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.
71824 Commits
32 Branches
380MB
Tree: 844799b033
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '844799b033'
${ noResults }
FFmpeg/libavcodec/vp9.c

4168 lines
164KB
Raw Blame History 

  1. /*

  2.  * VP9 compatible video decoder

  3.  *

  4.  * Copyright (C) 2013 Ronald S. Bultje <rsbultje gmail com>

  5.  * Copyright (C) 2013 Clément Bœsch <u pkh me>

  6.  *

  7.  * This file is part of FFmpeg.

  8.  *

  9.  * FFmpeg is free software; you can redistribute it and/or

  10.  * modify it under the terms of the GNU Lesser General Public

  11.  * License as published by the Free Software Foundation; either

  12.  * version 2.1 of the License, or (at your option) any later version.

  13.  *

  14.  * FFmpeg is distributed in the hope that it will be useful,

  15.  * but WITHOUT ANY WARRANTY; without even the implied warranty of

  16.  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU

  17.  * Lesser General Public License for more details.

  18.  *

  19.  * You should have received a copy of the GNU Lesser General Public

  20.  * License along with FFmpeg; if not, write to the Free Software

  21.  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA

  22.  */

  23. 

  24. #include "avcodec.h"

  25. #include "get_bits.h"

  26. #include "internal.h"

  27. #include "thread.h"

  28. #include "videodsp.h"

  29. #include "vp56.h"

  30. #include "vp9.h"

  31. #include "vp9data.h"

  32. #include "vp9dsp.h"

  33. #include "libavutil/avassert.h"

  34. 

  35. #define VP9_SYNCCODE 0x498342

  36. 

  37. enum CompPredMode {

  38.     PRED_SINGLEREF,

  39.     PRED_COMPREF,

  40.     PRED_SWITCHABLE,

  41. };

  42. 

  43. enum BlockLevel {

  44.     BL_64X64,

  45.     BL_32X32,

  46.     BL_16X16,

  47.     BL_8X8,

  48. };

  49. 

  50. enum BlockSize {

  51.     BS_64x64,

  52.     BS_64x32,

  53.     BS_32x64,

  54.     BS_32x32,

  55.     BS_32x16,

  56.     BS_16x32,

  57.     BS_16x16,

  58.     BS_16x8,

  59.     BS_8x16,

  60.     BS_8x8,

  61.     BS_8x4,

  62.     BS_4x8,

  63.     BS_4x4,

  64.     N_BS_SIZES,

  65. };

  66. 

  67. struct VP9mvrefPair {

  68.     VP56mv mv[2];

  69.     int8_t ref[2];

  70. };

  71. 

  72. typedef struct VP9Frame {

  73.     ThreadFrame tf;

  74.     AVBufferRef *extradata;

  75.     uint8_t *segmentation_map;

  76.     struct VP9mvrefPair *mv;

  77.     int uses_2pass;

  78. } VP9Frame;

  79. 

  80. struct VP9Filter {

  81.     uint8_t level[8 * 8];

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

  83.                               [8 /* rows */][4 /* 0=16, 1=8, 2=4, 3=inner4 */];

  84. };

  85. 

  86. typedef struct VP9Block {

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

  88.     enum FilterMode filter;

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

  90.     enum BlockSize bs;

  91.     enum TxfmMode tx, uvtx;

  92.     enum BlockLevel bl;

  93.     enum BlockPartition bp;

  94. } VP9Block;

  95. 

  96. typedef struct VP9Context {

  97.     VP9DSPContext dsp;

  98.     VideoDSPContext vdsp;

  99.     GetBitContext gb;

  100.     VP56RangeCoder c;

  101.     VP56RangeCoder *c_b;

  102.     unsigned c_b_size;

  103.     VP9Block *b_base, *b;

  104.     int pass;

  105.     int row, row7, col, col7;

  106.     uint8_t *dst[3];

  107.     ptrdiff_t y_stride, uv_stride;

  108. 

  109.     // bitstream header

  110.     uint8_t profile;

  111.     uint8_t keyframe, last_keyframe;

  112.     uint8_t invisible;

  113.     uint8_t use_last_frame_mvs;

  114.     uint8_t errorres;

  115.     uint8_t ss_h, ss_v;

  116.     uint8_t intraonly;

  117.     uint8_t resetctx;

  118.     uint8_t refreshrefmask;

  119.     uint8_t highprecisionmvs;

  120.     enum FilterMode filtermode;

  121.     uint8_t allowcompinter;

  122.     uint8_t fixcompref;

  123.     uint8_t refreshctx;

  124.     uint8_t parallelmode;

  125.     uint8_t framectxid;

  126.     uint8_t refidx[3];

  127.     uint8_t signbias[3];

  128.     uint8_t varcompref[2];

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

  130. #define CUR_FRAME 0

  131. #define REF_FRAME_MVPAIR 1

  132. #define REF_FRAME_SEGMAP 2

  133.     VP9Frame frames[3];

  134. 

  135.     struct {

  136.         uint8_t level;

  137.         int8_t sharpness;

  138.         uint8_t lim_lut[64];

  139.         uint8_t mblim_lut[64];

  140.     } filter;

  141.     struct {

  142.         uint8_t enabled;

  143.         int8_t mode[2];

  144.         int8_t ref[4];

  145.     } lf_delta;

  146.     uint8_t yac_qi;

  147.     int8_t ydc_qdelta, uvdc_qdelta, uvac_qdelta;

  148.     uint8_t lossless;

  149. #define MAX_SEGMENT 8

  150.     struct {

  151.         uint8_t enabled;

  152.         uint8_t temporal;

  153.         uint8_t absolute_vals;

  154.         uint8_t update_map;

  155.         struct {

  156.             uint8_t q_enabled;

  157.             uint8_t lf_enabled;

  158.             uint8_t ref_enabled;

  159.             uint8_t skip_enabled;

  160.             uint8_t ref_val;

  161.             int16_t q_val;

  162.             int8_t lf_val;

  163.             int16_t qmul[2][2];

  164.             uint8_t lflvl[4][2];

  165.         } feat[MAX_SEGMENT];

  166.     } segmentation;

  167.     struct {

  168.         unsigned log2_tile_cols, log2_tile_rows;

  169.         unsigned tile_cols, tile_rows;

  170.         unsigned tile_row_start, tile_row_end, tile_col_start, tile_col_end;

  171.     } tiling;

  172.     unsigned sb_cols, sb_rows, rows, cols;

  173.     struct {

  174.         prob_context p;

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

  176.     } prob_ctx[4];

  177.     struct {

  178.         prob_context p;

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

  180.         uint8_t seg[7];

  181.         uint8_t segpred[3];

  182.     } prob;

  183.     struct {

  184.         unsigned y_mode[4][10];

  185.         unsigned uv_mode[10][10];

  186.         unsigned filter[4][3];

  187.         unsigned mv_mode[7][4];

  188.         unsigned intra[4][2];

  189.         unsigned comp[5][2];

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

  191.         unsigned comp_ref[5][2];

  192.         unsigned tx32p[2][4];

  193.         unsigned tx16p[2][3];

  194.         unsigned tx8p[2][2];

  195.         unsigned skip[3][2];

  196.         unsigned mv_joint[4];

  197.         struct {

  198.             unsigned sign[2];

  199.             unsigned classes[11];

  200.             unsigned class0[2];

  201.             unsigned bits[10][2];

  202.             unsigned class0_fp[2][4];

  203.             unsigned fp[4];

  204.             unsigned class0_hp[2];

  205.             unsigned hp[2];

  206.         } mv_comp[2];

  207.         unsigned partition[4][4][4];

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

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

  210.     } counts;

  211.     enum TxfmMode txfmmode;

  212.     enum CompPredMode comppredmode;

  213. 

  214.     // contextual (left/above) cache

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

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

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

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

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

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

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

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

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

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

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

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

  227.     uint8_t *above_partition_ctx;

  228.     uint8_t *above_mode_ctx;

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

  230.     uint8_t *above_y_nnz_ctx;

  231.     uint8_t *above_uv_nnz_ctx[2];

  232.     uint8_t *above_skip_ctx; // 1bit

  233.     uint8_t *above_txfm_ctx; // 2bit

  234.     uint8_t *above_segpred_ctx; // 1bit

  235.     uint8_t *above_intra_ctx; // 1bit

  236.     uint8_t *above_comp_ctx; // 1bit

  237.     uint8_t *above_ref_ctx; // 2bit

  238.     uint8_t *above_filter_ctx;

  239.     VP56mv (*above_mv_ctx)[2];

  240. 

  241.     // whole-frame cache

  242.     uint8_t *intra_pred_data[3];

  243.     struct VP9Filter *lflvl;

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

  245. 

  246.     // block reconstruction intermediates

  247.     int block_alloc_using_2pass;

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

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

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

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

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

  253.     uint16_t mvscale[3][2];

  254.     uint8_t mvstep[3][2];

  255. } VP9Context;

  256. 

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

  258.     {

  259.         { 16, 16 }, { 16, 8 }, { 8, 16 }, { 8, 8 }, { 8, 4 }, { 4, 8 },

  260.         { 4, 4 }, { 4, 2 }, { 2, 4 }, { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 },

  261.     }, {

  262.         { 8, 8 }, { 8, 4 }, { 4, 8 }, { 4, 4 }, { 4, 2 }, { 2, 4 },

  263.         { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 }, { 1, 1 }, { 1, 1 }, { 1, 1 },

  264.     }

  265. };

  266. 

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

  268. {

  269.     VP9Context *s = ctx->priv_data;

  270.     int ret, sz;

  271. 

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

  273.         return ret;

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

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

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

  277.         return AVERROR(ENOMEM);

  278.     }

  279. 

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

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

  282. 

  283.     return 0;

  284. }

  285. 

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

  287. {

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

  289.     av_buffer_unref(&f->extradata);

  290. }

  291. 

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

  293. {

  294.     int res;

  295. 

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

  297.         return res;

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

  299.         vp9_unref_frame(ctx, dst);

  300.         return AVERROR(ENOMEM);

  301.     }

  302. 

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

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

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

  306. 

  307.     return 0;

  308. }

  309. 

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

  311. {

  312.     VP9Context *s = ctx->priv_data;

  313.     uint8_t *p;

  314. 

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

  316. 

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

  318.         return 0;

  319. 

  320.     ctx->width   = w;

  321.     ctx->height  = h;

  322.     ctx->pix_fmt = fmt;

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

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

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

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

  327. 

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

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

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

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

  332.     p = av_malloc(s->sb_cols * (320 + sizeof(*s->lflvl) + 16 * sizeof(*s->above_mv_ctx)));

  333.     if (!p)

  334.         return AVERROR(ENOMEM);

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  352. #undef assign

  353. 

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

  355.     av_freep(&s->b_base);

  356.     av_freep(&s->block_base);

  357. 

  358.     return 0;

  359. }

  360. 

  361. static int update_block_buffers(AVCodecContext *ctx)

  362. {

  363.     VP9Context *s = ctx->priv_data;

  364.     int chroma_blocks, chroma_eobs;

  365. 

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

  367.         return 0;

  368. 

  369.     av_free(s->b_base);

  370.     av_free(s->block_base);

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

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

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

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

  375. 

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

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

  378.                                     16 * 16 + 2 * chroma_eobs) * sbs);

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

  380.             return AVERROR(ENOMEM);

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

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

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

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

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

  386.     } else {

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

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

  389.                                    16 * 16 + 2 * chroma_eobs);

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

  391.             return AVERROR(ENOMEM);

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

  393.         s->uvblock_base[1] = s->uvblock_base[0] + chroma_blocks;

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

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

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

  397.     }

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

  399. 

  400.     return 0;

  401. }

  402. 

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

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

  405. {

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

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

  408. }

  409. 

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

  411. {

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

  413. }

  414. 

  415. // differential forward probability updates

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

  417. {

  418.     static const int inv_map_table[254] = {

  419.           7,  20,  33,  46,  59,  72,  85,  98, 111, 124, 137, 150, 163, 176,

  420.         189, 202, 215, 228, 241, 254,   1,   2,   3,   4,   5,   6,   8,   9,

  421.          10,  11,  12,  13,  14,  15,  16,  17,  18,  19,  21,  22,  23,  24,

  422.          25,  26,  27,  28,  29,  30,  31,  32,  34,  35,  36,  37,  38,  39,

  423.          40,  41,  42,  43,  44,  45,  47,  48,  49,  50,  51,  52,  53,  54,

  424.          55,  56,  57,  58,  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,

  425.          70,  71,  73,  74,  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,

  426.          86,  87,  88,  89,  90,  91,  92,  93,  94,  95,  96,  97,  99, 100,

  427.         101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 112, 113, 114, 115,

  428.         116, 117, 118, 119, 120, 121, 122, 123, 125, 126, 127, 128, 129, 130,

  429.         131, 132, 133, 134, 135, 136, 138, 139, 140, 141, 142, 143, 144, 145,

  430.         146, 147, 148, 149, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160,

  431.         161, 162, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175,

  432.         177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 190, 191,

  433.         192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 204, 205, 206,

  434.         207, 208, 209, 210, 211, 212, 213, 214, 216, 217, 218, 219, 220, 221,

  435.         222, 223, 224, 225, 226, 227, 229, 230, 231, 232, 233, 234, 235, 236,

  436.         237, 238, 239, 240, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251,

  437.         252, 253,

  438.     };

  439.     int d;

  440. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  455. 

  456.     if (!vp8_rac_get(c)) {

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

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

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

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

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

  462.     } else {

  463.         d = vp8_rac_get_uint(c, 7);

  464.         if (d >= 65)

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

  466.         d += 64;

  467.     }

  468. 

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

  470.                     255 - inv_recenter_nonneg(inv_map_table[d], 255 - p);

  471. }

  472. 

  473. static enum AVPixelFormat read_colorspace_details(AVCodecContext *ctx)

  474. {

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

  476.         AVCOL_SPC_UNSPECIFIED, AVCOL_SPC_BT470BG, AVCOL_SPC_BT709, AVCOL_SPC_SMPTE170M,

  477.         AVCOL_SPC_SMPTE240M, AVCOL_SPC_BT2020_NCL, AVCOL_SPC_RESERVED, AVCOL_SPC_RGB,

  478.     };

  479.     VP9Context *s = ctx->priv_data;

  480.     enum AVPixelFormat res;

  481. 

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

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

  484.         if (s->profile == 1) {

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

  486.             res = AV_PIX_FMT_GBRP;

  487.             ctx->color_range = AVCOL_RANGE_JPEG;

  488.         } else {

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

  490.             return AVERROR_INVALIDDATA;

  491.         }

  492.     } else {

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

  494.             { AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV422P },

  495.             { AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV420P },

  496.         };

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

  498.         if (s->profile == 1) {

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

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

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

  502.                 av_log(ctx, AV_LOG_ERROR, "YUV 4:2:0 not supported in profile 1\n");

  503.                 return AVERROR_INVALIDDATA;

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

  505.                 av_log(ctx, AV_LOG_ERROR, "Profile 1 color details reserved bit set\n");

  506.                 return AVERROR_INVALIDDATA;

  507.             }

  508.         } else {

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

  510.             res = AV_PIX_FMT_YUV420P;

  511.         }

  512.     }

  513. 

  514.     return res;

  515. }

  516. 

  517. static int decode_frame_header(AVCodecContext *ctx,

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

  519. {

  520.     VP9Context *s = ctx->priv_data;

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

  522.     enum AVPixelFormat fmt = ctx->pix_fmt;

  523.     int last_invisible;

  524.     const uint8_t *data2;

  525. 

  526.     /* general header */

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

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

  529.         return res;

  530.     }

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

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

  533.         return AVERROR_INVALIDDATA;

  534.     }

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

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

  537.     if (s->profile > 1) {

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

  539.         return AVERROR_INVALIDDATA;

  540.     }

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

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

  543.         return 0;

  544.     }

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

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

  547.     last_invisible    = s->invisible;

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

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

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

  551.     if (s->keyframe) {

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

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

  554.             return AVERROR_INVALIDDATA;

  555.         }

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

  557.             return fmt;

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

  559.         s->refreshrefmask = 0xff;

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

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

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

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

  564.     } else {

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

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

  567.         if (s->intraonly) {

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

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

  570.                 return AVERROR_INVALIDDATA;

  571.             }

  572.             if (s->profile == 1) {

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

  574.                     return fmt;

  575.             } else {

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

  577.                 fmt = AV_PIX_FMT_YUV420P;

  578.                 ctx->colorspace = AVCOL_SPC_BT470BG;

  579.                 ctx->color_range = AVCOL_RANGE_JPEG;

  580.             }

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

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

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

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

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

  586.         } else {

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

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

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

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

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

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

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

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

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

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

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

  598.                 return AVERROR_INVALIDDATA;

  599.             }

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

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

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

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

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

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

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

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

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

  609.             } else {

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

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

  612.             }

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

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

  615.             // the _last_ frame

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

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

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

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

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

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

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

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

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

  625.             if (s->allowcompinter) {

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

  627.                     s->fixcompref    = 2;

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

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

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

  631.                     s->fixcompref    = 1;

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

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

  634.                 } else {

  635.                     s->fixcompref    = 0;

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

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

  638.                 }

  639.             }

  640. 

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

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

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

  644. 

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

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

  647.                 } else {

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

  649.                         av_log(ctx, AV_LOG_ERROR,

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

  651.                                refw, refh, w, h);

  652.                         return AVERROR_INVALIDDATA;

  653.                     }

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

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

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

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

  658.                 }

  659.             }

  660.         }

  661.     }

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

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

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

  665. 

  666.     /* loopfilter header data */

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

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

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

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

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

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

  673.     s->filter.sharpness = sharp;

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

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

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

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

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

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

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

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

  682.         }

  683.     }

  684. 

  685.     /* quantization header data */

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

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

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

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

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

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

  692. 

  693.     /* segmentation header info */

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

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

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

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

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

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

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

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

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

  703.             }

  704.         }

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

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

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

  708.             av_log(ctx, AV_LOG_ERROR,

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

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

  711.             return AVERROR_INVALIDDATA;

  712.         }

  713. 

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

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

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

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

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

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

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

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

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

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

  724.             }

  725.         }

  726.     } else {

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

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

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

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

  731.     }

  732. 

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

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

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

  736. 

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

  738.             if (s->segmentation.absolute_vals)

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

  740.             else

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

  742.         } else {

  743.             qyac  = s->yac_qi;

  744.         }

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

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

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

  748.         qyac  = av_clip_uintp2(qyac, 8);

  749. 

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

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

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

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

  754. 

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

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

  757.             if (s->segmentation.absolute_vals)

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

  759.             else

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

  761.         } else {

  762.             lflvl  = s->filter.level;

  763.         }

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

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

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

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

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

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

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

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

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

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

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

  775.             }

  776.         } else {

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

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

  779.         }

  780.     }

  781. 

  782.     /* tiling info */

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

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

  785.         return res;

  786.     }

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

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

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

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

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

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

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

  794.             s->tiling.log2_tile_cols++;

  795.         else

  796.             break;

  797.     }

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

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

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

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

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

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

  804.         if (!s->c_b) {

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

  806.             return AVERROR(ENOMEM);

  807.         }

  808.     }

  809. 

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

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

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

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

  814.                sizeof(vp9_default_coef_probs));

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

  816.                sizeof(vp9_default_coef_probs));

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

  818.                sizeof(vp9_default_coef_probs));

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

  820.                sizeof(vp9_default_coef_probs));

  821.     }

  822. 

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

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

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

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

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

  828.         return AVERROR_INVALIDDATA;

  829.     }

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

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

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

  833.         return AVERROR_INVALIDDATA;

  834.     }

  835. 

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

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

  838.     } else {

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

  840.     }

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

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

  843.     // fw update)?

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

  845. 

  846.     // txfm updates

  847.     if (s->lossless) {

  848.         s->txfmmode = TX_4X4;

  849.     } else {

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

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

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

  853. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  868.         }

  869.     }

  870. 

  871.     // coef updates

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

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

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

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

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

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

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

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

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

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

  882.                                 break;

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

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

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

  886.                                 } else {

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

  888.                                 }

  889.                             }

  890.                             p[3] = 0;

  891.                         }

  892.         } else {

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

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

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

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

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

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

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

  900.                                 break;

  901.                             memcpy(p, r, 3);

  902.                             p[3] = 0;

  903.                         }

  904.         }

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

  906.             break;

  907.     }

  908. 

  909.     // mode updates

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

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

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

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

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

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

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

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

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

  919. 

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

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

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

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

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

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

  926. 

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

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

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

  930. 

  931.         if (s->allowcompinter) {

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

  933.             if (s->comppredmode)

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

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

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

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

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

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

  940.         } else {

  941.             s->comppredmode = PRED_SINGLEREF;

  942.         }

  943. 

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

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

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

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

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

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

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

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

  952.             }

  953.         }

  954. 

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

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

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

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

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

  960.         }

  961. 

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

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

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

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

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

  967. 

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

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

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

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

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

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

  974. 

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

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

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

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

  979. 

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

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

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

  983. 

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

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

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

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

  988. 

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

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

  991. 

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

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

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

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

  996.         }

  997. 

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

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

  1000.                 for (k = 0; k < 3; k++)

  1001.                     if (vp56_rac_get_prob_branchy(&s->c, 252))

  1002.                         s->prob.p.mv_comp[i].class0_fp[j][k] =

  1003.                             (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  1004. 

  1005.             for (j = 0; j < 3; j++)

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

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

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

  1009.         }

  1010. 

  1011.         if (s->highprecisionmvs) {

  1012.             for (i = 0; i < 2; i++) {

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

  1014.                     s->prob.p.mv_comp[i].class0_hp =

  1015.                         (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  1016. 

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

  1018.                     s->prob.p.mv_comp[i].hp =

  1019.                         (vp8_rac_get_uint(&s->c, 7) << 1) | 1;

  1020.             }

  1021.         }

  1022.     }

  1023. 

  1024.     return (data2 - data) + size2;

  1025. }

  1026. 

  1027. static av_always_inline void clamp_mv(VP56mv *dst, const VP56mv *src,

  1028.                                       VP9Context *s)

  1029. {

  1030.     dst->x = av_clip(src->x, s->min_mv.x, s->max_mv.x);

  1031.     dst->y = av_clip(src->y, s->min_mv.y, s->max_mv.y);

  1032. }

  1033. 

  1034. static void find_ref_mvs(VP9Context *s,

  1035.                          VP56mv *pmv, int ref, int z, int idx, int sb)

  1036. {

  1037.     static const int8_t mv_ref_blk_off[N_BS_SIZES][8][2] = {

  1038.         [BS_64x64] = {{  3, -1 }, { -1,  3 }, {  4, -1 }, { -1,  4 },

  1039.                       { -1, -1 }, {  0, -1 }, { -1,  0 }, {  6, -1 }},

  1040.         [BS_64x32] = {{  0, -1 }, { -1,  0 }, {  4, -1 }, { -1,  2 },

  1041.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, {  2, -1 }},

  1042.         [BS_32x64] = {{ -1,  0 }, {  0, -1 }, { -1,  4 }, {  2, -1 },

  1043.                       { -1, -1 }, { -3,  0 }, {  0, -3 }, { -1,  2 }},

  1044.         [BS_32x32] = {{  1, -1 }, { -1,  1 }, {  2, -1 }, { -1,  2 },

  1045.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

  1046.         [BS_32x16] = {{  0, -1 }, { -1,  0 }, {  2, -1 }, { -1, -1 },

  1047.                       { -1,  1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

  1048.         [BS_16x32] = {{ -1,  0 }, {  0, -1 }, { -1,  2 }, { -1, -1 },

  1049.                       {  1, -1 }, { -3,  0 }, {  0, -3 }, { -3, -3 }},

  1050.         [BS_16x16] = {{  0, -1 }, { -1,  0 }, {  1, -1 }, { -1,  1 },

  1051.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

  1052.         [BS_16x8]  = {{  0, -1 }, { -1,  0 }, {  1, -1 }, { -1, -1 },

  1053.                       {  0, -2 }, { -2,  0 }, { -2, -1 }, { -1, -2 }},

  1054.         [BS_8x16]  = {{ -1,  0 }, {  0, -1 }, { -1,  1 }, { -1, -1 },

  1055.                       { -2,  0 }, {  0, -2 }, { -1, -2 }, { -2, -1 }},

  1056.         [BS_8x8]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1057.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1058.         [BS_8x4]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1059.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1060.         [BS_4x8]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1061.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1062.         [BS_4x4]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1063.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1064.     };

  1065.     VP9Block *b = s->b;

  1066.     int row = s->row, col = s->col, row7 = s->row7;

  1067.     const int8_t (*p)[2] = mv_ref_blk_off[b->bs];

  1068. #define INVALID_MV 0x80008000U

  1069.     uint32_t mem = INVALID_MV;

  1070.     int i;

  1071. 

  1072. #define RETURN_DIRECT_MV(mv) \

  1073.     do { \

  1074.         uint32_t m = AV_RN32A(&mv); \

  1075.         if (!idx) { \

  1076.             AV_WN32A(pmv, m); \

  1077.             return; \

  1078.         } else if (mem == INVALID_MV) { \

  1079.             mem = m; \

  1080.         } else if (m != mem) { \

  1081.             AV_WN32A(pmv, m); \

  1082.             return; \

  1083.         } \

  1084.     } while (0)

  1085. 

  1086.     if (sb >= 0) {

  1087.         if (sb == 2 || sb == 1) {

  1088.             RETURN_DIRECT_MV(b->mv[0][z]);

  1089.         } else if (sb == 3) {

  1090.             RETURN_DIRECT_MV(b->mv[2][z]);

  1091.             RETURN_DIRECT_MV(b->mv[1][z]);

  1092.             RETURN_DIRECT_MV(b->mv[0][z]);

  1093.         }

  1094. 

  1095. #define RETURN_MV(mv) \

  1096.     do { \

  1097.         if (sb > 0) { \

  1098.             VP56mv tmp; \

  1099.             uint32_t m; \

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

  1101.             m = AV_RN32A(&tmp); \

  1102.             if (!idx) { \

  1103.                 AV_WN32A(pmv, m); \

  1104.                 return; \

  1105.             } else if (mem == INVALID_MV) { \

  1106.                 mem = m; \

  1107.             } else if (m != mem) { \

  1108.                 AV_WN32A(pmv, m); \

  1109.                 return; \

  1110.             } \

  1111.         } else { \

  1112.             uint32_t m = AV_RN32A(&mv); \

  1113.             if (!idx) { \

  1114.                 clamp_mv(pmv, &mv, s); \

  1115.                 return; \

  1116.             } else if (mem == INVALID_MV) { \

  1117.                 mem = m; \

  1118.             } else if (m != mem) { \

  1119.                 clamp_mv(pmv, &mv, s); \

  1120.                 return; \

  1121.             } \

  1122.         } \

  1123.     } while (0)

  1124. 

  1125.         if (row > 0) {

  1126.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[(row - 1) * s->sb_cols * 8 + col];

  1127.             if (mv->ref[0] == ref) {

  1128.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][0]);

  1129.             } else if (mv->ref[1] == ref) {

  1130.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][1]);

  1131.             }

  1132.         }

  1133.         if (col > s->tiling.tile_col_start) {

  1134.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[row * s->sb_cols * 8 + col - 1];

  1135.             if (mv->ref[0] == ref) {

  1136.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][0]);

  1137.             } else if (mv->ref[1] == ref) {

  1138.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][1]);

  1139.             }

  1140.         }

  1141.         i = 2;

  1142.     } else {

  1143.         i = 0;

  1144.     }

  1145. 

  1146.     // previously coded MVs in this neighbourhood, using same reference frame

  1147.     for (; i < 8; i++) {

  1148.         int c = p[i][0] + col, r = p[i][1] + row;

  1149. 

  1150.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1151.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1152. 

  1153.             if (mv->ref[0] == ref) {

  1154.                 RETURN_MV(mv->mv[0]);

  1155.             } else if (mv->ref[1] == ref) {

  1156.                 RETURN_MV(mv->mv[1]);

  1157.             }

  1158.         }

  1159.     }

  1160. 

  1161.     // MV at this position in previous frame, using same reference frame

  1162.     if (s->use_last_frame_mvs) {

  1163.         struct VP9mvrefPair *mv = &s->frames[REF_FRAME_MVPAIR].mv[row * s->sb_cols * 8 + col];

  1164. 

  1165.         if (!s->frames[REF_FRAME_MVPAIR].uses_2pass)

  1166.             ff_thread_await_progress(&s->frames[REF_FRAME_MVPAIR].tf, row >> 3, 0);

  1167.         if (mv->ref[0] == ref) {

  1168.             RETURN_MV(mv->mv[0]);

  1169.         } else if (mv->ref[1] == ref) {

  1170.             RETURN_MV(mv->mv[1]);

  1171.         }

  1172.     }

  1173. 

  1174. #define RETURN_SCALE_MV(mv, scale) \

  1175.     do { \

  1176.         if (scale) { \

  1177.             VP56mv mv_temp = { -mv.x, -mv.y }; \

  1178.             RETURN_MV(mv_temp); \

  1179.         } else { \

  1180.             RETURN_MV(mv); \

  1181.         } \

  1182.     } while (0)

  1183. 

  1184.     // previously coded MVs in this neighbourhood, using different reference frame

  1185.     for (i = 0; i < 8; i++) {

  1186.         int c = p[i][0] + col, r = p[i][1] + row;

  1187. 

  1188.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1189.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1190. 

  1191.             if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1192.                 RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1193.             }

  1194.             if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1195.                 // BUG - libvpx has this condition regardless of whether

  1196.                 // we used the first ref MV and pre-scaling

  1197.                 AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1198.                 RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1199.             }

  1200.         }

  1201.     }

  1202. 

  1203.     // MV at this position in previous frame, using different reference frame

  1204.     if (s->use_last_frame_mvs) {

  1205.         struct VP9mvrefPair *mv = &s->frames[REF_FRAME_MVPAIR].mv[row * s->sb_cols * 8 + col];

  1206. 

  1207.         // no need to await_progress, because we already did that above

  1208.         if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1209.             RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1210.         }

  1211.         if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1212.             // BUG - libvpx has this condition regardless of whether

  1213.             // we used the first ref MV and pre-scaling

  1214.             AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1215.             RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1216.         }

  1217.     }

  1218. 

  1219.     AV_ZERO32(pmv);

  1220. #undef INVALID_MV

  1221. #undef RETURN_MV

  1222. #undef RETURN_SCALE_MV

  1223. }

  1224. 

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

  1226. {

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

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

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

  1230. 

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

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

  1233.     if (c) {

  1234.         int m;

  1235. 

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

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

  1238.             n |= bit << m;

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

  1240.         }

  1241.         n <<= 3;

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

  1243.         n |= bit << 1;

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

  1245.         if (hp) {

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

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

  1248.             n |= bit;

  1249.         } else {

  1250.             n |= 1;

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

  1252.             // bit wasn't coded

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

  1254.         }

  1255.         n += 8 << c;

  1256.     } else {

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

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

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

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

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

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

  1263.         if (hp) {

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

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

  1266.             n |= bit;

  1267.         } else {

  1268.             n |= 1;

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

  1270.             // bit wasn't coded

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

  1272.         }

  1273.     }

  1274. 

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

  1276. }

  1277. 

  1278. static void fill_mv(VP9Context *s,

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

  1280. {

  1281.     VP9Block *b = s->b;

  1282. 

  1283.     if (mode == ZEROMV) {

  1284.         AV_ZERO64(mv);

  1285.     } else {

  1286.         int hp;

  1287. 

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

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

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

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

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

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

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

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

  1296.                     mv[0].y++;

  1297.                 else

  1298.                     mv[0].y--;

  1299.             }

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

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

  1302.                     mv[0].x++;

  1303.                 else

  1304.                     mv[0].x--;

  1305.             }

  1306.         }

  1307.         if (mode == NEWMV) {

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

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

  1310. 

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

  1312.             if (j >= MV_JOINT_V)

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

  1314.             if (j & 1)

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

  1316.         }

  1317. 

  1318.         if (b->comp) {

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

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

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

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

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

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

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

  1326.                         mv[1].y++;

  1327.                     else

  1328.                         mv[1].y--;

  1329.                 }

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

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

  1332.                         mv[1].x++;

  1333.                     else

  1334.                         mv[1].x--;

  1335.                 }

  1336.             }

  1337.             if (mode == NEWMV) {

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

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

  1340. 

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

  1342.                 if (j >= MV_JOINT_V)

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

  1344.                 if (j & 1)

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

  1346.             }

  1347.         }

  1348.     }

  1349. }

  1350. 

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

  1352.                                        ptrdiff_t stride, int v)

  1353. {

  1354.     switch (w) {

  1355.     case 1:

  1356.         do {

  1357.             *ptr = v;

  1358.             ptr += stride;

  1359.         } while (--h);

  1360.         break;

  1361.     case 2: {

  1362.         int v16 = v * 0x0101;

  1363.         do {

  1364.             AV_WN16A(ptr, v16);

  1365.             ptr += stride;

  1366.         } while (--h);

  1367.         break;

  1368.     }

  1369.     case 4: {

  1370.         uint32_t v32 = v * 0x01010101;

  1371.         do {

  1372.             AV_WN32A(ptr, v32);

  1373.             ptr += stride;

  1374.         } while (--h);

  1375.         break;

  1376.     }

  1377.     case 8: {

  1378. #if HAVE_FAST_64BIT

  1379.         uint64_t v64 = v * 0x0101010101010101ULL;

  1380.         do {

  1381.             AV_WN64A(ptr, v64);

  1382.             ptr += stride;

  1383.         } while (--h);

  1384. #else

  1385.         uint32_t v32 = v * 0x01010101;

  1386.         do {

  1387.             AV_WN32A(ptr,     v32);

  1388.             AV_WN32A(ptr + 4, v32);

  1389.             ptr += stride;

  1390.         } while (--h);

  1391. #endif

  1392.         break;

  1393.     }

  1394.     }

  1395. }

  1396. 

  1397. static void decode_mode(AVCodecContext *ctx)

  1398. {

  1399.     static const uint8_t left_ctx[N_BS_SIZES] = {

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

  1401.     };

  1402.     static const uint8_t above_ctx[N_BS_SIZES] = {

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

  1404.     };

  1405.     static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {

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

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

  1408.     };

  1409.     VP9Context *s = ctx->priv_data;

  1410.     VP9Block *b = s->b;

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

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

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

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

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

  1416.     int vref, filter_id;

  1417. 

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

  1419.         b->seg_id = 0;

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

  1421.         b->seg_id = vp8_rac_get_tree(&s->c, vp9_segmentation_tree, s->prob.seg);

  1422.     } else if (!s->segmentation.update_map ||

  1423.                (s->segmentation.temporal &&

  1424.                 vp56_rac_get_prob_branchy(&s->c,

  1425.                     s->prob.segpred[s->above_segpred_ctx[col] +

  1426.                                     s->left_segpred_ctx[row7]]))) {

  1427.         if (!s->errorres) {

  1428.             int pred = 8, x;

  1429.             uint8_t *refsegmap = s->frames[REF_FRAME_SEGMAP].segmentation_map;

  1430. 

  1431.             if (!s->frames[REF_FRAME_SEGMAP].uses_2pass)

  1432.                 ff_thread_await_progress(&s->frames[REF_FRAME_SEGMAP].tf, row >> 3, 0);

  1433.             for (y = 0; y < h4; y++) {

  1434.                 int idx_base = (y + row) * 8 * s->sb_cols + col;

  1435.                 for (x = 0; x < w4; x++)

  1436.                     pred = FFMIN(pred, refsegmap[idx_base + x]);

  1437.             }

  1438.             av_assert1(pred < 8);

  1439.             b->seg_id = pred;

  1440.         } else {

  1441.             b->seg_id = 0;

  1442.         }

  1443. 

  1444.         memset(&s->above_segpred_ctx[col], 1, w4);

  1445.         memset(&s->left_segpred_ctx[row7], 1, h4);

  1446.     } else {

  1447.         b->seg_id = vp8_rac_get_tree(&s->c, vp9_segmentation_tree,

  1448.                                      s->prob.seg);

  1449. 

  1450.         memset(&s->above_segpred_ctx[col], 0, w4);

  1451.         memset(&s->left_segpred_ctx[row7], 0, h4);

  1452.     }

  1453.     if (s->segmentation.enabled &&

  1454.         (s->segmentation.update_map || s->keyframe || s->intraonly)) {

  1455.         setctx_2d(&s->frames[CUR_FRAME].segmentation_map[row * 8 * s->sb_cols + col],

  1456.                   bw4, bh4, 8 * s->sb_cols, b->seg_id);

  1457.     }

  1458. 

  1459.     b->skip = s->segmentation.enabled &&

  1460.         s->segmentation.feat[b->seg_id].skip_enabled;

  1461.     if (!b->skip) {

  1462.         int c = s->left_skip_ctx[row7] + s->above_skip_ctx[col];

  1463.         b->skip = vp56_rac_get_prob(&s->c, s->prob.p.skip[c]);

  1464.         s->counts.skip[c][b->skip]++;

  1465.     }

  1466. 

  1467.     if (s->keyframe || s->intraonly) {

  1468.         b->intra = 1;

  1469.     } else if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1470.         b->intra = !s->segmentation.feat[b->seg_id].ref_val;

  1471.     } else {

  1472.         int c, bit;

  1473. 

  1474.         if (have_a && have_l) {

  1475.             c = s->above_intra_ctx[col] + s->left_intra_ctx[row7];

  1476.             c += (c == 2);

  1477.         } else {

  1478.             c = have_a ? 2 * s->above_intra_ctx[col] :

  1479.                 have_l ? 2 * s->left_intra_ctx[row7] : 0;

  1480.         }

  1481.         bit = vp56_rac_get_prob(&s->c, s->prob.p.intra[c]);

  1482.         s->counts.intra[c][bit]++;

  1483.         b->intra = !bit;

  1484.     }

  1485. 

  1486.     if ((b->intra || !b->skip) && s->txfmmode == TX_SWITCHABLE) {

  1487.         int c;

  1488.         if (have_a) {

  1489.             if (have_l) {

  1490.                 c = (s->above_skip_ctx[col] ? max_tx :

  1491.                      s->above_txfm_ctx[col]) +

  1492.                     (s->left_skip_ctx[row7] ? max_tx :

  1493.                      s->left_txfm_ctx[row7]) > max_tx;

  1494.             } else {

  1495.                 c = s->above_skip_ctx[col] ? 1 :

  1496.                     (s->above_txfm_ctx[col] * 2 > max_tx);

  1497.             }

  1498.         } else if (have_l) {

  1499.             c = s->left_skip_ctx[row7] ? 1 :

  1500.                 (s->left_txfm_ctx[row7] * 2 > max_tx);

  1501.         } else {

  1502.             c = 1;

  1503.         }

  1504.         switch (max_tx) {

  1505.         case TX_32X32:

  1506.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][0]);

  1507.             if (b->tx) {

  1508.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][1]);

  1509.                 if (b->tx == 2)

  1510.                     b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][2]);

  1511.             }

  1512.             s->counts.tx32p[c][b->tx]++;

  1513.             break;

  1514.         case TX_16X16:

  1515.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][0]);

  1516.             if (b->tx)

  1517.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][1]);

  1518.             s->counts.tx16p[c][b->tx]++;

  1519.             break;

  1520.         case TX_8X8:

  1521.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx8p[c]);

  1522.             s->counts.tx8p[c][b->tx]++;

  1523.             break;

  1524.         case TX_4X4:

  1525.             b->tx = TX_4X4;

  1526.             break;

  1527.         }

  1528.     } else {

  1529.         b->tx = FFMIN(max_tx, s->txfmmode);

  1530.     }

  1531. 

  1532.     if (s->keyframe || s->intraonly) {

  1533.         uint8_t *a = &s->above_mode_ctx[col * 2];

  1534.         uint8_t *l = &s->left_mode_ctx[(row7) << 1];

  1535. 

  1536.         b->comp = 0;

  1537.         if (b->bs > BS_8x8) {

  1538.             // FIXME the memory storage intermediates here aren't really

  1539.             // necessary, they're just there to make the code slightly

  1540.             // simpler for now

  1541.             b->mode[0] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1542.                                     vp9_default_kf_ymode_probs[a[0]][l[0]]);

  1543.             if (b->bs != BS_8x4) {

  1544.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1545.                                  vp9_default_kf_ymode_probs[a[1]][b->mode[0]]);

  1546.                 l[0] = a[1] = b->mode[1];

  1547.             } else {

  1548.                 l[0] = a[1] = b->mode[1] = b->mode[0];

  1549.             }

  1550.             if (b->bs != BS_4x8) {

  1551.                 b->mode[2] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1552.                                         vp9_default_kf_ymode_probs[a[0]][l[1]]);

  1553.                 if (b->bs != BS_8x4) {

  1554.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1555.                                   vp9_default_kf_ymode_probs[a[1]][b->mode[2]]);

  1556.                     l[1] = a[1] = b->mode[3];

  1557.                 } else {

  1558.                     l[1] = a[1] = b->mode[3] = b->mode[2];

  1559.                 }

  1560.             } else {

  1561.                 b->mode[2] = b->mode[0];

  1562.                 l[1] = a[1] = b->mode[3] = b->mode[1];

  1563.             }

  1564.         } else {

  1565.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1566.                                           vp9_default_kf_ymode_probs[*a][*l]);

  1567.             b->mode[3] = b->mode[2] = b->mode[1] = b->mode[0];

  1568.             // FIXME this can probably be optimized

  1569.             memset(a, b->mode[0], bwh_tab[0][b->bs][0]);

  1570.             memset(l, b->mode[0], bwh_tab[0][b->bs][1]);

  1571.         }

  1572.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1573.                                      vp9_default_kf_uvmode_probs[b->mode[3]]);

  1574.     } else if (b->intra) {

  1575.         b->comp = 0;

  1576.         if (b->bs > BS_8x8) {

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

  1578.                                           s->prob.p.y_mode[0]);

  1579.             s->counts.y_mode[0][b->mode[0]]++;

  1580.             if (b->bs != BS_8x4) {

  1581.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1582.                                               s->prob.p.y_mode[0]);

  1583.                 s->counts.y_mode[0][b->mode[1]]++;

  1584.             } else {

  1585.                 b->mode[1] = b->mode[0];

  1586.             }

  1587.             if (b->bs != BS_4x8) {

  1588.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1589.                                               s->prob.p.y_mode[0]);

  1590.                 s->counts.y_mode[0][b->mode[2]]++;

  1591.                 if (b->bs != BS_8x4) {

  1592.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1593.                                                   s->prob.p.y_mode[0]);

  1594.                     s->counts.y_mode[0][b->mode[3]]++;

  1595.                 } else {

  1596.                     b->mode[3] = b->mode[2];

  1597.                 }

  1598.             } else {

  1599.                 b->mode[2] = b->mode[0];

  1600.                 b->mode[3] = b->mode[1];

  1601.             }

  1602.         } else {

  1603.             static const uint8_t size_group[10] = {

  1604.                 3, 3, 3, 3, 2, 2, 2, 1, 1, 1

  1605.             };

  1606.             int sz = size_group[b->bs];

  1607. 

  1608.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1609.                                           s->prob.p.y_mode[sz]);

  1610.             b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1611.             s->counts.y_mode[sz][b->mode[3]]++;

  1612.         }

  1613.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1614.                                      s->prob.p.uv_mode[b->mode[3]]);

  1615.         s->counts.uv_mode[b->mode[3]][b->uvmode]++;

  1616.     } else {

  1617.         static const uint8_t inter_mode_ctx_lut[14][14] = {

  1618.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1619.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1620.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1621.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1622.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1623.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1624.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1625.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1626.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1627.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1628.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1629.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1630.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },

  1631.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },

  1632.         };

  1633. 

  1634.         if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1635.             av_assert2(s->segmentation.feat[b->seg_id].ref_val != 0);

  1636.             b->comp = 0;

  1637.             b->ref[0] = s->segmentation.feat[b->seg_id].ref_val - 1;

  1638.         } else {

  1639.             // read comp_pred flag

  1640.             if (s->comppredmode != PRED_SWITCHABLE) {

  1641.                 b->comp = s->comppredmode == PRED_COMPREF;

  1642.             } else {

  1643.                 int c;

  1644. 

  1645.                 // FIXME add intra as ref=0xff (or -1) to make these easier?

  1646.                 if (have_a) {

  1647.                     if (have_l) {

  1648.                         if (s->above_comp_ctx[col] && s->left_comp_ctx[row7]) {

  1649.                             c = 4;

  1650.                         } else if (s->above_comp_ctx[col]) {

  1651.                             c = 2 + (s->left_intra_ctx[row7] ||

  1652.                                      s->left_ref_ctx[row7] == s->fixcompref);

  1653.                         } else if (s->left_comp_ctx[row7]) {

  1654.                             c = 2 + (s->above_intra_ctx[col] ||

  1655.                                      s->above_ref_ctx[col] == s->fixcompref);

  1656.                         } else {

  1657.                             c = (!s->above_intra_ctx[col] &&

  1658.                                  s->above_ref_ctx[col] == s->fixcompref) ^

  1659.                             (!s->left_intra_ctx[row7] &&

  1660.                              s->left_ref_ctx[row & 7] == s->fixcompref);

  1661.                         }

  1662.                     } else {

  1663.                         c = s->above_comp_ctx[col] ? 3 :

  1664.                         (!s->above_intra_ctx[col] && s->above_ref_ctx[col] == s->fixcompref);

  1665.                     }

  1666.                 } else if (have_l) {

  1667.                     c = s->left_comp_ctx[row7] ? 3 :

  1668.                     (!s->left_intra_ctx[row7] && s->left_ref_ctx[row7] == s->fixcompref);

  1669.                 } else {

  1670.                     c = 1;

  1671.                 }

  1672.                 b->comp = vp56_rac_get_prob(&s->c, s->prob.p.comp[c]);

  1673.                 s->counts.comp[c][b->comp]++;

  1674.             }

  1675. 

  1676.             // read actual references

  1677.             // FIXME probably cache a few variables here to prevent repetitive

  1678.             // memory accesses below

  1679.             if (b->comp) /* two references */ {

  1680.                 int fix_idx = s->signbias[s->fixcompref], var_idx = !fix_idx, c, bit;

  1681. 

  1682.                 b->ref[fix_idx] = s->fixcompref;

  1683.                 // FIXME can this codeblob be replaced by some sort of LUT?

  1684.                 if (have_a) {

  1685.                     if (have_l) {

  1686.                         if (s->above_intra_ctx[col]) {

  1687.                             if (s->left_intra_ctx[row7]) {

  1688.                                 c = 2;

  1689.                             } else {

  1690.                                 c = 1 + 2 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1691.                             }

  1692.                         } else if (s->left_intra_ctx[row7]) {

  1693.                             c = 1 + 2 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1694.                         } else {

  1695.                             int refl = s->left_ref_ctx[row7], refa = s->above_ref_ctx[col];

  1696. 

  1697.                             if (refl == refa && refa == s->varcompref[1]) {

  1698.                                 c = 0;

  1699.                             } else if (!s->left_comp_ctx[row7] && !s->above_comp_ctx[col]) {

  1700.                                 if ((refa == s->fixcompref && refl == s->varcompref[0]) ||

  1701.                                     (refl == s->fixcompref && refa == s->varcompref[0])) {

  1702.                                     c = 4;

  1703.                                 } else {

  1704.                                     c = (refa == refl) ? 3 : 1;

  1705.                                 }

  1706.                             } else if (!s->left_comp_ctx[row7]) {

  1707.                                 if (refa == s->varcompref[1] && refl != s->varcompref[1]) {

  1708.                                     c = 1;

  1709.                                 } else {

  1710.                                     c = (refl == s->varcompref[1] &&

  1711.                                          refa != s->varcompref[1]) ? 2 : 4;

  1712.                                 }

  1713.                             } else if (!s->above_comp_ctx[col]) {

  1714.                                 if (refl == s->varcompref[1] && refa != s->varcompref[1]) {

  1715.                                     c = 1;

  1716.                                 } else {

  1717.                                     c = (refa == s->varcompref[1] &&

  1718.                                          refl != s->varcompref[1]) ? 2 : 4;

  1719.                                 }

  1720.                             } else {

  1721.                                 c = (refl == refa) ? 4 : 2;

  1722.                             }

  1723.                         }

  1724.                     } else {

  1725.                         if (s->above_intra_ctx[col]) {

  1726.                             c = 2;

  1727.                         } else if (s->above_comp_ctx[col]) {

  1728.                             c = 4 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1729.                         } else {

  1730.                             c = 3 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1731.                         }

  1732.                     }

  1733.                 } else if (have_l) {

  1734.                     if (s->left_intra_ctx[row7]) {

  1735.                         c = 2;

  1736.                     } else if (s->left_comp_ctx[row7]) {

  1737.                         c = 4 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1738.                     } else {

  1739.                         c = 3 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1740.                     }

  1741.                 } else {

  1742.                     c = 2;

  1743.                 }

  1744.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.comp_ref[c]);

  1745.                 b->ref[var_idx] = s->varcompref[bit];

  1746.                 s->counts.comp_ref[c][bit]++;

  1747.             } else /* single reference */ {

  1748.                 int bit, c;

  1749. 

  1750.                 if (have_a && !s->above_intra_ctx[col]) {

  1751.                     if (have_l && !s->left_intra_ctx[row7]) {

  1752.                         if (s->left_comp_ctx[row7]) {

  1753.                             if (s->above_comp_ctx[col]) {

  1754.                                 c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7] ||

  1755.                                          !s->above_ref_ctx[col]);

  1756.                             } else {

  1757.                                 c = (3 * !s->above_ref_ctx[col]) +

  1758.                                     (!s->fixcompref || !s->left_ref_ctx[row7]);

  1759.                             }

  1760.                         } else if (s->above_comp_ctx[col]) {

  1761.                             c = (3 * !s->left_ref_ctx[row7]) +

  1762.                                 (!s->fixcompref || !s->above_ref_ctx[col]);

  1763.                         } else {

  1764.                             c = 2 * !s->left_ref_ctx[row7] + 2 * !s->above_ref_ctx[col];

  1765.                         }

  1766.                     } else if (s->above_intra_ctx[col]) {

  1767.                         c = 2;

  1768.                     } else if (s->above_comp_ctx[col]) {

  1769.                         c = 1 + (!s->fixcompref || !s->above_ref_ctx[col]);

  1770.                     } else {

  1771.                         c = 4 * (!s->above_ref_ctx[col]);

  1772.                     }

  1773.                 } else if (have_l && !s->left_intra_ctx[row7]) {

  1774.                     if (s->left_intra_ctx[row7]) {

  1775.                         c = 2;

  1776.                     } else if (s->left_comp_ctx[row7]) {

  1777.                         c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7]);

  1778.                     } else {

  1779.                         c = 4 * (!s->left_ref_ctx[row7]);

  1780.                     }

  1781.                 } else {

  1782.                     c = 2;

  1783.                 }

  1784.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][0]);

  1785.                 s->counts.single_ref[c][0][bit]++;

  1786.                 if (!bit) {

  1787.                     b->ref[0] = 0;

  1788.                 } else {

  1789.                     // FIXME can this codeblob be replaced by some sort of LUT?

  1790.                     if (have_a) {

  1791.                         if (have_l) {

  1792.                             if (s->left_intra_ctx[row7]) {

  1793.                                 if (s->above_intra_ctx[col]) {

  1794.                                     c = 2;

  1795.                                 } else if (s->above_comp_ctx[col]) {

  1796.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1797.                                                  s->above_ref_ctx[col] == 1);

  1798.                                 } else if (!s->above_ref_ctx[col]) {

  1799.                                     c = 3;

  1800.                                 } else {

  1801.                                     c = 4 * (s->above_ref_ctx[col] == 1);

  1802.                                 }

  1803.                             } else if (s->above_intra_ctx[col]) {

  1804.                                 if (s->left_intra_ctx[row7]) {

  1805.                                     c = 2;

  1806.                                 } else if (s->left_comp_ctx[row7]) {

  1807.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1808.                                                  s->left_ref_ctx[row7] == 1);

  1809.                                 } else if (!s->left_ref_ctx[row7]) {

  1810.                                     c = 3;

  1811.                                 } else {

  1812.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1813.                                 }

  1814.                             } else if (s->above_comp_ctx[col]) {

  1815.                                 if (s->left_comp_ctx[row7]) {

  1816.                                     if (s->left_ref_ctx[row7] == s->above_ref_ctx[col]) {

  1817.                                         c = 3 * (s->fixcompref == 1 ||

  1818.                                                  s->left_ref_ctx[row7] == 1);

  1819.                                     } else {

  1820.                                         c = 2;

  1821.                                     }

  1822.                                 } else if (!s->left_ref_ctx[row7]) {

  1823.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1824.                                                  s->above_ref_ctx[col] == 1);

  1825.                                 } else {

  1826.                                     c = 3 * (s->left_ref_ctx[row7] == 1) +

  1827.                                     (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1828.                                 }

  1829.                             } else if (s->left_comp_ctx[row7]) {

  1830.                                 if (!s->above_ref_ctx[col]) {

  1831.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1832.                                                  s->left_ref_ctx[row7] == 1);

  1833.                                 } else {

  1834.                                     c = 3 * (s->above_ref_ctx[col] == 1) +

  1835.                                     (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1836.                                 }

  1837.                             } else if (!s->above_ref_ctx[col]) {

  1838.                                 if (!s->left_ref_ctx[row7]) {

  1839.                                     c = 3;

  1840.                                 } else {

  1841.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1842.                                 }

  1843.                             } else if (!s->left_ref_ctx[row7]) {

  1844.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1845.                             } else {

  1846.                                 c = 2 * (s->left_ref_ctx[row7] == 1) +

  1847.                                 2 * (s->above_ref_ctx[col] == 1);

  1848.                             }

  1849.                         } else {

  1850.                             if (s->above_intra_ctx[col] ||

  1851.                                 (!s->above_comp_ctx[col] && !s->above_ref_ctx[col])) {

  1852.                                 c = 2;

  1853.                             } else if (s->above_comp_ctx[col]) {

  1854.                                 c = 3 * (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1855.                             } else {

  1856.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1857.                             }

  1858.                         }

  1859.                     } else if (have_l) {

  1860.                         if (s->left_intra_ctx[row7] ||

  1861.                             (!s->left_comp_ctx[row7] && !s->left_ref_ctx[row7])) {

  1862.                             c = 2;

  1863.                         } else if (s->left_comp_ctx[row7]) {

  1864.                             c = 3 * (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1865.                         } else {

  1866.                             c = 4 * (s->left_ref_ctx[row7] == 1);

  1867.                         }

  1868.                     } else {

  1869.                         c = 2;

  1870.                     }

  1871.                     bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][1]);

  1872.                     s->counts.single_ref[c][1][bit]++;

  1873.                     b->ref[0] = 1 + bit;

  1874.                 }

  1875.             }

  1876.         }

  1877. 

  1878.         if (b->bs <= BS_8x8) {

  1879.             if (s->segmentation.feat[b->seg_id].skip_enabled) {

  1880.                 b->mode[0] = b->mode[1] = b->mode[2] = b->mode[3] = ZEROMV;

  1881.             } else {

  1882.                 static const uint8_t off[10] = {

  1883.                     3, 0, 0, 1, 0, 0, 0, 0, 0, 0

  1884.                 };

  1885. 

  1886.                 // FIXME this needs to use the LUT tables from find_ref_mvs

  1887.                 // because not all are -1,0/0,-1

  1888.                 int c = inter_mode_ctx_lut[s->above_mode_ctx[col + off[b->bs]]]

  1889.                                           [s->left_mode_ctx[row7 + off[b->bs]]];

  1890. 

  1891.                 b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1892.                                               s->prob.p.mv_mode[c]);

  1893.                 b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1894.                 s->counts.mv_mode[c][b->mode[0] - 10]++;

  1895.             }

  1896.         }

  1897. 

  1898.         if (s->filtermode == FILTER_SWITCHABLE) {

  1899.             int c;

  1900. 

  1901.             if (have_a && s->above_mode_ctx[col] >= NEARESTMV) {

  1902.                 if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1903.                     c = s->above_filter_ctx[col] == s->left_filter_ctx[row7] ?

  1904.                         s->left_filter_ctx[row7] : 3;

  1905.                 } else {

  1906.                     c = s->above_filter_ctx[col];

  1907.                 }

  1908.             } else if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1909.                 c = s->left_filter_ctx[row7];

  1910.             } else {

  1911.                 c = 3;

  1912.             }

  1913. 

  1914.             filter_id = vp8_rac_get_tree(&s->c, vp9_filter_tree,

  1915.                                          s->prob.p.filter[c]);

  1916.             s->counts.filter[c][filter_id]++;

  1917.             b->filter = vp9_filter_lut[filter_id];

  1918.         } else {

  1919.             b->filter = s->filtermode;

  1920.         }

  1921. 

  1922.         if (b->bs > BS_8x8) {

  1923.             int c = inter_mode_ctx_lut[s->above_mode_ctx[col]][s->left_mode_ctx[row7]];

  1924. 

  1925.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1926.                                           s->prob.p.mv_mode[c]);

  1927.             s->counts.mv_mode[c][b->mode[0] - 10]++;

  1928.             fill_mv(s, b->mv[0], b->mode[0], 0);

  1929. 

  1930.             if (b->bs != BS_8x4) {

  1931.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1932.                                               s->prob.p.mv_mode[c]);

  1933.                 s->counts.mv_mode[c][b->mode[1] - 10]++;

  1934.                 fill_mv(s, b->mv[1], b->mode[1], 1);

  1935.             } else {

  1936.                 b->mode[1] = b->mode[0];

  1937.                 AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1938.                 AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1939.             }

  1940. 

  1941.             if (b->bs != BS_4x8) {

  1942.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1943.                                               s->prob.p.mv_mode[c]);

  1944.                 s->counts.mv_mode[c][b->mode[2] - 10]++;

  1945.                 fill_mv(s, b->mv[2], b->mode[2], 2);

  1946. 

  1947.                 if (b->bs != BS_8x4) {

  1948.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1949.                                                   s->prob.p.mv_mode[c]);

  1950.                     s->counts.mv_mode[c][b->mode[3] - 10]++;

  1951.                     fill_mv(s, b->mv[3], b->mode[3], 3);

  1952.                 } else {

  1953.                     b->mode[3] = b->mode[2];

  1954.                     AV_COPY32(&b->mv[3][0], &b->mv[2][0]);

  1955.                     AV_COPY32(&b->mv[3][1], &b->mv[2][1]);

  1956.                 }

  1957.             } else {

  1958.                 b->mode[2] = b->mode[0];

  1959.                 AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1960.                 AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1961.                 b->mode[3] = b->mode[1];

  1962.                 AV_COPY32(&b->mv[3][0], &b->mv[1][0]);

  1963.                 AV_COPY32(&b->mv[3][1], &b->mv[1][1]);

  1964.             }

  1965.         } else {

  1966.             fill_mv(s, b->mv[0], b->mode[0], -1);

  1967.             AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1968.             AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1969.             AV_COPY32(&b->mv[3][0], &b->mv[0][0]);

  1970.             AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1971.             AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1972.             AV_COPY32(&b->mv[3][1], &b->mv[0][1]);

  1973.         }

  1974. 

  1975.         vref = b->ref[b->comp ? s->signbias[s->varcompref[0]] : 0];

  1976.     }

  1977. 

  1978. #if HAVE_FAST_64BIT

  1979. #define SPLAT_CTX(var, val, n) \

  1980.     switch (n) { \

  1981.     case 1:  var = val;                                    break; \

  1982.     case 2:  AV_WN16A(&var, val *             0x0101);     break; \

  1983.     case 4:  AV_WN32A(&var, val *         0x01010101);     break; \

  1984.     case 8:  AV_WN64A(&var, val * 0x0101010101010101ULL);  break; \

  1985.     case 16: { \

  1986.         uint64_t v64 = val * 0x0101010101010101ULL; \

  1987.         AV_WN64A(              &var,     v64); \

  1988.         AV_WN64A(&((uint8_t *) &var)[8], v64); \

  1989.         break; \

  1990.     } \

  1991.     }

  1992. #else

  1993. #define SPLAT_CTX(var, val, n) \

  1994.     switch (n) { \

  1995.     case 1:  var = val;                         break; \

  1996.     case 2:  AV_WN16A(&var, val *     0x0101);  break; \

  1997.     case 4:  AV_WN32A(&var, val * 0x01010101);  break; \

  1998.     case 8: { \

  1999.         uint32_t v32 = val * 0x01010101; \

  2000.         AV_WN32A(              &var,     v32); \

  2001.         AV_WN32A(&((uint8_t *) &var)[4], v32); \

  2002.         break; \

  2003.     } \

  2004.     case 16: { \

  2005.         uint32_t v32 = val * 0x01010101; \

  2006.         AV_WN32A(              &var,      v32); \

  2007.         AV_WN32A(&((uint8_t *) &var)[4],  v32); \

  2008.         AV_WN32A(&((uint8_t *) &var)[8],  v32); \

  2009.         AV_WN32A(&((uint8_t *) &var)[12], v32); \

  2010.         break; \

  2011.     } \

  2012.     }

  2013. #endif

  2014. 

  2015.     switch (bwh_tab[1][b->bs][0]) {

  2016. #define SET_CTXS(dir, off, n) \

  2017.     do { \

  2018.         SPLAT_CTX(s->dir##_skip_ctx[off],      b->skip,          n); \

  2019.         SPLAT_CTX(s->dir##_txfm_ctx[off],      b->tx,            n); \

  2020.         SPLAT_CTX(s->dir##_partition_ctx[off], dir##_ctx[b->bs], n); \

  2021.         if (!s->keyframe && !s->intraonly) { \

  2022.             SPLAT_CTX(s->dir##_intra_ctx[off], b->intra,   n); \

  2023.             SPLAT_CTX(s->dir##_comp_ctx[off],  b->comp,    n); \

  2024.             SPLAT_CTX(s->dir##_mode_ctx[off],  b->mode[3], n); \

  2025.             if (!b->intra) { \

  2026.                 SPLAT_CTX(s->dir##_ref_ctx[off], vref, n); \

  2027.                 if (s->filtermode == FILTER_SWITCHABLE) { \

  2028.                     SPLAT_CTX(s->dir##_filter_ctx[off], filter_id, n); \

  2029.                 } \

  2030.             } \

  2031.         } \

  2032.     } while (0)

  2033.     case 1: SET_CTXS(above, col, 1); break;

  2034.     case 2: SET_CTXS(above, col, 2); break;

  2035.     case 4: SET_CTXS(above, col, 4); break;

  2036.     case 8: SET_CTXS(above, col, 8); break;

  2037.     }

  2038.     switch (bwh_tab[1][b->bs][1]) {

  2039.     case 1: SET_CTXS(left, row7, 1); break;

  2040.     case 2: SET_CTXS(left, row7, 2); break;

  2041.     case 4: SET_CTXS(left, row7, 4); break;

  2042.     case 8: SET_CTXS(left, row7, 8); break;

  2043.     }

  2044. #undef SPLAT_CTX

  2045. #undef SET_CTXS

  2046. 

  2047.     if (!s->keyframe && !s->intraonly) {

  2048.         if (b->bs > BS_8x8) {

  2049.             int mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  2050. 

  2051.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][0], &b->mv[1][0]);

  2052.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][1], &b->mv[1][1]);

  2053.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][0], mv0);

  2054.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][1], mv1);

  2055.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][0], &b->mv[2][0]);

  2056.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][1], &b->mv[2][1]);

  2057.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][0], mv0);

  2058.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][1], mv1);

  2059.         } else {

  2060.             int n, mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  2061. 

  2062.             for (n = 0; n < w4 * 2; n++) {

  2063.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][0], mv0);

  2064.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][1], mv1);

  2065.             }

  2066.             for (n = 0; n < h4 * 2; n++) {

  2067.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][0], mv0);

  2068.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][1], mv1);

  2069.             }

  2070.         }

  2071.     }

  2072. 

  2073.     // FIXME kinda ugly

  2074.     for (y = 0; y < h4; y++) {

  2075.         int x, o = (row + y) * s->sb_cols * 8 + col;

  2076.         struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[o];

  2077. 

  2078.         if (b->intra) {

  2079.             for (x = 0; x < w4; x++) {

  2080.                 mv[x].ref[0] =

  2081.                 mv[x].ref[1] = -1;

  2082.             }

  2083.         } else if (b->comp) {

  2084.             for (x = 0; x < w4; x++) {

  2085.                 mv[x].ref[0] = b->ref[0];

  2086.                 mv[x].ref[1] = b->ref[1];

  2087.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  2088.                 AV_COPY32(&mv[x].mv[1], &b->mv[3][1]);

  2089.             }

  2090.         } else {

  2091.             for (x = 0; x < w4; x++) {

  2092.                 mv[x].ref[0] = b->ref[0];

  2093.                 mv[x].ref[1] = -1;

  2094.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  2095.             }

  2096.         }

  2097.     }

  2098. }

  2099. 

  2100. // FIXME merge cnt/eob arguments?

  2101. static av_always_inline int

  2102. decode_coeffs_b_generic(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  2103.                         int is_tx32x32, unsigned (*cnt)[6][3],

  2104.                         unsigned (*eob)[6][2], uint8_t (*p)[6][11],

  2105.                         int nnz, const int16_t *scan, const int16_t (*nb)[2],

  2106.                         const int16_t *band_counts, const int16_t *qmul)

  2107. {

  2108.     int i = 0, band = 0, band_left = band_counts[band];

  2109.     uint8_t *tp = p[0][nnz];

  2110.     uint8_t cache[1024];

  2111. 

  2112.     do {

  2113.         int val, rc;

  2114. 

  2115.         val = vp56_rac_get_prob_branchy(c, tp[0]); // eob

  2116.         eob[band][nnz][val]++;

  2117.         if (!val)

  2118.             break;

  2119. 

  2120.     skip_eob:

  2121.         if (!vp56_rac_get_prob_branchy(c, tp[1])) { // zero

  2122.             cnt[band][nnz][0]++;

  2123.             if (!--band_left)

  2124.                 band_left = band_counts[++band];

  2125.             cache[scan[i]] = 0;

  2126.             nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  2127.             tp = p[band][nnz];

  2128.             if (++i == n_coeffs)

  2129.                 break; //invalid input; blocks should end with EOB

  2130.             goto skip_eob;

  2131.         }

  2132. 

  2133.         rc = scan[i];

  2134.         if (!vp56_rac_get_prob_branchy(c, tp[2])) { // one

  2135.             cnt[band][nnz][1]++;

  2136.             val = 1;

  2137.             cache[rc] = 1;

  2138.         } else {

  2139.             // fill in p[3-10] (model fill) - only once per frame for each pos

  2140.             if (!tp[3])

  2141.                 memcpy(&tp[3], vp9_model_pareto8[tp[2]], 8);

  2142. 

  2143.             cnt[band][nnz][2]++;

  2144.             if (!vp56_rac_get_prob_branchy(c, tp[3])) { // 2, 3, 4

  2145.                 if (!vp56_rac_get_prob_branchy(c, tp[4])) {

  2146.                     cache[rc] = val = 2;

  2147.                 } else {

  2148.                     val = 3 + vp56_rac_get_prob(c, tp[5]);

  2149.                     cache[rc] = 3;

  2150.                 }

  2151.             } else if (!vp56_rac_get_prob_branchy(c, tp[6])) { // cat1/2

  2152.                 cache[rc] = 4;

  2153.                 if (!vp56_rac_get_prob_branchy(c, tp[7])) {

  2154.                     val = 5 + vp56_rac_get_prob(c, 159);

  2155.                 } else {

  2156.                     val  = 7 + (vp56_rac_get_prob(c, 165) << 1);

  2157.                     val +=      vp56_rac_get_prob(c, 145);

  2158.                 }

  2159.             } else { // cat 3-6

  2160.                 cache[rc] = 5;

  2161.                 if (!vp56_rac_get_prob_branchy(c, tp[8])) {

  2162.                     if (!vp56_rac_get_prob_branchy(c, tp[9])) {

  2163.                         val  = 11 + (vp56_rac_get_prob(c, 173) << 2);

  2164.                         val +=      (vp56_rac_get_prob(c, 148) << 1);

  2165.                         val +=       vp56_rac_get_prob(c, 140);

  2166.                     } else {

  2167.                         val  = 19 + (vp56_rac_get_prob(c, 176) << 3);

  2168.                         val +=      (vp56_rac_get_prob(c, 155) << 2);

  2169.                         val +=      (vp56_rac_get_prob(c, 140) << 1);

  2170.                         val +=       vp56_rac_get_prob(c, 135);

  2171.                     }

  2172.                 } else if (!vp56_rac_get_prob_branchy(c, tp[10])) {

  2173.                     val  = 35 + (vp56_rac_get_prob(c, 180) << 4);

  2174.                     val +=      (vp56_rac_get_prob(c, 157) << 3);

  2175.                     val +=      (vp56_rac_get_prob(c, 141) << 2);

  2176.                     val +=      (vp56_rac_get_prob(c, 134) << 1);

  2177.                     val +=       vp56_rac_get_prob(c, 130);

  2178.                 } else {

  2179.                     val  = 67 + (vp56_rac_get_prob(c, 254) << 13);

  2180.                     val +=      (vp56_rac_get_prob(c, 254) << 12);

  2181.                     val +=      (vp56_rac_get_prob(c, 254) << 11);

  2182.                     val +=      (vp56_rac_get_prob(c, 252) << 10);

  2183.                     val +=      (vp56_rac_get_prob(c, 249) << 9);

  2184.                     val +=      (vp56_rac_get_prob(c, 243) << 8);

  2185.                     val +=      (vp56_rac_get_prob(c, 230) << 7);

  2186.                     val +=      (vp56_rac_get_prob(c, 196) << 6);

  2187.                     val +=      (vp56_rac_get_prob(c, 177) << 5);

  2188.                     val +=      (vp56_rac_get_prob(c, 153) << 4);

  2189.                     val +=      (vp56_rac_get_prob(c, 140) << 3);

  2190.                     val +=      (vp56_rac_get_prob(c, 133) << 2);

  2191.                     val +=      (vp56_rac_get_prob(c, 130) << 1);

  2192.                     val +=       vp56_rac_get_prob(c, 129);

  2193.                 }

  2194.             }

  2195.         }

  2196.         if (!--band_left)

  2197.             band_left = band_counts[++band];

  2198.         if (is_tx32x32)

  2199.             coef[rc] = ((vp8_rac_get(c) ? -val : val) * qmul[!!i]) / 2;

  2200.         else

  2201.             coef[rc] = (vp8_rac_get(c) ? -val : val) * qmul[!!i];

  2202.         nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  2203.         tp = p[band][nnz];

  2204.     } while (++i < n_coeffs);

  2205. 

  2206.     return i;

  2207. }

  2208. 

  2209. static int decode_coeffs_b(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  2210.                            unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  2211.                            uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  2212.                            const int16_t (*nb)[2], const int16_t *band_counts,

  2213.                            const int16_t *qmul)

  2214. {

  2215.     return decode_coeffs_b_generic(c, coef, n_coeffs, 0, cnt, eob, p,

  2216.                                    nnz, scan, nb, band_counts, qmul);

  2217. }

  2218. 

  2219. static int decode_coeffs_b32(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  2220.                              unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  2221.                              uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  2222.                              const int16_t (*nb)[2], const int16_t *band_counts,

  2223.                              const int16_t *qmul)

  2224. {

  2225.     return decode_coeffs_b_generic(c, coef, n_coeffs, 1, cnt, eob, p,

  2226.                                    nnz, scan, nb, band_counts, qmul);

  2227. }

  2228. 

  2229. static void decode_coeffs(AVCodecContext *ctx)

  2230. {

  2231.     VP9Context *s = ctx->priv_data;

  2232.     VP9Block *b = s->b;

  2233.     int row = s->row, col = s->col;

  2234.     uint8_t (*p)[6][11] = s->prob.coef[b->tx][0 /* y */][!b->intra];

  2235.     unsigned (*c)[6][3] = s->counts.coef[b->tx][0 /* y */][!b->intra];

  2236.     unsigned (*e)[6][2] = s->counts.eob[b->tx][0 /* y */][!b->intra];

  2237.     int w4 = bwh_tab[1][b->bs][0] << 1, h4 = bwh_tab[1][b->bs][1] << 1;

  2238.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2239.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2240.     int n, pl, x, y, res;

  2241.     int16_t (*qmul)[2] = s->segmentation.feat[b->seg_id].qmul;

  2242.     int tx = 4 * s->lossless + b->tx;

  2243.     const int16_t * const *yscans = vp9_scans[tx];

  2244.     const int16_t (* const *ynbs)[2] = vp9_scans_nb[tx];

  2245.     const int16_t *uvscan = vp9_scans[b->uvtx][DCT_DCT];

  2246.     const int16_t (*uvnb)[2] = vp9_scans_nb[b->uvtx][DCT_DCT];

  2247.     uint8_t *a = &s->above_y_nnz_ctx[col * 2];

  2248.     uint8_t *l = &s->left_y_nnz_ctx[(row & 7) << 1];

  2249.     static const int16_t band_counts[4][8] = {

  2250.         { 1, 2, 3, 4,  3,   16 - 13 },

  2251.         { 1, 2, 3, 4, 11,   64 - 21 },

  2252.         { 1, 2, 3, 4, 11,  256 - 21 },

  2253.         { 1, 2, 3, 4, 11, 1024 - 21 },

  2254.     };

  2255.     const int16_t *y_band_counts = band_counts[b->tx];

  2256.     const int16_t *uv_band_counts = band_counts[b->uvtx];

  2257. 

  2258. #define MERGE(la, end, step, rd) \

  2259.     for (n = 0; n < end; n += step) \

  2260.         la[n] = !!rd(&la[n])

  2261. #define MERGE_CTX(step, rd) \

  2262.     do { \

  2263.         MERGE(l, end_y, step, rd); \

  2264.         MERGE(a, end_x, step, rd); \

  2265.     } while (0)

  2266. 

  2267. #define DECODE_Y_COEF_LOOP(step, mode_index, v) \

  2268.     for (n = 0, y = 0; y < end_y; y += step) { \

  2269.         for (x = 0; x < end_x; x += step, n += step * step) { \

  2270.             enum TxfmType txtp = vp9_intra_txfm_type[b->mode[mode_index]]; \

  2271.             res = decode_coeffs_b##v(&s->c, s->block + 16 * n, 16 * step * step, \

  2272.                                      c, e, p, a[x] + l[y], yscans[txtp], \

  2273.                                      ynbs[txtp], y_band_counts, qmul[0]); \

  2274.             a[x] = l[y] = !!res; \

  2275.             if (step >= 4) { \

  2276.                 AV_WN16A(&s->eob[n], res); \

  2277.             } else { \

  2278.                 s->eob[n] = res; \

  2279.             } \

  2280.         } \

  2281.     }

  2282. 

  2283. #define SPLAT(la, end, step, cond) \

  2284.     if (step == 2) { \

  2285.         for (n = 1; n < end; n += step) \

  2286.             la[n] = la[n - 1]; \

  2287.     } else if (step == 4) { \

  2288.         if (cond) { \

  2289.             for (n = 0; n < end; n += step) \

  2290.                 AV_WN32A(&la[n], la[n] * 0x01010101); \

  2291.         } else { \

  2292.             for (n = 0; n < end; n += step) \

  2293.                 memset(&la[n + 1], la[n], FFMIN(end - n - 1, 3)); \

  2294.         } \

  2295.     } else /* step == 8 */ { \

  2296.         if (cond) { \

  2297.             if (HAVE_FAST_64BIT) { \

  2298.                 for (n = 0; n < end; n += step) \

  2299.                     AV_WN64A(&la[n], la[n] * 0x0101010101010101ULL); \

  2300.             } else { \

  2301.                 for (n = 0; n < end; n += step) { \

  2302.                     uint32_t v32 = la[n] * 0x01010101; \

  2303.                     AV_WN32A(&la[n],     v32); \

  2304.                     AV_WN32A(&la[n + 4], v32); \

  2305.                 } \

  2306.             } \

  2307.         } else { \

  2308.             for (n = 0; n < end; n += step) \

  2309.                 memset(&la[n + 1], la[n], FFMIN(end - n - 1, 7)); \

  2310.         } \

  2311.     }

  2312. #define SPLAT_CTX(step) \

  2313.     do { \

  2314.         SPLAT(a, end_x, step, end_x == w4); \

  2315.         SPLAT(l, end_y, step, end_y == h4); \

  2316.     } while (0)

  2317. 

  2318.     /* y tokens */

  2319.     switch (b->tx) {

  2320.     case TX_4X4:

  2321.         DECODE_Y_COEF_LOOP(1, b->bs > BS_8x8 ? n : 0,);

  2322.         break;

  2323.     case TX_8X8:

  2324.         MERGE_CTX(2, AV_RN16A);

  2325.         DECODE_Y_COEF_LOOP(2, 0,);

  2326.         SPLAT_CTX(2);

  2327.         break;

  2328.     case TX_16X16:

  2329.         MERGE_CTX(4, AV_RN32A);

  2330.         DECODE_Y_COEF_LOOP(4, 0,);

  2331.         SPLAT_CTX(4);

  2332.         break;

  2333.     case TX_32X32:

  2334.         MERGE_CTX(8, AV_RN64A);

  2335.         DECODE_Y_COEF_LOOP(8, 0, 32);

  2336.         SPLAT_CTX(8);

  2337.         break;

  2338.     }

  2339. 

  2340. #define DECODE_UV_COEF_LOOP(step, decode_coeffs_fn) \

  2341.     for (n = 0, y = 0; y < end_y; y += step) { \

  2342.         for (x = 0; x < end_x; x += step, n += step * step) { \

  2343.             res = decode_coeffs_fn(&s->c, s->uvblock[pl] + 16 * n, \

  2344.                                    16 * step * step, c, e, p, a[x] + l[y], \

  2345.                                    uvscan, uvnb, uv_band_counts, qmul[1]); \

  2346.             a[x] = l[y] = !!res; \

  2347.             if (step >= 4) { \

  2348.                 AV_WN16A(&s->uveob[pl][n], res); \

  2349.             } else { \

  2350.                 s->uveob[pl][n] = res; \

  2351.             } \

  2352.         } \

  2353.     }

  2354. 

  2355.     p = s->prob.coef[b->uvtx][1 /* uv */][!b->intra];

  2356.     c = s->counts.coef[b->uvtx][1 /* uv */][!b->intra];

  2357.     e = s->counts.eob[b->uvtx][1 /* uv */][!b->intra];

  2358.     w4 >>= s->ss_h;

  2359.     end_x >>= s->ss_h;

  2360.     h4 >>= s->ss_v;

  2361.     end_y >>= s->ss_v;

  2362.     for (pl = 0; pl < 2; pl++) {

  2363.         a = &s->above_uv_nnz_ctx[pl][col << !s->ss_h];

  2364.         l = &s->left_uv_nnz_ctx[pl][(row & 7) << !s->ss_v];

  2365.         switch (b->uvtx) {

  2366.         case TX_4X4:

  2367.             DECODE_UV_COEF_LOOP(1, decode_coeffs_b);

  2368.             break;

  2369.         case TX_8X8:

  2370.             MERGE_CTX(2, AV_RN16A);

  2371.             DECODE_UV_COEF_LOOP(2, decode_coeffs_b);

  2372.             SPLAT_CTX(2);

  2373.             break;

  2374.         case TX_16X16:

  2375.             MERGE_CTX(4, AV_RN32A);

  2376.             DECODE_UV_COEF_LOOP(4, decode_coeffs_b);

  2377.             SPLAT_CTX(4);

  2378.             break;

  2379.         case TX_32X32:

  2380.             MERGE_CTX(8, AV_RN64A);

  2381.             DECODE_UV_COEF_LOOP(8, decode_coeffs_b32);

  2382.             SPLAT_CTX(8);

  2383.             break;

  2384.         }

  2385.     }

  2386. }

  2387. 

  2388. static av_always_inline int check_intra_mode(VP9Context *s, int mode, uint8_t **a,

  2389.                                              uint8_t *dst_edge, ptrdiff_t stride_edge,

  2390.                                              uint8_t *dst_inner, ptrdiff_t stride_inner,

  2391.                                              uint8_t *l, int col, int x, int w,

  2392.                                              int row, int y, enum TxfmMode tx,

  2393.                                              int p, int ss_h, int ss_v)

  2394. {

  2395.     int have_top = row > 0 || y > 0;

  2396.     int have_left = col > s->tiling.tile_col_start || x > 0;

  2397.     int have_right = x < w - 1;

  2398.     static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {

  2399.         [VERT_PRED]            = { { DC_127_PRED,          VERT_PRED },

  2400.                                    { DC_127_PRED,          VERT_PRED } },

  2401.         [HOR_PRED]             = { { DC_129_PRED,          DC_129_PRED },

  2402.                                    { HOR_PRED,             HOR_PRED } },

  2403.         [DC_PRED]              = { { DC_128_PRED,          TOP_DC_PRED },

  2404.                                    { LEFT_DC_PRED,         DC_PRED } },

  2405.         [DIAG_DOWN_LEFT_PRED]  = { { DC_127_PRED,          DIAG_DOWN_LEFT_PRED },

  2406.                                    { DC_127_PRED,          DIAG_DOWN_LEFT_PRED } },

  2407.         [DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },

  2408.                                    { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },

  2409.         [VERT_RIGHT_PRED]      = { { VERT_RIGHT_PRED,      VERT_RIGHT_PRED },

  2410.                                    { VERT_RIGHT_PRED,      VERT_RIGHT_PRED } },

  2411.         [HOR_DOWN_PRED]        = { { HOR_DOWN_PRED,        HOR_DOWN_PRED },

  2412.                                    { HOR_DOWN_PRED,        HOR_DOWN_PRED } },

  2413.         [VERT_LEFT_PRED]       = { { DC_127_PRED,          VERT_LEFT_PRED },

  2414.                                    { DC_127_PRED,          VERT_LEFT_PRED } },

  2415.         [HOR_UP_PRED]          = { { DC_129_PRED,          DC_129_PRED },

  2416.                                    { HOR_UP_PRED,          HOR_UP_PRED } },

  2417.         [TM_VP8_PRED]          = { { DC_129_PRED,          VERT_PRED },

  2418.                                    { HOR_PRED,             TM_VP8_PRED } },

  2419.     };

  2420.     static const struct {

  2421.         uint8_t needs_left:1;

  2422.         uint8_t needs_top:1;

  2423.         uint8_t needs_topleft:1;

  2424.         uint8_t needs_topright:1;

  2425.         uint8_t invert_left:1;

  2426.     } edges[N_INTRA_PRED_MODES] = {

  2427.         [VERT_PRED]            = { .needs_top  = 1 },

  2428.         [HOR_PRED]             = { .needs_left = 1 },

  2429.         [DC_PRED]              = { .needs_top  = 1, .needs_left = 1 },

  2430.         [DIAG_DOWN_LEFT_PRED]  = { .needs_top  = 1, .needs_topright = 1 },

  2431.         [DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2432.         [VERT_RIGHT_PRED]      = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2433.         [HOR_DOWN_PRED]        = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2434.         [VERT_LEFT_PRED]       = { .needs_top  = 1, .needs_topright = 1 },

  2435.         [HOR_UP_PRED]          = { .needs_left = 1, .invert_left = 1 },

  2436.         [TM_VP8_PRED]          = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2437.         [LEFT_DC_PRED]         = { .needs_left = 1 },

  2438.         [TOP_DC_PRED]          = { .needs_top  = 1 },

  2439.         [DC_128_PRED]          = { 0 },

  2440.         [DC_127_PRED]          = { 0 },

  2441.         [DC_129_PRED]          = { 0 }

  2442.     };

  2443. 

  2444.     av_assert2(mode >= 0 && mode < 10);

  2445.     mode = mode_conv[mode][have_left][have_top];

  2446.     if (edges[mode].needs_top) {

  2447.         uint8_t *top, *topleft;

  2448.         int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !ss_h) - x) * 4;

  2449.         int n_px_need_tr = 0;

  2450. 

  2451.         if (tx == TX_4X4 && edges[mode].needs_topright && have_right)

  2452.             n_px_need_tr = 4;

  2453. 

  2454.         // if top of sb64-row, use s->intra_pred_data[] instead of

  2455.         // dst[-stride] for intra prediction (it contains pre- instead of

  2456.         // post-loopfilter data)

  2457.         if (have_top) {

  2458.             top = !(row & 7) && !y ?

  2459.                 s->intra_pred_data[p] + col * (8 >> ss_h) + x * 4 :

  2460.                 y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];

  2461.             if (have_left)

  2462.                 topleft = !(row & 7) && !y ?

  2463.                     s->intra_pred_data[p] + col * (8 >> ss_h) + x * 4 :

  2464.                     y == 0 || x == 0 ? &dst_edge[-stride_edge] :

  2465.                     &dst_inner[-stride_inner];

  2466.         }

  2467. 

  2468.         if (have_top &&

  2469.             (!edges[mode].needs_topleft || (have_left && top == topleft)) &&

  2470.             (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&

  2471.             n_px_need + n_px_need_tr <= n_px_have) {

  2472.             *a = top;

  2473.         } else {

  2474.             if (have_top) {

  2475.                 if (n_px_need <= n_px_have) {

  2476.                     memcpy(*a, top, n_px_need);

  2477.                 } else {

  2478.                     memcpy(*a, top, n_px_have);

  2479.                     memset(&(*a)[n_px_have], (*a)[n_px_have - 1],

  2480.                            n_px_need - n_px_have);

  2481.                 }

  2482.             } else {

  2483.                 memset(*a, 127, n_px_need);

  2484.             }

  2485.             if (edges[mode].needs_topleft) {

  2486.                 if (have_left && have_top) {

  2487.                     (*a)[-1] = topleft[-1];

  2488.                 } else {

  2489.                     (*a)[-1] = have_top ? 129 : 127;

  2490.                 }

  2491.             }

  2492.             if (tx == TX_4X4 && edges[mode].needs_topright) {

  2493.                 if (have_top && have_right &&

  2494.                     n_px_need + n_px_need_tr <= n_px_have) {

  2495.                     memcpy(&(*a)[4], &top[4], 4);

  2496.                 } else {

  2497.                     memset(&(*a)[4], (*a)[3], 4);

  2498.                 }

  2499.             }

  2500.         }

  2501.     }

  2502.     if (edges[mode].needs_left) {

  2503.         if (have_left) {

  2504.             int n_px_need = 4 << tx, i, n_px_have = (((s->rows - row) << !ss_v) - y) * 4;

  2505.             uint8_t *dst = x == 0 ? dst_edge : dst_inner;

  2506.             ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;

  2507. 

  2508.             if (edges[mode].invert_left) {

  2509.                 if (n_px_need <= n_px_have) {

  2510.                     for (i = 0; i < n_px_need; i++)

  2511.                         l[i] = dst[i * stride - 1];

  2512.                 } else {

  2513.                     for (i = 0; i < n_px_have; i++)

  2514.                         l[i] = dst[i * stride - 1];

  2515.                     memset(&l[n_px_have], l[n_px_have - 1], n_px_need - n_px_have);

  2516.                 }

  2517.             } else {

  2518.                 if (n_px_need <= n_px_have) {

  2519.                     for (i = 0; i < n_px_need; i++)

  2520.                         l[n_px_need - 1 - i] = dst[i * stride - 1];

  2521.                 } else {

  2522.                     for (i = 0; i < n_px_have; i++)

  2523.                         l[n_px_need - 1 - i] = dst[i * stride - 1];

  2524.                     memset(l, l[n_px_need - n_px_have], n_px_need - n_px_have);

  2525.                 }

  2526.             }

  2527.         } else {

  2528.             memset(l, 129, 4 << tx);

  2529.         }

  2530.     }

  2531. 

  2532.     return mode;

  2533. }

  2534. 

  2535. static void intra_recon(AVCodecContext *ctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  2536. {

  2537.     VP9Context *s = ctx->priv_data;

  2538.     VP9Block *b = s->b;

  2539.     int row = s->row, col = s->col;

  2540.     int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2541.     int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2542.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2543.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2544.     int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2545.     int uvstep1d = 1 << b->uvtx, p;

  2546.     uint8_t *dst = s->dst[0], *dst_r = s->frames[CUR_FRAME].tf.f->data[0] + y_off;

  2547.     LOCAL_ALIGNED_32(uint8_t, a_buf, [64]);

  2548.     LOCAL_ALIGNED_32(uint8_t, l, [32]);

  2549. 

  2550.     for (n = 0, y = 0; y < end_y; y += step1d) {

  2551.         uint8_t *ptr = dst, *ptr_r = dst_r;

  2552.         for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d,

  2553.                                ptr_r += 4 * step1d, n += step) {

  2554.             int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?

  2555.                                y * 2 + x : 0];

  2556.             uint8_t *a = &a_buf[32];

  2557.             enum TxfmType txtp = vp9_intra_txfm_type[mode];

  2558.             int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2559. 

  2560.             mode = check_intra_mode(s, mode, &a, ptr_r,

  2561.                                     s->frames[CUR_FRAME].tf.f->linesize[0],

  2562.                                     ptr, s->y_stride, l,

  2563.                                     col, x, w4, row, y, b->tx, 0, 0, 0);

  2564.             s->dsp.intra_pred[b->tx][mode](ptr, s->y_stride, l, a);

  2565.             if (eob)

  2566.                 s->dsp.itxfm_add[tx][txtp](ptr, s->y_stride,

  2567.                                            s->block + 16 * n, eob);

  2568.         }

  2569.         dst_r += 4 * step1d * s->frames[CUR_FRAME].tf.f->linesize[0];

  2570.         dst   += 4 * step1d * s->y_stride;

  2571.     }

  2572. 

  2573.     // U/V

  2574.     w4 >>= s->ss_h;

  2575.     end_x >>= s->ss_h;

  2576.     end_y >>= s->ss_v;

  2577.     step = 1 << (b->uvtx * 2);

  2578.     for (p = 0; p < 2; p++) {

  2579.         dst   = s->dst[1 + p];

  2580.         dst_r = s->frames[CUR_FRAME].tf.f->data[1 + p] + uv_off;

  2581.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2582.             uint8_t *ptr = dst, *ptr_r = dst_r;

  2583.             for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d,

  2584.                                    ptr_r += 4 * uvstep1d, n += step) {

  2585.                 int mode = b->uvmode;

  2586.                 uint8_t *a = &a_buf[32];

  2587.                 int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2588. 

  2589.                 mode = check_intra_mode(s, mode, &a, ptr_r,

  2590.                                         s->frames[CUR_FRAME].tf.f->linesize[1],

  2591.                                         ptr, s->uv_stride, l, col, x, w4, row, y,

  2592.                                         b->uvtx, p + 1, s->ss_h, s->ss_v);

  2593.                 s->dsp.intra_pred[b->uvtx][mode](ptr, s->uv_stride, l, a);

  2594.                 if (eob)

  2595.                     s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2596.                                                     s->uvblock[p] + 16 * n, eob);

  2597.             }

  2598.             dst_r += 4 * uvstep1d * s->frames[CUR_FRAME].tf.f->linesize[1];

  2599.             dst   += 4 * uvstep1d * s->uv_stride;

  2600.         }

  2601.     }

  2602. }

  2603. 

  2604. static av_always_inline void mc_luma_scaled(VP9Context *s, vp9_scaled_mc_func smc,

  2605.                                             uint8_t *dst, ptrdiff_t dst_stride,

  2606.                                             const uint8_t *ref, ptrdiff_t ref_stride,

  2607.                                             ThreadFrame *ref_frame,

  2608.                                             ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2609.                                             int bw, int bh, int w, int h,

  2610.                                             const uint16_t *scale, const uint8_t *step)

  2611. {

  2612. #define scale_mv(n, dim) (((int64_t)(n) * scale[dim]) >> 14)

  2613.     // BUG libvpx seems to scale the two components separately. This introduces

  2614.     // rounding errors but we have to reproduce them to be exactly compatible

  2615.     // with the output from libvpx...

  2616.     int mx = scale_mv(mv->x * 2, 0) + scale_mv(x * 16, 0);

  2617.     int my = scale_mv(mv->y * 2, 1) + scale_mv(y * 16, 1);

  2618.     int refbw_m1, refbh_m1;

  2619.     int th;

  2620. 

  2621.     y = my >> 4;

  2622.     x = mx >> 4;

  2623.     ref += y * ref_stride + x;

  2624.     mx &= 15;

  2625.     my &= 15;

  2626.     refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;

  2627.     refbh_m1 = ((bh - 1) * step[1] + my) >> 4;

  2628.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2629.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2630.     // the longest loopfilter of the next sbrow

  2631.     th = (y + refbh_m1 + 4 + 7) >> 6;

  2632.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2633.     if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 4 >= h - refbh_m1) {

  2634.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2635.                                  ref - 3 * ref_stride - 3,

  2636.                                  144, ref_stride,

  2637.                                  refbw_m1 + 8, refbh_m1 + 8,

  2638.                                  x - 3, y - 3, w, h);

  2639.         ref = s->edge_emu_buffer + 3 * 144 + 3;

  2640.         ref_stride = 144;

  2641.     }

  2642.     smc(dst, dst_stride, ref, ref_stride, bh, mx, my, step[0], step[1]);

  2643. }

  2644. 

  2645. static av_always_inline void mc_chroma_scaled(VP9Context *s, vp9_scaled_mc_func smc,

  2646.                                               uint8_t *dst_u, uint8_t *dst_v,

  2647.                                               ptrdiff_t dst_stride,

  2648.                                               const uint8_t *ref_u, ptrdiff_t src_stride_u,

  2649.                                               const uint8_t *ref_v, ptrdiff_t src_stride_v,

  2650.                                               ThreadFrame *ref_frame,

  2651.                                               ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2652.                                               int bw, int bh, int w, int h,

  2653.                                               const uint16_t *scale, const uint8_t *step)

  2654. {

  2655.     // BUG https://code.google.com/p/webm/issues/detail?id=820

  2656.     int mx = scale_mv(mv->x << !s->ss_h, 0) + (scale_mv(x * 16, 0) & ~15) + (scale_mv(x * 32, 0) & 15);

  2657.     int my = scale_mv(mv->y << !s->ss_v, 1) + (scale_mv(y * 16, 1) & ~15) + (scale_mv(y * 32, 1) & 15);

  2658. #undef scale_mv

  2659.     int refbw_m1, refbh_m1;

  2660.     int th;

  2661. 

  2662.     y = my >> 4;

  2663.     x = mx >> 4;

  2664.     ref_u += y * src_stride_u + x;

  2665.     ref_v += y * src_stride_v + x;

  2666.     mx &= 15;

  2667.     my &= 15;

  2668.     refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;

  2669.     refbh_m1 = ((bh - 1) * step[1] + my) >> 4;

  2670.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2671.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2672.     // the longest loopfilter of the next sbrow

  2673.     th = (y + refbh_m1 + 4 + 7) >> (6 - s->ss_v);

  2674.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2675.     if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 4 >= h - refbh_m1) {

  2676.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2677.                                  ref_u - 3 * src_stride_u - 3,

  2678.                                  144, src_stride_u,

  2679.                                  refbw_m1 + 8, refbh_m1 + 8,

  2680.                                  x - 3, y - 3, w, h);

  2681.         ref_u = s->edge_emu_buffer + 3 * 144 + 3;

  2682.         smc(dst_u, dst_stride, ref_u, 144, bh, mx, my, step[0], step[1]);

  2683. 

  2684.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2685.                                  ref_v - 3 * src_stride_v - 3,

  2686.                                  144, src_stride_v,

  2687.                                  refbw_m1 + 8, refbh_m1 + 8,

  2688.                                  x - 3, y - 3, w, h);

  2689.         ref_v = s->edge_emu_buffer + 3 * 144 + 3;

  2690.         smc(dst_v, dst_stride, ref_v, 144, bh, mx, my, step[0], step[1]);

  2691.     } else {

  2692.         smc(dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my, step[0], step[1]);

  2693.         smc(dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my, step[0], step[1]);

  2694.     }

  2695. }

  2696. 

  2697. #define FN(x) x##_scaled

  2698. #define mc_luma_dir(s, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, bw, bh, w, h, i) \

  2699.     mc_luma_scaled(s, s->dsp.s##mc, dst, dst_ls, src, src_ls, tref, row, col, \

  2700.                    mv, bw, bh, w, h, s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])

  2701. #define mc_chroma_dir(s, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2702.                       row, col, mv, bw, bh, w, h, i) \

  2703.     mc_chroma_scaled(s, s->dsp.s##mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2704.                      row, col, mv, bw, bh, w, h, s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])

  2705. #include "vp9_mc_template.c"

  2706. #undef mc_luma_dir

  2707. #undef mc_chroma_dir

  2708. #undef FN

  2709. 

  2710. static av_always_inline void mc_luma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2],

  2711.                                               uint8_t *dst, ptrdiff_t dst_stride,

  2712.                                               const uint8_t *ref, ptrdiff_t ref_stride,

  2713.                                               ThreadFrame *ref_frame,

  2714.                                               ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2715.                                               int bw, int bh, int w, int h)

  2716. {

  2717.     int mx = mv->x, my = mv->y, th;

  2718. 

  2719.     y += my >> 3;

  2720.     x += mx >> 3;

  2721.     ref += y * ref_stride + x;

  2722.     mx &= 7;

  2723.     my &= 7;

  2724.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2725.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2726.     // the longest loopfilter of the next sbrow

  2727.     th = (y + bh + 4 * !!my + 7) >> 6;

  2728.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2729.     if (x < !!mx * 3 || y < !!my * 3 ||

  2730.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2731.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2732.                                  ref - !!my * 3 * ref_stride - !!mx * 3,

  2733.                                  80, ref_stride,

  2734.                                  bw + !!mx * 7, bh + !!my * 7,

  2735.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2736.         ref = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2737.         ref_stride = 80;

  2738.     }

  2739.     mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);

  2740. }

  2741. 

  2742. static av_always_inline void mc_chroma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2],

  2743.                                                 uint8_t *dst_u, uint8_t *dst_v,

  2744.                                                 ptrdiff_t dst_stride,

  2745.                                                 const uint8_t *ref_u, ptrdiff_t src_stride_u,

  2746.                                                 const uint8_t *ref_v, ptrdiff_t src_stride_v,

  2747.                                                 ThreadFrame *ref_frame,

  2748.                                                 ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2749.                                                 int bw, int bh, int w, int h)

  2750. {

  2751.     int mx = mv->x << !s->ss_h, my = mv->y << !s->ss_v, th;

  2752. 

  2753.     y += my >> 4;

  2754.     x += mx >> 4;

  2755.     ref_u += y * src_stride_u + x;

  2756.     ref_v += y * src_stride_v + x;

  2757.     mx &= 15;

  2758.     my &= 15;

  2759.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2760.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2761.     // the longest loopfilter of the next sbrow

  2762.     th = (y + bh + 4 * !!my + 7) >> (6 - s->ss_v);

  2763.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2764.     if (x < !!mx * 3 || y < !!my * 3 ||

  2765.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2766.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2767.                                  ref_u - !!my * 3 * src_stride_u - !!mx * 3,

  2768.                                  80, src_stride_u,

  2769.                                  bw + !!mx * 7, bh + !!my * 7,

  2770.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2771.         ref_u = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2772.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, 80, bh, mx, my);

  2773. 

  2774.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2775.                                  ref_v - !!my * 3 * src_stride_v - !!mx * 3,

  2776.                                  80, src_stride_v,

  2777.                                  bw + !!mx * 7, bh + !!my * 7,

  2778.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2779.         ref_v = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2780.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, 80, bh, mx, my);

  2781.     } else {

  2782.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);

  2783.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);

  2784.     }

  2785. }

  2786. 

  2787. #define FN(x) x

  2788. #define mc_luma_dir(s, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, bw, bh, w, h, i) \

  2789.     mc_luma_unscaled(s, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \

  2790.                      mv, bw, bh, w, h)

  2791. #define mc_chroma_dir(s, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2792.                       row, col, mv, bw, bh, w, h, i) \

  2793.     mc_chroma_unscaled(s, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2794.                        row, col, mv, bw, bh, w, h)

  2795. #include "vp9_mc_template.c"

  2796. #undef mc_luma_dir_dir

  2797. #undef mc_chroma_dir_dir

  2798. #undef FN

  2799. 

  2800. static void inter_recon(AVCodecContext *ctx)

  2801. {

  2802.     VP9Context *s = ctx->priv_data;

  2803.     VP9Block *b = s->b;

  2804.     int row = s->row, col = s->col;

  2805. 

  2806.     if (s->mvscale[b->ref[0]][0] || (b->comp && s->mvscale[b->ref[1]][0])) {

  2807.         inter_pred_scaled(ctx);

  2808.     } else {

  2809.         inter_pred(ctx);

  2810.     }

  2811.     if (!b->skip) {

  2812.         /* mostly copied intra_recon() */

  2813. 

  2814.         int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2815.         int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2816.         int end_x = FFMIN(2 * (s->cols - col), w4);

  2817.         int end_y = FFMIN(2 * (s->rows - row), h4);

  2818.         int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2819.         int uvstep1d = 1 << b->uvtx, p;

  2820.         uint8_t *dst = s->dst[0];

  2821. 

  2822.         // y itxfm add

  2823.         for (n = 0, y = 0; y < end_y; y += step1d) {

  2824.             uint8_t *ptr = dst;

  2825.             for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d, n += step) {

  2826.                 int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2827. 

  2828.                 if (eob)

  2829.                     s->dsp.itxfm_add[tx][DCT_DCT](ptr, s->y_stride,

  2830.                                                   s->block + 16 * n, eob);

  2831.             }

  2832.             dst += 4 * s->y_stride * step1d;

  2833.         }

  2834. 

  2835.         // uv itxfm add

  2836.         end_x >>= s->ss_h;

  2837.         end_y >>= s->ss_v;

  2838.         step = 1 << (b->uvtx * 2);

  2839.         for (p = 0; p < 2; p++) {

  2840.             dst = s->dst[p + 1];

  2841.             for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2842.                 uint8_t *ptr = dst;

  2843.                 for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d, n += step) {

  2844.                     int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2845. 

  2846.                     if (eob)

  2847.                         s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2848.                                                         s->uvblock[p] + 16 * n, eob);

  2849.                 }

  2850.                 dst += 4 * uvstep1d * s->uv_stride;

  2851.             }

  2852.         }

  2853.     }

  2854. }

  2855. 

  2856. static av_always_inline void mask_edges(struct VP9Filter *lflvl, int is_uv,

  2857.                                         int row_and_7, int col_and_7,

  2858.                                         int w, int h, int col_end, int row_end,

  2859.                                         enum TxfmMode tx, int skip_inter)

  2860. {

  2861.     // FIXME I'm pretty sure all loops can be replaced by a single LUT if

  2862.     // we make VP9Filter.mask uint64_t (i.e. row/col all single variable)

  2863.     // and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then

  2864.     // use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)

  2865. 

  2866.     // the intended behaviour of the vp9 loopfilter is to work on 8-pixel

  2867.     // edges. This means that for UV, we work on two subsampled blocks at

  2868.     // a time, and we only use the topleft block's mode information to set

  2869.     // things like block strength. Thus, for any block size smaller than

  2870.     // 16x16, ignore the odd portion of the block.

  2871.     if (tx == TX_4X4 && is_uv) {

  2872.         if (h == 1) {

  2873.             if (row_and_7 & 1)

  2874.                 return;

  2875.             if (!row_end)

  2876.                 h += 1;

  2877.         }

  2878.         if (w == 1) {

  2879.             if (col_and_7 & 1)

  2880.                 return;

  2881.             if (!col_end)

  2882.                 w += 1;

  2883.         }

  2884.     }

  2885. 

  2886.     if (tx == TX_4X4 && !skip_inter) {

  2887.         int t = 1 << col_and_7, m_col = (t << w) - t, y;

  2888.         int m_col_odd = (t << (w - 1)) - t;

  2889. 

  2890.         // on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide

  2891.         if (is_uv) {

  2892.             int m_row_8 = m_col & 0x01, m_row_4 = m_col - m_row_8;

  2893. 

  2894.             for (y = row_and_7; y < h + row_and_7; y++) {

  2895.                 int col_mask_id = 2 - !(y & 7);

  2896. 

  2897.                 lflvl->mask[is_uv][0][y][1] |= m_row_8;

  2898.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  2899.                 // for odd lines, if the odd col is not being filtered,

  2900.                 // skip odd row also:

  2901.                 // .---. <-- a

  2902.                 // |   |

  2903.                 // |___| <-- b

  2904.                 // ^   ^

  2905.                 // c   d

  2906.                 //

  2907.                 // if a/c are even row/col and b/d are odd, and d is skipped,

  2908.                 // e.g. right edge of size-66x66.webm, then skip b also (bug)

  2909.                 if ((col_end & 1) && (y & 1)) {

  2910.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col_odd;

  2911.                 } else {

  2912.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col;

  2913.                 }

  2914.             }

  2915.         } else {

  2916.             int m_row_8 = m_col & 0x11, m_row_4 = m_col - m_row_8;

  2917. 

  2918.             for (y = row_and_7; y < h + row_and_7; y++) {

  2919.                 int col_mask_id = 2 - !(y & 3);

  2920. 

  2921.                 lflvl->mask[is_uv][0][y][1] |= m_row_8; // row edge

  2922.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  2923.                 lflvl->mask[is_uv][1][y][col_mask_id] |= m_col; // col edge

  2924.                 lflvl->mask[is_uv][0][y][3] |= m_col;

  2925.                 lflvl->mask[is_uv][1][y][3] |= m_col;

  2926.             }

  2927.         }

  2928.     } else {

  2929.         int y, t = 1 << col_and_7, m_col = (t << w) - t;

  2930. 

  2931.         if (!skip_inter) {

  2932.             int mask_id = (tx == TX_8X8);

  2933.             int l2 = tx + is_uv - 1, step1d = 1 << l2;

  2934.             static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };

  2935.             int m_row = m_col & masks[l2];

  2936. 

  2937.             // at odd UV col/row edges tx16/tx32 loopfilter edges, force

  2938.             // 8wd loopfilter to prevent going off the visible edge.

  2939.             if (is_uv && tx > TX_8X8 && (w ^ (w - 1)) == 1) {

  2940.                 int m_row_16 = ((t << (w - 1)) - t) & masks[l2];

  2941.                 int m_row_8 = m_row - m_row_16;

  2942. 

  2943.                 for (y = row_and_7; y < h + row_and_7; y++) {

  2944.                     lflvl->mask[is_uv][0][y][0] |= m_row_16;

  2945.                     lflvl->mask[is_uv][0][y][1] |= m_row_8;

  2946.                 }

  2947.             } else {

  2948.                 for (y = row_and_7; y < h + row_and_7; y++)

  2949.                     lflvl->mask[is_uv][0][y][mask_id] |= m_row;

  2950.             }

  2951. 

  2952.             if (is_uv && tx > TX_8X8 && (h ^ (h - 1)) == 1) {

  2953.                 for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)

  2954.                     lflvl->mask[is_uv][1][y][0] |= m_col;

  2955.                 if (y - row_and_7 == h - 1)

  2956.                     lflvl->mask[is_uv][1][y][1] |= m_col;

  2957.             } else {

  2958.                 for (y = row_and_7; y < h + row_and_7; y += step1d)

  2959.                     lflvl->mask[is_uv][1][y][mask_id] |= m_col;

  2960.             }

  2961.         } else if (tx != TX_4X4) {

  2962.             int mask_id;

  2963. 

  2964.             mask_id = (tx == TX_8X8) || (is_uv && h == 1);

  2965.             lflvl->mask[is_uv][1][row_and_7][mask_id] |= m_col;

  2966.             mask_id = (tx == TX_8X8) || (is_uv && w == 1);

  2967.             for (y = row_and_7; y < h + row_and_7; y++)

  2968.                 lflvl->mask[is_uv][0][y][mask_id] |= t;

  2969.         } else if (is_uv) {

  2970.             int t8 = t & 0x01, t4 = t - t8;

  2971. 

  2972.             for (y = row_and_7; y < h + row_and_7; y++) {

  2973.                 lflvl->mask[is_uv][0][y][2] |= t4;

  2974.                 lflvl->mask[is_uv][0][y][1] |= t8;

  2975.             }

  2976.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 7)] |= m_col;

  2977.         } else {

  2978.             int t8 = t & 0x11, t4 = t - t8;

  2979. 

  2980.             for (y = row_and_7; y < h + row_and_7; y++) {

  2981.                 lflvl->mask[is_uv][0][y][2] |= t4;

  2982.                 lflvl->mask[is_uv][0][y][1] |= t8;

  2983.             }

  2984.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 3)] |= m_col;

  2985.         }

  2986.     }

  2987. }

  2988. 

  2989. static void decode_b(AVCodecContext *ctx, int row, int col,

  2990.                      struct VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,

  2991.                      enum BlockLevel bl, enum BlockPartition bp)

  2992. {

  2993.     VP9Context *s = ctx->priv_data;

  2994.     VP9Block *b = s->b;

  2995.     enum BlockSize bs = bl * 3 + bp;

  2996.     int w4 = bwh_tab[1][bs][0], h4 = bwh_tab[1][bs][1], lvl;

  2997.     int emu[2];

  2998.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2999. 

  3000.     s->row = row;

  3001.     s->row7 = row & 7;

  3002.     s->col = col;

  3003.     s->col7 = col & 7;

  3004.     s->min_mv.x = -(128 + col * 64);

  3005.     s->min_mv.y = -(128 + row * 64);

  3006.     s->max_mv.x = 128 + (s->cols - col - w4) * 64;

  3007.     s->max_mv.y = 128 + (s->rows - row - h4) * 64;

  3008.     if (s->pass < 2) {

  3009.         b->bs = bs;

  3010.         b->bl = bl;

  3011.         b->bp = bp;

  3012.         decode_mode(ctx);

  3013.         b->uvtx = b->tx - ((s->ss_h && w4 * 2 == (1 << b->tx)) ||

  3014.                            (s->ss_v && h4 * 2 == (1 << b->tx)));

  3015. 

  3016.         if (!b->skip) {

  3017.             decode_coeffs(ctx);

  3018.         } else {

  3019.             int row7 = s->row7;

  3020. 

  3021. #define SPLAT_ZERO_CTX(v, n) \

  3022.     switch (n) { \

  3023.     case 1:  v = 0;          break; \

  3024.     case 2:  AV_ZERO16(&v);  break; \

  3025.     case 4:  AV_ZERO32(&v);  break; \

  3026.     case 8:  AV_ZERO64(&v);  break; \

  3027.     case 16: AV_ZERO128(&v); break; \

  3028.     }

  3029. #define SPLAT_ZERO_YUV(dir, var, off, n, dir2) \

  3030.     do { \

  3031.         SPLAT_ZERO_CTX(s->dir##_y_##var[off * 2], n * 2); \

  3032.         if (s->ss_##dir2) { \

  3033.             SPLAT_ZERO_CTX(s->dir##_uv_##var[0][off], n); \

  3034.             SPLAT_ZERO_CTX(s->dir##_uv_##var[1][off], n); \

  3035.         } else { \

  3036.             SPLAT_ZERO_CTX(s->dir##_uv_##var[0][off * 2], n * 2); \

  3037.             SPLAT_ZERO_CTX(s->dir##_uv_##var[1][off * 2], n * 2); \

  3038.         } \

  3039.     } while (0)

  3040. 

  3041.             switch (w4) {

  3042.             case 1: SPLAT_ZERO_YUV(above, nnz_ctx, col, 1, h); break;

  3043.             case 2: SPLAT_ZERO_YUV(above, nnz_ctx, col, 2, h); break;

  3044.             case 4: SPLAT_ZERO_YUV(above, nnz_ctx, col, 4, h); break;

  3045.             case 8: SPLAT_ZERO_YUV(above, nnz_ctx, col, 8, h); break;

  3046.             }

  3047.             switch (h4) {

  3048.             case 1: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 1, v); break;

  3049.             case 2: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 2, v); break;

  3050.             case 4: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 4, v); break;

  3051.             case 8: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 8, v); break;

  3052.             }

  3053.         }

  3054.         if (s->pass == 1) {

  3055.             s->b++;

  3056.             s->block += w4 * h4 * 64;

  3057.             s->uvblock[0] += w4 * h4 * 64 >> (s->ss_h + s->ss_v);

  3058.             s->uvblock[1] += w4 * h4 * 64 >> (s->ss_h + s->ss_v);

  3059.             s->eob += 4 * w4 * h4;

  3060.             s->uveob[0] += 4 * w4 * h4 >> (s->ss_h + s->ss_v);

  3061.             s->uveob[1] += 4 * w4 * h4 >> (s->ss_h + s->ss_v);

  3062. 

  3063.             return;

  3064.         }

  3065.     }

  3066. 

  3067.     // emulated overhangs if the stride of the target buffer can't hold. This

  3068.     // allows to support emu-edge and so on even if we have large block

  3069.     // overhangs

  3070.     emu[0] = (col + w4) * 8 > f->linesize[0] ||

  3071.              (row + h4) > s->rows;

  3072.     emu[1] = (col + w4) * 4 > f->linesize[1] ||

  3073.              (row + h4) > s->rows;

  3074.     if (emu[0]) {

  3075.         s->dst[0] = s->tmp_y;

  3076.         s->y_stride = 64;

  3077.     } else {

  3078.         s->dst[0] = f->data[0] + yoff;

  3079.         s->y_stride = f->linesize[0];

  3080.     }

  3081.     if (emu[1]) {

  3082.         s->dst[1] = s->tmp_uv[0];

  3083.         s->dst[2] = s->tmp_uv[1];

  3084.         s->uv_stride = 32;

  3085.     } else {

  3086.         s->dst[1] = f->data[1] + uvoff;

  3087.         s->dst[2] = f->data[2] + uvoff;

  3088.         s->uv_stride = f->linesize[1];

  3089.     }

  3090.     if (b->intra) {

  3091.         intra_recon(ctx, yoff, uvoff);

  3092.     } else {

  3093.         inter_recon(ctx);

  3094.     }

  3095.     if (emu[0]) {

  3096.         int w = FFMIN(s->cols - col, w4) * 8, h = FFMIN(s->rows - row, h4) * 8, n, o = 0;

  3097. 

  3098.         for (n = 0; o < w; n++) {

  3099.             int bw = 64 >> n;

  3100. 

  3101.             av_assert2(n <= 4);

  3102.             if (w & bw) {

  3103.                 s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o, f->linesize[0],

  3104.                                          s->tmp_y + o, 64, h, 0, 0);

  3105.                 o += bw;

  3106.             }

  3107.         }

  3108.     }

  3109.     if (emu[1]) {

  3110.         int w = FFMIN(s->cols - col, w4) * 4, h = FFMIN(s->rows - row, h4) * 4, n, o = 0;

  3111. 

  3112.         for (n = 1; o < w; n++) {

  3113.             int bw = 64 >> n;

  3114. 

  3115.             av_assert2(n <= 4);

  3116.             if (w & bw) {

  3117.                 s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o, f->linesize[1],

  3118.                                          s->tmp_uv[0] + o, 32, h, 0, 0);

  3119.                 s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o, f->linesize[2],

  3120.                                          s->tmp_uv[1] + o, 32, h, 0, 0);

  3121.                 o += bw;

  3122.             }

  3123.         }

  3124.     }

  3125. 

  3126.     // pick filter level and find edges to apply filter to

  3127.     if (s->filter.level &&

  3128.         (lvl = s->segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]

  3129.                                                     [b->mode[3] != ZEROMV]) > 0) {

  3130.         int x_end = FFMIN(s->cols - col, w4), y_end = FFMIN(s->rows - row, h4);

  3131.         int skip_inter = !b->intra && b->skip, col7 = s->col7, row7 = s->row7;

  3132. 

  3133.         setctx_2d(&lflvl->level[row7 * 8 + col7], w4, h4, 8, lvl);

  3134.         mask_edges(lflvl, 0, row7, col7, x_end, y_end, 0, 0, b->tx, skip_inter);

  3135.         mask_edges(lflvl, 1, row7, col7, x_end, y_end,

  3136.                    s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,

  3137.                    s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,

  3138.                    b->uvtx, skip_inter);

  3139. 

  3140.         if (!s->filter.lim_lut[lvl]) {

  3141.             int sharp = s->filter.sharpness;

  3142.             int limit = lvl;

  3143. 

  3144.             if (sharp > 0) {

  3145.                 limit >>= (sharp + 3) >> 2;

  3146.                 limit = FFMIN(limit, 9 - sharp);

  3147.             }

  3148.             limit = FFMAX(limit, 1);

  3149. 

  3150.             s->filter.lim_lut[lvl] = limit;

  3151.             s->filter.mblim_lut[lvl] = 2 * (lvl + 2) + limit;

  3152.         }

  3153.     }

  3154. 

  3155.     if (s->pass == 2) {

  3156.         s->b++;

  3157.         s->block += w4 * h4 * 64;

  3158.         s->uvblock[0] += w4 * h4 * 64 >> (s->ss_v + s->ss_h);

  3159.         s->uvblock[1] += w4 * h4 * 64 >> (s->ss_v + s->ss_h);

  3160.         s->eob += 4 * w4 * h4;

  3161.         s->uveob[0] += 4 * w4 * h4 >> (s->ss_v + s->ss_h);

  3162.         s->uveob[1] += 4 * w4 * h4 >> (s->ss_v + s->ss_h);

  3163.     }

  3164. }

  3165. 

  3166. static void decode_sb(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  3167.                       ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  3168. {

  3169.     VP9Context *s = ctx->priv_data;

  3170.     int c = ((s->above_partition_ctx[col] >> (3 - bl)) & 1) |

  3171.             (((s->left_partition_ctx[row & 0x7] >> (3 - bl)) & 1) << 1);

  3172.     const uint8_t *p = s->keyframe || s->intraonly ? vp9_default_kf_partition_probs[bl][c] :

  3173.                                                      s->prob.p.partition[bl][c];

  3174.     enum BlockPartition bp;

  3175.     ptrdiff_t hbs = 4 >> bl;

  3176.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  3177.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  3178. 

  3179.     if (bl == BL_8X8) {

  3180.         bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  3181.         decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3182.     } else if (col + hbs < s->cols) { // FIXME why not <=?

  3183.         if (row + hbs < s->rows) { // FIXME why not <=?

  3184.             bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  3185.             switch (bp) {

  3186.             case PARTITION_NONE:

  3187.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3188.                 break;

  3189.             case PARTITION_H:

  3190.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3191.                 yoff  += hbs * 8 * y_stride;

  3192.                 uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3193.                 decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, bl, bp);

  3194.                 break;

  3195.             case PARTITION_V:

  3196.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3197.                 yoff  += hbs * 8;

  3198.                 uvoff += hbs * 8 >> s->ss_h;

  3199.                 decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, bl, bp);

  3200.                 break;

  3201.             case PARTITION_SPLIT:

  3202.                 decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3203.                 decode_sb(ctx, row, col + hbs, lflvl,

  3204.                           yoff + 8 * hbs, uvoff + (8 * hbs >> s->ss_h), bl + 1);

  3205.                 yoff  += hbs * 8 * y_stride;

  3206.                 uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3207.                 decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3208.                 decode_sb(ctx, row + hbs, col + hbs, lflvl,

  3209.                           yoff + 8 * hbs, uvoff + (8 * hbs >> s->ss_h), bl + 1);

  3210.                 break;

  3211.             default:

  3212.                 av_assert0(0);

  3213.             }

  3214.         } else if (vp56_rac_get_prob_branchy(&s->c, p[1])) {

  3215.             bp = PARTITION_SPLIT;

  3216.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3217.             decode_sb(ctx, row, col + hbs, lflvl,

  3218.                       yoff + 8 * hbs, uvoff + (8 * hbs >> s->ss_h), bl + 1);

  3219.         } else {

  3220.             bp = PARTITION_H;

  3221.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3222.         }

  3223.     } else if (row + hbs < s->rows) { // FIXME why not <=?

  3224.         if (vp56_rac_get_prob_branchy(&s->c, p[2])) {

  3225.             bp = PARTITION_SPLIT;

  3226.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3227.             yoff  += hbs * 8 * y_stride;

  3228.             uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3229.             decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3230.         } else {

  3231.             bp = PARTITION_V;

  3232.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3233.         }

  3234.     } else {

  3235.         bp = PARTITION_SPLIT;

  3236.         decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3237.     }

  3238.     s->counts.partition[bl][c][bp]++;

  3239. }

  3240. 

  3241. static void decode_sb_mem(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  3242.                           ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  3243. {

  3244.     VP9Context *s = ctx->priv_data;

  3245.     VP9Block *b = s->b;

  3246.     ptrdiff_t hbs = 4 >> bl;

  3247.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  3248.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  3249. 

  3250.     if (bl == BL_8X8) {

  3251.         av_assert2(b->bl == BL_8X8);

  3252.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  3253.     } else if (s->b->bl == bl) {

  3254.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  3255.         if (b->bp == PARTITION_H && row + hbs < s->rows) {

  3256.             yoff  += hbs * 8 * y_stride;

  3257.             uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3258.             decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, b->bl, b->bp);

  3259.         } else if (b->bp == PARTITION_V && col + hbs < s->cols) {

  3260.             yoff  += hbs * 8;

  3261.             uvoff += hbs * 8 >> s->ss_h;

  3262.             decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, b->bl, b->bp);

  3263.         }

  3264.     } else {

  3265.         decode_sb_mem(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3266.         if (col + hbs < s->cols) { // FIXME why not <=?

  3267.             if (row + hbs < s->rows) {

  3268.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff + 8 * hbs,

  3269.                               uvoff + (8 * hbs >> s->ss_h), bl + 1);

  3270.                 yoff  += hbs * 8 * y_stride;

  3271.                 uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3272.                 decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3273.                 decode_sb_mem(ctx, row + hbs, col + hbs, lflvl,

  3274.                                     yoff + 8 * hbs, uvoff + (8 * hbs >> s->ss_h), bl + 1);

  3275.             } else {

  3276.                 yoff  += hbs * 8;

  3277.                 uvoff += hbs * 8 >> s->ss_h;

  3278.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff, uvoff, bl + 1);

  3279.             }

  3280.         } else if (row + hbs < s->rows) {

  3281.             yoff  += hbs * 8 * y_stride;

  3282.             uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3283.             decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3284.         }

  3285.     }

  3286. }

  3287. 

  3288. static void loopfilter_sb(AVCodecContext *ctx, struct VP9Filter *lflvl,

  3289.                           int row, int col, ptrdiff_t yoff, ptrdiff_t uvoff)

  3290. {

  3291.     VP9Context *s = ctx->priv_data;

  3292.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  3293.     uint8_t *dst = f->data[0] + yoff, *lvl = lflvl->level;

  3294.     ptrdiff_t ls_y = f->linesize[0], ls_uv = f->linesize[1];

  3295.     int y, x, p;

  3296. 

  3297.     // FIXME in how far can we interleave the v/h loopfilter calls? E.g.

  3298.     // if you think of them as acting on a 8x8 block max, we can interleave

  3299.     // each v/h within the single x loop, but that only works if we work on

  3300.     // 8 pixel blocks, and we won't always do that (we want at least 16px

  3301.     // to use SSE2 optimizations, perhaps 32 for AVX2)

  3302. 

  3303.     // filter edges between columns, Y plane (e.g. block1 | block2)

  3304.     for (y = 0; y < 8; y += 2, dst += 16 * ls_y, lvl += 16) {

  3305.         uint8_t *ptr = dst, *l = lvl, *hmask1 = lflvl->mask[0][0][y];

  3306.         uint8_t *hmask2 = lflvl->mask[0][0][y + 1];

  3307.         unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2], hm13 = hmask1[3];

  3308.         unsigned hm2 = hmask2[1] | hmask2[2], hm23 = hmask2[3];

  3309.         unsigned hm = hm1 | hm2 | hm13 | hm23;

  3310. 

  3311.         for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 8, l++) {

  3312.             if (col || x > 1) {

  3313.                 if (hm1 & x) {

  3314.                     int L = *l, H = L >> 4;

  3315.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3316. 

  3317.                     if (hmask1[0] & x) {

  3318.                         if (hmask2[0] & x) {

  3319.                             av_assert2(l[8] == L);

  3320.                             s->dsp.loop_filter_16[0](ptr, ls_y, E, I, H);

  3321.                         } else {

  3322.                             s->dsp.loop_filter_8[2][0](ptr, ls_y, E, I, H);

  3323.                         }

  3324.                     } else if (hm2 & x) {

  3325.                         L = l[8];

  3326.                         H |= (L >> 4) << 8;

  3327.                         E |= s->filter.mblim_lut[L] << 8;

  3328.                         I |= s->filter.lim_lut[L] << 8;

  3329.                         s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3330.                                                [!!(hmask2[1] & x)]

  3331.                                                [0](ptr, ls_y, E, I, H);

  3332.                     } else {

  3333.                         s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3334.                                             [0](ptr, ls_y, E, I, H);

  3335.                     }

  3336.                 } else if (hm2 & x) {

  3337.                     int L = l[8], H = L >> 4;

  3338.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3339. 

  3340.                     s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3341.                                         [0](ptr + 8 * ls_y, ls_y, E, I, H);

  3342.                 }

  3343.             }

  3344.             if (hm13 & x) {

  3345.                 int L = *l, H = L >> 4;

  3346.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3347. 

  3348.                 if (hm23 & x) {

  3349.                     L = l[8];

  3350.                     H |= (L >> 4) << 8;

  3351.                     E |= s->filter.mblim_lut[L] << 8;

  3352.                     I |= s->filter.lim_lut[L] << 8;

  3353.                     s->dsp.loop_filter_mix2[0][0][0](ptr + 4, ls_y, E, I, H);

  3354.                 } else {

  3355.                     s->dsp.loop_filter_8[0][0](ptr + 4, ls_y, E, I, H);

  3356.                 }

  3357.             } else if (hm23 & x) {

  3358.                 int L = l[8], H = L >> 4;

  3359.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3360. 

  3361.                 s->dsp.loop_filter_8[0][0](ptr + 8 * ls_y + 4, ls_y, E, I, H);

  3362.             }

  3363.         }

  3364.     }

  3365. 

  3366.     //                                          block1

  3367.     // filter edges between rows, Y plane (e.g. ------)

  3368.     //                                          block2

  3369.     dst = f->data[0] + yoff;

  3370.     lvl = lflvl->level;

  3371.     for (y = 0; y < 8; y++, dst += 8 * ls_y, lvl += 8) {

  3372.         uint8_t *ptr = dst, *l = lvl, *vmask = lflvl->mask[0][1][y];

  3373.         unsigned vm = vmask[0] | vmask[1] | vmask[2], vm3 = vmask[3];

  3374. 

  3375.         for (x = 1; vm & ~(x - 1); x <<= 2, ptr += 16, l += 2) {

  3376.             if (row || y) {

  3377.                 if (vm & x) {

  3378.                     int L = *l, H = L >> 4;

  3379.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3380. 

  3381.                     if (vmask[0] & x) {

  3382.                         if (vmask[0] & (x << 1)) {

  3383.                             av_assert2(l[1] == L);

  3384.                             s->dsp.loop_filter_16[1](ptr, ls_y, E, I, H);

  3385.                         } else {

  3386.                             s->dsp.loop_filter_8[2][1](ptr, ls_y, E, I, H);

  3387.                         }

  3388.                     } else if (vm & (x << 1)) {

  3389.                         L = l[1];

  3390.                         H |= (L >> 4) << 8;

  3391.                         E |= s->filter.mblim_lut[L] << 8;

  3392.                         I |= s->filter.lim_lut[L] << 8;

  3393.                         s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3394.                                                [!!(vmask[1] & (x << 1))]

  3395.                                                [1](ptr, ls_y, E, I, H);

  3396.                     } else {

  3397.                         s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3398.                                             [1](ptr, ls_y, E, I, H);

  3399.                     }

  3400.                 } else if (vm & (x << 1)) {

  3401.                     int L = l[1], H = L >> 4;

  3402.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3403. 

  3404.                     s->dsp.loop_filter_8[!!(vmask[1] & (x << 1))]

  3405.                                         [1](ptr + 8, ls_y, E, I, H);

  3406.                 }

  3407.             }

  3408.             if (vm3 & x) {

  3409.                 int L = *l, H = L >> 4;

  3410.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3411. 

  3412.                 if (vm3 & (x << 1)) {

  3413.                     L = l[1];

  3414.                     H |= (L >> 4) << 8;

  3415.                     E |= s->filter.mblim_lut[L] << 8;

  3416.                     I |= s->filter.lim_lut[L] << 8;

  3417.                     s->dsp.loop_filter_mix2[0][0][1](ptr + ls_y * 4, ls_y, E, I, H);

  3418.                 } else {

  3419.                     s->dsp.loop_filter_8[0][1](ptr + ls_y * 4, ls_y, E, I, H);

  3420.                 }

  3421.             } else if (vm3 & (x << 1)) {

  3422.                 int L = l[1], H = L >> 4;

  3423.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3424. 

  3425.                 s->dsp.loop_filter_8[0][1](ptr + ls_y * 4 + 8, ls_y, E, I, H);

  3426.             }

  3427.         }

  3428.     }

  3429. 

  3430.     // same principle but for U/V planes

  3431.     for (p = 0; p < 2; p++) {

  3432.         lvl = lflvl->level;

  3433.         dst = f->data[1 + p] + uvoff;

  3434.         for (y = 0; y < 8; y += 4, dst += 16 * ls_uv, lvl += 32) {

  3435.             uint8_t *ptr = dst, *l = lvl, *hmask1 = lflvl->mask[1][0][y];

  3436.             uint8_t *hmask2 = lflvl->mask[1][0][y + 2];

  3437.             unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2];

  3438.             unsigned hm2 = hmask2[1] | hmask2[2], hm = hm1 | hm2;

  3439. 

  3440.             for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 4) {

  3441.                 if (col || x > 1) {

  3442.                     if (hm1 & x) {

  3443.                         int L = *l, H = L >> 4;

  3444.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3445. 

  3446.                         if (hmask1[0] & x) {

  3447.                             if (hmask2[0] & x) {

  3448.                                 av_assert2(l[16] == L);

  3449.                                 s->dsp.loop_filter_16[0](ptr, ls_uv, E, I, H);

  3450.                             } else {

  3451.                                 s->dsp.loop_filter_8[2][0](ptr, ls_uv, E, I, H);

  3452.                             }

  3453.                         } else if (hm2 & x) {

  3454.                             L = l[16];

  3455.                             H |= (L >> 4) << 8;

  3456.                             E |= s->filter.mblim_lut[L] << 8;

  3457.                             I |= s->filter.lim_lut[L] << 8;

  3458.                             s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3459.                                                    [!!(hmask2[1] & x)]

  3460.                                                    [0](ptr, ls_uv, E, I, H);

  3461.                         } else {

  3462.                             s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3463.                                                 [0](ptr, ls_uv, E, I, H);

  3464.                         }

  3465.                     } else if (hm2 & x) {

  3466.                         int L = l[16], H = L >> 4;

  3467.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3468. 

  3469.                         s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3470.                                             [0](ptr + 8 * ls_uv, ls_uv, E, I, H);

  3471.                     }

  3472.                 }

  3473.                 if (x & 0xAA)

  3474.                     l += 2;

  3475.             }

  3476.         }

  3477.         lvl = lflvl->level;

  3478.         dst = f->data[1 + p] + uvoff;

  3479.         for (y = 0; y < 8; y++, dst += 4 * ls_uv) {

  3480.             uint8_t *ptr = dst, *l = lvl, *vmask = lflvl->mask[1][1][y];

  3481.             unsigned vm = vmask[0] | vmask[1] | vmask[2];

  3482. 

  3483.             for (x = 1; vm & ~(x - 1); x <<= 4, ptr += 16, l += 4) {

  3484.                 if (row || y) {

  3485.                     if (vm & x) {

  3486.                         int L = *l, H = L >> 4;

  3487.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3488. 

  3489.                         if (vmask[0] & x) {

  3490.                             if (vmask[0] & (x << 2)) {

  3491.                                 av_assert2(l[2] == L);

  3492.                                 s->dsp.loop_filter_16[1](ptr, ls_uv, E, I, H);

  3493.                             } else {

  3494.                                 s->dsp.loop_filter_8[2][1](ptr, ls_uv, E, I, H);

  3495.                             }

  3496.                         } else if (vm & (x << 2)) {

  3497.                             L = l[2];

  3498.                             H |= (L >> 4) << 8;

  3499.                             E |= s->filter.mblim_lut[L] << 8;

  3500.                             I |= s->filter.lim_lut[L] << 8;

  3501.                             s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3502.                                                    [!!(vmask[1] & (x << 2))]

  3503.                                                    [1](ptr, ls_uv, E, I, H);

  3504.                         } else {

  3505.                             s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3506.                                                 [1](ptr, ls_uv, E, I, H);

  3507.                         }

  3508.                     } else if (vm & (x << 2)) {

  3509.                         int L = l[2], H = L >> 4;

  3510.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3511. 

  3512.                         s->dsp.loop_filter_8[!!(vmask[1] & (x << 2))]

  3513.                                             [1](ptr + 8, ls_uv, E, I, H);

  3514.                     }

  3515.                 }

  3516.             }

  3517.             if (y & 1)

  3518.                 lvl += 16;

  3519.         }

  3520.     }

  3521. }

  3522. 

  3523. static void set_tile_offset(int *start, int *end, int idx, int log2_n, int n)

  3524. {

  3525.     int sb_start = ( idx      * n) >> log2_n;

  3526.     int sb_end   = ((idx + 1) * n) >> log2_n;

  3527.     *start = FFMIN(sb_start, n) << 3;

  3528.     *end   = FFMIN(sb_end,   n) << 3;

  3529. }

  3530. 

  3531. static av_always_inline void adapt_prob(uint8_t *p, unsigned ct0, unsigned ct1,

  3532.                                         int max_count, int update_factor)

  3533. {

  3534.     unsigned ct = ct0 + ct1, p2, p1;

  3535. 

  3536.     if (!ct)

  3537.         return;

  3538. 

  3539.     p1 = *p;

  3540.     p2 = ((ct0 << 8) + (ct >> 1)) / ct;

  3541.     p2 = av_clip(p2, 1, 255);

  3542.     ct = FFMIN(ct, max_count);

  3543.     update_factor = FASTDIV(update_factor * ct, max_count);

  3544. 

  3545.     // (p1 * (256 - update_factor) + p2 * update_factor + 128) >> 8

  3546.     *p = p1 + (((p2 - p1) * update_factor + 128) >> 8);

  3547. }

  3548. 

  3549. static void adapt_probs(VP9Context *s)

  3550. {

  3551.     int i, j, k, l, m;

  3552.     prob_context *p = &s->prob_ctx[s->framectxid].p;

  3553.     int uf = (s->keyframe || s->intraonly || !s->last_keyframe) ? 112 : 128;

  3554. 

  3555.     // coefficients

  3556.     for (i = 0; i < 4; i++)

  3557.         for (j = 0; j < 2; j++)

  3558.             for (k = 0; k < 2; k++)

  3559.                 for (l = 0; l < 6; l++)

  3560.                     for (m = 0; m < 6; m++) {

  3561.                         uint8_t *pp = s->prob_ctx[s->framectxid].coef[i][j][k][l][m];

  3562.                         unsigned *e = s->counts.eob[i][j][k][l][m];

  3563.                         unsigned *c = s->counts.coef[i][j][k][l][m];

  3564. 

  3565.                         if (l == 0 && m >= 3) // dc only has 3 pt

  3566.                             break;

  3567. 

  3568.                         adapt_prob(&pp[0], e[0], e[1], 24, uf);

  3569.                         adapt_prob(&pp[1], c[0], c[1] + c[2], 24, uf);

  3570.                         adapt_prob(&pp[2], c[1], c[2], 24, uf);

  3571.                     }

  3572. 

  3573.     if (s->keyframe || s->intraonly) {

  3574.         memcpy(p->skip,  s->prob.p.skip,  sizeof(p->skip));

  3575.         memcpy(p->tx32p, s->prob.p.tx32p, sizeof(p->tx32p));

  3576.         memcpy(p->tx16p, s->prob.p.tx16p, sizeof(p->tx16p));

  3577.         memcpy(p->tx8p,  s->prob.p.tx8p,  sizeof(p->tx8p));

  3578.         return;

  3579.     }

  3580. 

  3581.     // skip flag

  3582.     for (i = 0; i < 3; i++)

  3583.         adapt_prob(&p->skip[i], s->counts.skip[i][0], s->counts.skip[i][1], 20, 128);

  3584. 

  3585.     // intra/inter flag

  3586.     for (i = 0; i < 4; i++)

  3587.         adapt_prob(&p->intra[i], s->counts.intra[i][0], s->counts.intra[i][1], 20, 128);

  3588. 

  3589.     // comppred flag

  3590.     if (s->comppredmode == PRED_SWITCHABLE) {

  3591.       for (i = 0; i < 5; i++)

  3592.           adapt_prob(&p->comp[i], s->counts.comp[i][0], s->counts.comp[i][1], 20, 128);

  3593.     }

  3594. 

  3595.     // reference frames

  3596.     if (s->comppredmode != PRED_SINGLEREF) {

  3597.       for (i = 0; i < 5; i++)

  3598.           adapt_prob(&p->comp_ref[i], s->counts.comp_ref[i][0],

  3599.                      s->counts.comp_ref[i][1], 20, 128);

  3600.     }

  3601. 

  3602.     if (s->comppredmode != PRED_COMPREF) {

  3603.       for (i = 0; i < 5; i++) {

  3604.           uint8_t *pp = p->single_ref[i];

  3605.           unsigned (*c)[2] = s->counts.single_ref[i];

  3606. 

  3607.           adapt_prob(&pp[0], c[0][0], c[0][1], 20, 128);

  3608.           adapt_prob(&pp[1], c[1][0], c[1][1], 20, 128);

  3609.       }

  3610.     }

  3611. 

  3612.     // block partitioning

  3613.     for (i = 0; i < 4; i++)

  3614.         for (j = 0; j < 4; j++) {

  3615.             uint8_t *pp = p->partition[i][j];

  3616.             unsigned *c = s->counts.partition[i][j];

  3617. 

  3618.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3619.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3620.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3621.         }

  3622. 

  3623.     // tx size

  3624.     if (s->txfmmode == TX_SWITCHABLE) {

  3625.       for (i = 0; i < 2; i++) {

  3626.           unsigned *c16 = s->counts.tx16p[i], *c32 = s->counts.tx32p[i];

  3627. 

  3628.           adapt_prob(&p->tx8p[i], s->counts.tx8p[i][0], s->counts.tx8p[i][1], 20, 128);

  3629.           adapt_prob(&p->tx16p[i][0], c16[0], c16[1] + c16[2], 20, 128);

  3630.           adapt_prob(&p->tx16p[i][1], c16[1], c16[2], 20, 128);

  3631.           adapt_prob(&p->tx32p[i][0], c32[0], c32[1] + c32[2] + c32[3], 20, 128);

  3632.           adapt_prob(&p->tx32p[i][1], c32[1], c32[2] + c32[3], 20, 128);

  3633.           adapt_prob(&p->tx32p[i][2], c32[2], c32[3], 20, 128);

  3634.       }

  3635.     }

  3636. 

  3637.     // interpolation filter

  3638.     if (s->filtermode == FILTER_SWITCHABLE) {

  3639.         for (i = 0; i < 4; i++) {

  3640.             uint8_t *pp = p->filter[i];

  3641.             unsigned *c = s->counts.filter[i];

  3642. 

  3643.             adapt_prob(&pp[0], c[0], c[1] + c[2], 20, 128);

  3644.             adapt_prob(&pp[1], c[1], c[2], 20, 128);

  3645.         }

  3646.     }

  3647. 

  3648.     // inter modes

  3649.     for (i = 0; i < 7; i++) {

  3650.         uint8_t *pp = p->mv_mode[i];

  3651.         unsigned *c = s->counts.mv_mode[i];

  3652. 

  3653.         adapt_prob(&pp[0], c[2], c[1] + c[0] + c[3], 20, 128);

  3654.         adapt_prob(&pp[1], c[0], c[1] + c[3], 20, 128);

  3655.         adapt_prob(&pp[2], c[1], c[3], 20, 128);

  3656.     }

  3657. 

  3658.     // mv joints

  3659.     {

  3660.         uint8_t *pp = p->mv_joint;

  3661.         unsigned *c = s->counts.mv_joint;

  3662. 

  3663.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3664.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3665.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3666.     }

  3667. 

  3668.     // mv components

  3669.     for (i = 0; i < 2; i++) {

  3670.         uint8_t *pp;

  3671.         unsigned *c, (*c2)[2], sum;

  3672. 

  3673.         adapt_prob(&p->mv_comp[i].sign, s->counts.mv_comp[i].sign[0],

  3674.                    s->counts.mv_comp[i].sign[1], 20, 128);

  3675. 

  3676.         pp = p->mv_comp[i].classes;

  3677.         c = s->counts.mv_comp[i].classes;

  3678.         sum = c[1] + c[2] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9] + c[10];

  3679.         adapt_prob(&pp[0], c[0], sum, 20, 128);

  3680.         sum -= c[1];

  3681.         adapt_prob(&pp[1], c[1], sum, 20, 128);

  3682.         sum -= c[2] + c[3];

  3683.         adapt_prob(&pp[2], c[2] + c[3], sum, 20, 128);

  3684.         adapt_prob(&pp[3], c[2], c[3], 20, 128);

  3685.         sum -= c[4] + c[5];

  3686.         adapt_prob(&pp[4], c[4] + c[5], sum, 20, 128);

  3687.         adapt_prob(&pp[5], c[4], c[5], 20, 128);

  3688.         sum -= c[6];

  3689.         adapt_prob(&pp[6], c[6], sum, 20, 128);

  3690.         adapt_prob(&pp[7], c[7] + c[8], c[9] + c[10], 20, 128);

  3691.         adapt_prob(&pp[8], c[7], c[8], 20, 128);

  3692.         adapt_prob(&pp[9], c[9], c[10], 20, 128);

  3693. 

  3694.         adapt_prob(&p->mv_comp[i].class0, s->counts.mv_comp[i].class0[0],

  3695.                    s->counts.mv_comp[i].class0[1], 20, 128);

  3696.         pp = p->mv_comp[i].bits;

  3697.         c2 = s->counts.mv_comp[i].bits;

  3698.         for (j = 0; j < 10; j++)

  3699.             adapt_prob(&pp[j], c2[j][0], c2[j][1], 20, 128);

  3700. 

  3701.         for (j = 0; j < 2; j++) {

  3702.             pp = p->mv_comp[i].class0_fp[j];

  3703.             c = s->counts.mv_comp[i].class0_fp[j];

  3704.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3705.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3706.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3707.         }

  3708.         pp = p->mv_comp[i].fp;

  3709.         c = s->counts.mv_comp[i].fp;

  3710.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3711.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3712.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3713. 

  3714.         if (s->highprecisionmvs) {

  3715.             adapt_prob(&p->mv_comp[i].class0_hp, s->counts.mv_comp[i].class0_hp[0],

  3716.                        s->counts.mv_comp[i].class0_hp[1], 20, 128);

  3717.             adapt_prob(&p->mv_comp[i].hp, s->counts.mv_comp[i].hp[0],

  3718.                        s->counts.mv_comp[i].hp[1], 20, 128);

  3719.         }

  3720.     }

  3721. 

  3722.     // y intra modes

  3723.     for (i = 0; i < 4; i++) {

  3724.         uint8_t *pp = p->y_mode[i];

  3725.         unsigned *c = s->counts.y_mode[i], sum, s2;

  3726. 

  3727.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3728.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3729.         sum -= c[TM_VP8_PRED];

  3730.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3731.         sum -= c[VERT_PRED];

  3732.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3733.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3734.         sum -= s2;

  3735.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3736.         s2 -= c[HOR_PRED];

  3737.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3738.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3739.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3740.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3741.         sum -= c[VERT_LEFT_PRED];

  3742.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3743.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3744.     }

  3745. 

  3746.     // uv intra modes

  3747.     for (i = 0; i < 10; i++) {

  3748.         uint8_t *pp = p->uv_mode[i];

  3749.         unsigned *c = s->counts.uv_mode[i], sum, s2;

  3750. 

  3751.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3752.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3753.         sum -= c[TM_VP8_PRED];

  3754.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3755.         sum -= c[VERT_PRED];

  3756.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3757.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3758.         sum -= s2;

  3759.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3760.         s2 -= c[HOR_PRED];

  3761.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3762.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3763.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3764.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3765.         sum -= c[VERT_LEFT_PRED];

  3766.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3767.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3768.     }

  3769. }

  3770. 

  3771. static void free_buffers(VP9Context *s)

  3772. {

  3773.     av_freep(&s->intra_pred_data[0]);

  3774.     av_freep(&s->b_base);

  3775.     av_freep(&s->block_base);

  3776. }

  3777. 

  3778. static av_cold int vp9_decode_free(AVCodecContext *ctx)

  3779. {

  3780.     VP9Context *s = ctx->priv_data;

  3781.     int i;

  3782. 

  3783.     for (i = 0; i < 3; i++) {

  3784.         if (s->frames[i].tf.f->data[0])

  3785.             vp9_unref_frame(ctx, &s->frames[i]);

  3786.         av_frame_free(&s->frames[i].tf.f);

  3787.     }

  3788.     for (i = 0; i < 8; i++) {

  3789.         if (s->refs[i].f->data[0])

  3790.             ff_thread_release_buffer(ctx, &s->refs[i]);

  3791.         av_frame_free(&s->refs[i].f);

  3792.         if (s->next_refs[i].f->data[0])

  3793.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3794.         av_frame_free(&s->next_refs[i].f);

  3795.     }

  3796.     free_buffers(s);

  3797.     av_freep(&s->c_b);

  3798.     s->c_b_size = 0;

  3799. 

  3800.     return 0;

  3801. }

  3802. 

  3803. 

  3804. static int vp9_decode_frame(AVCodecContext *ctx, void *frame,

  3805.                             int *got_frame, AVPacket *pkt)

  3806. {

  3807.     const uint8_t *data = pkt->data;

  3808.     int size = pkt->size;

  3809.     VP9Context *s = ctx->priv_data;

  3810.     int res, tile_row, tile_col, i, ref, row, col;

  3811.     int retain_segmap_ref = s->segmentation.enabled && !s->segmentation.update_map;

  3812.     ptrdiff_t yoff, uvoff, ls_y, ls_uv;

  3813.     AVFrame *f;

  3814. 

  3815.     if ((res = decode_frame_header(ctx, data, size, &ref)) < 0) {

  3816.         return res;

  3817.     } else if (res == 0) {

  3818.         if (!s->refs[ref].f->data[0]) {

  3819.             av_log(ctx, AV_LOG_ERROR, "Requested reference %d not available\n", ref);

  3820.             return AVERROR_INVALIDDATA;

  3821.         }

  3822.         if ((res = av_frame_ref(frame, s->refs[ref].f)) < 0)

  3823.             return res;

  3824.         ((AVFrame *)frame)->pkt_pts = pkt->pts;

  3825.         ((AVFrame *)frame)->pkt_dts = pkt->dts;

  3826.         for (i = 0; i < 8; i++) {

  3827.             if (s->next_refs[i].f->data[0])

  3828.                 ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3829.             if (s->refs[i].f->data[0] &&

  3830.                 (res = ff_thread_ref_frame(&s->next_refs[i], &s->refs[i])) < 0)

  3831.                 return res;

  3832.         }

  3833.         *got_frame = 1;

  3834.         return pkt->size;

  3835.     }

  3836.     data += res;

  3837.     size -= res;

  3838. 

  3839.     if (!retain_segmap_ref) {

  3840.         if (s->frames[REF_FRAME_SEGMAP].tf.f->data[0])

  3841.             vp9_unref_frame(ctx, &s->frames[REF_FRAME_SEGMAP]);

  3842.         if (!s->keyframe && !s->intraonly && !s->errorres && s->frames[CUR_FRAME].tf.f->data[0] &&

  3843.             (res = vp9_ref_frame(ctx, &s->frames[REF_FRAME_SEGMAP], &s->frames[CUR_FRAME])) < 0)

  3844.             return res;

  3845.     }

  3846.     if (s->frames[REF_FRAME_MVPAIR].tf.f->data[0])

  3847.         vp9_unref_frame(ctx, &s->frames[REF_FRAME_MVPAIR]);

  3848.     if (!s->intraonly && !s->keyframe && !s->errorres && s->frames[CUR_FRAME].tf.f->data[0] &&

  3849.         (res = vp9_ref_frame(ctx, &s->frames[REF_FRAME_MVPAIR], &s->frames[CUR_FRAME])) < 0)

  3850.         return res;

  3851.     if (s->frames[CUR_FRAME].tf.f->data[0])

  3852.         vp9_unref_frame(ctx, &s->frames[CUR_FRAME]);

  3853.     if ((res = vp9_alloc_frame(ctx, &s->frames[CUR_FRAME])) < 0)

  3854.         return res;

  3855.     f = s->frames[CUR_FRAME].tf.f;

  3856.     f->key_frame = s->keyframe;

  3857.     f->pict_type = (s->keyframe || s->intraonly) ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;

  3858.     ls_y = f->linesize[0];

  3859.     ls_uv =f->linesize[1];

  3860. 

  3861.     // ref frame setup

  3862.     for (i = 0; i < 8; i++) {

  3863.         if (s->next_refs[i].f->data[0])

  3864.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3865.         if (s->refreshrefmask & (1 << i)) {

  3866.             res = ff_thread_ref_frame(&s->next_refs[i], &s->frames[CUR_FRAME].tf);

  3867.         } else if (s->refs[i].f->data[0]) {

  3868.             res = ff_thread_ref_frame(&s->next_refs[i], &s->refs[i]);

  3869.         }

  3870.         if (res < 0)

  3871.             return res;

  3872.     }

  3873. 

  3874.     // main tile decode loop

  3875.     memset(s->above_partition_ctx, 0, s->cols);

  3876.     memset(s->above_skip_ctx, 0, s->cols);

  3877.     if (s->keyframe || s->intraonly) {

  3878.         memset(s->above_mode_ctx, DC_PRED, s->cols * 2);

  3879.     } else {

  3880.         memset(s->above_mode_ctx, NEARESTMV, s->cols);

  3881.     }

  3882.     memset(s->above_y_nnz_ctx, 0, s->sb_cols * 16);

  3883.     memset(s->above_uv_nnz_ctx[0], 0, s->sb_cols * 16 >> s->ss_h);

  3884.     memset(s->above_uv_nnz_ctx[1], 0, s->sb_cols * 16 >> s->ss_h);

  3885.     memset(s->above_segpred_ctx, 0, s->cols);

  3886.     s->pass = s->frames[CUR_FRAME].uses_2pass =

  3887.         ctx->active_thread_type == FF_THREAD_FRAME && s->refreshctx && !s->parallelmode;

  3888.     if ((res = update_block_buffers(ctx)) < 0) {

  3889.         av_log(ctx, AV_LOG_ERROR,

  3890.                "Failed to allocate block buffers\n");

  3891.         return res;

  3892.     }

  3893.     if (s->refreshctx && s->parallelmode) {

  3894.         int j, k, l, m;

  3895. 

  3896.         for (i = 0; i < 4; i++) {

  3897.             for (j = 0; j < 2; j++)

  3898.                 for (k = 0; k < 2; k++)

  3899.                     for (l = 0; l < 6; l++)

  3900.                         for (m = 0; m < 6; m++)

  3901.                             memcpy(s->prob_ctx[s->framectxid].coef[i][j][k][l][m],

  3902.                                    s->prob.coef[i][j][k][l][m], 3);

  3903.             if (s->txfmmode == i)

  3904.                 break;

  3905.         }

  3906.         s->prob_ctx[s->framectxid].p = s->prob.p;

  3907.         ff_thread_finish_setup(ctx);

  3908.     } else if (!s->refreshctx) {

  3909.         ff_thread_finish_setup(ctx);

  3910.     }

  3911. 

  3912.     do {

  3913.         yoff = uvoff = 0;

  3914.         s->b = s->b_base;

  3915.         s->block = s->block_base;

  3916.         s->uvblock[0] = s->uvblock_base[0];

  3917.         s->uvblock[1] = s->uvblock_base[1];

  3918.         s->eob = s->eob_base;

  3919.         s->uveob[0] = s->uveob_base[0];

  3920.         s->uveob[1] = s->uveob_base[1];

  3921. 

  3922.         for (tile_row = 0; tile_row < s->tiling.tile_rows; tile_row++) {

  3923.             set_tile_offset(&s->tiling.tile_row_start, &s->tiling.tile_row_end,

  3924.                             tile_row, s->tiling.log2_tile_rows, s->sb_rows);

  3925.             if (s->pass != 2) {

  3926.                 for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  3927.                     unsigned tile_size;

  3928. 

  3929.                     if (tile_col == s->tiling.tile_cols - 1 &&

  3930.                         tile_row == s->tiling.tile_rows - 1) {

  3931.                         tile_size = size;

  3932.                     } else {

  3933.                         tile_size = AV_RB32(data);

  3934.                         data += 4;

  3935.                         size -= 4;

  3936.                     }

  3937.                     if (tile_size > size) {

  3938.                         ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  3939.                         return AVERROR_INVALIDDATA;

  3940.                     }

  3941.                     ff_vp56_init_range_decoder(&s->c_b[tile_col], data, tile_size);

  3942.                     if (vp56_rac_get_prob_branchy(&s->c_b[tile_col], 128)) { // marker bit

  3943.                         ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  3944.                         return AVERROR_INVALIDDATA;

  3945.                     }

  3946.                     data += tile_size;

  3947.                     size -= tile_size;

  3948.                 }

  3949.             }

  3950. 

  3951.             for (row = s->tiling.tile_row_start; row < s->tiling.tile_row_end;

  3952.                  row += 8, yoff += ls_y * 64, uvoff += ls_uv * 64 >> s->ss_v) {

  3953.                 struct VP9Filter *lflvl_ptr = s->lflvl;

  3954.                 ptrdiff_t yoff2 = yoff, uvoff2 = uvoff;

  3955. 

  3956.                 for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  3957.                     set_tile_offset(&s->tiling.tile_col_start, &s->tiling.tile_col_end,

  3958.                                     tile_col, s->tiling.log2_tile_cols, s->sb_cols);

  3959. 

  3960.                     if (s->pass != 2) {

  3961.                         memset(s->left_partition_ctx, 0, 8);

  3962.                         memset(s->left_skip_ctx, 0, 8);

  3963.                         if (s->keyframe || s->intraonly) {

  3964.                             memset(s->left_mode_ctx, DC_PRED, 16);

  3965.                         } else {

  3966.                             memset(s->left_mode_ctx, NEARESTMV, 8);

  3967.                         }

  3968.                         memset(s->left_y_nnz_ctx, 0, 16);

  3969.                         memset(s->left_uv_nnz_ctx, 0, 32);

  3970.                         memset(s->left_segpred_ctx, 0, 8);

  3971. 

  3972.                         memcpy(&s->c, &s->c_b[tile_col], sizeof(s->c));

  3973.                     }

  3974. 

  3975.                     for (col = s->tiling.tile_col_start;

  3976.                          col < s->tiling.tile_col_end;

  3977.                          col += 8, yoff2 += 64, uvoff2 += 64 >> s->ss_h, lflvl_ptr++) {

  3978.                         // FIXME integrate with lf code (i.e. zero after each

  3979.                         // use, similar to invtxfm coefficients, or similar)

  3980.                         if (s->pass != 1) {

  3981.                             memset(lflvl_ptr->mask, 0, sizeof(lflvl_ptr->mask));

  3982.                         }

  3983. 

  3984.                         if (s->pass == 2) {

  3985.                             decode_sb_mem(ctx, row, col, lflvl_ptr,

  3986.                                           yoff2, uvoff2, BL_64X64);

  3987.                         } else {

  3988.                             decode_sb(ctx, row, col, lflvl_ptr,

  3989.                                       yoff2, uvoff2, BL_64X64);

  3990.                         }

  3991.                     }

  3992.                     if (s->pass != 2) {

  3993.                         memcpy(&s->c_b[tile_col], &s->c, sizeof(s->c));

  3994.                     }

  3995.                 }

  3996. 

  3997.                 if (s->pass == 1) {

  3998.                     continue;

  3999.                 }

  4000. 

  4001.                 // backup pre-loopfilter reconstruction data for intra

  4002.                 // prediction of next row of sb64s

  4003.                 if (row + 8 < s->rows) {

  4004.                     memcpy(s->intra_pred_data[0],

  4005.                            f->data[0] + yoff + 63 * ls_y,

  4006.                            8 * s->cols);

  4007.                     memcpy(s->intra_pred_data[1],

  4008.                            f->data[1] + uvoff + ((64 >> s->ss_v) - 1) * ls_uv,

  4009.                            8 * s->cols >> s->ss_h);

  4010.                     memcpy(s->intra_pred_data[2],

  4011.                            f->data[2] + uvoff + ((64 >> s->ss_v) - 1) * ls_uv,

  4012.                            8 * s->cols >> s->ss_h);

  4013.                 }

  4014. 

  4015.                 // loopfilter one row

  4016.                 if (s->filter.level) {

  4017.                     yoff2 = yoff;

  4018.                     uvoff2 = uvoff;

  4019.                     lflvl_ptr = s->lflvl;

  4020.                     for (col = 0; col < s->cols;

  4021.                          col += 8, yoff2 += 64, uvoff2 += 64 >> s->ss_h, lflvl_ptr++) {

  4022.                         loopfilter_sb(ctx, lflvl_ptr, row, col, yoff2, uvoff2);

  4023.                     }

  4024.                 }

  4025. 

  4026.                 // FIXME maybe we can make this more finegrained by running the

  4027.                 // loopfilter per-block instead of after each sbrow

  4028.                 // In fact that would also make intra pred left preparation easier?

  4029.                 ff_thread_report_progress(&s->frames[CUR_FRAME].tf, row >> 3, 0);

  4030.             }

  4031.         }

  4032. 

  4033.         if (s->pass < 2 && s->refreshctx && !s->parallelmode) {

  4034.             adapt_probs(s);

  4035.             ff_thread_finish_setup(ctx);

  4036.         }

  4037.     } while (s->pass++ == 1);

  4038.     ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  4039. 

  4040.     // ref frame setup

  4041.     for (i = 0; i < 8; i++) {

  4042.         if (s->refs[i].f->data[0])

  4043.             ff_thread_release_buffer(ctx, &s->refs[i]);

  4044.         ff_thread_ref_frame(&s->refs[i], &s->next_refs[i]);

  4045.     }

  4046. 

  4047.     if (!s->invisible) {

  4048.         if ((res = av_frame_ref(frame, s->frames[CUR_FRAME].tf.f)) < 0)

  4049.             return res;

  4050.         *got_frame = 1;

  4051.     }

  4052. 

  4053.     return pkt->size;

  4054. }

  4055. 

  4056. static void vp9_decode_flush(AVCodecContext *ctx)

  4057. {

  4058.     VP9Context *s = ctx->priv_data;

  4059.     int i;

  4060. 

  4061.     for (i = 0; i < 3; i++)

  4062.         vp9_unref_frame(ctx, &s->frames[i]);

  4063.     for (i = 0; i < 8; i++)

  4064.         ff_thread_release_buffer(ctx, &s->refs[i]);

  4065. }

  4066. 

  4067. static int init_frames(AVCodecContext *ctx)

  4068. {

  4069.     VP9Context *s = ctx->priv_data;

  4070.     int i;

  4071. 

  4072.     for (i = 0; i < 3; i++) {

  4073.         s->frames[i].tf.f = av_frame_alloc();

  4074.         if (!s->frames[i].tf.f) {

  4075.             vp9_decode_free(ctx);

  4076.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  4077.             return AVERROR(ENOMEM);

  4078.         }

  4079.     }

  4080.     for (i = 0; i < 8; i++) {

  4081.         s->refs[i].f = av_frame_alloc();

  4082.         s->next_refs[i].f = av_frame_alloc();

  4083.         if (!s->refs[i].f || !s->next_refs[i].f) {

  4084.             vp9_decode_free(ctx);

  4085.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  4086.             return AVERROR(ENOMEM);

  4087.         }

  4088.     }

  4089. 

  4090.     return 0;

  4091. }

  4092. 

  4093. static av_cold int vp9_decode_init(AVCodecContext *ctx)

  4094. {

  4095.     VP9Context *s = ctx->priv_data;

  4096. 

  4097.     ctx->internal->allocate_progress = 1;

  4098.     ff_vp9dsp_init(&s->dsp);

  4099.     ff_videodsp_init(&s->vdsp, 8);

  4100.     s->filter.sharpness = -1;

  4101. 

  4102.     return init_frames(ctx);

  4103. }

  4104. 

  4105. static av_cold int vp9_decode_init_thread_copy(AVCodecContext *avctx)

  4106. {

  4107.     return init_frames(avctx);

  4108. }

  4109. 

  4110. static int vp9_decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src)

  4111. {

  4112.     int i, res;

  4113.     VP9Context *s = dst->priv_data, *ssrc = src->priv_data;

  4114. 

  4115.     // detect size changes in other threads

  4116.     if (s->intra_pred_data[0] &&

  4117.         (!ssrc->intra_pred_data[0] || s->cols != ssrc->cols || s->rows != ssrc->rows)) {

  4118.         free_buffers(s);

  4119.     }

  4120. 

  4121.     for (i = 0; i < 3; i++) {

  4122.         if (s->frames[i].tf.f->data[0])

  4123.             vp9_unref_frame(dst, &s->frames[i]);

  4124.         if (ssrc->frames[i].tf.f->data[0]) {

  4125.             if ((res = vp9_ref_frame(dst, &s->frames[i], &ssrc->frames[i])) < 0)

  4126.                 return res;

  4127.         }

  4128.     }

  4129.     for (i = 0; i < 8; i++) {

  4130.         if (s->refs[i].f->data[0])

  4131.             ff_thread_release_buffer(dst, &s->refs[i]);

  4132.         if (ssrc->next_refs[i].f->data[0]) {

  4133.             if ((res = ff_thread_ref_frame(&s->refs[i], &ssrc->next_refs[i])) < 0)

  4134.                 return res;

  4135.         }

  4136.     }

  4137. 

  4138.     s->invisible = ssrc->invisible;

  4139.     s->keyframe = ssrc->keyframe;

  4140.     s->ss_v = ssrc->ss_v;

  4141.     s->ss_h = ssrc->ss_h;

  4142.     s->segmentation.enabled = ssrc->segmentation.enabled;

  4143.     s->segmentation.update_map = ssrc->segmentation.update_map;

  4144.     memcpy(&s->prob_ctx, &ssrc->prob_ctx, sizeof(s->prob_ctx));

  4145.     memcpy(&s->lf_delta, &ssrc->lf_delta, sizeof(s->lf_delta));

  4146.     if (ssrc->segmentation.enabled) {

  4147.         memcpy(&s->segmentation.feat, &ssrc->segmentation.feat,

  4148.                sizeof(s->segmentation.feat));

  4149.     }

  4150. 

  4151.     return 0;

  4152. }

  4153. 

  4154. AVCodec ff_vp9_decoder = {

  4155.     .name                  = "vp9",

  4156.     .long_name             = NULL_IF_CONFIG_SMALL("Google VP9"),

  4157.     .type                  = AVMEDIA_TYPE_VIDEO,

  4158.     .id                    = AV_CODEC_ID_VP9,

  4159.     .priv_data_size        = sizeof(VP9Context),

  4160.     .init                  = vp9_decode_init,

  4161.     .close                 = vp9_decode_free,

  4162.     .decode                = vp9_decode_frame,

  4163.     .capabilities          = CODEC_CAP_DR1 | CODEC_CAP_FRAME_THREADS,

  4164.     .flush                 = vp9_decode_flush,

  4165.     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(vp9_decode_init_thread_copy),

  4166.     .update_thread_context = ONLY_IF_THREADS_ENABLED(vp9_decode_update_thread_context),

  4167. };