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.
60197 Commits
32 Branches
303MB
Tree: c2871568cf
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'c2871568cf'
${ noResults }
FFmpeg/libavcodec/vp9.c

3834 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.             s->use_last_frame_mvs &= s->frames[LAST_FRAME].tf.f->width == w &&

  529.                                      s->frames[LAST_FRAME].tf.f->height == h;

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

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

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

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

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

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

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

  537.             if (s->allowcompinter) {

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

  539.                     s->fixcompref    = 2;

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

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

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

  543.                     s->fixcompref    = 1;

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

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

  546.                 } else {

  547.                     s->fixcompref    = 0;

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

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

  550.                 }

  551.             }

  552.         }

  553.     }

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

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

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

  557. 

  558.     /* loopfilter header data */

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

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

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

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

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

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

  565.     s->filter.sharpness = sharp;

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

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

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

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

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

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

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

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

  574.         }

  575.     } else {

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

  577.     }

  578. 

  579.     /* quantization header data */

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

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

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

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

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

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

  586. 

  587.     /* segmentation header info */

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

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

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

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

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

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

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

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

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

  597.         }

  598. 

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

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

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

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

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

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

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

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

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

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

  609.             }

  610.         }

  611.     } else {

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

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

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

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

  616.     }

  617. 

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

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

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

  621. 

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

  623.             if (s->segmentation.absolute_vals)

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

  625.             else

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

  627.         } else {

  628.             qyac  = s->yac_qi;

  629.         }

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

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

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

  633.         qyac  = av_clip_uintp2(qyac, 8);

  634. 

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

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

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

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

  639. 

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

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

  642.             if (s->segmentation.absolute_vals)

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

  644.             else

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

  646.         } else {

  647.             lflvl  = s->filter.level;

  648.         }

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

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

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

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

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

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

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

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

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

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

  659.         }

  660.     }

  661. 

  662.     /* tiling info */

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

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

  665.         return res;

  666.     }

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

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

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

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

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

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

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

  674.             s->tiling.log2_tile_cols++;

  675.         else

  676.             break;

  677.     }

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

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

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

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

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

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

  684.         if (!s->c_b) {

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

  686.             return AVERROR(ENOMEM);

  687.         }

  688.     }

  689. 

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

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

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

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

  694.                sizeof(vp9_default_coef_probs));

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

  696.                sizeof(vp9_default_coef_probs));

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

  698.                sizeof(vp9_default_coef_probs));

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

  700.                sizeof(vp9_default_coef_probs));

  701.     }

  702. 

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

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

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

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

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

  708.         return AVERROR_INVALIDDATA;

  709.     }

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

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

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

  713.         return AVERROR_INVALIDDATA;

  714.     }

  715. 

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

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

  718.     } else {

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

  720.     }

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

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

  723.     // fw update)?

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

  725. 

  726.     // txfm updates

  727.     if (s->lossless) {

  728.         s->txfmmode = TX_4X4;

  729.     } else {

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

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

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

  733. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  748.         }

  749.     }

  750. 

  751.     // coef updates

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

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

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

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

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

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

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

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

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

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

  762.                                 break;

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

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

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

  766.                                 } else {

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

  768.                                 }

  769.                             }

  770.                             p[3] = 0;

  771.                         }

  772.         } else {

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

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

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

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

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

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

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

  780.                                 break;

  781.                             memcpy(p, r, 3);

  782.                             p[3] = 0;

  783.                         }

  784.         }

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

  786.             break;

  787.     }

  788. 

  789.     // mode updates

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

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

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

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

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

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

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

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

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

  799. 

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

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

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

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

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

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

  806. 

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

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

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

  810. 

  811.         if (s->allowcompinter) {

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

  813.             if (s->comppredmode)

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

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

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

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

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

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

  820.         } else {

  821.             s->comppredmode = PRED_SINGLEREF;

  822.         }

  823. 

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

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

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

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

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

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

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

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

  832.             }

  833.         }

  834. 

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

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

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

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

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

  840.         }

  841. 

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

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

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

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

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

  847. 

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

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

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

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

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

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

  854. 

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

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

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

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

  859. 

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

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

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

  863. 

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

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

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

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

  868. 

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

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

  871. 

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

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

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

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

  876.         }

  877. 

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

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

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

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

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

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

  884. 

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

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

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

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

  889.         }

  890. 

  891.         if (s->highprecisionmvs) {

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

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

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

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

  896. 

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

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

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

  900.             }

  901.         }

  902.     }

  903. 

  904.     return (data2 - data) + size2;

  905. }

  906. 

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

  908.                                       VP9Context *s)

  909. {

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

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

  912. }

  913. 

  914. static void find_ref_mvs(VP9Context *s,

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

  916. {

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

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

  919.                       { -1, -1 }, {  0, -1 }, { -1,  0 }, {  6, -1 }},

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

  921.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, {  2, -1 }},

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

  923.                       { -1, -1 }, { -3,  0 }, {  0, -3 }, { -1,  2 }},

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

  925.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  927.                       { -1,  1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  929.                       {  1, -1 }, { -3,  0 }, {  0, -3 }, { -3, -3 }},

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

  931.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  933.                       {  0, -2 }, { -2,  0 }, { -2, -1 }, { -1, -2 }},

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

  935.                       { -2,  0 }, {  0, -2 }, { -1, -2 }, { -2, -1 }},

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

  937.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

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

  939.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

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

  941.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

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

  943.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  944.     };

  945.     VP9Block *b = s->b;

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

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

  948. #define INVALID_MV 0x80008000U

  949.     uint32_t mem = INVALID_MV;

  950.     int i;

  951. 

  952. #define RETURN_DIRECT_MV(mv) \

  953.     do { \

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

  955.         if (!idx) { \

  956.             AV_WN32A(pmv, m); \

  957.             return; \

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

  959.             mem = m; \

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

  961.             AV_WN32A(pmv, m); \

  962.             return; \

  963.         } \

  964.     } while (0)

  965. 

  966.     if (sb >= 0) {

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

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

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

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

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

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

  973.         }

  974. 

  975. #define RETURN_MV(mv) \

  976.     do { \

  977.         if (sb > 0) { \

  978.             VP56mv tmp; \

  979.             uint32_t m; \

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

  981.             m = AV_RN32A(&tmp); \

  982.             if (!idx) { \

  983.                 AV_WN32A(pmv, m); \

  984.                 return; \

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

  986.                 mem = m; \

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

  988.                 AV_WN32A(pmv, m); \

  989.                 return; \

  990.             } \

  991.         } else { \

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

  993.             if (!idx) { \

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

  995.                 return; \

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

  997.                 mem = m; \

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

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

  1000.                 return; \

  1001.             } \

  1002.         } \

  1003.     } while (0)

  1004. 

  1005.         if (row > 0) {

  1006.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[(row - 1) * s->sb_cols * 8 + col];

  1007.             if (mv->ref[0] == ref) {

  1008.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][0]);

  1009.             } else if (mv->ref[1] == ref) {

  1010.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][1]);

  1011.             }

  1012.         }

  1013.         if (col > s->tiling.tile_col_start) {

  1014.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[row * s->sb_cols * 8 + col - 1];

  1015.             if (mv->ref[0] == ref) {

  1016.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][0]);

  1017.             } else if (mv->ref[1] == ref) {

  1018.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][1]);

  1019.             }

  1020.         }

  1021.         i = 2;

  1022.     } else {

  1023.         i = 0;

  1024.     }

  1025. 

  1026.     // previously coded MVs in this neighbourhood, using same reference frame

  1027.     for (; i < 8; i++) {

  1028.         int c = p[i][0] + col, r = p[i][1] + row;

  1029. 

  1030.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1031.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1032. 

  1033.             if (mv->ref[0] == ref) {

  1034.                 RETURN_MV(mv->mv[0]);

  1035.             } else if (mv->ref[1] == ref) {

  1036.                 RETURN_MV(mv->mv[1]);

  1037.             }

  1038.         }

  1039.     }

  1040. 

  1041.     // MV at this position in previous frame, using same reference frame

  1042.     if (s->use_last_frame_mvs) {

  1043.         struct VP9mvrefPair *mv = &s->frames[LAST_FRAME].mv[row * s->sb_cols * 8 + col];

  1044. 

  1045.         if (!s->last_uses_2pass)

  1046.             ff_thread_await_progress(&s->frames[LAST_FRAME].tf, row >> 3, 0);

  1047.         if (mv->ref[0] == ref) {

  1048.             RETURN_MV(mv->mv[0]);

  1049.         } else if (mv->ref[1] == ref) {

  1050.             RETURN_MV(mv->mv[1]);

  1051.         }

  1052.     }

  1053. 

  1054. #define RETURN_SCALE_MV(mv, scale) \

  1055.     do { \

  1056.         if (scale) { \

  1057.             VP56mv mv_temp = { -mv.x, -mv.y }; \

  1058.             RETURN_MV(mv_temp); \

  1059.         } else { \

  1060.             RETURN_MV(mv); \

  1061.         } \

  1062.     } while (0)

  1063. 

  1064.     // previously coded MVs in this neighbourhood, using different reference frame

  1065.     for (i = 0; i < 8; i++) {

  1066.         int c = p[i][0] + col, r = p[i][1] + row;

  1067. 

  1068.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1069.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1070. 

  1071.             if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1072.                 RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1073.             }

  1074.             if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1075.                 // BUG - libvpx has this condition regardless of whether

  1076.                 // we used the first ref MV and pre-scaling

  1077.                 AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1078.                 RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1079.             }

  1080.         }

  1081.     }

  1082. 

  1083.     // MV at this position in previous frame, using different reference frame

  1084.     if (s->use_last_frame_mvs) {

  1085.         struct VP9mvrefPair *mv = &s->frames[LAST_FRAME].mv[row * s->sb_cols * 8 + col];

  1086. 

  1087.         // no need to await_progress, because we already did that above

  1088.         if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1089.             RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1090.         }

  1091.         if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1092.             // BUG - libvpx has this condition regardless of whether

  1093.             // we used the first ref MV and pre-scaling

  1094.             AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1095.             RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1096.         }

  1097.     }

  1098. 

  1099.     AV_ZERO32(pmv);

  1100. #undef INVALID_MV

  1101. #undef RETURN_MV

  1102. #undef RETURN_SCALE_MV

  1103. }

  1104. 

  1105. static av_always_inline int read_mv_component(VP9Context *s, int idx, int hp)

  1106. {

  1107.     int bit, sign = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].sign);

  1108.     int n, c = vp8_rac_get_tree(&s->c, vp9_mv_class_tree,

  1109.                                 s->prob.p.mv_comp[idx].classes);

  1110. 

  1111.     s->counts.mv_comp[idx].sign[sign]++;

  1112.     s->counts.mv_comp[idx].classes[c]++;

  1113.     if (c) {

  1114.         int m;

  1115. 

  1116.         for (n = 0, m = 0; m < c; m++) {

  1117.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].bits[m]);

  1118.             n |= bit << m;

  1119.             s->counts.mv_comp[idx].bits[m][bit]++;

  1120.         }

  1121.         n <<= 3;

  1122.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree, s->prob.p.mv_comp[idx].fp);

  1123.         n |= bit << 1;

  1124.         s->counts.mv_comp[idx].fp[bit]++;

  1125.         if (hp) {

  1126.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].hp);

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

  1128.             n |= bit;

  1129.         } else {

  1130.             n |= 1;

  1131.             // bug in libvpx - we count for bw entropy purposes even if the

  1132.             // bit wasn't coded

  1133.             s->counts.mv_comp[idx].hp[1]++;

  1134.         }

  1135.         n += 8 << c;

  1136.     } else {

  1137.         n = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0);

  1138.         s->counts.mv_comp[idx].class0[n]++;

  1139.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree,

  1140.                                s->prob.p.mv_comp[idx].class0_fp[n]);

  1141.         s->counts.mv_comp[idx].class0_fp[n][bit]++;

  1142.         n = (n << 3) | (bit << 1);

  1143.         if (hp) {

  1144.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0_hp);

  1145.             s->counts.mv_comp[idx].class0_hp[bit]++;

  1146.             n |= bit;

  1147.         } else {

  1148.             n |= 1;

  1149.             // bug in libvpx - we count for bw entropy purposes even if the

  1150.             // bit wasn't coded

  1151.             s->counts.mv_comp[idx].class0_hp[1]++;

  1152.         }

  1153.     }

  1154. 

  1155.     return sign ? -(n + 1) : (n + 1);

  1156. }

  1157. 

  1158. static void fill_mv(VP9Context *s,

  1159.                     VP56mv *mv, int mode, int sb)

  1160. {

  1161.     VP9Block *b = s->b;

  1162. 

  1163.     if (mode == ZEROMV) {

  1164.         memset(mv, 0, sizeof(*mv) * 2);

  1165.     } else {

  1166.         int hp;

  1167. 

  1168.         // FIXME cache this value and reuse for other subblocks

  1169.         find_ref_mvs(s, &mv[0], b->ref[0], 0, mode == NEARMV,

  1170.                      mode == NEWMV ? -1 : sb);

  1171.         // FIXME maybe move this code into find_ref_mvs()

  1172.         if ((mode == NEWMV || sb == -1) &&

  1173.             !(hp = s->highprecisionmvs && abs(mv[0].x) < 64 && abs(mv[0].y) < 64)) {

  1174.             if (mv[0].y & 1) {

  1175.                 if (mv[0].y < 0)

  1176.                     mv[0].y++;

  1177.                 else

  1178.                     mv[0].y--;

  1179.             }

  1180.             if (mv[0].x & 1) {

  1181.                 if (mv[0].x < 0)

  1182.                     mv[0].x++;

  1183.                 else

  1184.                     mv[0].x--;

  1185.             }

  1186.         }

  1187.         if (mode == NEWMV) {

  1188.             enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1189.                                               s->prob.p.mv_joint);

  1190. 

  1191.             s->counts.mv_joint[j]++;

  1192.             if (j >= MV_JOINT_V)

  1193.                 mv[0].y += read_mv_component(s, 0, hp);

  1194.             if (j & 1)

  1195.                 mv[0].x += read_mv_component(s, 1, hp);

  1196.         }

  1197. 

  1198.         if (b->comp) {

  1199.             // FIXME cache this value and reuse for other subblocks

  1200.             find_ref_mvs(s, &mv[1], b->ref[1], 1, mode == NEARMV,

  1201.                          mode == NEWMV ? -1 : sb);

  1202.             if ((mode == NEWMV || sb == -1) &&

  1203.                 !(hp = s->highprecisionmvs && abs(mv[1].x) < 64 && abs(mv[1].y) < 64)) {

  1204.                 if (mv[1].y & 1) {

  1205.                     if (mv[1].y < 0)

  1206.                         mv[1].y++;

  1207.                     else

  1208.                         mv[1].y--;

  1209.                 }

  1210.                 if (mv[1].x & 1) {

  1211.                     if (mv[1].x < 0)

  1212.                         mv[1].x++;

  1213.                     else

  1214.                         mv[1].x--;

  1215.                 }

  1216.             }

  1217.             if (mode == NEWMV) {

  1218.                 enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1219.                                                   s->prob.p.mv_joint);

  1220. 

  1221.                 s->counts.mv_joint[j]++;

  1222.                 if (j >= MV_JOINT_V)

  1223.                     mv[1].y += read_mv_component(s, 0, hp);

  1224.                 if (j & 1)

  1225.                     mv[1].x += read_mv_component(s, 1, hp);

  1226.             }

  1227.         }

  1228.     }

  1229. }

  1230. 

  1231. static void decode_mode(AVCodecContext *ctx)

  1232. {

  1233.     static const uint8_t left_ctx[N_BS_SIZES] = {

  1234.         0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf

  1235.     };

  1236.     static const uint8_t above_ctx[N_BS_SIZES] = {

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

  1238.     };

  1239.     static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {

  1240.         TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,

  1241.         TX_16X16, TX_8X8, TX_8X8, TX_8X8, TX_4X4, TX_4X4, TX_4X4

  1242.     };

  1243.     VP9Context *s = ctx->priv_data;

  1244.     VP9Block *b = s->b;

  1245.     int row = s->row, col = s->col, row7 = s->row7;

  1246.     enum TxfmMode max_tx = max_tx_for_bl_bp[b->bs];

  1247.     int w4 = FFMIN(s->cols - col, bwh_tab[1][b->bs][0]);

  1248.     int h4 = FFMIN(s->rows - row, bwh_tab[1][b->bs][1]), y;

  1249.     int have_a = row > 0, have_l = col > s->tiling.tile_col_start;

  1250. 

  1251.     if (!s->segmentation.enabled) {

  1252.         b->seg_id = 0;

  1253.     } else if (s->keyframe || s->intraonly) {

  1254.         b->seg_id = s->segmentation.update_map ?

  1255.             vp8_rac_get_tree(&s->c, vp9_segmentation_tree, s->prob.seg) : 0;

  1256.     } else if (!s->segmentation.update_map ||

  1257.                (s->segmentation.temporal &&

  1258.                 vp56_rac_get_prob_branchy(&s->c,

  1259.                     s->prob.segpred[s->above_segpred_ctx[col] +

  1260.                                     s->left_segpred_ctx[row7]]))) {

  1261.         int pred = 8, x;

  1262.         uint8_t *refsegmap = s->frames[LAST_FRAME].segmentation_map;

  1263. 

  1264.         if (!s->last_uses_2pass)

  1265.             ff_thread_await_progress(&s->frames[LAST_FRAME].tf, row >> 3, 0);

  1266.         for (y = 0; y < h4; y++)

  1267.             for (x = 0; x < w4; x++)

  1268.                 pred = FFMIN(pred, refsegmap[(y + row) * 8 * s->sb_cols + x + col]);

  1269.         av_assert1(pred < 8);

  1270.         b->seg_id = pred;

  1271. 

  1272.         memset(&s->above_segpred_ctx[col], 1, w4);

  1273.         memset(&s->left_segpred_ctx[row7], 1, h4);

  1274.     } else {

  1275.         b->seg_id = vp8_rac_get_tree(&s->c, vp9_segmentation_tree,

  1276.                                      s->prob.seg);

  1277. 

  1278.         memset(&s->above_segpred_ctx[col], 0, w4);

  1279.         memset(&s->left_segpred_ctx[row7], 0, h4);

  1280.     }

  1281.     if ((s->segmentation.enabled && s->segmentation.update_map) || s->keyframe) {

  1282.         uint8_t *segmap = s->frames[CUR_FRAME].segmentation_map;

  1283. 

  1284.         for (y = 0; y < h4; y++)

  1285.             memset(&segmap[(y + row) * 8 * s->sb_cols + col], b->seg_id, w4);

  1286.     }

  1287. 

  1288.     b->skip = s->segmentation.enabled &&

  1289.         s->segmentation.feat[b->seg_id].skip_enabled;

  1290.     if (!b->skip) {

  1291.         int c = s->left_skip_ctx[row7] + s->above_skip_ctx[col];

  1292.         b->skip = vp56_rac_get_prob(&s->c, s->prob.p.skip[c]);

  1293.         s->counts.skip[c][b->skip]++;

  1294.     }

  1295. 

  1296.     if (s->keyframe || s->intraonly) {

  1297.         b->intra = 1;

  1298.     } else if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1299.         b->intra = !s->segmentation.feat[b->seg_id].ref_val;

  1300.     } else {

  1301.         int c, bit;

  1302. 

  1303.         if (have_a && have_l) {

  1304.             c = s->above_intra_ctx[col] + s->left_intra_ctx[row7];

  1305.             c += (c == 2);

  1306.         } else {

  1307.             c = have_a ? 2 * s->above_intra_ctx[col] :

  1308.                 have_l ? 2 * s->left_intra_ctx[row7] : 0;

  1309.         }

  1310.         bit = vp56_rac_get_prob(&s->c, s->prob.p.intra[c]);

  1311.         s->counts.intra[c][bit]++;

  1312.         b->intra = !bit;

  1313.     }

  1314. 

  1315.     if ((b->intra || !b->skip) && s->txfmmode == TX_SWITCHABLE) {

  1316.         int c;

  1317.         if (have_a) {

  1318.             if (have_l) {

  1319.                 c = (s->above_skip_ctx[col] ? max_tx :

  1320.                      s->above_txfm_ctx[col]) +

  1321.                     (s->left_skip_ctx[row7] ? max_tx :

  1322.                      s->left_txfm_ctx[row7]) > max_tx;

  1323.             } else {

  1324.                 c = s->above_skip_ctx[col] ? 1 :

  1325.                     (s->above_txfm_ctx[col] * 2 > max_tx);

  1326.             }

  1327.         } else if (have_l) {

  1328.             c = s->left_skip_ctx[row7] ? 1 :

  1329.                 (s->left_txfm_ctx[row7] * 2 > max_tx);

  1330.         } else {

  1331.             c = 1;

  1332.         }

  1333.         switch (max_tx) {

  1334.         case TX_32X32:

  1335.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][0]);

  1336.             if (b->tx) {

  1337.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][1]);

  1338.                 if (b->tx == 2)

  1339.                     b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][2]);

  1340.             }

  1341.             s->counts.tx32p[c][b->tx]++;

  1342.             break;

  1343.         case TX_16X16:

  1344.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][0]);

  1345.             if (b->tx)

  1346.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][1]);

  1347.             s->counts.tx16p[c][b->tx]++;

  1348.             break;

  1349.         case TX_8X8:

  1350.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx8p[c]);

  1351.             s->counts.tx8p[c][b->tx]++;

  1352.             break;

  1353.         case TX_4X4:

  1354.             b->tx = TX_4X4;

  1355.             break;

  1356.         }

  1357.     } else {

  1358.         b->tx = FFMIN(max_tx, s->txfmmode);

  1359.     }

  1360. 

  1361.     if (s->keyframe || s->intraonly) {

  1362.         uint8_t *a = &s->above_mode_ctx[col * 2];

  1363.         uint8_t *l = &s->left_mode_ctx[(row7) << 1];

  1364. 

  1365.         b->comp = 0;

  1366.         if (b->bs > BS_8x8) {

  1367.             // FIXME the memory storage intermediates here aren't really

  1368.             // necessary, they're just there to make the code slightly

  1369.             // simpler for now

  1370.             b->mode[0] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1371.                                     vp9_default_kf_ymode_probs[a[0]][l[0]]);

  1372.             if (b->bs != BS_8x4) {

  1373.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1374.                                  vp9_default_kf_ymode_probs[a[1]][b->mode[0]]);

  1375.                 l[0] = a[1] = b->mode[1];

  1376.             } else {

  1377.                 l[0] = a[1] = b->mode[1] = b->mode[0];

  1378.             }

  1379.             if (b->bs != BS_4x8) {

  1380.                 b->mode[2] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1381.                                         vp9_default_kf_ymode_probs[a[0]][l[1]]);

  1382.                 if (b->bs != BS_8x4) {

  1383.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1384.                                   vp9_default_kf_ymode_probs[a[1]][b->mode[2]]);

  1385.                     l[1] = a[1] = b->mode[3];

  1386.                 } else {

  1387.                     l[1] = a[1] = b->mode[3] = b->mode[2];

  1388.                 }

  1389.             } else {

  1390.                 b->mode[2] = b->mode[0];

  1391.                 l[1] = a[1] = b->mode[3] = b->mode[1];

  1392.             }

  1393.         } else {

  1394.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1395.                                           vp9_default_kf_ymode_probs[*a][*l]);

  1396.             b->mode[3] = b->mode[2] = b->mode[1] = b->mode[0];

  1397.             // FIXME this can probably be optimized

  1398.             memset(a, b->mode[0], bwh_tab[0][b->bs][0]);

  1399.             memset(l, b->mode[0], bwh_tab[0][b->bs][1]);

  1400.         }

  1401.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1402.                                      vp9_default_kf_uvmode_probs[b->mode[3]]);

  1403.     } else if (b->intra) {

  1404.         b->comp = 0;

  1405.         if (b->bs > BS_8x8) {

  1406.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1407.                                           s->prob.p.y_mode[0]);

  1408.             s->counts.y_mode[0][b->mode[0]]++;

  1409.             if (b->bs != BS_8x4) {

  1410.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1411.                                               s->prob.p.y_mode[0]);

  1412.                 s->counts.y_mode[0][b->mode[1]]++;

  1413.             } else {

  1414.                 b->mode[1] = b->mode[0];

  1415.             }

  1416.             if (b->bs != BS_4x8) {

  1417.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1418.                                               s->prob.p.y_mode[0]);

  1419.                 s->counts.y_mode[0][b->mode[2]]++;

  1420.                 if (b->bs != BS_8x4) {

  1421.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1422.                                                   s->prob.p.y_mode[0]);

  1423.                     s->counts.y_mode[0][b->mode[3]]++;

  1424.                 } else {

  1425.                     b->mode[3] = b->mode[2];

  1426.                 }

  1427.             } else {

  1428.                 b->mode[2] = b->mode[0];

  1429.                 b->mode[3] = b->mode[1];

  1430.             }

  1431.         } else {

  1432.             static const uint8_t size_group[10] = {

  1433.                 3, 3, 3, 3, 2, 2, 2, 1, 1, 1

  1434.             };

  1435.             int sz = size_group[b->bs];

  1436. 

  1437.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1438.                                           s->prob.p.y_mode[sz]);

  1439.             b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1440.             s->counts.y_mode[sz][b->mode[3]]++;

  1441.         }

  1442.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1443.                                      s->prob.p.uv_mode[b->mode[3]]);

  1444.         s->counts.uv_mode[b->mode[3]][b->uvmode]++;

  1445.     } else {

  1446.         static const uint8_t inter_mode_ctx_lut[14][14] = {

  1447.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1448.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1449.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

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

  1458.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1459.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },

  1460.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },

  1461.         };

  1462. 

  1463.         if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1464.             av_assert2(s->segmentation.feat[b->seg_id].ref_val != 0);

  1465.             b->comp = 0;

  1466.             b->ref[0] = s->segmentation.feat[b->seg_id].ref_val - 1;

  1467.         } else {

  1468.             // read comp_pred flag

  1469.             if (s->comppredmode != PRED_SWITCHABLE) {

  1470.                 b->comp = s->comppredmode == PRED_COMPREF;

  1471.             } else {

  1472.                 int c;

  1473. 

  1474.                 // FIXME add intra as ref=0xff (or -1) to make these easier?

  1475.                 if (have_a) {

  1476.                     if (have_l) {

  1477.                         if (s->above_comp_ctx[col] && s->left_comp_ctx[row7]) {

  1478.                             c = 4;

  1479.                         } else if (s->above_comp_ctx[col]) {

  1480.                             c = 2 + (s->left_intra_ctx[row7] ||

  1481.                                      s->left_ref_ctx[row7] == s->fixcompref);

  1482.                         } else if (s->left_comp_ctx[row7]) {

  1483.                             c = 2 + (s->above_intra_ctx[col] ||

  1484.                                      s->above_ref_ctx[col] == s->fixcompref);

  1485.                         } else {

  1486.                             c = (!s->above_intra_ctx[col] &&

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

  1488.                             (!s->left_intra_ctx[row7] &&

  1489.                              s->left_ref_ctx[row & 7] == s->fixcompref);

  1490.                         }

  1491.                     } else {

  1492.                         c = s->above_comp_ctx[col] ? 3 :

  1493.                         (!s->above_intra_ctx[col] && s->above_ref_ctx[col] == s->fixcompref);

  1494.                     }

  1495.                 } else if (have_l) {

  1496.                     c = s->left_comp_ctx[row7] ? 3 :

  1497.                     (!s->left_intra_ctx[row7] && s->left_ref_ctx[row7] == s->fixcompref);

  1498.                 } else {

  1499.                     c = 1;

  1500.                 }

  1501.                 b->comp = vp56_rac_get_prob(&s->c, s->prob.p.comp[c]);

  1502.                 s->counts.comp[c][b->comp]++;

  1503.             }

  1504. 

  1505.             // read actual references

  1506.             // FIXME probably cache a few variables here to prevent repetitive

  1507.             // memory accesses below

  1508.             if (b->comp) /* two references */ {

  1509.                 int fix_idx = s->signbias[s->fixcompref], var_idx = !fix_idx, c, bit;

  1510. 

  1511.                 b->ref[fix_idx] = s->fixcompref;

  1512.                 // FIXME can this codeblob be replaced by some sort of LUT?

  1513.                 if (have_a) {

  1514.                     if (have_l) {

  1515.                         if (s->above_intra_ctx[col]) {

  1516.                             if (s->left_intra_ctx[row7]) {

  1517.                                 c = 2;

  1518.                             } else {

  1519.                                 c = 1 + 2 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1520.                             }

  1521.                         } else if (s->left_intra_ctx[row7]) {

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

  1523.                         } else {

  1524.                             int refl = s->left_ref_ctx[row7], refa = s->above_ref_ctx[col];

  1525. 

  1526.                             if (refl == refa && refa == s->varcompref[1]) {

  1527.                                 c = 0;

  1528.                             } else if (!s->left_comp_ctx[row7] && !s->above_comp_ctx[col]) {

  1529.                                 if ((refa == s->fixcompref && refl == s->varcompref[0]) ||

  1530.                                     (refl == s->fixcompref && refa == s->varcompref[0])) {

  1531.                                     c = 4;

  1532.                                 } else {

  1533.                                     c = (refa == refl) ? 3 : 1;

  1534.                                 }

  1535.                             } else if (!s->left_comp_ctx[row7]) {

  1536.                                 if (refa == s->varcompref[1] && refl != s->varcompref[1]) {

  1537.                                     c = 1;

  1538.                                 } else {

  1539.                                     c = (refl == s->varcompref[1] &&

  1540.                                          refa != s->varcompref[1]) ? 2 : 4;

  1541.                                 }

  1542.                             } else if (!s->above_comp_ctx[col]) {

  1543.                                 if (refl == s->varcompref[1] && refa != s->varcompref[1]) {

  1544.                                     c = 1;

  1545.                                 } else {

  1546.                                     c = (refa == s->varcompref[1] &&

  1547.                                          refl != s->varcompref[1]) ? 2 : 4;

  1548.                                 }

  1549.                             } else {

  1550.                                 c = (refl == refa) ? 4 : 2;

  1551.                             }

  1552.                         }

  1553.                     } else {

  1554.                         if (s->above_intra_ctx[col]) {

  1555.                             c = 2;

  1556.                         } else if (s->above_comp_ctx[col]) {

  1557.                             c = 4 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1558.                         } else {

  1559.                             c = 3 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1560.                         }

  1561.                     }

  1562.                 } else if (have_l) {

  1563.                     if (s->left_intra_ctx[row7]) {

  1564.                         c = 2;

  1565.                     } else if (s->left_comp_ctx[row7]) {

  1566.                         c = 4 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1567.                     } else {

  1568.                         c = 3 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1569.                     }

  1570.                 } else {

  1571.                     c = 2;

  1572.                 }

  1573.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.comp_ref[c]);

  1574.                 b->ref[var_idx] = s->varcompref[bit];

  1575.                 s->counts.comp_ref[c][bit]++;

  1576.             } else /* single reference */ {

  1577.                 int bit, c;

  1578. 

  1579.                 if (have_a && !s->above_intra_ctx[col]) {

  1580.                     if (have_l && !s->left_intra_ctx[row7]) {

  1581.                         if (s->left_comp_ctx[row7]) {

  1582.                             if (s->above_comp_ctx[col]) {

  1583.                                 c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7] ||

  1584.                                          !s->above_ref_ctx[col]);

  1585.                             } else {

  1586.                                 c = (3 * !s->above_ref_ctx[col]) +

  1587.                                     (!s->fixcompref || !s->left_ref_ctx[row7]);

  1588.                             }

  1589.                         } else if (s->above_comp_ctx[col]) {

  1590.                             c = (3 * !s->left_ref_ctx[row7]) +

  1591.                                 (!s->fixcompref || !s->above_ref_ctx[col]);

  1592.                         } else {

  1593.                             c = 2 * !s->left_ref_ctx[row7] + 2 * !s->above_ref_ctx[col];

  1594.                         }

  1595.                     } else if (s->above_intra_ctx[col]) {

  1596.                         c = 2;

  1597.                     } else if (s->above_comp_ctx[col]) {

  1598.                         c = 1 + (!s->fixcompref || !s->above_ref_ctx[col]);

  1599.                     } else {

  1600.                         c = 4 * (!s->above_ref_ctx[col]);

  1601.                     }

  1602.                 } else if (have_l && !s->left_intra_ctx[row7]) {

  1603.                     if (s->left_intra_ctx[row7]) {

  1604.                         c = 2;

  1605.                     } else if (s->left_comp_ctx[row7]) {

  1606.                         c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7]);

  1607.                     } else {

  1608.                         c = 4 * (!s->left_ref_ctx[row7]);

  1609.                     }

  1610.                 } else {

  1611.                     c = 2;

  1612.                 }

  1613.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][0]);

  1614.                 s->counts.single_ref[c][0][bit]++;

  1615.                 if (!bit) {

  1616.                     b->ref[0] = 0;

  1617.                 } else {

  1618.                     // FIXME can this codeblob be replaced by some sort of LUT?

  1619.                     if (have_a) {

  1620.                         if (have_l) {

  1621.                             if (s->left_intra_ctx[row7]) {

  1622.                                 if (s->above_intra_ctx[col]) {

  1623.                                     c = 2;

  1624.                                 } else if (s->above_comp_ctx[col]) {

  1625.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1626.                                                  s->above_ref_ctx[col] == 1);

  1627.                                 } else if (!s->above_ref_ctx[col]) {

  1628.                                     c = 3;

  1629.                                 } else {

  1630.                                     c = 4 * (s->above_ref_ctx[col] == 1);

  1631.                                 }

  1632.                             } else if (s->above_intra_ctx[col]) {

  1633.                                 if (s->left_intra_ctx[row7]) {

  1634.                                     c = 2;

  1635.                                 } else if (s->left_comp_ctx[row7]) {

  1636.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1637.                                                  s->left_ref_ctx[row7] == 1);

  1638.                                 } else if (!s->left_ref_ctx[row7]) {

  1639.                                     c = 3;

  1640.                                 } else {

  1641.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1642.                                 }

  1643.                             } else if (s->above_comp_ctx[col]) {

  1644.                                 if (s->left_comp_ctx[row7]) {

  1645.                                     if (s->left_ref_ctx[row7] == s->above_ref_ctx[col]) {

  1646.                                         c = 3 * (s->fixcompref == 1 ||

  1647.                                                  s->left_ref_ctx[row7] == 1);

  1648.                                     } else {

  1649.                                         c = 2;

  1650.                                     }

  1651.                                 } else if (!s->left_ref_ctx[row7]) {

  1652.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1653.                                                  s->above_ref_ctx[col] == 1);

  1654.                                 } else {

  1655.                                     c = 3 * (s->left_ref_ctx[row7] == 1) +

  1656.                                     (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1657.                                 }

  1658.                             } else if (s->left_comp_ctx[row7]) {

  1659.                                 if (!s->above_ref_ctx[col]) {

  1660.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1661.                                                  s->left_ref_ctx[row7] == 1);

  1662.                                 } else {

  1663.                                     c = 3 * (s->above_ref_ctx[col] == 1) +

  1664.                                     (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1665.                                 }

  1666.                             } else if (!s->above_ref_ctx[col]) {

  1667.                                 if (!s->left_ref_ctx[row7]) {

  1668.                                     c = 3;

  1669.                                 } else {

  1670.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1671.                                 }

  1672.                             } else if (!s->left_ref_ctx[row7]) {

  1673.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1674.                             } else {

  1675.                                 c = 2 * (s->left_ref_ctx[row7] == 1) +

  1676.                                 2 * (s->above_ref_ctx[col] == 1);

  1677.                             }

  1678.                         } else {

  1679.                             if (s->above_intra_ctx[col] ||

  1680.                                 (!s->above_comp_ctx[col] && !s->above_ref_ctx[col])) {

  1681.                                 c = 2;

  1682.                             } else if (s->above_comp_ctx[col]) {

  1683.                                 c = 3 * (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1684.                             } else {

  1685.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1686.                             }

  1687.                         }

  1688.                     } else if (have_l) {

  1689.                         if (s->left_intra_ctx[row7] ||

  1690.                             (!s->left_comp_ctx[row7] && !s->left_ref_ctx[row7])) {

  1691.                             c = 2;

  1692.                         } else if (s->left_comp_ctx[row7]) {

  1693.                             c = 3 * (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1694.                         } else {

  1695.                             c = 4 * (s->left_ref_ctx[row7] == 1);

  1696.                         }

  1697.                     } else {

  1698.                         c = 2;

  1699.                     }

  1700.                     bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][1]);

  1701.                     s->counts.single_ref[c][1][bit]++;

  1702.                     b->ref[0] = 1 + bit;

  1703.                 }

  1704.             }

  1705.         }

  1706. 

  1707.         if (b->bs <= BS_8x8) {

  1708.             if (s->segmentation.feat[b->seg_id].skip_enabled) {

  1709.                 b->mode[0] = b->mode[1] = b->mode[2] = b->mode[3] = ZEROMV;

  1710.             } else {

  1711.                 static const uint8_t off[10] = {

  1712.                     3, 0, 0, 1, 0, 0, 0, 0, 0, 0

  1713.                 };

  1714. 

  1715.                 // FIXME this needs to use the LUT tables from find_ref_mvs

  1716.                 // because not all are -1,0/0,-1

  1717.                 int c = inter_mode_ctx_lut[s->above_mode_ctx[col + off[b->bs]]]

  1718.                                           [s->left_mode_ctx[row7 + off[b->bs]]];

  1719. 

  1720.                 b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1721.                                               s->prob.p.mv_mode[c]);

  1722.                 b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1723.                 s->counts.mv_mode[c][b->mode[0] - 10]++;

  1724.             }

  1725.         }

  1726. 

  1727.         if (s->filtermode == FILTER_SWITCHABLE) {

  1728.             int c;

  1729. 

  1730.             if (have_a && s->above_mode_ctx[col] >= NEARESTMV) {

  1731.                 if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1732.                     c = s->above_filter_ctx[col] == s->left_filter_ctx[row7] ?

  1733.                         s->left_filter_ctx[row7] : 3;

  1734.                 } else {

  1735.                     c = s->above_filter_ctx[col];

  1736.                 }

  1737.             } else if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1738.                 c = s->left_filter_ctx[row7];

  1739.             } else {

  1740.                 c = 3;

  1741.             }

  1742. 

  1743.             b->filter = vp8_rac_get_tree(&s->c, vp9_filter_tree,

  1744.                                          s->prob.p.filter[c]);

  1745.             s->counts.filter[c][b->filter]++;

  1746.         } else {

  1747.             b->filter = s->filtermode;

  1748.         }

  1749. 

  1750.         if (b->bs > BS_8x8) {

  1751.             int c = inter_mode_ctx_lut[s->above_mode_ctx[col]][s->left_mode_ctx[row7]];

  1752. 

  1753.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1754.                                           s->prob.p.mv_mode[c]);

  1755.             s->counts.mv_mode[c][b->mode[0] - 10]++;

  1756.             fill_mv(s, b->mv[0], b->mode[0], 0);

  1757. 

  1758.             if (b->bs != BS_8x4) {

  1759.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1760.                                               s->prob.p.mv_mode[c]);

  1761.                 s->counts.mv_mode[c][b->mode[1] - 10]++;

  1762.                 fill_mv(s, b->mv[1], b->mode[1], 1);

  1763.             } else {

  1764.                 b->mode[1] = b->mode[0];

  1765.                 AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1766.                 AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1767.             }

  1768. 

  1769.             if (b->bs != BS_4x8) {

  1770.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1771.                                               s->prob.p.mv_mode[c]);

  1772.                 s->counts.mv_mode[c][b->mode[2] - 10]++;

  1773.                 fill_mv(s, b->mv[2], b->mode[2], 2);

  1774. 

  1775.                 if (b->bs != BS_8x4) {

  1776.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1777.                                                   s->prob.p.mv_mode[c]);

  1778.                     s->counts.mv_mode[c][b->mode[3] - 10]++;

  1779.                     fill_mv(s, b->mv[3], b->mode[3], 3);

  1780.                 } else {

  1781.                     b->mode[3] = b->mode[2];

  1782.                     AV_COPY32(&b->mv[3][0], &b->mv[2][0]);

  1783.                     AV_COPY32(&b->mv[3][1], &b->mv[2][1]);

  1784.                 }

  1785.             } else {

  1786.                 b->mode[2] = b->mode[0];

  1787.                 AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1788.                 AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1789.                 b->mode[3] = b->mode[1];

  1790.                 AV_COPY32(&b->mv[3][0], &b->mv[1][0]);

  1791.                 AV_COPY32(&b->mv[3][1], &b->mv[1][1]);

  1792.             }

  1793.         } else {

  1794.             fill_mv(s, b->mv[0], b->mode[0], -1);

  1795.             AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1796.             AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1797.             AV_COPY32(&b->mv[3][0], &b->mv[0][0]);

  1798.             AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1799.             AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1800.             AV_COPY32(&b->mv[3][1], &b->mv[0][1]);

  1801.         }

  1802.     }

  1803. 

  1804.     // FIXME this can probably be optimized

  1805.     memset(&s->above_skip_ctx[col], b->skip, w4);

  1806.     memset(&s->left_skip_ctx[row7], b->skip, h4);

  1807.     memset(&s->above_txfm_ctx[col], b->tx, w4);

  1808.     memset(&s->left_txfm_ctx[row7], b->tx, h4);

  1809.     memset(&s->above_partition_ctx[col], above_ctx[b->bs], w4);

  1810.     memset(&s->left_partition_ctx[row7], left_ctx[b->bs], h4);

  1811.     if (!s->keyframe && !s->intraonly) {

  1812.         memset(&s->above_intra_ctx[col], b->intra, w4);

  1813.         memset(&s->left_intra_ctx[row7], b->intra, h4);

  1814.         memset(&s->above_comp_ctx[col], b->comp, w4);

  1815.         memset(&s->left_comp_ctx[row7], b->comp, h4);

  1816.         memset(&s->above_mode_ctx[col], b->mode[3], w4);

  1817.         memset(&s->left_mode_ctx[row7], b->mode[3], h4);

  1818.         if (s->filtermode == FILTER_SWITCHABLE && !b->intra ) {

  1819.             memset(&s->above_filter_ctx[col], b->filter, w4);

  1820.             memset(&s->left_filter_ctx[row7], b->filter, h4);

  1821.             b->filter = vp9_filter_lut[b->filter];

  1822.         }

  1823.         if (b->bs > BS_8x8) {

  1824.             int mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  1825. 

  1826.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][0], &b->mv[1][0]);

  1827.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][1], &b->mv[1][1]);

  1828.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][0], mv0);

  1829.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][1], mv1);

  1830.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][0], &b->mv[2][0]);

  1831.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][1], &b->mv[2][1]);

  1832.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][0], mv0);

  1833.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][1], mv1);

  1834.         } else {

  1835.             int n, mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  1836. 

  1837.             for (n = 0; n < w4 * 2; n++) {

  1838.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][0], mv0);

  1839.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][1], mv1);

  1840.             }

  1841.             for (n = 0; n < h4 * 2; n++) {

  1842.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][0], mv0);

  1843.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][1], mv1);

  1844.             }

  1845.         }

  1846. 

  1847.         if (!b->intra) { // FIXME write 0xff or -1 if intra, so we can use this

  1848.                          // as a direct check in above branches

  1849.             int vref = b->ref[b->comp ? s->signbias[s->varcompref[0]] : 0];

  1850. 

  1851.             memset(&s->above_ref_ctx[col], vref, w4);

  1852.             memset(&s->left_ref_ctx[row7], vref, h4);

  1853.         }

  1854.     }

  1855. 

  1856.     // FIXME kinda ugly

  1857.     for (y = 0; y < h4; y++) {

  1858.         int x, o = (row + y) * s->sb_cols * 8 + col;

  1859.         struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[o];

  1860. 

  1861.         if (b->intra) {

  1862.             for (x = 0; x < w4; x++) {

  1863.                 mv[x].ref[0] =

  1864.                 mv[x].ref[1] = -1;

  1865.             }

  1866.         } else if (b->comp) {

  1867.             for (x = 0; x < w4; x++) {

  1868.                 mv[x].ref[0] = b->ref[0];

  1869.                 mv[x].ref[1] = b->ref[1];

  1870.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  1871.                 AV_COPY32(&mv[x].mv[1], &b->mv[3][1]);

  1872.             }

  1873.         } else {

  1874.             for (x = 0; x < w4; x++) {

  1875.                 mv[x].ref[0] = b->ref[0];

  1876.                 mv[x].ref[1] = -1;

  1877.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  1878.             }

  1879.         }

  1880.     }

  1881. }

  1882. 

  1883. // FIXME remove tx argument, and merge cnt/eob arguments?

  1884. static int decode_coeffs_b(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  1885.                            enum TxfmMode tx, unsigned (*cnt)[6][3],

  1886.                            unsigned (*eob)[6][2], uint8_t (*p)[6][11],

  1887.                            int nnz, const int16_t *scan, const int16_t (*nb)[2],

  1888.                            const int16_t *band_counts, const int16_t *qmul)

  1889. {

  1890.     int i = 0, band = 0, band_left = band_counts[band];

  1891.     uint8_t *tp = p[0][nnz];

  1892.     uint8_t cache[1024];

  1893. 

  1894.     do {

  1895.         int val, rc;

  1896. 

  1897.         val = vp56_rac_get_prob_branchy(c, tp[0]); // eob

  1898.         eob[band][nnz][val]++;

  1899.         if (!val)

  1900.             break;

  1901. 

  1902.     skip_eob:

  1903.         if (!vp56_rac_get_prob_branchy(c, tp[1])) { // zero

  1904.             cnt[band][nnz][0]++;

  1905.             if (!--band_left)

  1906.                 band_left = band_counts[++band];

  1907.             cache[scan[i]] = 0;

  1908.             nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  1909.             tp = p[band][nnz];

  1910.             if (++i == n_coeffs)

  1911.                 break; //invalid input; blocks should end with EOB

  1912.             goto skip_eob;

  1913.         }

  1914. 

  1915.         rc = scan[i];

  1916.         if (!vp56_rac_get_prob_branchy(c, tp[2])) { // one

  1917.             cnt[band][nnz][1]++;

  1918.             val = 1;

  1919.             cache[rc] = 1;

  1920.         } else {

  1921.             // fill in p[3-10] (model fill) - only once per frame for each pos

  1922.             if (!tp[3])

  1923.                 memcpy(&tp[3], vp9_model_pareto8[tp[2]], 8);

  1924. 

  1925.             cnt[band][nnz][2]++;

  1926.             if (!vp56_rac_get_prob_branchy(c, tp[3])) { // 2, 3, 4

  1927.                 if (!vp56_rac_get_prob_branchy(c, tp[4])) {

  1928.                     cache[rc] = val = 2;

  1929.                 } else {

  1930.                     val = 3 + vp56_rac_get_prob(c, tp[5]);

  1931.                     cache[rc] = 3;

  1932.                 }

  1933.             } else if (!vp56_rac_get_prob_branchy(c, tp[6])) { // cat1/2

  1934.                 cache[rc] = 4;

  1935.                 if (!vp56_rac_get_prob_branchy(c, tp[7])) {

  1936.                     val = 5 + vp56_rac_get_prob(c, 159);

  1937.                 } else {

  1938.                     val  = 7 + (vp56_rac_get_prob(c, 165) << 1);

  1939.                     val +=      vp56_rac_get_prob(c, 145);

  1940.                 }

  1941.             } else { // cat 3-6

  1942.                 cache[rc] = 5;

  1943.                 if (!vp56_rac_get_prob_branchy(c, tp[8])) {

  1944.                     if (!vp56_rac_get_prob_branchy(c, tp[9])) {

  1945.                         val  = 11 + (vp56_rac_get_prob(c, 173) << 2);

  1946.                         val +=      (vp56_rac_get_prob(c, 148) << 1);

  1947.                         val +=       vp56_rac_get_prob(c, 140);

  1948.                     } else {

  1949.                         val  = 19 + (vp56_rac_get_prob(c, 176) << 3);

  1950.                         val +=      (vp56_rac_get_prob(c, 155) << 2);

  1951.                         val +=      (vp56_rac_get_prob(c, 140) << 1);

  1952.                         val +=       vp56_rac_get_prob(c, 135);

  1953.                     }

  1954.                 } else if (!vp56_rac_get_prob_branchy(c, tp[10])) {

  1955.                     val  = 35 + (vp56_rac_get_prob(c, 180) << 4);

  1956.                     val +=      (vp56_rac_get_prob(c, 157) << 3);

  1957.                     val +=      (vp56_rac_get_prob(c, 141) << 2);

  1958.                     val +=      (vp56_rac_get_prob(c, 134) << 1);

  1959.                     val +=       vp56_rac_get_prob(c, 130);

  1960.                 } else {

  1961.                     val  = 67 + (vp56_rac_get_prob(c, 254) << 13);

  1962.                     val +=      (vp56_rac_get_prob(c, 254) << 12);

  1963.                     val +=      (vp56_rac_get_prob(c, 254) << 11);

  1964.                     val +=      (vp56_rac_get_prob(c, 252) << 10);

  1965.                     val +=      (vp56_rac_get_prob(c, 249) << 9);

  1966.                     val +=      (vp56_rac_get_prob(c, 243) << 8);

  1967.                     val +=      (vp56_rac_get_prob(c, 230) << 7);

  1968.                     val +=      (vp56_rac_get_prob(c, 196) << 6);

  1969.                     val +=      (vp56_rac_get_prob(c, 177) << 5);

  1970.                     val +=      (vp56_rac_get_prob(c, 153) << 4);

  1971.                     val +=      (vp56_rac_get_prob(c, 140) << 3);

  1972.                     val +=      (vp56_rac_get_prob(c, 133) << 2);

  1973.                     val +=      (vp56_rac_get_prob(c, 130) << 1);

  1974.                     val +=       vp56_rac_get_prob(c, 129);

  1975.                 }

  1976.             }

  1977.         }

  1978.         if (!--band_left)

  1979.             band_left = band_counts[++band];

  1980.         if (tx == TX_32X32) // FIXME slow

  1981.             coef[rc] = ((vp8_rac_get(c) ? -val : val) * qmul[!!i]) / 2;

  1982.         else

  1983.             coef[rc] = (vp8_rac_get(c) ? -val : val) * qmul[!!i];

  1984.         nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  1985.         tp = p[band][nnz];

  1986.     } while (++i < n_coeffs);

  1987. 

  1988.     return i;

  1989. }

  1990. 

  1991. static void decode_coeffs(AVCodecContext *ctx)

  1992. {

  1993.     VP9Context *s = ctx->priv_data;

  1994.     VP9Block *b = s->b;

  1995.     int row = s->row, col = s->col;

  1996.     uint8_t (*p)[6][11] = s->prob.coef[b->tx][0 /* y */][!b->intra];

  1997.     unsigned (*c)[6][3] = s->counts.coef[b->tx][0 /* y */][!b->intra];

  1998.     unsigned (*e)[6][2] = s->counts.eob[b->tx][0 /* y */][!b->intra];

  1999.     int w4 = bwh_tab[1][b->bs][0] << 1, h4 = bwh_tab[1][b->bs][1] << 1;

  2000.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2001.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2002.     int n, pl, x, y, step1d = 1 << b->tx, step = 1 << (b->tx * 2);

  2003.     int uvstep1d = 1 << b->uvtx, uvstep = 1 << (b->uvtx * 2), res;

  2004.     int16_t (*qmul)[2] = s->segmentation.feat[b->seg_id].qmul;

  2005.     int tx = 4 * s->lossless + b->tx;

  2006.     const int16_t * const *yscans = vp9_scans[tx];

  2007.     const int16_t (* const *ynbs)[2] = vp9_scans_nb[tx];

  2008.     const int16_t *uvscan = vp9_scans[b->uvtx][DCT_DCT];

  2009.     const int16_t (*uvnb)[2] = vp9_scans_nb[b->uvtx][DCT_DCT];

  2010.     uint8_t *a = &s->above_y_nnz_ctx[col * 2];

  2011.     uint8_t *l = &s->left_y_nnz_ctx[(row & 7) << 1];

  2012.     static const int16_t band_counts[4][8] = {

  2013.         { 1, 2, 3, 4,  3,   16 - 13 },

  2014.         { 1, 2, 3, 4, 11,   64 - 21 },

  2015.         { 1, 2, 3, 4, 11,  256 - 21 },

  2016.         { 1, 2, 3, 4, 11, 1024 - 21 },

  2017.     };

  2018.     const int16_t *y_band_counts = band_counts[b->tx];

  2019.     const int16_t *uv_band_counts = band_counts[b->uvtx];

  2020. 

  2021.     /* y tokens */

  2022.     if (b->tx > TX_4X4) { // FIXME slow

  2023.         for (y = 0; y < end_y; y += step1d)

  2024.             for (x = 1; x < step1d; x++)

  2025.                 l[y] |= l[y + x];

  2026.         for (x = 0; x < end_x; x += step1d)

  2027.             for (y = 1; y < step1d; y++)

  2028.                 a[x] |= a[x + y];

  2029.     }

  2030.     for (n = 0, y = 0; y < end_y; y += step1d) {

  2031.         for (x = 0; x < end_x; x += step1d, n += step) {

  2032.             enum TxfmType txtp = vp9_intra_txfm_type[b->mode[b->tx == TX_4X4 &&

  2033.                                                              b->bs > BS_8x8 ?

  2034.                                                              n : 0]];

  2035.             int nnz = a[x] + l[y];

  2036.             res = decode_coeffs_b(&s->c, s->block + 16 * n, 16 * step,

  2037.                                   b->tx, c, e, p, nnz, yscans[txtp],

  2038.                                   ynbs[txtp], y_band_counts, qmul[0]);

  2039.             a[x] = l[y] = !!res;

  2040.             if (b->tx > TX_8X8) {

  2041.                 AV_WN16A(&s->eob[n], res);

  2042.             } else {

  2043.                 s->eob[n] = res;

  2044.             }

  2045.         }

  2046.     }

  2047.     if (b->tx > TX_4X4) { // FIXME slow

  2048.         for (y = 0; y < end_y; y += step1d)

  2049.             memset(&l[y + 1], l[y], FFMIN(end_y - y - 1, step1d - 1));

  2050.         for (x = 0; x < end_x; x += step1d)

  2051.             memset(&a[x + 1], a[x], FFMIN(end_x - x - 1, step1d - 1));

  2052.     }

  2053. 

  2054.     p = s->prob.coef[b->uvtx][1 /* uv */][!b->intra];

  2055.     c = s->counts.coef[b->uvtx][1 /* uv */][!b->intra];

  2056.     e = s->counts.eob[b->uvtx][1 /* uv */][!b->intra];

  2057.     w4 >>= 1;

  2058.     h4 >>= 1;

  2059.     end_x >>= 1;

  2060.     end_y >>= 1;

  2061.     for (pl = 0; pl < 2; pl++) {

  2062.         a = &s->above_uv_nnz_ctx[pl][col];

  2063.         l = &s->left_uv_nnz_ctx[pl][row & 7];

  2064.         if (b->uvtx > TX_4X4) { // FIXME slow

  2065.             for (y = 0; y < end_y; y += uvstep1d)

  2066.                 for (x = 1; x < uvstep1d; x++)

  2067.                     l[y] |= l[y + x];

  2068.             for (x = 0; x < end_x; x += uvstep1d)

  2069.                 for (y = 1; y < uvstep1d; y++)

  2070.                     a[x] |= a[x + y];

  2071.         }

  2072.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2073.             for (x = 0; x < end_x; x += uvstep1d, n += uvstep) {

  2074.                 int nnz = a[x] + l[y];

  2075.                 res = decode_coeffs_b(&s->c, s->uvblock[pl] + 16 * n,

  2076.                                       16 * uvstep, b->uvtx, c, e, p, nnz,

  2077.                                       uvscan, uvnb, uv_band_counts, qmul[1]);

  2078.                 a[x] = l[y] = !!res;

  2079.                 if (b->uvtx > TX_8X8) {

  2080.                     AV_WN16A(&s->uveob[pl][n], res);

  2081.                 } else {

  2082.                     s->uveob[pl][n] = res;

  2083.                 }

  2084.             }

  2085.         }

  2086.         if (b->uvtx > TX_4X4) { // FIXME slow

  2087.             for (y = 0; y < end_y; y += uvstep1d)

  2088.                 memset(&l[y + 1], l[y], FFMIN(end_y - y - 1, uvstep1d - 1));

  2089.             for (x = 0; x < end_x; x += uvstep1d)

  2090.                 memset(&a[x + 1], a[x], FFMIN(end_x - x - 1, uvstep1d - 1));

  2091.         }

  2092.     }

  2093. }

  2094. 

  2095. static av_always_inline int check_intra_mode(VP9Context *s, int mode, uint8_t **a,

  2096.                                              uint8_t *dst_edge, ptrdiff_t stride_edge,

  2097.                                              uint8_t *dst_inner, ptrdiff_t stride_inner,

  2098.                                              uint8_t *l, int col, int x, int w,

  2099.                                              int row, int y, enum TxfmMode tx,

  2100.                                              int p)

  2101. {

  2102.     int have_top = row > 0 || y > 0;

  2103.     int have_left = col > s->tiling.tile_col_start || x > 0;

  2104.     int have_right = x < w - 1;

  2105.     static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {

  2106.         [VERT_PRED]            = { { DC_127_PRED,          VERT_PRED },

  2107.                                    { DC_127_PRED,          VERT_PRED } },

  2108.         [HOR_PRED]             = { { DC_129_PRED,          DC_129_PRED },

  2109.                                    { HOR_PRED,             HOR_PRED } },

  2110.         [DC_PRED]              = { { DC_128_PRED,          TOP_DC_PRED },

  2111.                                    { LEFT_DC_PRED,         DC_PRED } },

  2112.         [DIAG_DOWN_LEFT_PRED]  = { { DC_127_PRED,          DIAG_DOWN_LEFT_PRED },

  2113.                                    { DC_127_PRED,          DIAG_DOWN_LEFT_PRED } },

  2114.         [DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },

  2115.                                    { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },

  2116.         [VERT_RIGHT_PRED]      = { { VERT_RIGHT_PRED,      VERT_RIGHT_PRED },

  2117.                                    { VERT_RIGHT_PRED,      VERT_RIGHT_PRED } },

  2118.         [HOR_DOWN_PRED]        = { { HOR_DOWN_PRED,        HOR_DOWN_PRED },

  2119.                                    { HOR_DOWN_PRED,        HOR_DOWN_PRED } },

  2120.         [VERT_LEFT_PRED]       = { { DC_127_PRED,          VERT_LEFT_PRED },

  2121.                                    { DC_127_PRED,          VERT_LEFT_PRED } },

  2122.         [HOR_UP_PRED]          = { { DC_129_PRED,          DC_129_PRED },

  2123.                                    { HOR_UP_PRED,          HOR_UP_PRED } },

  2124.         [TM_VP8_PRED]          = { { DC_129_PRED,          VERT_PRED },

  2125.                                    { HOR_PRED,             TM_VP8_PRED } },

  2126.     };

  2127.     static const struct {

  2128.         uint8_t needs_left:1;

  2129.         uint8_t needs_top:1;

  2130.         uint8_t needs_topleft:1;

  2131.         uint8_t needs_topright:1;

  2132.     } edges[N_INTRA_PRED_MODES] = {

  2133.         [VERT_PRED]            = { .needs_top  = 1 },

  2134.         [HOR_PRED]             = { .needs_left = 1 },

  2135.         [DC_PRED]              = { .needs_top  = 1, .needs_left = 1 },

  2136.         [DIAG_DOWN_LEFT_PRED]  = { .needs_top  = 1, .needs_topright = 1 },

  2137.         [DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2138.         [VERT_RIGHT_PRED]      = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2139.         [HOR_DOWN_PRED]        = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2140.         [VERT_LEFT_PRED]       = { .needs_top  = 1, .needs_topright = 1 },

  2141.         [HOR_UP_PRED]          = { .needs_left = 1 },

  2142.         [TM_VP8_PRED]          = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2143.         [LEFT_DC_PRED]         = { .needs_left = 1 },

  2144.         [TOP_DC_PRED]          = { .needs_top  = 1 },

  2145.         [DC_128_PRED]          = { 0 },

  2146.         [DC_127_PRED]          = { 0 },

  2147.         [DC_129_PRED]          = { 0 }

  2148.     };

  2149. 

  2150.     av_assert2(mode >= 0 && mode < 10);

  2151.     mode = mode_conv[mode][have_left][have_top];

  2152.     if (edges[mode].needs_top) {

  2153.         uint8_t *top, *topleft;

  2154.         int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !p) - x) * 4;

  2155.         int n_px_need_tr = 0;

  2156. 

  2157.         if (tx == TX_4X4 && edges[mode].needs_topright && have_right)

  2158.             n_px_need_tr = 4;

  2159. 

  2160.         // if top of sb64-row, use s->intra_pred_data[] instead of

  2161.         // dst[-stride] for intra prediction (it contains pre- instead of

  2162.         // post-loopfilter data)

  2163.         if (have_top) {

  2164.             top = !(row & 7) && !y ?

  2165.                 s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  2166.                 y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];

  2167.             if (have_left)

  2168.                 topleft = !(row & 7) && !y ?

  2169.                     s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  2170.                     y == 0 || x == 0 ? &dst_edge[-stride_edge] :

  2171.                     &dst_inner[-stride_inner];

  2172.         }

  2173. 

  2174.         if (have_top &&

  2175.             (!edges[mode].needs_topleft || (have_left && top == topleft)) &&

  2176.             (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&

  2177.             n_px_need + n_px_need_tr <= n_px_have) {

  2178.             *a = top;

  2179.         } else {

  2180.             if (have_top) {

  2181.                 if (n_px_need <= n_px_have) {

  2182.                     memcpy(*a, top, n_px_need);

  2183.                 } else {

  2184.                     memcpy(*a, top, n_px_have);

  2185.                     memset(&(*a)[n_px_have], (*a)[n_px_have - 1],

  2186.                            n_px_need - n_px_have);

  2187.                 }

  2188.             } else {

  2189.                 memset(*a, 127, n_px_need);

  2190.             }

  2191.             if (edges[mode].needs_topleft) {

  2192.                 if (have_left && have_top) {

  2193.                     (*a)[-1] = topleft[-1];

  2194.                 } else {

  2195.                     (*a)[-1] = have_top ? 129 : 127;

  2196.                 }

  2197.             }

  2198.             if (tx == TX_4X4 && edges[mode].needs_topright) {

  2199.                 if (have_top && have_right &&

  2200.                     n_px_need + n_px_need_tr <= n_px_have) {

  2201.                     memcpy(&(*a)[4], &top[4], 4);

  2202.                 } else {

  2203.                     memset(&(*a)[4], (*a)[3], 4);

  2204.                 }

  2205.             }

  2206.         }

  2207.     }

  2208.     if (edges[mode].needs_left) {

  2209.         if (have_left) {

  2210.             int n_px_need = 4 << tx, i, n_px_have = (((s->rows - row) << !p) - y) * 4;

  2211.             uint8_t *dst = x == 0 ? dst_edge : dst_inner;

  2212.             ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;

  2213. 

  2214.             if (n_px_need <= n_px_have) {

  2215.                 for (i = 0; i < n_px_need; i++)

  2216.                     l[i] = dst[i * stride - 1];

  2217.             } else {

  2218.                 for (i = 0; i < n_px_have; i++)

  2219.                     l[i] = dst[i * stride - 1];

  2220.                 memset(&l[i], l[i - 1], n_px_need - n_px_have);

  2221.             }

  2222.         } else {

  2223.             memset(l, 129, 4 << tx);

  2224.         }

  2225.     }

  2226. 

  2227.     return mode;

  2228. }

  2229. 

  2230. static void intra_recon(AVCodecContext *ctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  2231. {

  2232.     VP9Context *s = ctx->priv_data;

  2233.     VP9Block *b = s->b;

  2234.     int row = s->row, col = s->col;

  2235.     int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2236.     int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2237.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2238.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2239.     int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2240.     int uvstep1d = 1 << b->uvtx, p;

  2241.     uint8_t *dst = s->dst[0], *dst_r = s->frames[CUR_FRAME].tf.f->data[0] + y_off;

  2242. 

  2243.     for (n = 0, y = 0; y < end_y; y += step1d) {

  2244.         uint8_t *ptr = dst, *ptr_r = dst_r;

  2245.         for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d,

  2246.                                ptr_r += 4 * step1d, n += step) {

  2247.             int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?

  2248.                                y * 2 + x : 0];

  2249.             LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);

  2250.             uint8_t *a = &a_buf[16], l[32];

  2251.             enum TxfmType txtp = vp9_intra_txfm_type[mode];

  2252.             int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2253. 

  2254.             mode = check_intra_mode(s, mode, &a, ptr_r,

  2255.                                     s->frames[CUR_FRAME].tf.f->linesize[0],

  2256.                                     ptr, s->y_stride, l,

  2257.                                     col, x, w4, row, y, b->tx, 0);

  2258.             s->dsp.intra_pred[b->tx][mode](ptr, s->y_stride, l, a);

  2259.             if (eob)

  2260.                 s->dsp.itxfm_add[tx][txtp](ptr, s->y_stride,

  2261.                                            s->block + 16 * n, eob);

  2262.         }

  2263.         dst_r += 4 * step1d * s->frames[CUR_FRAME].tf.f->linesize[0];

  2264.         dst   += 4 * step1d * s->y_stride;

  2265.     }

  2266. 

  2267.     // U/V

  2268.     h4 >>= 1;

  2269.     w4 >>= 1;

  2270.     end_x >>= 1;

  2271.     end_y >>= 1;

  2272.     step = 1 << (b->uvtx * 2);

  2273.     for (p = 0; p < 2; p++) {

  2274.         dst   = s->dst[1 + p];

  2275.         dst_r = s->frames[CUR_FRAME].tf.f->data[1 + p] + uv_off;

  2276.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2277.             uint8_t *ptr = dst, *ptr_r = dst_r;

  2278.             for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d,

  2279.                                    ptr_r += 4 * uvstep1d, n += step) {

  2280.                 int mode = b->uvmode;

  2281.                 LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);

  2282.                 uint8_t *a = &a_buf[16], l[32];

  2283.                 int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2284. 

  2285.                 mode = check_intra_mode(s, mode, &a, ptr_r,

  2286.                                         s->frames[CUR_FRAME].tf.f->linesize[1],

  2287.                                         ptr, s->uv_stride, l,

  2288.                                         col, x, w4, row, y, b->uvtx, p + 1);

  2289.                 s->dsp.intra_pred[b->uvtx][mode](ptr, s->uv_stride, l, a);

  2290.                 if (eob)

  2291.                     s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2292.                                                     s->uvblock[p] + 16 * n, eob);

  2293.             }

  2294.             dst_r += 4 * uvstep1d * s->frames[CUR_FRAME].tf.f->linesize[1];

  2295.             dst   += 4 * uvstep1d * s->uv_stride;

  2296.         }

  2297.     }

  2298. }

  2299. 

  2300. static av_always_inline void mc_luma_dir(VP9Context *s, vp9_mc_func (*mc)[2],

  2301.                                          uint8_t *dst, ptrdiff_t dst_stride,

  2302.                                          const uint8_t *ref, ptrdiff_t ref_stride,

  2303.                                          ThreadFrame *ref_frame,

  2304.                                          ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2305.                                          int bw, int bh, int w, int h)

  2306. {

  2307.     int mx = mv->x, my = mv->y, th;

  2308. 

  2309.     y += my >> 3;

  2310.     x += mx >> 3;

  2311.     ref += y * ref_stride + x;

  2312.     mx &= 7;

  2313.     my &= 7;

  2314.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2315.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2316.     // the longest loopfilter of the next sbrow

  2317.     th = (y + bh + 4 * !!my + 7) >> 6;

  2318.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2319.     if (x < !!mx * 3 || y < !!my * 3 ||

  2320.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2321.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2322.                                  ref - !!my * 3 * ref_stride - !!mx * 3,

  2323.                                  80, ref_stride,

  2324.                                  bw + !!mx * 7, bh + !!my * 7,

  2325.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2326.         ref = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2327.         ref_stride = 80;

  2328.     }

  2329.     mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);

  2330. }

  2331. 

  2332. static av_always_inline void mc_chroma_dir(VP9Context *s, vp9_mc_func (*mc)[2],

  2333.                                            uint8_t *dst_u, uint8_t *dst_v,

  2334.                                            ptrdiff_t dst_stride,

  2335.                                            const uint8_t *ref_u, ptrdiff_t src_stride_u,

  2336.                                            const uint8_t *ref_v, ptrdiff_t src_stride_v,

  2337.                                            ThreadFrame *ref_frame,

  2338.                                            ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2339.                                            int bw, int bh, int w, int h)

  2340. {

  2341.     int mx = mv->x, my = mv->y, th;

  2342. 

  2343.     y += my >> 4;

  2344.     x += mx >> 4;

  2345.     ref_u += y * src_stride_u + x;

  2346.     ref_v += y * src_stride_v + x;

  2347.     mx &= 15;

  2348.     my &= 15;

  2349.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2350.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2351.     // the longest loopfilter of the next sbrow

  2352.     th = (y + bh + 4 * !!my + 7) >> 5;

  2353.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2354.     if (x < !!mx * 3 || y < !!my * 3 ||

  2355.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2356.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2357.                                  ref_u - !!my * 3 * src_stride_u - !!mx * 3,

  2358.                                  80, src_stride_u,

  2359.                                  bw + !!mx * 7, bh + !!my * 7,

  2360.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2361.         ref_u = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2362.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, 80, bh, mx, my);

  2363. 

  2364.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2365.                                  ref_v - !!my * 3 * src_stride_v - !!mx * 3,

  2366.                                  80, src_stride_v,

  2367.                                  bw + !!mx * 7, bh + !!my * 7,

  2368.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2369.         ref_v = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2370.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, 80, bh, mx, my);

  2371.     } else {

  2372.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);

  2373.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);

  2374.     }

  2375. }

  2376. 

  2377. static void inter_recon(AVCodecContext *ctx)

  2378. {

  2379.     static const uint8_t bwlog_tab[2][N_BS_SIZES] = {

  2380.         { 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4 },

  2381.         { 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 4, 4 },

  2382.     };

  2383.     VP9Context *s = ctx->priv_data;

  2384.     VP9Block *b = s->b;

  2385.     int row = s->row, col = s->col;

  2386.     ThreadFrame *tref1 = &s->refs[s->refidx[b->ref[0]]], *tref2;

  2387.     AVFrame *ref1 = tref1->f, *ref2;

  2388.     int w1 = ref1->width, h1 = ref1->height, w2, h2;

  2389.     ptrdiff_t ls_y = s->y_stride, ls_uv = s->uv_stride;

  2390. 

  2391.     if (b->comp) {

  2392.         tref2 = &s->refs[s->refidx[b->ref[1]]];

  2393.         ref2 = tref2->f;

  2394.         w2 = ref2->width;

  2395.         h2 = ref2->height;

  2396.     }

  2397. 

  2398.     // y inter pred

  2399.     if (b->bs > BS_8x8) {

  2400.         if (b->bs == BS_8x4) {

  2401.             mc_luma_dir(s, s->dsp.mc[3][b->filter][0], s->dst[0], ls_y,

  2402.                         ref1->data[0], ref1->linesize[0], tref1,

  2403.                         row << 3, col << 3, &b->mv[0][0], 8, 4, w1, h1);

  2404.             mc_luma_dir(s, s->dsp.mc[3][b->filter][0],

  2405.                         s->dst[0] + 4 * ls_y, ls_y,

  2406.                         ref1->data[0], ref1->linesize[0], tref1,

  2407.                         (row << 3) + 4, col << 3, &b->mv[2][0], 8, 4, w1, h1);

  2408. 

  2409.             if (b->comp) {

  2410.                 mc_luma_dir(s, s->dsp.mc[3][b->filter][1], s->dst[0], ls_y,

  2411.                             ref2->data[0], ref2->linesize[0], tref2,

  2412.                             row << 3, col << 3, &b->mv[0][1], 8, 4, w2, h2);

  2413.                 mc_luma_dir(s, s->dsp.mc[3][b->filter][1],

  2414.                             s->dst[0] + 4 * ls_y, ls_y,

  2415.                             ref2->data[0], ref2->linesize[0], tref2,

  2416.                             (row << 3) + 4, col << 3, &b->mv[2][1], 8, 4, w2, h2);

  2417.             }

  2418.         } else if (b->bs == BS_4x8) {

  2419.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0], ls_y,

  2420.                         ref1->data[0], ref1->linesize[0], tref1,

  2421.                         row << 3, col << 3, &b->mv[0][0], 4, 8, w1, h1);

  2422.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0] + 4, ls_y,

  2423.                         ref1->data[0], ref1->linesize[0], tref1,

  2424.                         row << 3, (col << 3) + 4, &b->mv[1][0], 4, 8, w1, h1);

  2425. 

  2426.             if (b->comp) {

  2427.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0], ls_y,

  2428.                             ref2->data[0], ref2->linesize[0], tref2,

  2429.                             row << 3, col << 3, &b->mv[0][1], 4, 8, w2, h2);

  2430.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0] + 4, ls_y,

  2431.                             ref2->data[0], ref2->linesize[0], tref2,

  2432.                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 8, w2, h2);

  2433.             }

  2434.         } else {

  2435.             av_assert2(b->bs == BS_4x4);

  2436. 

  2437.             // FIXME if two horizontally adjacent blocks have the same MV,

  2438.             // do a w8 instead of a w4 call

  2439.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0], ls_y,

  2440.                         ref1->data[0], ref1->linesize[0], tref1,

  2441.                         row << 3, col << 3, &b->mv[0][0], 4, 4, w1, h1);

  2442.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0] + 4, ls_y,

  2443.                         ref1->data[0], ref1->linesize[0], tref1,

  2444.                         row << 3, (col << 3) + 4, &b->mv[1][0], 4, 4, w1, h1);

  2445.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0],

  2446.                         s->dst[0] + 4 * ls_y, ls_y,

  2447.                         ref1->data[0], ref1->linesize[0], tref1,

  2448.                         (row << 3) + 4, col << 3, &b->mv[2][0], 4, 4, w1, h1);

  2449.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0],

  2450.                         s->dst[0] + 4 * ls_y + 4, ls_y,

  2451.                         ref1->data[0], ref1->linesize[0], tref1,

  2452.                         (row << 3) + 4, (col << 3) + 4, &b->mv[3][0], 4, 4, w1, h1);

  2453. 

  2454.             if (b->comp) {

  2455.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0], ls_y,

  2456.                             ref2->data[0], ref2->linesize[0], tref2,

  2457.                             row << 3, col << 3, &b->mv[0][1], 4, 4, w2, h2);

  2458.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0] + 4, ls_y,

  2459.                             ref2->data[0], ref2->linesize[0], tref2,

  2460.                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 4, w2, h2);

  2461.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1],

  2462.                             s->dst[0] + 4 * ls_y, ls_y,

  2463.                             ref2->data[0], ref2->linesize[0], tref2,

  2464.                             (row << 3) + 4, col << 3, &b->mv[2][1], 4, 4, w2, h2);

  2465.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1],

  2466.                             s->dst[0] + 4 * ls_y + 4, ls_y,

  2467.                             ref2->data[0], ref2->linesize[0], tref2,

  2468.                             (row << 3) + 4, (col << 3) + 4, &b->mv[3][1], 4, 4, w2, h2);

  2469.             }

  2470.         }

  2471.     } else {

  2472.         int bwl = bwlog_tab[0][b->bs];

  2473.         int bw = bwh_tab[0][b->bs][0] * 4, bh = bwh_tab[0][b->bs][1] * 4;

  2474. 

  2475.         mc_luma_dir(s, s->dsp.mc[bwl][b->filter][0], s->dst[0], ls_y,

  2476.                     ref1->data[0], ref1->linesize[0], tref1,

  2477.                     row << 3, col << 3, &b->mv[0][0],bw, bh, w1, h1);

  2478. 

  2479.         if (b->comp)

  2480.             mc_luma_dir(s, s->dsp.mc[bwl][b->filter][1], s->dst[0], ls_y,

  2481.                         ref2->data[0], ref2->linesize[0], tref2,

  2482.                         row << 3, col << 3, &b->mv[0][1], bw, bh, w2, h2);

  2483.     }

  2484. 

  2485.     // uv inter pred

  2486.     {

  2487.         int bwl = bwlog_tab[1][b->bs];

  2488.         int bw = bwh_tab[1][b->bs][0] * 4, bh = bwh_tab[1][b->bs][1] * 4;

  2489.         VP56mv mvuv;

  2490. 

  2491.         w1 = (w1 + 1) >> 1;

  2492.         h1 = (h1 + 1) >> 1;

  2493.         if (b->comp) {

  2494.             w2 = (w2 + 1) >> 1;

  2495.             h2 = (h2 + 1) >> 1;

  2496.         }

  2497.         if (b->bs > BS_8x8) {

  2498.             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);

  2499.             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);

  2500.         } else {

  2501.             mvuv = b->mv[0][0];

  2502.         }

  2503. 

  2504.         mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][0],

  2505.                       s->dst[1], s->dst[2], ls_uv,

  2506.                       ref1->data[1], ref1->linesize[1],

  2507.                       ref1->data[2], ref1->linesize[2], tref1,

  2508.                       row << 2, col << 2, &mvuv, bw, bh, w1, h1);

  2509. 

  2510.         if (b->comp) {

  2511.             if (b->bs > BS_8x8) {

  2512.                 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);

  2513.                 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);

  2514.             } else {

  2515.                 mvuv = b->mv[0][1];

  2516.             }

  2517.             mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][1],

  2518.                           s->dst[1], s->dst[2], ls_uv,

  2519.                           ref2->data[1], ref2->linesize[1],

  2520.                           ref2->data[2], ref2->linesize[2], tref2,

  2521.                           row << 2, col << 2, &mvuv, bw, bh, w2, h2);

  2522.         }

  2523.     }

  2524. 

  2525.     if (!b->skip) {

  2526.         /* mostly copied intra_reconn() */

  2527. 

  2528.         int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2529.         int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2530.         int end_x = FFMIN(2 * (s->cols - col), w4);

  2531.         int end_y = FFMIN(2 * (s->rows - row), h4);

  2532.         int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2533.         int uvstep1d = 1 << b->uvtx, p;

  2534.         uint8_t *dst = s->dst[0];

  2535. 

  2536.         // y itxfm add

  2537.         for (n = 0, y = 0; y < end_y; y += step1d) {

  2538.             uint8_t *ptr = dst;

  2539.             for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d, n += step) {

  2540.                 int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2541. 

  2542.                 if (eob)

  2543.                     s->dsp.itxfm_add[tx][DCT_DCT](ptr, s->y_stride,

  2544.                                                   s->block + 16 * n, eob);

  2545.             }

  2546.             dst += 4 * s->y_stride * step1d;

  2547.         }

  2548. 

  2549.         // uv itxfm add

  2550.         h4 >>= 1;

  2551.         w4 >>= 1;

  2552.         end_x >>= 1;

  2553.         end_y >>= 1;

  2554.         step = 1 << (b->uvtx * 2);

  2555.         for (p = 0; p < 2; p++) {

  2556.             dst = s->dst[p + 1];

  2557.             for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2558.                 uint8_t *ptr = dst;

  2559.                 for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d, n += step) {

  2560.                     int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2561. 

  2562.                     if (eob)

  2563.                         s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2564.                                                         s->uvblock[p] + 16 * n, eob);

  2565.                 }

  2566.                 dst += 4 * uvstep1d * s->uv_stride;

  2567.             }

  2568.         }

  2569.     }

  2570. }

  2571. 

  2572. static av_always_inline void mask_edges(struct VP9Filter *lflvl, int is_uv,

  2573.                                         int row_and_7, int col_and_7,

  2574.                                         int w, int h, int col_end, int row_end,

  2575.                                         enum TxfmMode tx, int skip_inter)

  2576. {

  2577.     // FIXME I'm pretty sure all loops can be replaced by a single LUT if

  2578.     // we make VP9Filter.mask uint64_t (i.e. row/col all single variable)

  2579.     // and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then

  2580.     // use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)

  2581. 

  2582.     // the intended behaviour of the vp9 loopfilter is to work on 8-pixel

  2583.     // edges. This means that for UV, we work on two subsampled blocks at

  2584.     // a time, and we only use the topleft block's mode information to set

  2585.     // things like block strength. Thus, for any block size smaller than

  2586.     // 16x16, ignore the odd portion of the block.

  2587.     if (tx == TX_4X4 && is_uv) {

  2588.         if (h == 1) {

  2589.             if (row_and_7 & 1)

  2590.                 return;

  2591.             if (!row_end)

  2592.                 h += 1;

  2593.         }

  2594.         if (w == 1) {

  2595.             if (col_and_7 & 1)

  2596.                 return;

  2597.             if (!col_end)

  2598.                 w += 1;

  2599.         }

  2600.     }

  2601. 

  2602.     if (tx == TX_4X4 && !skip_inter) {

  2603.         int t = 1 << col_and_7, m_col = (t << w) - t, y;

  2604.         int m_col_odd = (t << (w - 1)) - t;

  2605. 

  2606.         // on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide

  2607.         if (is_uv) {

  2608.             int m_row_8 = m_col & 0x01, m_row_4 = m_col - m_row_8;

  2609. 

  2610.             for (y = row_and_7; y < h + row_and_7; y++) {

  2611.                 int col_mask_id = 2 - !(y & 7);

  2612. 

  2613.                 lflvl->mask[is_uv][0][y][1] |= m_row_8;

  2614.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  2615.                 // for odd lines, if the odd col is not being filtered,

  2616.                 // skip odd row also:

  2617.                 // .---. <-- a

  2618.                 // |   |

  2619.                 // |___| <-- b

  2620.                 // ^   ^

  2621.                 // c   d

  2622.                 //

  2623.                 // if a/c are even row/col and b/d are odd, and d is skipped,

  2624.                 // e.g. right edge of size-66x66.webm, then skip b also (bug)

  2625.                 if ((col_end & 1) && (y & 1)) {

  2626.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col_odd;

  2627.                 } else {

  2628.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col;

  2629.                 }

  2630.             }

  2631.         } else {

  2632.             int m_row_8 = m_col & 0x11, m_row_4 = m_col - m_row_8;

  2633. 

  2634.             for (y = row_and_7; y < h + row_and_7; y++) {

  2635.                 int col_mask_id = 2 - !(y & 3);

  2636. 

  2637.                 lflvl->mask[is_uv][0][y][1] |= m_row_8; // row edge

  2638.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  2639.                 lflvl->mask[is_uv][1][y][col_mask_id] |= m_col; // col edge

  2640.                 lflvl->mask[is_uv][0][y][3] |= m_col;

  2641.                 lflvl->mask[is_uv][1][y][3] |= m_col;

  2642.             }

  2643.         }

  2644.     } else {

  2645.         int y, t = 1 << col_and_7, m_col = (t << w) - t;

  2646. 

  2647.         if (!skip_inter) {

  2648.             int mask_id = (tx == TX_8X8);

  2649.             int l2 = tx + is_uv - 1, step1d = 1 << l2;

  2650.             static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };

  2651.             int m_row = m_col & masks[l2];

  2652. 

  2653.             // at odd UV col/row edges tx16/tx32 loopfilter edges, force

  2654.             // 8wd loopfilter to prevent going off the visible edge.

  2655.             if (is_uv && tx > TX_8X8 && (w ^ (w - 1)) == 1) {

  2656.                 int m_row_16 = ((t << (w - 1)) - t) & masks[l2];

  2657.                 int m_row_8 = m_row - m_row_16;

  2658. 

  2659.                 for (y = row_and_7; y < h + row_and_7; y++) {

  2660.                     lflvl->mask[is_uv][0][y][0] |= m_row_16;

  2661.                     lflvl->mask[is_uv][0][y][1] |= m_row_8;

  2662.                 }

  2663.             } else {

  2664.                 for (y = row_and_7; y < h + row_and_7; y++)

  2665.                     lflvl->mask[is_uv][0][y][mask_id] |= m_row;

  2666.             }

  2667. 

  2668.             if (is_uv && tx > TX_8X8 && (h ^ (h - 1)) == 1) {

  2669.                 for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)

  2670.                     lflvl->mask[is_uv][1][y][0] |= m_col;

  2671.                 if (y - row_and_7 == h - 1)

  2672.                     lflvl->mask[is_uv][1][y][1] |= m_col;

  2673.             } else {

  2674.                 for (y = row_and_7; y < h + row_and_7; y += step1d)

  2675.                     lflvl->mask[is_uv][1][y][mask_id] |= m_col;

  2676.             }

  2677.         } else if (tx != TX_4X4) {

  2678.             int mask_id;

  2679. 

  2680.             mask_id = (tx == TX_8X8) || (is_uv && h == 1);

  2681.             lflvl->mask[is_uv][1][row_and_7][mask_id] |= m_col;

  2682.             mask_id = (tx == TX_8X8) || (is_uv && w == 1);

  2683.             for (y = row_and_7; y < h + row_and_7; y++)

  2684.                 lflvl->mask[is_uv][0][y][mask_id] |= t;

  2685.         } else if (is_uv) {

  2686.             int t8 = t & 0x01, t4 = t - t8;

  2687. 

  2688.             for (y = row_and_7; y < h + row_and_7; y++) {

  2689.                 lflvl->mask[is_uv][0][y][2] |= t4;

  2690.                 lflvl->mask[is_uv][0][y][1] |= t8;

  2691.             }

  2692.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 7)] |= m_col;

  2693.         } else {

  2694.             int t8 = t & 0x11, t4 = t - t8;

  2695. 

  2696.             for (y = row_and_7; y < h + row_and_7; y++) {

  2697.                 lflvl->mask[is_uv][0][y][2] |= t4;

  2698.                 lflvl->mask[is_uv][0][y][1] |= t8;

  2699.             }

  2700.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 3)] |= m_col;

  2701.         }

  2702.     }

  2703. }

  2704. 

  2705. static void decode_b(AVCodecContext *ctx, int row, int col,

  2706.                      struct VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,

  2707.                      enum BlockLevel bl, enum BlockPartition bp)

  2708. {

  2709.     VP9Context *s = ctx->priv_data;

  2710.     VP9Block *b = s->b;

  2711.     enum BlockSize bs = bl * 3 + bp;

  2712.     int y, w4 = bwh_tab[1][bs][0], h4 = bwh_tab[1][bs][1], lvl;

  2713.     int emu[2];

  2714.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2715. 

  2716.     s->row = row;

  2717.     s->row7 = row & 7;

  2718.     s->col = col;

  2719.     s->col7 = col & 7;

  2720.     s->min_mv.x = -(128 + col * 64);

  2721.     s->min_mv.y = -(128 + row * 64);

  2722.     s->max_mv.x = 128 + (s->cols - col - w4) * 64;

  2723.     s->max_mv.y = 128 + (s->rows - row - h4) * 64;

  2724.     if (s->pass < 2) {

  2725.         b->bs = bs;

  2726.         b->bl = bl;

  2727.         b->bp = bp;

  2728.         decode_mode(ctx);

  2729.         b->uvtx = b->tx - (w4 * 2 == (1 << b->tx) || h4 * 2 == (1 << b->tx));

  2730. 

  2731.         if (!b->skip) {

  2732.             decode_coeffs(ctx);

  2733.         } else {

  2734.             int pl;

  2735. 

  2736.             memset(&s->above_y_nnz_ctx[col * 2], 0, w4 * 2);

  2737.             memset(&s->left_y_nnz_ctx[(row & 7) << 1], 0, h4 * 2);

  2738.             for (pl = 0; pl < 2; pl++) {

  2739.                 memset(&s->above_uv_nnz_ctx[pl][col], 0, w4);

  2740.                 memset(&s->left_uv_nnz_ctx[pl][row & 7], 0, h4);

  2741.             }

  2742.         }

  2743.         if (s->pass == 1) {

  2744.             s->b++;

  2745.             s->block += w4 * h4 * 64;

  2746.             s->uvblock[0] += w4 * h4 * 16;

  2747.             s->uvblock[1] += w4 * h4 * 16;

  2748.             s->eob += 4 * w4 * h4;

  2749.             s->uveob[0] += w4 * h4;

  2750.             s->uveob[1] += w4 * h4;

  2751. 

  2752.             return;

  2753.         }

  2754.     }

  2755. 

  2756.     // emulated overhangs if the stride of the target buffer can't hold. This

  2757.     // allows to support emu-edge and so on even if we have large block

  2758.     // overhangs

  2759.     emu[0] = (col + w4) * 8 > f->linesize[0] ||

  2760.              (row + h4) > s->rows;

  2761.     emu[1] = (col + w4) * 4 > f->linesize[1] ||

  2762.              (row + h4) > s->rows;

  2763.     if (emu[0]) {

  2764.         s->dst[0] = s->tmp_y;

  2765.         s->y_stride = 64;

  2766.     } else {

  2767.         s->dst[0] = f->data[0] + yoff;

  2768.         s->y_stride = f->linesize[0];

  2769.     }

  2770.     if (emu[1]) {

  2771.         s->dst[1] = s->tmp_uv[0];

  2772.         s->dst[2] = s->tmp_uv[1];

  2773.         s->uv_stride = 32;

  2774.     } else {

  2775.         s->dst[1] = f->data[1] + uvoff;

  2776.         s->dst[2] = f->data[2] + uvoff;

  2777.         s->uv_stride = f->linesize[1];

  2778.     }

  2779.     if (b->intra) {

  2780.         intra_recon(ctx, yoff, uvoff);

  2781.     } else {

  2782.         inter_recon(ctx);

  2783.     }

  2784.     if (emu[0]) {

  2785.         int w = FFMIN(s->cols - col, w4) * 8, h = FFMIN(s->rows - row, h4) * 8, n, o = 0;

  2786. 

  2787.         for (n = 0; o < w; n++) {

  2788.             int bw = 64 >> n;

  2789. 

  2790.             av_assert2(n <= 4);

  2791.             if (w & bw) {

  2792.                 s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o, f->linesize[0],

  2793.                                          s->tmp_y + o, 64, h, 0, 0);

  2794.                 o += bw;

  2795.             }

  2796.         }

  2797.     }

  2798.     if (emu[1]) {

  2799.         int w = FFMIN(s->cols - col, w4) * 4, h = FFMIN(s->rows - row, h4) * 4, n, o = 0;

  2800. 

  2801.         for (n = 1; o < w; n++) {

  2802.             int bw = 64 >> n;

  2803. 

  2804.             av_assert2(n <= 4);

  2805.             if (w & bw) {

  2806.                 s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o, f->linesize[1],

  2807.                                          s->tmp_uv[0] + o, 32, h, 0, 0);

  2808.                 s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o, f->linesize[2],

  2809.                                          s->tmp_uv[1] + o, 32, h, 0, 0);

  2810.                 o += bw;

  2811.             }

  2812.         }

  2813.     }

  2814. 

  2815.     // pick filter level and find edges to apply filter to

  2816.     if (s->filter.level &&

  2817.         (lvl = s->segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]

  2818.                                                     [b->mode[3] != ZEROMV]) > 0) {

  2819.         int x_end = FFMIN(s->cols - col, w4), y_end = FFMIN(s->rows - row, h4);

  2820.         int skip_inter = !b->intra && b->skip;

  2821. 

  2822.         for (y = 0; y < h4; y++)

  2823.             memset(&lflvl->level[((row & 7) + y) * 8 + (col & 7)], lvl, w4);

  2824.         mask_edges(lflvl, 0, row & 7, col & 7, x_end, y_end, 0, 0, b->tx, skip_inter);

  2825.         mask_edges(lflvl, 1, row & 7, col & 7, x_end, y_end,

  2826.                    s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,

  2827.                    s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,

  2828.                    b->uvtx, skip_inter);

  2829. 

  2830.         if (!s->filter.lim_lut[lvl]) {

  2831.             int sharp = s->filter.sharpness;

  2832.             int limit = lvl;

  2833. 

  2834.             if (sharp > 0) {

  2835.                 limit >>= (sharp + 3) >> 2;

  2836.                 limit = FFMIN(limit, 9 - sharp);

  2837.             }

  2838.             limit = FFMAX(limit, 1);

  2839. 

  2840.             s->filter.lim_lut[lvl] = limit;

  2841.             s->filter.mblim_lut[lvl] = 2 * (lvl + 2) + limit;

  2842.         }

  2843.     }

  2844. 

  2845.     if (s->pass == 2) {

  2846.         s->b++;

  2847.         s->block += w4 * h4 * 64;

  2848.         s->uvblock[0] += w4 * h4 * 16;

  2849.         s->uvblock[1] += w4 * h4 * 16;

  2850.         s->eob += 4 * w4 * h4;

  2851.         s->uveob[0] += w4 * h4;

  2852.         s->uveob[1] += w4 * h4;

  2853.     }

  2854. }

  2855. 

  2856. static void decode_sb(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  2857.                       ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  2858. {

  2859.     VP9Context *s = ctx->priv_data;

  2860.     int c = ((s->above_partition_ctx[col] >> (3 - bl)) & 1) |

  2861.             (((s->left_partition_ctx[row & 0x7] >> (3 - bl)) & 1) << 1);

  2862.     const uint8_t *p = s->keyframe ? vp9_default_kf_partition_probs[bl][c] :

  2863.                                      s->prob.p.partition[bl][c];

  2864.     enum BlockPartition bp;

  2865.     ptrdiff_t hbs = 4 >> bl;

  2866.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2867.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  2868. 

  2869.     if (bl == BL_8X8) {

  2870.         bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  2871.         decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2872.     } else if (col + hbs < s->cols) { // FIXME why not <=?

  2873.         if (row + hbs < s->rows) { // FIXME why not <=?

  2874.             bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  2875.             switch (bp) {

  2876.             case PARTITION_NONE:

  2877.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2878.                 break;

  2879.             case PARTITION_H:

  2880.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2881.                 yoff  += hbs * 8 * y_stride;

  2882.                 uvoff += hbs * 4 * uv_stride;

  2883.                 decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, bl, bp);

  2884.                 break;

  2885.             case PARTITION_V:

  2886.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2887.                 yoff  += hbs * 8;

  2888.                 uvoff += hbs * 4;

  2889.                 decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, bl, bp);

  2890.                 break;

  2891.             case PARTITION_SPLIT:

  2892.                 decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2893.                 decode_sb(ctx, row, col + hbs, lflvl,

  2894.                           yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2895.                 yoff  += hbs * 8 * y_stride;

  2896.                 uvoff += hbs * 4 * uv_stride;

  2897.                 decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  2898.                 decode_sb(ctx, row + hbs, col + hbs, lflvl,

  2899.                           yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2900.                 break;

  2901.             default:

  2902.                 av_assert0(0);

  2903.             }

  2904.         } else if (vp56_rac_get_prob_branchy(&s->c, p[1])) {

  2905.             bp = PARTITION_SPLIT;

  2906.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2907.             decode_sb(ctx, row, col + hbs, lflvl,

  2908.                       yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2909.         } else {

  2910.             bp = PARTITION_H;

  2911.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2912.         }

  2913.     } else if (row + hbs < s->rows) { // FIXME why not <=?

  2914.         if (vp56_rac_get_prob_branchy(&s->c, p[2])) {

  2915.             bp = PARTITION_SPLIT;

  2916.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2917.             yoff  += hbs * 8 * y_stride;

  2918.             uvoff += hbs * 4 * uv_stride;

  2919.             decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  2920.         } else {

  2921.             bp = PARTITION_V;

  2922.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2923.         }

  2924.     } else {

  2925.         bp = PARTITION_SPLIT;

  2926.         decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2927.     }

  2928.     s->counts.partition[bl][c][bp]++;

  2929. }

  2930. 

  2931. static void decode_sb_mem(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  2932.                           ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  2933. {

  2934.     VP9Context *s = ctx->priv_data;

  2935.     VP9Block *b = s->b;

  2936.     ptrdiff_t hbs = 4 >> bl;

  2937.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2938.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  2939. 

  2940.     if (bl == BL_8X8) {

  2941.         av_assert2(b->bl == BL_8X8);

  2942.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  2943.     } else if (s->b->bl == bl) {

  2944.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  2945.         if (b->bp == PARTITION_H && row + hbs < s->rows) {

  2946.             yoff  += hbs * 8 * y_stride;

  2947.             uvoff += hbs * 4 * uv_stride;

  2948.             decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, b->bl, b->bp);

  2949.         } else if (b->bp == PARTITION_V && col + hbs < s->cols) {

  2950.             yoff  += hbs * 8;

  2951.             uvoff += hbs * 4;

  2952.             decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, b->bl, b->bp);

  2953.         }

  2954.     } else {

  2955.         decode_sb_mem(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2956.         if (col + hbs < s->cols) { // FIXME why not <=?

  2957.             if (row + hbs < s->rows) {

  2958.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff + 8 * hbs,

  2959.                               uvoff + 4 * hbs, bl + 1);

  2960.                 yoff  += hbs * 8 * y_stride;

  2961.                 uvoff += hbs * 4 * uv_stride;

  2962.                 decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  2963.                 decode_sb_mem(ctx, row + hbs, col + hbs, lflvl,

  2964.                                     yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2965.             } else {

  2966.                 yoff  += hbs * 8;

  2967.                 uvoff += hbs * 4;

  2968.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff, uvoff, bl + 1);

  2969.             }

  2970.         } else if (row + hbs < s->rows) {

  2971.             yoff  += hbs * 8 * y_stride;

  2972.             uvoff += hbs * 4 * uv_stride;

  2973.             decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  2974.         }

  2975.     }

  2976. }

  2977. 

  2978. static void loopfilter_sb(AVCodecContext *ctx, struct VP9Filter *lflvl,

  2979.                           int row, int col, ptrdiff_t yoff, ptrdiff_t uvoff)

  2980. {

  2981.     VP9Context *s = ctx->priv_data;

  2982.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2983.     uint8_t *dst = f->data[0] + yoff, *lvl = lflvl->level;

  2984.     ptrdiff_t ls_y = f->linesize[0], ls_uv = f->linesize[1];

  2985.     int y, x, p;

  2986. 

  2987.     // FIXME in how far can we interleave the v/h loopfilter calls? E.g.

  2988.     // if you think of them as acting on a 8x8 block max, we can interleave

  2989.     // each v/h within the single x loop, but that only works if we work on

  2990.     // 8 pixel blocks, and we won't always do that (we want at least 16px

  2991.     // to use SSE2 optimizations, perhaps 32 for AVX2)

  2992. 

  2993.     // filter edges between columns, Y plane (e.g. block1 | block2)

  2994.     for (y = 0; y < 8; y += 2, dst += 16 * ls_y, lvl += 16) {

  2995.         uint8_t *ptr = dst, *l = lvl, *hmask1 = lflvl->mask[0][0][y];

  2996.         uint8_t *hmask2 = lflvl->mask[0][0][y + 1];

  2997.         unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2], hm13 = hmask1[3];

  2998.         unsigned hm2 = hmask2[1] | hmask2[2], hm23 = hmask2[3];

  2999.         unsigned hm = hm1 | hm2 | hm13 | hm23;

  3000. 

  3001.         for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 8, l++) {

  3002.             if (hm1 & x) {

  3003.                 int L = *l, H = L >> 4;

  3004.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3005. 

  3006.                 if (col || x > 1) {

  3007.                     if (hmask1[0] & x) {

  3008.                         if (hmask2[0] & x) {

  3009.                             av_assert2(l[8] == L);

  3010.                             s->dsp.loop_filter_16[0](ptr, ls_y, E, I, H);

  3011.                         } else {

  3012.                             s->dsp.loop_filter_8[2][0](ptr, ls_y, E, I, H);

  3013.                         }

  3014.                     } else if (hm2 & x) {

  3015.                         L = l[8];

  3016.                         H |= (L >> 4) << 8;

  3017.                         E |= s->filter.mblim_lut[L] << 8;

  3018.                         I |= s->filter.lim_lut[L] << 8;

  3019.                         s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3020.                                                [!!(hmask2[1] & x)]

  3021.                                                [0](ptr, ls_y, E, I, H);

  3022.                     } else {

  3023.                         s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3024.                                             [0](ptr, ls_y, E, I, H);

  3025.                     }

  3026.                 }

  3027.             } else if (hm2 & x) {

  3028.                 int L = l[8], H = L >> 4;

  3029.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3030. 

  3031.                 if (col || x > 1) {

  3032.                     s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3033.                                         [0](ptr + 8 * ls_y, ls_y, E, I, H);

  3034.                 }

  3035.             }

  3036.             if (hm13 & x) {

  3037.                 int L = *l, H = L >> 4;

  3038.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3039. 

  3040.                 if (hm23 & x) {

  3041.                     L = l[8];

  3042.                     H |= (L >> 4) << 8;

  3043.                     E |= s->filter.mblim_lut[L] << 8;

  3044.                     I |= s->filter.lim_lut[L] << 8;

  3045.                     s->dsp.loop_filter_mix2[0][0][0](ptr + 4, ls_y, E, I, H);

  3046.                 } else {

  3047.                     s->dsp.loop_filter_8[0][0](ptr + 4, ls_y, E, I, H);

  3048.                 }

  3049.             } else if (hm23 & x) {

  3050.                 int L = l[8], H = L >> 4;

  3051.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3052. 

  3053.                 s->dsp.loop_filter_8[0][0](ptr + 8 * ls_y + 4, ls_y, E, I, H);

  3054.             }

  3055.         }

  3056.     }

  3057. 

  3058.     //                                          block1

  3059.     // filter edges between rows, Y plane (e.g. ------)

  3060.     //                                          block2

  3061.     dst = f->data[0] + yoff;

  3062.     lvl = lflvl->level;

  3063.     for (y = 0; y < 8; y++, dst += 8 * ls_y, lvl += 8) {

  3064.         uint8_t *ptr = dst, *l = lvl, *vmask = lflvl->mask[0][1][y];

  3065.         unsigned vm = vmask[0] | vmask[1] | vmask[2], vm3 = vmask[3];

  3066. 

  3067.         for (x = 1; vm & ~(x - 1); x <<= 2, ptr += 16, l += 2) {

  3068.             if (row || y) {

  3069.                 if (vm & x) {

  3070.                     int L = *l, H = L >> 4;

  3071.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3072. 

  3073.                     if (vmask[0] & x) {

  3074.                         if (vmask[0] & (x << 1)) {

  3075.                             av_assert2(l[1] == L);

  3076.                             s->dsp.loop_filter_16[1](ptr, ls_y, E, I, H);

  3077.                         } else {

  3078.                             s->dsp.loop_filter_8[2][1](ptr, ls_y, E, I, H);

  3079.                         }

  3080.                     } else if (vm & (x << 1)) {

  3081.                         L = l[1];

  3082.                         H |= (L >> 4) << 8;

  3083.                         E |= s->filter.mblim_lut[L] << 8;

  3084.                         I |= s->filter.lim_lut[L] << 8;

  3085.                         s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3086.                                                [!!(vmask[1] & (x << 1))]

  3087.                                                [1](ptr, ls_y, E, I, H);

  3088.                     } else {

  3089.                         s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3090.                                             [1](ptr, ls_y, E, I, H);

  3091.                     }

  3092.                 } else if (vm & (x << 1)) {

  3093.                     int L = l[1], H = L >> 4;

  3094.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3095. 

  3096.                     s->dsp.loop_filter_8[!!(vmask[1] & (x << 1))]

  3097.                                         [1](ptr + 8, ls_y, E, I, H);

  3098.                 }

  3099.             }

  3100.             if (vm3 & x) {

  3101.                 int L = *l, H = L >> 4;

  3102.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3103. 

  3104.                 if (vm3 & (x << 1)) {

  3105.                     L = l[1];

  3106.                     H |= (L >> 4) << 8;

  3107.                     E |= s->filter.mblim_lut[L] << 8;

  3108.                     I |= s->filter.lim_lut[L] << 8;

  3109.                     s->dsp.loop_filter_mix2[0][0][1](ptr + ls_y * 4, ls_y, E, I, H);

  3110.                 } else {

  3111.                     s->dsp.loop_filter_8[0][1](ptr + ls_y * 4, ls_y, E, I, H);

  3112.                 }

  3113.             } else if (vm3 & (x << 1)) {

  3114.                 int L = l[1], H = L >> 4;

  3115.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3116. 

  3117.                 s->dsp.loop_filter_8[0][1](ptr + ls_y * 4 + 8, ls_y, E, I, H);

  3118.             }

  3119.         }

  3120.     }

  3121. 

  3122.     // same principle but for U/V planes

  3123.     for (p = 0; p < 2; p++) {

  3124.         lvl = lflvl->level;

  3125.         dst = f->data[1 + p] + uvoff;

  3126.         for (y = 0; y < 8; y += 4, dst += 16 * ls_uv, lvl += 32) {

  3127.             uint8_t *ptr = dst, *l = lvl, *hmask1 = lflvl->mask[1][0][y];

  3128.             uint8_t *hmask2 = lflvl->mask[1][0][y + 2];

  3129.             unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2];

  3130.             unsigned hm2 = hmask2[1] | hmask2[2], hm = hm1 | hm2;

  3131. 

  3132.             for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 4) {

  3133.                 if (col || x > 1) {

  3134.                     if (hm1 & x) {

  3135.                         int L = *l, H = L >> 4;

  3136.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3137. 

  3138.                         if (hmask1[0] & x) {

  3139.                             if (hmask2[0] & x) {

  3140.                                 av_assert2(l[16] == L);

  3141.                                 s->dsp.loop_filter_16[0](ptr, ls_uv, E, I, H);

  3142.                             } else {

  3143.                                 s->dsp.loop_filter_8[2][0](ptr, ls_uv, E, I, H);

  3144.                             }

  3145.                         } else if (hm2 & x) {

  3146.                             L = l[16];

  3147.                             H |= (L >> 4) << 8;

  3148.                             E |= s->filter.mblim_lut[L] << 8;

  3149.                             I |= s->filter.lim_lut[L] << 8;

  3150.                             s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3151.                                                    [!!(hmask2[1] & x)]

  3152.                                                    [0](ptr, ls_uv, E, I, H);

  3153.                         } else {

  3154.                             s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3155.                                                 [0](ptr, ls_uv, E, I, H);

  3156.                         }

  3157.                     } else if (hm2 & x) {

  3158.                         int L = l[16], H = L >> 4;

  3159.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3160. 

  3161.                         s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3162.                                             [0](ptr + 8 * ls_uv, ls_uv, E, I, H);

  3163.                     }

  3164.                 }

  3165.                 if (x & 0xAA)

  3166.                     l += 2;

  3167.             }

  3168.         }

  3169.         lvl = lflvl->level;

  3170.         dst = f->data[1 + p] + uvoff;

  3171.         for (y = 0; y < 8; y++, dst += 4 * ls_uv) {

  3172.             uint8_t *ptr = dst, *l = lvl, *vmask = lflvl->mask[1][1][y];

  3173.             unsigned vm = vmask[0] | vmask[1] | vmask[2];

  3174. 

  3175.             for (x = 1; vm & ~(x - 1); x <<= 4, ptr += 16, l += 4) {

  3176.                 if (row || y) {

  3177.                     if (vm & x) {

  3178.                         int L = *l, H = L >> 4;

  3179.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3180. 

  3181.                         if (vmask[0] & x) {

  3182.                             if (vmask[0] & (x << 2)) {

  3183.                                 av_assert2(l[2] == L);

  3184.                                 s->dsp.loop_filter_16[1](ptr, ls_uv, E, I, H);

  3185.                             } else {

  3186.                                 s->dsp.loop_filter_8[2][1](ptr, ls_uv, E, I, H);

  3187.                             }

  3188.                         } else if (vm & (x << 2)) {

  3189.                             L = l[2];

  3190.                             H |= (L >> 4) << 8;

  3191.                             E |= s->filter.mblim_lut[L] << 8;

  3192.                             I |= s->filter.lim_lut[L] << 8;

  3193.                             s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3194.                                                    [!!(vmask[1] & (x << 2))]

  3195.                                                    [1](ptr, ls_uv, E, I, H);

  3196.                         } else {

  3197.                             s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3198.                                                 [1](ptr, ls_uv, E, I, H);

  3199.                         }

  3200.                     } else if (vm & (x << 2)) {

  3201.                         int L = l[2], H = L >> 4;

  3202.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3203. 

  3204.                         s->dsp.loop_filter_8[!!(vmask[1] & (x << 2))]

  3205.                                             [1](ptr + 8, ls_uv, E, I, H);

  3206.                     }

  3207.                 }

  3208.             }

  3209.             if (y & 1)

  3210.                 lvl += 16;

  3211.         }

  3212.     }

  3213. }

  3214. 

  3215. static void set_tile_offset(int *start, int *end, int idx, int log2_n, int n)

  3216. {

  3217.     int sb_start = ( idx      * n) >> log2_n;

  3218.     int sb_end   = ((idx + 1) * n) >> log2_n;

  3219.     *start = FFMIN(sb_start, n) << 3;

  3220.     *end   = FFMIN(sb_end,   n) << 3;

  3221. }

  3222. 

  3223. static av_always_inline void adapt_prob(uint8_t *p, unsigned ct0, unsigned ct1,

  3224.                                         int max_count, int update_factor)

  3225. {

  3226.     unsigned ct = ct0 + ct1, p2, p1;

  3227. 

  3228.     if (!ct)

  3229.         return;

  3230. 

  3231.     p1 = *p;

  3232.     p2 = ((ct0 << 8) + (ct >> 1)) / ct;

  3233.     p2 = av_clip(p2, 1, 255);

  3234.     ct = FFMIN(ct, max_count);

  3235.     update_factor = FASTDIV(update_factor * ct, max_count);

  3236. 

  3237.     // (p1 * (256 - update_factor) + p2 * update_factor + 128) >> 8

  3238.     *p = p1 + (((p2 - p1) * update_factor + 128) >> 8);

  3239. }

  3240. 

  3241. static void adapt_probs(VP9Context *s)

  3242. {

  3243.     int i, j, k, l, m;

  3244.     prob_context *p = &s->prob_ctx[s->framectxid].p;

  3245.     int uf = (s->keyframe || s->intraonly || !s->last_keyframe) ? 112 : 128;

  3246. 

  3247.     // coefficients

  3248.     for (i = 0; i < 4; i++)

  3249.         for (j = 0; j < 2; j++)

  3250.             for (k = 0; k < 2; k++)

  3251.                 for (l = 0; l < 6; l++)

  3252.                     for (m = 0; m < 6; m++) {

  3253.                         uint8_t *pp = s->prob_ctx[s->framectxid].coef[i][j][k][l][m];

  3254.                         unsigned *e = s->counts.eob[i][j][k][l][m];

  3255.                         unsigned *c = s->counts.coef[i][j][k][l][m];

  3256. 

  3257.                         if (l == 0 && m >= 3) // dc only has 3 pt

  3258.                             break;

  3259. 

  3260.                         adapt_prob(&pp[0], e[0], e[1], 24, uf);

  3261.                         adapt_prob(&pp[1], c[0], c[1] + c[2], 24, uf);

  3262.                         adapt_prob(&pp[2], c[1], c[2], 24, uf);

  3263.                     }

  3264. 

  3265.     if (s->keyframe || s->intraonly) {

  3266.         memcpy(p->skip,  s->prob.p.skip,  sizeof(p->skip));

  3267.         memcpy(p->tx32p, s->prob.p.tx32p, sizeof(p->tx32p));

  3268.         memcpy(p->tx16p, s->prob.p.tx16p, sizeof(p->tx16p));

  3269.         memcpy(p->tx8p,  s->prob.p.tx8p,  sizeof(p->tx8p));

  3270.         return;

  3271.     }

  3272. 

  3273.     // skip flag

  3274.     for (i = 0; i < 3; i++)

  3275.         adapt_prob(&p->skip[i], s->counts.skip[i][0], s->counts.skip[i][1], 20, 128);

  3276. 

  3277.     // intra/inter flag

  3278.     for (i = 0; i < 4; i++)

  3279.         adapt_prob(&p->intra[i], s->counts.intra[i][0], s->counts.intra[i][1], 20, 128);

  3280. 

  3281.     // comppred flag

  3282.     if (s->comppredmode == PRED_SWITCHABLE) {

  3283.       for (i = 0; i < 5; i++)

  3284.           adapt_prob(&p->comp[i], s->counts.comp[i][0], s->counts.comp[i][1], 20, 128);

  3285.     }

  3286. 

  3287.     // reference frames

  3288.     if (s->comppredmode != PRED_SINGLEREF) {

  3289.       for (i = 0; i < 5; i++)

  3290.           adapt_prob(&p->comp_ref[i], s->counts.comp_ref[i][0],

  3291.                      s->counts.comp_ref[i][1], 20, 128);

  3292.     }

  3293. 

  3294.     if (s->comppredmode != PRED_COMPREF) {

  3295.       for (i = 0; i < 5; i++) {

  3296.           uint8_t *pp = p->single_ref[i];

  3297.           unsigned (*c)[2] = s->counts.single_ref[i];

  3298. 

  3299.           adapt_prob(&pp[0], c[0][0], c[0][1], 20, 128);

  3300.           adapt_prob(&pp[1], c[1][0], c[1][1], 20, 128);

  3301.       }

  3302.     }

  3303. 

  3304.     // block partitioning

  3305.     for (i = 0; i < 4; i++)

  3306.         for (j = 0; j < 4; j++) {

  3307.             uint8_t *pp = p->partition[i][j];

  3308.             unsigned *c = s->counts.partition[i][j];

  3309. 

  3310.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3311.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3312.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3313.         }

  3314. 

  3315.     // tx size

  3316.     if (s->txfmmode == TX_SWITCHABLE) {

  3317.       for (i = 0; i < 2; i++) {

  3318.           unsigned *c16 = s->counts.tx16p[i], *c32 = s->counts.tx32p[i];

  3319. 

  3320.           adapt_prob(&p->tx8p[i], s->counts.tx8p[i][0], s->counts.tx8p[i][1], 20, 128);

  3321.           adapt_prob(&p->tx16p[i][0], c16[0], c16[1] + c16[2], 20, 128);

  3322.           adapt_prob(&p->tx16p[i][1], c16[1], c16[2], 20, 128);

  3323.           adapt_prob(&p->tx32p[i][0], c32[0], c32[1] + c32[2] + c32[3], 20, 128);

  3324.           adapt_prob(&p->tx32p[i][1], c32[1], c32[2] + c32[3], 20, 128);

  3325.           adapt_prob(&p->tx32p[i][2], c32[2], c32[3], 20, 128);

  3326.       }

  3327.     }

  3328. 

  3329.     // interpolation filter

  3330.     if (s->filtermode == FILTER_SWITCHABLE) {

  3331.         for (i = 0; i < 4; i++) {

  3332.             uint8_t *pp = p->filter[i];

  3333.             unsigned *c = s->counts.filter[i];

  3334. 

  3335.             adapt_prob(&pp[0], c[0], c[1] + c[2], 20, 128);

  3336.             adapt_prob(&pp[1], c[1], c[2], 20, 128);

  3337.         }

  3338.     }

  3339. 

  3340.     // inter modes

  3341.     for (i = 0; i < 7; i++) {

  3342.         uint8_t *pp = p->mv_mode[i];

  3343.         unsigned *c = s->counts.mv_mode[i];

  3344. 

  3345.         adapt_prob(&pp[0], c[2], c[1] + c[0] + c[3], 20, 128);

  3346.         adapt_prob(&pp[1], c[0], c[1] + c[3], 20, 128);

  3347.         adapt_prob(&pp[2], c[1], c[3], 20, 128);

  3348.     }

  3349. 

  3350.     // mv joints

  3351.     {

  3352.         uint8_t *pp = p->mv_joint;

  3353.         unsigned *c = s->counts.mv_joint;

  3354. 

  3355.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3356.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3357.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3358.     }

  3359. 

  3360.     // mv components

  3361.     for (i = 0; i < 2; i++) {

  3362.         uint8_t *pp;

  3363.         unsigned *c, (*c2)[2], sum;

  3364. 

  3365.         adapt_prob(&p->mv_comp[i].sign, s->counts.mv_comp[i].sign[0],

  3366.                    s->counts.mv_comp[i].sign[1], 20, 128);

  3367. 

  3368.         pp = p->mv_comp[i].classes;

  3369.         c = s->counts.mv_comp[i].classes;

  3370.         sum = c[1] + c[2] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9] + c[10];

  3371.         adapt_prob(&pp[0], c[0], sum, 20, 128);

  3372.         sum -= c[1];

  3373.         adapt_prob(&pp[1], c[1], sum, 20, 128);

  3374.         sum -= c[2] + c[3];

  3375.         adapt_prob(&pp[2], c[2] + c[3], sum, 20, 128);

  3376.         adapt_prob(&pp[3], c[2], c[3], 20, 128);

  3377.         sum -= c[4] + c[5];

  3378.         adapt_prob(&pp[4], c[4] + c[5], sum, 20, 128);

  3379.         adapt_prob(&pp[5], c[4], c[5], 20, 128);

  3380.         sum -= c[6];

  3381.         adapt_prob(&pp[6], c[6], sum, 20, 128);

  3382.         adapt_prob(&pp[7], c[7] + c[8], c[9] + c[10], 20, 128);

  3383.         adapt_prob(&pp[8], c[7], c[8], 20, 128);

  3384.         adapt_prob(&pp[9], c[9], c[10], 20, 128);

  3385. 

  3386.         adapt_prob(&p->mv_comp[i].class0, s->counts.mv_comp[i].class0[0],

  3387.                    s->counts.mv_comp[i].class0[1], 20, 128);

  3388.         pp = p->mv_comp[i].bits;

  3389.         c2 = s->counts.mv_comp[i].bits;

  3390.         for (j = 0; j < 10; j++)

  3391.             adapt_prob(&pp[j], c2[j][0], c2[j][1], 20, 128);

  3392. 

  3393.         for (j = 0; j < 2; j++) {

  3394.             pp = p->mv_comp[i].class0_fp[j];

  3395.             c = s->counts.mv_comp[i].class0_fp[j];

  3396.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3397.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3398.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3399.         }

  3400.         pp = p->mv_comp[i].fp;

  3401.         c = s->counts.mv_comp[i].fp;

  3402.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3403.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3404.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3405. 

  3406.         if (s->highprecisionmvs) {

  3407.             adapt_prob(&p->mv_comp[i].class0_hp, s->counts.mv_comp[i].class0_hp[0],

  3408.                        s->counts.mv_comp[i].class0_hp[1], 20, 128);

  3409.             adapt_prob(&p->mv_comp[i].hp, s->counts.mv_comp[i].hp[0],

  3410.                        s->counts.mv_comp[i].hp[1], 20, 128);

  3411.         }

  3412.     }

  3413. 

  3414.     // y intra modes

  3415.     for (i = 0; i < 4; i++) {

  3416.         uint8_t *pp = p->y_mode[i];

  3417.         unsigned *c = s->counts.y_mode[i], sum, s2;

  3418. 

  3419.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3420.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3421.         sum -= c[TM_VP8_PRED];

  3422.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3423.         sum -= c[VERT_PRED];

  3424.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3425.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3426.         sum -= s2;

  3427.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3428.         s2 -= c[HOR_PRED];

  3429.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3430.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3431.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3432.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3433.         sum -= c[VERT_LEFT_PRED];

  3434.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3435.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3436.     }

  3437. 

  3438.     // uv intra modes

  3439.     for (i = 0; i < 10; i++) {

  3440.         uint8_t *pp = p->uv_mode[i];

  3441.         unsigned *c = s->counts.uv_mode[i], sum, s2;

  3442. 

  3443.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3444.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3445.         sum -= c[TM_VP8_PRED];

  3446.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3447.         sum -= c[VERT_PRED];

  3448.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3449.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3450.         sum -= s2;

  3451.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3452.         s2 -= c[HOR_PRED];

  3453.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3454.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3455.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3456.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3457.         sum -= c[VERT_LEFT_PRED];

  3458.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3459.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3460.     }

  3461. }

  3462. 

  3463. static void free_buffers(VP9Context *s)

  3464. {

  3465.     av_freep(&s->above_partition_ctx);

  3466.     av_freep(&s->b_base);

  3467.     av_freep(&s->block_base);

  3468. }

  3469. 

  3470. static av_cold int vp9_decode_free(AVCodecContext *ctx)

  3471. {

  3472.     VP9Context *s = ctx->priv_data;

  3473.     int i;

  3474. 

  3475.     for (i = 0; i < 2; i++) {

  3476.         if (s->frames[i].tf.f->data[0])

  3477.             vp9_unref_frame(ctx, &s->frames[i]);

  3478.         av_frame_free(&s->frames[i].tf.f);

  3479.     }

  3480.     for (i = 0; i < 8; i++) {

  3481.         if (s->refs[i].f->data[0])

  3482.             ff_thread_release_buffer(ctx, &s->refs[i]);

  3483.         av_frame_free(&s->refs[i].f);

  3484.         if (s->next_refs[i].f->data[0])

  3485.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3486.         av_frame_free(&s->next_refs[i].f);

  3487.     }

  3488.     free_buffers(s);

  3489.     av_freep(&s->c_b);

  3490.     s->c_b_size = 0;

  3491. 

  3492.     return 0;

  3493. }

  3494. 

  3495. 

  3496. static int vp9_decode_frame(AVCodecContext *ctx, void *frame,

  3497.                             int *got_frame, AVPacket *pkt)

  3498. {

  3499.     const uint8_t *data = pkt->data;

  3500.     int size = pkt->size;

  3501.     VP9Context *s = ctx->priv_data;

  3502.     int res, tile_row, tile_col, i, ref, row, col;

  3503.     ptrdiff_t yoff, uvoff, ls_y, ls_uv;

  3504.     AVFrame *f;

  3505. 

  3506.     if ((res = decode_frame_header(ctx, data, size, &ref)) < 0) {

  3507.         return res;

  3508.     } else if (res == 0) {

  3509.         if (!s->refs[ref].f->data[0]) {

  3510.             av_log(ctx, AV_LOG_ERROR, "Requested reference %d not available\n", ref);

  3511.             return AVERROR_INVALIDDATA;

  3512.         }

  3513.         if ((res = av_frame_ref(frame, s->refs[ref].f)) < 0)

  3514.             return res;

  3515.         *got_frame = 1;

  3516.         return 0;

  3517.     }

  3518.     data += res;

  3519.     size -= res;

  3520. 

  3521.     if (s->frames[LAST_FRAME].tf.f->data[0])

  3522.         vp9_unref_frame(ctx, &s->frames[LAST_FRAME]);

  3523.     if (!s->keyframe && s->frames[CUR_FRAME].tf.f->data[0] &&

  3524.         (res = vp9_ref_frame(ctx, &s->frames[LAST_FRAME], &s->frames[CUR_FRAME])) < 0)

  3525.         return res;

  3526.     if (s->frames[CUR_FRAME].tf.f->data[0])

  3527.         vp9_unref_frame(ctx, &s->frames[CUR_FRAME]);

  3528.     if ((res = vp9_alloc_frame(ctx, &s->frames[CUR_FRAME])) < 0)

  3529.         return res;

  3530.     f = s->frames[CUR_FRAME].tf.f;

  3531.     f->key_frame = s->keyframe;

  3532.     f->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;

  3533.     ls_y = f->linesize[0];

  3534.     ls_uv =f->linesize[1];

  3535. 

  3536.     // ref frame setup

  3537.     for (i = 0; i < 8; i++) {

  3538.         if (s->next_refs[i].f->data[0])

  3539.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3540.         if (s->refreshrefmask & (1 << i)) {

  3541.             res = ff_thread_ref_frame(&s->next_refs[i], &s->frames[CUR_FRAME].tf);

  3542.         } else {

  3543.             res = ff_thread_ref_frame(&s->next_refs[i], &s->refs[i]);

  3544.         }

  3545.         if (res < 0)

  3546.             return res;

  3547.     }

  3548. 

  3549.     // main tile decode loop

  3550.     memset(s->above_partition_ctx, 0, s->cols);

  3551.     memset(s->above_skip_ctx, 0, s->cols);

  3552.     if (s->keyframe || s->intraonly) {

  3553.         memset(s->above_mode_ctx, DC_PRED, s->cols * 2);

  3554.     } else {

  3555.         memset(s->above_mode_ctx, NEARESTMV, s->cols);

  3556.     }

  3557.     memset(s->above_y_nnz_ctx, 0, s->sb_cols * 16);

  3558.     memset(s->above_uv_nnz_ctx[0], 0, s->sb_cols * 8);

  3559.     memset(s->above_uv_nnz_ctx[1], 0, s->sb_cols * 8);

  3560.     memset(s->above_segpred_ctx, 0, s->cols);

  3561.     s->pass = s->uses_2pass =

  3562.         ctx->active_thread_type == FF_THREAD_FRAME && s->refreshctx && !s->parallelmode;

  3563.     if (s->refreshctx && s->parallelmode) {

  3564.         int j, k, l, m;

  3565. 

  3566.         for (i = 0; i < 4; i++) {

  3567.             for (j = 0; j < 2; j++)

  3568.                 for (k = 0; k < 2; k++)

  3569.                     for (l = 0; l < 6; l++)

  3570.                         for (m = 0; m < 6; m++)

  3571.                             memcpy(s->prob_ctx[s->framectxid].coef[i][j][k][l][m],

  3572.                                    s->prob.coef[i][j][k][l][m], 3);

  3573.             if (s->txfmmode == i)

  3574.                 break;

  3575.         }

  3576.         s->prob_ctx[s->framectxid].p = s->prob.p;

  3577.         ff_thread_finish_setup(ctx);

  3578.     }

  3579. 

  3580.     do {

  3581.         yoff = uvoff = 0;

  3582.         s->b = s->b_base;

  3583.         s->block = s->block_base;

  3584.         s->uvblock[0] = s->uvblock_base[0];

  3585.         s->uvblock[1] = s->uvblock_base[1];

  3586.         s->eob = s->eob_base;

  3587.         s->uveob[0] = s->uveob_base[0];

  3588.         s->uveob[1] = s->uveob_base[1];

  3589. 

  3590.         for (tile_row = 0; tile_row < s->tiling.tile_rows; tile_row++) {

  3591.             set_tile_offset(&s->tiling.tile_row_start, &s->tiling.tile_row_end,

  3592.                             tile_row, s->tiling.log2_tile_rows, s->sb_rows);

  3593.             if (s->pass != 2) {

  3594.                 for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  3595.                     unsigned tile_size;

  3596. 

  3597.                     if (tile_col == s->tiling.tile_cols - 1 &&

  3598.                         tile_row == s->tiling.tile_rows - 1) {

  3599.                         tile_size = size;

  3600.                     } else {

  3601.                         tile_size = AV_RB32(data);

  3602.                         data += 4;

  3603.                         size -= 4;

  3604.                     }

  3605.                     if (tile_size > size) {

  3606.                         ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  3607.                         return AVERROR_INVALIDDATA;

  3608.                     }

  3609.                     ff_vp56_init_range_decoder(&s->c_b[tile_col], data, tile_size);

  3610.                     if (vp56_rac_get_prob_branchy(&s->c_b[tile_col], 128)) { // marker bit

  3611.                         ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  3612.                         return AVERROR_INVALIDDATA;

  3613.                     }

  3614.                     data += tile_size;

  3615.                     size -= tile_size;

  3616.                 }

  3617.             }

  3618. 

  3619.             for (row = s->tiling.tile_row_start; row < s->tiling.tile_row_end;

  3620.                  row += 8, yoff += ls_y * 64, uvoff += ls_uv * 32) {

  3621.                 struct VP9Filter *lflvl_ptr = s->lflvl;

  3622.                 ptrdiff_t yoff2 = yoff, uvoff2 = uvoff;

  3623. 

  3624.                 for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  3625.                     set_tile_offset(&s->tiling.tile_col_start, &s->tiling.tile_col_end,

  3626.                                     tile_col, s->tiling.log2_tile_cols, s->sb_cols);

  3627. 

  3628.                     if (s->pass != 2) {

  3629.                         memset(s->left_partition_ctx, 0, 8);

  3630.                         memset(s->left_skip_ctx, 0, 8);

  3631.                         if (s->keyframe || s->intraonly) {

  3632.                             memset(s->left_mode_ctx, DC_PRED, 16);

  3633.                         } else {

  3634.                             memset(s->left_mode_ctx, NEARESTMV, 8);

  3635.                         }

  3636.                         memset(s->left_y_nnz_ctx, 0, 16);

  3637.                         memset(s->left_uv_nnz_ctx, 0, 16);

  3638.                         memset(s->left_segpred_ctx, 0, 8);

  3639. 

  3640.                         memcpy(&s->c, &s->c_b[tile_col], sizeof(s->c));

  3641.                     }

  3642. 

  3643.                     for (col = s->tiling.tile_col_start;

  3644.                          col < s->tiling.tile_col_end;

  3645.                          col += 8, yoff2 += 64, uvoff2 += 32, lflvl_ptr++) {

  3646.                         // FIXME integrate with lf code (i.e. zero after each

  3647.                         // use, similar to invtxfm coefficients, or similar)

  3648.                         if (s->pass != 1) {

  3649.                             memset(lflvl_ptr->mask, 0, sizeof(lflvl_ptr->mask));

  3650.                         }

  3651. 

  3652.                         if (s->pass == 2) {

  3653.                             decode_sb_mem(ctx, row, col, lflvl_ptr,

  3654.                                           yoff2, uvoff2, BL_64X64);

  3655.                         } else {

  3656.                             decode_sb(ctx, row, col, lflvl_ptr,

  3657.                                       yoff2, uvoff2, BL_64X64);

  3658.                         }

  3659.                     }

  3660.                     if (s->pass != 2) {

  3661.                         memcpy(&s->c_b[tile_col], &s->c, sizeof(s->c));

  3662.                     }

  3663.                 }

  3664. 

  3665.                 if (s->pass == 1) {

  3666.                     continue;

  3667.                 }

  3668. 

  3669.                 // backup pre-loopfilter reconstruction data for intra

  3670.                 // prediction of next row of sb64s

  3671.                 if (row + 8 < s->rows) {

  3672.                     memcpy(s->intra_pred_data[0],

  3673.                            f->data[0] + yoff + 63 * ls_y,

  3674.                            8 * s->cols);

  3675.                     memcpy(s->intra_pred_data[1],

  3676.                            f->data[1] + uvoff + 31 * ls_uv,

  3677.                            4 * s->cols);

  3678.                     memcpy(s->intra_pred_data[2],

  3679.                            f->data[2] + uvoff + 31 * ls_uv,

  3680.                            4 * s->cols);

  3681.                 }

  3682. 

  3683.                 // loopfilter one row

  3684.                 if (s->filter.level) {

  3685.                     yoff2 = yoff;

  3686.                     uvoff2 = uvoff;

  3687.                     lflvl_ptr = s->lflvl;

  3688.                     for (col = 0; col < s->cols;

  3689.                          col += 8, yoff2 += 64, uvoff2 += 32, lflvl_ptr++) {

  3690.                         loopfilter_sb(ctx, lflvl_ptr, row, col, yoff2, uvoff2);

  3691.                     }

  3692.                 }

  3693. 

  3694.                 // FIXME maybe we can make this more finegrained by running the

  3695.                 // loopfilter per-block instead of after each sbrow

  3696.                 // In fact that would also make intra pred left preparation easier?

  3697.                 ff_thread_report_progress(&s->frames[CUR_FRAME].tf, row >> 3, 0);

  3698.             }

  3699.         }

  3700. 

  3701.         if (s->pass < 2 && s->refreshctx && !s->parallelmode) {

  3702.             adapt_probs(s);

  3703.             ff_thread_finish_setup(ctx);

  3704.         }

  3705.     } while (s->pass++ == 1);

  3706.     ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  3707. 

  3708.     // ref frame setup

  3709.     for (i = 0; i < 8; i++) {

  3710.         if (s->refs[i].f->data[0])

  3711.             ff_thread_release_buffer(ctx, &s->refs[i]);

  3712.         ff_thread_ref_frame(&s->refs[i], &s->next_refs[i]);

  3713.     }

  3714. 

  3715.     if (!s->invisible) {

  3716.         if ((res = av_frame_ref(frame, s->frames[CUR_FRAME].tf.f)) < 0)

  3717.             return res;

  3718.         *got_frame = 1;

  3719.     }

  3720. 

  3721.     return 0;

  3722. }

  3723. 

  3724. static void vp9_decode_flush(AVCodecContext *ctx)

  3725. {

  3726.     VP9Context *s = ctx->priv_data;

  3727.     int i;

  3728. 

  3729.     for (i = 0; i < 2; i++)

  3730.         vp9_unref_frame(ctx, &s->frames[i]);

  3731.     for (i = 0; i < 8; i++)

  3732.         ff_thread_release_buffer(ctx, &s->refs[i]);

  3733. }

  3734. 

  3735. static int init_frames(AVCodecContext *ctx)

  3736. {

  3737.     VP9Context *s = ctx->priv_data;

  3738.     int i;

  3739. 

  3740.     for (i = 0; i < 2; i++) {

  3741.         s->frames[i].tf.f = av_frame_alloc();

  3742.         if (!s->frames[i].tf.f) {

  3743.             vp9_decode_free(ctx);

  3744.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  3745.             return AVERROR(ENOMEM);

  3746.         }

  3747.     }

  3748.     for (i = 0; i < 8; i++) {

  3749.         s->refs[i].f = av_frame_alloc();

  3750.         s->next_refs[i].f = av_frame_alloc();

  3751.         if (!s->refs[i].f || !s->next_refs[i].f) {

  3752.             vp9_decode_free(ctx);

  3753.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  3754.             return AVERROR(ENOMEM);

  3755.         }

  3756.     }

  3757. 

  3758.     return 0;

  3759. }

  3760. 

  3761. static av_cold int vp9_decode_init(AVCodecContext *ctx)

  3762. {

  3763.     VP9Context *s = ctx->priv_data;

  3764. 

  3765.     ctx->internal->allocate_progress = 1;

  3766.     ctx->pix_fmt = AV_PIX_FMT_YUV420P;

  3767.     ff_vp9dsp_init(&s->dsp);

  3768.     ff_videodsp_init(&s->vdsp, 8);

  3769.     s->filter.sharpness = -1;

  3770. 

  3771.     return init_frames(ctx);

  3772. }

  3773. 

  3774. static av_cold int vp9_decode_init_thread_copy(AVCodecContext *avctx)

  3775. {

  3776.     return init_frames(avctx);

  3777. }

  3778. 

  3779. static int vp9_decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src)

  3780. {

  3781.     int i, res;

  3782.     VP9Context *s = dst->priv_data, *ssrc = src->priv_data;

  3783. 

  3784.     // detect size changes in other threads

  3785.     if (s->above_partition_ctx &&

  3786.         (!ssrc->above_partition_ctx || s->cols != ssrc->cols || s->rows != ssrc->rows)) {

  3787.         free_buffers(s);

  3788.     }

  3789. 

  3790.     for (i = 0; i < 2; i++) {

  3791.         if (s->frames[i].tf.f->data[0])

  3792.             vp9_unref_frame(dst, &s->frames[i]);

  3793.         if (ssrc->frames[i].tf.f->data[0]) {

  3794.             if ((res = vp9_ref_frame(dst, &s->frames[i], &ssrc->frames[i])) < 0)

  3795.                 return res;

  3796.         }

  3797.     }

  3798.     for (i = 0; i < 8; i++) {

  3799.         if (s->refs[i].f->data[0])

  3800.             ff_thread_release_buffer(dst, &s->refs[i]);

  3801.         if (ssrc->next_refs[i].f->data[0]) {

  3802.             if ((res = ff_thread_ref_frame(&s->refs[i], &ssrc->next_refs[i])) < 0)

  3803.                 return res;

  3804.         }

  3805.     }

  3806. 

  3807.     s->invisible = ssrc->invisible;

  3808.     s->keyframe = ssrc->keyframe;

  3809.     s->uses_2pass = ssrc->uses_2pass;

  3810.     memcpy(&s->prob_ctx, &ssrc->prob_ctx, sizeof(s->prob_ctx));

  3811.     memcpy(&s->lf_delta, &ssrc->lf_delta, sizeof(s->lf_delta));

  3812.     if (ssrc->segmentation.enabled) {

  3813.         memcpy(&s->segmentation.feat, &ssrc->segmentation.feat,

  3814.                sizeof(s->segmentation.feat));

  3815.     }

  3816. 

  3817.     return 0;

  3818. }

  3819. 

  3820. AVCodec ff_vp9_decoder = {

  3821.     .name                  = "vp9",

  3822.     .long_name             = NULL_IF_CONFIG_SMALL("Google VP9"),

  3823.     .type                  = AVMEDIA_TYPE_VIDEO,

  3824.     .id                    = AV_CODEC_ID_VP9,

  3825.     .priv_data_size        = sizeof(VP9Context),

  3826.     .init                  = vp9_decode_init,

  3827.     .close                 = vp9_decode_free,

  3828.     .decode                = vp9_decode_frame,

  3829.     .capabilities          = CODEC_CAP_DR1 | CODEC_CAP_FRAME_THREADS,

  3830.     .flush                 = vp9_decode_flush,

  3831.     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(vp9_decode_init_thread_copy),

  3832.     .update_thread_context = ONLY_IF_THREADS_ENABLED(vp9_decode_update_thread_context),

  3833. };