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.
59315 Commits
32 Branches
303MB
Tree: 4a55bffad3
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '4a55bffad3'
${ noResults }
FFmpeg/libavcodec/vp9.c

3810 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.     VP56mv 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.         memcpy(f->segmentation_map, s->frames[LAST_FRAME].segmentation_map, sz);

  269.     }

  270. 

  271.     return 0;

  272. }

  273. 

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

  275. {

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

  277.     av_buffer_unref(&f->extradata);

  278. }

  279. 

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

  281. {

  282.     int res;

  283. 

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

  285.         return res;

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

  287.         vp9_unref_frame(ctx, dst);

  288.         return AVERROR(ENOMEM);

  289.     }

  290. 

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

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

  293. 

  294.     return 0;

  295. }

  296. 

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

  298. {

  299.     VP9Context *s = ctx->priv_data;

  300.     uint8_t *p;

  301. 

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

  303. 

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

  305.         return 0;

  306. 

  307.     ctx->width  = w;

  308.     ctx->height = h;

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

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

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

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

  313. 

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

  315.     av_freep(&s->above_partition_ctx);

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

  317.     if (!p)

  318.         return AVERROR(ENOMEM);

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  336. #undef assign

  337. 

  338.     av_free(s->b_base);

  339.     av_free(s->block_base);

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

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

  342. 

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

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

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

  346.             return AVERROR(ENOMEM);

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

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

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

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

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

  352.     } else {

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

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

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

  356.             return AVERROR(ENOMEM);

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

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

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

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

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

  362.     }

  363. 

  364.     return 0;

  365. }

  366. 

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

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

  369. {

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

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

  372. }

  373. 

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

  375. {

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

  377. }

  378. 

  379. // differential forward probability updates

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

  381. {

  382.     static const int inv_map_table[254] = {

  383.           7,  20,  33,  46,  59,  72,  85,  98, 111, 124, 137, 150, 163, 176,

  384.         189, 202, 215, 228, 241, 254,   1,   2,   3,   4,   5,   6,   8,   9,

  385.          10,  11,  12,  13,  14,  15,  16,  17,  18,  19,  21,  22,  23,  24,

  386.          25,  26,  27,  28,  29,  30,  31,  32,  34,  35,  36,  37,  38,  39,

  387.          40,  41,  42,  43,  44,  45,  47,  48,  49,  50,  51,  52,  53,  54,

  388.          55,  56,  57,  58,  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,

  389.          70,  71,  73,  74,  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,

  390.          86,  87,  88,  89,  90,  91,  92,  93,  94,  95,  96,  97,  99, 100,

  391.         101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 112, 113, 114, 115,

  392.         116, 117, 118, 119, 120, 121, 122, 123, 125, 126, 127, 128, 129, 130,

  393.         131, 132, 133, 134, 135, 136, 138, 139, 140, 141, 142, 143, 144, 145,

  394.         146, 147, 148, 149, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160,

  395.         161, 162, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175,

  396.         177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 190, 191,

  397.         192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 204, 205, 206,

  398.         207, 208, 209, 210, 211, 212, 213, 214, 216, 217, 218, 219, 220, 221,

  399.         222, 223, 224, 225, 226, 227, 229, 230, 231, 232, 233, 234, 235, 236,

  400.         237, 238, 239, 240, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251,

  401.         252, 253,

  402.     };

  403.     int d;

  404. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  419. 

  420.     if (!vp8_rac_get(c)) {

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

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

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

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

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

  426.     } else {

  427.         d = vp8_rac_get_uint(c, 7);

  428.         if (d >= 65)

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

  430.         d += 64;

  431.     }

  432. 

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

  434.                     255 - inv_recenter_nonneg(inv_map_table[d], 255 - p);

  435. }

  436. 

  437. static int decode_frame_header(AVCodecContext *ctx,

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

  439. {

  440.     VP9Context *s = ctx->priv_data;

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

  442.     int last_invisible;

  443.     const uint8_t *data2;

  444. 

  445.     /* general header */

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

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

  448.         return res;

  449.     }

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

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

  452.         return AVERROR_INVALIDDATA;

  453.     }

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

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

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

  457.         return AVERROR_INVALIDDATA;

  458.     }

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

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

  461.         return 0;

  462.     }

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

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

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

  466.     last_invisible    = s->invisible;

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

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

  469.     // FIXME disable this upon resolution change

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

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

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

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

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

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

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

  535.             if (s->allowcompinter) {

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

  537.                     s->fixcompref    = 2;

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

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

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

  541.                     s->fixcompref    = 1;

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

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

  544.                 } else {

  545.                     s->fixcompref    = 0;

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

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

  548.                 }

  549.             }

  550.         }

  551.     }

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

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

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

  555. 

  556.     /* loopfilter header data */

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

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

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

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

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

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

  563.     s->filter.sharpness = sharp;

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

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

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

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

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

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

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

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

  572.         }

  573.     } else {

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

  575.     }

  576. 

  577.     /* quantization header data */

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

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

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

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

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

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

  584. 

  585.     /* segmentation header info */

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

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

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

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

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

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

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

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

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

  595.         }

  596. 

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

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

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

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

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

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

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

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

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

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

  607.             }

  608.         }

  609.     } else {

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

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

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

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

  614.     }

  615. 

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

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

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

  619. 

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

  621.             if (s->segmentation.absolute_vals)

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

  623.             else

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

  625.         } else {

  626.             qyac  = s->yac_qi;

  627.         }

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

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

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

  631.         qyac  = av_clip_uintp2(qyac, 8);

  632. 

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

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

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

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

  637. 

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

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

  640.             if (s->segmentation.absolute_vals)

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

  642.             else

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

  644.         } else {

  645.             lflvl  = s->filter.level;

  646.         }

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

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

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

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

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

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

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

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

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

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

  657.         }

  658.     }

  659. 

  660.     /* tiling info */

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

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

  663.         return res;

  664.     }

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

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

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

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

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

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

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

  672.             s->tiling.log2_tile_cols++;

  673.         else

  674.             break;

  675.     }

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

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

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

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

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

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

  682.         if (!s->c_b) {

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

  684.             return AVERROR(ENOMEM);

  685.         }

  686.     }

  687. 

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

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

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

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

  692.                sizeof(vp9_default_coef_probs));

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

  694.                sizeof(vp9_default_coef_probs));

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

  696.                sizeof(vp9_default_coef_probs));

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

  698.                sizeof(vp9_default_coef_probs));

  699.     }

  700. 

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

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

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

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

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

  706.         return AVERROR_INVALIDDATA;

  707.     }

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

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

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

  711.         return AVERROR_INVALIDDATA;

  712.     }

  713. 

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

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

  716.     } else {

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

  718.     }

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

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

  721.     // fw update)?

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

  723. 

  724.     // txfm updates

  725.     if (s->lossless) {

  726.         s->txfmmode = TX_4X4;

  727.     } else {

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

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

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

  731. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  746.         }

  747.     }

  748. 

  749.     // coef updates

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

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

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

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

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

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

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

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

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

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

  760.                                 break;

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

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

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

  764.                                 } else {

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

  766.                                 }

  767.                             }

  768.                             p[3] = 0;

  769.                         }

  770.         } else {

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

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

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

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

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

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

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

  778.                                 break;

  779.                             memcpy(p, r, 3);

  780.                             p[3] = 0;

  781.                         }

  782.         }

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

  784.             break;

  785.     }

  786. 

  787.     // mode updates

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

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

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

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

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

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

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

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

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

  797. 

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

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

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

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

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

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

  804. 

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

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

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

  808. 

  809.         if (s->allowcompinter) {

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

  811.             if (s->comppredmode)

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

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

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

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

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

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

  818.         } else {

  819.             s->comppredmode = PRED_SINGLEREF;

  820.         }

  821. 

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

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

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

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

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

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

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

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

  830.             }

  831.         }

  832. 

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

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

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

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

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

  838.         }

  839. 

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

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

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

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

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

  845. 

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

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

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

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

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

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

  852. 

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

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

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

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

  857. 

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

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

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

  861. 

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

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

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

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

  866. 

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

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

  869. 

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

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

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

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

  874.         }

  875. 

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

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

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

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

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

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

  882. 

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

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

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

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

  887.         }

  888. 

  889.         if (s->highprecisionmvs) {

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

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

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

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

  894. 

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

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

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

  898.             }

  899.         }

  900.     }

  901. 

  902.     return (data2 - data) + size2;

  903. }

  904. 

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

  906.                                       VP9Context *s)

  907. {

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

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

  910. }

  911. 

  912. static void find_ref_mvs(VP9Context *s,

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

  914. {

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

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

  917.                       { -1, -1 }, {  0, -1 }, { -1,  0 }, {  6, -1 }},

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

  919.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, {  2, -1 }},

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

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

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

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

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

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

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

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

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

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

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

  931.                       {  0, -2 }, { -2,  0 }, { -2, -1 }, { -1, -2 }},

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

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

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

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

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

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

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

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

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

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

  942.     };

  943.     VP9Block *b = s->b;

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

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

  946. #define INVALID_MV 0x80008000U

  947.     uint32_t mem = INVALID_MV;

  948.     int i;

  949. 

  950. #define RETURN_DIRECT_MV(mv) \

  951.     do { \

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

  953.         if (!idx) { \

  954.             AV_WN32A(pmv, m); \

  955.             return; \

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

  957.             mem = m; \

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

  959.             AV_WN32A(pmv, m); \

  960.             return; \

  961.         } \

  962.     } while (0)

  963. 

  964.     if (sb >= 0) {

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

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

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

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

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

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

  971.         }

  972. 

  973. #define RETURN_MV(mv) \

  974.     do { \

  975.         if (sb > 0) { \

  976.             VP56mv tmp; \

  977.             uint32_t m; \

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

  979.             m = AV_RN32A(&tmp); \

  980.             if (!idx) { \

  981.                 AV_WN32A(pmv, m); \

  982.                 return; \

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

  984.                 mem = m; \

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

  986.                 AV_WN32A(pmv, m); \

  987.                 return; \

  988.             } \

  989.         } else { \

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

  991.             if (!idx) { \

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

  993.                 return; \

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

  995.                 mem = m; \

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

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

  998.                 return; \

  999.             } \

  1000.         } \

  1001.     } while (0)

  1002. 

  1003.         if (row > 0) {

  1004.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[(row - 1) * s->sb_cols * 8 + col];

  1005.             if (mv->ref[0] == ref) {

  1006.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][0]);

  1007.             } else if (mv->ref[1] == ref) {

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

  1009.             }

  1010.         }

  1011.         if (col > s->tiling.tile_col_start) {

  1012.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[row * s->sb_cols * 8 + col - 1];

  1013.             if (mv->ref[0] == ref) {

  1014.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][0]);

  1015.             } else if (mv->ref[1] == ref) {

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

  1017.             }

  1018.         }

  1019.         i = 2;

  1020.     } else {

  1021.         i = 0;

  1022.     }

  1023. 

  1024.     // previously coded MVs in this neighbourhood, using same reference frame

  1025.     for (; i < 8; i++) {

  1026.         int c = p[i][0] + col, r = p[i][1] + row;

  1027. 

  1028.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1029.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1030. 

  1031.             if (mv->ref[0] == ref) {

  1032.                 RETURN_MV(mv->mv[0]);

  1033.             } else if (mv->ref[1] == ref) {

  1034.                 RETURN_MV(mv->mv[1]);

  1035.             }

  1036.         }

  1037.     }

  1038. 

  1039.     // MV at this position in previous frame, using same reference frame

  1040.     if (s->use_last_frame_mvs) {

  1041.         struct VP9mvrefPair *mv = &s->frames[LAST_FRAME].mv[row * s->sb_cols * 8 + col];

  1042. 

  1043.         if (!s->last_uses_2pass)

  1044.             ff_thread_await_progress(&s->frames[LAST_FRAME].tf, row >> 3, 0);

  1045.         if (mv->ref[0] == ref) {

  1046.             RETURN_MV(mv->mv[0]);

  1047.         } else if (mv->ref[1] == ref) {

  1048.             RETURN_MV(mv->mv[1]);

  1049.         }

  1050.     }

  1051. 

  1052. #define RETURN_SCALE_MV(mv, scale) \

  1053.     do { \

  1054.         if (scale) { \

  1055.             VP56mv mv_temp = { -mv.x, -mv.y }; \

  1056.             RETURN_MV(mv_temp); \

  1057.         } else { \

  1058.             RETURN_MV(mv); \

  1059.         } \

  1060.     } while (0)

  1061. 

  1062.     // previously coded MVs in this neighbourhood, using different reference frame

  1063.     for (i = 0; i < 8; i++) {

  1064.         int c = p[i][0] + col, r = p[i][1] + row;

  1065. 

  1066.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1067.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1068. 

  1069.             if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1070.                 RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1071.             }

  1072.             if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1073.                 // BUG - libvpx has this condition regardless of whether

  1074.                 // we used the first ref MV and pre-scaling

  1075.                 AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1076.                 RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1077.             }

  1078.         }

  1079.     }

  1080. 

  1081.     // MV at this position in previous frame, using different reference frame

  1082.     if (s->use_last_frame_mvs) {

  1083.         struct VP9mvrefPair *mv = &s->frames[LAST_FRAME].mv[row * s->sb_cols * 8 + col];

  1084. 

  1085.         // no need to await_progress, because we already did that above

  1086.         if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1087.             RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1088.         }

  1089.         if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1090.             // BUG - libvpx has this condition regardless of whether

  1091.             // we used the first ref MV and pre-scaling

  1092.             AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1093.             RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1094.         }

  1095.     }

  1096. 

  1097.     AV_ZERO32(pmv);

  1098. #undef INVALID_MV

  1099. #undef RETURN_MV

  1100. #undef RETURN_SCALE_MV

  1101. }

  1102. 

  1103. static av_always_inline int read_mv_component(VP9Context *s, int idx, int hp)

  1104. {

  1105.     int bit, sign = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].sign);

  1106.     int n, c = vp8_rac_get_tree(&s->c, vp9_mv_class_tree,

  1107.                                 s->prob.p.mv_comp[idx].classes);

  1108. 

  1109.     s->counts.mv_comp[idx].sign[sign]++;

  1110.     s->counts.mv_comp[idx].classes[c]++;

  1111.     if (c) {

  1112.         int m;

  1113. 

  1114.         for (n = 0, m = 0; m < c; m++) {

  1115.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].bits[m]);

  1116.             n |= bit << m;

  1117.             s->counts.mv_comp[idx].bits[m][bit]++;

  1118.         }

  1119.         n <<= 3;

  1120.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree, s->prob.p.mv_comp[idx].fp);

  1121.         n |= bit << 1;

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

  1123.         if (hp) {

  1124.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].hp);

  1125.             s->counts.mv_comp[idx].hp[bit]++;

  1126.             n |= bit;

  1127.         } else {

  1128.             n |= 1;

  1129.             // bug in libvpx - we count for bw entropy purposes even if the

  1130.             // bit wasn't coded

  1131.             s->counts.mv_comp[idx].hp[1]++;

  1132.         }

  1133.         n += 8 << c;

  1134.     } else {

  1135.         n = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0);

  1136.         s->counts.mv_comp[idx].class0[n]++;

  1137.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree,

  1138.                                s->prob.p.mv_comp[idx].class0_fp[n]);

  1139.         s->counts.mv_comp[idx].class0_fp[n][bit]++;

  1140.         n = (n << 3) | (bit << 1);

  1141.         if (hp) {

  1142.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0_hp);

  1143.             s->counts.mv_comp[idx].class0_hp[bit]++;

  1144.             n |= bit;

  1145.         } else {

  1146.             n |= 1;

  1147.             // bug in libvpx - we count for bw entropy purposes even if the

  1148.             // bit wasn't coded

  1149.             s->counts.mv_comp[idx].class0_hp[1]++;

  1150.         }

  1151.     }

  1152. 

  1153.     return sign ? -(n + 1) : (n + 1);

  1154. }

  1155. 

  1156. static void fill_mv(VP9Context *s,

  1157.                     VP56mv *mv, int mode, int sb)

  1158. {

  1159.     VP9Block *b = s->b;

  1160. 

  1161.     if (mode == ZEROMV) {

  1162.         memset(mv, 0, sizeof(*mv) * 2);

  1163.     } else {

  1164.         int hp;

  1165. 

  1166.         // FIXME cache this value and reuse for other subblocks

  1167.         find_ref_mvs(s, &mv[0], b->ref[0], 0, mode == NEARMV,

  1168.                      mode == NEWMV ? -1 : sb);

  1169.         // FIXME maybe move this code into find_ref_mvs()

  1170.         if ((mode == NEWMV || sb == -1) &&

  1171.             !(hp = s->highprecisionmvs && abs(mv[0].x) < 64 && abs(mv[0].y) < 64)) {

  1172.             if (mv[0].y & 1) {

  1173.                 if (mv[0].y < 0)

  1174.                     mv[0].y++;

  1175.                 else

  1176.                     mv[0].y--;

  1177.             }

  1178.             if (mv[0].x & 1) {

  1179.                 if (mv[0].x < 0)

  1180.                     mv[0].x++;

  1181.                 else

  1182.                     mv[0].x--;

  1183.             }

  1184.         }

  1185.         if (mode == NEWMV) {

  1186.             enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1187.                                               s->prob.p.mv_joint);

  1188. 

  1189.             s->counts.mv_joint[j]++;

  1190.             if (j >= MV_JOINT_V)

  1191.                 mv[0].y += read_mv_component(s, 0, hp);

  1192.             if (j & 1)

  1193.                 mv[0].x += read_mv_component(s, 1, hp);

  1194.         }

  1195. 

  1196.         if (b->comp) {

  1197.             // FIXME cache this value and reuse for other subblocks

  1198.             find_ref_mvs(s, &mv[1], b->ref[1], 1, mode == NEARMV,

  1199.                          mode == NEWMV ? -1 : sb);

  1200.             if ((mode == NEWMV || sb == -1) &&

  1201.                 !(hp = s->highprecisionmvs && abs(mv[1].x) < 64 && abs(mv[1].y) < 64)) {

  1202.                 if (mv[1].y & 1) {

  1203.                     if (mv[1].y < 0)

  1204.                         mv[1].y++;

  1205.                     else

  1206.                         mv[1].y--;

  1207.                 }

  1208.                 if (mv[1].x & 1) {

  1209.                     if (mv[1].x < 0)

  1210.                         mv[1].x++;

  1211.                     else

  1212.                         mv[1].x--;

  1213.                 }

  1214.             }

  1215.             if (mode == NEWMV) {

  1216.                 enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1217.                                                   s->prob.p.mv_joint);

  1218. 

  1219.                 s->counts.mv_joint[j]++;

  1220.                 if (j >= MV_JOINT_V)

  1221.                     mv[1].y += read_mv_component(s, 0, hp);

  1222.                 if (j & 1)

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

  1224.             }

  1225.         }

  1226.     }

  1227. }

  1228. 

  1229. static void decode_mode(AVCodecContext *ctx)

  1230. {

  1231.     static const uint8_t left_ctx[N_BS_SIZES] = {

  1232.         0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf

  1233.     };

  1234.     static const uint8_t above_ctx[N_BS_SIZES] = {

  1235.         0x0, 0x0, 0x8, 0x8, 0x8, 0xc, 0xc, 0xc, 0xe, 0xe, 0xe, 0xf, 0xf

  1236.     };

  1237.     static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {

  1238.         TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,

  1239.         TX_16X16, TX_8X8, TX_8X8, TX_8X8, TX_4X4, TX_4X4, TX_4X4

  1240.     };

  1241.     VP9Context *s = ctx->priv_data;

  1242.     VP9Block *b = s->b;

  1243.     int row = s->row, col = s->col, row7 = s->row7;

  1244.     enum TxfmMode max_tx = max_tx_for_bl_bp[b->bs];

  1245.     int w4 = FFMIN(s->cols - col, bwh_tab[1][b->bs][0]);

  1246.     int h4 = FFMIN(s->rows - row, bwh_tab[1][b->bs][1]), y;

  1247.     int have_a = row > 0, have_l = col > s->tiling.tile_col_start;

  1248. 

  1249.     if (!s->segmentation.enabled) {

  1250.         b->seg_id = 0;

  1251.     } else if (s->keyframe || s->intraonly) {

  1252.         b->seg_id = s->segmentation.update_map ?

  1253.             vp8_rac_get_tree(&s->c, vp9_segmentation_tree, s->prob.seg) : 0;

  1254.     } else if (!s->segmentation.update_map ||

  1255.                (s->segmentation.temporal &&

  1256.                 vp56_rac_get_prob_branchy(&s->c,

  1257.                     s->prob.segpred[s->above_segpred_ctx[col] +

  1258.                                     s->left_segpred_ctx[row7]]))) {

  1259.         int pred = 8, x;

  1260.         uint8_t *refsegmap = s->frames[LAST_FRAME].segmentation_map;

  1261. 

  1262.         if (!s->last_uses_2pass)

  1263.             ff_thread_await_progress(&s->frames[LAST_FRAME].tf, row >> 3, 0);

  1264.         for (y = 0; y < h4; y++)

  1265.             for (x = 0; x < w4; x++)

  1266.                 pred = FFMIN(pred, refsegmap[(y + row) * 8 * s->sb_cols + x + col]);

  1267.         av_assert1(pred < 8);

  1268.         b->seg_id = pred;

  1269. 

  1270.         memset(&s->above_segpred_ctx[col], 1, w4);

  1271.         memset(&s->left_segpred_ctx[row7], 1, h4);

  1272.     } else {

  1273.         b->seg_id = vp8_rac_get_tree(&s->c, vp9_segmentation_tree,

  1274.                                      s->prob.seg);

  1275. 

  1276.         memset(&s->above_segpred_ctx[col], 0, w4);

  1277.         memset(&s->left_segpred_ctx[row7], 0, h4);

  1278.     }

  1279.     if ((s->segmentation.enabled && s->segmentation.update_map) || s->keyframe) {

  1280.         uint8_t *segmap = s->frames[CUR_FRAME].segmentation_map;

  1281. 

  1282.         for (y = 0; y < h4; y++)

  1283.             memset(&segmap[(y + row) * 8 * s->sb_cols + col], b->seg_id, w4);

  1284.     }

  1285. 

  1286.     b->skip = s->segmentation.enabled &&

  1287.         s->segmentation.feat[b->seg_id].skip_enabled;

  1288.     if (!b->skip) {

  1289.         int c = s->left_skip_ctx[row7] + s->above_skip_ctx[col];

  1290.         b->skip = vp56_rac_get_prob(&s->c, s->prob.p.skip[c]);

  1291.         s->counts.skip[c][b->skip]++;

  1292.     }

  1293. 

  1294.     if (s->keyframe || s->intraonly) {

  1295.         b->intra = 1;

  1296.     } else if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1297.         b->intra = !s->segmentation.feat[b->seg_id].ref_val;

  1298.     } else {

  1299.         int c, bit;

  1300. 

  1301.         if (have_a && have_l) {

  1302.             c = s->above_intra_ctx[col] + s->left_intra_ctx[row7];

  1303.             c += (c == 2);

  1304.         } else {

  1305.             c = have_a ? 2 * s->above_intra_ctx[col] :

  1306.                 have_l ? 2 * s->left_intra_ctx[row7] : 0;

  1307.         }

  1308.         bit = vp56_rac_get_prob(&s->c, s->prob.p.intra[c]);

  1309.         s->counts.intra[c][bit]++;

  1310.         b->intra = !bit;

  1311.     }

  1312. 

  1313.     if ((b->intra || !b->skip) && s->txfmmode == TX_SWITCHABLE) {

  1314.         int c;

  1315.         if (have_a) {

  1316.             if (have_l) {

  1317.                 c = (s->above_skip_ctx[col] ? max_tx :

  1318.                      s->above_txfm_ctx[col]) +

  1319.                     (s->left_skip_ctx[row7] ? max_tx :

  1320.                      s->left_txfm_ctx[row7]) > max_tx;

  1321.             } else {

  1322.                 c = s->above_skip_ctx[col] ? 1 :

  1323.                     (s->above_txfm_ctx[col] * 2 > max_tx);

  1324.             }

  1325.         } else if (have_l) {

  1326.             c = s->left_skip_ctx[row7] ? 1 :

  1327.                 (s->left_txfm_ctx[row7] * 2 > max_tx);

  1328.         } else {

  1329.             c = 1;

  1330.         }

  1331.         switch (max_tx) {

  1332.         case TX_32X32:

  1333.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][0]);

  1334.             if (b->tx) {

  1335.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][1]);

  1336.                 if (b->tx == 2)

  1337.                     b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][2]);

  1338.             }

  1339.             s->counts.tx32p[c][b->tx]++;

  1340.             break;

  1341.         case TX_16X16:

  1342.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][0]);

  1343.             if (b->tx)

  1344.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][1]);

  1345.             s->counts.tx16p[c][b->tx]++;

  1346.             break;

  1347.         case TX_8X8:

  1348.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx8p[c]);

  1349.             s->counts.tx8p[c][b->tx]++;

  1350.             break;

  1351.         case TX_4X4:

  1352.             b->tx = TX_4X4;

  1353.             break;

  1354.         }

  1355.     } else {

  1356.         b->tx = FFMIN(max_tx, s->txfmmode);

  1357.     }

  1358. 

  1359.     if (s->keyframe || s->intraonly) {

  1360.         uint8_t *a = &s->above_mode_ctx[col * 2];

  1361.         uint8_t *l = &s->left_mode_ctx[(row7) << 1];

  1362. 

  1363.         b->comp = 0;

  1364.         if (b->bs > BS_8x8) {

  1365.             // FIXME the memory storage intermediates here aren't really

  1366.             // necessary, they're just there to make the code slightly

  1367.             // simpler for now

  1368.             b->mode[0] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1369.                                     vp9_default_kf_ymode_probs[a[0]][l[0]]);

  1370.             if (b->bs != BS_8x4) {

  1371.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1372.                                  vp9_default_kf_ymode_probs[a[1]][b->mode[0]]);

  1373.                 l[0] = a[1] = b->mode[1];

  1374.             } else {

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

  1376.             }

  1377.             if (b->bs != BS_4x8) {

  1378.                 b->mode[2] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1379.                                         vp9_default_kf_ymode_probs[a[0]][l[1]]);

  1380.                 if (b->bs != BS_8x4) {

  1381.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1382.                                   vp9_default_kf_ymode_probs[a[1]][b->mode[2]]);

  1383.                     l[1] = a[1] = b->mode[3];

  1384.                 } else {

  1385.                     l[1] = a[1] = b->mode[3] = b->mode[2];

  1386.                 }

  1387.             } else {

  1388.                 b->mode[2] = b->mode[0];

  1389.                 l[1] = a[1] = b->mode[3] = b->mode[1];

  1390.             }

  1391.         } else {

  1392.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1393.                                           vp9_default_kf_ymode_probs[*a][*l]);

  1394.             b->mode[3] = b->mode[2] = b->mode[1] = b->mode[0];

  1395.             // FIXME this can probably be optimized

  1396.             memset(a, b->mode[0], bwh_tab[0][b->bs][0]);

  1397.             memset(l, b->mode[0], bwh_tab[0][b->bs][1]);

  1398.         }

  1399.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1400.                                      vp9_default_kf_uvmode_probs[b->mode[3]]);

  1401.     } else if (b->intra) {

  1402.         b->comp = 0;

  1403.         if (b->bs > BS_8x8) {

  1404.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1405.                                           s->prob.p.y_mode[0]);

  1406.             s->counts.y_mode[0][b->mode[0]]++;

  1407.             if (b->bs != BS_8x4) {

  1408.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1409.                                               s->prob.p.y_mode[0]);

  1410.                 s->counts.y_mode[0][b->mode[1]]++;

  1411.             } else {

  1412.                 b->mode[1] = b->mode[0];

  1413.             }

  1414.             if (b->bs != BS_4x8) {

  1415.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1416.                                               s->prob.p.y_mode[0]);

  1417.                 s->counts.y_mode[0][b->mode[2]]++;

  1418.                 if (b->bs != BS_8x4) {

  1419.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1420.                                                   s->prob.p.y_mode[0]);

  1421.                     s->counts.y_mode[0][b->mode[3]]++;

  1422.                 } else {

  1423.                     b->mode[3] = b->mode[2];

  1424.                 }

  1425.             } else {

  1426.                 b->mode[2] = b->mode[0];

  1427.                 b->mode[3] = b->mode[1];

  1428.             }

  1429.         } else {

  1430.             static const uint8_t size_group[10] = {

  1431.                 3, 3, 3, 3, 2, 2, 2, 1, 1, 1

  1432.             };

  1433.             int sz = size_group[b->bs];

  1434. 

  1435.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1436.                                           s->prob.p.y_mode[sz]);

  1437.             b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1438.             s->counts.y_mode[sz][b->mode[3]]++;

  1439.         }

  1440.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1441.                                      s->prob.p.uv_mode[b->mode[3]]);

  1442.         s->counts.uv_mode[b->mode[3]][b->uvmode]++;

  1443.     } else {

  1444.         static const uint8_t inter_mode_ctx_lut[14][14] = {

  1445.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1446.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

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

  1456.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1457.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },

  1458.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },

  1459.         };

  1460. 

  1461.         if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1462.             av_assert2(s->segmentation.feat[b->seg_id].ref_val != 0);

  1463.             b->comp = 0;

  1464.             b->ref[0] = s->segmentation.feat[b->seg_id].ref_val - 1;

  1465.         } else {

  1466.             // read comp_pred flag

  1467.             if (s->comppredmode != PRED_SWITCHABLE) {

  1468.                 b->comp = s->comppredmode == PRED_COMPREF;

  1469.             } else {

  1470.                 int c;

  1471. 

  1472.                 // FIXME add intra as ref=0xff (or -1) to make these easier?

  1473.                 if (have_a) {

  1474.                     if (have_l) {

  1475.                         if (s->above_comp_ctx[col] && s->left_comp_ctx[row7]) {

  1476.                             c = 4;

  1477.                         } else if (s->above_comp_ctx[col]) {

  1478.                             c = 2 + (s->left_intra_ctx[row7] ||

  1479.                                      s->left_ref_ctx[row7] == s->fixcompref);

  1480.                         } else if (s->left_comp_ctx[row7]) {

  1481.                             c = 2 + (s->above_intra_ctx[col] ||

  1482.                                      s->above_ref_ctx[col] == s->fixcompref);

  1483.                         } else {

  1484.                             c = (!s->above_intra_ctx[col] &&

  1485.                                  s->above_ref_ctx[col] == s->fixcompref) ^

  1486.                             (!s->left_intra_ctx[row7] &&

  1487.                              s->left_ref_ctx[row & 7] == s->fixcompref);

  1488.                         }

  1489.                     } else {

  1490.                         c = s->above_comp_ctx[col] ? 3 :

  1491.                         (!s->above_intra_ctx[col] && s->above_ref_ctx[col] == s->fixcompref);

  1492.                     }

  1493.                 } else if (have_l) {

  1494.                     c = s->left_comp_ctx[row7] ? 3 :

  1495.                     (!s->left_intra_ctx[row7] && s->left_ref_ctx[row7] == s->fixcompref);

  1496.                 } else {

  1497.                     c = 1;

  1498.                 }

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

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

  1501.             }

  1502. 

  1503.             // read actual references

  1504.             // FIXME probably cache a few variables here to prevent repetitive

  1505.             // memory accesses below

  1506.             if (b->comp) /* two references */ {

  1507.                 int fix_idx = s->signbias[s->fixcompref], var_idx = !fix_idx, c, bit;

  1508. 

  1509.                 b->ref[fix_idx] = s->fixcompref;

  1510.                 // FIXME can this codeblob be replaced by some sort of LUT?

  1511.                 if (have_a) {

  1512.                     if (have_l) {

  1513.                         if (s->above_intra_ctx[col]) {

  1514.                             if (s->left_intra_ctx[row7]) {

  1515.                                 c = 2;

  1516.                             } else {

  1517.                                 c = 1 + 2 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1518.                             }

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

  1520.                             c = 1 + 2 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1521.                         } else {

  1522.                             int refl = s->left_ref_ctx[row7], refa = s->above_ref_ctx[col];

  1523. 

  1524.                             if (refl == refa && refa == s->varcompref[1]) {

  1525.                                 c = 0;

  1526.                             } else if (!s->left_comp_ctx[row7] && !s->above_comp_ctx[col]) {

  1527.                                 if ((refa == s->fixcompref && refl == s->varcompref[0]) ||

  1528.                                     (refl == s->fixcompref && refa == s->varcompref[0])) {

  1529.                                     c = 4;

  1530.                                 } else {

  1531.                                     c = (refa == refl) ? 3 : 1;

  1532.                                 }

  1533.                             } else if (!s->left_comp_ctx[row7]) {

  1534.                                 if (refa == s->varcompref[1] && refl != s->varcompref[1]) {

  1535.                                     c = 1;

  1536.                                 } else {

  1537.                                     c = (refl == s->varcompref[1] &&

  1538.                                          refa != s->varcompref[1]) ? 2 : 4;

  1539.                                 }

  1540.                             } else if (!s->above_comp_ctx[col]) {

  1541.                                 if (refl == s->varcompref[1] && refa != s->varcompref[1]) {

  1542.                                     c = 1;

  1543.                                 } else {

  1544.                                     c = (refa == s->varcompref[1] &&

  1545.                                          refl != s->varcompref[1]) ? 2 : 4;

  1546.                                 }

  1547.                             } else {

  1548.                                 c = (refl == refa) ? 4 : 2;

  1549.                             }

  1550.                         }

  1551.                     } else {

  1552.                         if (s->above_intra_ctx[col]) {

  1553.                             c = 2;

  1554.                         } else if (s->above_comp_ctx[col]) {

  1555.                             c = 4 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1556.                         } else {

  1557.                             c = 3 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1558.                         }

  1559.                     }

  1560.                 } else if (have_l) {

  1561.                     if (s->left_intra_ctx[row7]) {

  1562.                         c = 2;

  1563.                     } else if (s->left_comp_ctx[row7]) {

  1564.                         c = 4 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1565.                     } else {

  1566.                         c = 3 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1567.                     }

  1568.                 } else {

  1569.                     c = 2;

  1570.                 }

  1571.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.comp_ref[c]);

  1572.                 b->ref[var_idx] = s->varcompref[bit];

  1573.                 s->counts.comp_ref[c][bit]++;

  1574.             } else /* single reference */ {

  1575.                 int bit, c;

  1576. 

  1577.                 if (have_a && !s->above_intra_ctx[col]) {

  1578.                     if (have_l && !s->left_intra_ctx[row7]) {

  1579.                         if (s->left_comp_ctx[row7]) {

  1580.                             if (s->above_comp_ctx[col]) {

  1581.                                 c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7] ||

  1582.                                          !s->above_ref_ctx[col]);

  1583.                             } else {

  1584.                                 c = (3 * !s->above_ref_ctx[col]) +

  1585.                                     (!s->fixcompref || !s->left_ref_ctx[row7]);

  1586.                             }

  1587.                         } else if (s->above_comp_ctx[col]) {

  1588.                             c = (3 * !s->left_ref_ctx[row7]) +

  1589.                                 (!s->fixcompref || !s->above_ref_ctx[col]);

  1590.                         } else {

  1591.                             c = 2 * !s->left_ref_ctx[row7] + 2 * !s->above_ref_ctx[col];

  1592.                         }

  1593.                     } else if (s->above_intra_ctx[col]) {

  1594.                         c = 2;

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

  1596.                         c = 1 + (!s->fixcompref || !s->above_ref_ctx[col]);

  1597.                     } else {

  1598.                         c = 4 * (!s->above_ref_ctx[col]);

  1599.                     }

  1600.                 } else if (have_l && !s->left_intra_ctx[row7]) {

  1601.                     if (s->left_intra_ctx[row7]) {

  1602.                         c = 2;

  1603.                     } else if (s->left_comp_ctx[row7]) {

  1604.                         c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7]);

  1605.                     } else {

  1606.                         c = 4 * (!s->left_ref_ctx[row7]);

  1607.                     }

  1608.                 } else {

  1609.                     c = 2;

  1610.                 }

  1611.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][0]);

  1612.                 s->counts.single_ref[c][0][bit]++;

  1613.                 if (!bit) {

  1614.                     b->ref[0] = 0;

  1615.                 } else {

  1616.                     // FIXME can this codeblob be replaced by some sort of LUT?

  1617.                     if (have_a) {

  1618.                         if (have_l) {

  1619.                             if (s->left_intra_ctx[row7]) {

  1620.                                 if (s->above_intra_ctx[col]) {

  1621.                                     c = 2;

  1622.                                 } else if (s->above_comp_ctx[col]) {

  1623.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1624.                                                  s->above_ref_ctx[col] == 1);

  1625.                                 } else if (!s->above_ref_ctx[col]) {

  1626.                                     c = 3;

  1627.                                 } else {

  1628.                                     c = 4 * (s->above_ref_ctx[col] == 1);

  1629.                                 }

  1630.                             } else if (s->above_intra_ctx[col]) {

  1631.                                 if (s->left_intra_ctx[row7]) {

  1632.                                     c = 2;

  1633.                                 } else if (s->left_comp_ctx[row7]) {

  1634.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1635.                                                  s->left_ref_ctx[row7] == 1);

  1636.                                 } else if (!s->left_ref_ctx[row7]) {

  1637.                                     c = 3;

  1638.                                 } else {

  1639.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1640.                                 }

  1641.                             } else if (s->above_comp_ctx[col]) {

  1642.                                 if (s->left_comp_ctx[row7]) {

  1643.                                     if (s->left_ref_ctx[row7] == s->above_ref_ctx[col]) {

  1644.                                         c = 3 * (s->fixcompref == 1 ||

  1645.                                                  s->left_ref_ctx[row7] == 1);

  1646.                                     } else {

  1647.                                         c = 2;

  1648.                                     }

  1649.                                 } else if (!s->left_ref_ctx[row7]) {

  1650.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1651.                                                  s->above_ref_ctx[col] == 1);

  1652.                                 } else {

  1653.                                     c = 3 * (s->left_ref_ctx[row7] == 1) +

  1654.                                     (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1655.                                 }

  1656.                             } else if (s->left_comp_ctx[row7]) {

  1657.                                 if (!s->above_ref_ctx[col]) {

  1658.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1659.                                                  s->left_ref_ctx[row7] == 1);

  1660.                                 } else {

  1661.                                     c = 3 * (s->above_ref_ctx[col] == 1) +

  1662.                                     (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1663.                                 }

  1664.                             } else if (!s->above_ref_ctx[col]) {

  1665.                                 if (!s->left_ref_ctx[row7]) {

  1666.                                     c = 3;

  1667.                                 } else {

  1668.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1669.                                 }

  1670.                             } else if (!s->left_ref_ctx[row7]) {

  1671.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1672.                             } else {

  1673.                                 c = 2 * (s->left_ref_ctx[row7] == 1) +

  1674.                                 2 * (s->above_ref_ctx[col] == 1);

  1675.                             }

  1676.                         } else {

  1677.                             if (s->above_intra_ctx[col] ||

  1678.                                 (!s->above_comp_ctx[col] && !s->above_ref_ctx[col])) {

  1679.                                 c = 2;

  1680.                             } else if (s->above_comp_ctx[col]) {

  1681.                                 c = 3 * (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1682.                             } else {

  1683.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1684.                             }

  1685.                         }

  1686.                     } else if (have_l) {

  1687.                         if (s->left_intra_ctx[row7] ||

  1688.                             (!s->left_comp_ctx[row7] && !s->left_ref_ctx[row7])) {

  1689.                             c = 2;

  1690.                         } else if (s->left_comp_ctx[row7]) {

  1691.                             c = 3 * (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1692.                         } else {

  1693.                             c = 4 * (s->left_ref_ctx[row7] == 1);

  1694.                         }

  1695.                     } else {

  1696.                         c = 2;

  1697.                     }

  1698.                     bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][1]);

  1699.                     s->counts.single_ref[c][1][bit]++;

  1700.                     b->ref[0] = 1 + bit;

  1701.                 }

  1702.             }

  1703.         }

  1704. 

  1705.         if (b->bs <= BS_8x8) {

  1706.             if (s->segmentation.feat[b->seg_id].skip_enabled) {

  1707.                 b->mode[0] = b->mode[1] = b->mode[2] = b->mode[3] = ZEROMV;

  1708.             } else {

  1709.                 static const uint8_t off[10] = {

  1710.                     3, 0, 0, 1, 0, 0, 0, 0, 0, 0

  1711.                 };

  1712. 

  1713.                 // FIXME this needs to use the LUT tables from find_ref_mvs

  1714.                 // because not all are -1,0/0,-1

  1715.                 int c = inter_mode_ctx_lut[s->above_mode_ctx[col + off[b->bs]]]

  1716.                                           [s->left_mode_ctx[row7 + off[b->bs]]];

  1717. 

  1718.                 b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1719.                                               s->prob.p.mv_mode[c]);

  1720.                 b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1721.                 s->counts.mv_mode[c][b->mode[0] - 10]++;

  1722.             }

  1723.         }

  1724. 

  1725.         if (s->filtermode == FILTER_SWITCHABLE) {

  1726.             int c;

  1727. 

  1728.             if (have_a && s->above_mode_ctx[col] >= NEARESTMV) {

  1729.                 if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1730.                     c = s->above_filter_ctx[col] == s->left_filter_ctx[row7] ?

  1731.                         s->left_filter_ctx[row7] : 3;

  1732.                 } else {

  1733.                     c = s->above_filter_ctx[col];

  1734.                 }

  1735.             } else if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1736.                 c = s->left_filter_ctx[row7];

  1737.             } else {

  1738.                 c = 3;

  1739.             }

  1740. 

  1741.             b->filter = vp8_rac_get_tree(&s->c, vp9_filter_tree,

  1742.                                          s->prob.p.filter[c]);

  1743.             s->counts.filter[c][b->filter]++;

  1744.         } else {

  1745.             b->filter = s->filtermode;

  1746.         }

  1747. 

  1748.         if (b->bs > BS_8x8) {

  1749.             int c = inter_mode_ctx_lut[s->above_mode_ctx[col]][s->left_mode_ctx[row7]];

  1750. 

  1751.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1752.                                           s->prob.p.mv_mode[c]);

  1753.             s->counts.mv_mode[c][b->mode[0] - 10]++;

  1754.             fill_mv(s, b->mv[0], b->mode[0], 0);

  1755. 

  1756.             if (b->bs != BS_8x4) {

  1757.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1758.                                               s->prob.p.mv_mode[c]);

  1759.                 s->counts.mv_mode[c][b->mode[1] - 10]++;

  1760.                 fill_mv(s, b->mv[1], b->mode[1], 1);

  1761.             } else {

  1762.                 b->mode[1] = b->mode[0];

  1763.                 AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1764.                 AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1765.             }

  1766. 

  1767.             if (b->bs != BS_4x8) {

  1768.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1769.                                               s->prob.p.mv_mode[c]);

  1770.                 s->counts.mv_mode[c][b->mode[2] - 10]++;

  1771.                 fill_mv(s, b->mv[2], b->mode[2], 2);

  1772. 

  1773.                 if (b->bs != BS_8x4) {

  1774.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1775.                                                   s->prob.p.mv_mode[c]);

  1776.                     s->counts.mv_mode[c][b->mode[3] - 10]++;

  1777.                     fill_mv(s, b->mv[3], b->mode[3], 3);

  1778.                 } else {

  1779.                     b->mode[3] = b->mode[2];

  1780.                     AV_COPY32(&b->mv[3][0], &b->mv[2][0]);

  1781.                     AV_COPY32(&b->mv[3][1], &b->mv[2][1]);

  1782.                 }

  1783.             } else {

  1784.                 b->mode[2] = b->mode[0];

  1785.                 AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1786.                 AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1787.                 b->mode[3] = b->mode[1];

  1788.                 AV_COPY32(&b->mv[3][0], &b->mv[1][0]);

  1789.                 AV_COPY32(&b->mv[3][1], &b->mv[1][1]);

  1790.             }

  1791.         } else {

  1792.             fill_mv(s, b->mv[0], b->mode[0], -1);

  1793.             AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1794.             AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1795.             AV_COPY32(&b->mv[3][0], &b->mv[0][0]);

  1796.             AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1797.             AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1798.             AV_COPY32(&b->mv[3][1], &b->mv[0][1]);

  1799.         }

  1800.     }

  1801. 

  1802.     // FIXME this can probably be optimized

  1803.     memset(&s->above_skip_ctx[col], b->skip, w4);

  1804.     memset(&s->left_skip_ctx[row7], b->skip, h4);

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

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

  1807.     memset(&s->above_partition_ctx[col], above_ctx[b->bs], w4);

  1808.     memset(&s->left_partition_ctx[row7], left_ctx[b->bs], h4);

  1809.     if (!s->keyframe && !s->intraonly) {

  1810.         memset(&s->above_intra_ctx[col], b->intra, w4);

  1811.         memset(&s->left_intra_ctx[row7], b->intra, h4);

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

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

  1814.         memset(&s->above_mode_ctx[col], b->mode[3], w4);

  1815.         memset(&s->left_mode_ctx[row7], b->mode[3], h4);

  1816.         if (s->filtermode == FILTER_SWITCHABLE && !b->intra ) {

  1817.             memset(&s->above_filter_ctx[col], b->filter, w4);

  1818.             memset(&s->left_filter_ctx[row7], b->filter, h4);

  1819.             b->filter = vp9_filter_lut[b->filter];

  1820.         }

  1821.         if (b->bs > BS_8x8) {

  1822.             int mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  1823. 

  1824.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][0], &b->mv[1][0]);

  1825.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][1], &b->mv[1][1]);

  1826.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][0], mv0);

  1827.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][1], mv1);

  1828.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][0], &b->mv[2][0]);

  1829.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][1], &b->mv[2][1]);

  1830.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][0], mv0);

  1831.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][1], mv1);

  1832.         } else {

  1833.             int n, mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  1834. 

  1835.             for (n = 0; n < w4 * 2; n++) {

  1836.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][0], mv0);

  1837.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][1], mv1);

  1838.             }

  1839.             for (n = 0; n < h4 * 2; n++) {

  1840.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][0], mv0);

  1841.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][1], mv1);

  1842.             }

  1843.         }

  1844. 

  1845.         if (!b->intra) { // FIXME write 0xff or -1 if intra, so we can use this

  1846.                          // as a direct check in above branches

  1847.             int vref = b->ref[b->comp ? s->signbias[s->varcompref[0]] : 0];

  1848. 

  1849.             memset(&s->above_ref_ctx[col], vref, w4);

  1850.             memset(&s->left_ref_ctx[row7], vref, h4);

  1851.         }

  1852.     }

  1853. 

  1854.     // FIXME kinda ugly

  1855.     for (y = 0; y < h4; y++) {

  1856.         int x, o = (row + y) * s->sb_cols * 8 + col;

  1857.         struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[o];

  1858. 

  1859.         if (b->intra) {

  1860.             for (x = 0; x < w4; x++) {

  1861.                 mv[x].ref[0] =

  1862.                 mv[x].ref[1] = -1;

  1863.             }

  1864.         } else if (b->comp) {

  1865.             for (x = 0; x < w4; x++) {

  1866.                 mv[x].ref[0] = b->ref[0];

  1867.                 mv[x].ref[1] = b->ref[1];

  1868.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  1869.                 AV_COPY32(&mv[x].mv[1], &b->mv[3][1]);

  1870.             }

  1871.         } else {

  1872.             for (x = 0; x < w4; x++) {

  1873.                 mv[x].ref[0] = b->ref[0];

  1874.                 mv[x].ref[1] = -1;

  1875.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  1876.             }

  1877.         }

  1878.     }

  1879. }

  1880. 

  1881. // FIXME remove tx argument, and merge cnt/eob arguments?

  1882. static int decode_coeffs_b(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  1883.                            enum TxfmMode tx, unsigned (*cnt)[6][3],

  1884.                            unsigned (*eob)[6][2], uint8_t (*p)[6][11],

  1885.                            int nnz, const int16_t *scan, const int16_t (*nb)[2],

  1886.                            const int16_t *band_counts, const int16_t *qmul)

  1887. {

  1888.     int i = 0, band = 0, band_left = band_counts[band];

  1889.     uint8_t *tp = p[0][nnz];

  1890.     uint8_t cache[1024];

  1891. 

  1892.     do {

  1893.         int val, rc;

  1894. 

  1895.         val = vp56_rac_get_prob_branchy(c, tp[0]); // eob

  1896.         eob[band][nnz][val]++;

  1897.         if (!val)

  1898.             break;

  1899. 

  1900.     skip_eob:

  1901.         if (!vp56_rac_get_prob_branchy(c, tp[1])) { // zero

  1902.             cnt[band][nnz][0]++;

  1903.             if (!--band_left)

  1904.                 band_left = band_counts[++band];

  1905.             cache[scan[i]] = 0;

  1906.             nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  1907.             tp = p[band][nnz];

  1908.             if (++i == n_coeffs)

  1909.                 break; //invalid input; blocks should end with EOB

  1910.             goto skip_eob;

  1911.         }

  1912. 

  1913.         rc = scan[i];

  1914.         if (!vp56_rac_get_prob_branchy(c, tp[2])) { // one

  1915.             cnt[band][nnz][1]++;

  1916.             val = 1;

  1917.             cache[rc] = 1;

  1918.         } else {

  1919.             // fill in p[3-10] (model fill) - only once per frame for each pos

  1920.             if (!tp[3])

  1921.                 memcpy(&tp[3], vp9_model_pareto8[tp[2]], 8);

  1922. 

  1923.             cnt[band][nnz][2]++;

  1924.             if (!vp56_rac_get_prob_branchy(c, tp[3])) { // 2, 3, 4

  1925.                 if (!vp56_rac_get_prob_branchy(c, tp[4])) {

  1926.                     cache[rc] = val = 2;

  1927.                 } else {

  1928.                     val = 3 + vp56_rac_get_prob(c, tp[5]);

  1929.                     cache[rc] = 3;

  1930.                 }

  1931.             } else if (!vp56_rac_get_prob_branchy(c, tp[6])) { // cat1/2

  1932.                 cache[rc] = 4;

  1933.                 if (!vp56_rac_get_prob_branchy(c, tp[7])) {

  1934.                     val = 5 + vp56_rac_get_prob(c, 159);

  1935.                 } else {

  1936.                     val  = 7 + (vp56_rac_get_prob(c, 165) << 1);

  1937.                     val +=      vp56_rac_get_prob(c, 145);

  1938.                 }

  1939.             } else { // cat 3-6

  1940.                 cache[rc] = 5;

  1941.                 if (!vp56_rac_get_prob_branchy(c, tp[8])) {

  1942.                     if (!vp56_rac_get_prob_branchy(c, tp[9])) {

  1943.                         val  = 11 + (vp56_rac_get_prob(c, 173) << 2);

  1944.                         val +=      (vp56_rac_get_prob(c, 148) << 1);

  1945.                         val +=       vp56_rac_get_prob(c, 140);

  1946.                     } else {

  1947.                         val  = 19 + (vp56_rac_get_prob(c, 176) << 3);

  1948.                         val +=      (vp56_rac_get_prob(c, 155) << 2);

  1949.                         val +=      (vp56_rac_get_prob(c, 140) << 1);

  1950.                         val +=       vp56_rac_get_prob(c, 135);

  1951.                     }

  1952.                 } else if (!vp56_rac_get_prob_branchy(c, tp[10])) {

  1953.                     val  = 35 + (vp56_rac_get_prob(c, 180) << 4);

  1954.                     val +=      (vp56_rac_get_prob(c, 157) << 3);

  1955.                     val +=      (vp56_rac_get_prob(c, 141) << 2);

  1956.                     val +=      (vp56_rac_get_prob(c, 134) << 1);

  1957.                     val +=       vp56_rac_get_prob(c, 130);

  1958.                 } else {

  1959.                     val  = 67 + (vp56_rac_get_prob(c, 254) << 13);

  1960.                     val +=      (vp56_rac_get_prob(c, 254) << 12);

  1961.                     val +=      (vp56_rac_get_prob(c, 254) << 11);

  1962.                     val +=      (vp56_rac_get_prob(c, 252) << 10);

  1963.                     val +=      (vp56_rac_get_prob(c, 249) << 9);

  1964.                     val +=      (vp56_rac_get_prob(c, 243) << 8);

  1965.                     val +=      (vp56_rac_get_prob(c, 230) << 7);

  1966.                     val +=      (vp56_rac_get_prob(c, 196) << 6);

  1967.                     val +=      (vp56_rac_get_prob(c, 177) << 5);

  1968.                     val +=      (vp56_rac_get_prob(c, 153) << 4);

  1969.                     val +=      (vp56_rac_get_prob(c, 140) << 3);

  1970.                     val +=      (vp56_rac_get_prob(c, 133) << 2);

  1971.                     val +=      (vp56_rac_get_prob(c, 130) << 1);

  1972.                     val +=       vp56_rac_get_prob(c, 129);

  1973.                 }

  1974.             }

  1975.         }

  1976.         if (!--band_left)

  1977.             band_left = band_counts[++band];

  1978.         if (tx == TX_32X32) // FIXME slow

  1979.             coef[rc] = ((vp8_rac_get(c) ? -val : val) * qmul[!!i]) / 2;

  1980.         else

  1981.             coef[rc] = (vp8_rac_get(c) ? -val : val) * qmul[!!i];

  1982.         nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  1983.         tp = p[band][nnz];

  1984.     } while (++i < n_coeffs);

  1985. 

  1986.     return i;

  1987. }

  1988. 

  1989. static void decode_coeffs(AVCodecContext *ctx)

  1990. {

  1991.     VP9Context *s = ctx->priv_data;

  1992.     VP9Block *b = s->b;

  1993.     int row = s->row, col = s->col;

  1994.     uint8_t (*p)[6][11] = s->prob.coef[b->tx][0 /* y */][!b->intra];

  1995.     unsigned (*c)[6][3] = s->counts.coef[b->tx][0 /* y */][!b->intra];

  1996.     unsigned (*e)[6][2] = s->counts.eob[b->tx][0 /* y */][!b->intra];

  1997.     int w4 = bwh_tab[1][b->bs][0] << 1, h4 = bwh_tab[1][b->bs][1] << 1;

  1998.     int end_x = FFMIN(2 * (s->cols - col), w4);

  1999.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2000.     int n, pl, x, y, step1d = 1 << b->tx, step = 1 << (b->tx * 2);

  2001.     int uvstep1d = 1 << b->uvtx, uvstep = 1 << (b->uvtx * 2), res;

  2002.     int16_t (*qmul)[2] = s->segmentation.feat[b->seg_id].qmul;

  2003.     int tx = 4 * s->lossless + b->tx;

  2004.     const int16_t * const *yscans = vp9_scans[tx];

  2005.     const int16_t (* const *ynbs)[2] = vp9_scans_nb[tx];

  2006.     const int16_t *uvscan = vp9_scans[b->uvtx][DCT_DCT];

  2007.     const int16_t (*uvnb)[2] = vp9_scans_nb[b->uvtx][DCT_DCT];

  2008.     uint8_t *a = &s->above_y_nnz_ctx[col * 2];

  2009.     uint8_t *l = &s->left_y_nnz_ctx[(row & 7) << 1];

  2010.     static const int16_t band_counts[4][8] = {

  2011.         { 1, 2, 3, 4,  3,   16 - 13 },

  2012.         { 1, 2, 3, 4, 11,   64 - 21 },

  2013.         { 1, 2, 3, 4, 11,  256 - 21 },

  2014.         { 1, 2, 3, 4, 11, 1024 - 21 },

  2015.     };

  2016.     const int16_t *y_band_counts = band_counts[b->tx];

  2017.     const int16_t *uv_band_counts = band_counts[b->uvtx];

  2018. 

  2019.     /* y tokens */

  2020.     if (b->tx > TX_4X4) { // FIXME slow

  2021.         for (y = 0; y < end_y; y += step1d)

  2022.             for (x = 1; x < step1d; x++)

  2023.                 l[y] |= l[y + x];

  2024.         for (x = 0; x < end_x; x += step1d)

  2025.             for (y = 1; y < step1d; y++)

  2026.                 a[x] |= a[x + y];

  2027.     }

  2028.     for (n = 0, y = 0; y < end_y; y += step1d) {

  2029.         for (x = 0; x < end_x; x += step1d, n += step) {

  2030.             enum TxfmType txtp = vp9_intra_txfm_type[b->mode[b->tx == TX_4X4 &&

  2031.                                                              b->bs > BS_8x8 ?

  2032.                                                              n : 0]];

  2033.             int nnz = a[x] + l[y];

  2034.             res = decode_coeffs_b(&s->c, s->block + 16 * n, 16 * step,

  2035.                                   b->tx, c, e, p, nnz, yscans[txtp],

  2036.                                   ynbs[txtp], y_band_counts, qmul[0]);

  2037.             a[x] = l[y] = !!res;

  2038.             if (b->tx > TX_8X8) {

  2039.                 AV_WN16A(&s->eob[n], res);

  2040.             } else {

  2041.                 s->eob[n] = res;

  2042.             }

  2043.         }

  2044.     }

  2045.     if (b->tx > TX_4X4) { // FIXME slow

  2046.         for (y = 0; y < end_y; y += step1d)

  2047.             memset(&l[y + 1], l[y], FFMIN(end_y - y - 1, step1d - 1));

  2048.         for (x = 0; x < end_x; x += step1d)

  2049.             memset(&a[x + 1], a[x], FFMIN(end_x - x - 1, step1d - 1));

  2050.     }

  2051. 

  2052.     p = s->prob.coef[b->uvtx][1 /* uv */][!b->intra];

  2053.     c = s->counts.coef[b->uvtx][1 /* uv */][!b->intra];

  2054.     e = s->counts.eob[b->uvtx][1 /* uv */][!b->intra];

  2055.     w4 >>= 1;

  2056.     h4 >>= 1;

  2057.     end_x >>= 1;

  2058.     end_y >>= 1;

  2059.     for (pl = 0; pl < 2; pl++) {

  2060.         a = &s->above_uv_nnz_ctx[pl][col];

  2061.         l = &s->left_uv_nnz_ctx[pl][row & 7];

  2062.         if (b->uvtx > TX_4X4) { // FIXME slow

  2063.             for (y = 0; y < end_y; y += uvstep1d)

  2064.                 for (x = 1; x < uvstep1d; x++)

  2065.                     l[y] |= l[y + x];

  2066.             for (x = 0; x < end_x; x += uvstep1d)

  2067.                 for (y = 1; y < uvstep1d; y++)

  2068.                     a[x] |= a[x + y];

  2069.         }

  2070.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2071.             for (x = 0; x < end_x; x += uvstep1d, n += uvstep) {

  2072.                 int nnz = a[x] + l[y];

  2073.                 res = decode_coeffs_b(&s->c, s->uvblock[pl] + 16 * n,

  2074.                                       16 * uvstep, b->uvtx, c, e, p, nnz,

  2075.                                       uvscan, uvnb, uv_band_counts, qmul[1]);

  2076.                 a[x] = l[y] = !!res;

  2077.                 if (b->uvtx > TX_8X8) {

  2078.                     AV_WN16A(&s->uveob[pl][n], res);

  2079.                 } else {

  2080.                     s->uveob[pl][n] = res;

  2081.                 }

  2082.             }

  2083.         }

  2084.         if (b->uvtx > TX_4X4) { // FIXME slow

  2085.             for (y = 0; y < end_y; y += uvstep1d)

  2086.                 memset(&l[y + 1], l[y], FFMIN(end_y - y - 1, uvstep1d - 1));

  2087.             for (x = 0; x < end_x; x += uvstep1d)

  2088.                 memset(&a[x + 1], a[x], FFMIN(end_x - x - 1, uvstep1d - 1));

  2089.         }

  2090.     }

  2091. }

  2092. 

  2093. static av_always_inline int check_intra_mode(VP9Context *s, int mode, uint8_t **a,

  2094.                                              uint8_t *dst_edge, ptrdiff_t stride_edge,

  2095.                                              uint8_t *dst_inner, ptrdiff_t stride_inner,

  2096.                                              uint8_t *l, int col, int x, int w,

  2097.                                              int row, int y, enum TxfmMode tx,

  2098.                                              int p)

  2099. {

  2100.     int have_top = row > 0 || y > 0;

  2101.     int have_left = col > s->tiling.tile_col_start || x > 0;

  2102.     int have_right = x < w - 1;

  2103.     static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {

  2104.         [VERT_PRED]            = { { DC_127_PRED,          VERT_PRED },

  2105.                                    { DC_127_PRED,          VERT_PRED } },

  2106.         [HOR_PRED]             = { { DC_129_PRED,          DC_129_PRED },

  2107.                                    { HOR_PRED,             HOR_PRED } },

  2108.         [DC_PRED]              = { { DC_128_PRED,          TOP_DC_PRED },

  2109.                                    { LEFT_DC_PRED,         DC_PRED } },

  2110.         [DIAG_DOWN_LEFT_PRED]  = { { DC_127_PRED,          DIAG_DOWN_LEFT_PRED },

  2111.                                    { DC_127_PRED,          DIAG_DOWN_LEFT_PRED } },

  2112.         [DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },

  2113.                                    { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },

  2114.         [VERT_RIGHT_PRED]      = { { VERT_RIGHT_PRED,      VERT_RIGHT_PRED },

  2115.                                    { VERT_RIGHT_PRED,      VERT_RIGHT_PRED } },

  2116.         [HOR_DOWN_PRED]        = { { HOR_DOWN_PRED,        HOR_DOWN_PRED },

  2117.                                    { HOR_DOWN_PRED,        HOR_DOWN_PRED } },

  2118.         [VERT_LEFT_PRED]       = { { DC_127_PRED,          VERT_LEFT_PRED },

  2119.                                    { DC_127_PRED,          VERT_LEFT_PRED } },

  2120.         [HOR_UP_PRED]          = { { DC_129_PRED,          DC_129_PRED },

  2121.                                    { HOR_UP_PRED,          HOR_UP_PRED } },

  2122.         [TM_VP8_PRED]          = { { DC_129_PRED,          VERT_PRED },

  2123.                                    { HOR_PRED,             TM_VP8_PRED } },

  2124.     };

  2125.     static const struct {

  2126.         uint8_t needs_left:1;

  2127.         uint8_t needs_top:1;

  2128.         uint8_t needs_topleft:1;

  2129.         uint8_t needs_topright:1;

  2130.     } edges[N_INTRA_PRED_MODES] = {

  2131.         [VERT_PRED]            = { .needs_top  = 1 },

  2132.         [HOR_PRED]             = { .needs_left = 1 },

  2133.         [DC_PRED]              = { .needs_top  = 1, .needs_left = 1 },

  2134.         [DIAG_DOWN_LEFT_PRED]  = { .needs_top  = 1, .needs_topright = 1 },

  2135.         [DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2136.         [VERT_RIGHT_PRED]      = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2137.         [HOR_DOWN_PRED]        = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2138.         [VERT_LEFT_PRED]       = { .needs_top  = 1, .needs_topright = 1 },

  2139.         [HOR_UP_PRED]          = { .needs_left = 1 },

  2140.         [TM_VP8_PRED]          = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2141.         [LEFT_DC_PRED]         = { .needs_left = 1 },

  2142.         [TOP_DC_PRED]          = { .needs_top  = 1 },

  2143.         [DC_128_PRED]          = { 0 },

  2144.         [DC_127_PRED]          = { 0 },

  2145.         [DC_129_PRED]          = { 0 }

  2146.     };

  2147. 

  2148.     av_assert2(mode >= 0 && mode < 10);

  2149.     mode = mode_conv[mode][have_left][have_top];

  2150.     if (edges[mode].needs_top) {

  2151.         uint8_t *top, *topleft;

  2152.         int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !p) - x) * 4;

  2153.         int n_px_need_tr = 0;

  2154. 

  2155.         if (tx == TX_4X4 && edges[mode].needs_topright && have_right)

  2156.             n_px_need_tr = 4;

  2157. 

  2158.         // if top of sb64-row, use s->intra_pred_data[] instead of

  2159.         // dst[-stride] for intra prediction (it contains pre- instead of

  2160.         // post-loopfilter data)

  2161.         if (have_top) {

  2162.             top = !(row & 7) && !y ?

  2163.                 s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  2164.                 y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];

  2165.             if (have_left)

  2166.                 topleft = !(row & 7) && !y ?

  2167.                     s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  2168.                     y == 0 || x == 0 ? &dst_edge[-stride_edge] :

  2169.                     &dst_inner[-stride_inner];

  2170.         }

  2171. 

  2172.         if (have_top &&

  2173.             (!edges[mode].needs_topleft || (have_left && top == topleft)) &&

  2174.             (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&

  2175.             n_px_need + n_px_need_tr <= n_px_have) {

  2176.             *a = top;

  2177.         } else {

  2178.             if (have_top) {

  2179.                 if (n_px_need <= n_px_have) {

  2180.                     memcpy(*a, top, n_px_need);

  2181.                 } else {

  2182.                     memcpy(*a, top, n_px_have);

  2183.                     memset(&(*a)[n_px_have], (*a)[n_px_have - 1],

  2184.                            n_px_need - n_px_have);

  2185.                 }

  2186.             } else {

  2187.                 memset(*a, 127, n_px_need);

  2188.             }

  2189.             if (edges[mode].needs_topleft) {

  2190.                 if (have_left && have_top) {

  2191.                     (*a)[-1] = topleft[-1];

  2192.                 } else {

  2193.                     (*a)[-1] = have_top ? 129 : 127;

  2194.                 }

  2195.             }

  2196.             if (tx == TX_4X4 && edges[mode].needs_topright) {

  2197.                 if (have_top && have_right &&

  2198.                     n_px_need + n_px_need_tr <= n_px_have) {

  2199.                     memcpy(&(*a)[4], &top[4], 4);

  2200.                 } else {

  2201.                     memset(&(*a)[4], (*a)[3], 4);

  2202.                 }

  2203.             }

  2204.         }

  2205.     }

  2206.     if (edges[mode].needs_left) {

  2207.         if (have_left) {

  2208.             int n_px_need = 4 << tx, i, n_px_have = (((s->rows - row) << !p) - y) * 4;

  2209.             uint8_t *dst = x == 0 ? dst_edge : dst_inner;

  2210.             ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;

  2211. 

  2212.             if (n_px_need <= n_px_have) {

  2213.                 for (i = 0; i < n_px_need; i++)

  2214.                     l[i] = dst[i * stride - 1];

  2215.             } else {

  2216.                 for (i = 0; i < n_px_have; i++)

  2217.                     l[i] = dst[i * stride - 1];

  2218.                 memset(&l[i], l[i - 1], n_px_need - n_px_have);

  2219.             }

  2220.         } else {

  2221.             memset(l, 129, 4 << tx);

  2222.         }

  2223.     }

  2224. 

  2225.     return mode;

  2226. }

  2227. 

  2228. static void intra_recon(AVCodecContext *ctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  2229. {

  2230.     VP9Context *s = ctx->priv_data;

  2231.     VP9Block *b = s->b;

  2232.     int row = s->row, col = s->col;

  2233.     int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2234.     int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2235.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2236.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2237.     int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2238.     int uvstep1d = 1 << b->uvtx, p;

  2239.     uint8_t *dst = s->dst[0], *dst_r = s->frames[CUR_FRAME].tf.f->data[0] + y_off;

  2240. 

  2241.     for (n = 0, y = 0; y < end_y; y += step1d) {

  2242.         uint8_t *ptr = dst, *ptr_r = dst_r;

  2243.         for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d,

  2244.                                ptr_r += 4 * step1d, n += step) {

  2245.             int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?

  2246.                                y * 2 + x : 0];

  2247.             LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);

  2248.             uint8_t *a = &a_buf[16], l[32];

  2249.             enum TxfmType txtp = vp9_intra_txfm_type[mode];

  2250.             int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2251. 

  2252.             mode = check_intra_mode(s, mode, &a, ptr_r,

  2253.                                     s->frames[CUR_FRAME].tf.f->linesize[0],

  2254.                                     ptr, s->y_stride, l,

  2255.                                     col, x, w4, row, y, b->tx, 0);

  2256.             s->dsp.intra_pred[b->tx][mode](ptr, s->y_stride, l, a);

  2257.             if (eob)

  2258.                 s->dsp.itxfm_add[tx][txtp](ptr, s->y_stride,

  2259.                                            s->block + 16 * n, eob);

  2260.         }

  2261.         dst_r += 4 * step1d * s->frames[CUR_FRAME].tf.f->linesize[0];

  2262.         dst   += 4 * step1d * s->y_stride;

  2263.     }

  2264. 

  2265.     // U/V

  2266.     h4 >>= 1;

  2267.     w4 >>= 1;

  2268.     end_x >>= 1;

  2269.     end_y >>= 1;

  2270.     step = 1 << (b->uvtx * 2);

  2271.     for (p = 0; p < 2; p++) {

  2272.         dst   = s->dst[1 + p];

  2273.         dst_r = s->frames[CUR_FRAME].tf.f->data[1 + p] + uv_off;

  2274.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2275.             uint8_t *ptr = dst, *ptr_r = dst_r;

  2276.             for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d,

  2277.                                    ptr_r += 4 * uvstep1d, n += step) {

  2278.                 int mode = b->uvmode;

  2279.                 LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);

  2280.                 uint8_t *a = &a_buf[16], l[32];

  2281.                 int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2282. 

  2283.                 mode = check_intra_mode(s, mode, &a, ptr_r,

  2284.                                         s->frames[CUR_FRAME].tf.f->linesize[1],

  2285.                                         ptr, s->uv_stride, l,

  2286.                                         col, x, w4, row, y, b->uvtx, p + 1);

  2287.                 s->dsp.intra_pred[b->uvtx][mode](ptr, s->uv_stride, l, a);

  2288.                 if (eob)

  2289.                     s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2290.                                                     s->uvblock[p] + 16 * n, eob);

  2291.             }

  2292.             dst_r += 4 * uvstep1d * s->frames[CUR_FRAME].tf.f->linesize[1];

  2293.             dst   += 4 * uvstep1d * s->uv_stride;

  2294.         }

  2295.     }

  2296. }

  2297. 

  2298. static av_always_inline void mc_luma_dir(VP9Context *s, vp9_mc_func (*mc)[2],

  2299.                                          uint8_t *dst, ptrdiff_t dst_stride,

  2300.                                          const uint8_t *ref, ptrdiff_t ref_stride,

  2301.                                          ThreadFrame *ref_frame,

  2302.                                          ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2303.                                          int bw, int bh, int w, int h)

  2304. {

  2305.     int mx = mv->x, my = mv->y, th;

  2306. 

  2307.     y += my >> 3;

  2308.     x += mx >> 3;

  2309.     ref += y * ref_stride + x;

  2310.     mx &= 7;

  2311.     my &= 7;

  2312.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2313.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2314.     // the longest loopfilter of the next sbrow

  2315.     th = (y + bh + 4 * !!my + 7) >> 6;

  2316.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2317.     if (x < !!mx * 3 || y < !!my * 3 ||

  2318.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2319.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2320.                                  ref - !!my * 3 * ref_stride - !!mx * 3,

  2321.                                  80, ref_stride,

  2322.                                  bw + !!mx * 7, bh + !!my * 7,

  2323.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2324.         ref = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2325.         ref_stride = 80;

  2326.     }

  2327.     mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);

  2328. }

  2329. 

  2330. static av_always_inline void mc_chroma_dir(VP9Context *s, vp9_mc_func (*mc)[2],

  2331.                                            uint8_t *dst_u, uint8_t *dst_v,

  2332.                                            ptrdiff_t dst_stride,

  2333.                                            const uint8_t *ref_u, ptrdiff_t src_stride_u,

  2334.                                            const uint8_t *ref_v, ptrdiff_t src_stride_v,

  2335.                                            ThreadFrame *ref_frame,

  2336.                                            ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2337.                                            int bw, int bh, int w, int h)

  2338. {

  2339.     int mx = mv->x, my = mv->y, th;

  2340. 

  2341.     y += my >> 4;

  2342.     x += mx >> 4;

  2343.     ref_u += y * src_stride_u + x;

  2344.     ref_v += y * src_stride_v + x;

  2345.     mx &= 15;

  2346.     my &= 15;

  2347.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2348.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2349.     // the longest loopfilter of the next sbrow

  2350.     th = (y + bh + 4 * !!my + 7) >> 5;

  2351.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2352.     if (x < !!mx * 3 || y < !!my * 3 ||

  2353.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2354.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2355.                                  ref_u - !!my * 3 * src_stride_u - !!mx * 3,

  2356.                                  80, src_stride_u,

  2357.                                  bw + !!mx * 7, bh + !!my * 7,

  2358.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2359.         ref_u = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2360.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, 80, bh, mx, my);

  2361. 

  2362.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2363.                                  ref_v - !!my * 3 * src_stride_v - !!mx * 3,

  2364.                                  80, src_stride_v,

  2365.                                  bw + !!mx * 7, bh + !!my * 7,

  2366.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2367.         ref_v = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2368.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, 80, bh, mx, my);

  2369.     } else {

  2370.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);

  2371.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);

  2372.     }

  2373. }

  2374. 

  2375. static void inter_recon(AVCodecContext *ctx)

  2376. {

  2377.     static const uint8_t bwlog_tab[2][N_BS_SIZES] = {

  2378.         { 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4 },

  2379.         { 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 4, 4 },

  2380.     };

  2381.     VP9Context *s = ctx->priv_data;

  2382.     VP9Block *b = s->b;

  2383.     int row = s->row, col = s->col;

  2384.     ThreadFrame *tref1 = &s->refs[s->refidx[b->ref[0]]];

  2385.     AVFrame *ref1 = tref1->f;

  2386.     ThreadFrame *tref2 = b->comp ? &s->refs[s->refidx[b->ref[1]]] : NULL;

  2387.     AVFrame *ref2 = b->comp ? tref2->f : NULL;

  2388.     int w = ctx->width, h = ctx->height;

  2389.     ptrdiff_t ls_y = s->y_stride, ls_uv = s->uv_stride;

  2390. 

  2391.     // y inter pred

  2392.     if (b->bs > BS_8x8) {

  2393.         if (b->bs == BS_8x4) {

  2394.             mc_luma_dir(s, s->dsp.mc[3][b->filter][0], s->dst[0], ls_y,

  2395.                         ref1->data[0], ref1->linesize[0], tref1,

  2396.                         row << 3, col << 3, &b->mv[0][0], 8, 4, w, h);

  2397.             mc_luma_dir(s, s->dsp.mc[3][b->filter][0],

  2398.                         s->dst[0] + 4 * ls_y, ls_y,

  2399.                         ref1->data[0], ref1->linesize[0], tref1,

  2400.                         (row << 3) + 4, col << 3, &b->mv[2][0], 8, 4, w, h);

  2401. 

  2402.             if (b->comp) {

  2403.                 mc_luma_dir(s, s->dsp.mc[3][b->filter][1], s->dst[0], ls_y,

  2404.                             ref2->data[0], ref2->linesize[0], tref2,

  2405.                             row << 3, col << 3, &b->mv[0][1], 8, 4, w, h);

  2406.                 mc_luma_dir(s, s->dsp.mc[3][b->filter][1],

  2407.                             s->dst[0] + 4 * ls_y, ls_y,

  2408.                             ref2->data[0], ref2->linesize[0], tref2,

  2409.                             (row << 3) + 4, col << 3, &b->mv[2][1], 8, 4, w, h);

  2410.             }

  2411.         } else if (b->bs == BS_4x8) {

  2412.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0], ls_y,

  2413.                         ref1->data[0], ref1->linesize[0], tref1,

  2414.                         row << 3, col << 3, &b->mv[0][0], 4, 8, w, h);

  2415.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0] + 4, ls_y,

  2416.                         ref1->data[0], ref1->linesize[0], tref1,

  2417.                         row << 3, (col << 3) + 4, &b->mv[1][0], 4, 8, w, h);

  2418. 

  2419.             if (b->comp) {

  2420.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0], ls_y,

  2421.                             ref2->data[0], ref2->linesize[0], tref2,

  2422.                             row << 3, col << 3, &b->mv[0][1], 4, 8, w, h);

  2423.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0] + 4, ls_y,

  2424.                             ref2->data[0], ref2->linesize[0], tref2,

  2425.                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 8, w, h);

  2426.             }

  2427.         } else {

  2428.             av_assert2(b->bs == BS_4x4);

  2429. 

  2430.             // FIXME if two horizontally adjacent blocks have the same MV,

  2431.             // do a w8 instead of a w4 call

  2432.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0], ls_y,

  2433.                         ref1->data[0], ref1->linesize[0], tref1,

  2434.                         row << 3, col << 3, &b->mv[0][0], 4, 4, w, h);

  2435.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0] + 4, ls_y,

  2436.                         ref1->data[0], ref1->linesize[0], tref1,

  2437.                         row << 3, (col << 3) + 4, &b->mv[1][0], 4, 4, w, h);

  2438.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0],

  2439.                         s->dst[0] + 4 * ls_y, ls_y,

  2440.                         ref1->data[0], ref1->linesize[0], tref1,

  2441.                         (row << 3) + 4, col << 3, &b->mv[2][0], 4, 4, w, h);

  2442.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0],

  2443.                         s->dst[0] + 4 * ls_y + 4, ls_y,

  2444.                         ref1->data[0], ref1->linesize[0], tref1,

  2445.                         (row << 3) + 4, (col << 3) + 4, &b->mv[3][0], 4, 4, w, h);

  2446. 

  2447.             if (b->comp) {

  2448.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0], ls_y,

  2449.                             ref2->data[0], ref2->linesize[0], tref2,

  2450.                             row << 3, col << 3, &b->mv[0][1], 4, 4, w, h);

  2451.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0] + 4, ls_y,

  2452.                             ref2->data[0], ref2->linesize[0], tref2,

  2453.                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 4, w, h);

  2454.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1],

  2455.                             s->dst[0] + 4 * ls_y, ls_y,

  2456.                             ref2->data[0], ref2->linesize[0], tref2,

  2457.                             (row << 3) + 4, col << 3, &b->mv[2][1], 4, 4, w, h);

  2458.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1],

  2459.                             s->dst[0] + 4 * ls_y + 4, ls_y,

  2460.                             ref2->data[0], ref2->linesize[0], tref2,

  2461.                             (row << 3) + 4, (col << 3) + 4, &b->mv[3][1], 4, 4, w, h);

  2462.             }

  2463.         }

  2464.     } else {

  2465.         int bwl = bwlog_tab[0][b->bs];

  2466.         int bw = bwh_tab[0][b->bs][0] * 4, bh = bwh_tab[0][b->bs][1] * 4;

  2467. 

  2468.         mc_luma_dir(s, s->dsp.mc[bwl][b->filter][0], s->dst[0], ls_y,

  2469.                     ref1->data[0], ref1->linesize[0], tref1,

  2470.                     row << 3, col << 3, &b->mv[0][0],bw, bh, w, h);

  2471. 

  2472.         if (b->comp)

  2473.             mc_luma_dir(s, s->dsp.mc[bwl][b->filter][1], s->dst[0], ls_y,

  2474.                         ref2->data[0], ref2->linesize[0], tref2,

  2475.                         row << 3, col << 3, &b->mv[0][1], bw, bh, w, h);

  2476.     }

  2477. 

  2478.     // uv inter pred

  2479.     {

  2480.         int bwl = bwlog_tab[1][b->bs];

  2481.         int bw = bwh_tab[1][b->bs][0] * 4, bh = bwh_tab[1][b->bs][1] * 4;

  2482.         VP56mv mvuv;

  2483. 

  2484.         w = (w + 1) >> 1;

  2485.         h = (h + 1) >> 1;

  2486.         if (b->bs > BS_8x8) {

  2487.             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);

  2488.             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);

  2489.         } else {

  2490.             mvuv = b->mv[0][0];

  2491.         }

  2492. 

  2493.         mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][0],

  2494.                       s->dst[1], s->dst[2], ls_uv,

  2495.                       ref1->data[1], ref1->linesize[1],

  2496.                       ref1->data[2], ref1->linesize[2], tref1,

  2497.                       row << 2, col << 2, &mvuv, bw, bh, w, h);

  2498. 

  2499.         if (b->comp) {

  2500.             if (b->bs > BS_8x8) {

  2501.                 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);

  2502.                 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);

  2503.             } else {

  2504.                 mvuv = b->mv[0][1];

  2505.             }

  2506.             mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][1],

  2507.                           s->dst[1], s->dst[2], ls_uv,

  2508.                           ref2->data[1], ref2->linesize[1],

  2509.                           ref2->data[2], ref2->linesize[2], tref2,

  2510.                           row << 2, col << 2, &mvuv, bw, bh, w, h);

  2511.         }

  2512.     }

  2513. 

  2514.     if (!b->skip) {

  2515.         /* mostly copied intra_reconn() */

  2516. 

  2517.         int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2518.         int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2519.         int end_x = FFMIN(2 * (s->cols - col), w4);

  2520.         int end_y = FFMIN(2 * (s->rows - row), h4);

  2521.         int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2522.         int uvstep1d = 1 << b->uvtx, p;

  2523.         uint8_t *dst = s->dst[0];

  2524. 

  2525.         // y itxfm add

  2526.         for (n = 0, y = 0; y < end_y; y += step1d) {

  2527.             uint8_t *ptr = dst;

  2528.             for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d, n += step) {

  2529.                 int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2530. 

  2531.                 if (eob)

  2532.                     s->dsp.itxfm_add[tx][DCT_DCT](ptr, s->y_stride,

  2533.                                                   s->block + 16 * n, eob);

  2534.             }

  2535.             dst += 4 * s->y_stride * step1d;

  2536.         }

  2537. 

  2538.         // uv itxfm add

  2539.         h4 >>= 1;

  2540.         w4 >>= 1;

  2541.         end_x >>= 1;

  2542.         end_y >>= 1;

  2543.         step = 1 << (b->uvtx * 2);

  2544.         for (p = 0; p < 2; p++) {

  2545.             dst = s->dst[p + 1];

  2546.             for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2547.                 uint8_t *ptr = dst;

  2548.                 for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d, n += step) {

  2549.                     int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2550. 

  2551.                     if (eob)

  2552.                         s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2553.                                                         s->uvblock[p] + 16 * n, eob);

  2554.                 }

  2555.                 dst += 4 * uvstep1d * s->uv_stride;

  2556.             }

  2557.         }

  2558.     }

  2559. }

  2560. 

  2561. static av_always_inline void mask_edges(struct VP9Filter *lflvl, int is_uv,

  2562.                                         int row_and_7, int col_and_7,

  2563.                                         int w, int h, int col_end, int row_end,

  2564.                                         enum TxfmMode tx, int skip_inter)

  2565. {

  2566.     // FIXME I'm pretty sure all loops can be replaced by a single LUT if

  2567.     // we make VP9Filter.mask uint64_t (i.e. row/col all single variable)

  2568.     // and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then

  2569.     // use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)

  2570. 

  2571.     // the intended behaviour of the vp9 loopfilter is to work on 8-pixel

  2572.     // edges. This means that for UV, we work on two subsampled blocks at

  2573.     // a time, and we only use the topleft block's mode information to set

  2574.     // things like block strength. Thus, for any block size smaller than

  2575.     // 16x16, ignore the odd portion of the block.

  2576.     if (tx == TX_4X4 && is_uv) {

  2577.         if (h == 1) {

  2578.             if (row_and_7 & 1)

  2579.                 return;

  2580.             if (!row_end)

  2581.                 h += 1;

  2582.         }

  2583.         if (w == 1) {

  2584.             if (col_and_7 & 1)

  2585.                 return;

  2586.             if (!col_end)

  2587.                 w += 1;

  2588.         }

  2589.     }

  2590. 

  2591.     if (tx == TX_4X4 && !skip_inter) {

  2592.         int t = 1 << col_and_7, m_col = (t << w) - t, y;

  2593.         int m_col_odd = (t << (w - 1)) - t;

  2594. 

  2595.         // on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide

  2596.         if (is_uv) {

  2597.             int m_row_8 = m_col & 0x01, m_row_4 = m_col - m_row_8;

  2598. 

  2599.             for (y = row_and_7; y < h + row_and_7; y++) {

  2600.                 int col_mask_id = 2 - !(y & 7);

  2601. 

  2602.                 lflvl->mask[is_uv][0][y][1] |= m_row_8;

  2603.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  2604.                 // for odd lines, if the odd col is not being filtered,

  2605.                 // skip odd row also:

  2606.                 // .---. <-- a

  2607.                 // |   |

  2608.                 // |___| <-- b

  2609.                 // ^   ^

  2610.                 // c   d

  2611.                 //

  2612.                 // if a/c are even row/col and b/d are odd, and d is skipped,

  2613.                 // e.g. right edge of size-66x66.webm, then skip b also (bug)

  2614.                 if ((col_end & 1) && (y & 1)) {

  2615.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col_odd;

  2616.                 } else {

  2617.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col;

  2618.                 }

  2619.             }

  2620.         } else {

  2621.             int m_row_8 = m_col & 0x11, m_row_4 = m_col - m_row_8;

  2622. 

  2623.             for (y = row_and_7; y < h + row_and_7; y++) {

  2624.                 int col_mask_id = 2 - !(y & 3);

  2625. 

  2626.                 lflvl->mask[is_uv][0][y][1] |= m_row_8; // row edge

  2627.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  2628.                 lflvl->mask[is_uv][1][y][col_mask_id] |= m_col; // col edge

  2629.                 lflvl->mask[is_uv][0][y][3] |= m_col;

  2630.                 lflvl->mask[is_uv][1][y][3] |= m_col;

  2631.             }

  2632.         }

  2633.     } else {

  2634.         int y, t = 1 << col_and_7, m_col = (t << w) - t;

  2635. 

  2636.         if (!skip_inter) {

  2637.             int mask_id = (tx == TX_8X8);

  2638.             int l2 = tx + is_uv - 1, step1d = 1 << l2;

  2639.             static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };

  2640.             int m_row = m_col & masks[l2];

  2641. 

  2642.             // at odd UV col/row edges tx16/tx32 loopfilter edges, force

  2643.             // 8wd loopfilter to prevent going off the visible edge.

  2644.             if (is_uv && tx > TX_8X8 && (w ^ (w - 1)) == 1) {

  2645.                 int m_row_16 = ((t << (w - 1)) - t) & masks[l2];

  2646.                 int m_row_8 = m_row - m_row_16;

  2647. 

  2648.                 for (y = row_and_7; y < h + row_and_7; y++) {

  2649.                     lflvl->mask[is_uv][0][y][0] |= m_row_16;

  2650.                     lflvl->mask[is_uv][0][y][1] |= m_row_8;

  2651.                 }

  2652.             } else {

  2653.                 for (y = row_and_7; y < h + row_and_7; y++)

  2654.                     lflvl->mask[is_uv][0][y][mask_id] |= m_row;

  2655.             }

  2656. 

  2657.             if (is_uv && tx > TX_8X8 && (h ^ (h - 1)) == 1) {

  2658.                 for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)

  2659.                     lflvl->mask[is_uv][1][y][0] |= m_col;

  2660.                 if (y - row_and_7 == h - 1)

  2661.                     lflvl->mask[is_uv][1][y][1] |= m_col;

  2662.             } else {

  2663.                 for (y = row_and_7; y < h + row_and_7; y += step1d)

  2664.                     lflvl->mask[is_uv][1][y][mask_id] |= m_col;

  2665.             }

  2666.         } else if (tx != TX_4X4) {

  2667.             int mask_id;

  2668. 

  2669.             mask_id = (tx == TX_8X8) || (is_uv && h == 1);

  2670.             lflvl->mask[is_uv][1][row_and_7][mask_id] |= m_col;

  2671.             mask_id = (tx == TX_8X8) || (is_uv && w == 1);

  2672.             for (y = row_and_7; y < h + row_and_7; y++)

  2673.                 lflvl->mask[is_uv][0][y][mask_id] |= t;

  2674.         } else if (is_uv) {

  2675.             int t8 = t & 0x01, t4 = t - t8;

  2676. 

  2677.             for (y = row_and_7; y < h + row_and_7; y++) {

  2678.                 lflvl->mask[is_uv][0][y][2] |= t4;

  2679.                 lflvl->mask[is_uv][0][y][1] |= t8;

  2680.             }

  2681.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 7)] |= m_col;

  2682.         } else {

  2683.             int t8 = t & 0x11, t4 = t - t8;

  2684. 

  2685.             for (y = row_and_7; y < h + row_and_7; y++) {

  2686.                 lflvl->mask[is_uv][0][y][2] |= t4;

  2687.                 lflvl->mask[is_uv][0][y][1] |= t8;

  2688.             }

  2689.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 3)] |= m_col;

  2690.         }

  2691.     }

  2692. }

  2693. 

  2694. static void decode_b(AVCodecContext *ctx, int row, int col,

  2695.                      struct VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,

  2696.                      enum BlockLevel bl, enum BlockPartition bp)

  2697. {

  2698.     VP9Context *s = ctx->priv_data;

  2699.     VP9Block *b = s->b;

  2700.     enum BlockSize bs = bl * 3 + bp;

  2701.     int y, w4 = bwh_tab[1][bs][0], h4 = bwh_tab[1][bs][1], lvl;

  2702.     int emu[2];

  2703.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2704. 

  2705.     s->row = row;

  2706.     s->row7 = row & 7;

  2707.     s->col = col;

  2708.     s->col7 = col & 7;

  2709.     s->min_mv.x = -(128 + col * 64);

  2710.     s->min_mv.y = -(128 + row * 64);

  2711.     s->max_mv.x = 128 + (s->cols - col - w4) * 64;

  2712.     s->max_mv.y = 128 + (s->rows - row - h4) * 64;

  2713.     if (s->pass < 2) {

  2714.         b->bs = bs;

  2715.         b->bl = bl;

  2716.         b->bp = bp;

  2717.         decode_mode(ctx);

  2718.         b->uvtx = b->tx - (w4 * 2 == (1 << b->tx) || h4 * 2 == (1 << b->tx));

  2719. 

  2720.         if (!b->skip) {

  2721.             decode_coeffs(ctx);

  2722.         } else {

  2723.             int pl;

  2724. 

  2725.             memset(&s->above_y_nnz_ctx[col * 2], 0, w4 * 2);

  2726.             memset(&s->left_y_nnz_ctx[(row & 7) << 1], 0, h4 * 2);

  2727.             for (pl = 0; pl < 2; pl++) {

  2728.                 memset(&s->above_uv_nnz_ctx[pl][col], 0, w4);

  2729.                 memset(&s->left_uv_nnz_ctx[pl][row & 7], 0, h4);

  2730.             }

  2731.         }

  2732.         if (s->pass == 1) {

  2733.             s->b++;

  2734.             s->block += w4 * h4 * 64;

  2735.             s->uvblock[0] += w4 * h4 * 16;

  2736.             s->uvblock[1] += w4 * h4 * 16;

  2737.             s->eob += 4 * w4 * h4;

  2738.             s->uveob[0] += w4 * h4;

  2739.             s->uveob[1] += w4 * h4;

  2740. 

  2741.             return;

  2742.         }

  2743.     }

  2744. 

  2745.     // emulated overhangs if the stride of the target buffer can't hold. This

  2746.     // allows to support emu-edge and so on even if we have large block

  2747.     // overhangs

  2748.     emu[0] = (col + w4) * 8 > f->linesize[0] ||

  2749.              (row + h4) > s->rows + 2 * !(ctx->flags & CODEC_FLAG_EMU_EDGE);

  2750.     emu[1] = (col + w4) * 4 > f->linesize[1] ||

  2751.              (row + h4) > s->rows + 2 * !(ctx->flags & CODEC_FLAG_EMU_EDGE);

  2752.     if (emu[0]) {

  2753.         s->dst[0] = s->tmp_y;

  2754.         s->y_stride = 64;

  2755.     } else {

  2756.         s->dst[0] = f->data[0] + yoff;

  2757.         s->y_stride = f->linesize[0];

  2758.     }

  2759.     if (emu[1]) {

  2760.         s->dst[1] = s->tmp_uv[0];

  2761.         s->dst[2] = s->tmp_uv[1];

  2762.         s->uv_stride = 32;

  2763.     } else {

  2764.         s->dst[1] = f->data[1] + uvoff;

  2765.         s->dst[2] = f->data[2] + uvoff;

  2766.         s->uv_stride = f->linesize[1];

  2767.     }

  2768.     if (b->intra) {

  2769.         intra_recon(ctx, yoff, uvoff);

  2770.     } else {

  2771.         inter_recon(ctx);

  2772.     }

  2773.     if (emu[0]) {

  2774.         int w = FFMIN(s->cols - col, w4) * 8, h = FFMIN(s->rows - row, h4) * 8, n, o = 0;

  2775. 

  2776.         for (n = 0; o < w; n++) {

  2777.             int bw = 64 >> n;

  2778. 

  2779.             av_assert2(n <= 4);

  2780.             if (w & bw) {

  2781.                 s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o, f->linesize[0],

  2782.                                          s->tmp_y + o, 64, h, 0, 0);

  2783.                 o += bw;

  2784.             }

  2785.         }

  2786.     }

  2787.     if (emu[1]) {

  2788.         int w = FFMIN(s->cols - col, w4) * 4, h = FFMIN(s->rows - row, h4) * 4, n, o = 0;

  2789. 

  2790.         for (n = 1; o < w; n++) {

  2791.             int bw = 64 >> n;

  2792. 

  2793.             av_assert2(n <= 4);

  2794.             if (w & bw) {

  2795.                 s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o, f->linesize[1],

  2796.                                          s->tmp_uv[0] + o, 32, h, 0, 0);

  2797.                 s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o, f->linesize[2],

  2798.                                          s->tmp_uv[1] + o, 32, h, 0, 0);

  2799.                 o += bw;

  2800.             }

  2801.         }

  2802.     }

  2803. 

  2804.     // pick filter level and find edges to apply filter to

  2805.     if (s->filter.level &&

  2806.         (lvl = s->segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]

  2807.                                                     [b->mode[3] != ZEROMV]) > 0) {

  2808.         int x_end = FFMIN(s->cols - col, w4), y_end = FFMIN(s->rows - row, h4);

  2809.         int skip_inter = !b->intra && b->skip;

  2810. 

  2811.         for (y = 0; y < h4; y++)

  2812.             memset(&lflvl->level[((row & 7) + y) * 8 + (col & 7)], lvl, w4);

  2813.         mask_edges(lflvl, 0, row & 7, col & 7, x_end, y_end, 0, 0, b->tx, skip_inter);

  2814.         mask_edges(lflvl, 1, row & 7, col & 7, x_end, y_end,

  2815.                    s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,

  2816.                    s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,

  2817.                    b->uvtx, skip_inter);

  2818. 

  2819.         if (!s->filter.lim_lut[lvl]) {

  2820.             int sharp = s->filter.sharpness;

  2821.             int limit = lvl;

  2822. 

  2823.             if (sharp > 0) {

  2824.                 limit >>= (sharp + 3) >> 2;

  2825.                 limit = FFMIN(limit, 9 - sharp);

  2826.             }

  2827.             limit = FFMAX(limit, 1);

  2828. 

  2829.             s->filter.lim_lut[lvl] = limit;

  2830.             s->filter.mblim_lut[lvl] = 2 * (lvl + 2) + limit;

  2831.         }

  2832.     }

  2833. 

  2834.     if (s->pass == 2) {

  2835.         s->b++;

  2836.         s->block += w4 * h4 * 64;

  2837.         s->uvblock[0] += w4 * h4 * 16;

  2838.         s->uvblock[1] += w4 * h4 * 16;

  2839.         s->eob += 4 * w4 * h4;

  2840.         s->uveob[0] += w4 * h4;

  2841.         s->uveob[1] += w4 * h4;

  2842.     }

  2843. }

  2844. 

  2845. static void decode_sb(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  2846.                       ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  2847. {

  2848.     VP9Context *s = ctx->priv_data;

  2849.     int c = ((s->above_partition_ctx[col] >> (3 - bl)) & 1) |

  2850.             (((s->left_partition_ctx[row & 0x7] >> (3 - bl)) & 1) << 1);

  2851.     const uint8_t *p = s->keyframe ? vp9_default_kf_partition_probs[bl][c] :

  2852.                                      s->prob.p.partition[bl][c];

  2853.     enum BlockPartition bp;

  2854.     ptrdiff_t hbs = 4 >> bl;

  2855.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2856.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  2857. 

  2858.     if (bl == BL_8X8) {

  2859.         bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  2860.         decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2861.     } else if (col + hbs < s->cols) { // FIXME why not <=?

  2862.         if (row + hbs < s->rows) { // FIXME why not <=?

  2863.             bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  2864.             switch (bp) {

  2865.             case PARTITION_NONE:

  2866.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2867.                 break;

  2868.             case PARTITION_H:

  2869.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2870.                 yoff  += hbs * 8 * y_stride;

  2871.                 uvoff += hbs * 4 * uv_stride;

  2872.                 decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, bl, bp);

  2873.                 break;

  2874.             case PARTITION_V:

  2875.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2876.                 yoff  += hbs * 8;

  2877.                 uvoff += hbs * 4;

  2878.                 decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, bl, bp);

  2879.                 break;

  2880.             case PARTITION_SPLIT:

  2881.                 decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2882.                 decode_sb(ctx, row, col + hbs, lflvl,

  2883.                           yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2884.                 yoff  += hbs * 8 * y_stride;

  2885.                 uvoff += hbs * 4 * uv_stride;

  2886.                 decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  2887.                 decode_sb(ctx, row + hbs, col + hbs, lflvl,

  2888.                           yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2889.                 break;

  2890.             default:

  2891.                 av_assert0(0);

  2892.             }

  2893.         } else if (vp56_rac_get_prob_branchy(&s->c, p[1])) {

  2894.             bp = PARTITION_SPLIT;

  2895.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2896.             decode_sb(ctx, row, col + hbs, lflvl,

  2897.                       yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2898.         } else {

  2899.             bp = PARTITION_H;

  2900.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2901.         }

  2902.     } else if (row + hbs < s->rows) { // FIXME why not <=?

  2903.         if (vp56_rac_get_prob_branchy(&s->c, p[2])) {

  2904.             bp = PARTITION_SPLIT;

  2905.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2906.             yoff  += hbs * 8 * y_stride;

  2907.             uvoff += hbs * 4 * uv_stride;

  2908.             decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  2909.         } else {

  2910.             bp = PARTITION_V;

  2911.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2912.         }

  2913.     } else {

  2914.         bp = PARTITION_SPLIT;

  2915.         decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2916.     }

  2917.     s->counts.partition[bl][c][bp]++;

  2918. }

  2919. 

  2920. static void decode_sb_mem(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  2921.                           ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  2922. {

  2923.     VP9Context *s = ctx->priv_data;

  2924.     VP9Block *b = s->b;

  2925.     ptrdiff_t hbs = 4 >> bl;

  2926.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2927.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  2928. 

  2929.     if (bl == BL_8X8) {

  2930.         av_assert2(b->bl == BL_8X8);

  2931.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  2932.     } else if (s->b->bl == bl) {

  2933.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  2934.         if (b->bp == PARTITION_H && row + hbs < s->rows) {

  2935.             yoff  += hbs * 8 * y_stride;

  2936.             uvoff += hbs * 4 * uv_stride;

  2937.             decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, b->bl, b->bp);

  2938.         } else if (b->bp == PARTITION_V && col + hbs < s->cols) {

  2939.             yoff  += hbs * 8;

  2940.             uvoff += hbs * 4;

  2941.             decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, b->bl, b->bp);

  2942.         }

  2943.     } else {

  2944.         decode_sb_mem(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2945.         if (col + hbs < s->cols) { // FIXME why not <=?

  2946.             if (row + hbs < s->rows) {

  2947.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff + 8 * hbs,

  2948.                               uvoff + 4 * hbs, bl + 1);

  2949.                 yoff  += hbs * 8 * y_stride;

  2950.                 uvoff += hbs * 4 * uv_stride;

  2951.                 decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  2952.                 decode_sb_mem(ctx, row + hbs, col + hbs, lflvl,

  2953.                                     yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2954.             } else {

  2955.                 yoff  += hbs * 8;

  2956.                 uvoff += hbs * 4;

  2957.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff, uvoff, bl + 1);

  2958.             }

  2959.         } else if (row + hbs < s->rows) {

  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.         }

  2964.     }

  2965. }

  2966. 

  2967. static void loopfilter_sb(AVCodecContext *ctx, struct VP9Filter *lflvl,

  2968.                           int row, int col, ptrdiff_t yoff, ptrdiff_t uvoff)

  2969. {

  2970.     VP9Context *s = ctx->priv_data;

  2971.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2972.     uint8_t *dst = f->data[0] + yoff, *lvl = lflvl->level;

  2973.     ptrdiff_t ls_y = f->linesize[0], ls_uv = f->linesize[1];

  2974.     int y, x, p;

  2975. 

  2976.     // FIXME in how far can we interleave the v/h loopfilter calls? E.g.

  2977.     // if you think of them as acting on a 8x8 block max, we can interleave

  2978.     // each v/h within the single x loop, but that only works if we work on

  2979.     // 8 pixel blocks, and we won't always do that (we want at least 16px

  2980.     // to use SSE2 optimizations, perhaps 32 for AVX2)

  2981. 

  2982.     // filter edges between columns, Y plane (e.g. block1 | block2)

  2983.     for (y = 0; y < 8; y += 2, dst += 16 * ls_y, lvl += 16) {

  2984.         uint8_t *ptr = dst, *l = lvl, *hmask1 = lflvl->mask[0][0][y];

  2985.         uint8_t *hmask2 = lflvl->mask[0][0][y + 1];

  2986.         unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2], hm13 = hmask1[3];

  2987.         unsigned hm2 = hmask2[1] | hmask2[2], hm23 = hmask2[3];

  2988.         unsigned hm = hm1 | hm2 | hm13 | hm23;

  2989. 

  2990.         for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 8, l++) {

  2991.             if (hm1 & x) {

  2992.                 int L = *l, H = L >> 4;

  2993.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  2994. 

  2995.                 if (col || x > 1) {

  2996.                     if (hmask1[0] & x) {

  2997.                         if (hmask2[0] & x) {

  2998.                             av_assert2(l[8] == L);

  2999.                             s->dsp.loop_filter_16[0](ptr, ls_y, E, I, H);

  3000.                         } else {

  3001.                             s->dsp.loop_filter_8[2][0](ptr, ls_y, E, I, H);

  3002.                         }

  3003.                     } else if (hm2 & x) {

  3004.                         L = l[8];

  3005.                         H |= (L >> 4) << 8;

  3006.                         E |= s->filter.mblim_lut[L] << 8;

  3007.                         I |= s->filter.lim_lut[L] << 8;

  3008.                         s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3009.                                                [!!(hmask2[1] & x)]

  3010.                                                [0](ptr, ls_y, E, I, H);

  3011.                     } else {

  3012.                         s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3013.                                             [0](ptr, ls_y, E, I, H);

  3014.                     }

  3015.                 }

  3016.             } else if (hm2 & x) {

  3017.                 int L = l[8], H = L >> 4;

  3018.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3019. 

  3020.                 if (col || x > 1) {

  3021.                     s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3022.                                         [0](ptr + 8 * ls_y, ls_y, E, I, H);

  3023.                 }

  3024.             }

  3025.             if (hm13 & x) {

  3026.                 int L = *l, H = L >> 4;

  3027.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3028. 

  3029.                 if (hm23 & x) {

  3030.                     L = l[8];

  3031.                     H |= (L >> 4) << 8;

  3032.                     E |= s->filter.mblim_lut[L] << 8;

  3033.                     I |= s->filter.lim_lut[L] << 8;

  3034.                     s->dsp.loop_filter_mix2[0][0][0](ptr + 4, ls_y, E, I, H);

  3035.                 } else {

  3036.                     s->dsp.loop_filter_8[0][0](ptr + 4, ls_y, E, I, H);

  3037.                 }

  3038.             } else if (hm23 & x) {

  3039.                 int L = l[8], H = L >> 4;

  3040.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3041. 

  3042.                 s->dsp.loop_filter_8[0][0](ptr + 8 * ls_y + 4, ls_y, E, I, H);

  3043.             }

  3044.         }

  3045.     }

  3046. 

  3047.     //                                          block1

  3048.     // filter edges between rows, Y plane (e.g. ------)

  3049.     //                                          block2

  3050.     dst = f->data[0] + yoff;

  3051.     lvl = lflvl->level;

  3052.     for (y = 0; y < 8; y++, dst += 8 * ls_y, lvl += 8) {

  3053.         uint8_t *ptr = dst, *l = lvl, *vmask = lflvl->mask[0][1][y];

  3054.         unsigned vm = vmask[0] | vmask[1] | vmask[2], vm3 = vmask[3];

  3055. 

  3056.         for (x = 1; vm & ~(x - 1); x <<= 2, ptr += 16, l += 2) {

  3057.             if (row || y) {

  3058.                 if (vm & x) {

  3059.                     int L = *l, H = L >> 4;

  3060.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3061. 

  3062.                     if (vmask[0] & x) {

  3063.                         if (vmask[0] & (x << 1)) {

  3064.                             av_assert2(l[1] == L);

  3065.                             s->dsp.loop_filter_16[1](ptr, ls_y, E, I, H);

  3066.                         } else {

  3067.                             s->dsp.loop_filter_8[2][1](ptr, ls_y, E, I, H);

  3068.                         }

  3069.                     } else if (vm & (x << 1)) {

  3070.                         L = l[1];

  3071.                         H |= (L >> 4) << 8;

  3072.                         E |= s->filter.mblim_lut[L] << 8;

  3073.                         I |= s->filter.lim_lut[L] << 8;

  3074.                         s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3075.                                                [!!(vmask[1] & (x << 1))]

  3076.                                                [1](ptr, ls_y, E, I, H);

  3077.                     } else {

  3078.                         s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3079.                                             [1](ptr, ls_y, E, I, H);

  3080.                     }

  3081.                 } else if (vm & (x << 1)) {

  3082.                     int L = l[1], H = L >> 4;

  3083.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3084. 

  3085.                     s->dsp.loop_filter_8[!!(vmask[1] & (x << 1))]

  3086.                                         [1](ptr + 8, ls_y, E, I, H);

  3087.                 }

  3088.             }

  3089.             if (vm3 & x) {

  3090.                 int L = *l, H = L >> 4;

  3091.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3092. 

  3093.                 if (vm3 & (x << 1)) {

  3094.                     L = l[1];

  3095.                     H |= (L >> 4) << 8;

  3096.                     E |= s->filter.mblim_lut[L] << 8;

  3097.                     I |= s->filter.lim_lut[L] << 8;

  3098.                     s->dsp.loop_filter_mix2[0][0][1](ptr + ls_y * 4, ls_y, E, I, H);

  3099.                 } else {

  3100.                     s->dsp.loop_filter_8[0][1](ptr + ls_y * 4, ls_y, E, I, H);

  3101.                 }

  3102.             } else if (vm3 & (x << 1)) {

  3103.                 int L = l[1], H = L >> 4;

  3104.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3105. 

  3106.                 s->dsp.loop_filter_8[0][1](ptr + ls_y * 4 + 8, ls_y, E, I, H);

  3107.             }

  3108.         }

  3109.     }

  3110. 

  3111.     // same principle but for U/V planes

  3112.     for (p = 0; p < 2; p++) {

  3113.         lvl = lflvl->level;

  3114.         dst = f->data[1 + p] + uvoff;

  3115.         for (y = 0; y < 8; y += 4, dst += 16 * ls_uv, lvl += 32) {

  3116.             uint8_t *ptr = dst, *l = lvl, *hmask1 = lflvl->mask[1][0][y];

  3117.             uint8_t *hmask2 = lflvl->mask[1][0][y + 2];

  3118.             unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2];

  3119.             unsigned hm2 = hmask2[1] | hmask2[2], hm = hm1 | hm2;

  3120. 

  3121.             for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 4) {

  3122.                 if (col || x > 1) {

  3123.                     if (hm1 & x) {

  3124.                         int L = *l, H = L >> 4;

  3125.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3126. 

  3127.                         if (hmask1[0] & x) {

  3128.                             if (hmask2[0] & x) {

  3129.                                 av_assert2(l[16] == L);

  3130.                                 s->dsp.loop_filter_16[0](ptr, ls_uv, E, I, H);

  3131.                             } else {

  3132.                                 s->dsp.loop_filter_8[2][0](ptr, ls_uv, E, I, H);

  3133.                             }

  3134.                         } else if (hm2 & x) {

  3135.                             L = l[16];

  3136.                             H |= (L >> 4) << 8;

  3137.                             E |= s->filter.mblim_lut[L] << 8;

  3138.                             I |= s->filter.lim_lut[L] << 8;

  3139.                             s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3140.                                                    [!!(hmask2[1] & x)]

  3141.                                                    [0](ptr, ls_uv, E, I, H);

  3142.                         } else {

  3143.                             s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3144.                                                 [0](ptr, ls_uv, E, I, H);

  3145.                         }

  3146.                     } else if (hm2 & x) {

  3147.                         int L = l[16], H = L >> 4;

  3148.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3149. 

  3150.                         s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3151.                                             [0](ptr + 8 * ls_uv, ls_uv, E, I, H);

  3152.                     }

  3153.                 }

  3154.                 if (x & 0xAA)

  3155.                     l += 2;

  3156.             }

  3157.         }

  3158.         lvl = lflvl->level;

  3159.         dst = f->data[1 + p] + uvoff;

  3160.         for (y = 0; y < 8; y++, dst += 4 * ls_uv) {

  3161.             uint8_t *ptr = dst, *l = lvl, *vmask = lflvl->mask[1][1][y];

  3162.             unsigned vm = vmask[0] | vmask[1] | vmask[2];

  3163. 

  3164.             for (x = 1; vm & ~(x - 1); x <<= 4, ptr += 16, l += 4) {

  3165.                 if (row || y) {

  3166.                     if (vm & x) {

  3167.                         int L = *l, H = L >> 4;

  3168.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3169. 

  3170.                         if (vmask[0] & x) {

  3171.                             if (vmask[0] & (x << 2)) {

  3172.                                 av_assert2(l[2] == L);

  3173.                                 s->dsp.loop_filter_16[1](ptr, ls_uv, E, I, H);

  3174.                             } else {

  3175.                                 s->dsp.loop_filter_8[2][1](ptr, ls_uv, E, I, H);

  3176.                             }

  3177.                         } else if (vm & (x << 2)) {

  3178.                             L = l[2];

  3179.                             H |= (L >> 4) << 8;

  3180.                             E |= s->filter.mblim_lut[L] << 8;

  3181.                             I |= s->filter.lim_lut[L] << 8;

  3182.                             s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3183.                                                    [!!(vmask[1] & (x << 2))]

  3184.                                                    [1](ptr, ls_uv, E, I, H);

  3185.                         } else {

  3186.                             s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3187.                                                 [1](ptr, ls_uv, E, I, H);

  3188.                         }

  3189.                     } else if (vm & (x << 2)) {

  3190.                         int L = l[2], H = L >> 4;

  3191.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3192. 

  3193.                         s->dsp.loop_filter_8[!!(vmask[1] & (x << 2))]

  3194.                                             [1](ptr + 8, ls_uv, E, I, H);

  3195.                     }

  3196.                 }

  3197.             }

  3198.             if (y & 1)

  3199.                 lvl += 16;

  3200.         }

  3201.     }

  3202. }

  3203. 

  3204. static void set_tile_offset(int *start, int *end, int idx, int log2_n, int n)

  3205. {

  3206.     int sb_start = ( idx      * n) >> log2_n;

  3207.     int sb_end   = ((idx + 1) * n) >> log2_n;

  3208.     *start = FFMIN(sb_start, n) << 3;

  3209.     *end   = FFMIN(sb_end,   n) << 3;

  3210. }

  3211. 

  3212. static av_always_inline void adapt_prob(uint8_t *p, unsigned ct0, unsigned ct1,

  3213.                                         int max_count, int update_factor)

  3214. {

  3215.     unsigned ct = ct0 + ct1, p2, p1;

  3216. 

  3217.     if (!ct)

  3218.         return;

  3219. 

  3220.     p1 = *p;

  3221.     p2 = ((ct0 << 8) + (ct >> 1)) / ct;

  3222.     p2 = av_clip(p2, 1, 255);

  3223.     ct = FFMIN(ct, max_count);

  3224.     update_factor = FASTDIV(update_factor * ct, max_count);

  3225. 

  3226.     // (p1 * (256 - update_factor) + p2 * update_factor + 128) >> 8

  3227.     *p = p1 + (((p2 - p1) * update_factor + 128) >> 8);

  3228. }

  3229. 

  3230. static void adapt_probs(VP9Context *s)

  3231. {

  3232.     int i, j, k, l, m;

  3233.     prob_context *p = &s->prob_ctx[s->framectxid].p;

  3234.     int uf = (s->keyframe || s->intraonly || !s->last_keyframe) ? 112 : 128;

  3235. 

  3236.     // coefficients

  3237.     for (i = 0; i < 4; i++)

  3238.         for (j = 0; j < 2; j++)

  3239.             for (k = 0; k < 2; k++)

  3240.                 for (l = 0; l < 6; l++)

  3241.                     for (m = 0; m < 6; m++) {

  3242.                         uint8_t *pp = s->prob_ctx[s->framectxid].coef[i][j][k][l][m];

  3243.                         unsigned *e = s->counts.eob[i][j][k][l][m];

  3244.                         unsigned *c = s->counts.coef[i][j][k][l][m];

  3245. 

  3246.                         if (l == 0 && m >= 3) // dc only has 3 pt

  3247.                             break;

  3248. 

  3249.                         adapt_prob(&pp[0], e[0], e[1], 24, uf);

  3250.                         adapt_prob(&pp[1], c[0], c[1] + c[2], 24, uf);

  3251.                         adapt_prob(&pp[2], c[1], c[2], 24, uf);

  3252.                     }

  3253. 

  3254.     if (s->keyframe || s->intraonly) {

  3255.         memcpy(p->skip,  s->prob.p.skip,  sizeof(p->skip));

  3256.         memcpy(p->tx32p, s->prob.p.tx32p, sizeof(p->tx32p));

  3257.         memcpy(p->tx16p, s->prob.p.tx16p, sizeof(p->tx16p));

  3258.         memcpy(p->tx8p,  s->prob.p.tx8p,  sizeof(p->tx8p));

  3259.         return;

  3260.     }

  3261. 

  3262.     // skip flag

  3263.     for (i = 0; i < 3; i++)

  3264.         adapt_prob(&p->skip[i], s->counts.skip[i][0], s->counts.skip[i][1], 20, 128);

  3265. 

  3266.     // intra/inter flag

  3267.     for (i = 0; i < 4; i++)

  3268.         adapt_prob(&p->intra[i], s->counts.intra[i][0], s->counts.intra[i][1], 20, 128);

  3269. 

  3270.     // comppred flag

  3271.     if (s->comppredmode == PRED_SWITCHABLE) {

  3272.       for (i = 0; i < 5; i++)

  3273.           adapt_prob(&p->comp[i], s->counts.comp[i][0], s->counts.comp[i][1], 20, 128);

  3274.     }

  3275. 

  3276.     // reference frames

  3277.     if (s->comppredmode != PRED_SINGLEREF) {

  3278.       for (i = 0; i < 5; i++)

  3279.           adapt_prob(&p->comp_ref[i], s->counts.comp_ref[i][0],

  3280.                      s->counts.comp_ref[i][1], 20, 128);

  3281.     }

  3282. 

  3283.     if (s->comppredmode != PRED_COMPREF) {

  3284.       for (i = 0; i < 5; i++) {

  3285.           uint8_t *pp = p->single_ref[i];

  3286.           unsigned (*c)[2] = s->counts.single_ref[i];

  3287. 

  3288.           adapt_prob(&pp[0], c[0][0], c[0][1], 20, 128);

  3289.           adapt_prob(&pp[1], c[1][0], c[1][1], 20, 128);

  3290.       }

  3291.     }

  3292. 

  3293.     // block partitioning

  3294.     for (i = 0; i < 4; i++)

  3295.         for (j = 0; j < 4; j++) {

  3296.             uint8_t *pp = p->partition[i][j];

  3297.             unsigned *c = s->counts.partition[i][j];

  3298. 

  3299.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3300.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3301.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3302.         }

  3303. 

  3304.     // tx size

  3305.     if (s->txfmmode == TX_SWITCHABLE) {

  3306.       for (i = 0; i < 2; i++) {

  3307.           unsigned *c16 = s->counts.tx16p[i], *c32 = s->counts.tx32p[i];

  3308. 

  3309.           adapt_prob(&p->tx8p[i], s->counts.tx8p[i][0], s->counts.tx8p[i][1], 20, 128);

  3310.           adapt_prob(&p->tx16p[i][0], c16[0], c16[1] + c16[2], 20, 128);

  3311.           adapt_prob(&p->tx16p[i][1], c16[1], c16[2], 20, 128);

  3312.           adapt_prob(&p->tx32p[i][0], c32[0], c32[1] + c32[2] + c32[3], 20, 128);

  3313.           adapt_prob(&p->tx32p[i][1], c32[1], c32[2] + c32[3], 20, 128);

  3314.           adapt_prob(&p->tx32p[i][2], c32[2], c32[3], 20, 128);

  3315.       }

  3316.     }

  3317. 

  3318.     // interpolation filter

  3319.     if (s->filtermode == FILTER_SWITCHABLE) {

  3320.         for (i = 0; i < 4; i++) {

  3321.             uint8_t *pp = p->filter[i];

  3322.             unsigned *c = s->counts.filter[i];

  3323. 

  3324.             adapt_prob(&pp[0], c[0], c[1] + c[2], 20, 128);

  3325.             adapt_prob(&pp[1], c[1], c[2], 20, 128);

  3326.         }

  3327.     }

  3328. 

  3329.     // inter modes

  3330.     for (i = 0; i < 7; i++) {

  3331.         uint8_t *pp = p->mv_mode[i];

  3332.         unsigned *c = s->counts.mv_mode[i];

  3333. 

  3334.         adapt_prob(&pp[0], c[2], c[1] + c[0] + c[3], 20, 128);

  3335.         adapt_prob(&pp[1], c[0], c[1] + c[3], 20, 128);

  3336.         adapt_prob(&pp[2], c[1], c[3], 20, 128);

  3337.     }

  3338. 

  3339.     // mv joints

  3340.     {

  3341.         uint8_t *pp = p->mv_joint;

  3342.         unsigned *c = s->counts.mv_joint;

  3343. 

  3344.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3345.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3346.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3347.     }

  3348. 

  3349.     // mv components

  3350.     for (i = 0; i < 2; i++) {

  3351.         uint8_t *pp;

  3352.         unsigned *c, (*c2)[2], sum;

  3353. 

  3354.         adapt_prob(&p->mv_comp[i].sign, s->counts.mv_comp[i].sign[0],

  3355.                    s->counts.mv_comp[i].sign[1], 20, 128);

  3356. 

  3357.         pp = p->mv_comp[i].classes;

  3358.         c = s->counts.mv_comp[i].classes;

  3359.         sum = c[1] + c[2] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9] + c[10];

  3360.         adapt_prob(&pp[0], c[0], sum, 20, 128);

  3361.         sum -= c[1];

  3362.         adapt_prob(&pp[1], c[1], sum, 20, 128);

  3363.         sum -= c[2] + c[3];

  3364.         adapt_prob(&pp[2], c[2] + c[3], sum, 20, 128);

  3365.         adapt_prob(&pp[3], c[2], c[3], 20, 128);

  3366.         sum -= c[4] + c[5];

  3367.         adapt_prob(&pp[4], c[4] + c[5], sum, 20, 128);

  3368.         adapt_prob(&pp[5], c[4], c[5], 20, 128);

  3369.         sum -= c[6];

  3370.         adapt_prob(&pp[6], c[6], sum, 20, 128);

  3371.         adapt_prob(&pp[7], c[7] + c[8], c[9] + c[10], 20, 128);

  3372.         adapt_prob(&pp[8], c[7], c[8], 20, 128);

  3373.         adapt_prob(&pp[9], c[9], c[10], 20, 128);

  3374. 

  3375.         adapt_prob(&p->mv_comp[i].class0, s->counts.mv_comp[i].class0[0],

  3376.                    s->counts.mv_comp[i].class0[1], 20, 128);

  3377.         pp = p->mv_comp[i].bits;

  3378.         c2 = s->counts.mv_comp[i].bits;

  3379.         for (j = 0; j < 10; j++)

  3380.             adapt_prob(&pp[j], c2[j][0], c2[j][1], 20, 128);

  3381. 

  3382.         for (j = 0; j < 2; j++) {

  3383.             pp = p->mv_comp[i].class0_fp[j];

  3384.             c = s->counts.mv_comp[i].class0_fp[j];

  3385.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3386.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3387.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3388.         }

  3389.         pp = p->mv_comp[i].fp;

  3390.         c = s->counts.mv_comp[i].fp;

  3391.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3392.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3393.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3394. 

  3395.         if (s->highprecisionmvs) {

  3396.             adapt_prob(&p->mv_comp[i].class0_hp, s->counts.mv_comp[i].class0_hp[0],

  3397.                        s->counts.mv_comp[i].class0_hp[1], 20, 128);

  3398.             adapt_prob(&p->mv_comp[i].hp, s->counts.mv_comp[i].hp[0],

  3399.                        s->counts.mv_comp[i].hp[1], 20, 128);

  3400.         }

  3401.     }

  3402. 

  3403.     // y intra modes

  3404.     for (i = 0; i < 4; i++) {

  3405.         uint8_t *pp = p->y_mode[i];

  3406.         unsigned *c = s->counts.y_mode[i], sum, s2;

  3407. 

  3408.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3409.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3410.         sum -= c[TM_VP8_PRED];

  3411.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3412.         sum -= c[VERT_PRED];

  3413.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3414.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3415.         sum -= s2;

  3416.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3417.         s2 -= c[HOR_PRED];

  3418.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3419.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3420.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3421.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3422.         sum -= c[VERT_LEFT_PRED];

  3423.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3424.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3425.     }

  3426. 

  3427.     // uv intra modes

  3428.     for (i = 0; i < 10; i++) {

  3429.         uint8_t *pp = p->uv_mode[i];

  3430.         unsigned *c = s->counts.uv_mode[i], sum, s2;

  3431. 

  3432.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3433.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3434.         sum -= c[TM_VP8_PRED];

  3435.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3436.         sum -= c[VERT_PRED];

  3437.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3438.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3439.         sum -= s2;

  3440.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3441.         s2 -= c[HOR_PRED];

  3442.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3443.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3444.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3445.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3446.         sum -= c[VERT_LEFT_PRED];

  3447.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3448.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3449.     }

  3450. }

  3451. 

  3452. static av_cold int vp9_decode_free(AVCodecContext *ctx)

  3453. {

  3454.     VP9Context *s = ctx->priv_data;

  3455.     int i;

  3456. 

  3457.     for (i = 0; i < 2; i++) {

  3458.         if (s->frames[i].tf.f->data[0])

  3459.             vp9_unref_frame(ctx, &s->frames[i]);

  3460.         av_frame_free(&s->frames[i].tf.f);

  3461.     }

  3462.     for (i = 0; i < 8; i++) {

  3463.         if (s->refs[i].f->data[0])

  3464.             ff_thread_release_buffer(ctx, &s->refs[i]);

  3465.         av_frame_free(&s->refs[i].f);

  3466.         if (s->next_refs[i].f->data[0])

  3467.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3468.         av_frame_free(&s->next_refs[i].f);

  3469.     }

  3470.     av_freep(&s->above_partition_ctx);

  3471.     av_freep(&s->c_b);

  3472.     s->c_b_size = 0;

  3473.     av_freep(&s->b_base);

  3474.     av_freep(&s->block_base);

  3475. 

  3476.     return 0;

  3477. }

  3478. 

  3479. 

  3480. static int vp9_decode_frame(AVCodecContext *ctx, void *frame,

  3481.                             int *got_frame, AVPacket *pkt)

  3482. {

  3483.     const uint8_t *data = pkt->data;

  3484.     int size = pkt->size;

  3485.     VP9Context *s = ctx->priv_data;

  3486.     int res, tile_row, tile_col, i, ref, row, col;

  3487.     ptrdiff_t yoff, uvoff, ls_y, ls_uv;

  3488.     AVFrame *f;

  3489. 

  3490.     if ((res = decode_frame_header(ctx, data, size, &ref)) < 0) {

  3491.         return res;

  3492.     } else if (res == 0) {

  3493.         if (!s->refs[ref].f->data[0]) {

  3494.             av_log(ctx, AV_LOG_ERROR, "Requested reference %d not available\n", ref);

  3495.             return AVERROR_INVALIDDATA;

  3496.         }

  3497.         if ((res = av_frame_ref(frame, s->refs[ref].f)) < 0)

  3498.             return res;

  3499.         *got_frame = 1;

  3500.         return 0;

  3501.     }

  3502.     data += res;

  3503.     size -= res;

  3504. 

  3505.     if (s->frames[LAST_FRAME].tf.f->data[0])

  3506.         vp9_unref_frame(ctx, &s->frames[LAST_FRAME]);

  3507.     if (!s->keyframe && s->frames[CUR_FRAME].tf.f->data[0] &&

  3508.         (res = vp9_ref_frame(ctx, &s->frames[LAST_FRAME], &s->frames[CUR_FRAME])) < 0)

  3509.         return res;

  3510.     if (s->frames[CUR_FRAME].tf.f->data[0])

  3511.         vp9_unref_frame(ctx, &s->frames[CUR_FRAME]);

  3512.     if ((res = vp9_alloc_frame(ctx, &s->frames[CUR_FRAME])) < 0)

  3513.         return res;

  3514.     f = s->frames[CUR_FRAME].tf.f;

  3515.     f->key_frame = s->keyframe;

  3516.     f->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;

  3517.     ls_y = f->linesize[0];

  3518.     ls_uv =f->linesize[1];

  3519. 

  3520.     // ref frame setup

  3521.     for (i = 0; i < 8; i++) {

  3522.         if (s->next_refs[i].f->data[0])

  3523.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3524.         if (s->refreshrefmask & (1 << i)) {

  3525.             res = ff_thread_ref_frame(&s->next_refs[i], &s->frames[CUR_FRAME].tf);

  3526.         } else {

  3527.             res = ff_thread_ref_frame(&s->next_refs[i], &s->refs[i]);

  3528.         }

  3529.         if (res < 0)

  3530.             return res;

  3531.     }

  3532. 

  3533.     // main tile decode loop

  3534.     memset(s->above_partition_ctx, 0, s->cols);

  3535.     memset(s->above_skip_ctx, 0, s->cols);

  3536.     if (s->keyframe || s->intraonly) {

  3537.         memset(s->above_mode_ctx, DC_PRED, s->cols * 2);

  3538.     } else {

  3539.         memset(s->above_mode_ctx, NEARESTMV, s->cols);

  3540.     }

  3541.     memset(s->above_y_nnz_ctx, 0, s->sb_cols * 16);

  3542.     memset(s->above_uv_nnz_ctx[0], 0, s->sb_cols * 8);

  3543.     memset(s->above_uv_nnz_ctx[1], 0, s->sb_cols * 8);

  3544.     memset(s->above_segpred_ctx, 0, s->cols);

  3545.     s->pass = s->uses_2pass =

  3546.         ctx->active_thread_type == FF_THREAD_FRAME && s->refreshctx && !s->parallelmode;

  3547.     if (s->refreshctx && s->parallelmode) {

  3548.         int j, k, l, m;

  3549. 

  3550.         for (i = 0; i < 4; i++) {

  3551.             for (j = 0; j < 2; j++)

  3552.                 for (k = 0; k < 2; k++)

  3553.                     for (l = 0; l < 6; l++)

  3554.                         for (m = 0; m < 6; m++)

  3555.                             memcpy(s->prob_ctx[s->framectxid].coef[i][j][k][l][m],

  3556.                                    s->prob.coef[i][j][k][l][m], 3);

  3557.             if (s->txfmmode == i)

  3558.                 break;

  3559.         }

  3560.         s->prob_ctx[s->framectxid].p = s->prob.p;

  3561.         ff_thread_finish_setup(ctx);

  3562.     }

  3563. 

  3564.     do {

  3565.         yoff = uvoff = 0;

  3566.         s->b = s->b_base;

  3567.         s->block = s->block_base;

  3568.         s->uvblock[0] = s->uvblock_base[0];

  3569.         s->uvblock[1] = s->uvblock_base[1];

  3570.         s->eob = s->eob_base;

  3571.         s->uveob[0] = s->uveob_base[0];

  3572.         s->uveob[1] = s->uveob_base[1];

  3573. 

  3574.         for (tile_row = 0; tile_row < s->tiling.tile_rows; tile_row++) {

  3575.             set_tile_offset(&s->tiling.tile_row_start, &s->tiling.tile_row_end,

  3576.                             tile_row, s->tiling.log2_tile_rows, s->sb_rows);

  3577.             if (s->pass != 2) {

  3578.                 for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  3579.                     unsigned tile_size;

  3580. 

  3581.                     if (tile_col == s->tiling.tile_cols - 1 &&

  3582.                         tile_row == s->tiling.tile_rows - 1) {

  3583.                         tile_size = size;

  3584.                     } else {

  3585.                         tile_size = AV_RB32(data);

  3586.                         data += 4;

  3587.                         size -= 4;

  3588.                     }

  3589.                     if (tile_size > size)

  3590.                         return AVERROR_INVALIDDATA;

  3591.                     ff_vp56_init_range_decoder(&s->c_b[tile_col], data, tile_size);

  3592.                     if (vp56_rac_get_prob_branchy(&s->c_b[tile_col], 128)) // marker bit

  3593.                         return AVERROR_INVALIDDATA;

  3594.                     data += tile_size;

  3595.                     size -= tile_size;

  3596.                 }

  3597.             }

  3598. 

  3599.             for (row = s->tiling.tile_row_start; row < s->tiling.tile_row_end;

  3600.                  row += 8, yoff += ls_y * 64, uvoff += ls_uv * 32) {

  3601.                 struct VP9Filter *lflvl_ptr = s->lflvl;

  3602.                 ptrdiff_t yoff2 = yoff, uvoff2 = uvoff;

  3603. 

  3604.                 for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  3605.                     set_tile_offset(&s->tiling.tile_col_start, &s->tiling.tile_col_end,

  3606.                                     tile_col, s->tiling.log2_tile_cols, s->sb_cols);

  3607. 

  3608.                     if (s->pass != 2) {

  3609.                         memset(s->left_partition_ctx, 0, 8);

  3610.                         memset(s->left_skip_ctx, 0, 8);

  3611.                         if (s->keyframe || s->intraonly) {

  3612.                             memset(s->left_mode_ctx, DC_PRED, 16);

  3613.                         } else {

  3614.                             memset(s->left_mode_ctx, NEARESTMV, 8);

  3615.                         }

  3616.                         memset(s->left_y_nnz_ctx, 0, 16);

  3617.                         memset(s->left_uv_nnz_ctx, 0, 16);

  3618.                         memset(s->left_segpred_ctx, 0, 8);

  3619. 

  3620.                         memcpy(&s->c, &s->c_b[tile_col], sizeof(s->c));

  3621.                     }

  3622. 

  3623.                     for (col = s->tiling.tile_col_start;

  3624.                          col < s->tiling.tile_col_end;

  3625.                          col += 8, yoff2 += 64, uvoff2 += 32, lflvl_ptr++) {

  3626.                         // FIXME integrate with lf code (i.e. zero after each

  3627.                         // use, similar to invtxfm coefficients, or similar)

  3628.                         if (s->pass != 1) {

  3629.                             memset(lflvl_ptr->mask, 0, sizeof(lflvl_ptr->mask));

  3630.                         }

  3631. 

  3632.                         if (s->pass == 2) {

  3633.                             decode_sb_mem(ctx, row, col, lflvl_ptr,

  3634.                                           yoff2, uvoff2, BL_64X64);

  3635.                         } else {

  3636.                             decode_sb(ctx, row, col, lflvl_ptr,

  3637.                                       yoff2, uvoff2, BL_64X64);

  3638.                         }

  3639.                     }

  3640.                     if (s->pass != 2) {

  3641.                         memcpy(&s->c_b[tile_col], &s->c, sizeof(s->c));

  3642.                     }

  3643.                 }

  3644. 

  3645.                 if (s->pass == 1) {

  3646.                     continue;

  3647.                 }

  3648. 

  3649.                 // backup pre-loopfilter reconstruction data for intra

  3650.                 // prediction of next row of sb64s

  3651.                 if (row + 8 < s->rows) {

  3652.                     memcpy(s->intra_pred_data[0],

  3653.                            f->data[0] + yoff + 63 * ls_y,

  3654.                            8 * s->cols);

  3655.                     memcpy(s->intra_pred_data[1],

  3656.                            f->data[1] + uvoff + 31 * ls_uv,

  3657.                            4 * s->cols);

  3658.                     memcpy(s->intra_pred_data[2],

  3659.                            f->data[2] + uvoff + 31 * ls_uv,

  3660.                            4 * s->cols);

  3661.                 }

  3662. 

  3663.                 // loopfilter one row

  3664.                 if (s->filter.level) {

  3665.                     yoff2 = yoff;

  3666.                     uvoff2 = uvoff;

  3667.                     lflvl_ptr = s->lflvl;

  3668.                     for (col = 0; col < s->cols;

  3669.                          col += 8, yoff2 += 64, uvoff2 += 32, lflvl_ptr++) {

  3670.                         loopfilter_sb(ctx, lflvl_ptr, row, col, yoff2, uvoff2);

  3671.                     }

  3672.                 }

  3673. 

  3674.                 // FIXME maybe we can make this more finegrained by running the

  3675.                 // loopfilter per-block instead of after each sbrow

  3676.                 // In fact that would also make intra pred left preparation easier?

  3677.                 ff_thread_report_progress(&s->frames[CUR_FRAME].tf, row >> 3, 0);

  3678.             }

  3679.         }

  3680. 

  3681.         if (s->pass < 2 && s->refreshctx && !s->parallelmode) {

  3682.             adapt_probs(s);

  3683.             ff_thread_finish_setup(ctx);

  3684.         }

  3685.     } while (s->pass++ == 1);

  3686.     ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  3687. 

  3688.     // ref frame setup

  3689.     for (i = 0; i < 8; i++) {

  3690.         if (s->refs[i].f->data[0])

  3691.             ff_thread_release_buffer(ctx, &s->refs[i]);

  3692.         ff_thread_ref_frame(&s->refs[i], &s->next_refs[i]);

  3693.     }

  3694. 

  3695.     if (!s->invisible) {

  3696.         if ((res = av_frame_ref(frame, s->frames[CUR_FRAME].tf.f)) < 0)

  3697.             return res;

  3698.         *got_frame = 1;

  3699.     }

  3700. 

  3701.     return 0;

  3702. }

  3703. 

  3704. static void vp9_decode_flush(AVCodecContext *ctx)

  3705. {

  3706.     VP9Context *s = ctx->priv_data;

  3707.     int i;

  3708. 

  3709.     for (i = 0; i < 2; i++)

  3710.         vp9_unref_frame(ctx, &s->frames[i]);

  3711.     for (i = 0; i < 8; i++)

  3712.         ff_thread_release_buffer(ctx, &s->refs[i]);

  3713. }

  3714. 

  3715. static int init_frames(AVCodecContext *ctx)

  3716. {

  3717.     VP9Context *s = ctx->priv_data;

  3718.     int i;

  3719. 

  3720.     for (i = 0; i < 2; i++) {

  3721.         s->frames[i].tf.f = av_frame_alloc();

  3722.         if (!s->frames[i].tf.f) {

  3723.             vp9_decode_free(ctx);

  3724.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  3725.             return AVERROR(ENOMEM);

  3726.         }

  3727.     }

  3728.     for (i = 0; i < 8; i++) {

  3729.         s->refs[i].f = av_frame_alloc();

  3730.         s->next_refs[i].f = av_frame_alloc();

  3731.         if (!s->refs[i].f || !s->next_refs[i].f) {

  3732.             vp9_decode_free(ctx);

  3733.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  3734.             return AVERROR(ENOMEM);

  3735.         }

  3736.     }

  3737. 

  3738.     return 0;

  3739. }

  3740. 

  3741. static av_cold int vp9_decode_init(AVCodecContext *ctx)

  3742. {

  3743.     VP9Context *s = ctx->priv_data;

  3744. 

  3745.     ctx->internal->allocate_progress = 1;

  3746.     ctx->pix_fmt = AV_PIX_FMT_YUV420P;

  3747.     ff_vp9dsp_init(&s->dsp);

  3748.     ff_videodsp_init(&s->vdsp, 8);

  3749.     s->filter.sharpness = -1;

  3750. 

  3751.     return init_frames(ctx);

  3752. }

  3753. 

  3754. static av_cold int vp9_decode_init_thread_copy(AVCodecContext *avctx)

  3755. {

  3756.     return init_frames(avctx);

  3757. }

  3758. 

  3759. static int vp9_decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src)

  3760. {

  3761.     int i, res;

  3762.     VP9Context *s = dst->priv_data, *ssrc = src->priv_data;

  3763. 

  3764.     // FIXME scalability, size, etc.

  3765. 

  3766.     for (i = 0; i < 2; i++) {

  3767.         if (s->frames[i].tf.f->data[0])

  3768.             vp9_unref_frame(dst, &s->frames[i]);

  3769.         if (ssrc->frames[i].tf.f->data[0]) {

  3770.             if ((res = vp9_ref_frame(dst, &s->frames[i], &ssrc->frames[i])) < 0)

  3771.                 return res;

  3772.         }

  3773.     }

  3774.     for (i = 0; i < 8; i++) {

  3775.         if (s->refs[i].f->data[0])

  3776.             ff_thread_release_buffer(dst, &s->refs[i]);

  3777.         if (ssrc->next_refs[i].f->data[0]) {

  3778.             if ((res = ff_thread_ref_frame(&s->refs[i], &ssrc->next_refs[i])) < 0)

  3779.                 return res;

  3780.         }

  3781.     }

  3782. 

  3783.     s->invisible = ssrc->invisible;

  3784.     s->keyframe = ssrc->keyframe;

  3785.     s->uses_2pass = ssrc->uses_2pass;

  3786.     memcpy(&s->prob_ctx, &ssrc->prob_ctx, sizeof(s->prob_ctx));

  3787.     memcpy(&s->lf_delta, &ssrc->lf_delta, sizeof(s->lf_delta));

  3788.     if (ssrc->segmentation.enabled) {

  3789.         memcpy(&s->segmentation.feat, &ssrc->segmentation.feat,

  3790.                sizeof(s->segmentation.feat));

  3791.     }

  3792. 

  3793.     return 0;

  3794. }

  3795. 

  3796. AVCodec ff_vp9_decoder = {

  3797.     .name                  = "vp9",

  3798.     .long_name             = NULL_IF_CONFIG_SMALL("Google VP9"),

  3799.     .type                  = AVMEDIA_TYPE_VIDEO,

  3800.     .id                    = AV_CODEC_ID_VP9,

  3801.     .priv_data_size        = sizeof(VP9Context),

  3802.     .init                  = vp9_decode_init,

  3803.     .close                 = vp9_decode_free,

  3804.     .decode                = vp9_decode_frame,

  3805.     .capabilities          = CODEC_CAP_DR1 | CODEC_CAP_FRAME_THREADS,

  3806.     .flush                 = vp9_decode_flush,

  3807.     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(vp9_decode_init_thread_copy),

  3808.     .update_thread_context = ONLY_IF_THREADS_ENABLED(vp9_decode_update_thread_context),

  3809. };