }

  4285. 

  4286.     return 0;

  4287. }

  4288. 

  4289. static const AVProfile profiles[] = {

  4290.     { FF_PROFILE_VP9_0, "Profile 0" },

  4291.     { FF_PROFILE_VP9_1, "Profile 1" },

  4292.     { FF_PROFILE_VP9_2, "Profile 2" },

  4293.     { FF_PROFILE_VP9_3, "Profile 3" },

  4294.     { FF_PROFILE_UNKNOWN },

  4295. };

  4296. 

  4297. AVCodec ff_vp9_decoder = {

  4298.     .name                  = "vp9",

  4299.     .long_name             = NULL_IF_CONFIG_SMALL("Google VP9"),

  4300.     .type                  = AVMEDIA_TYPE_VIDEO,

  4301.     .id                    = AV_CODEC_ID_VP9,

  4302.     .priv_data_size        = sizeof(VP9Context),

  4303.     .init                  = vp9_decode_init,

  4304.     .close                 = vp9_decode_free,

  4305.     .decode                = vp9_decode_frame,

  4306.     .capabilities          = CODEC_CAP_DR1 | CODEC_CAP_FRAME_THREADS,

  4307.     .flush                 = vp9_decode_flush,

  4308.     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(vp9_decode_init_thread_copy),

  4309.     .update_thread_context = ONLY_IF_THREADS_ENABLED(vp9_decode_update_thread_context),

  4310.     .profiles              = NULL_IF_CONFIG_SMALL(profiles),

  4311. };