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.
60425 Commits
32 Branches
293MB
Tree: bbc3425fa2
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'bbc3425fa2'
${ noResults }
FFmpeg/libavcodec/vp9.c

3837 lines
152KB
Raw Blame History 

  1. /*

  2.  * VP9 compatible video decoder

  3.  *

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

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

  6.  *

  7.  * This file is part of FFmpeg.

  8.  *

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

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

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

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

  13.  *

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

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

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

  17.  * Lesser General Public License for more details.

  18.  *

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

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

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

  22.  */

  23. 

  24. #include "avcodec.h"

  25. #include "get_bits.h"

  26. #include "internal.h"

  27. #include "thread.h"

  28. #include "videodsp.h"

  29. #include "vp56.h"

  30. #include "vp9.h"

  31. #include "vp9data.h"

  32. #include "vp9dsp.h"

  33. #include "libavutil/avassert.h"

  34. 

  35. #define VP9_SYNCCODE 0x498342

  36. 

  37. enum CompPredMode {

  38.     PRED_SINGLEREF,

  39.     PRED_COMPREF,

  40.     PRED_SWITCHABLE,

  41. };

  42. 

  43. enum BlockLevel {

  44.     BL_64X64,

  45.     BL_32X32,

  46.     BL_16X16,

  47.     BL_8X8,

  48. };

  49. 

  50. enum BlockSize {

  51.     BS_64x64,

  52.     BS_64x32,

  53.     BS_32x64,

  54.     BS_32x32,

  55.     BS_32x16,

  56.     BS_16x32,

  57.     BS_16x16,

  58.     BS_16x8,

  59.     BS_8x16,

  60.     BS_8x8,

  61.     BS_8x4,

  62.     BS_4x8,

  63.     BS_4x4,

  64.     N_BS_SIZES,

  65. };

  66. 

  67. struct VP9mvrefPair {

  68.     VP56mv mv[2];

  69.     int8_t ref[2];

  70. };

  71. 

  72. typedef struct VP9Frame {

  73.     ThreadFrame tf;

  74.     AVBufferRef *extradata;

  75.     uint8_t *segmentation_map;

  76.     struct VP9mvrefPair *mv;

  77. } VP9Frame;

  78. 

  79. struct VP9Filter {

  80.     uint8_t level[8 * 8];

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

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

  83. };

  84. 

  85. typedef struct VP9Block {

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

  87.     enum FilterMode filter;

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

  89.     enum BlockSize bs;

  90.     enum TxfmMode tx, uvtx;

  91.     enum BlockLevel bl;

  92.     enum BlockPartition bp;

  93. } VP9Block;

  94. 

  95. typedef struct VP9Context {

  96.     VP9DSPContext dsp;

  97.     VideoDSPContext vdsp;

  98.     GetBitContext gb;

  99.     VP56RangeCoder c;

  100.     VP56RangeCoder *c_b;

  101.     unsigned c_b_size;

  102.     VP9Block *b_base, *b;

  103.     int pass, uses_2pass, last_uses_2pass;

  104.     int row, row7, col, col7;

  105.     uint8_t *dst[3];

  106.     ptrdiff_t y_stride, uv_stride;

  107. 

  108.     // bitstream header

  109.     uint8_t profile;

  110.     uint8_t keyframe, last_keyframe;

  111.     uint8_t invisible;

  112.     uint8_t use_last_frame_mvs;

  113.     uint8_t errorres;

  114.     uint8_t colorspace;

  115.     uint8_t fullrange;

  116.     uint8_t intraonly;

  117.     uint8_t resetctx;

  118.     uint8_t refreshrefmask;

  119.     uint8_t highprecisionmvs;

  120.     enum FilterMode filtermode;

  121.     uint8_t allowcompinter;

  122.     uint8_t fixcompref;

  123.     uint8_t refreshctx;

  124.     uint8_t parallelmode;

  125.     uint8_t framectxid;

  126.     uint8_t refidx[3];

  127.     uint8_t signbias[3];

  128.     uint8_t varcompref[2];

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

  130. #define CUR_FRAME 0

  131. #define LAST_FRAME 1

  132.     VP9Frame frames[2];

  133. 

  134.     struct {

  135.         uint8_t level;

  136.         int8_t sharpness;

  137.         uint8_t lim_lut[64];

  138.         uint8_t mblim_lut[64];

  139.     } filter;

  140.     struct {

  141.         uint8_t enabled;

  142.         int8_t mode[2];

  143.         int8_t ref[4];

  144.     } lf_delta;

  145.     uint8_t yac_qi;

  146.     int8_t ydc_qdelta, uvdc_qdelta, uvac_qdelta;

  147.     uint8_t lossless;

  148.     struct {

  149.         uint8_t enabled;

  150.         uint8_t temporal;

  151.         uint8_t absolute_vals;

  152.         uint8_t update_map;

  153.         struct {

  154.             uint8_t q_enabled;

  155.             uint8_t lf_enabled;

  156.             uint8_t ref_enabled;

  157.             uint8_t skip_enabled;

  158.             uint8_t ref_val;

  159.             int16_t q_val;

  160.             int8_t lf_val;

  161.             int16_t qmul[2][2];

  162.             uint8_t lflvl[4][2];

  163.         } feat[8];

  164.     } segmentation;

  165.     struct {

  166.         unsigned log2_tile_cols, log2_tile_rows;

  167.         unsigned tile_cols, tile_rows;

  168.         unsigned tile_row_start, tile_row_end, tile_col_start, tile_col_end;

  169.     } tiling;

  170.     unsigned sb_cols, sb_rows, rows, cols;

  171.     struct {

  172.         prob_context p;

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

  174.     } prob_ctx[4];

  175.     struct {

  176.         prob_context p;

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

  178.         uint8_t seg[7];

  179.         uint8_t segpred[3];

  180.     } prob;

  181.     struct {

  182.         unsigned y_mode[4][10];

  183.         unsigned uv_mode[10][10];

  184.         unsigned filter[4][3];

  185.         unsigned mv_mode[7][4];

  186.         unsigned intra[4][2];

  187.         unsigned comp[5][2];

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

  189.         unsigned comp_ref[5][2];

  190.         unsigned tx32p[2][4];

  191.         unsigned tx16p[2][3];

  192.         unsigned tx8p[2][2];

  193.         unsigned skip[3][2];

  194.         unsigned mv_joint[4];

  195.         struct {

  196.             unsigned sign[2];

  197.             unsigned classes[11];

  198.             unsigned class0[2];

  199.             unsigned bits[10][2];

  200.             unsigned class0_fp[2][4];

  201.             unsigned fp[4];

  202.             unsigned class0_hp[2];

  203.             unsigned hp[2];

  204.         } mv_comp[2];

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

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

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

  208.     } counts;

  209.     enum TxfmMode txfmmode;

  210.     enum CompPredMode comppredmode;

  211. 

  212.     // contextual (left/above) cache

  213.     uint8_t left_partition_ctx[8], *above_partition_ctx;

  214.     uint8_t left_mode_ctx[16], *above_mode_ctx;

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

  216.     uint8_t left_y_nnz_ctx[16], *above_y_nnz_ctx;

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

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

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

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

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

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

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

  224.     uint8_t left_filter_ctx[8], *above_filter_ctx;

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

  226. 

  227.     // whole-frame cache

  228.     uint8_t *intra_pred_data[3];

  229.     struct VP9Filter *lflvl;

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

  231. 

  232.     // block reconstruction intermediates

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

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

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

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

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

  238. } VP9Context;

  239. 

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

  241.     {

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

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

  244.     }, {

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

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

  247.     }

  248. };

  249. 

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

  251. {

  252.     VP9Context *s = ctx->priv_data;

  253.     int ret, sz;

  254. 

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

  256.         return ret;

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

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

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

  260.         return AVERROR(ENOMEM);

  261.     }

  262. 

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

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

  265. 

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

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

  268.         !s->keyframe && !s->intraonly) {

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

  270.     }

  271. 

  272.     return 0;

  273. }

  274. 

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

  276. {

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

  278.     av_buffer_unref(&f->extradata);

  279. }

  280. 

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

  282. {

  283.     int res;

  284. 

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

  286.         return res;

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

  288.         vp9_unref_frame(ctx, dst);

  289.         return AVERROR(ENOMEM);

  290.     }

  291. 

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

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

  294. 

  295.     return 0;

  296. }

  297. 

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

  299. {

  300.     VP9Context *s = ctx->priv_data;

  301.     uint8_t *p;

  302. 

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

  304. 

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

  306.         return 0;

  307. 

  308.     ctx->width  = w;

  309.     ctx->height = h;

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

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

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

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

  314. 

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

  316.     av_freep(&s->above_partition_ctx);

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

  318.     if (!p)

  319.         return AVERROR(ENOMEM);

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  337. #undef assign

  338. 

  339.     av_free(s->b_base);

  340.     av_free(s->block_base);

  341.     if (ctx->active_thread_type == FF_THREAD_FRAME && s->refreshctx && !s->parallelmode) {

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

  343. 

  344.         s->b_base = av_malloc(sizeof(VP9Block) * s->cols * s->rows);

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

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

  347.             return AVERROR(ENOMEM);

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

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

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

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

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

  353.     } else {

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

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

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

  357.             return AVERROR(ENOMEM);

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

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

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

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

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

  363.     }

  364. 

  365.     return 0;

  366. }

  367. 

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

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

  370. {

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

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

  373. }

  374. 

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

  376. {

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

  378. }

  379. 

  380. // differential forward probability updates

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

  382. {

  383.     static const int inv_map_table[254] = {

  384.           7,  20,  33,  46,  59,  72,  85,  98, 111, 124, 137, 150, 163, 176,

  385.         189, 202, 215, 228, 241, 254,   1,   2,   3,   4,   5,   6,   8,   9,

  386.          10,  11,  12,  13,  14,  15,  16,  17,  18,  19,  21,  22,  23,  24,

  387.          25,  26,  27,  28,  29,  30,  31,  32,  34,  35,  36,  37,  38,  39,

  388.          40,  41,  42,  43,  44,  45,  47,  48,  49,  50,  51,  52,  53,  54,

  389.          55,  56,  57,  58,  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,

  390.          70,  71,  73,  74,  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,

  391.          86,  87,  88,  89,  90,  91,  92,  93,  94,  95,  96,  97,  99, 100,

  392.         101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 112, 113, 114, 115,

  393.         116, 117, 118, 119, 120, 121, 122, 123, 125, 126, 127, 128, 129, 130,

  394.         131, 132, 133, 134, 135, 136, 138, 139, 140, 141, 142, 143, 144, 145,

  395.         146, 147, 148, 149, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160,

  396.         161, 162, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175,

  397.         177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 190, 191,

  398.         192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 204, 205, 206,

  399.         207, 208, 209, 210, 211, 212, 213, 214, 216, 217, 218, 219, 220, 221,

  400.         222, 223, 224, 225, 226, 227, 229, 230, 231, 232, 233, 234, 235, 236,

  401.         237, 238, 239, 240, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251,

  402.         252, 253,

  403.     };

  404.     int d;

  405. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  420. 

  421.     if (!vp8_rac_get(c)) {

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

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

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

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

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

  427.     } else {

  428.         d = vp8_rac_get_uint(c, 7);

  429.         if (d >= 65)

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

  431.         d += 64;

  432.     }

  433. 

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

  435.                     255 - inv_recenter_nonneg(inv_map_table[d], 255 - p);

  436. }

  437. 

  438. static int decode_frame_header(AVCodecContext *ctx,

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

  440. {

  441.     VP9Context *s = ctx->priv_data;

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

  443.     int last_invisible;

  444.     const uint8_t *data2;

  445. 

  446.     /* general header */

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

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

  449.         return res;

  450.     }

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

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

  453.         return AVERROR_INVALIDDATA;

  454.     }

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

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

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

  458.         return AVERROR_INVALIDDATA;

  459.     }

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

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

  462.         return 0;

  463.     }

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

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

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

  467.     last_invisible    = s->invisible;

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

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

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

  471.     if (s->keyframe) {

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

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

  474.             return AVERROR_INVALIDDATA;

  475.         }

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

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

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

  479.             return AVERROR_INVALIDDATA;

  480.         }

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

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

  483.         s->refreshrefmask = 0xff;

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

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

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

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

  488.     } else {

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

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

  491.         if (s->intraonly) {

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

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

  494.                 return AVERROR_INVALIDDATA;

  495.             }

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

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

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

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

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

  501.         } else {

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

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

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

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

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

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

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

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

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

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

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

  513.                 return AVERROR_INVALIDDATA;

  514.             }

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

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

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

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

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

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

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

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

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

  524.             } else {

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

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

  527.             }

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

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

  530.             // the _last_ frame

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

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

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

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

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

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

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

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

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

  540.             if (s->allowcompinter) {

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

  542.                     s->fixcompref    = 2;

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

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

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

  546.                     s->fixcompref    = 1;

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

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

  549.                 } else {

  550.                     s->fixcompref    = 0;

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

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

  553.                 }

  554.             }

  555.         }

  556.     }

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

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

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

  560. 

  561.     /* loopfilter header data */

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

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

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

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

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

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

  568.     s->filter.sharpness = sharp;

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

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

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

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

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

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

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

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

  577.         }

  578.     } else {

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

  580.     }

  581. 

  582.     /* quantization header data */

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

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

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

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

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

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

  589. 

  590.     /* segmentation header info */

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

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

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

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

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

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

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

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

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

  600.         }

  601. 

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

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

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

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

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

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

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

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

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

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

  612.             }

  613.         }

  614.     } else {

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

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

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

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

  619.     }

  620. 

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

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

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

  624. 

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

  626.             if (s->segmentation.absolute_vals)

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

  628.             else

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

  630.         } else {

  631.             qyac  = s->yac_qi;

  632.         }

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

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

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

  636.         qyac  = av_clip_uintp2(qyac, 8);

  637. 

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

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

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

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

  642. 

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

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

  645.             if (s->segmentation.absolute_vals)

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

  647.             else

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

  649.         } else {

  650.             lflvl  = s->filter.level;

  651.         }

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

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

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

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

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

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

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

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

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

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

  662.         }

  663.     }

  664. 

  665.     /* tiling info */

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

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

  668.         return res;

  669.     }

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

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

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

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

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

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

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

  677.             s->tiling.log2_tile_cols++;

  678.         else

  679.             break;

  680.     }

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

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

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

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

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

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

  687.         if (!s->c_b) {

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

  689.             return AVERROR(ENOMEM);

  690.         }

  691.     }

  692. 

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

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

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

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

  697.                sizeof(vp9_default_coef_probs));

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

  699.                sizeof(vp9_default_coef_probs));

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

  701.                sizeof(vp9_default_coef_probs));

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

  703.                sizeof(vp9_default_coef_probs));

  704.     }

  705. 

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

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

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

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

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

  711.         return AVERROR_INVALIDDATA;

  712.     }

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

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

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

  716.         return AVERROR_INVALIDDATA;

  717.     }

  718. 

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

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

  721.     } else {

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

  723.     }

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

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

  726.     // fw update)?

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

  728. 

  729.     // txfm updates

  730.     if (s->lossless) {

  731.         s->txfmmode = TX_4X4;

  732.     } else {

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

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

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

  736. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  751.         }

  752.     }

  753. 

  754.     // coef updates

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

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

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

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

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

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

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

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

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

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

  765.                                 break;

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

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

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

  769.                                 } else {

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

  771.                                 }

  772.                             }

  773.                             p[3] = 0;

  774.                         }

  775.         } else {

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

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

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

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

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

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

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

  783.                                 break;

  784.                             memcpy(p, r, 3);

  785.                             p[3] = 0;

  786.                         }

  787.         }

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

  789.             break;

  790.     }

  791. 

  792.     // mode updates

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

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

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

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

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

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

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

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

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

  802. 

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

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

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

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

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

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

  809. 

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

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

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

  813. 

  814.         if (s->allowcompinter) {

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

  816.             if (s->comppredmode)

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

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

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

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

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

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

  823.         } else {

  824.             s->comppredmode = PRED_SINGLEREF;

  825.         }

  826. 

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

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

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

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

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

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

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

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

  835.             }

  836.         }

  837. 

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

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

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

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

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

  843.         }

  844. 

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

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

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

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

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

  850. 

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

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

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

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

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

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

  857. 

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

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

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

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

  862. 

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

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

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

  866. 

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

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

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

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

  871. 

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

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

  874. 

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

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

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

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

  879.         }

  880. 

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

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

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

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

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

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

  887. 

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

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

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

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

  892.         }

  893. 

  894.         if (s->highprecisionmvs) {

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

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

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

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

  899. 

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

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

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

  903.             }

  904.         }

  905.     }

  906. 

  907.     return (data2 - data) + size2;

  908. }

  909. 

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

  911.                                       VP9Context *s)

  912. {

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

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

  915. }

  916. 

  917. static void find_ref_mvs(VP9Context *s,

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

  919. {

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

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

  922.                       { -1, -1 }, {  0, -1 }, { -1,  0 }, {  6, -1 }},

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

  924.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, {  2, -1 }},

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

  926.                       { -1, -1 }, { -3,  0 }, {  0, -3 }, { -1,  2 }},

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

  928.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  930.                       { -1,  1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  932.                       {  1, -1 }, { -3,  0 }, {  0, -3 }, { -3, -3 }},

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

  934.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  936.                       {  0, -2 }, { -2,  0 }, { -2, -1 }, { -1, -2 }},

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

  938.                       { -2,  0 }, {  0, -2 }, { -1, -2 }, { -2, -1 }},

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

  940.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

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

  942.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

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

  944.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

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

  946.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  947.     };

  948.     VP9Block *b = s->b;

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

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

  951. #define INVALID_MV 0x80008000U

  952.     uint32_t mem = INVALID_MV;

  953.     int i;

  954. 

  955. #define RETURN_DIRECT_MV(mv) \

  956.     do { \

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

  958.         if (!idx) { \

  959.             AV_WN32A(pmv, m); \

  960.             return; \

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

  962.             mem = m; \

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

  964.             AV_WN32A(pmv, m); \

  965.             return; \

  966.         } \

  967.     } while (0)

  968. 

  969.     if (sb >= 0) {

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

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

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

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

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

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

  976.         }

  977. 

  978. #define RETURN_MV(mv) \

  979.     do { \

  980.         if (sb > 0) { \

  981.             VP56mv tmp; \

  982.             uint32_t m; \

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

  984.             m = AV_RN32A(&tmp); \

  985.             if (!idx) { \

  986.                 AV_WN32A(pmv, m); \

  987.                 return; \

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

  989.                 mem = m; \

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

  991.                 AV_WN32A(pmv, m); \

  992.                 return; \

  993.             } \

  994.         } else { \

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

  996.             if (!idx) { \

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

  998.                 return; \

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

  1000.                 mem = m; \

  1001.             } else if (m != mem) { \

  1002.                 clamp_mv(pmv, &mv, s); \

  1003.                 return; \

  1004.             } \

  1005.         } \

  1006.     } while (0)

  1007. 

  1008.         if (row > 0) {

  1009.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[(row - 1) * s->sb_cols * 8 + col];

  1010.             if (mv->ref[0] == ref) {

  1011.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][0]);

  1012.             } else if (mv->ref[1] == ref) {

  1013.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][1]);

  1014.             }

  1015.         }

  1016.         if (col > s->tiling.tile_col_start) {

  1017.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[row * s->sb_cols * 8 + col - 1];

  1018.             if (mv->ref[0] == ref) {

  1019.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][0]);

  1020.             } else if (mv->ref[1] == ref) {

  1021.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][1]);

  1022.             }

  1023.         }

  1024.         i = 2;

  1025.     } else {

  1026.         i = 0;

  1027.     }

  1028. 

  1029.     // previously coded MVs in this neighbourhood, using same reference frame

  1030.     for (; i < 8; i++) {

  1031.         int c = p[i][0] + col, r = p[i][1] + row;

  1032. 

  1033.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1034.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1035. 

  1036.             if (mv->ref[0] == ref) {

  1037.                 RETURN_MV(mv->mv[0]);

  1038.             } else if (mv->ref[1] == ref) {

  1039.                 RETURN_MV(mv->mv[1]);

  1040.             }

  1041.         }

  1042.     }

  1043. 

  1044.     // MV at this position in previous frame, using same reference frame

  1045.     if (s->use_last_frame_mvs) {

  1046.         struct VP9mvrefPair *mv = &s->frames[LAST_FRAME].mv[row * s->sb_cols * 8 + col];

  1047. 

  1048.         if (!s->last_uses_2pass)

  1049.             ff_thread_await_progress(&s->frames[LAST_FRAME].tf, row >> 3, 0);

  1050.         if (mv->ref[0] == ref) {

  1051.             RETURN_MV(mv->mv[0]);

  1052.         } else if (mv->ref[1] == ref) {

  1053.             RETURN_MV(mv->mv[1]);

  1054.         }

  1055.     }

  1056. 

  1057. #define RETURN_SCALE_MV(mv, scale) \

  1058.     do { \

  1059.         if (scale) { \

  1060.             VP56mv mv_temp = { -mv.x, -mv.y }; \

  1061.             RETURN_MV(mv_temp); \

  1062.         } else { \

  1063.             RETURN_MV(mv); \

  1064.         } \

  1065.     } while (0)

  1066. 

  1067.     // previously coded MVs in this neighbourhood, using different reference frame

  1068.     for (i = 0; i < 8; i++) {

  1069.         int c = p[i][0] + col, r = p[i][1] + row;

  1070. 

  1071.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1072.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1073. 

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

  1075.                 RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1076.             }

  1077.             if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1078.                 // BUG - libvpx has this condition regardless of whether

  1079.                 // we used the first ref MV and pre-scaling

  1080.                 AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1081.                 RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1082.             }

  1083.         }

  1084.     }

  1085. 

  1086.     // MV at this position in previous frame, using different reference frame

  1087.     if (s->use_last_frame_mvs) {

  1088.         struct VP9mvrefPair *mv = &s->frames[LAST_FRAME].mv[row * s->sb_cols * 8 + col];

  1089. 

  1090.         // no need to await_progress, because we already did that above

  1091.         if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1092.             RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1093.         }

  1094.         if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1095.             // BUG - libvpx has this condition regardless of whether

  1096.             // we used the first ref MV and pre-scaling

  1097.             AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1098.             RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1099.         }

  1100.     }

  1101. 

  1102.     AV_ZERO32(pmv);

  1103. #undef INVALID_MV

  1104. #undef RETURN_MV

  1105. #undef RETURN_SCALE_MV

  1106. }

  1107. 

  1108. static av_always_inline int read_mv_component(VP9Context *s, int idx, int hp)

  1109. {

  1110.     int bit, sign = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].sign);

  1111.     int n, c = vp8_rac_get_tree(&s->c, vp9_mv_class_tree,

  1112.                                 s->prob.p.mv_comp[idx].classes);

  1113. 

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

  1115.     s->counts.mv_comp[idx].classes[c]++;

  1116.     if (c) {

  1117.         int m;

  1118. 

  1119.         for (n = 0, m = 0; m < c; m++) {

  1120.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].bits[m]);

  1121.             n |= bit << m;

  1122.             s->counts.mv_comp[idx].bits[m][bit]++;

  1123.         }

  1124.         n <<= 3;

  1125.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree, s->prob.p.mv_comp[idx].fp);

  1126.         n |= bit << 1;

  1127.         s->counts.mv_comp[idx].fp[bit]++;

  1128.         if (hp) {

  1129.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].hp);

  1130.             s->counts.mv_comp[idx].hp[bit]++;

  1131.             n |= bit;

  1132.         } else {

  1133.             n |= 1;

  1134.             // bug in libvpx - we count for bw entropy purposes even if the

  1135.             // bit wasn't coded

  1136.             s->counts.mv_comp[idx].hp[1]++;

  1137.         }

  1138.         n += 8 << c;

  1139.     } else {

  1140.         n = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0);

  1141.         s->counts.mv_comp[idx].class0[n]++;

  1142.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree,

  1143.                                s->prob.p.mv_comp[idx].class0_fp[n]);

  1144.         s->counts.mv_comp[idx].class0_fp[n][bit]++;

  1145.         n = (n << 3) | (bit << 1);

  1146.         if (hp) {

  1147.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0_hp);

  1148.             s->counts.mv_comp[idx].class0_hp[bit]++;

  1149.             n |= bit;

  1150.         } else {

  1151.             n |= 1;

  1152.             // bug in libvpx - we count for bw entropy purposes even if the

  1153.             // bit wasn't coded

  1154.             s->counts.mv_comp[idx].class0_hp[1]++;

  1155.         }

  1156.     }

  1157. 

  1158.     return sign ? -(n + 1) : (n + 1);

  1159. }

  1160. 

  1161. static void fill_mv(VP9Context *s,

  1162.                     VP56mv *mv, int mode, int sb)

  1163. {

  1164.     VP9Block *b = s->b;

  1165. 

  1166.     if (mode == ZEROMV) {

  1167.         memset(mv, 0, sizeof(*mv) * 2);

  1168.     } else {

  1169.         int hp;

  1170. 

  1171.         // FIXME cache this value and reuse for other subblocks

  1172.         find_ref_mvs(s, &mv[0], b->ref[0], 0, mode == NEARMV,

  1173.                      mode == NEWMV ? -1 : sb);

  1174.         // FIXME maybe move this code into find_ref_mvs()

  1175.         if ((mode == NEWMV || sb == -1) &&

  1176.             !(hp = s->highprecisionmvs && abs(mv[0].x) < 64 && abs(mv[0].y) < 64)) {

  1177.             if (mv[0].y & 1) {

  1178.                 if (mv[0].y < 0)

  1179.                     mv[0].y++;

  1180.                 else

  1181.                     mv[0].y--;

  1182.             }

  1183.             if (mv[0].x & 1) {

  1184.                 if (mv[0].x < 0)

  1185.                     mv[0].x++;

  1186.                 else

  1187.                     mv[0].x--;

  1188.             }

  1189.         }

  1190.         if (mode == NEWMV) {

  1191.             enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1192.                                               s->prob.p.mv_joint);

  1193. 

  1194.             s->counts.mv_joint[j]++;

  1195.             if (j >= MV_JOINT_V)

  1196.                 mv[0].y += read_mv_component(s, 0, hp);

  1197.             if (j & 1)

  1198.                 mv[0].x += read_mv_component(s, 1, hp);

  1199.         }

  1200. 

  1201.         if (b->comp) {

  1202.             // FIXME cache this value and reuse for other subblocks

  1203.             find_ref_mvs(s, &mv[1], b->ref[1], 1, mode == NEARMV,

  1204.                          mode == NEWMV ? -1 : sb);

  1205.             if ((mode == NEWMV || sb == -1) &&

  1206.                 !(hp = s->highprecisionmvs && abs(mv[1].x) < 64 && abs(mv[1].y) < 64)) {

  1207.                 if (mv[1].y & 1) {

  1208.                     if (mv[1].y < 0)

  1209.                         mv[1].y++;

  1210.                     else

  1211.                         mv[1].y--;

  1212.                 }

  1213.                 if (mv[1].x & 1) {

  1214.                     if (mv[1].x < 0)

  1215.                         mv[1].x++;

  1216.                     else

  1217.                         mv[1].x--;

  1218.                 }

  1219.             }

  1220.             if (mode == NEWMV) {

  1221.                 enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1222.                                                   s->prob.p.mv_joint);

  1223. 

  1224.                 s->counts.mv_joint[j]++;

  1225.                 if (j >= MV_JOINT_V)

  1226.                     mv[1].y += read_mv_component(s, 0, hp);

  1227.                 if (j & 1)

  1228.                     mv[1].x += read_mv_component(s, 1, hp);

  1229.             }

  1230.         }

  1231.     }

  1232. }

  1233. 

  1234. static void decode_mode(AVCodecContext *ctx)

  1235. {

  1236.     static const uint8_t left_ctx[N_BS_SIZES] = {

  1237.         0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf

  1238.     };

  1239.     static const uint8_t above_ctx[N_BS_SIZES] = {

  1240.         0x0, 0x0, 0x8, 0x8, 0x8, 0xc, 0xc, 0xc, 0xe, 0xe, 0xe, 0xf, 0xf

  1241.     };

  1242.     static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {

  1243.         TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,

  1244.         TX_16X16, TX_8X8, TX_8X8, TX_8X8, TX_4X4, TX_4X4, TX_4X4

  1245.     };

  1246.     VP9Context *s = ctx->priv_data;

  1247.     VP9Block *b = s->b;

  1248.     int row = s->row, col = s->col, row7 = s->row7;

  1249.     enum TxfmMode max_tx = max_tx_for_bl_bp[b->bs];

  1250.     int w4 = FFMIN(s->cols - col, bwh_tab[1][b->bs][0]);

  1251.     int h4 = FFMIN(s->rows - row, bwh_tab[1][b->bs][1]), y;

  1252.     int have_a = row > 0, have_l = col > s->tiling.tile_col_start;

  1253. 

  1254.     if (!s->segmentation.enabled) {

  1255.         b->seg_id = 0;

  1256.     } else if (s->keyframe || s->intraonly) {

  1257.         b->seg_id = s->segmentation.update_map ?

  1258.             vp8_rac_get_tree(&s->c, vp9_segmentation_tree, s->prob.seg) : 0;

  1259.     } else if (!s->segmentation.update_map ||

  1260.                (s->segmentation.temporal &&

  1261.                 vp56_rac_get_prob_branchy(&s->c,

  1262.                     s->prob.segpred[s->above_segpred_ctx[col] +

  1263.                                     s->left_segpred_ctx[row7]]))) {

  1264.         int pred = 8, x;

  1265.         uint8_t *refsegmap = s->frames[LAST_FRAME].segmentation_map;

  1266. 

  1267.         if (!s->last_uses_2pass)

  1268.             ff_thread_await_progress(&s->frames[LAST_FRAME].tf, row >> 3, 0);

  1269.         for (y = 0; y < h4; y++)

  1270.             for (x = 0; x < w4; x++)

  1271.                 pred = FFMIN(pred, refsegmap[(y + row) * 8 * s->sb_cols + x + col]);

  1272.         av_assert1(pred < 8);

  1273.         b->seg_id = pred;

  1274. 

  1275.         memset(&s->above_segpred_ctx[col], 1, w4);

  1276.         memset(&s->left_segpred_ctx[row7], 1, h4);

  1277.     } else {

  1278.         b->seg_id = vp8_rac_get_tree(&s->c, vp9_segmentation_tree,

  1279.                                      s->prob.seg);

  1280. 

  1281.         memset(&s->above_segpred_ctx[col], 0, w4);

  1282.         memset(&s->left_segpred_ctx[row7], 0, h4);

  1283.     }

  1284.     if ((s->segmentation.enabled && s->segmentation.update_map) || s->keyframe) {

  1285.         uint8_t *segmap = s->frames[CUR_FRAME].segmentation_map;

  1286. 

  1287.         for (y = 0; y < h4; y++)

  1288.             memset(&segmap[(y + row) * 8 * s->sb_cols + col], b->seg_id, w4);

  1289.     }

  1290. 

  1291.     b->skip = s->segmentation.enabled &&

  1292.         s->segmentation.feat[b->seg_id].skip_enabled;

  1293.     if (!b->skip) {

  1294.         int c = s->left_skip_ctx[row7] + s->above_skip_ctx[col];

  1295.         b->skip = vp56_rac_get_prob(&s->c, s->prob.p.skip[c]);

  1296.         s->counts.skip[c][b->skip]++;

  1297.     }

  1298. 

  1299.     if (s->keyframe || s->intraonly) {

  1300.         b->intra = 1;

  1301.     } else if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1302.         b->intra = !s->segmentation.feat[b->seg_id].ref_val;

  1303.     } else {

  1304.         int c, bit;

  1305. 

  1306.         if (have_a && have_l) {

  1307.             c = s->above_intra_ctx[col] + s->left_intra_ctx[row7];

  1308.             c += (c == 2);

  1309.         } else {

  1310.             c = have_a ? 2 * s->above_intra_ctx[col] :

  1311.                 have_l ? 2 * s->left_intra_ctx[row7] : 0;

  1312.         }

  1313.         bit = vp56_rac_get_prob(&s->c, s->prob.p.intra[c]);

  1314.         s->counts.intra[c][bit]++;

  1315.         b->intra = !bit;

  1316.     }

  1317. 

  1318.     if ((b->intra || !b->skip) && s->txfmmode == TX_SWITCHABLE) {

  1319.         int c;

  1320.         if (have_a) {

  1321.             if (have_l) {

  1322.                 c = (s->above_skip_ctx[col] ? max_tx :

  1323.                      s->above_txfm_ctx[col]) +

  1324.                     (s->left_skip_ctx[row7] ? max_tx :

  1325.                      s->left_txfm_ctx[row7]) > max_tx;

  1326.             } else {

  1327.                 c = s->above_skip_ctx[col] ? 1 :

  1328.                     (s->above_txfm_ctx[col] * 2 > max_tx);

  1329.             }

  1330.         } else if (have_l) {

  1331.             c = s->left_skip_ctx[row7] ? 1 :

  1332.                 (s->left_txfm_ctx[row7] * 2 > max_tx);

  1333.         } else {

  1334.             c = 1;

  1335.         }

  1336.         switch (max_tx) {

  1337.         case TX_32X32:

  1338.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][0]);

  1339.             if (b->tx) {

  1340.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][1]);

  1341.                 if (b->tx == 2)

  1342.                     b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][2]);

  1343.             }

  1344.             s->counts.tx32p[c][b->tx]++;

  1345.             break;

  1346.         case TX_16X16:

  1347.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][0]);

  1348.             if (b->tx)

  1349.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][1]);

  1350.             s->counts.tx16p[c][b->tx]++;

  1351.             break;

  1352.         case TX_8X8:

  1353.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx8p[c]);

  1354.             s->counts.tx8p[c][b->tx]++;

  1355.             break;

  1356.         case TX_4X4:

  1357.             b->tx = TX_4X4;

  1358.             break;

  1359.         }

  1360.     } else {

  1361.         b->tx = FFMIN(max_tx, s->txfmmode);

  1362.     }

  1363. 

  1364.     if (s->keyframe || s->intraonly) {

  1365.         uint8_t *a = &s->above_mode_ctx[col * 2];

  1366.         uint8_t *l = &s->left_mode_ctx[(row7) << 1];

  1367. 

  1368.         b->comp = 0;

  1369.         if (b->bs > BS_8x8) {

  1370.             // FIXME the memory storage intermediates here aren't really

  1371.             // necessary, they're just there to make the code slightly

  1372.             // simpler for now

  1373.             b->mode[0] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1374.                                     vp9_default_kf_ymode_probs[a[0]][l[0]]);

  1375.             if (b->bs != BS_8x4) {

  1376.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1377.                                  vp9_default_kf_ymode_probs[a[1]][b->mode[0]]);

  1378.                 l[0] = a[1] = b->mode[1];

  1379.             } else {

  1380.                 l[0] = a[1] = b->mode[1] = b->mode[0];

  1381.             }

  1382.             if (b->bs != BS_4x8) {

  1383.                 b->mode[2] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1384.                                         vp9_default_kf_ymode_probs[a[0]][l[1]]);

  1385.                 if (b->bs != BS_8x4) {

  1386.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1387.                                   vp9_default_kf_ymode_probs[a[1]][b->mode[2]]);

  1388.                     l[1] = a[1] = b->mode[3];

  1389.                 } else {

  1390.                     l[1] = a[1] = b->mode[3] = b->mode[2];

  1391.                 }

  1392.             } else {

  1393.                 b->mode[2] = b->mode[0];

  1394.                 l[1] = a[1] = b->mode[3] = b->mode[1];

  1395.             }

  1396.         } else {

  1397.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1398.                                           vp9_default_kf_ymode_probs[*a][*l]);

  1399.             b->mode[3] = b->mode[2] = b->mode[1] = b->mode[0];

  1400.             // FIXME this can probably be optimized

  1401.             memset(a, b->mode[0], bwh_tab[0][b->bs][0]);

  1402.             memset(l, b->mode[0], bwh_tab[0][b->bs][1]);

  1403.         }

  1404.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1405.                                      vp9_default_kf_uvmode_probs[b->mode[3]]);

  1406.     } else if (b->intra) {

  1407.         b->comp = 0;

  1408.         if (b->bs > BS_8x8) {

  1409.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1410.                                           s->prob.p.y_mode[0]);

  1411.             s->counts.y_mode[0][b->mode[0]]++;

  1412.             if (b->bs != BS_8x4) {

  1413.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1414.                                               s->prob.p.y_mode[0]);

  1415.                 s->counts.y_mode[0][b->mode[1]]++;

  1416.             } else {

  1417.                 b->mode[1] = b->mode[0];

  1418.             }

  1419.             if (b->bs != BS_4x8) {

  1420.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1421.                                               s->prob.p.y_mode[0]);

  1422.                 s->counts.y_mode[0][b->mode[2]]++;

  1423.                 if (b->bs != BS_8x4) {

  1424.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1425.                                                   s->prob.p.y_mode[0]);

  1426.                     s->counts.y_mode[0][b->mode[3]]++;

  1427.                 } else {

  1428.                     b->mode[3] = b->mode[2];

  1429.                 }

  1430.             } else {

  1431.                 b->mode[2] = b->mode[0];

  1432.                 b->mode[3] = b->mode[1];

  1433.             }

  1434.         } else {

  1435.             static const uint8_t size_group[10] = {

  1436.                 3, 3, 3, 3, 2, 2, 2, 1, 1, 1

  1437.             };

  1438.             int sz = size_group[b->bs];

  1439. 

  1440.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1441.                                           s->prob.p.y_mode[sz]);

  1442.             b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1443.             s->counts.y_mode[sz][b->mode[3]]++;

  1444.         }

  1445.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1446.                                      s->prob.p.uv_mode[b->mode[3]]);

  1447.         s->counts.uv_mode[b->mode[3]][b->uvmode]++;

  1448.     } else {

  1449.         static const uint8_t inter_mode_ctx_lut[14][14] = {

  1450.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1451.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1452.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1453.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1454.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1455.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1456.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1457.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1458.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1459.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1460.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1461.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1462.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },

  1463.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },

  1464.         };

  1465. 

  1466.         if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1467.             av_assert2(s->segmentation.feat[b->seg_id].ref_val != 0);

  1468.             b->comp = 0;

  1469.             b->ref[0] = s->segmentation.feat[b->seg_id].ref_val - 1;

  1470.         } else {

  1471.             // read comp_pred flag

  1472.             if (s->comppredmode != PRED_SWITCHABLE) {

  1473.                 b->comp = s->comppredmode == PRED_COMPREF;

  1474.             } else {

  1475.                 int c;

  1476. 

  1477.                 // FIXME add intra as ref=0xff (or -1) to make these easier?

  1478.                 if (have_a) {

  1479.                     if (have_l) {

  1480.                         if (s->above_comp_ctx[col] && s->left_comp_ctx[row7]) {

  1481.                             c = 4;

  1482.                         } else if (s->above_comp_ctx[col]) {

  1483.                             c = 2 + (s->left_intra_ctx[row7] ||

  1484.                                      s->left_ref_ctx[row7] == s->fixcompref);

  1485.                         } else if (s->left_comp_ctx[row7]) {

  1486.                             c = 2 + (s->above_intra_ctx[col] ||

  1487.                                      s->above_ref_ctx[col] == s->fixcompref);

  1488.                         } else {

  1489.                             c = (!s->above_intra_ctx[col] &&

  1490.                                  s->above_ref_ctx[col] == s->fixcompref) ^

  1491.                             (!s->left_intra_ctx[row7] &&

  1492.                              s->left_ref_ctx[row & 7] == s->fixcompref);

  1493.                         }

  1494.                     } else {

  1495.                         c = s->above_comp_ctx[col] ? 3 :

  1496.                         (!s->above_intra_ctx[col] && s->above_ref_ctx[col] == s->fixcompref);

  1497.                     }

  1498.                 } else if (have_l) {

  1499.                     c = s->left_comp_ctx[row7] ? 3 :

  1500.                     (!s->left_intra_ctx[row7] && s->left_ref_ctx[row7] == s->fixcompref);

  1501.                 } else {

  1502.                     c = 1;

  1503.                 }

  1504.                 b->comp = vp56_rac_get_prob(&s->c, s->prob.p.comp[c]);

  1505.                 s->counts.comp[c][b->comp]++;

  1506.             }

  1507. 

  1508.             // read actual references

  1509.             // FIXME probably cache a few variables here to prevent repetitive

  1510.             // memory accesses below

  1511.             if (b->comp) /* two references */ {

  1512.                 int fix_idx = s->signbias[s->fixcompref], var_idx = !fix_idx, c, bit;

  1513. 

  1514.                 b->ref[fix_idx] = s->fixcompref;

  1515.                 // FIXME can this codeblob be replaced by some sort of LUT?

  1516.                 if (have_a) {

  1517.                     if (have_l) {

  1518.                         if (s->above_intra_ctx[col]) {

  1519.                             if (s->left_intra_ctx[row7]) {

  1520.                                 c = 2;

  1521.                             } else {

  1522.                                 c = 1 + 2 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1523.                             }

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

  1525.                             c = 1 + 2 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1526.                         } else {

  1527.                             int refl = s->left_ref_ctx[row7], refa = s->above_ref_ctx[col];

  1528. 

  1529.                             if (refl == refa && refa == s->varcompref[1]) {

  1530.                                 c = 0;

  1531.                             } else if (!s->left_comp_ctx[row7] && !s->above_comp_ctx[col]) {

  1532.                                 if ((refa == s->fixcompref && refl == s->varcompref[0]) ||

  1533.                                     (refl == s->fixcompref && refa == s->varcompref[0])) {

  1534.                                     c = 4;

  1535.                                 } else {

  1536.                                     c = (refa == refl) ? 3 : 1;

  1537.                                 }

  1538.                             } else if (!s->left_comp_ctx[row7]) {

  1539.                                 if (refa == s->varcompref[1] && refl != s->varcompref[1]) {

  1540.                                     c = 1;

  1541.                                 } else {

  1542.                                     c = (refl == s->varcompref[1] &&

  1543.                                          refa != s->varcompref[1]) ? 2 : 4;

  1544.                                 }

  1545.                             } else if (!s->above_comp_ctx[col]) {

  1546.                                 if (refl == s->varcompref[1] && refa != s->varcompref[1]) {

  1547.                                     c = 1;

  1548.                                 } else {

  1549.                                     c = (refa == s->varcompref[1] &&

  1550.                                          refl != s->varcompref[1]) ? 2 : 4;

  1551.                                 }

  1552.                             } else {

  1553.                                 c = (refl == refa) ? 4 : 2;

  1554.                             }

  1555.                         }

  1556.                     } else {

  1557.                         if (s->above_intra_ctx[col]) {

  1558.                             c = 2;

  1559.                         } else if (s->above_comp_ctx[col]) {

  1560.                             c = 4 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1561.                         } else {

  1562.                             c = 3 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1563.                         }

  1564.                     }

  1565.                 } else if (have_l) {

  1566.                     if (s->left_intra_ctx[row7]) {

  1567.                         c = 2;

  1568.                     } else if (s->left_comp_ctx[row7]) {

  1569.                         c = 4 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1570.                     } else {

  1571.                         c = 3 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1572.                     }

  1573.                 } else {

  1574.                     c = 2;

  1575.                 }

  1576.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.comp_ref[c]);

  1577.                 b->ref[var_idx] = s->varcompref[bit];

  1578.                 s->counts.comp_ref[c][bit]++;

  1579.             } else /* single reference */ {

  1580.                 int bit, c;

  1581. 

  1582.                 if (have_a && !s->above_intra_ctx[col]) {

  1583.                     if (have_l && !s->left_intra_ctx[row7]) {

  1584.                         if (s->left_comp_ctx[row7]) {

  1585.                             if (s->above_comp_ctx[col]) {

  1586.                                 c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7] ||

  1587.                                          !s->above_ref_ctx[col]);

  1588.                             } else {

  1589.                                 c = (3 * !s->above_ref_ctx[col]) +

  1590.                                     (!s->fixcompref || !s->left_ref_ctx[row7]);

  1591.                             }

  1592.                         } else if (s->above_comp_ctx[col]) {

  1593.                             c = (3 * !s->left_ref_ctx[row7]) +

  1594.                                 (!s->fixcompref || !s->above_ref_ctx[col]);

  1595.                         } else {

  1596.                             c = 2 * !s->left_ref_ctx[row7] + 2 * !s->above_ref_ctx[col];

  1597.                         }

  1598.                     } else if (s->above_intra_ctx[col]) {

  1599.                         c = 2;

  1600.                     } else if (s->above_comp_ctx[col]) {

  1601.                         c = 1 + (!s->fixcompref || !s->above_ref_ctx[col]);

  1602.                     } else {

  1603.                         c = 4 * (!s->above_ref_ctx[col]);

  1604.                     }

  1605.                 } else if (have_l && !s->left_intra_ctx[row7]) {

  1606.                     if (s->left_intra_ctx[row7]) {

  1607.                         c = 2;

  1608.                     } else if (s->left_comp_ctx[row7]) {

  1609.                         c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7]);

  1610.                     } else {

  1611.                         c = 4 * (!s->left_ref_ctx[row7]);

  1612.                     }

  1613.                 } else {

  1614.                     c = 2;

  1615.                 }

  1616.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][0]);

  1617.                 s->counts.single_ref[c][0][bit]++;

  1618.                 if (!bit) {

  1619.                     b->ref[0] = 0;

  1620.                 } else {

  1621.                     // FIXME can this codeblob be replaced by some sort of LUT?

  1622.                     if (have_a) {

  1623.                         if (have_l) {

  1624.                             if (s->left_intra_ctx[row7]) {

  1625.                                 if (s->above_intra_ctx[col]) {

  1626.                                     c = 2;

  1627.                                 } else if (s->above_comp_ctx[col]) {

  1628.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1629.                                                  s->above_ref_ctx[col] == 1);

  1630.                                 } else if (!s->above_ref_ctx[col]) {

  1631.                                     c = 3;

  1632.                                 } else {

  1633.                                     c = 4 * (s->above_ref_ctx[col] == 1);

  1634.                                 }

  1635.                             } else if (s->above_intra_ctx[col]) {

  1636.                                 if (s->left_intra_ctx[row7]) {

  1637.                                     c = 2;

  1638.                                 } else if (s->left_comp_ctx[row7]) {

  1639.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1640.                                                  s->left_ref_ctx[row7] == 1);

  1641.                                 } else if (!s->left_ref_ctx[row7]) {

  1642.                                     c = 3;

  1643.                                 } else {

  1644.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1645.                                 }

  1646.                             } else if (s->above_comp_ctx[col]) {

  1647.                                 if (s->left_comp_ctx[row7]) {

  1648.                                     if (s->left_ref_ctx[row7] == s->above_ref_ctx[col]) {

  1649.                                         c = 3 * (s->fixcompref == 1 ||

  1650.                                                  s->left_ref_ctx[row7] == 1);

  1651.                                     } else {

  1652.                                         c = 2;

  1653.                                     }

  1654.                                 } else if (!s->left_ref_ctx[row7]) {

  1655.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1656.                                                  s->above_ref_ctx[col] == 1);

  1657.                                 } else {

  1658.                                     c = 3 * (s->left_ref_ctx[row7] == 1) +

  1659.                                     (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1660.                                 }

  1661.                             } else if (s->left_comp_ctx[row7]) {

  1662.                                 if (!s->above_ref_ctx[col]) {

  1663.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1664.                                                  s->left_ref_ctx[row7] == 1);

  1665.                                 } else {

  1666.                                     c = 3 * (s->above_ref_ctx[col] == 1) +

  1667.                                     (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1668.                                 }

  1669.                             } else if (!s->above_ref_ctx[col]) {

  1670.                                 if (!s->left_ref_ctx[row7]) {

  1671.                                     c = 3;

  1672.                                 } else {

  1673.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1674.                                 }

  1675.                             } else if (!s->left_ref_ctx[row7]) {

  1676.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1677.                             } else {

  1678.                                 c = 2 * (s->left_ref_ctx[row7] == 1) +

  1679.                                 2 * (s->above_ref_ctx[col] == 1);

  1680.                             }

  1681.                         } else {

  1682.                             if (s->above_intra_ctx[col] ||

  1683.                                 (!s->above_comp_ctx[col] && !s->above_ref_ctx[col])) {

  1684.                                 c = 2;

  1685.                             } else if (s->above_comp_ctx[col]) {

  1686.                                 c = 3 * (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1687.                             } else {

  1688.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1689.                             }

  1690.                         }

  1691.                     } else if (have_l) {

  1692.                         if (s->left_intra_ctx[row7] ||

  1693.                             (!s->left_comp_ctx[row7] && !s->left_ref_ctx[row7])) {

  1694.                             c = 2;

  1695.                         } else if (s->left_comp_ctx[row7]) {

  1696.                             c = 3 * (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1697.                         } else {

  1698.                             c = 4 * (s->left_ref_ctx[row7] == 1);

  1699.                         }

  1700.                     } else {

  1701.                         c = 2;

  1702.                     }

  1703.                     bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][1]);

  1704.                     s->counts.single_ref[c][1][bit]++;

  1705.                     b->ref[0] = 1 + bit;

  1706.                 }

  1707.             }

  1708.         }

  1709. 

  1710.         if (b->bs <= BS_8x8) {

  1711.             if (s->segmentation.feat[b->seg_id].skip_enabled) {

  1712.                 b->mode[0] = b->mode[1] = b->mode[2] = b->mode[3] = ZEROMV;

  1713.             } else {

  1714.                 static const uint8_t off[10] = {

  1715.                     3, 0, 0, 1, 0, 0, 0, 0, 0, 0

  1716.                 };

  1717. 

  1718.                 // FIXME this needs to use the LUT tables from find_ref_mvs

  1719.                 // because not all are -1,0/0,-1

  1720.                 int c = inter_mode_ctx_lut[s->above_mode_ctx[col + off[b->bs]]]

  1721.                                           [s->left_mode_ctx[row7 + off[b->bs]]];

  1722. 

  1723.                 b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1724.                                               s->prob.p.mv_mode[c]);

  1725.                 b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1726.                 s->counts.mv_mode[c][b->mode[0] - 10]++;

  1727.             }

  1728.         }

  1729. 

  1730.         if (s->filtermode == FILTER_SWITCHABLE) {

  1731.             int c;

  1732. 

  1733.             if (have_a && s->above_mode_ctx[col] >= NEARESTMV) {

  1734.                 if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1735.                     c = s->above_filter_ctx[col] == s->left_filter_ctx[row7] ?

  1736.                         s->left_filter_ctx[row7] : 3;

  1737.                 } else {

  1738.                     c = s->above_filter_ctx[col];

  1739.                 }

  1740.             } else if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1741.                 c = s->left_filter_ctx[row7];

  1742.             } else {

  1743.                 c = 3;

  1744.             }

  1745. 

  1746.             b->filter = vp8_rac_get_tree(&s->c, vp9_filter_tree,

  1747.                                          s->prob.p.filter[c]);

  1748.             s->counts.filter[c][b->filter]++;

  1749.         } else {

  1750.             b->filter = s->filtermode;

  1751.         }

  1752. 

  1753.         if (b->bs > BS_8x8) {

  1754.             int c = inter_mode_ctx_lut[s->above_mode_ctx[col]][s->left_mode_ctx[row7]];

  1755. 

  1756.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1757.                                           s->prob.p.mv_mode[c]);

  1758.             s->counts.mv_mode[c][b->mode[0] - 10]++;

  1759.             fill_mv(s, b->mv[0], b->mode[0], 0);

  1760. 

  1761.             if (b->bs != BS_8x4) {

  1762.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1763.                                               s->prob.p.mv_mode[c]);

  1764.                 s->counts.mv_mode[c][b->mode[1] - 10]++;

  1765.                 fill_mv(s, b->mv[1], b->mode[1], 1);

  1766.             } else {

  1767.                 b->mode[1] = b->mode[0];

  1768.                 AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1769.                 AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1770.             }

  1771. 

  1772.             if (b->bs != BS_4x8) {

  1773.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1774.                                               s->prob.p.mv_mode[c]);

  1775.                 s->counts.mv_mode[c][b->mode[2] - 10]++;

  1776.                 fill_mv(s, b->mv[2], b->mode[2], 2);

  1777. 

  1778.                 if (b->bs != BS_8x4) {

  1779.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1780.                                                   s->prob.p.mv_mode[c]);

  1781.                     s->counts.mv_mode[c][b->mode[3] - 10]++;

  1782.                     fill_mv(s, b->mv[3], b->mode[3], 3);

  1783.                 } else {

  1784.                     b->mode[3] = b->mode[2];

  1785.                     AV_COPY32(&b->mv[3][0], &b->mv[2][0]);

  1786.                     AV_COPY32(&b->mv[3][1], &b->mv[2][1]);

  1787.                 }

  1788.             } else {

  1789.                 b->mode[2] = b->mode[0];

  1790.                 AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1791.                 AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1792.                 b->mode[3] = b->mode[1];

  1793.                 AV_COPY32(&b->mv[3][0], &b->mv[1][0]);

  1794.                 AV_COPY32(&b->mv[3][1], &b->mv[1][1]);

  1795.             }

  1796.         } else {

  1797.             fill_mv(s, b->mv[0], b->mode[0], -1);

  1798.             AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1799.             AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1800.             AV_COPY32(&b->mv[3][0], &b->mv[0][0]);

  1801.             AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1802.             AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1803.             AV_COPY32(&b->mv[3][1], &b->mv[0][1]);

  1804.         }

  1805.     }

  1806. 

  1807.     // FIXME this can probably be optimized

  1808.     memset(&s->above_skip_ctx[col], b->skip, w4);

  1809.     memset(&s->left_skip_ctx[row7], b->skip, h4);

  1810.     memset(&s->above_txfm_ctx[col], b->tx, w4);

  1811.     memset(&s->left_txfm_ctx[row7], b->tx, h4);

  1812.     memset(&s->above_partition_ctx[col], above_ctx[b->bs], w4);

  1813.     memset(&s->left_partition_ctx[row7], left_ctx[b->bs], h4);

  1814.     if (!s->keyframe && !s->intraonly) {

  1815.         memset(&s->above_intra_ctx[col], b->intra, w4);

  1816.         memset(&s->left_intra_ctx[row7], b->intra, h4);

  1817.         memset(&s->above_comp_ctx[col], b->comp, w4);

  1818.         memset(&s->left_comp_ctx[row7], b->comp, h4);

  1819.         memset(&s->above_mode_ctx[col], b->mode[3], w4);

  1820.         memset(&s->left_mode_ctx[row7], b->mode[3], h4);

  1821.         if (s->filtermode == FILTER_SWITCHABLE && !b->intra ) {

  1822.             memset(&s->above_filter_ctx[col], b->filter, w4);

  1823.             memset(&s->left_filter_ctx[row7], b->filter, h4);

  1824.             b->filter = vp9_filter_lut[b->filter];

  1825.         }

  1826.         if (b->bs > BS_8x8) {

  1827.             int mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  1828. 

  1829.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][0], &b->mv[1][0]);

  1830.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][1], &b->mv[1][1]);

  1831.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][0], mv0);

  1832.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][1], mv1);

  1833.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][0], &b->mv[2][0]);

  1834.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][1], &b->mv[2][1]);

  1835.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][0], mv0);

  1836.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][1], mv1);

  1837.         } else {

  1838.             int n, mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  1839. 

  1840.             for (n = 0; n < w4 * 2; n++) {

  1841.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][0], mv0);

  1842.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][1], mv1);

  1843.             }

  1844.             for (n = 0; n < h4 * 2; n++) {

  1845.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][0], mv0);

  1846.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][1], mv1);

  1847.             }

  1848.         }

  1849. 

  1850.         if (!b->intra) { // FIXME write 0xff or -1 if intra, so we can use this

  1851.                          // as a direct check in above branches

  1852.             int vref = b->ref[b->comp ? s->signbias[s->varcompref[0]] : 0];

  1853. 

  1854.             memset(&s->above_ref_ctx[col], vref, w4);

  1855.             memset(&s->left_ref_ctx[row7], vref, h4);

  1856.         }

  1857.     }

  1858. 

  1859.     // FIXME kinda ugly

  1860.     for (y = 0; y < h4; y++) {

  1861.         int x, o = (row + y) * s->sb_cols * 8 + col;

  1862.         struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[o];

  1863. 

  1864.         if (b->intra) {

  1865.             for (x = 0; x < w4; x++) {

  1866.                 mv[x].ref[0] =

  1867.                 mv[x].ref[1] = -1;

  1868.             }

  1869.         } else if (b->comp) {

  1870.             for (x = 0; x < w4; x++) {

  1871.                 mv[x].ref[0] = b->ref[0];

  1872.                 mv[x].ref[1] = b->ref[1];

  1873.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  1874.                 AV_COPY32(&mv[x].mv[1], &b->mv[3][1]);

  1875.             }

  1876.         } else {

  1877.             for (x = 0; x < w4; x++) {

  1878.                 mv[x].ref[0] = b->ref[0];

  1879.                 mv[x].ref[1] = -1;

  1880.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  1881.             }

  1882.         }

  1883.     }

  1884. }

  1885. 

  1886. // FIXME remove tx argument, and merge cnt/eob arguments?

  1887. static int decode_coeffs_b(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  1888.                            enum TxfmMode tx, unsigned (*cnt)[6][3],

  1889.                            unsigned (*eob)[6][2], uint8_t (*p)[6][11],

  1890.                            int nnz, const int16_t *scan, const int16_t (*nb)[2],

  1891.                            const int16_t *band_counts, const int16_t *qmul)

  1892. {

  1893.     int i = 0, band = 0, band_left = band_counts[band];

  1894.     uint8_t *tp = p[0][nnz];

  1895.     uint8_t cache[1024];

  1896. 

  1897.     do {

  1898.         int val, rc;

  1899. 

  1900.         val = vp56_rac_get_prob_branchy(c, tp[0]); // eob

  1901.         eob[band][nnz][val]++;

  1902.         if (!val)

  1903.             break;

  1904. 

  1905.     skip_eob:

  1906.         if (!vp56_rac_get_prob_branchy(c, tp[1])) { // zero

  1907.             cnt[band][nnz][0]++;

  1908.             if (!--band_left)

  1909.                 band_left = band_counts[++band];

  1910.             cache[scan[i]] = 0;

  1911.             nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  1912.             tp = p[band][nnz];

  1913.             if (++i == n_coeffs)

  1914.                 break; //invalid input; blocks should end with EOB

  1915.             goto skip_eob;

  1916.         }

  1917. 

  1918.         rc = scan[i];

  1919.         if (!vp56_rac_get_prob_branchy(c, tp[2])) { // one

  1920.             cnt[band][nnz][1]++;

  1921.             val = 1;

  1922.             cache[rc] = 1;

  1923.         } else {

  1924.             // fill in p[3-10] (model fill) - only once per frame for each pos

  1925.             if (!tp[3])

  1926.                 memcpy(&tp[3], vp9_model_pareto8[tp[2]], 8);

  1927. 

  1928.             cnt[band][nnz][2]++;

  1929.             if (!vp56_rac_get_prob_branchy(c, tp[3])) { // 2, 3, 4

  1930.                 if (!vp56_rac_get_prob_branchy(c, tp[4])) {

  1931.                     cache[rc] = val = 2;

  1932.                 } else {

  1933.                     val = 3 + vp56_rac_get_prob(c, tp[5]);

  1934.                     cache[rc] = 3;

  1935.                 }

  1936.             } else if (!vp56_rac_get_prob_branchy(c, tp[6])) { // cat1/2

  1937.                 cache[rc] = 4;

  1938.                 if (!vp56_rac_get_prob_branchy(c, tp[7])) {

  1939.                     val = 5 + vp56_rac_get_prob(c, 159);

  1940.                 } else {

  1941.                     val  = 7 + (vp56_rac_get_prob(c, 165) << 1);

  1942.                     val +=      vp56_rac_get_prob(c, 145);

  1943.                 }

  1944.             } else { // cat 3-6

  1945.                 cache[rc] = 5;

  1946.                 if (!vp56_rac_get_prob_branchy(c, tp[8])) {

  1947.                     if (!vp56_rac_get_prob_branchy(c, tp[9])) {

  1948.                         val  = 11 + (vp56_rac_get_prob(c, 173) << 2);

  1949.                         val +=      (vp56_rac_get_prob(c, 148) << 1);

  1950.                         val +=       vp56_rac_get_prob(c, 140);

  1951.                     } else {

  1952.                         val  = 19 + (vp56_rac_get_prob(c, 176) << 3);

  1953.                         val +=      (vp56_rac_get_prob(c, 155) << 2);

  1954.                         val +=      (vp56_rac_get_prob(c, 140) << 1);

  1955.                         val +=       vp56_rac_get_prob(c, 135);

  1956.                     }

  1957.                 } else if (!vp56_rac_get_prob_branchy(c, tp[10])) {

  1958.                     val  = 35 + (vp56_rac_get_prob(c, 180) << 4);

  1959.                     val +=      (vp56_rac_get_prob(c, 157) << 3);

  1960.                     val +=      (vp56_rac_get_prob(c, 141) << 2);

  1961.                     val +=      (vp56_rac_get_prob(c, 134) << 1);

  1962.                     val +=       vp56_rac_get_prob(c, 130);

  1963.                 } else {

  1964.                     val  = 67 + (vp56_rac_get_prob(c, 254) << 13);

  1965.                     val +=      (vp56_rac_get_prob(c, 254) << 12);

  1966.                     val +=      (vp56_rac_get_prob(c, 254) << 11);

  1967.                     val +=      (vp56_rac_get_prob(c, 252) << 10);

  1968.                     val +=      (vp56_rac_get_prob(c, 249) << 9);

  1969.                     val +=      (vp56_rac_get_prob(c, 243) << 8);

  1970.                     val +=      (vp56_rac_get_prob(c, 230) << 7);

  1971.                     val +=      (vp56_rac_get_prob(c, 196) << 6);

  1972.                     val +=      (vp56_rac_get_prob(c, 177) << 5);

  1973.                     val +=      (vp56_rac_get_prob(c, 153) << 4);

  1974.                     val +=      (vp56_rac_get_prob(c, 140) << 3);

  1975.                     val +=      (vp56_rac_get_prob(c, 133) << 2);

  1976.                     val +=      (vp56_rac_get_prob(c, 130) << 1);

  1977.                     val +=       vp56_rac_get_prob(c, 129);

  1978.                 }

  1979.             }

  1980.         }

  1981.         if (!--band_left)

  1982.             band_left = band_counts[++band];

  1983.         if (tx == TX_32X32) // FIXME slow

  1984.             coef[rc] = ((vp8_rac_get(c) ? -val : val) * qmul[!!i]) / 2;

  1985.         else

  1986.             coef[rc] = (vp8_rac_get(c) ? -val : val) * qmul[!!i];

  1987.         nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  1988.         tp = p[band][nnz];

  1989.     } while (++i < n_coeffs);

  1990. 

  1991.     return i;

  1992. }

  1993. 

  1994. static void decode_coeffs(AVCodecContext *ctx)

  1995. {

  1996.     VP9Context *s = ctx->priv_data;

  1997.     VP9Block *b = s->b;

  1998.     int row = s->row, col = s->col;

  1999.     uint8_t (*p)[6][11] = s->prob.coef[b->tx][0 /* y */][!b->intra];

  2000.     unsigned (*c)[6][3] = s->counts.coef[b->tx][0 /* y */][!b->intra];

  2001.     unsigned (*e)[6][2] = s->counts.eob[b->tx][0 /* y */][!b->intra];

  2002.     int w4 = bwh_tab[1][b->bs][0] << 1, h4 = bwh_tab[1][b->bs][1] << 1;

  2003.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2004.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2005.     int n, pl, x, y, step1d = 1 << b->tx, step = 1 << (b->tx * 2);

  2006.     int uvstep1d = 1 << b->uvtx, uvstep = 1 << (b->uvtx * 2), res;

  2007.     int16_t (*qmul)[2] = s->segmentation.feat[b->seg_id].qmul;

  2008.     int tx = 4 * s->lossless + b->tx;

  2009.     const int16_t * const *yscans = vp9_scans[tx];

  2010.     const int16_t (* const *ynbs)[2] = vp9_scans_nb[tx];

  2011.     const int16_t *uvscan = vp9_scans[b->uvtx][DCT_DCT];

  2012.     const int16_t (*uvnb)[2] = vp9_scans_nb[b->uvtx][DCT_DCT];

  2013.     uint8_t *a = &s->above_y_nnz_ctx[col * 2];

  2014.     uint8_t *l = &s->left_y_nnz_ctx[(row & 7) << 1];

  2015.     static const int16_t band_counts[4][8] = {

  2016.         { 1, 2, 3, 4,  3,   16 - 13 },

  2017.         { 1, 2, 3, 4, 11,   64 - 21 },

  2018.         { 1, 2, 3, 4, 11,  256 - 21 },

  2019.         { 1, 2, 3, 4, 11, 1024 - 21 },

  2020.     };

  2021.     const int16_t *y_band_counts = band_counts[b->tx];

  2022.     const int16_t *uv_band_counts = band_counts[b->uvtx];

  2023. 

  2024.     /* y tokens */

  2025.     if (b->tx > TX_4X4) { // FIXME slow

  2026.         for (y = 0; y < end_y; y += step1d)

  2027.             for (x = 1; x < step1d; x++)

  2028.                 l[y] |= l[y + x];

  2029.         for (x = 0; x < end_x; x += step1d)

  2030.             for (y = 1; y < step1d; y++)

  2031.                 a[x] |= a[x + y];

  2032.     }

  2033.     for (n = 0, y = 0; y < end_y; y += step1d) {

  2034.         for (x = 0; x < end_x; x += step1d, n += step) {

  2035.             enum TxfmType txtp = vp9_intra_txfm_type[b->mode[b->tx == TX_4X4 &&

  2036.                                                              b->bs > BS_8x8 ?

  2037.                                                              n : 0]];

  2038.             int nnz = a[x] + l[y];

  2039.             res = decode_coeffs_b(&s->c, s->block + 16 * n, 16 * step,

  2040.                                   b->tx, c, e, p, nnz, yscans[txtp],

  2041.                                   ynbs[txtp], y_band_counts, qmul[0]);

  2042.             a[x] = l[y] = !!res;

  2043.             if (b->tx > TX_8X8) {

  2044.                 AV_WN16A(&s->eob[n], res);

  2045.             } else {

  2046.                 s->eob[n] = res;

  2047.             }

  2048.         }

  2049.     }

  2050.     if (b->tx > TX_4X4) { // FIXME slow

  2051.         for (y = 0; y < end_y; y += step1d)

  2052.             memset(&l[y + 1], l[y], FFMIN(end_y - y - 1, step1d - 1));

  2053.         for (x = 0; x < end_x; x += step1d)

  2054.             memset(&a[x + 1], a[x], FFMIN(end_x - x - 1, step1d - 1));

  2055.     }

  2056. 

  2057.     p = s->prob.coef[b->uvtx][1 /* uv */][!b->intra];

  2058.     c = s->counts.coef[b->uvtx][1 /* uv */][!b->intra];

  2059.     e = s->counts.eob[b->uvtx][1 /* uv */][!b->intra];

  2060.     w4 >>= 1;

  2061.     h4 >>= 1;

  2062.     end_x >>= 1;

  2063.     end_y >>= 1;

  2064.     for (pl = 0; pl < 2; pl++) {

  2065.         a = &s->above_uv_nnz_ctx[pl][col];

  2066.         l = &s->left_uv_nnz_ctx[pl][row & 7];

  2067.         if (b->uvtx > TX_4X4) { // FIXME slow

  2068.             for (y = 0; y < end_y; y += uvstep1d)

  2069.                 for (x = 1; x < uvstep1d; x++)

  2070.                     l[y] |= l[y + x];

  2071.             for (x = 0; x < end_x; x += uvstep1d)

  2072.                 for (y = 1; y < uvstep1d; y++)

  2073.                     a[x] |= a[x + y];

  2074.         }

  2075.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2076.             for (x = 0; x < end_x; x += uvstep1d, n += uvstep) {

  2077.                 int nnz = a[x] + l[y];

  2078.                 res = decode_coeffs_b(&s->c, s->uvblock[pl] + 16 * n,

  2079.                                       16 * uvstep, b->uvtx, c, e, p, nnz,

  2080.                                       uvscan, uvnb, uv_band_counts, qmul[1]);

  2081.                 a[x] = l[y] = !!res;

  2082.                 if (b->uvtx > TX_8X8) {

  2083.                     AV_WN16A(&s->uveob[pl][n], res);

  2084.                 } else {

  2085.                     s->uveob[pl][n] = res;

  2086.                 }

  2087.             }

  2088.         }

  2089.         if (b->uvtx > TX_4X4) { // FIXME slow

  2090.             for (y = 0; y < end_y; y += uvstep1d)

  2091.                 memset(&l[y + 1], l[y], FFMIN(end_y - y - 1, uvstep1d - 1));

  2092.             for (x = 0; x < end_x; x += uvstep1d)

  2093.                 memset(&a[x + 1], a[x], FFMIN(end_x - x - 1, uvstep1d - 1));

  2094.         }

  2095.     }

  2096. }

  2097. 

  2098. static av_always_inline int check_intra_mode(VP9Context *s, int mode, uint8_t **a,

  2099.                                              uint8_t *dst_edge, ptrdiff_t stride_edge,

  2100.                                              uint8_t *dst_inner, ptrdiff_t stride_inner,

  2101.                                              uint8_t *l, int col, int x, int w,

  2102.                                              int row, int y, enum TxfmMode tx,

  2103.                                              int p)

  2104. {

  2105.     int have_top = row > 0 || y > 0;

  2106.     int have_left = col > s->tiling.tile_col_start || x > 0;

  2107.     int have_right = x < w - 1;

  2108.     static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {

  2109.         [VERT_PRED]            = { { DC_127_PRED,          VERT_PRED },

  2110.                                    { DC_127_PRED,          VERT_PRED } },

  2111.         [HOR_PRED]             = { { DC_129_PRED,          DC_129_PRED },

  2112.                                    { HOR_PRED,             HOR_PRED } },

  2113.         [DC_PRED]              = { { DC_128_PRED,          TOP_DC_PRED },

  2114.                                    { LEFT_DC_PRED,         DC_PRED } },

  2115.         [DIAG_DOWN_LEFT_PRED]  = { { DC_127_PRED,          DIAG_DOWN_LEFT_PRED },

  2116.                                    { DC_127_PRED,          DIAG_DOWN_LEFT_PRED } },

  2117.         [DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },

  2118.                                    { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },

  2119.         [VERT_RIGHT_PRED]      = { { VERT_RIGHT_PRED,      VERT_RIGHT_PRED },

  2120.                                    { VERT_RIGHT_PRED,      VERT_RIGHT_PRED } },

  2121.         [HOR_DOWN_PRED]        = { { HOR_DOWN_PRED,        HOR_DOWN_PRED },

  2122.                                    { HOR_DOWN_PRED,        HOR_DOWN_PRED } },

  2123.         [VERT_LEFT_PRED]       = { { DC_127_PRED,          VERT_LEFT_PRED },

  2124.                                    { DC_127_PRED,          VERT_LEFT_PRED } },

  2125.         [HOR_UP_PRED]          = { { DC_129_PRED,          DC_129_PRED },

  2126.                                    { HOR_UP_PRED,          HOR_UP_PRED } },

  2127.         [TM_VP8_PRED]          = { { DC_129_PRED,          VERT_PRED },

  2128.                                    { HOR_PRED,             TM_VP8_PRED } },

  2129.     };

  2130.     static const struct {

  2131.         uint8_t needs_left:1;

  2132.         uint8_t needs_top:1;

  2133.         uint8_t needs_topleft:1;

  2134.         uint8_t needs_topright:1;

  2135.     } edges[N_INTRA_PRED_MODES] = {

  2136.         [VERT_PRED]            = { .needs_top  = 1 },

  2137.         [HOR_PRED]             = { .needs_left = 1 },

  2138.         [DC_PRED]              = { .needs_top  = 1, .needs_left = 1 },

  2139.         [DIAG_DOWN_LEFT_PRED]  = { .needs_top  = 1, .needs_topright = 1 },

  2140.         [DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2141.         [VERT_RIGHT_PRED]      = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2142.         [HOR_DOWN_PRED]        = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2143.         [VERT_LEFT_PRED]       = { .needs_top  = 1, .needs_topright = 1 },

  2144.         [HOR_UP_PRED]          = { .needs_left = 1 },

  2145.         [TM_VP8_PRED]          = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2146.         [LEFT_DC_PRED]         = { .needs_left = 1 },

  2147.         [TOP_DC_PRED]          = { .needs_top  = 1 },

  2148.         [DC_128_PRED]          = { 0 },

  2149.         [DC_127_PRED]          = { 0 },

  2150.         [DC_129_PRED]          = { 0 }

  2151.     };

  2152. 

  2153.     av_assert2(mode >= 0 && mode < 10);

  2154.     mode = mode_conv[mode][have_left][have_top];

  2155.     if (edges[mode].needs_top) {

  2156.         uint8_t *top, *topleft;

  2157.         int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !p) - x) * 4;

  2158.         int n_px_need_tr = 0;

  2159. 

  2160.         if (tx == TX_4X4 && edges[mode].needs_topright && have_right)

  2161.             n_px_need_tr = 4;

  2162. 

  2163.         // if top of sb64-row, use s->intra_pred_data[] instead of

  2164.         // dst[-stride] for intra prediction (it contains pre- instead of

  2165.         // post-loopfilter data)

  2166.         if (have_top) {

  2167.             top = !(row & 7) && !y ?

  2168.                 s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  2169.                 y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];

  2170.             if (have_left)

  2171.                 topleft = !(row & 7) && !y ?

  2172.                     s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  2173.                     y == 0 || x == 0 ? &dst_edge[-stride_edge] :

  2174.                     &dst_inner[-stride_inner];

  2175.         }

  2176. 

  2177.         if (have_top &&

  2178.             (!edges[mode].needs_topleft || (have_left && top == topleft)) &&

  2179.             (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&

  2180.             n_px_need + n_px_need_tr <= n_px_have) {

  2181.             *a = top;

  2182.         } else {

  2183.             if (have_top) {

  2184.                 if (n_px_need <= n_px_have) {

  2185.                     memcpy(*a, top, n_px_need);

  2186.                 } else {

  2187.                     memcpy(*a, top, n_px_have);

  2188.                     memset(&(*a)[n_px_have], (*a)[n_px_have - 1],

  2189.                            n_px_need - n_px_have);

  2190.                 }

  2191.             } else {

  2192.                 memset(*a, 127, n_px_need);

  2193.             }

  2194.             if (edges[mode].needs_topleft) {

  2195.                 if (have_left && have_top) {

  2196.                     (*a)[-1] = topleft[-1];

  2197.                 } else {

  2198.                     (*a)[-1] = have_top ? 129 : 127;

  2199.                 }

  2200.             }

  2201.             if (tx == TX_4X4 && edges[mode].needs_topright) {

  2202.                 if (have_top && have_right &&

  2203.                     n_px_need + n_px_need_tr <= n_px_have) {

  2204.                     memcpy(&(*a)[4], &top[4], 4);

  2205.                 } else {

  2206.                     memset(&(*a)[4], (*a)[3], 4);

  2207.                 }

  2208.             }

  2209.         }

  2210.     }

  2211.     if (edges[mode].needs_left) {

  2212.         if (have_left) {

  2213.             int n_px_need = 4 << tx, i, n_px_have = (((s->rows - row) << !p) - y) * 4;

  2214.             uint8_t *dst = x == 0 ? dst_edge : dst_inner;

  2215.             ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;

  2216. 

  2217.             if (n_px_need <= n_px_have) {

  2218.                 for (i = 0; i < n_px_need; i++)

  2219.                     l[i] = dst[i * stride - 1];

  2220.             } else {

  2221.                 for (i = 0; i < n_px_have; i++)

  2222.                     l[i] = dst[i * stride - 1];

  2223.                 memset(&l[i], l[i - 1], n_px_need - n_px_have);

  2224.             }

  2225.         } else {

  2226.             memset(l, 129, 4 << tx);

  2227.         }

  2228.     }

  2229. 

  2230.     return mode;

  2231. }

  2232. 

  2233. static void intra_recon(AVCodecContext *ctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  2234. {

  2235.     VP9Context *s = ctx->priv_data;

  2236.     VP9Block *b = s->b;

  2237.     int row = s->row, col = s->col;

  2238.     int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2239.     int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2240.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2241.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2242.     int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2243.     int uvstep1d = 1 << b->uvtx, p;

  2244.     uint8_t *dst = s->dst[0], *dst_r = s->frames[CUR_FRAME].tf.f->data[0] + y_off;

  2245. 

  2246.     for (n = 0, y = 0; y < end_y; y += step1d) {

  2247.         uint8_t *ptr = dst, *ptr_r = dst_r;

  2248.         for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d,

  2249.                                ptr_r += 4 * step1d, n += step) {

  2250.             int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?

  2251.                                y * 2 + x : 0];

  2252.             LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);

  2253.             uint8_t *a = &a_buf[16], l[32];

  2254.             enum TxfmType txtp = vp9_intra_txfm_type[mode];

  2255.             int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2256. 

  2257.             mode = check_intra_mode(s, mode, &a, ptr_r,

  2258.                                     s->frames[CUR_FRAME].tf.f->linesize[0],

  2259.                                     ptr, s->y_stride, l,

  2260.                                     col, x, w4, row, y, b->tx, 0);

  2261.             s->dsp.intra_pred[b->tx][mode](ptr, s->y_stride, l, a);

  2262.             if (eob)

  2263.                 s->dsp.itxfm_add[tx][txtp](ptr, s->y_stride,

  2264.                                            s->block + 16 * n, eob);

  2265.         }

  2266.         dst_r += 4 * step1d * s->frames[CUR_FRAME].tf.f->linesize[0];

  2267.         dst   += 4 * step1d * s->y_stride;

  2268.     }

  2269. 

  2270.     // U/V

  2271.     h4 >>= 1;

  2272.     w4 >>= 1;

  2273.     end_x >>= 1;

  2274.     end_y >>= 1;

  2275.     step = 1 << (b->uvtx * 2);

  2276.     for (p = 0; p < 2; p++) {

  2277.         dst   = s->dst[1 + p];

  2278.         dst_r = s->frames[CUR_FRAME].tf.f->data[1 + p] + uv_off;

  2279.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2280.             uint8_t *ptr = dst, *ptr_r = dst_r;

  2281.             for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d,

  2282.                                    ptr_r += 4 * uvstep1d, n += step) {

  2283.                 int mode = b->uvmode;

  2284.                 LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);

  2285.                 uint8_t *a = &a_buf[16], l[32];

  2286.                 int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2287. 

  2288.                 mode = check_intra_mode(s, mode, &a, ptr_r,

  2289.                                         s->frames[CUR_FRAME].tf.f->linesize[1],

  2290.                                         ptr, s->uv_stride, l,

  2291.                                         col, x, w4, row, y, b->uvtx, p + 1);

  2292.                 s->dsp.intra_pred[b->uvtx][mode](ptr, s->uv_stride, l, a);

  2293.                 if (eob)

  2294.                     s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2295.                                                     s->uvblock[p] + 16 * n, eob);

  2296.             }

  2297.             dst_r += 4 * uvstep1d * s->frames[CUR_FRAME].tf.f->linesize[1];

  2298.             dst   += 4 * uvstep1d * s->uv_stride;

  2299.         }

  2300.     }

  2301. }

  2302. 

  2303. static av_always_inline void mc_luma_dir(VP9Context *s, vp9_mc_func (*mc)[2],

  2304.                                          uint8_t *dst, ptrdiff_t dst_stride,

  2305.                                          const uint8_t *ref, ptrdiff_t ref_stride,

  2306.                                          ThreadFrame *ref_frame,

  2307.                                          ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2308.                                          int bw, int bh, int w, int h)

  2309. {

  2310.     int mx = mv->x, my = mv->y, th;

  2311. 

  2312.     y += my >> 3;

  2313.     x += mx >> 3;

  2314.     ref += y * ref_stride + x;

  2315.     mx &= 7;

  2316.     my &= 7;

  2317.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2318.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2319.     // the longest loopfilter of the next sbrow

  2320.     th = (y + bh + 4 * !!my + 7) >> 6;

  2321.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2322.     if (x < !!mx * 3 || y < !!my * 3 ||

  2323.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2324.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2325.                                  ref - !!my * 3 * ref_stride - !!mx * 3,

  2326.                                  80, ref_stride,

  2327.                                  bw + !!mx * 7, bh + !!my * 7,

  2328.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2329.         ref = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2330.         ref_stride = 80;

  2331.     }

  2332.     mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);

  2333. }

  2334. 

  2335. static av_always_inline void mc_chroma_dir(VP9Context *s, vp9_mc_func (*mc)[2],

  2336.                                            uint8_t *dst_u, uint8_t *dst_v,

  2337.                                            ptrdiff_t dst_stride,

  2338.                                            const uint8_t *ref_u, ptrdiff_t src_stride_u,

  2339.                                            const uint8_t *ref_v, ptrdiff_t src_stride_v,

  2340.                                            ThreadFrame *ref_frame,

  2341.                                            ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2342.                                            int bw, int bh, int w, int h)

  2343. {

  2344.     int mx = mv->x, my = mv->y, th;

  2345. 

  2346.     y += my >> 4;

  2347.     x += mx >> 4;

  2348.     ref_u += y * src_stride_u + x;

  2349.     ref_v += y * src_stride_v + x;

  2350.     mx &= 15;

  2351.     my &= 15;

  2352.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2353.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2354.     // the longest loopfilter of the next sbrow

  2355.     th = (y + bh + 4 * !!my + 7) >> 5;

  2356.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2357.     if (x < !!mx * 3 || y < !!my * 3 ||

  2358.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2359.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2360.                                  ref_u - !!my * 3 * src_stride_u - !!mx * 3,

  2361.                                  80, src_stride_u,

  2362.                                  bw + !!mx * 7, bh + !!my * 7,

  2363.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2364.         ref_u = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2365.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, 80, bh, mx, my);

  2366. 

  2367.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2368.                                  ref_v - !!my * 3 * src_stride_v - !!mx * 3,

  2369.                                  80, src_stride_v,

  2370.                                  bw + !!mx * 7, bh + !!my * 7,

  2371.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2372.         ref_v = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2373.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, 80, bh, mx, my);

  2374.     } else {

  2375.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);

  2376.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);

  2377.     }

  2378. }

  2379. 

  2380. static void inter_recon(AVCodecContext *ctx)

  2381. {

  2382.     static const uint8_t bwlog_tab[2][N_BS_SIZES] = {

  2383.         { 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4 },

  2384.         { 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 4, 4 },

  2385.     };

  2386.     VP9Context *s = ctx->priv_data;

  2387.     VP9Block *b = s->b;

  2388.     int row = s->row, col = s->col;

  2389.     ThreadFrame *tref1 = &s->refs[s->refidx[b->ref[0]]], *tref2;

  2390.     AVFrame *ref1 = tref1->f, *ref2;

  2391.     int w1 = ref1->width, h1 = ref1->height, w2, h2;

  2392.     ptrdiff_t ls_y = s->y_stride, ls_uv = s->uv_stride;

  2393. 

  2394.     if (b->comp) {

  2395.         tref2 = &s->refs[s->refidx[b->ref[1]]];

  2396.         ref2 = tref2->f;

  2397.         w2 = ref2->width;

  2398.         h2 = ref2->height;

  2399.     }

  2400. 

  2401.     // y inter pred

  2402.     if (b->bs > BS_8x8) {

  2403.         if (b->bs == BS_8x4) {

  2404.             mc_luma_dir(s, s->dsp.mc[3][b->filter][0], s->dst[0], ls_y,

  2405.                         ref1->data[0], ref1->linesize[0], tref1,

  2406.                         row << 3, col << 3, &b->mv[0][0], 8, 4, w1, h1);

  2407.             mc_luma_dir(s, s->dsp.mc[3][b->filter][0],

  2408.                         s->dst[0] + 4 * ls_y, ls_y,

  2409.                         ref1->data[0], ref1->linesize[0], tref1,

  2410.                         (row << 3) + 4, col << 3, &b->mv[2][0], 8, 4, w1, h1);

  2411. 

  2412.             if (b->comp) {

  2413.                 mc_luma_dir(s, s->dsp.mc[3][b->filter][1], s->dst[0], ls_y,

  2414.                             ref2->data[0], ref2->linesize[0], tref2,

  2415.                             row << 3, col << 3, &b->mv[0][1], 8, 4, w2, h2);

  2416.                 mc_luma_dir(s, s->dsp.mc[3][b->filter][1],

  2417.                             s->dst[0] + 4 * ls_y, ls_y,

  2418.                             ref2->data[0], ref2->linesize[0], tref2,

  2419.                             (row << 3) + 4, col << 3, &b->mv[2][1], 8, 4, w2, h2);

  2420.             }

  2421.         } else if (b->bs == BS_4x8) {

  2422.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0], ls_y,

  2423.                         ref1->data[0], ref1->linesize[0], tref1,

  2424.                         row << 3, col << 3, &b->mv[0][0], 4, 8, w1, h1);

  2425.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0] + 4, ls_y,

  2426.                         ref1->data[0], ref1->linesize[0], tref1,

  2427.                         row << 3, (col << 3) + 4, &b->mv[1][0], 4, 8, w1, h1);

  2428. 

  2429.             if (b->comp) {

  2430.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0], ls_y,

  2431.                             ref2->data[0], ref2->linesize[0], tref2,

  2432.                             row << 3, col << 3, &b->mv[0][1], 4, 8, w2, h2);

  2433.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0] + 4, ls_y,

  2434.                             ref2->data[0], ref2->linesize[0], tref2,

  2435.                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 8, w2, h2);

  2436.             }

  2437.         } else {

  2438.             av_assert2(b->bs == BS_4x4);

  2439. 

  2440.             // FIXME if two horizontally adjacent blocks have the same MV,

  2441.             // do a w8 instead of a w4 call

  2442.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0], ls_y,

  2443.                         ref1->data[0], ref1->linesize[0], tref1,

  2444.                         row << 3, col << 3, &b->mv[0][0], 4, 4, w1, h1);

  2445.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0] + 4, ls_y,

  2446.                         ref1->data[0], ref1->linesize[0], tref1,

  2447.                         row << 3, (col << 3) + 4, &b->mv[1][0], 4, 4, w1, h1);

  2448.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0],

  2449.                         s->dst[0] + 4 * ls_y, ls_y,

  2450.                         ref1->data[0], ref1->linesize[0], tref1,

  2451.                         (row << 3) + 4, col << 3, &b->mv[2][0], 4, 4, w1, h1);

  2452.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0],

  2453.                         s->dst[0] + 4 * ls_y + 4, ls_y,

  2454.                         ref1->data[0], ref1->linesize[0], tref1,

  2455.                         (row << 3) + 4, (col << 3) + 4, &b->mv[3][0], 4, 4, w1, h1);

  2456. 

  2457.             if (b->comp) {

  2458.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0], ls_y,

  2459.                             ref2->data[0], ref2->linesize[0], tref2,

  2460.                             row << 3, col << 3, &b->mv[0][1], 4, 4, w2, h2);

  2461.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0] + 4, ls_y,

  2462.                             ref2->data[0], ref2->linesize[0], tref2,

  2463.                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 4, w2, h2);

  2464.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1],

  2465.                             s->dst[0] + 4 * ls_y, ls_y,

  2466.                             ref2->data[0], ref2->linesize[0], tref2,

  2467.                             (row << 3) + 4, col << 3, &b->mv[2][1], 4, 4, w2, h2);

  2468.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1],

  2469.                             s->dst[0] + 4 * ls_y + 4, ls_y,

  2470.                             ref2->data[0], ref2->linesize[0], tref2,

  2471.                             (row << 3) + 4, (col << 3) + 4, &b->mv[3][1], 4, 4, w2, h2);

  2472.             }

  2473.         }

  2474.     } else {

  2475.         int bwl = bwlog_tab[0][b->bs];

  2476.         int bw = bwh_tab[0][b->bs][0] * 4, bh = bwh_tab[0][b->bs][1] * 4;

  2477. 

  2478.         mc_luma_dir(s, s->dsp.mc[bwl][b->filter][0], s->dst[0], ls_y,

  2479.                     ref1->data[0], ref1->linesize[0], tref1,

  2480.                     row << 3, col << 3, &b->mv[0][0],bw, bh, w1, h1);

  2481. 

  2482.         if (b->comp)

  2483.             mc_luma_dir(s, s->dsp.mc[bwl][b->filter][1], s->dst[0], ls_y,

  2484.                         ref2->data[0], ref2->linesize[0], tref2,

  2485.                         row << 3, col << 3, &b->mv[0][1], bw, bh, w2, h2);

  2486.     }

  2487. 

  2488.     // uv inter pred

  2489.     {

  2490.         int bwl = bwlog_tab[1][b->bs];

  2491.         int bw = bwh_tab[1][b->bs][0] * 4, bh = bwh_tab[1][b->bs][1] * 4;

  2492.         VP56mv mvuv;

  2493. 

  2494.         w1 = (w1 + 1) >> 1;

  2495.         h1 = (h1 + 1) >> 1;

  2496.         if (b->comp) {

  2497.             w2 = (w2 + 1) >> 1;

  2498.             h2 = (h2 + 1) >> 1;

  2499.         }

  2500.         if (b->bs > BS_8x8) {

  2501.             mvuv.x = ROUNDED_DIV(b->mv[0][0].x + b->mv[1][0].x + b->mv[2][0].x + b->mv[3][0].x, 4);

  2502.             mvuv.y = ROUNDED_DIV(b->mv[0][0].y + b->mv[1][0].y + b->mv[2][0].y + b->mv[3][0].y, 4);

  2503.         } else {

  2504.             mvuv = b->mv[0][0];

  2505.         }

  2506. 

  2507.         mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][0],

  2508.                       s->dst[1], s->dst[2], ls_uv,

  2509.                       ref1->data[1], ref1->linesize[1],

  2510.                       ref1->data[2], ref1->linesize[2], tref1,

  2511.                       row << 2, col << 2, &mvuv, bw, bh, w1, h1);

  2512. 

  2513.         if (b->comp) {

  2514.             if (b->bs > BS_8x8) {

  2515.                 mvuv.x = ROUNDED_DIV(b->mv[0][1].x + b->mv[1][1].x + b->mv[2][1].x + b->mv[3][1].x, 4);

  2516.                 mvuv.y = ROUNDED_DIV(b->mv[0][1].y + b->mv[1][1].y + b->mv[2][1].y + b->mv[3][1].y, 4);

  2517.             } else {

  2518.                 mvuv = b->mv[0][1];

  2519.             }

  2520.             mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][1],

  2521.                           s->dst[1], s->dst[2], ls_uv,

  2522.                           ref2->data[1], ref2->linesize[1],

  2523.                           ref2->data[2], ref2->linesize[2], tref2,

  2524.                           row << 2, col << 2, &mvuv, bw, bh, w2, h2);

  2525.         }

  2526.     }

  2527. 

  2528.     if (!b->skip) {

  2529.         /* mostly copied intra_reconn() */

  2530. 

  2531.         int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2532.         int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2533.         int end_x = FFMIN(2 * (s->cols - col), w4);

  2534.         int end_y = FFMIN(2 * (s->rows - row), h4);

  2535.         int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2536.         int uvstep1d = 1 << b->uvtx, p;

  2537.         uint8_t *dst = s->dst[0];

  2538. 

  2539.         // y itxfm add

  2540.         for (n = 0, y = 0; y < end_y; y += step1d) {

  2541.             uint8_t *ptr = dst;

  2542.             for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d, n += step) {

  2543.                 int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2544. 

  2545.                 if (eob)

  2546.                     s->dsp.itxfm_add[tx][DCT_DCT](ptr, s->y_stride,

  2547.                                                   s->block + 16 * n, eob);

  2548.             }

  2549.             dst += 4 * s->y_stride * step1d;

  2550.         }

  2551. 

  2552.         // uv itxfm add

  2553.         h4 >>= 1;

  2554.         w4 >>= 1;

  2555.         end_x >>= 1;

  2556.         end_y >>= 1;

  2557.         step = 1 << (b->uvtx * 2);

  2558.         for (p = 0; p < 2; p++) {

  2559.             dst = s->dst[p + 1];

  2560.             for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2561.                 uint8_t *ptr = dst;

  2562.                 for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d, n += step) {

  2563.                     int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2564. 

  2565.                     if (eob)

  2566.                         s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2567.                                                         s->uvblock[p] + 16 * n, eob);

  2568.                 }

  2569.                 dst += 4 * uvstep1d * s->uv_stride;

  2570.             }

  2571.         }

  2572.     }

  2573. }

  2574. 

  2575. static av_always_inline void mask_edges(struct VP9Filter *lflvl, int is_uv,

  2576.                                         int row_and_7, int col_and_7,

  2577.                                         int w, int h, int col_end, int row_end,

  2578.                                         enum TxfmMode tx, int skip_inter)

  2579. {

  2580.     // FIXME I'm pretty sure all loops can be replaced by a single LUT if

  2581.     // we make VP9Filter.mask uint64_t (i.e. row/col all single variable)

  2582.     // and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then

  2583.     // use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)

  2584. 

  2585.     // the intended behaviour of the vp9 loopfilter is to work on 8-pixel

  2586.     // edges. This means that for UV, we work on two subsampled blocks at

  2587.     // a time, and we only use the topleft block's mode information to set

  2588.     // things like block strength. Thus, for any block size smaller than

  2589.     // 16x16, ignore the odd portion of the block.

  2590.     if (tx == TX_4X4 && is_uv) {

  2591.         if (h == 1) {

  2592.             if (row_and_7 & 1)

  2593.                 return;

  2594.             if (!row_end)

  2595.                 h += 1;

  2596.         }

  2597.         if (w == 1) {

  2598.             if (col_and_7 & 1)

  2599.                 return;

  2600.             if (!col_end)

  2601.                 w += 1;

  2602.         }

  2603.     }

  2604. 

  2605.     if (tx == TX_4X4 && !skip_inter) {

  2606.         int t = 1 << col_and_7, m_col = (t << w) - t, y;

  2607.         int m_col_odd = (t << (w - 1)) - t;

  2608. 

  2609.         // on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide

  2610.         if (is_uv) {

  2611.             int m_row_8 = m_col & 0x01, m_row_4 = m_col - m_row_8;

  2612. 

  2613.             for (y = row_and_7; y < h + row_and_7; y++) {

  2614.                 int col_mask_id = 2 - !(y & 7);

  2615. 

  2616.                 lflvl->mask[is_uv][0][y][1] |= m_row_8;

  2617.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  2618.                 // for odd lines, if the odd col is not being filtered,

  2619.                 // skip odd row also:

  2620.                 // .---. <-- a

  2621.                 // |   |

  2622.                 // |___| <-- b

  2623.                 // ^   ^

  2624.                 // c   d

  2625.                 //

  2626.                 // if a/c are even row/col and b/d are odd, and d is skipped,

  2627.                 // e.g. right edge of size-66x66.webm, then skip b also (bug)

  2628.                 if ((col_end & 1) && (y & 1)) {

  2629.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col_odd;

  2630.                 } else {

  2631.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col;

  2632.                 }

  2633.             }

  2634.         } else {

  2635.             int m_row_8 = m_col & 0x11, m_row_4 = m_col - m_row_8;

  2636. 

  2637.             for (y = row_and_7; y < h + row_and_7; y++) {

  2638.                 int col_mask_id = 2 - !(y & 3);

  2639. 

  2640.                 lflvl->mask[is_uv][0][y][1] |= m_row_8; // row edge

  2641.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  2642.                 lflvl->mask[is_uv][1][y][col_mask_id] |= m_col; // col edge

  2643.                 lflvl->mask[is_uv][0][y][3] |= m_col;

  2644.                 lflvl->mask[is_uv][1][y][3] |= m_col;

  2645.             }

  2646.         }

  2647.     } else {

  2648.         int y, t = 1 << col_and_7, m_col = (t << w) - t;

  2649. 

  2650.         if (!skip_inter) {

  2651.             int mask_id = (tx == TX_8X8);

  2652.             int l2 = tx + is_uv - 1, step1d = 1 << l2;

  2653.             static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };

  2654.             int m_row = m_col & masks[l2];

  2655. 

  2656.             // at odd UV col/row edges tx16/tx32 loopfilter edges, force

  2657.             // 8wd loopfilter to prevent going off the visible edge.

  2658.             if (is_uv && tx > TX_8X8 && (w ^ (w - 1)) == 1) {

  2659.                 int m_row_16 = ((t << (w - 1)) - t) & masks[l2];

  2660.                 int m_row_8 = m_row - m_row_16;

  2661. 

  2662.                 for (y = row_and_7; y < h + row_and_7; y++) {

  2663.                     lflvl->mask[is_uv][0][y][0] |= m_row_16;

  2664.                     lflvl->mask[is_uv][0][y][1] |= m_row_8;

  2665.                 }

  2666.             } else {

  2667.                 for (y = row_and_7; y < h + row_and_7; y++)

  2668.                     lflvl->mask[is_uv][0][y][mask_id] |= m_row;

  2669.             }

  2670. 

  2671.             if (is_uv && tx > TX_8X8 && (h ^ (h - 1)) == 1) {

  2672.                 for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)

  2673.                     lflvl->mask[is_uv][1][y][0] |= m_col;

  2674.                 if (y - row_and_7 == h - 1)

  2675.                     lflvl->mask[is_uv][1][y][1] |= m_col;

  2676.             } else {

  2677.                 for (y = row_and_7; y < h + row_and_7; y += step1d)

  2678.                     lflvl->mask[is_uv][1][y][mask_id] |= m_col;

  2679.             }

  2680.         } else if (tx != TX_4X4) {

  2681.             int mask_id;

  2682. 

  2683.             mask_id = (tx == TX_8X8) || (is_uv && h == 1);

  2684.             lflvl->mask[is_uv][1][row_and_7][mask_id] |= m_col;

  2685.             mask_id = (tx == TX_8X8) || (is_uv && w == 1);

  2686.             for (y = row_and_7; y < h + row_and_7; y++)

  2687.                 lflvl->mask[is_uv][0][y][mask_id] |= t;

  2688.         } else if (is_uv) {

  2689.             int t8 = t & 0x01, t4 = t - t8;

  2690. 

  2691.             for (y = row_and_7; y < h + row_and_7; y++) {

  2692.                 lflvl->mask[is_uv][0][y][2] |= t4;

  2693.                 lflvl->mask[is_uv][0][y][1] |= t8;

  2694.             }

  2695.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 7)] |= m_col;

  2696.         } else {

  2697.             int t8 = t & 0x11, t4 = t - t8;

  2698. 

  2699.             for (y = row_and_7; y < h + row_and_7; y++) {

  2700.                 lflvl->mask[is_uv][0][y][2] |= t4;

  2701.                 lflvl->mask[is_uv][0][y][1] |= t8;

  2702.             }

  2703.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 3)] |= m_col;

  2704.         }

  2705.     }

  2706. }

  2707. 

  2708. static void decode_b(AVCodecContext *ctx, int row, int col,

  2709.                      struct VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,

  2710.                      enum BlockLevel bl, enum BlockPartition bp)

  2711. {

  2712.     VP9Context *s = ctx->priv_data;

  2713.     VP9Block *b = s->b;

  2714.     enum BlockSize bs = bl * 3 + bp;

  2715.     int y, w4 = bwh_tab[1][bs][0], h4 = bwh_tab[1][bs][1], lvl;

  2716.     int emu[2];

  2717.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2718. 

  2719.     s->row = row;

  2720.     s->row7 = row & 7;

  2721.     s->col = col;

  2722.     s->col7 = col & 7;

  2723.     s->min_mv.x = -(128 + col * 64);

  2724.     s->min_mv.y = -(128 + row * 64);

  2725.     s->max_mv.x = 128 + (s->cols - col - w4) * 64;

  2726.     s->max_mv.y = 128 + (s->rows - row - h4) * 64;

  2727.     if (s->pass < 2) {

  2728.         b->bs = bs;

  2729.         b->bl = bl;

  2730.         b->bp = bp;

  2731.         decode_mode(ctx);

  2732.         b->uvtx = b->tx - (w4 * 2 == (1 << b->tx) || h4 * 2 == (1 << b->tx));

  2733. 

  2734.         if (!b->skip) {

  2735.             decode_coeffs(ctx);

  2736.         } else {

  2737.             int pl;

  2738. 

  2739.             memset(&s->above_y_nnz_ctx[col * 2], 0, w4 * 2);

  2740.             memset(&s->left_y_nnz_ctx[(row & 7) << 1], 0, h4 * 2);

  2741.             for (pl = 0; pl < 2; pl++) {

  2742.                 memset(&s->above_uv_nnz_ctx[pl][col], 0, w4);

  2743.                 memset(&s->left_uv_nnz_ctx[pl][row & 7], 0, h4);

  2744.             }

  2745.         }

  2746.         if (s->pass == 1) {

  2747.             s->b++;

  2748.             s->block += w4 * h4 * 64;

  2749.             s->uvblock[0] += w4 * h4 * 16;

  2750.             s->uvblock[1] += w4 * h4 * 16;

  2751.             s->eob += 4 * w4 * h4;

  2752.             s->uveob[0] += w4 * h4;

  2753.             s->uveob[1] += w4 * h4;

  2754. 

  2755.             return;

  2756.         }

  2757.     }

  2758. 

  2759.     // emulated overhangs if the stride of the target buffer can't hold. This

  2760.     // allows to support emu-edge and so on even if we have large block

  2761.     // overhangs

  2762.     emu[0] = (col + w4) * 8 > f->linesize[0] ||

  2763.              (row + h4) > s->rows;

  2764.     emu[1] = (col + w4) * 4 > f->linesize[1] ||

  2765.              (row + h4) > s->rows;

  2766.     if (emu[0]) {

  2767.         s->dst[0] = s->tmp_y;

  2768.         s->y_stride = 64;

  2769.     } else {

  2770.         s->dst[0] = f->data[0] + yoff;

  2771.         s->y_stride = f->linesize[0];

  2772.     }

  2773.     if (emu[1]) {

  2774.         s->dst[1] = s->tmp_uv[0];

  2775.         s->dst[2] = s->tmp_uv[1];

  2776.         s->uv_stride = 32;

  2777.     } else {

  2778.         s->dst[1] = f->data[1] + uvoff;

  2779.         s->dst[2] = f->data[2] + uvoff;

  2780.         s->uv_stride = f->linesize[1];

  2781.     }

  2782.     if (b->intra) {

  2783.         intra_recon(ctx, yoff, uvoff);

  2784.     } else {

  2785.         inter_recon(ctx);

  2786.     }

  2787.     if (emu[0]) {

  2788.         int w = FFMIN(s->cols - col, w4) * 8, h = FFMIN(s->rows - row, h4) * 8, n, o = 0;

  2789. 

  2790.         for (n = 0; o < w; n++) {

  2791.             int bw = 64 >> n;

  2792. 

  2793.             av_assert2(n <= 4);

  2794.             if (w & bw) {

  2795.                 s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o, f->linesize[0],

  2796.                                          s->tmp_y + o, 64, h, 0, 0);

  2797.                 o += bw;

  2798.             }

  2799.         }

  2800.     }

  2801.     if (emu[1]) {

  2802.         int w = FFMIN(s->cols - col, w4) * 4, h = FFMIN(s->rows - row, h4) * 4, n, o = 0;

  2803. 

  2804.         for (n = 1; o < w; n++) {

  2805.             int bw = 64 >> n;

  2806. 

  2807.             av_assert2(n <= 4);

  2808.             if (w & bw) {

  2809.                 s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o, f->linesize[1],

  2810.                                          s->tmp_uv[0] + o, 32, h, 0, 0);

  2811.                 s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o, f->linesize[2],

  2812.                                          s->tmp_uv[1] + o, 32, h, 0, 0);

  2813.                 o += bw;

  2814.             }

  2815.         }

  2816.     }

  2817. 

  2818.     // pick filter level and find edges to apply filter to

  2819.     if (s->filter.level &&

  2820.         (lvl = s->segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]

  2821.                                                     [b->mode[3] != ZEROMV]) > 0) {

  2822.         int x_end = FFMIN(s->cols - col, w4), y_end = FFMIN(s->rows - row, h4);

  2823.         int skip_inter = !b->intra && b->skip;

  2824. 

  2825.         for (y = 0; y < h4; y++)

  2826.             memset(&lflvl->level[((row & 7) + y) * 8 + (col & 7)], lvl, w4);

  2827.         mask_edges(lflvl, 0, row & 7, col & 7, x_end, y_end, 0, 0, b->tx, skip_inter);

  2828.         mask_edges(lflvl, 1, row & 7, col & 7, x_end, y_end,

  2829.                    s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,

  2830.                    s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,

  2831.                    b->uvtx, skip_inter);

  2832. 

  2833.         if (!s->filter.lim_lut[lvl]) {

  2834.             int sharp = s->filter.sharpness;

  2835.             int limit = lvl;

  2836. 

  2837.             if (sharp > 0) {

  2838.                 limit >>= (sharp + 3) >> 2;

  2839.                 limit = FFMIN(limit, 9 - sharp);

  2840.             }

  2841.             limit = FFMAX(limit, 1);

  2842. 

  2843.             s->filter.lim_lut[lvl] = limit;

  2844.             s->filter.mblim_lut[lvl] = 2 * (lvl + 2) + limit;

  2845.         }

  2846.     }

  2847. 

  2848.     if (s->pass == 2) {

  2849.         s->b++;

  2850.         s->block += w4 * h4 * 64;

  2851.         s->uvblock[0] += w4 * h4 * 16;

  2852.         s->uvblock[1] += w4 * h4 * 16;

  2853.         s->eob += 4 * w4 * h4;

  2854.         s->uveob[0] += w4 * h4;

  2855.         s->uveob[1] += w4 * h4;

  2856.     }

  2857. }

  2858. 

  2859. static void decode_sb(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  2860.                       ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  2861. {

  2862.     VP9Context *s = ctx->priv_data;

  2863.     int c = ((s->above_partition_ctx[col] >> (3 - bl)) & 1) |

  2864.             (((s->left_partition_ctx[row & 0x7] >> (3 - bl)) & 1) << 1);

  2865.     const uint8_t *p = s->keyframe ? vp9_default_kf_partition_probs[bl][c] :

  2866.                                      s->prob.p.partition[bl][c];

  2867.     enum BlockPartition bp;

  2868.     ptrdiff_t hbs = 4 >> bl;

  2869.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2870.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  2871. 

  2872.     if (bl == BL_8X8) {

  2873.         bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  2874.         decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2875.     } else if (col + hbs < s->cols) { // FIXME why not <=?

  2876.         if (row + hbs < s->rows) { // FIXME why not <=?

  2877.             bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  2878.             switch (bp) {

  2879.             case PARTITION_NONE:

  2880.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2881.                 break;

  2882.             case PARTITION_H:

  2883.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2884.                 yoff  += hbs * 8 * y_stride;

  2885.                 uvoff += hbs * 4 * uv_stride;

  2886.                 decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, bl, bp);

  2887.                 break;

  2888.             case PARTITION_V:

  2889.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2890.                 yoff  += hbs * 8;

  2891.                 uvoff += hbs * 4;

  2892.                 decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, bl, bp);

  2893.                 break;

  2894.             case PARTITION_SPLIT:

  2895.                 decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2896.                 decode_sb(ctx, row, col + hbs, lflvl,

  2897.                           yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2898.                 yoff  += hbs * 8 * y_stride;

  2899.                 uvoff += hbs * 4 * uv_stride;

  2900.                 decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  2901.                 decode_sb(ctx, row + hbs, col + hbs, lflvl,

  2902.                           yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2903.                 break;

  2904.             default:

  2905.                 av_assert0(0);

  2906.             }

  2907.         } else if (vp56_rac_get_prob_branchy(&s->c, p[1])) {

  2908.             bp = PARTITION_SPLIT;

  2909.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2910.             decode_sb(ctx, row, col + hbs, lflvl,

  2911.                       yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2912.         } else {

  2913.             bp = PARTITION_H;

  2914.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2915.         }

  2916.     } else if (row + hbs < s->rows) { // FIXME why not <=?

  2917.         if (vp56_rac_get_prob_branchy(&s->c, p[2])) {

  2918.             bp = PARTITION_SPLIT;

  2919.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2920.             yoff  += hbs * 8 * y_stride;

  2921.             uvoff += hbs * 4 * uv_stride;

  2922.             decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  2923.         } else {

  2924.             bp = PARTITION_V;

  2925.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2926.         }

  2927.     } else {

  2928.         bp = PARTITION_SPLIT;

  2929.         decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2930.     }

  2931.     s->counts.partition[bl][c][bp]++;

  2932. }

  2933. 

  2934. static void decode_sb_mem(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  2935.                           ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  2936. {

  2937.     VP9Context *s = ctx->priv_data;

  2938.     VP9Block *b = s->b;

  2939.     ptrdiff_t hbs = 4 >> bl;

  2940.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2941.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  2942. 

  2943.     if (bl == BL_8X8) {

  2944.         av_assert2(b->bl == BL_8X8);

  2945.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  2946.     } else if (s->b->bl == bl) {

  2947.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  2948.         if (b->bp == PARTITION_H && row + hbs < s->rows) {

  2949.             yoff  += hbs * 8 * y_stride;

  2950.             uvoff += hbs * 4 * uv_stride;

  2951.             decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, b->bl, b->bp);

  2952.         } else if (b->bp == PARTITION_V && col + hbs < s->cols) {

  2953.             yoff  += hbs * 8;

  2954.             uvoff += hbs * 4;

  2955.             decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, b->bl, b->bp);

  2956.         }

  2957.     } else {

  2958.         decode_sb_mem(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2959.         if (col + hbs < s->cols) { // FIXME why not <=?

  2960.             if (row + hbs < s->rows) {

  2961.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff + 8 * hbs,

  2962.                               uvoff + 4 * hbs, bl + 1);

  2963.                 yoff  += hbs * 8 * y_stride;

  2964.                 uvoff += hbs * 4 * uv_stride;

  2965.                 decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  2966.                 decode_sb_mem(ctx, row + hbs, col + hbs, lflvl,

  2967.                                     yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2968.             } else {

  2969.                 yoff  += hbs * 8;

  2970.                 uvoff += hbs * 4;

  2971.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff, uvoff, bl + 1);

  2972.             }

  2973.         } else if (row + hbs < s->rows) {

  2974.             yoff  += hbs * 8 * y_stride;

  2975.             uvoff += hbs * 4 * uv_stride;

  2976.             decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  2977.         }

  2978.     }

  2979. }

  2980. 

  2981. static void loopfilter_sb(AVCodecContext *ctx, struct VP9Filter *lflvl,

  2982.                           int row, int col, ptrdiff_t yoff, ptrdiff_t uvoff)

  2983. {

  2984.     VP9Context *s = ctx->priv_data;

  2985.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2986.     uint8_t *dst = f->data[0] + yoff, *lvl = lflvl->level;

  2987.     ptrdiff_t ls_y = f->linesize[0], ls_uv = f->linesize[1];

  2988.     int y, x, p;

  2989. 

  2990.     // FIXME in how far can we interleave the v/h loopfilter calls? E.g.

  2991.     // if you think of them as acting on a 8x8 block max, we can interleave

  2992.     // each v/h within the single x loop, but that only works if we work on

  2993.     // 8 pixel blocks, and we won't always do that (we want at least 16px

  2994.     // to use SSE2 optimizations, perhaps 32 for AVX2)

  2995. 

  2996.     // filter edges between columns, Y plane (e.g. block1 | block2)

  2997.     for (y = 0; y < 8; y += 2, dst += 16 * ls_y, lvl += 16) {

  2998.         uint8_t *ptr = dst, *l = lvl, *hmask1 = lflvl->mask[0][0][y];

  2999.         uint8_t *hmask2 = lflvl->mask[0][0][y + 1];

  3000.         unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2], hm13 = hmask1[3];

  3001.         unsigned hm2 = hmask2[1] | hmask2[2], hm23 = hmask2[3];

  3002.         unsigned hm = hm1 | hm2 | hm13 | hm23;

  3003. 

  3004.         for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 8, l++) {

  3005.             if (hm1 & x) {

  3006.                 int L = *l, H = L >> 4;

  3007.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3008. 

  3009.                 if (col || x > 1) {

  3010.                     if (hmask1[0] & x) {

  3011.                         if (hmask2[0] & x) {

  3012.                             av_assert2(l[8] == L);

  3013.                             s->dsp.loop_filter_16[0](ptr, ls_y, E, I, H);

  3014.                         } else {

  3015.                             s->dsp.loop_filter_8[2][0](ptr, ls_y, E, I, H);

  3016.                         }

  3017.                     } else if (hm2 & x) {

  3018.                         L = l[8];

  3019.                         H |= (L >> 4) << 8;

  3020.                         E |= s->filter.mblim_lut[L] << 8;

  3021.                         I |= s->filter.lim_lut[L] << 8;

  3022.                         s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3023.                                                [!!(hmask2[1] & x)]

  3024.                                                [0](ptr, ls_y, E, I, H);

  3025.                     } else {

  3026.                         s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3027.                                             [0](ptr, ls_y, E, I, H);

  3028.                     }

  3029.                 }

  3030.             } else if (hm2 & x) {

  3031.                 int L = l[8], H = L >> 4;

  3032.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3033. 

  3034.                 if (col || x > 1) {

  3035.                     s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3036.                                         [0](ptr + 8 * ls_y, ls_y, E, I, H);

  3037.                 }

  3038.             }

  3039.             if (hm13 & x) {

  3040.                 int L = *l, H = L >> 4;

  3041.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3042. 

  3043.                 if (hm23 & x) {

  3044.                     L = l[8];

  3045.                     H |= (L >> 4) << 8;

  3046.                     E |= s->filter.mblim_lut[L] << 8;

  3047.                     I |= s->filter.lim_lut[L] << 8;

  3048.                     s->dsp.loop_filter_mix2[0][0][0](ptr + 4, ls_y, E, I, H);

  3049.                 } else {

  3050.                     s->dsp.loop_filter_8[0][0](ptr + 4, ls_y, E, I, H);

  3051.                 }

  3052.             } else if (hm23 & x) {

  3053.                 int L = l[8], H = L >> 4;

  3054.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3055. 

  3056.                 s->dsp.loop_filter_8[0][0](ptr + 8 * ls_y + 4, ls_y, E, I, H);

  3057.             }

  3058.         }

  3059.     }

  3060. 

  3061.     //                                          block1

  3062.     // filter edges between rows, Y plane (e.g. ------)

  3063.     //                                          block2

  3064.     dst = f->data[0] + yoff;

  3065.     lvl = lflvl->level;

  3066.     for (y = 0; y < 8; y++, dst += 8 * ls_y, lvl += 8) {

  3067.         uint8_t *ptr = dst, *l = lvl, *vmask = lflvl->mask[0][1][y];

  3068.         unsigned vm = vmask[0] | vmask[1] | vmask[2], vm3 = vmask[3];

  3069. 

  3070.         for (x = 1; vm & ~(x - 1); x <<= 2, ptr += 16, l += 2) {

  3071.             if (row || y) {

  3072.                 if (vm & x) {

  3073.                     int L = *l, H = L >> 4;

  3074.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3075. 

  3076.                     if (vmask[0] & x) {

  3077.                         if (vmask[0] & (x << 1)) {

  3078.                             av_assert2(l[1] == L);

  3079.                             s->dsp.loop_filter_16[1](ptr, ls_y, E, I, H);

  3080.                         } else {

  3081.                             s->dsp.loop_filter_8[2][1](ptr, ls_y, E, I, H);

  3082.                         }

  3083.                     } else if (vm & (x << 1)) {

  3084.                         L = l[1];

  3085.                         H |= (L >> 4) << 8;

  3086.                         E |= s->filter.mblim_lut[L] << 8;

  3087.                         I |= s->filter.lim_lut[L] << 8;

  3088.                         s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3089.                                                [!!(vmask[1] & (x << 1))]

  3090.                                                [1](ptr, ls_y, E, I, H);

  3091.                     } else {

  3092.                         s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3093.                                             [1](ptr, ls_y, E, I, H);

  3094.                     }

  3095.                 } else if (vm & (x << 1)) {

  3096.                     int L = l[1], H = L >> 4;

  3097.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3098. 

  3099.                     s->dsp.loop_filter_8[!!(vmask[1] & (x << 1))]

  3100.                                         [1](ptr + 8, ls_y, E, I, H);

  3101.                 }

  3102.             }

  3103.             if (vm3 & x) {

  3104.                 int L = *l, H = L >> 4;

  3105.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3106. 

  3107.                 if (vm3 & (x << 1)) {

  3108.                     L = l[1];

  3109.                     H |= (L >> 4) << 8;

  3110.                     E |= s->filter.mblim_lut[L] << 8;

  3111.                     I |= s->filter.lim_lut[L] << 8;

  3112.                     s->dsp.loop_filter_mix2[0][0][1](ptr + ls_y * 4, ls_y, E, I, H);

  3113.                 } else {

  3114.                     s->dsp.loop_filter_8[0][1](ptr + ls_y * 4, ls_y, E, I, H);

  3115.                 }

  3116.             } else if (vm3 & (x << 1)) {

  3117.                 int L = l[1], H = L >> 4;

  3118.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3119. 

  3120.                 s->dsp.loop_filter_8[0][1](ptr + ls_y * 4 + 8, ls_y, E, I, H);

  3121.             }

  3122.         }

  3123.     }

  3124. 

  3125.     // same principle but for U/V planes

  3126.     for (p = 0; p < 2; p++) {

  3127.         lvl = lflvl->level;

  3128.         dst = f->data[1 + p] + uvoff;

  3129.         for (y = 0; y < 8; y += 4, dst += 16 * ls_uv, lvl += 32) {

  3130.             uint8_t *ptr = dst, *l = lvl, *hmask1 = lflvl->mask[1][0][y];

  3131.             uint8_t *hmask2 = lflvl->mask[1][0][y + 2];

  3132.             unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2];

  3133.             unsigned hm2 = hmask2[1] | hmask2[2], hm = hm1 | hm2;

  3134. 

  3135.             for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 4) {

  3136.                 if (col || x > 1) {

  3137.                     if (hm1 & x) {

  3138.                         int L = *l, H = L >> 4;

  3139.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3140. 

  3141.                         if (hmask1[0] & x) {

  3142.                             if (hmask2[0] & x) {

  3143.                                 av_assert2(l[16] == L);

  3144.                                 s->dsp.loop_filter_16[0](ptr, ls_uv, E, I, H);

  3145.                             } else {

  3146.                                 s->dsp.loop_filter_8[2][0](ptr, ls_uv, E, I, H);

  3147.                             }

  3148.                         } else if (hm2 & x) {

  3149.                             L = l[16];

  3150.                             H |= (L >> 4) << 8;

  3151.                             E |= s->filter.mblim_lut[L] << 8;

  3152.                             I |= s->filter.lim_lut[L] << 8;

  3153.                             s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3154.                                                    [!!(hmask2[1] & x)]

  3155.                                                    [0](ptr, ls_uv, E, I, H);

  3156.                         } else {

  3157.                             s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3158.                                                 [0](ptr, ls_uv, E, I, H);

  3159.                         }

  3160.                     } else if (hm2 & x) {

  3161.                         int L = l[16], H = L >> 4;

  3162.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3163. 

  3164.                         s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3165.                                             [0](ptr + 8 * ls_uv, ls_uv, E, I, H);

  3166.                     }

  3167.                 }

  3168.                 if (x & 0xAA)

  3169.                     l += 2;

  3170.             }

  3171.         }

  3172.         lvl = lflvl->level;

  3173.         dst = f->data[1 + p] + uvoff;

  3174.         for (y = 0; y < 8; y++, dst += 4 * ls_uv) {

  3175.             uint8_t *ptr = dst, *l = lvl, *vmask = lflvl->mask[1][1][y];

  3176.             unsigned vm = vmask[0] | vmask[1] | vmask[2];

  3177. 

  3178.             for (x = 1; vm & ~(x - 1); x <<= 4, ptr += 16, l += 4) {

  3179.                 if (row || y) {

  3180.                     if (vm & x) {

  3181.                         int L = *l, H = L >> 4;

  3182.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3183. 

  3184.                         if (vmask[0] & x) {

  3185.                             if (vmask[0] & (x << 2)) {

  3186.                                 av_assert2(l[2] == L);

  3187.                                 s->dsp.loop_filter_16[1](ptr, ls_uv, E, I, H);

  3188.                             } else {

  3189.                                 s->dsp.loop_filter_8[2][1](ptr, ls_uv, E, I, H);

  3190.                             }

  3191.                         } else if (vm & (x << 2)) {

  3192.                             L = l[2];

  3193.                             H |= (L >> 4) << 8;

  3194.                             E |= s->filter.mblim_lut[L] << 8;

  3195.                             I |= s->filter.lim_lut[L] << 8;

  3196.                             s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3197.                                                    [!!(vmask[1] & (x << 2))]

  3198.                                                    [1](ptr, ls_uv, E, I, H);

  3199.                         } else {

  3200.                             s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3201.                                                 [1](ptr, ls_uv, E, I, H);

  3202.                         }

  3203.                     } else if (vm & (x << 2)) {

  3204.                         int L = l[2], H = L >> 4;

  3205.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3206. 

  3207.                         s->dsp.loop_filter_8[!!(vmask[1] & (x << 2))]

  3208.                                             [1](ptr + 8, ls_uv, E, I, H);

  3209.                     }

  3210.                 }

  3211.             }

  3212.             if (y & 1)

  3213.                 lvl += 16;

  3214.         }

  3215.     }

  3216. }

  3217. 

  3218. static void set_tile_offset(int *start, int *end, int idx, int log2_n, int n)

  3219. {

  3220.     int sb_start = ( idx      * n) >> log2_n;

  3221.     int sb_end   = ((idx + 1) * n) >> log2_n;

  3222.     *start = FFMIN(sb_start, n) << 3;

  3223.     *end   = FFMIN(sb_end,   n) << 3;

  3224. }

  3225. 

  3226. static av_always_inline void adapt_prob(uint8_t *p, unsigned ct0, unsigned ct1,

  3227.                                         int max_count, int update_factor)

  3228. {

  3229.     unsigned ct = ct0 + ct1, p2, p1;

  3230. 

  3231.     if (!ct)

  3232.         return;

  3233. 

  3234.     p1 = *p;

  3235.     p2 = ((ct0 << 8) + (ct >> 1)) / ct;

  3236.     p2 = av_clip(p2, 1, 255);

  3237.     ct = FFMIN(ct, max_count);

  3238.     update_factor = FASTDIV(update_factor * ct, max_count);

  3239. 

  3240.     // (p1 * (256 - update_factor) + p2 * update_factor + 128) >> 8

  3241.     *p = p1 + (((p2 - p1) * update_factor + 128) >> 8);

  3242. }

  3243. 

  3244. static void adapt_probs(VP9Context *s)

  3245. {

  3246.     int i, j, k, l, m;

  3247.     prob_context *p = &s->prob_ctx[s->framectxid].p;

  3248.     int uf = (s->keyframe || s->intraonly || !s->last_keyframe) ? 112 : 128;

  3249. 

  3250.     // coefficients

  3251.     for (i = 0; i < 4; i++)

  3252.         for (j = 0; j < 2; j++)

  3253.             for (k = 0; k < 2; k++)

  3254.                 for (l = 0; l < 6; l++)

  3255.                     for (m = 0; m < 6; m++) {

  3256.                         uint8_t *pp = s->prob_ctx[s->framectxid].coef[i][j][k][l][m];

  3257.                         unsigned *e = s->counts.eob[i][j][k][l][m];

  3258.                         unsigned *c = s->counts.coef[i][j][k][l][m];

  3259. 

  3260.                         if (l == 0 && m >= 3) // dc only has 3 pt

  3261.                             break;

  3262. 

  3263.                         adapt_prob(&pp[0], e[0], e[1], 24, uf);

  3264.                         adapt_prob(&pp[1], c[0], c[1] + c[2], 24, uf);

  3265.                         adapt_prob(&pp[2], c[1], c[2], 24, uf);

  3266.                     }

  3267. 

  3268.     if (s->keyframe || s->intraonly) {

  3269.         memcpy(p->skip,  s->prob.p.skip,  sizeof(p->skip));

  3270.         memcpy(p->tx32p, s->prob.p.tx32p, sizeof(p->tx32p));

  3271.         memcpy(p->tx16p, s->prob.p.tx16p, sizeof(p->tx16p));

  3272.         memcpy(p->tx8p,  s->prob.p.tx8p,  sizeof(p->tx8p));

  3273.         return;

  3274.     }

  3275. 

  3276.     // skip flag

  3277.     for (i = 0; i < 3; i++)

  3278.         adapt_prob(&p->skip[i], s->counts.skip[i][0], s->counts.skip[i][1], 20, 128);

  3279. 

  3280.     // intra/inter flag

  3281.     for (i = 0; i < 4; i++)

  3282.         adapt_prob(&p->intra[i], s->counts.intra[i][0], s->counts.intra[i][1], 20, 128);

  3283. 

  3284.     // comppred flag

  3285.     if (s->comppredmode == PRED_SWITCHABLE) {

  3286.       for (i = 0; i < 5; i++)

  3287.           adapt_prob(&p->comp[i], s->counts.comp[i][0], s->counts.comp[i][1], 20, 128);

  3288.     }

  3289. 

  3290.     // reference frames

  3291.     if (s->comppredmode != PRED_SINGLEREF) {

  3292.       for (i = 0; i < 5; i++)

  3293.           adapt_prob(&p->comp_ref[i], s->counts.comp_ref[i][0],

  3294.                      s->counts.comp_ref[i][1], 20, 128);

  3295.     }

  3296. 

  3297.     if (s->comppredmode != PRED_COMPREF) {

  3298.       for (i = 0; i < 5; i++) {

  3299.           uint8_t *pp = p->single_ref[i];

  3300.           unsigned (*c)[2] = s->counts.single_ref[i];

  3301. 

  3302.           adapt_prob(&pp[0], c[0][0], c[0][1], 20, 128);

  3303.           adapt_prob(&pp[1], c[1][0], c[1][1], 20, 128);

  3304.       }

  3305.     }

  3306. 

  3307.     // block partitioning

  3308.     for (i = 0; i < 4; i++)

  3309.         for (j = 0; j < 4; j++) {

  3310.             uint8_t *pp = p->partition[i][j];

  3311.             unsigned *c = s->counts.partition[i][j];

  3312. 

  3313.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3314.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3315.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3316.         }

  3317. 

  3318.     // tx size

  3319.     if (s->txfmmode == TX_SWITCHABLE) {

  3320.       for (i = 0; i < 2; i++) {

  3321.           unsigned *c16 = s->counts.tx16p[i], *c32 = s->counts.tx32p[i];

  3322. 

  3323.           adapt_prob(&p->tx8p[i], s->counts.tx8p[i][0], s->counts.tx8p[i][1], 20, 128);

  3324.           adapt_prob(&p->tx16p[i][0], c16[0], c16[1] + c16[2], 20, 128);

  3325.           adapt_prob(&p->tx16p[i][1], c16[1], c16[2], 20, 128);

  3326.           adapt_prob(&p->tx32p[i][0], c32[0], c32[1] + c32[2] + c32[3], 20, 128);

  3327.           adapt_prob(&p->tx32p[i][1], c32[1], c32[2] + c32[3], 20, 128);

  3328.           adapt_prob(&p->tx32p[i][2], c32[2], c32[3], 20, 128);

  3329.       }

  3330.     }

  3331. 

  3332.     // interpolation filter

  3333.     if (s->filtermode == FILTER_SWITCHABLE) {

  3334.         for (i = 0; i < 4; i++) {

  3335.             uint8_t *pp = p->filter[i];

  3336.             unsigned *c = s->counts.filter[i];

  3337. 

  3338.             adapt_prob(&pp[0], c[0], c[1] + c[2], 20, 128);

  3339.             adapt_prob(&pp[1], c[1], c[2], 20, 128);

  3340.         }

  3341.     }

  3342. 

  3343.     // inter modes

  3344.     for (i = 0; i < 7; i++) {

  3345.         uint8_t *pp = p->mv_mode[i];

  3346.         unsigned *c = s->counts.mv_mode[i];

  3347. 

  3348.         adapt_prob(&pp[0], c[2], c[1] + c[0] + c[3], 20, 128);

  3349.         adapt_prob(&pp[1], c[0], c[1] + c[3], 20, 128);

  3350.         adapt_prob(&pp[2], c[1], c[3], 20, 128);

  3351.     }

  3352. 

  3353.     // mv joints

  3354.     {

  3355.         uint8_t *pp = p->mv_joint;

  3356.         unsigned *c = s->counts.mv_joint;

  3357. 

  3358.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3359.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3360.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3361.     }

  3362. 

  3363.     // mv components

  3364.     for (i = 0; i < 2; i++) {

  3365.         uint8_t *pp;

  3366.         unsigned *c, (*c2)[2], sum;

  3367. 

  3368.         adapt_prob(&p->mv_comp[i].sign, s->counts.mv_comp[i].sign[0],

  3369.                    s->counts.mv_comp[i].sign[1], 20, 128);

  3370. 

  3371.         pp = p->mv_comp[i].classes;

  3372.         c = s->counts.mv_comp[i].classes;

  3373.         sum = c[1] + c[2] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9] + c[10];

  3374.         adapt_prob(&pp[0], c[0], sum, 20, 128);

  3375.         sum -= c[1];

  3376.         adapt_prob(&pp[1], c[1], sum, 20, 128);

  3377.         sum -= c[2] + c[3];

  3378.         adapt_prob(&pp[2], c[2] + c[3], sum, 20, 128);

  3379.         adapt_prob(&pp[3], c[2], c[3], 20, 128);

  3380.         sum -= c[4] + c[5];

  3381.         adapt_prob(&pp[4], c[4] + c[5], sum, 20, 128);

  3382.         adapt_prob(&pp[5], c[4], c[5], 20, 128);

  3383.         sum -= c[6];

  3384.         adapt_prob(&pp[6], c[6], sum, 20, 128);

  3385.         adapt_prob(&pp[7], c[7] + c[8], c[9] + c[10], 20, 128);

  3386.         adapt_prob(&pp[8], c[7], c[8], 20, 128);

  3387.         adapt_prob(&pp[9], c[9], c[10], 20, 128);

  3388. 

  3389.         adapt_prob(&p->mv_comp[i].class0, s->counts.mv_comp[i].class0[0],

  3390.                    s->counts.mv_comp[i].class0[1], 20, 128);

  3391.         pp = p->mv_comp[i].bits;

  3392.         c2 = s->counts.mv_comp[i].bits;

  3393.         for (j = 0; j < 10; j++)

  3394.             adapt_prob(&pp[j], c2[j][0], c2[j][1], 20, 128);

  3395. 

  3396.         for (j = 0; j < 2; j++) {

  3397.             pp = p->mv_comp[i].class0_fp[j];

  3398.             c = s->counts.mv_comp[i].class0_fp[j];

  3399.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3400.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3401.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3402.         }

  3403.         pp = p->mv_comp[i].fp;

  3404.         c = s->counts.mv_comp[i].fp;

  3405.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3406.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3407.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3408. 

  3409.         if (s->highprecisionmvs) {

  3410.             adapt_prob(&p->mv_comp[i].class0_hp, s->counts.mv_comp[i].class0_hp[0],

  3411.                        s->counts.mv_comp[i].class0_hp[1], 20, 128);

  3412.             adapt_prob(&p->mv_comp[i].hp, s->counts.mv_comp[i].hp[0],

  3413.                        s->counts.mv_comp[i].hp[1], 20, 128);

  3414.         }

  3415.     }

  3416. 

  3417.     // y intra modes

  3418.     for (i = 0; i < 4; i++) {

  3419.         uint8_t *pp = p->y_mode[i];

  3420.         unsigned *c = s->counts.y_mode[i], sum, s2;

  3421. 

  3422.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3423.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3424.         sum -= c[TM_VP8_PRED];

  3425.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3426.         sum -= c[VERT_PRED];

  3427.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3428.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3429.         sum -= s2;

  3430.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3431.         s2 -= c[HOR_PRED];

  3432.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3433.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3434.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3435.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3436.         sum -= c[VERT_LEFT_PRED];

  3437.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3438.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3439.     }

  3440. 

  3441.     // uv intra modes

  3442.     for (i = 0; i < 10; i++) {

  3443.         uint8_t *pp = p->uv_mode[i];

  3444.         unsigned *c = s->counts.uv_mode[i], sum, s2;

  3445. 

  3446.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3447.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3448.         sum -= c[TM_VP8_PRED];

  3449.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3450.         sum -= c[VERT_PRED];

  3451.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3452.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3453.         sum -= s2;

  3454.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3455.         s2 -= c[HOR_PRED];

  3456.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3457.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3458.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3459.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3460.         sum -= c[VERT_LEFT_PRED];

  3461.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3462.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3463.     }

  3464. }

  3465. 

  3466. static void free_buffers(VP9Context *s)

  3467. {

  3468.     av_freep(&s->above_partition_ctx);

  3469.     av_freep(&s->b_base);

  3470.     av_freep(&s->block_base);

  3471. }

  3472. 

  3473. static av_cold int vp9_decode_free(AVCodecContext *ctx)

  3474. {

  3475.     VP9Context *s = ctx->priv_data;

  3476.     int i;

  3477. 

  3478.     for (i = 0; i < 2; i++) {

  3479.         if (s->frames[i].tf.f->data[0])

  3480.             vp9_unref_frame(ctx, &s->frames[i]);

  3481.         av_frame_free(&s->frames[i].tf.f);

  3482.     }

  3483.     for (i = 0; i < 8; i++) {

  3484.         if (s->refs[i].f->data[0])

  3485.             ff_thread_release_buffer(ctx, &s->refs[i]);

  3486.         av_frame_free(&s->refs[i].f);

  3487.         if (s->next_refs[i].f->data[0])

  3488.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3489.         av_frame_free(&s->next_refs[i].f);

  3490.     }

  3491.     free_buffers(s);

  3492.     av_freep(&s->c_b);

  3493.     s->c_b_size = 0;

  3494. 

  3495.     return 0;

  3496. }

  3497. 

  3498. 

  3499. static int vp9_decode_frame(AVCodecContext *ctx, void *frame,

  3500.                             int *got_frame, AVPacket *pkt)

  3501. {

  3502.     const uint8_t *data = pkt->data;

  3503.     int size = pkt->size;

  3504.     VP9Context *s = ctx->priv_data;

  3505.     int res, tile_row, tile_col, i, ref, row, col;

  3506.     ptrdiff_t yoff, uvoff, ls_y, ls_uv;

  3507.     AVFrame *f;

  3508. 

  3509.     if ((res = decode_frame_header(ctx, data, size, &ref)) < 0) {

  3510.         return res;

  3511.     } else if (res == 0) {

  3512.         if (!s->refs[ref].f->data[0]) {

  3513.             av_log(ctx, AV_LOG_ERROR, "Requested reference %d not available\n", ref);

  3514.             return AVERROR_INVALIDDATA;

  3515.         }

  3516.         if ((res = av_frame_ref(frame, s->refs[ref].f)) < 0)

  3517.             return res;

  3518.         *got_frame = 1;

  3519.         return 0;

  3520.     }

  3521.     data += res;

  3522.     size -= res;

  3523. 

  3524.     if (s->frames[LAST_FRAME].tf.f->data[0])

  3525.         vp9_unref_frame(ctx, &s->frames[LAST_FRAME]);

  3526.     if (!s->keyframe && s->frames[CUR_FRAME].tf.f->data[0] &&

  3527.         (res = vp9_ref_frame(ctx, &s->frames[LAST_FRAME], &s->frames[CUR_FRAME])) < 0)

  3528.         return res;

  3529.     if (s->frames[CUR_FRAME].tf.f->data[0])

  3530.         vp9_unref_frame(ctx, &s->frames[CUR_FRAME]);

  3531.     if ((res = vp9_alloc_frame(ctx, &s->frames[CUR_FRAME])) < 0)

  3532.         return res;

  3533.     f = s->frames[CUR_FRAME].tf.f;

  3534.     f->key_frame = s->keyframe;

  3535.     f->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;

  3536.     ls_y = f->linesize[0];

  3537.     ls_uv =f->linesize[1];

  3538. 

  3539.     // ref frame setup

  3540.     for (i = 0; i < 8; i++) {

  3541.         if (s->next_refs[i].f->data[0])

  3542.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3543.         if (s->refreshrefmask & (1 << i)) {

  3544.             res = ff_thread_ref_frame(&s->next_refs[i], &s->frames[CUR_FRAME].tf);

  3545.         } else {

  3546.             res = ff_thread_ref_frame(&s->next_refs[i], &s->refs[i]);

  3547.         }

  3548.         if (res < 0)

  3549.             return res;

  3550.     }

  3551. 

  3552.     // main tile decode loop

  3553.     memset(s->above_partition_ctx, 0, s->cols);

  3554.     memset(s->above_skip_ctx, 0, s->cols);

  3555.     if (s->keyframe || s->intraonly) {

  3556.         memset(s->above_mode_ctx, DC_PRED, s->cols * 2);

  3557.     } else {

  3558.         memset(s->above_mode_ctx, NEARESTMV, s->cols);

  3559.     }

  3560.     memset(s->above_y_nnz_ctx, 0, s->sb_cols * 16);

  3561.     memset(s->above_uv_nnz_ctx[0], 0, s->sb_cols * 8);

  3562.     memset(s->above_uv_nnz_ctx[1], 0, s->sb_cols * 8);

  3563.     memset(s->above_segpred_ctx, 0, s->cols);

  3564.     s->pass = s->uses_2pass =

  3565.         ctx->active_thread_type == FF_THREAD_FRAME && s->refreshctx && !s->parallelmode;

  3566.     if (s->refreshctx && s->parallelmode) {

  3567.         int j, k, l, m;

  3568. 

  3569.         for (i = 0; i < 4; i++) {

  3570.             for (j = 0; j < 2; j++)

  3571.                 for (k = 0; k < 2; k++)

  3572.                     for (l = 0; l < 6; l++)

  3573.                         for (m = 0; m < 6; m++)

  3574.                             memcpy(s->prob_ctx[s->framectxid].coef[i][j][k][l][m],

  3575.                                    s->prob.coef[i][j][k][l][m], 3);

  3576.             if (s->txfmmode == i)

  3577.                 break;

  3578.         }

  3579.         s->prob_ctx[s->framectxid].p = s->prob.p;

  3580.         ff_thread_finish_setup(ctx);

  3581.     }

  3582. 

  3583.     do {

  3584.         yoff = uvoff = 0;

  3585.         s->b = s->b_base;

  3586.         s->block = s->block_base;

  3587.         s->uvblock[0] = s->uvblock_base[0];

  3588.         s->uvblock[1] = s->uvblock_base[1];

  3589.         s->eob = s->eob_base;

  3590.         s->uveob[0] = s->uveob_base[0];

  3591.         s->uveob[1] = s->uveob_base[1];

  3592. 

  3593.         for (tile_row = 0; tile_row < s->tiling.tile_rows; tile_row++) {

  3594.             set_tile_offset(&s->tiling.tile_row_start, &s->tiling.tile_row_end,

  3595.                             tile_row, s->tiling.log2_tile_rows, s->sb_rows);

  3596.             if (s->pass != 2) {

  3597.                 for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  3598.                     unsigned tile_size;

  3599. 

  3600.                     if (tile_col == s->tiling.tile_cols - 1 &&

  3601.                         tile_row == s->tiling.tile_rows - 1) {

  3602.                         tile_size = size;

  3603.                     } else {

  3604.                         tile_size = AV_RB32(data);

  3605.                         data += 4;

  3606.                         size -= 4;

  3607.                     }

  3608.                     if (tile_size > size) {

  3609.                         ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  3610.                         return AVERROR_INVALIDDATA;

  3611.                     }

  3612.                     ff_vp56_init_range_decoder(&s->c_b[tile_col], data, tile_size);

  3613.                     if (vp56_rac_get_prob_branchy(&s->c_b[tile_col], 128)) { // marker bit

  3614.                         ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  3615.                         return AVERROR_INVALIDDATA;

  3616.                     }

  3617.                     data += tile_size;

  3618.                     size -= tile_size;

  3619.                 }

  3620.             }

  3621. 

  3622.             for (row = s->tiling.tile_row_start; row < s->tiling.tile_row_end;

  3623.                  row += 8, yoff += ls_y * 64, uvoff += ls_uv * 32) {

  3624.                 struct VP9Filter *lflvl_ptr = s->lflvl;

  3625.                 ptrdiff_t yoff2 = yoff, uvoff2 = uvoff;

  3626. 

  3627.                 for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  3628.                     set_tile_offset(&s->tiling.tile_col_start, &s->tiling.tile_col_end,

  3629.                                     tile_col, s->tiling.log2_tile_cols, s->sb_cols);

  3630. 

  3631.                     if (s->pass != 2) {

  3632.                         memset(s->left_partition_ctx, 0, 8);

  3633.                         memset(s->left_skip_ctx, 0, 8);

  3634.                         if (s->keyframe || s->intraonly) {

  3635.                             memset(s->left_mode_ctx, DC_PRED, 16);

  3636.                         } else {

  3637.                             memset(s->left_mode_ctx, NEARESTMV, 8);

  3638.                         }

  3639.                         memset(s->left_y_nnz_ctx, 0, 16);

  3640.                         memset(s->left_uv_nnz_ctx, 0, 16);

  3641.                         memset(s->left_segpred_ctx, 0, 8);

  3642. 

  3643.                         memcpy(&s->c, &s->c_b[tile_col], sizeof(s->c));

  3644.                     }

  3645. 

  3646.                     for (col = s->tiling.tile_col_start;

  3647.                          col < s->tiling.tile_col_end;

  3648.                          col += 8, yoff2 += 64, uvoff2 += 32, lflvl_ptr++) {

  3649.                         // FIXME integrate with lf code (i.e. zero after each

  3650.                         // use, similar to invtxfm coefficients, or similar)

  3651.                         if (s->pass != 1) {

  3652.                             memset(lflvl_ptr->mask, 0, sizeof(lflvl_ptr->mask));

  3653.                         }

  3654. 

  3655.                         if (s->pass == 2) {

  3656.                             decode_sb_mem(ctx, row, col, lflvl_ptr,

  3657.                                           yoff2, uvoff2, BL_64X64);

  3658.                         } else {

  3659.                             decode_sb(ctx, row, col, lflvl_ptr,

  3660.                                       yoff2, uvoff2, BL_64X64);

  3661.                         }

  3662.                     }

  3663.                     if (s->pass != 2) {

  3664.                         memcpy(&s->c_b[tile_col], &s->c, sizeof(s->c));

  3665.                     }

  3666.                 }

  3667. 

  3668.                 if (s->pass == 1) {

  3669.                     continue;

  3670.                 }

  3671. 

  3672.                 // backup pre-loopfilter reconstruction data for intra

  3673.                 // prediction of next row of sb64s

  3674.                 if (row + 8 < s->rows) {

  3675.                     memcpy(s->intra_pred_data[0],

  3676.                            f->data[0] + yoff + 63 * ls_y,

  3677.                            8 * s->cols);

  3678.                     memcpy(s->intra_pred_data[1],

  3679.                            f->data[1] + uvoff + 31 * ls_uv,

  3680.                            4 * s->cols);

  3681.                     memcpy(s->intra_pred_data[2],

  3682.                            f->data[2] + uvoff + 31 * ls_uv,

  3683.                            4 * s->cols);

  3684.                 }

  3685. 

  3686.                 // loopfilter one row

  3687.                 if (s->filter.level) {

  3688.                     yoff2 = yoff;

  3689.                     uvoff2 = uvoff;

  3690.                     lflvl_ptr = s->lflvl;

  3691.                     for (col = 0; col < s->cols;

  3692.                          col += 8, yoff2 += 64, uvoff2 += 32, lflvl_ptr++) {

  3693.                         loopfilter_sb(ctx, lflvl_ptr, row, col, yoff2, uvoff2);

  3694.                     }

  3695.                 }

  3696. 

  3697.                 // FIXME maybe we can make this more finegrained by running the

  3698.                 // loopfilter per-block instead of after each sbrow

  3699.                 // In fact that would also make intra pred left preparation easier?

  3700.                 ff_thread_report_progress(&s->frames[CUR_FRAME].tf, row >> 3, 0);

  3701.             }

  3702.         }

  3703. 

  3704.         if (s->pass < 2 && s->refreshctx && !s->parallelmode) {

  3705.             adapt_probs(s);

  3706.             ff_thread_finish_setup(ctx);

  3707.         }

  3708.     } while (s->pass++ == 1);

  3709.     ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  3710. 

  3711.     // ref frame setup

  3712.     for (i = 0; i < 8; i++) {

  3713.         if (s->refs[i].f->data[0])

  3714.             ff_thread_release_buffer(ctx, &s->refs[i]);

  3715.         ff_thread_ref_frame(&s->refs[i], &s->next_refs[i]);

  3716.     }

  3717. 

  3718.     if (!s->invisible) {

  3719.         if ((res = av_frame_ref(frame, s->frames[CUR_FRAME].tf.f)) < 0)

  3720.             return res;

  3721.         *got_frame = 1;

  3722.     }

  3723. 

  3724.     return 0;

  3725. }

  3726. 

  3727. static void vp9_decode_flush(AVCodecContext *ctx)

  3728. {

  3729.     VP9Context *s = ctx->priv_data;

  3730.     int i;

  3731. 

  3732.     for (i = 0; i < 2; i++)

  3733.         vp9_unref_frame(ctx, &s->frames[i]);

  3734.     for (i = 0; i < 8; i++)

  3735.         ff_thread_release_buffer(ctx, &s->refs[i]);

  3736. }

  3737. 

  3738. static int init_frames(AVCodecContext *ctx)

  3739. {

  3740.     VP9Context *s = ctx->priv_data;

  3741.     int i;

  3742. 

  3743.     for (i = 0; i < 2; i++) {

  3744.         s->frames[i].tf.f = av_frame_alloc();

  3745.         if (!s->frames[i].tf.f) {

  3746.             vp9_decode_free(ctx);

  3747.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  3748.             return AVERROR(ENOMEM);

  3749.         }

  3750.     }

  3751.     for (i = 0; i < 8; i++) {

  3752.         s->refs[i].f = av_frame_alloc();

  3753.         s->next_refs[i].f = av_frame_alloc();

  3754.         if (!s->refs[i].f || !s->next_refs[i].f) {

  3755.             vp9_decode_free(ctx);

  3756.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  3757.             return AVERROR(ENOMEM);

  3758.         }

  3759.     }

  3760. 

  3761.     return 0;

  3762. }

  3763. 

  3764. static av_cold int vp9_decode_init(AVCodecContext *ctx)

  3765. {

  3766.     VP9Context *s = ctx->priv_data;

  3767. 

  3768.     ctx->internal->allocate_progress = 1;

  3769.     ctx->pix_fmt = AV_PIX_FMT_YUV420P;

  3770.     ff_vp9dsp_init(&s->dsp);

  3771.     ff_videodsp_init(&s->vdsp, 8);

  3772.     s->filter.sharpness = -1;

  3773. 

  3774.     return init_frames(ctx);

  3775. }

  3776. 

  3777. static av_cold int vp9_decode_init_thread_copy(AVCodecContext *avctx)

  3778. {

  3779.     return init_frames(avctx);

  3780. }

  3781. 

  3782. static int vp9_decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src)

  3783. {

  3784.     int i, res;

  3785.     VP9Context *s = dst->priv_data, *ssrc = src->priv_data;

  3786. 

  3787.     // detect size changes in other threads

  3788.     if (s->above_partition_ctx &&

  3789.         (!ssrc->above_partition_ctx || s->cols != ssrc->cols || s->rows != ssrc->rows)) {

  3790.         free_buffers(s);

  3791.     }

  3792. 

  3793.     for (i = 0; i < 2; i++) {

  3794.         if (s->frames[i].tf.f->data[0])

  3795.             vp9_unref_frame(dst, &s->frames[i]);

  3796.         if (ssrc->frames[i].tf.f->data[0]) {

  3797.             if ((res = vp9_ref_frame(dst, &s->frames[i], &ssrc->frames[i])) < 0)

  3798.                 return res;

  3799.         }

  3800.     }

  3801.     for (i = 0; i < 8; i++) {

  3802.         if (s->refs[i].f->data[0])

  3803.             ff_thread_release_buffer(dst, &s->refs[i]);

  3804.         if (ssrc->next_refs[i].f->data[0]) {

  3805.             if ((res = ff_thread_ref_frame(&s->refs[i], &ssrc->next_refs[i])) < 0)

  3806.                 return res;

  3807.         }

  3808.     }

  3809. 

  3810.     s->invisible = ssrc->invisible;

  3811.     s->keyframe = ssrc->keyframe;

  3812.     s->uses_2pass = ssrc->uses_2pass;

  3813.     memcpy(&s->prob_ctx, &ssrc->prob_ctx, sizeof(s->prob_ctx));

  3814.     memcpy(&s->lf_delta, &ssrc->lf_delta, sizeof(s->lf_delta));

  3815.     if (ssrc->segmentation.enabled) {

  3816.         memcpy(&s->segmentation.feat, &ssrc->segmentation.feat,

  3817.                sizeof(s->segmentation.feat));

  3818.     }

  3819. 

  3820.     return 0;

  3821. }

  3822. 

  3823. AVCodec ff_vp9_decoder = {

  3824.     .name                  = "vp9",

  3825.     .long_name             = NULL_IF_CONFIG_SMALL("Google VP9"),

  3826.     .type                  = AVMEDIA_TYPE_VIDEO,

  3827.     .id                    = AV_CODEC_ID_VP9,

  3828.     .priv_data_size        = sizeof(VP9Context),

  3829.     .init                  = vp9_decode_init,

  3830.     .close                 = vp9_decode_free,

  3831.     .decode                = vp9_decode_frame,

  3832.     .capabilities          = CODEC_CAP_DR1 | CODEC_CAP_FRAME_THREADS,

  3833.     .flush                 = vp9_decode_flush,

  3834.     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(vp9_decode_init_thread_copy),

  3835.     .update_thread_context = ONLY_IF_THREADS_ENABLED(vp9_decode_update_thread_context),

  3836. };