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.
60461 Commits
32 Branches
326MB
Tree: 9aeca1c572
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '9aeca1c572'
${ noResults }
FFmpeg/libavcodec/vp9.c

3878 lines
154KB
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.     DECLARE_ALIGNED(16, uint8_t, left_y_nnz_ctx)[16];

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

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

  216.     DECLARE_ALIGNED(8, uint8_t, left_uv_nnz_ctx)[2][8];

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

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

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

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

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

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

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

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

  225.     uint8_t *above_partition_ctx;

  226.     uint8_t *above_mode_ctx;

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

  228.     uint8_t *above_y_nnz_ctx;

  229.     uint8_t *above_uv_nnz_ctx[2];

  230.     uint8_t *above_skip_ctx; // 1bit

  231.     uint8_t *above_txfm_ctx; // 2bit

  232.     uint8_t *above_segpred_ctx; // 1bit

  233.     uint8_t *above_intra_ctx; // 1bit

  234.     uint8_t *above_comp_ctx; // 1bit

  235.     uint8_t *above_ref_ctx; // 2bit

  236.     uint8_t *above_filter_ctx;

  237.     VP56mv (*above_mv_ctx)[2];

  238. 

  239.     // whole-frame cache

  240.     uint8_t *intra_pred_data[3];

  241.     struct VP9Filter *lflvl;

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

  243. 

  244.     // block reconstruction intermediates

  245.     int block_alloc_using_2pass;

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

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

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

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

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

  251. } VP9Context;

  252. 

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

  254.     {

  255.         { 16, 16 }, { 16, 8 }, { 8, 16 }, { 8, 8 }, { 8, 4 }, { 4, 8 },

  256.         { 4, 4 }, { 4, 2 }, { 2, 4 }, { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 },

  257.     }, {

  258.         { 8, 8 }, { 8, 4 }, { 4, 8 }, { 4, 4 }, { 4, 2 }, { 2, 4 },

  259.         { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 }, { 1, 1 }, { 1, 1 }, { 1, 1 },

  260.     }

  261. };

  262. 

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

  264. {

  265.     VP9Context *s = ctx->priv_data;

  266.     int ret, sz;

  267. 

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

  269.         return ret;

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

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

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

  273.         return AVERROR(ENOMEM);

  274.     }

  275. 

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

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

  278. 

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

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

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

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

  283.     }

  284. 

  285.     return 0;

  286. }

  287. 

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

  289. {

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

  291.     av_buffer_unref(&f->extradata);

  292. }

  293. 

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

  295. {

  296.     int res;

  297. 

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

  299.         return res;

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

  301.         vp9_unref_frame(ctx, dst);

  302.         return AVERROR(ENOMEM);

  303.     }

  304. 

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

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

  307. 

  308.     return 0;

  309. }

  310. 

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

  312. {

  313.     VP9Context *s = ctx->priv_data;

  314.     uint8_t *p;

  315. 

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

  317. 

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

  319.         return 0;

  320. 

  321.     ctx->width  = w;

  322.     ctx->height = h;

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

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

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

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

  327. 

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

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

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

  331.     if (!p)

  332.         return AVERROR(ENOMEM);

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  350. #undef assign

  351. 

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

  353.     av_freep(&s->b_base);

  354.     av_freep(&s->block_base);

  355. 

  356.     return 0;

  357. }

  358. 

  359. static int update_block_buffers(AVCodecContext *ctx)

  360. {

  361.     VP9Context *s = ctx->priv_data;

  362. 

  363.     if (s->b_base && s->block_base && s->block_alloc_using_2pass == s->uses_2pass)

  364.         return 0;

  365. 

  366.     av_free(s->b_base);

  367.     av_free(s->block_base);

  368.     if (s->uses_2pass) {

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

  370. 

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

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

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

  374.             return AVERROR(ENOMEM);

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

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

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

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

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

  380.     } else {

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

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

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

  384.             return AVERROR(ENOMEM);

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

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

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

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

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

  390.     }

  391.     s->block_alloc_using_2pass = s->uses_2pass;

  392. 

  393.     return 0;

  394. }

  395. 

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

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

  398. {

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

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

  401. }

  402. 

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

  404. {

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

  406. }

  407. 

  408. // differential forward probability updates

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

  410. {

  411.     static const int inv_map_table[254] = {

  412.           7,  20,  33,  46,  59,  72,  85,  98, 111, 124, 137, 150, 163, 176,

  413.         189, 202, 215, 228, 241, 254,   1,   2,   3,   4,   5,   6,   8,   9,

  414.          10,  11,  12,  13,  14,  15,  16,  17,  18,  19,  21,  22,  23,  24,

  415.          25,  26,  27,  28,  29,  30,  31,  32,  34,  35,  36,  37,  38,  39,

  416.          40,  41,  42,  43,  44,  45,  47,  48,  49,  50,  51,  52,  53,  54,

  417.          55,  56,  57,  58,  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,

  418.          70,  71,  73,  74,  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,

  419.          86,  87,  88,  89,  90,  91,  92,  93,  94,  95,  96,  97,  99, 100,

  420.         101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 112, 113, 114, 115,

  421.         116, 117, 118, 119, 120, 121, 122, 123, 125, 126, 127, 128, 129, 130,

  422.         131, 132, 133, 134, 135, 136, 138, 139, 140, 141, 142, 143, 144, 145,

  423.         146, 147, 148, 149, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160,

  424.         161, 162, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175,

  425.         177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 190, 191,

  426.         192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 204, 205, 206,

  427.         207, 208, 209, 210, 211, 212, 213, 214, 216, 217, 218, 219, 220, 221,

  428.         222, 223, 224, 225, 226, 227, 229, 230, 231, 232, 233, 234, 235, 236,

  429.         237, 238, 239, 240, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251,

  430.         252, 253,

  431.     };

  432.     int d;

  433. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  448. 

  449.     if (!vp8_rac_get(c)) {

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

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

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

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

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

  455.     } else {

  456.         d = vp8_rac_get_uint(c, 7);

  457.         if (d >= 65)

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

  459.         d += 64;

  460.     }

  461. 

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

  463.                     255 - inv_recenter_nonneg(inv_map_table[d], 255 - p);

  464. }

  465. 

  466. static int decode_frame_header(AVCodecContext *ctx,

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

  468. {

  469.     VP9Context *s = ctx->priv_data;

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

  471.     int last_invisible;

  472.     const uint8_t *data2;

  473. 

  474.     /* general header */

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

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

  477.         return res;

  478.     }

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

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

  481.         return AVERROR_INVALIDDATA;

  482.     }

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

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

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

  486.         return AVERROR_INVALIDDATA;

  487.     }

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

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

  490.         return 0;

  491.     }

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

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

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

  495.     last_invisible    = s->invisible;

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

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

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

  499.     if (s->keyframe) {

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

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

  502.             return AVERROR_INVALIDDATA;

  503.         }

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

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

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

  507.             return AVERROR_INVALIDDATA;

  508.         }

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

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

  511.         s->refreshrefmask = 0xff;

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

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

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

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

  516.     } else {

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

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

  519.         if (s->intraonly) {

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

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

  522.                 return AVERROR_INVALIDDATA;

  523.             }

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

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

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

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

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

  529.         } else {

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

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

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

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

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

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

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

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

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

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

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

  541.                 return AVERROR_INVALIDDATA;

  542.             }

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

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

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

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

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

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

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

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

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

  552.             } else {

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

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

  555.             }

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

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

  558.             // the _last_ frame

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

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

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

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

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

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

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

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

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

  568.             if (s->allowcompinter) {

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

  570.                     s->fixcompref    = 2;

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

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

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

  574.                     s->fixcompref    = 1;

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

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

  577.                 } else {

  578.                     s->fixcompref    = 0;

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

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

  581.                 }

  582.             }

  583.         }

  584.     }

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

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

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

  588. 

  589.     /* loopfilter header data */

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

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

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

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

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

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

  596.     s->filter.sharpness = sharp;

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

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

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

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

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

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

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

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

  605.         }

  606.     } else {

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

  608.     }

  609. 

  610.     /* quantization header data */

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

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

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

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

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

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

  617. 

  618.     /* segmentation header info */

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

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

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

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

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

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

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

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

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

  628.             }

  629.         }

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

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

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

  633.             av_log(ctx, AV_LOG_ERROR,

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

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

  636.             return AVERROR_INVALIDDATA;

  637.         }

  638. 

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

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

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

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

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

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

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

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

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

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

  649.             }

  650.         }

  651.     } else {

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

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

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

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

  656.     }

  657. 

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

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

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

  661. 

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

  663.             if (s->segmentation.absolute_vals)

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

  665.             else

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

  667.         } else {

  668.             qyac  = s->yac_qi;

  669.         }

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

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

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

  673.         qyac  = av_clip_uintp2(qyac, 8);

  674. 

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

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

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

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

  679. 

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

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

  682.             if (s->segmentation.absolute_vals)

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

  684.             else

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

  686.         } else {

  687.             lflvl  = s->filter.level;

  688.         }

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

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

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

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

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

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

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

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

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

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

  699.         }

  700.     }

  701. 

  702.     /* tiling info */

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

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

  705.         return res;

  706.     }

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

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

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

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

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

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

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

  714.             s->tiling.log2_tile_cols++;

  715.         else

  716.             break;

  717.     }

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

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

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

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

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

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

  724.         if (!s->c_b) {

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

  726.             return AVERROR(ENOMEM);

  727.         }

  728.     }

  729. 

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

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

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

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

  734.                sizeof(vp9_default_coef_probs));

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

  736.                sizeof(vp9_default_coef_probs));

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

  738.                sizeof(vp9_default_coef_probs));

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

  740.                sizeof(vp9_default_coef_probs));

  741.     }

  742. 

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

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

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

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

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

  748.         return AVERROR_INVALIDDATA;

  749.     }

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

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

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

  753.         return AVERROR_INVALIDDATA;

  754.     }

  755. 

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

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

  758.     } else {

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

  760.     }

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

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

  763.     // fw update)?

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

  765. 

  766.     // txfm updates

  767.     if (s->lossless) {

  768.         s->txfmmode = TX_4X4;

  769.     } else {

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

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

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

  773. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  788.         }

  789.     }

  790. 

  791.     // coef updates

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

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

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

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

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

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

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

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

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

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

  802.                                 break;

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

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

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

  806.                                 } else {

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

  808.                                 }

  809.                             }

  810.                             p[3] = 0;

  811.                         }

  812.         } else {

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

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

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

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

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

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

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

  820.                                 break;

  821.                             memcpy(p, r, 3);

  822.                             p[3] = 0;

  823.                         }

  824.         }

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

  826.             break;

  827.     }

  828. 

  829.     // mode updates

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

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

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

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

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

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

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

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

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

  839. 

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

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

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

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

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

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

  846. 

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

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

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

  850. 

  851.         if (s->allowcompinter) {

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

  853.             if (s->comppredmode)

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

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

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

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

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

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

  860.         } else {

  861.             s->comppredmode = PRED_SINGLEREF;

  862.         }

  863. 

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

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

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

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

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

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

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

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

  872.             }

  873.         }

  874. 

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

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

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

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

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

  880.         }

  881. 

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

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

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

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

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

  887. 

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

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

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

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

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

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

  894. 

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

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

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

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

  899. 

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

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

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

  903. 

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

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

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

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

  908. 

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

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

  911. 

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

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

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

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

  916.         }

  917. 

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

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

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

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

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

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

  924. 

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

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

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

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

  929.         }

  930. 

  931.         if (s->highprecisionmvs) {

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

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

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

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

  936. 

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

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

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

  940.             }

  941.         }

  942.     }

  943. 

  944.     return (data2 - data) + size2;

  945. }

  946. 

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

  948.                                       VP9Context *s)

  949. {

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

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

  952. }

  953. 

  954. static void find_ref_mvs(VP9Context *s,

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

  956. {

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

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

  959.                       { -1, -1 }, {  0, -1 }, { -1,  0 }, {  6, -1 }},

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

  961.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, {  2, -1 }},

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

  963.                       { -1, -1 }, { -3,  0 }, {  0, -3 }, { -1,  2 }},

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

  965.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  967.                       { -1,  1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  969.                       {  1, -1 }, { -3,  0 }, {  0, -3 }, { -3, -3 }},

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

  971.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  973.                       {  0, -2 }, { -2,  0 }, { -2, -1 }, { -1, -2 }},

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

  975.                       { -2,  0 }, {  0, -2 }, { -1, -2 }, { -2, -1 }},

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

  977.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

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

  979.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

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

  981.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

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

  983.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  984.     };

  985.     VP9Block *b = s->b;

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

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

  988. #define INVALID_MV 0x80008000U

  989.     uint32_t mem = INVALID_MV;

  990.     int i;

  991. 

  992. #define RETURN_DIRECT_MV(mv) \

  993.     do { \

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

  995.         if (!idx) { \

  996.             AV_WN32A(pmv, m); \

  997.             return; \

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

  999.             mem = m; \

  1000.         } else if (m != mem) { \

  1001.             AV_WN32A(pmv, m); \

  1002.             return; \

  1003.         } \

  1004.     } while (0)

  1005. 

  1006.     if (sb >= 0) {

  1007.         if (sb == 2 || sb == 1) {

  1008.             RETURN_DIRECT_MV(b->mv[0][z]);

  1009.         } else if (sb == 3) {

  1010.             RETURN_DIRECT_MV(b->mv[2][z]);

  1011.             RETURN_DIRECT_MV(b->mv[1][z]);

  1012.             RETURN_DIRECT_MV(b->mv[0][z]);

  1013.         }

  1014. 

  1015. #define RETURN_MV(mv) \

  1016.     do { \

  1017.         if (sb > 0) { \

  1018.             VP56mv tmp; \

  1019.             uint32_t m; \

  1020.             clamp_mv(&tmp, &mv, s); \

  1021.             m = AV_RN32A(&tmp); \

  1022.             if (!idx) { \

  1023.                 AV_WN32A(pmv, m); \

  1024.                 return; \

  1025.             } else if (mem == INVALID_MV) { \

  1026.                 mem = m; \

  1027.             } else if (m != mem) { \

  1028.                 AV_WN32A(pmv, m); \

  1029.                 return; \

  1030.             } \

  1031.         } else { \

  1032.             uint32_t m = AV_RN32A(&mv); \

  1033.             if (!idx) { \

  1034.                 clamp_mv(pmv, &mv, s); \

  1035.                 return; \

  1036.             } else if (mem == INVALID_MV) { \

  1037.                 mem = m; \

  1038.             } else if (m != mem) { \

  1039.                 clamp_mv(pmv, &mv, s); \

  1040.                 return; \

  1041.             } \

  1042.         } \

  1043.     } while (0)

  1044. 

  1045.         if (row > 0) {

  1046.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[(row - 1) * s->sb_cols * 8 + col];

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

  1048.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][0]);

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

  1050.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][1]);

  1051.             }

  1052.         }

  1053.         if (col > s->tiling.tile_col_start) {

  1054.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[row * s->sb_cols * 8 + col - 1];

  1055.             if (mv->ref[0] == ref) {

  1056.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][0]);

  1057.             } else if (mv->ref[1] == ref) {

  1058.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][1]);

  1059.             }

  1060.         }

  1061.         i = 2;

  1062.     } else {

  1063.         i = 0;

  1064.     }

  1065. 

  1066.     // previously coded MVs in this neighbourhood, using same reference frame

  1067.     for (; i < 8; i++) {

  1068.         int c = p[i][0] + col, r = p[i][1] + row;

  1069. 

  1070.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1071.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1072. 

  1073.             if (mv->ref[0] == ref) {

  1074.                 RETURN_MV(mv->mv[0]);

  1075.             } else if (mv->ref[1] == ref) {

  1076.                 RETURN_MV(mv->mv[1]);

  1077.             }

  1078.         }

  1079.     }

  1080. 

  1081.     // MV at this position in previous frame, using same 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.         if (!s->last_uses_2pass)

  1086.             ff_thread_await_progress(&s->frames[LAST_FRAME].tf, row >> 3, 0);

  1087.         if (mv->ref[0] == ref) {

  1088.             RETURN_MV(mv->mv[0]);

  1089.         } else if (mv->ref[1] == ref) {

  1090.             RETURN_MV(mv->mv[1]);

  1091.         }

  1092.     }

  1093. 

  1094. #define RETURN_SCALE_MV(mv, scale) \

  1095.     do { \

  1096.         if (scale) { \

  1097.             VP56mv mv_temp = { -mv.x, -mv.y }; \

  1098.             RETURN_MV(mv_temp); \

  1099.         } else { \

  1100.             RETURN_MV(mv); \

  1101.         } \

  1102.     } while (0)

  1103. 

  1104.     // previously coded MVs in this neighbourhood, using different reference frame

  1105.     for (i = 0; i < 8; i++) {

  1106.         int c = p[i][0] + col, r = p[i][1] + row;

  1107. 

  1108.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1109.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1110. 

  1111.             if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1112.                 RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1113.             }

  1114.             if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1115.                 // BUG - libvpx has this condition regardless of whether

  1116.                 // we used the first ref MV and pre-scaling

  1117.                 AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1118.                 RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1119.             }

  1120.         }

  1121.     }

  1122. 

  1123.     // MV at this position in previous frame, using different reference frame

  1124.     if (s->use_last_frame_mvs) {

  1125.         struct VP9mvrefPair *mv = &s->frames[LAST_FRAME].mv[row * s->sb_cols * 8 + col];

  1126. 

  1127.         // no need to await_progress, because we already did that above

  1128.         if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1129.             RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1130.         }

  1131.         if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1132.             // BUG - libvpx has this condition regardless of whether

  1133.             // we used the first ref MV and pre-scaling

  1134.             AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1135.             RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1136.         }

  1137.     }

  1138. 

  1139.     AV_ZERO32(pmv);

  1140. #undef INVALID_MV

  1141. #undef RETURN_MV

  1142. #undef RETURN_SCALE_MV

  1143. }

  1144. 

  1145. static av_always_inline int read_mv_component(VP9Context *s, int idx, int hp)

  1146. {

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

  1148.     int n, c = vp8_rac_get_tree(&s->c, vp9_mv_class_tree,

  1149.                                 s->prob.p.mv_comp[idx].classes);

  1150. 

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

  1152.     s->counts.mv_comp[idx].classes[c]++;

  1153.     if (c) {

  1154.         int m;

  1155. 

  1156.         for (n = 0, m = 0; m < c; m++) {

  1157.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].bits[m]);

  1158.             n |= bit << m;

  1159.             s->counts.mv_comp[idx].bits[m][bit]++;

  1160.         }

  1161.         n <<= 3;

  1162.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree, s->prob.p.mv_comp[idx].fp);

  1163.         n |= bit << 1;

  1164.         s->counts.mv_comp[idx].fp[bit]++;

  1165.         if (hp) {

  1166.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].hp);

  1167.             s->counts.mv_comp[idx].hp[bit]++;

  1168.             n |= bit;

  1169.         } else {

  1170.             n |= 1;

  1171.             // bug in libvpx - we count for bw entropy purposes even if the

  1172.             // bit wasn't coded

  1173.             s->counts.mv_comp[idx].hp[1]++;

  1174.         }

  1175.         n += 8 << c;

  1176.     } else {

  1177.         n = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0);

  1178.         s->counts.mv_comp[idx].class0[n]++;

  1179.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree,

  1180.                                s->prob.p.mv_comp[idx].class0_fp[n]);

  1181.         s->counts.mv_comp[idx].class0_fp[n][bit]++;

  1182.         n = (n << 3) | (bit << 1);

  1183.         if (hp) {

  1184.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0_hp);

  1185.             s->counts.mv_comp[idx].class0_hp[bit]++;

  1186.             n |= bit;

  1187.         } else {

  1188.             n |= 1;

  1189.             // bug in libvpx - we count for bw entropy purposes even if the

  1190.             // bit wasn't coded

  1191.             s->counts.mv_comp[idx].class0_hp[1]++;

  1192.         }

  1193.     }

  1194. 

  1195.     return sign ? -(n + 1) : (n + 1);

  1196. }

  1197. 

  1198. static void fill_mv(VP9Context *s,

  1199.                     VP56mv *mv, int mode, int sb)

  1200. {

  1201.     VP9Block *b = s->b;

  1202. 

  1203.     if (mode == ZEROMV) {

  1204.         memset(mv, 0, sizeof(*mv) * 2);

  1205.     } else {

  1206.         int hp;

  1207. 

  1208.         // FIXME cache this value and reuse for other subblocks

  1209.         find_ref_mvs(s, &mv[0], b->ref[0], 0, mode == NEARMV,

  1210.                      mode == NEWMV ? -1 : sb);

  1211.         // FIXME maybe move this code into find_ref_mvs()

  1212.         if ((mode == NEWMV || sb == -1) &&

  1213.             !(hp = s->highprecisionmvs && abs(mv[0].x) < 64 && abs(mv[0].y) < 64)) {

  1214.             if (mv[0].y & 1) {

  1215.                 if (mv[0].y < 0)

  1216.                     mv[0].y++;

  1217.                 else

  1218.                     mv[0].y--;

  1219.             }

  1220.             if (mv[0].x & 1) {

  1221.                 if (mv[0].x < 0)

  1222.                     mv[0].x++;

  1223.                 else

  1224.                     mv[0].x--;

  1225.             }

  1226.         }

  1227.         if (mode == NEWMV) {

  1228.             enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1229.                                               s->prob.p.mv_joint);

  1230. 

  1231.             s->counts.mv_joint[j]++;

  1232.             if (j >= MV_JOINT_V)

  1233.                 mv[0].y += read_mv_component(s, 0, hp);

  1234.             if (j & 1)

  1235.                 mv[0].x += read_mv_component(s, 1, hp);

  1236.         }

  1237. 

  1238.         if (b->comp) {

  1239.             // FIXME cache this value and reuse for other subblocks

  1240.             find_ref_mvs(s, &mv[1], b->ref[1], 1, mode == NEARMV,

  1241.                          mode == NEWMV ? -1 : sb);

  1242.             if ((mode == NEWMV || sb == -1) &&

  1243.                 !(hp = s->highprecisionmvs && abs(mv[1].x) < 64 && abs(mv[1].y) < 64)) {

  1244.                 if (mv[1].y & 1) {

  1245.                     if (mv[1].y < 0)

  1246.                         mv[1].y++;

  1247.                     else

  1248.                         mv[1].y--;

  1249.                 }

  1250.                 if (mv[1].x & 1) {

  1251.                     if (mv[1].x < 0)

  1252.                         mv[1].x++;

  1253.                     else

  1254.                         mv[1].x--;

  1255.                 }

  1256.             }

  1257.             if (mode == NEWMV) {

  1258.                 enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1259.                                                   s->prob.p.mv_joint);

  1260. 

  1261.                 s->counts.mv_joint[j]++;

  1262.                 if (j >= MV_JOINT_V)

  1263.                     mv[1].y += read_mv_component(s, 0, hp);

  1264.                 if (j & 1)

  1265.                     mv[1].x += read_mv_component(s, 1, hp);

  1266.             }

  1267.         }

  1268.     }

  1269. }

  1270. 

  1271. static void decode_mode(AVCodecContext *ctx)

  1272. {

  1273.     static const uint8_t left_ctx[N_BS_SIZES] = {

  1274.         0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf

  1275.     };

  1276.     static const uint8_t above_ctx[N_BS_SIZES] = {

  1277.         0x0, 0x0, 0x8, 0x8, 0x8, 0xc, 0xc, 0xc, 0xe, 0xe, 0xe, 0xf, 0xf

  1278.     };

  1279.     static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {

  1280.         TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,

  1281.         TX_16X16, TX_8X8, TX_8X8, TX_8X8, TX_4X4, TX_4X4, TX_4X4

  1282.     };

  1283.     VP9Context *s = ctx->priv_data;

  1284.     VP9Block *b = s->b;

  1285.     int row = s->row, col = s->col, row7 = s->row7;

  1286.     enum TxfmMode max_tx = max_tx_for_bl_bp[b->bs];

  1287.     int w4 = FFMIN(s->cols - col, bwh_tab[1][b->bs][0]);

  1288.     int h4 = FFMIN(s->rows - row, bwh_tab[1][b->bs][1]), y;

  1289.     int have_a = row > 0, have_l = col > s->tiling.tile_col_start;

  1290. 

  1291.     if (!s->segmentation.enabled) {

  1292.         b->seg_id = 0;

  1293.     } else if (s->keyframe || s->intraonly) {

  1294.         b->seg_id = vp8_rac_get_tree(&s->c, vp9_segmentation_tree, s->prob.seg);

  1295.     } else if (!s->segmentation.update_map ||

  1296.                (s->segmentation.temporal &&

  1297.                 vp56_rac_get_prob_branchy(&s->c,

  1298.                     s->prob.segpred[s->above_segpred_ctx[col] +

  1299.                                     s->left_segpred_ctx[row7]]))) {

  1300.         int pred = 8, x;

  1301.         uint8_t *refsegmap = s->frames[LAST_FRAME].segmentation_map;

  1302. 

  1303.         if (!s->last_uses_2pass)

  1304.             ff_thread_await_progress(&s->frames[LAST_FRAME].tf, row >> 3, 0);

  1305.         for (y = 0; y < h4; y++)

  1306.             for (x = 0; x < w4; x++)

  1307.                 pred = FFMIN(pred, refsegmap[(y + row) * 8 * s->sb_cols + x + col]);

  1308.         av_assert1(pred < 8);

  1309.         b->seg_id = pred;

  1310. 

  1311.         memset(&s->above_segpred_ctx[col], 1, w4);

  1312.         memset(&s->left_segpred_ctx[row7], 1, h4);

  1313.     } else {

  1314.         b->seg_id = vp8_rac_get_tree(&s->c, vp9_segmentation_tree,

  1315.                                      s->prob.seg);

  1316. 

  1317.         memset(&s->above_segpred_ctx[col], 0, w4);

  1318.         memset(&s->left_segpred_ctx[row7], 0, h4);

  1319.     }

  1320.     if ((s->segmentation.enabled && s->segmentation.update_map) || s->keyframe) {

  1321.         uint8_t *segmap = s->frames[CUR_FRAME].segmentation_map;

  1322. 

  1323.         for (y = 0; y < h4; y++)

  1324.             memset(&segmap[(y + row) * 8 * s->sb_cols + col], b->seg_id, w4);

  1325.     }

  1326. 

  1327.     b->skip = s->segmentation.enabled &&

  1328.         s->segmentation.feat[b->seg_id].skip_enabled;

  1329.     if (!b->skip) {

  1330.         int c = s->left_skip_ctx[row7] + s->above_skip_ctx[col];

  1331.         b->skip = vp56_rac_get_prob(&s->c, s->prob.p.skip[c]);

  1332.         s->counts.skip[c][b->skip]++;

  1333.     }

  1334. 

  1335.     if (s->keyframe || s->intraonly) {

  1336.         b->intra = 1;

  1337.     } else if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1338.         b->intra = !s->segmentation.feat[b->seg_id].ref_val;

  1339.     } else {

  1340.         int c, bit;

  1341. 

  1342.         if (have_a && have_l) {

  1343.             c = s->above_intra_ctx[col] + s->left_intra_ctx[row7];

  1344.             c += (c == 2);

  1345.         } else {

  1346.             c = have_a ? 2 * s->above_intra_ctx[col] :

  1347.                 have_l ? 2 * s->left_intra_ctx[row7] : 0;

  1348.         }

  1349.         bit = vp56_rac_get_prob(&s->c, s->prob.p.intra[c]);

  1350.         s->counts.intra[c][bit]++;

  1351.         b->intra = !bit;

  1352.     }

  1353. 

  1354.     if ((b->intra || !b->skip) && s->txfmmode == TX_SWITCHABLE) {

  1355.         int c;

  1356.         if (have_a) {

  1357.             if (have_l) {

  1358.                 c = (s->above_skip_ctx[col] ? max_tx :

  1359.                      s->above_txfm_ctx[col]) +

  1360.                     (s->left_skip_ctx[row7] ? max_tx :

  1361.                      s->left_txfm_ctx[row7]) > max_tx;

  1362.             } else {

  1363.                 c = s->above_skip_ctx[col] ? 1 :

  1364.                     (s->above_txfm_ctx[col] * 2 > max_tx);

  1365.             }

  1366.         } else if (have_l) {

  1367.             c = s->left_skip_ctx[row7] ? 1 :

  1368.                 (s->left_txfm_ctx[row7] * 2 > max_tx);

  1369.         } else {

  1370.             c = 1;

  1371.         }

  1372.         switch (max_tx) {

  1373.         case TX_32X32:

  1374.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][0]);

  1375.             if (b->tx) {

  1376.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][1]);

  1377.                 if (b->tx == 2)

  1378.                     b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][2]);

  1379.             }

  1380.             s->counts.tx32p[c][b->tx]++;

  1381.             break;

  1382.         case TX_16X16:

  1383.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][0]);

  1384.             if (b->tx)

  1385.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][1]);

  1386.             s->counts.tx16p[c][b->tx]++;

  1387.             break;

  1388.         case TX_8X8:

  1389.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx8p[c]);

  1390.             s->counts.tx8p[c][b->tx]++;

  1391.             break;

  1392.         case TX_4X4:

  1393.             b->tx = TX_4X4;

  1394.             break;

  1395.         }

  1396.     } else {

  1397.         b->tx = FFMIN(max_tx, s->txfmmode);

  1398.     }

  1399. 

  1400.     if (s->keyframe || s->intraonly) {

  1401.         uint8_t *a = &s->above_mode_ctx[col * 2];

  1402.         uint8_t *l = &s->left_mode_ctx[(row7) << 1];

  1403. 

  1404.         b->comp = 0;

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

  1406.             // FIXME the memory storage intermediates here aren't really

  1407.             // necessary, they're just there to make the code slightly

  1408.             // simpler for now

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

  1410.                                     vp9_default_kf_ymode_probs[a[0]][l[0]]);

  1411.             if (b->bs != BS_8x4) {

  1412.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1413.                                  vp9_default_kf_ymode_probs[a[1]][b->mode[0]]);

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

  1415.             } else {

  1416.                 l[0] = a[1] = b->mode[1] = b->mode[0];

  1417.             }

  1418.             if (b->bs != BS_4x8) {

  1419.                 b->mode[2] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1420.                                         vp9_default_kf_ymode_probs[a[0]][l[1]]);

  1421.                 if (b->bs != BS_8x4) {

  1422.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1423.                                   vp9_default_kf_ymode_probs[a[1]][b->mode[2]]);

  1424.                     l[1] = a[1] = b->mode[3];

  1425.                 } else {

  1426.                     l[1] = a[1] = b->mode[3] = b->mode[2];

  1427.                 }

  1428.             } else {

  1429.                 b->mode[2] = b->mode[0];

  1430.                 l[1] = a[1] = b->mode[3] = b->mode[1];

  1431.             }

  1432.         } else {

  1433.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1434.                                           vp9_default_kf_ymode_probs[*a][*l]);

  1435.             b->mode[3] = b->mode[2] = b->mode[1] = b->mode[0];

  1436.             // FIXME this can probably be optimized

  1437.             memset(a, b->mode[0], bwh_tab[0][b->bs][0]);

  1438.             memset(l, b->mode[0], bwh_tab[0][b->bs][1]);

  1439.         }

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

  1441.                                      vp9_default_kf_uvmode_probs[b->mode[3]]);

  1442.     } else if (b->intra) {

  1443.         b->comp = 0;

  1444.         if (b->bs > BS_8x8) {

  1445.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1446.                                           s->prob.p.y_mode[0]);

  1447.             s->counts.y_mode[0][b->mode[0]]++;

  1448.             if (b->bs != BS_8x4) {

  1449.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1450.                                               s->prob.p.y_mode[0]);

  1451.                 s->counts.y_mode[0][b->mode[1]]++;

  1452.             } else {

  1453.                 b->mode[1] = b->mode[0];

  1454.             }

  1455.             if (b->bs != BS_4x8) {

  1456.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1457.                                               s->prob.p.y_mode[0]);

  1458.                 s->counts.y_mode[0][b->mode[2]]++;

  1459.                 if (b->bs != BS_8x4) {

  1460.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1461.                                                   s->prob.p.y_mode[0]);

  1462.                     s->counts.y_mode[0][b->mode[3]]++;

  1463.                 } else {

  1464.                     b->mode[3] = b->mode[2];

  1465.                 }

  1466.             } else {

  1467.                 b->mode[2] = b->mode[0];

  1468.                 b->mode[3] = b->mode[1];

  1469.             }

  1470.         } else {

  1471.             static const uint8_t size_group[10] = {

  1472.                 3, 3, 3, 3, 2, 2, 2, 1, 1, 1

  1473.             };

  1474.             int sz = size_group[b->bs];

  1475. 

  1476.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1477.                                           s->prob.p.y_mode[sz]);

  1478.             b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1479.             s->counts.y_mode[sz][b->mode[3]]++;

  1480.         }

  1481.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1482.                                      s->prob.p.uv_mode[b->mode[3]]);

  1483.         s->counts.uv_mode[b->mode[3]][b->uvmode]++;

  1484.     } else {

  1485.         static const uint8_t inter_mode_ctx_lut[14][14] = {

  1486.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1487.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1488.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1489.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1490.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1491.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1492.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1493.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1494.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1495.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1496.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1497.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1498.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },

  1499.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },

  1500.         };

  1501. 

  1502.         if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1503.             av_assert2(s->segmentation.feat[b->seg_id].ref_val != 0);

  1504.             b->comp = 0;

  1505.             b->ref[0] = s->segmentation.feat[b->seg_id].ref_val - 1;

  1506.         } else {

  1507.             // read comp_pred flag

  1508.             if (s->comppredmode != PRED_SWITCHABLE) {

  1509.                 b->comp = s->comppredmode == PRED_COMPREF;

  1510.             } else {

  1511.                 int c;

  1512. 

  1513.                 // FIXME add intra as ref=0xff (or -1) to make these easier?

  1514.                 if (have_a) {

  1515.                     if (have_l) {

  1516.                         if (s->above_comp_ctx[col] && s->left_comp_ctx[row7]) {

  1517.                             c = 4;

  1518.                         } else if (s->above_comp_ctx[col]) {

  1519.                             c = 2 + (s->left_intra_ctx[row7] ||

  1520.                                      s->left_ref_ctx[row7] == s->fixcompref);

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

  1522.                             c = 2 + (s->above_intra_ctx[col] ||

  1523.                                      s->above_ref_ctx[col] == s->fixcompref);

  1524.                         } else {

  1525.                             c = (!s->above_intra_ctx[col] &&

  1526.                                  s->above_ref_ctx[col] == s->fixcompref) ^

  1527.                             (!s->left_intra_ctx[row7] &&

  1528.                              s->left_ref_ctx[row & 7] == s->fixcompref);

  1529.                         }

  1530.                     } else {

  1531.                         c = s->above_comp_ctx[col] ? 3 :

  1532.                         (!s->above_intra_ctx[col] && s->above_ref_ctx[col] == s->fixcompref);

  1533.                     }

  1534.                 } else if (have_l) {

  1535.                     c = s->left_comp_ctx[row7] ? 3 :

  1536.                     (!s->left_intra_ctx[row7] && s->left_ref_ctx[row7] == s->fixcompref);

  1537.                 } else {

  1538.                     c = 1;

  1539.                 }

  1540.                 b->comp = vp56_rac_get_prob(&s->c, s->prob.p.comp[c]);

  1541.                 s->counts.comp[c][b->comp]++;

  1542.             }

  1543. 

  1544.             // read actual references

  1545.             // FIXME probably cache a few variables here to prevent repetitive

  1546.             // memory accesses below

  1547.             if (b->comp) /* two references */ {

  1548.                 int fix_idx = s->signbias[s->fixcompref], var_idx = !fix_idx, c, bit;

  1549. 

  1550.                 b->ref[fix_idx] = s->fixcompref;

  1551.                 // FIXME can this codeblob be replaced by some sort of LUT?

  1552.                 if (have_a) {

  1553.                     if (have_l) {

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

  1555.                             if (s->left_intra_ctx[row7]) {

  1556.                                 c = 2;

  1557.                             } else {

  1558.                                 c = 1 + 2 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1559.                             }

  1560.                         } else if (s->left_intra_ctx[row7]) {

  1561.                             c = 1 + 2 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1562.                         } else {

  1563.                             int refl = s->left_ref_ctx[row7], refa = s->above_ref_ctx[col];

  1564. 

  1565.                             if (refl == refa && refa == s->varcompref[1]) {

  1566.                                 c = 0;

  1567.                             } else if (!s->left_comp_ctx[row7] && !s->above_comp_ctx[col]) {

  1568.                                 if ((refa == s->fixcompref && refl == s->varcompref[0]) ||

  1569.                                     (refl == s->fixcompref && refa == s->varcompref[0])) {

  1570.                                     c = 4;

  1571.                                 } else {

  1572.                                     c = (refa == refl) ? 3 : 1;

  1573.                                 }

  1574.                             } else if (!s->left_comp_ctx[row7]) {

  1575.                                 if (refa == s->varcompref[1] && refl != s->varcompref[1]) {

  1576.                                     c = 1;

  1577.                                 } else {

  1578.                                     c = (refl == s->varcompref[1] &&

  1579.                                          refa != s->varcompref[1]) ? 2 : 4;

  1580.                                 }

  1581.                             } else if (!s->above_comp_ctx[col]) {

  1582.                                 if (refl == s->varcompref[1] && refa != s->varcompref[1]) {

  1583.                                     c = 1;

  1584.                                 } else {

  1585.                                     c = (refa == s->varcompref[1] &&

  1586.                                          refl != s->varcompref[1]) ? 2 : 4;

  1587.                                 }

  1588.                             } else {

  1589.                                 c = (refl == refa) ? 4 : 2;

  1590.                             }

  1591.                         }

  1592.                     } else {

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

  1594.                             c = 2;

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

  1596.                             c = 4 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1597.                         } else {

  1598.                             c = 3 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1599.                         }

  1600.                     }

  1601.                 } else if (have_l) {

  1602.                     if (s->left_intra_ctx[row7]) {

  1603.                         c = 2;

  1604.                     } else if (s->left_comp_ctx[row7]) {

  1605.                         c = 4 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1606.                     } else {

  1607.                         c = 3 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1608.                     }

  1609.                 } else {

  1610.                     c = 2;

  1611.                 }

  1612.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.comp_ref[c]);

  1613.                 b->ref[var_idx] = s->varcompref[bit];

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

  1615.             } else /* single reference */ {

  1616.                 int bit, c;

  1617. 

  1618.                 if (have_a && !s->above_intra_ctx[col]) {

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

  1620.                         if (s->left_comp_ctx[row7]) {

  1621.                             if (s->above_comp_ctx[col]) {

  1622.                                 c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7] ||

  1623.                                          !s->above_ref_ctx[col]);

  1624.                             } else {

  1625.                                 c = (3 * !s->above_ref_ctx[col]) +

  1626.                                     (!s->fixcompref || !s->left_ref_ctx[row7]);

  1627.                             }

  1628.                         } else if (s->above_comp_ctx[col]) {

  1629.                             c = (3 * !s->left_ref_ctx[row7]) +

  1630.                                 (!s->fixcompref || !s->above_ref_ctx[col]);

  1631.                         } else {

  1632.                             c = 2 * !s->left_ref_ctx[row7] + 2 * !s->above_ref_ctx[col];

  1633.                         }

  1634.                     } else if (s->above_intra_ctx[col]) {

  1635.                         c = 2;

  1636.                     } else if (s->above_comp_ctx[col]) {

  1637.                         c = 1 + (!s->fixcompref || !s->above_ref_ctx[col]);

  1638.                     } else {

  1639.                         c = 4 * (!s->above_ref_ctx[col]);

  1640.                     }

  1641.                 } else if (have_l && !s->left_intra_ctx[row7]) {

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

  1643.                         c = 2;

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

  1645.                         c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7]);

  1646.                     } else {

  1647.                         c = 4 * (!s->left_ref_ctx[row7]);

  1648.                     }

  1649.                 } else {

  1650.                     c = 2;

  1651.                 }

  1652.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][0]);

  1653.                 s->counts.single_ref[c][0][bit]++;

  1654.                 if (!bit) {

  1655.                     b->ref[0] = 0;

  1656.                 } else {

  1657.                     // FIXME can this codeblob be replaced by some sort of LUT?

  1658.                     if (have_a) {

  1659.                         if (have_l) {

  1660.                             if (s->left_intra_ctx[row7]) {

  1661.                                 if (s->above_intra_ctx[col]) {

  1662.                                     c = 2;

  1663.                                 } else if (s->above_comp_ctx[col]) {

  1664.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1665.                                                  s->above_ref_ctx[col] == 1);

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

  1667.                                     c = 3;

  1668.                                 } else {

  1669.                                     c = 4 * (s->above_ref_ctx[col] == 1);

  1670.                                 }

  1671.                             } else if (s->above_intra_ctx[col]) {

  1672.                                 if (s->left_intra_ctx[row7]) {

  1673.                                     c = 2;

  1674.                                 } else if (s->left_comp_ctx[row7]) {

  1675.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1676.                                                  s->left_ref_ctx[row7] == 1);

  1677.                                 } else if (!s->left_ref_ctx[row7]) {

  1678.                                     c = 3;

  1679.                                 } else {

  1680.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1681.                                 }

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

  1683.                                 if (s->left_comp_ctx[row7]) {

  1684.                                     if (s->left_ref_ctx[row7] == s->above_ref_ctx[col]) {

  1685.                                         c = 3 * (s->fixcompref == 1 ||

  1686.                                                  s->left_ref_ctx[row7] == 1);

  1687.                                     } else {

  1688.                                         c = 2;

  1689.                                     }

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

  1691.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1692.                                                  s->above_ref_ctx[col] == 1);

  1693.                                 } else {

  1694.                                     c = 3 * (s->left_ref_ctx[row7] == 1) +

  1695.                                     (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1696.                                 }

  1697.                             } else if (s->left_comp_ctx[row7]) {

  1698.                                 if (!s->above_ref_ctx[col]) {

  1699.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1700.                                                  s->left_ref_ctx[row7] == 1);

  1701.                                 } else {

  1702.                                     c = 3 * (s->above_ref_ctx[col] == 1) +

  1703.                                     (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1704.                                 }

  1705.                             } else if (!s->above_ref_ctx[col]) {

  1706.                                 if (!s->left_ref_ctx[row7]) {

  1707.                                     c = 3;

  1708.                                 } else {

  1709.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1710.                                 }

  1711.                             } else if (!s->left_ref_ctx[row7]) {

  1712.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1713.                             } else {

  1714.                                 c = 2 * (s->left_ref_ctx[row7] == 1) +

  1715.                                 2 * (s->above_ref_ctx[col] == 1);

  1716.                             }

  1717.                         } else {

  1718.                             if (s->above_intra_ctx[col] ||

  1719.                                 (!s->above_comp_ctx[col] && !s->above_ref_ctx[col])) {

  1720.                                 c = 2;

  1721.                             } else if (s->above_comp_ctx[col]) {

  1722.                                 c = 3 * (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1723.                             } else {

  1724.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1725.                             }

  1726.                         }

  1727.                     } else if (have_l) {

  1728.                         if (s->left_intra_ctx[row7] ||

  1729.                             (!s->left_comp_ctx[row7] && !s->left_ref_ctx[row7])) {

  1730.                             c = 2;

  1731.                         } else if (s->left_comp_ctx[row7]) {

  1732.                             c = 3 * (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1733.                         } else {

  1734.                             c = 4 * (s->left_ref_ctx[row7] == 1);

  1735.                         }

  1736.                     } else {

  1737.                         c = 2;

  1738.                     }

  1739.                     bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][1]);

  1740.                     s->counts.single_ref[c][1][bit]++;

  1741.                     b->ref[0] = 1 + bit;

  1742.                 }

  1743.             }

  1744.         }

  1745. 

  1746.         if (b->bs <= BS_8x8) {

  1747.             if (s->segmentation.feat[b->seg_id].skip_enabled) {

  1748.                 b->mode[0] = b->mode[1] = b->mode[2] = b->mode[3] = ZEROMV;

  1749.             } else {

  1750.                 static const uint8_t off[10] = {

  1751.                     3, 0, 0, 1, 0, 0, 0, 0, 0, 0

  1752.                 };

  1753. 

  1754.                 // FIXME this needs to use the LUT tables from find_ref_mvs

  1755.                 // because not all are -1,0/0,-1

  1756.                 int c = inter_mode_ctx_lut[s->above_mode_ctx[col + off[b->bs]]]

  1757.                                           [s->left_mode_ctx[row7 + off[b->bs]]];

  1758. 

  1759.                 b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

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

  1761.                 b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1762.                 s->counts.mv_mode[c][b->mode[0] - 10]++;

  1763.             }

  1764.         }

  1765. 

  1766.         if (s->filtermode == FILTER_SWITCHABLE) {

  1767.             int c;

  1768. 

  1769.             if (have_a && s->above_mode_ctx[col] >= NEARESTMV) {

  1770.                 if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1771.                     c = s->above_filter_ctx[col] == s->left_filter_ctx[row7] ?

  1772.                         s->left_filter_ctx[row7] : 3;

  1773.                 } else {

  1774.                     c = s->above_filter_ctx[col];

  1775.                 }

  1776.             } else if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1777.                 c = s->left_filter_ctx[row7];

  1778.             } else {

  1779.                 c = 3;

  1780.             }

  1781. 

  1782.             b->filter = vp8_rac_get_tree(&s->c, vp9_filter_tree,

  1783.                                          s->prob.p.filter[c]);

  1784.             s->counts.filter[c][b->filter]++;

  1785.         } else {

  1786.             b->filter = s->filtermode;

  1787.         }

  1788. 

  1789.         if (b->bs > BS_8x8) {

  1790.             int c = inter_mode_ctx_lut[s->above_mode_ctx[col]][s->left_mode_ctx[row7]];

  1791. 

  1792.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1793.                                           s->prob.p.mv_mode[c]);

  1794.             s->counts.mv_mode[c][b->mode[0] - 10]++;

  1795.             fill_mv(s, b->mv[0], b->mode[0], 0);

  1796. 

  1797.             if (b->bs != BS_8x4) {

  1798.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1799.                                               s->prob.p.mv_mode[c]);

  1800.                 s->counts.mv_mode[c][b->mode[1] - 10]++;

  1801.                 fill_mv(s, b->mv[1], b->mode[1], 1);

  1802.             } else {

  1803.                 b->mode[1] = b->mode[0];

  1804.                 AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1805.                 AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1806.             }

  1807. 

  1808.             if (b->bs != BS_4x8) {

  1809.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1810.                                               s->prob.p.mv_mode[c]);

  1811.                 s->counts.mv_mode[c][b->mode[2] - 10]++;

  1812.                 fill_mv(s, b->mv[2], b->mode[2], 2);

  1813. 

  1814.                 if (b->bs != BS_8x4) {

  1815.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1816.                                                   s->prob.p.mv_mode[c]);

  1817.                     s->counts.mv_mode[c][b->mode[3] - 10]++;

  1818.                     fill_mv(s, b->mv[3], b->mode[3], 3);

  1819.                 } else {

  1820.                     b->mode[3] = b->mode[2];

  1821.                     AV_COPY32(&b->mv[3][0], &b->mv[2][0]);

  1822.                     AV_COPY32(&b->mv[3][1], &b->mv[2][1]);

  1823.                 }

  1824.             } else {

  1825.                 b->mode[2] = b->mode[0];

  1826.                 AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1827.                 AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1828.                 b->mode[3] = b->mode[1];

  1829.                 AV_COPY32(&b->mv[3][0], &b->mv[1][0]);

  1830.                 AV_COPY32(&b->mv[3][1], &b->mv[1][1]);

  1831.             }

  1832.         } else {

  1833.             fill_mv(s, b->mv[0], b->mode[0], -1);

  1834.             AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1835.             AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1836.             AV_COPY32(&b->mv[3][0], &b->mv[0][0]);

  1837.             AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1838.             AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1839.             AV_COPY32(&b->mv[3][1], &b->mv[0][1]);

  1840.         }

  1841.     }

  1842. 

  1843.     // FIXME this can probably be optimized

  1844.     memset(&s->above_skip_ctx[col], b->skip, w4);

  1845.     memset(&s->left_skip_ctx[row7], b->skip, h4);

  1846.     memset(&s->above_txfm_ctx[col], b->tx, w4);

  1847.     memset(&s->left_txfm_ctx[row7], b->tx, h4);

  1848.     memset(&s->above_partition_ctx[col], above_ctx[b->bs], w4);

  1849.     memset(&s->left_partition_ctx[row7], left_ctx[b->bs], h4);

  1850.     if (!s->keyframe && !s->intraonly) {

  1851.         memset(&s->above_intra_ctx[col], b->intra, w4);

  1852.         memset(&s->left_intra_ctx[row7], b->intra, h4);

  1853.         memset(&s->above_comp_ctx[col], b->comp, w4);

  1854.         memset(&s->left_comp_ctx[row7], b->comp, h4);

  1855.         memset(&s->above_mode_ctx[col], b->mode[3], w4);

  1856.         memset(&s->left_mode_ctx[row7], b->mode[3], h4);

  1857.         if (s->filtermode == FILTER_SWITCHABLE && !b->intra ) {

  1858.             memset(&s->above_filter_ctx[col], b->filter, w4);

  1859.             memset(&s->left_filter_ctx[row7], b->filter, h4);

  1860.             b->filter = vp9_filter_lut[b->filter];

  1861.         }

  1862.         if (b->bs > BS_8x8) {

  1863.             int mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  1864. 

  1865.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][0], &b->mv[1][0]);

  1866.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][1], &b->mv[1][1]);

  1867.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][0], mv0);

  1868.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][1], mv1);

  1869.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][0], &b->mv[2][0]);

  1870.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][1], &b->mv[2][1]);

  1871.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][0], mv0);

  1872.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][1], mv1);

  1873.         } else {

  1874.             int n, mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  1875. 

  1876.             for (n = 0; n < w4 * 2; n++) {

  1877.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][0], mv0);

  1878.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][1], mv1);

  1879.             }

  1880.             for (n = 0; n < h4 * 2; n++) {

  1881.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][0], mv0);

  1882.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][1], mv1);

  1883.             }

  1884.         }

  1885. 

  1886.         if (!b->intra) { // FIXME write 0xff or -1 if intra, so we can use this

  1887.                          // as a direct check in above branches

  1888.             int vref = b->ref[b->comp ? s->signbias[s->varcompref[0]] : 0];

  1889. 

  1890.             memset(&s->above_ref_ctx[col], vref, w4);

  1891.             memset(&s->left_ref_ctx[row7], vref, h4);

  1892.         }

  1893.     }

  1894. 

  1895.     // FIXME kinda ugly

  1896.     for (y = 0; y < h4; y++) {

  1897.         int x, o = (row + y) * s->sb_cols * 8 + col;

  1898.         struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[o];

  1899. 

  1900.         if (b->intra) {

  1901.             for (x = 0; x < w4; x++) {

  1902.                 mv[x].ref[0] =

  1903.                 mv[x].ref[1] = -1;

  1904.             }

  1905.         } else if (b->comp) {

  1906.             for (x = 0; x < w4; x++) {

  1907.                 mv[x].ref[0] = b->ref[0];

  1908.                 mv[x].ref[1] = b->ref[1];

  1909.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  1910.                 AV_COPY32(&mv[x].mv[1], &b->mv[3][1]);

  1911.             }

  1912.         } else {

  1913.             for (x = 0; x < w4; x++) {

  1914.                 mv[x].ref[0] = b->ref[0];

  1915.                 mv[x].ref[1] = -1;

  1916.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  1917.             }

  1918.         }

  1919.     }

  1920. }

  1921. 

  1922. // FIXME remove tx argument, and merge cnt/eob arguments?

  1923. static int decode_coeffs_b(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  1924.                            enum TxfmMode tx, unsigned (*cnt)[6][3],

  1925.                            unsigned (*eob)[6][2], uint8_t (*p)[6][11],

  1926.                            int nnz, const int16_t *scan, const int16_t (*nb)[2],

  1927.                            const int16_t *band_counts, const int16_t *qmul)

  1928. {

  1929.     int i = 0, band = 0, band_left = band_counts[band];

  1930.     uint8_t *tp = p[0][nnz];

  1931.     uint8_t cache[1024];

  1932. 

  1933.     do {

  1934.         int val, rc;

  1935. 

  1936.         val = vp56_rac_get_prob_branchy(c, tp[0]); // eob

  1937.         eob[band][nnz][val]++;

  1938.         if (!val)

  1939.             break;

  1940. 

  1941.     skip_eob:

  1942.         if (!vp56_rac_get_prob_branchy(c, tp[1])) { // zero

  1943.             cnt[band][nnz][0]++;

  1944.             if (!--band_left)

  1945.                 band_left = band_counts[++band];

  1946.             cache[scan[i]] = 0;

  1947.             nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  1948.             tp = p[band][nnz];

  1949.             if (++i == n_coeffs)

  1950.                 break; //invalid input; blocks should end with EOB

  1951.             goto skip_eob;

  1952.         }

  1953. 

  1954.         rc = scan[i];

  1955.         if (!vp56_rac_get_prob_branchy(c, tp[2])) { // one

  1956.             cnt[band][nnz][1]++;

  1957.             val = 1;

  1958.             cache[rc] = 1;

  1959.         } else {

  1960.             // fill in p[3-10] (model fill) - only once per frame for each pos

  1961.             if (!tp[3])

  1962.                 memcpy(&tp[3], vp9_model_pareto8[tp[2]], 8);

  1963. 

  1964.             cnt[band][nnz][2]++;

  1965.             if (!vp56_rac_get_prob_branchy(c, tp[3])) { // 2, 3, 4

  1966.                 if (!vp56_rac_get_prob_branchy(c, tp[4])) {

  1967.                     cache[rc] = val = 2;

  1968.                 } else {

  1969.                     val = 3 + vp56_rac_get_prob(c, tp[5]);

  1970.                     cache[rc] = 3;

  1971.                 }

  1972.             } else if (!vp56_rac_get_prob_branchy(c, tp[6])) { // cat1/2

  1973.                 cache[rc] = 4;

  1974.                 if (!vp56_rac_get_prob_branchy(c, tp[7])) {

  1975.                     val = 5 + vp56_rac_get_prob(c, 159);

  1976.                 } else {

  1977.                     val  = 7 + (vp56_rac_get_prob(c, 165) << 1);

  1978.                     val +=      vp56_rac_get_prob(c, 145);

  1979.                 }

  1980.             } else { // cat 3-6

  1981.                 cache[rc] = 5;

  1982.                 if (!vp56_rac_get_prob_branchy(c, tp[8])) {

  1983.                     if (!vp56_rac_get_prob_branchy(c, tp[9])) {

  1984.                         val  = 11 + (vp56_rac_get_prob(c, 173) << 2);

  1985.                         val +=      (vp56_rac_get_prob(c, 148) << 1);

  1986.                         val +=       vp56_rac_get_prob(c, 140);

  1987.                     } else {

  1988.                         val  = 19 + (vp56_rac_get_prob(c, 176) << 3);

  1989.                         val +=      (vp56_rac_get_prob(c, 155) << 2);

  1990.                         val +=      (vp56_rac_get_prob(c, 140) << 1);

  1991.                         val +=       vp56_rac_get_prob(c, 135);

  1992.                     }

  1993.                 } else if (!vp56_rac_get_prob_branchy(c, tp[10])) {

  1994.                     val  = 35 + (vp56_rac_get_prob(c, 180) << 4);

  1995.                     val +=      (vp56_rac_get_prob(c, 157) << 3);

  1996.                     val +=      (vp56_rac_get_prob(c, 141) << 2);

  1997.                     val +=      (vp56_rac_get_prob(c, 134) << 1);

  1998.                     val +=       vp56_rac_get_prob(c, 130);

  1999.                 } else {

  2000.                     val  = 67 + (vp56_rac_get_prob(c, 254) << 13);

  2001.                     val +=      (vp56_rac_get_prob(c, 254) << 12);

  2002.                     val +=      (vp56_rac_get_prob(c, 254) << 11);

  2003.                     val +=      (vp56_rac_get_prob(c, 252) << 10);

  2004.                     val +=      (vp56_rac_get_prob(c, 249) << 9);

  2005.                     val +=      (vp56_rac_get_prob(c, 243) << 8);

  2006.                     val +=      (vp56_rac_get_prob(c, 230) << 7);

  2007.                     val +=      (vp56_rac_get_prob(c, 196) << 6);

  2008.                     val +=      (vp56_rac_get_prob(c, 177) << 5);

  2009.                     val +=      (vp56_rac_get_prob(c, 153) << 4);

  2010.                     val +=      (vp56_rac_get_prob(c, 140) << 3);

  2011.                     val +=      (vp56_rac_get_prob(c, 133) << 2);

  2012.                     val +=      (vp56_rac_get_prob(c, 130) << 1);

  2013.                     val +=       vp56_rac_get_prob(c, 129);

  2014.                 }

  2015.             }

  2016.         }

  2017.         if (!--band_left)

  2018.             band_left = band_counts[++band];

  2019.         if (tx == TX_32X32) // FIXME slow

  2020.             coef[rc] = ((vp8_rac_get(c) ? -val : val) * qmul[!!i]) / 2;

  2021.         else

  2022.             coef[rc] = (vp8_rac_get(c) ? -val : val) * qmul[!!i];

  2023.         nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  2024.         tp = p[band][nnz];

  2025.     } while (++i < n_coeffs);

  2026. 

  2027.     return i;

  2028. }

  2029. 

  2030. static void decode_coeffs(AVCodecContext *ctx)

  2031. {

  2032.     VP9Context *s = ctx->priv_data;

  2033.     VP9Block *b = s->b;

  2034.     int row = s->row, col = s->col;

  2035.     uint8_t (*p)[6][11] = s->prob.coef[b->tx][0 /* y */][!b->intra];

  2036.     unsigned (*c)[6][3] = s->counts.coef[b->tx][0 /* y */][!b->intra];

  2037.     unsigned (*e)[6][2] = s->counts.eob[b->tx][0 /* y */][!b->intra];

  2038.     int w4 = bwh_tab[1][b->bs][0] << 1, h4 = bwh_tab[1][b->bs][1] << 1;

  2039.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2040.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2041.     int n, pl, x, y, step1d = 1 << b->tx, step = 1 << (b->tx * 2);

  2042.     int uvstep1d = 1 << b->uvtx, uvstep = 1 << (b->uvtx * 2), res;

  2043.     int16_t (*qmul)[2] = s->segmentation.feat[b->seg_id].qmul;

  2044.     int tx = 4 * s->lossless + b->tx;

  2045.     const int16_t * const *yscans = vp9_scans[tx];

  2046.     const int16_t (* const *ynbs)[2] = vp9_scans_nb[tx];

  2047.     const int16_t *uvscan = vp9_scans[b->uvtx][DCT_DCT];

  2048.     const int16_t (*uvnb)[2] = vp9_scans_nb[b->uvtx][DCT_DCT];

  2049.     uint8_t *a = &s->above_y_nnz_ctx[col * 2];

  2050.     uint8_t *l = &s->left_y_nnz_ctx[(row & 7) << 1];

  2051.     static const int16_t band_counts[4][8] = {

  2052.         { 1, 2, 3, 4,  3,   16 - 13 },

  2053.         { 1, 2, 3, 4, 11,   64 - 21 },

  2054.         { 1, 2, 3, 4, 11,  256 - 21 },

  2055.         { 1, 2, 3, 4, 11, 1024 - 21 },

  2056.     };

  2057.     const int16_t *y_band_counts = band_counts[b->tx];

  2058.     const int16_t *uv_band_counts = band_counts[b->uvtx];

  2059. 

  2060.     /* y tokens */

  2061.     if (b->tx > TX_4X4) { // FIXME slow

  2062.         for (y = 0; y < end_y; y += step1d)

  2063.             for (x = 1; x < step1d; x++)

  2064.                 l[y] |= l[y + x];

  2065.         for (x = 0; x < end_x; x += step1d)

  2066.             for (y = 1; y < step1d; y++)

  2067.                 a[x] |= a[x + y];

  2068.     }

  2069.     for (n = 0, y = 0; y < end_y; y += step1d) {

  2070.         for (x = 0; x < end_x; x += step1d, n += step) {

  2071.             enum TxfmType txtp = vp9_intra_txfm_type[b->mode[b->tx == TX_4X4 &&

  2072.                                                              b->bs > BS_8x8 ?

  2073.                                                              n : 0]];

  2074.             int nnz = a[x] + l[y];

  2075.             res = decode_coeffs_b(&s->c, s->block + 16 * n, 16 * step,

  2076.                                   b->tx, c, e, p, nnz, yscans[txtp],

  2077.                                   ynbs[txtp], y_band_counts, qmul[0]);

  2078.             a[x] = l[y] = !!res;

  2079.             if (b->tx > TX_8X8) {

  2080.                 AV_WN16A(&s->eob[n], res);

  2081.             } else {

  2082.                 s->eob[n] = res;

  2083.             }

  2084.         }

  2085.     }

  2086.     if (b->tx > TX_4X4) { // FIXME slow

  2087.         for (y = 0; y < end_y; y += step1d)

  2088.             memset(&l[y + 1], l[y], FFMIN(end_y - y - 1, step1d - 1));

  2089.         for (x = 0; x < end_x; x += step1d)

  2090.             memset(&a[x + 1], a[x], FFMIN(end_x - x - 1, step1d - 1));

  2091.     }

  2092. 

  2093.     p = s->prob.coef[b->uvtx][1 /* uv */][!b->intra];

  2094.     c = s->counts.coef[b->uvtx][1 /* uv */][!b->intra];

  2095.     e = s->counts.eob[b->uvtx][1 /* uv */][!b->intra];

  2096.     w4 >>= 1;

  2097.     h4 >>= 1;

  2098.     end_x >>= 1;

  2099.     end_y >>= 1;

  2100.     for (pl = 0; pl < 2; pl++) {

  2101.         a = &s->above_uv_nnz_ctx[pl][col];

  2102.         l = &s->left_uv_nnz_ctx[pl][row & 7];

  2103.         if (b->uvtx > TX_4X4) { // FIXME slow

  2104.             for (y = 0; y < end_y; y += uvstep1d)

  2105.                 for (x = 1; x < uvstep1d; x++)

  2106.                     l[y] |= l[y + x];

  2107.             for (x = 0; x < end_x; x += uvstep1d)

  2108.                 for (y = 1; y < uvstep1d; y++)

  2109.                     a[x] |= a[x + y];

  2110.         }

  2111.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2112.             for (x = 0; x < end_x; x += uvstep1d, n += uvstep) {

  2113.                 int nnz = a[x] + l[y];

  2114.                 res = decode_coeffs_b(&s->c, s->uvblock[pl] + 16 * n,

  2115.                                       16 * uvstep, b->uvtx, c, e, p, nnz,

  2116.                                       uvscan, uvnb, uv_band_counts, qmul[1]);

  2117.                 a[x] = l[y] = !!res;

  2118.                 if (b->uvtx > TX_8X8) {

  2119.                     AV_WN16A(&s->uveob[pl][n], res);

  2120.                 } else {

  2121.                     s->uveob[pl][n] = res;

  2122.                 }

  2123.             }

  2124.         }

  2125.         if (b->uvtx > TX_4X4) { // FIXME slow

  2126.             for (y = 0; y < end_y; y += uvstep1d)

  2127.                 memset(&l[y + 1], l[y], FFMIN(end_y - y - 1, uvstep1d - 1));

  2128.             for (x = 0; x < end_x; x += uvstep1d)

  2129.                 memset(&a[x + 1], a[x], FFMIN(end_x - x - 1, uvstep1d - 1));

  2130.         }

  2131.     }

  2132. }

  2133. 

  2134. static av_always_inline int check_intra_mode(VP9Context *s, int mode, uint8_t **a,

  2135.                                              uint8_t *dst_edge, ptrdiff_t stride_edge,

  2136.                                              uint8_t *dst_inner, ptrdiff_t stride_inner,

  2137.                                              uint8_t *l, int col, int x, int w,

  2138.                                              int row, int y, enum TxfmMode tx,

  2139.                                              int p)

  2140. {

  2141.     int have_top = row > 0 || y > 0;

  2142.     int have_left = col > s->tiling.tile_col_start || x > 0;

  2143.     int have_right = x < w - 1;

  2144.     static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {

  2145.         [VERT_PRED]            = { { DC_127_PRED,          VERT_PRED },

  2146.                                    { DC_127_PRED,          VERT_PRED } },

  2147.         [HOR_PRED]             = { { DC_129_PRED,          DC_129_PRED },

  2148.                                    { HOR_PRED,             HOR_PRED } },

  2149.         [DC_PRED]              = { { DC_128_PRED,          TOP_DC_PRED },

  2150.                                    { LEFT_DC_PRED,         DC_PRED } },

  2151.         [DIAG_DOWN_LEFT_PRED]  = { { DC_127_PRED,          DIAG_DOWN_LEFT_PRED },

  2152.                                    { DC_127_PRED,          DIAG_DOWN_LEFT_PRED } },

  2153.         [DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },

  2154.                                    { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },

  2155.         [VERT_RIGHT_PRED]      = { { VERT_RIGHT_PRED,      VERT_RIGHT_PRED },

  2156.                                    { VERT_RIGHT_PRED,      VERT_RIGHT_PRED } },

  2157.         [HOR_DOWN_PRED]        = { { HOR_DOWN_PRED,        HOR_DOWN_PRED },

  2158.                                    { HOR_DOWN_PRED,        HOR_DOWN_PRED } },

  2159.         [VERT_LEFT_PRED]       = { { DC_127_PRED,          VERT_LEFT_PRED },

  2160.                                    { DC_127_PRED,          VERT_LEFT_PRED } },

  2161.         [HOR_UP_PRED]          = { { DC_129_PRED,          DC_129_PRED },

  2162.                                    { HOR_UP_PRED,          HOR_UP_PRED } },

  2163.         [TM_VP8_PRED]          = { { DC_129_PRED,          VERT_PRED },

  2164.                                    { HOR_PRED,             TM_VP8_PRED } },

  2165.     };

  2166.     static const struct {

  2167.         uint8_t needs_left:1;

  2168.         uint8_t needs_top:1;

  2169.         uint8_t needs_topleft:1;

  2170.         uint8_t needs_topright:1;

  2171.     } edges[N_INTRA_PRED_MODES] = {

  2172.         [VERT_PRED]            = { .needs_top  = 1 },

  2173.         [HOR_PRED]             = { .needs_left = 1 },

  2174.         [DC_PRED]              = { .needs_top  = 1, .needs_left = 1 },

  2175.         [DIAG_DOWN_LEFT_PRED]  = { .needs_top  = 1, .needs_topright = 1 },

  2176.         [DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2177.         [VERT_RIGHT_PRED]      = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2178.         [HOR_DOWN_PRED]        = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2179.         [VERT_LEFT_PRED]       = { .needs_top  = 1, .needs_topright = 1 },

  2180.         [HOR_UP_PRED]          = { .needs_left = 1 },

  2181.         [TM_VP8_PRED]          = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2182.         [LEFT_DC_PRED]         = { .needs_left = 1 },

  2183.         [TOP_DC_PRED]          = { .needs_top  = 1 },

  2184.         [DC_128_PRED]          = { 0 },

  2185.         [DC_127_PRED]          = { 0 },

  2186.         [DC_129_PRED]          = { 0 }

  2187.     };

  2188. 

  2189.     av_assert2(mode >= 0 && mode < 10);

  2190.     mode = mode_conv[mode][have_left][have_top];

  2191.     if (edges[mode].needs_top) {

  2192.         uint8_t *top, *topleft;

  2193.         int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !p) - x) * 4;

  2194.         int n_px_need_tr = 0;

  2195. 

  2196.         if (tx == TX_4X4 && edges[mode].needs_topright && have_right)

  2197.             n_px_need_tr = 4;

  2198. 

  2199.         // if top of sb64-row, use s->intra_pred_data[] instead of

  2200.         // dst[-stride] for intra prediction (it contains pre- instead of

  2201.         // post-loopfilter data)

  2202.         if (have_top) {

  2203.             top = !(row & 7) && !y ?

  2204.                 s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  2205.                 y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];

  2206.             if (have_left)

  2207.                 topleft = !(row & 7) && !y ?

  2208.                     s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  2209.                     y == 0 || x == 0 ? &dst_edge[-stride_edge] :

  2210.                     &dst_inner[-stride_inner];

  2211.         }

  2212. 

  2213.         if (have_top &&

  2214.             (!edges[mode].needs_topleft || (have_left && top == topleft)) &&

  2215.             (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&

  2216.             n_px_need + n_px_need_tr <= n_px_have) {

  2217.             *a = top;

  2218.         } else {

  2219.             if (have_top) {

  2220.                 if (n_px_need <= n_px_have) {

  2221.                     memcpy(*a, top, n_px_need);

  2222.                 } else {

  2223.                     memcpy(*a, top, n_px_have);

  2224.                     memset(&(*a)[n_px_have], (*a)[n_px_have - 1],

  2225.                            n_px_need - n_px_have);

  2226.                 }

  2227.             } else {

  2228.                 memset(*a, 127, n_px_need);

  2229.             }

  2230.             if (edges[mode].needs_topleft) {

  2231.                 if (have_left && have_top) {

  2232.                     (*a)[-1] = topleft[-1];

  2233.                 } else {

  2234.                     (*a)[-1] = have_top ? 129 : 127;

  2235.                 }

  2236.             }

  2237.             if (tx == TX_4X4 && edges[mode].needs_topright) {

  2238.                 if (have_top && have_right &&

  2239.                     n_px_need + n_px_need_tr <= n_px_have) {

  2240.                     memcpy(&(*a)[4], &top[4], 4);

  2241.                 } else {

  2242.                     memset(&(*a)[4], (*a)[3], 4);

  2243.                 }

  2244.             }

  2245.         }

  2246.     }

  2247.     if (edges[mode].needs_left) {

  2248.         if (have_left) {

  2249.             int n_px_need = 4 << tx, i, n_px_have = (((s->rows - row) << !p) - y) * 4;

  2250.             uint8_t *dst = x == 0 ? dst_edge : dst_inner;

  2251.             ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;

  2252. 

  2253.             if (n_px_need <= n_px_have) {

  2254.                 for (i = 0; i < n_px_need; i++)

  2255.                     l[i] = dst[i * stride - 1];

  2256.             } else {

  2257.                 for (i = 0; i < n_px_have; i++)

  2258.                     l[i] = dst[i * stride - 1];

  2259.                 memset(&l[i], l[i - 1], n_px_need - n_px_have);

  2260.             }

  2261.         } else {

  2262.             memset(l, 129, 4 << tx);

  2263.         }

  2264.     }

  2265. 

  2266.     return mode;

  2267. }

  2268. 

  2269. static void intra_recon(AVCodecContext *ctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  2270. {

  2271.     VP9Context *s = ctx->priv_data;

  2272.     VP9Block *b = s->b;

  2273.     int row = s->row, col = s->col;

  2274.     int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2275.     int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2276.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2277.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2278.     int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2279.     int uvstep1d = 1 << b->uvtx, p;

  2280.     uint8_t *dst = s->dst[0], *dst_r = s->frames[CUR_FRAME].tf.f->data[0] + y_off;

  2281. 

  2282.     for (n = 0, y = 0; y < end_y; y += step1d) {

  2283.         uint8_t *ptr = dst, *ptr_r = dst_r;

  2284.         for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d,

  2285.                                ptr_r += 4 * step1d, n += step) {

  2286.             int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?

  2287.                                y * 2 + x : 0];

  2288.             LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);

  2289.             uint8_t *a = &a_buf[16], l[32];

  2290.             enum TxfmType txtp = vp9_intra_txfm_type[mode];

  2291.             int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2292. 

  2293.             mode = check_intra_mode(s, mode, &a, ptr_r,

  2294.                                     s->frames[CUR_FRAME].tf.f->linesize[0],

  2295.                                     ptr, s->y_stride, l,

  2296.                                     col, x, w4, row, y, b->tx, 0);

  2297.             s->dsp.intra_pred[b->tx][mode](ptr, s->y_stride, l, a);

  2298.             if (eob)

  2299.                 s->dsp.itxfm_add[tx][txtp](ptr, s->y_stride,

  2300.                                            s->block + 16 * n, eob);

  2301.         }

  2302.         dst_r += 4 * step1d * s->frames[CUR_FRAME].tf.f->linesize[0];

  2303.         dst   += 4 * step1d * s->y_stride;

  2304.     }

  2305. 

  2306.     // U/V

  2307.     h4 >>= 1;

  2308.     w4 >>= 1;

  2309.     end_x >>= 1;

  2310.     end_y >>= 1;

  2311.     step = 1 << (b->uvtx * 2);

  2312.     for (p = 0; p < 2; p++) {

  2313.         dst   = s->dst[1 + p];

  2314.         dst_r = s->frames[CUR_FRAME].tf.f->data[1 + p] + uv_off;

  2315.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2316.             uint8_t *ptr = dst, *ptr_r = dst_r;

  2317.             for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d,

  2318.                                    ptr_r += 4 * uvstep1d, n += step) {

  2319.                 int mode = b->uvmode;

  2320.                 LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);

  2321.                 uint8_t *a = &a_buf[16], l[32];

  2322.                 int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2323. 

  2324.                 mode = check_intra_mode(s, mode, &a, ptr_r,

  2325.                                         s->frames[CUR_FRAME].tf.f->linesize[1],

  2326.                                         ptr, s->uv_stride, l,

  2327.                                         col, x, w4, row, y, b->uvtx, p + 1);

  2328.                 s->dsp.intra_pred[b->uvtx][mode](ptr, s->uv_stride, l, a);

  2329.                 if (eob)

  2330.                     s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2331.                                                     s->uvblock[p] + 16 * n, eob);

  2332.             }

  2333.             dst_r += 4 * uvstep1d * s->frames[CUR_FRAME].tf.f->linesize[1];

  2334.             dst   += 4 * uvstep1d * s->uv_stride;

  2335.         }

  2336.     }

  2337. }

  2338. 

  2339. static av_always_inline void mc_luma_dir(VP9Context *s, vp9_mc_func (*mc)[2],

  2340.                                          uint8_t *dst, ptrdiff_t dst_stride,

  2341.                                          const uint8_t *ref, ptrdiff_t ref_stride,

  2342.                                          ThreadFrame *ref_frame,

  2343.                                          ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2344.                                          int bw, int bh, int w, int h)

  2345. {

  2346.     int mx = mv->x, my = mv->y, th;

  2347. 

  2348.     y += my >> 3;

  2349.     x += mx >> 3;

  2350.     ref += y * ref_stride + x;

  2351.     mx &= 7;

  2352.     my &= 7;

  2353.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2354.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2355.     // the longest loopfilter of the next sbrow

  2356.     th = (y + bh + 4 * !!my + 7) >> 6;

  2357.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2358.     if (x < !!mx * 3 || y < !!my * 3 ||

  2359.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2360.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2361.                                  ref - !!my * 3 * ref_stride - !!mx * 3,

  2362.                                  80, ref_stride,

  2363.                                  bw + !!mx * 7, bh + !!my * 7,

  2364.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2365.         ref = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2366.         ref_stride = 80;

  2367.     }

  2368.     mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);

  2369. }

  2370. 

  2371. static av_always_inline void mc_chroma_dir(VP9Context *s, vp9_mc_func (*mc)[2],

  2372.                                            uint8_t *dst_u, uint8_t *dst_v,

  2373.                                            ptrdiff_t dst_stride,

  2374.                                            const uint8_t *ref_u, ptrdiff_t src_stride_u,

  2375.                                            const uint8_t *ref_v, ptrdiff_t src_stride_v,

  2376.                                            ThreadFrame *ref_frame,

  2377.                                            ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2378.                                            int bw, int bh, int w, int h)

  2379. {

  2380.     int mx = mv->x, my = mv->y, th;

  2381. 

  2382.     y += my >> 4;

  2383.     x += mx >> 4;

  2384.     ref_u += y * src_stride_u + x;

  2385.     ref_v += y * src_stride_v + x;

  2386.     mx &= 15;

  2387.     my &= 15;

  2388.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2389.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2390.     // the longest loopfilter of the next sbrow

  2391.     th = (y + bh + 4 * !!my + 7) >> 5;

  2392.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2393.     if (x < !!mx * 3 || y < !!my * 3 ||

  2394.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2395.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2396.                                  ref_u - !!my * 3 * src_stride_u - !!mx * 3,

  2397.                                  80, src_stride_u,

  2398.                                  bw + !!mx * 7, bh + !!my * 7,

  2399.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2400.         ref_u = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2401.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, 80, bh, mx, my);

  2402. 

  2403.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2404.                                  ref_v - !!my * 3 * src_stride_v - !!mx * 3,

  2405.                                  80, src_stride_v,

  2406.                                  bw + !!mx * 7, bh + !!my * 7,

  2407.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2408.         ref_v = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2409.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, 80, bh, mx, my);

  2410.     } else {

  2411.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);

  2412.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);

  2413.     }

  2414. }

  2415. 

  2416. static void inter_recon(AVCodecContext *ctx)

  2417. {

  2418.     static const uint8_t bwlog_tab[2][N_BS_SIZES] = {

  2419.         { 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4 },

  2420.         { 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 4, 4 },

  2421.     };

  2422.     VP9Context *s = ctx->priv_data;

  2423.     VP9Block *b = s->b;

  2424.     int row = s->row, col = s->col;

  2425.     ThreadFrame *tref1 = &s->refs[s->refidx[b->ref[0]]], *tref2;

  2426.     AVFrame *ref1 = tref1->f, *ref2;

  2427.     int w1 = ref1->width, h1 = ref1->height, w2, h2;

  2428.     ptrdiff_t ls_y = s->y_stride, ls_uv = s->uv_stride;

  2429. 

  2430.     if (b->comp) {

  2431.         tref2 = &s->refs[s->refidx[b->ref[1]]];

  2432.         ref2 = tref2->f;

  2433.         w2 = ref2->width;

  2434.         h2 = ref2->height;

  2435.     }

  2436. 

  2437.     // y inter pred

  2438.     if (b->bs > BS_8x8) {

  2439.         if (b->bs == BS_8x4) {

  2440.             mc_luma_dir(s, s->dsp.mc[3][b->filter][0], s->dst[0], ls_y,

  2441.                         ref1->data[0], ref1->linesize[0], tref1,

  2442.                         row << 3, col << 3, &b->mv[0][0], 8, 4, w1, h1);

  2443.             mc_luma_dir(s, s->dsp.mc[3][b->filter][0],

  2444.                         s->dst[0] + 4 * ls_y, ls_y,

  2445.                         ref1->data[0], ref1->linesize[0], tref1,

  2446.                         (row << 3) + 4, col << 3, &b->mv[2][0], 8, 4, w1, h1);

  2447. 

  2448.             if (b->comp) {

  2449.                 mc_luma_dir(s, s->dsp.mc[3][b->filter][1], s->dst[0], ls_y,

  2450.                             ref2->data[0], ref2->linesize[0], tref2,

  2451.                             row << 3, col << 3, &b->mv[0][1], 8, 4, w2, h2);

  2452.                 mc_luma_dir(s, s->dsp.mc[3][b->filter][1],

  2453.                             s->dst[0] + 4 * ls_y, ls_y,

  2454.                             ref2->data[0], ref2->linesize[0], tref2,

  2455.                             (row << 3) + 4, col << 3, &b->mv[2][1], 8, 4, w2, h2);

  2456.             }

  2457.         } else if (b->bs == BS_4x8) {

  2458.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0], ls_y,

  2459.                         ref1->data[0], ref1->linesize[0], tref1,

  2460.                         row << 3, col << 3, &b->mv[0][0], 4, 8, w1, h1);

  2461.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0] + 4, ls_y,

  2462.                         ref1->data[0], ref1->linesize[0], tref1,

  2463.                         row << 3, (col << 3) + 4, &b->mv[1][0], 4, 8, w1, h1);

  2464. 

  2465.             if (b->comp) {

  2466.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0], ls_y,

  2467.                             ref2->data[0], ref2->linesize[0], tref2,

  2468.                             row << 3, col << 3, &b->mv[0][1], 4, 8, w2, h2);

  2469.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0] + 4, ls_y,

  2470.                             ref2->data[0], ref2->linesize[0], tref2,

  2471.                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 8, w2, h2);

  2472.             }

  2473.         } else {

  2474.             av_assert2(b->bs == BS_4x4);

  2475. 

  2476.             // FIXME if two horizontally adjacent blocks have the same MV,

  2477.             // do a w8 instead of a w4 call

  2478.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0], ls_y,

  2479.                         ref1->data[0], ref1->linesize[0], tref1,

  2480.                         row << 3, col << 3, &b->mv[0][0], 4, 4, w1, h1);

  2481.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0] + 4, ls_y,

  2482.                         ref1->data[0], ref1->linesize[0], tref1,

  2483.                         row << 3, (col << 3) + 4, &b->mv[1][0], 4, 4, w1, h1);

  2484.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0],

  2485.                         s->dst[0] + 4 * ls_y, ls_y,

  2486.                         ref1->data[0], ref1->linesize[0], tref1,

  2487.                         (row << 3) + 4, col << 3, &b->mv[2][0], 4, 4, w1, h1);

  2488.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0],

  2489.                         s->dst[0] + 4 * ls_y + 4, ls_y,

  2490.                         ref1->data[0], ref1->linesize[0], tref1,

  2491.                         (row << 3) + 4, (col << 3) + 4, &b->mv[3][0], 4, 4, w1, h1);

  2492. 

  2493.             if (b->comp) {

  2494.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0], ls_y,

  2495.                             ref2->data[0], ref2->linesize[0], tref2,

  2496.                             row << 3, col << 3, &b->mv[0][1], 4, 4, w2, h2);

  2497.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0] + 4, ls_y,

  2498.                             ref2->data[0], ref2->linesize[0], tref2,

  2499.                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 4, w2, h2);

  2500.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1],

  2501.                             s->dst[0] + 4 * ls_y, ls_y,

  2502.                             ref2->data[0], ref2->linesize[0], tref2,

  2503.                             (row << 3) + 4, col << 3, &b->mv[2][1], 4, 4, w2, h2);

  2504.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1],

  2505.                             s->dst[0] + 4 * ls_y + 4, ls_y,

  2506.                             ref2->data[0], ref2->linesize[0], tref2,

  2507.                             (row << 3) + 4, (col << 3) + 4, &b->mv[3][1], 4, 4, w2, h2);

  2508.             }

  2509.         }

  2510.     } else {

  2511.         int bwl = bwlog_tab[0][b->bs];

  2512.         int bw = bwh_tab[0][b->bs][0] * 4, bh = bwh_tab[0][b->bs][1] * 4;

  2513. 

  2514.         mc_luma_dir(s, s->dsp.mc[bwl][b->filter][0], s->dst[0], ls_y,

  2515.                     ref1->data[0], ref1->linesize[0], tref1,

  2516.                     row << 3, col << 3, &b->mv[0][0],bw, bh, w1, h1);

  2517. 

  2518.         if (b->comp)

  2519.             mc_luma_dir(s, s->dsp.mc[bwl][b->filter][1], s->dst[0], ls_y,

  2520.                         ref2->data[0], ref2->linesize[0], tref2,

  2521.                         row << 3, col << 3, &b->mv[0][1], bw, bh, w2, h2);

  2522.     }

  2523. 

  2524.     // uv inter pred

  2525.     {

  2526.         int bwl = bwlog_tab[1][b->bs];

  2527.         int bw = bwh_tab[1][b->bs][0] * 4, bh = bwh_tab[1][b->bs][1] * 4;

  2528.         VP56mv mvuv;

  2529. 

  2530.         w1 = (w1 + 1) >> 1;

  2531.         h1 = (h1 + 1) >> 1;

  2532.         if (b->comp) {

  2533.             w2 = (w2 + 1) >> 1;

  2534.             h2 = (h2 + 1) >> 1;

  2535.         }

  2536.         if (b->bs > BS_8x8) {

  2537.             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);

  2538.             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);

  2539.         } else {

  2540.             mvuv = b->mv[0][0];

  2541.         }

  2542. 

  2543.         mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][0],

  2544.                       s->dst[1], s->dst[2], ls_uv,

  2545.                       ref1->data[1], ref1->linesize[1],

  2546.                       ref1->data[2], ref1->linesize[2], tref1,

  2547.                       row << 2, col << 2, &mvuv, bw, bh, w1, h1);

  2548. 

  2549.         if (b->comp) {

  2550.             if (b->bs > BS_8x8) {

  2551.                 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);

  2552.                 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);

  2553.             } else {

  2554.                 mvuv = b->mv[0][1];

  2555.             }

  2556.             mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][1],

  2557.                           s->dst[1], s->dst[2], ls_uv,

  2558.                           ref2->data[1], ref2->linesize[1],

  2559.                           ref2->data[2], ref2->linesize[2], tref2,

  2560.                           row << 2, col << 2, &mvuv, bw, bh, w2, h2);

  2561.         }

  2562.     }

  2563. 

  2564.     if (!b->skip) {

  2565.         /* mostly copied intra_reconn() */

  2566. 

  2567.         int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2568.         int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2569.         int end_x = FFMIN(2 * (s->cols - col), w4);

  2570.         int end_y = FFMIN(2 * (s->rows - row), h4);

  2571.         int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2572.         int uvstep1d = 1 << b->uvtx, p;

  2573.         uint8_t *dst = s->dst[0];

  2574. 

  2575.         // y itxfm add

  2576.         for (n = 0, y = 0; y < end_y; y += step1d) {

  2577.             uint8_t *ptr = dst;

  2578.             for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d, n += step) {

  2579.                 int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2580. 

  2581.                 if (eob)

  2582.                     s->dsp.itxfm_add[tx][DCT_DCT](ptr, s->y_stride,

  2583.                                                   s->block + 16 * n, eob);

  2584.             }

  2585.             dst += 4 * s->y_stride * step1d;

  2586.         }

  2587. 

  2588.         // uv itxfm add

  2589.         h4 >>= 1;

  2590.         w4 >>= 1;

  2591.         end_x >>= 1;

  2592.         end_y >>= 1;

  2593.         step = 1 << (b->uvtx * 2);

  2594.         for (p = 0; p < 2; p++) {

  2595.             dst = s->dst[p + 1];

  2596.             for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2597.                 uint8_t *ptr = dst;

  2598.                 for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d, n += step) {

  2599.                     int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2600. 

  2601.                     if (eob)

  2602.                         s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2603.                                                         s->uvblock[p] + 16 * n, eob);

  2604.                 }

  2605.                 dst += 4 * uvstep1d * s->uv_stride;

  2606.             }

  2607.         }

  2608.     }

  2609. }

  2610. 

  2611. static av_always_inline void mask_edges(struct VP9Filter *lflvl, int is_uv,

  2612.                                         int row_and_7, int col_and_7,

  2613.                                         int w, int h, int col_end, int row_end,

  2614.                                         enum TxfmMode tx, int skip_inter)

  2615. {

  2616.     // FIXME I'm pretty sure all loops can be replaced by a single LUT if

  2617.     // we make VP9Filter.mask uint64_t (i.e. row/col all single variable)

  2618.     // and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then

  2619.     // use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)

  2620. 

  2621.     // the intended behaviour of the vp9 loopfilter is to work on 8-pixel

  2622.     // edges. This means that for UV, we work on two subsampled blocks at

  2623.     // a time, and we only use the topleft block's mode information to set

  2624.     // things like block strength. Thus, for any block size smaller than

  2625.     // 16x16, ignore the odd portion of the block.

  2626.     if (tx == TX_4X4 && is_uv) {

  2627.         if (h == 1) {

  2628.             if (row_and_7 & 1)

  2629.                 return;

  2630.             if (!row_end)

  2631.                 h += 1;

  2632.         }

  2633.         if (w == 1) {

  2634.             if (col_and_7 & 1)

  2635.                 return;

  2636.             if (!col_end)

  2637.                 w += 1;

  2638.         }

  2639.     }

  2640. 

  2641.     if (tx == TX_4X4 && !skip_inter) {

  2642.         int t = 1 << col_and_7, m_col = (t << w) - t, y;

  2643.         int m_col_odd = (t << (w - 1)) - t;

  2644. 

  2645.         // on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide

  2646.         if (is_uv) {

  2647.             int m_row_8 = m_col & 0x01, m_row_4 = m_col - m_row_8;

  2648. 

  2649.             for (y = row_and_7; y < h + row_and_7; y++) {

  2650.                 int col_mask_id = 2 - !(y & 7);

  2651. 

  2652.                 lflvl->mask[is_uv][0][y][1] |= m_row_8;

  2653.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  2654.                 // for odd lines, if the odd col is not being filtered,

  2655.                 // skip odd row also:

  2656.                 // .---. <-- a

  2657.                 // |   |

  2658.                 // |___| <-- b

  2659.                 // ^   ^

  2660.                 // c   d

  2661.                 //

  2662.                 // if a/c are even row/col and b/d are odd, and d is skipped,

  2663.                 // e.g. right edge of size-66x66.webm, then skip b also (bug)

  2664.                 if ((col_end & 1) && (y & 1)) {

  2665.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col_odd;

  2666.                 } else {

  2667.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col;

  2668.                 }

  2669.             }

  2670.         } else {

  2671.             int m_row_8 = m_col & 0x11, m_row_4 = m_col - m_row_8;

  2672. 

  2673.             for (y = row_and_7; y < h + row_and_7; y++) {

  2674.                 int col_mask_id = 2 - !(y & 3);

  2675. 

  2676.                 lflvl->mask[is_uv][0][y][1] |= m_row_8; // row edge

  2677.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  2678.                 lflvl->mask[is_uv][1][y][col_mask_id] |= m_col; // col edge

  2679.                 lflvl->mask[is_uv][0][y][3] |= m_col;

  2680.                 lflvl->mask[is_uv][1][y][3] |= m_col;

  2681.             }

  2682.         }

  2683.     } else {

  2684.         int y, t = 1 << col_and_7, m_col = (t << w) - t;

  2685. 

  2686.         if (!skip_inter) {

  2687.             int mask_id = (tx == TX_8X8);

  2688.             int l2 = tx + is_uv - 1, step1d = 1 << l2;

  2689.             static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };

  2690.             int m_row = m_col & masks[l2];

  2691. 

  2692.             // at odd UV col/row edges tx16/tx32 loopfilter edges, force

  2693.             // 8wd loopfilter to prevent going off the visible edge.

  2694.             if (is_uv && tx > TX_8X8 && (w ^ (w - 1)) == 1) {

  2695.                 int m_row_16 = ((t << (w - 1)) - t) & masks[l2];

  2696.                 int m_row_8 = m_row - m_row_16;

  2697. 

  2698.                 for (y = row_and_7; y < h + row_and_7; y++) {

  2699.                     lflvl->mask[is_uv][0][y][0] |= m_row_16;

  2700.                     lflvl->mask[is_uv][0][y][1] |= m_row_8;

  2701.                 }

  2702.             } else {

  2703.                 for (y = row_and_7; y < h + row_and_7; y++)

  2704.                     lflvl->mask[is_uv][0][y][mask_id] |= m_row;

  2705.             }

  2706. 

  2707.             if (is_uv && tx > TX_8X8 && (h ^ (h - 1)) == 1) {

  2708.                 for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)

  2709.                     lflvl->mask[is_uv][1][y][0] |= m_col;

  2710.                 if (y - row_and_7 == h - 1)

  2711.                     lflvl->mask[is_uv][1][y][1] |= m_col;

  2712.             } else {

  2713.                 for (y = row_and_7; y < h + row_and_7; y += step1d)

  2714.                     lflvl->mask[is_uv][1][y][mask_id] |= m_col;

  2715.             }

  2716.         } else if (tx != TX_4X4) {

  2717.             int mask_id;

  2718. 

  2719.             mask_id = (tx == TX_8X8) || (is_uv && h == 1);

  2720.             lflvl->mask[is_uv][1][row_and_7][mask_id] |= m_col;

  2721.             mask_id = (tx == TX_8X8) || (is_uv && w == 1);

  2722.             for (y = row_and_7; y < h + row_and_7; y++)

  2723.                 lflvl->mask[is_uv][0][y][mask_id] |= t;

  2724.         } else if (is_uv) {

  2725.             int t8 = t & 0x01, t4 = t - t8;

  2726. 

  2727.             for (y = row_and_7; y < h + row_and_7; y++) {

  2728.                 lflvl->mask[is_uv][0][y][2] |= t4;

  2729.                 lflvl->mask[is_uv][0][y][1] |= t8;

  2730.             }

  2731.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 7)] |= m_col;

  2732.         } else {

  2733.             int t8 = t & 0x11, t4 = t - t8;

  2734. 

  2735.             for (y = row_and_7; y < h + row_and_7; y++) {

  2736.                 lflvl->mask[is_uv][0][y][2] |= t4;

  2737.                 lflvl->mask[is_uv][0][y][1] |= t8;

  2738.             }

  2739.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 3)] |= m_col;

  2740.         }

  2741.     }

  2742. }

  2743. 

  2744. static void decode_b(AVCodecContext *ctx, int row, int col,

  2745.                      struct VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,

  2746.                      enum BlockLevel bl, enum BlockPartition bp)

  2747. {

  2748.     VP9Context *s = ctx->priv_data;

  2749.     VP9Block *b = s->b;

  2750.     enum BlockSize bs = bl * 3 + bp;

  2751.     int y, w4 = bwh_tab[1][bs][0], h4 = bwh_tab[1][bs][1], lvl;

  2752.     int emu[2];

  2753.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2754. 

  2755.     s->row = row;

  2756.     s->row7 = row & 7;

  2757.     s->col = col;

  2758.     s->col7 = col & 7;

  2759.     s->min_mv.x = -(128 + col * 64);

  2760.     s->min_mv.y = -(128 + row * 64);

  2761.     s->max_mv.x = 128 + (s->cols - col - w4) * 64;

  2762.     s->max_mv.y = 128 + (s->rows - row - h4) * 64;

  2763.     if (s->pass < 2) {

  2764.         b->bs = bs;

  2765.         b->bl = bl;

  2766.         b->bp = bp;

  2767.         decode_mode(ctx);

  2768.         b->uvtx = b->tx - (w4 * 2 == (1 << b->tx) || h4 * 2 == (1 << b->tx));

  2769. 

  2770.         if (!b->skip) {

  2771.             decode_coeffs(ctx);

  2772.         } else {

  2773.             int pl;

  2774. 

  2775.             memset(&s->above_y_nnz_ctx[col * 2], 0, w4 * 2);

  2776.             memset(&s->left_y_nnz_ctx[(row & 7) << 1], 0, h4 * 2);

  2777.             for (pl = 0; pl < 2; pl++) {

  2778.                 memset(&s->above_uv_nnz_ctx[pl][col], 0, w4);

  2779.                 memset(&s->left_uv_nnz_ctx[pl][row & 7], 0, h4);

  2780.             }

  2781.         }

  2782.         if (s->pass == 1) {

  2783.             s->b++;

  2784.             s->block += w4 * h4 * 64;

  2785.             s->uvblock[0] += w4 * h4 * 16;

  2786.             s->uvblock[1] += w4 * h4 * 16;

  2787.             s->eob += 4 * w4 * h4;

  2788.             s->uveob[0] += w4 * h4;

  2789.             s->uveob[1] += w4 * h4;

  2790. 

  2791.             return;

  2792.         }

  2793.     }

  2794. 

  2795.     // emulated overhangs if the stride of the target buffer can't hold. This

  2796.     // allows to support emu-edge and so on even if we have large block

  2797.     // overhangs

  2798.     emu[0] = (col + w4) * 8 > f->linesize[0] ||

  2799.              (row + h4) > s->rows;

  2800.     emu[1] = (col + w4) * 4 > f->linesize[1] ||

  2801.              (row + h4) > s->rows;

  2802.     if (emu[0]) {

  2803.         s->dst[0] = s->tmp_y;

  2804.         s->y_stride = 64;

  2805.     } else {

  2806.         s->dst[0] = f->data[0] + yoff;

  2807.         s->y_stride = f->linesize[0];

  2808.     }

  2809.     if (emu[1]) {

  2810.         s->dst[1] = s->tmp_uv[0];

  2811.         s->dst[2] = s->tmp_uv[1];

  2812.         s->uv_stride = 32;

  2813.     } else {

  2814.         s->dst[1] = f->data[1] + uvoff;

  2815.         s->dst[2] = f->data[2] + uvoff;

  2816.         s->uv_stride = f->linesize[1];

  2817.     }

  2818.     if (b->intra) {

  2819.         intra_recon(ctx, yoff, uvoff);

  2820.     } else {

  2821.         inter_recon(ctx);

  2822.     }

  2823.     if (emu[0]) {

  2824.         int w = FFMIN(s->cols - col, w4) * 8, h = FFMIN(s->rows - row, h4) * 8, n, o = 0;

  2825. 

  2826.         for (n = 0; o < w; n++) {

  2827.             int bw = 64 >> n;

  2828. 

  2829.             av_assert2(n <= 4);

  2830.             if (w & bw) {

  2831.                 s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o, f->linesize[0],

  2832.                                          s->tmp_y + o, 64, h, 0, 0);

  2833.                 o += bw;

  2834.             }

  2835.         }

  2836.     }

  2837.     if (emu[1]) {

  2838.         int w = FFMIN(s->cols - col, w4) * 4, h = FFMIN(s->rows - row, h4) * 4, n, o = 0;

  2839. 

  2840.         for (n = 1; o < w; n++) {

  2841.             int bw = 64 >> n;

  2842. 

  2843.             av_assert2(n <= 4);

  2844.             if (w & bw) {

  2845.                 s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o, f->linesize[1],

  2846.                                          s->tmp_uv[0] + o, 32, h, 0, 0);

  2847.                 s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o, f->linesize[2],

  2848.                                          s->tmp_uv[1] + o, 32, h, 0, 0);

  2849.                 o += bw;

  2850.             }

  2851.         }

  2852.     }

  2853. 

  2854.     // pick filter level and find edges to apply filter to

  2855.     if (s->filter.level &&

  2856.         (lvl = s->segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]

  2857.                                                     [b->mode[3] != ZEROMV]) > 0) {

  2858.         int x_end = FFMIN(s->cols - col, w4), y_end = FFMIN(s->rows - row, h4);

  2859.         int skip_inter = !b->intra && b->skip;

  2860. 

  2861.         for (y = 0; y < h4; y++)

  2862.             memset(&lflvl->level[((row & 7) + y) * 8 + (col & 7)], lvl, w4);

  2863.         mask_edges(lflvl, 0, row & 7, col & 7, x_end, y_end, 0, 0, b->tx, skip_inter);

  2864.         mask_edges(lflvl, 1, row & 7, col & 7, x_end, y_end,

  2865.                    s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,

  2866.                    s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,

  2867.                    b->uvtx, skip_inter);

  2868. 

  2869.         if (!s->filter.lim_lut[lvl]) {

  2870.             int sharp = s->filter.sharpness;

  2871.             int limit = lvl;

  2872. 

  2873.             if (sharp > 0) {

  2874.                 limit >>= (sharp + 3) >> 2;

  2875.                 limit = FFMIN(limit, 9 - sharp);

  2876.             }

  2877.             limit = FFMAX(limit, 1);

  2878. 

  2879.             s->filter.lim_lut[lvl] = limit;

  2880.             s->filter.mblim_lut[lvl] = 2 * (lvl + 2) + limit;

  2881.         }

  2882.     }

  2883. 

  2884.     if (s->pass == 2) {

  2885.         s->b++;

  2886.         s->block += w4 * h4 * 64;

  2887.         s->uvblock[0] += w4 * h4 * 16;

  2888.         s->uvblock[1] += w4 * h4 * 16;

  2889.         s->eob += 4 * w4 * h4;

  2890.         s->uveob[0] += w4 * h4;

  2891.         s->uveob[1] += w4 * h4;

  2892.     }

  2893. }

  2894. 

  2895. static void decode_sb(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  2896.                       ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  2897. {

  2898.     VP9Context *s = ctx->priv_data;

  2899.     int c = ((s->above_partition_ctx[col] >> (3 - bl)) & 1) |

  2900.             (((s->left_partition_ctx[row & 0x7] >> (3 - bl)) & 1) << 1);

  2901.     const uint8_t *p = s->keyframe ? vp9_default_kf_partition_probs[bl][c] :

  2902.                                      s->prob.p.partition[bl][c];

  2903.     enum BlockPartition bp;

  2904.     ptrdiff_t hbs = 4 >> bl;

  2905.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2906.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  2907. 

  2908.     if (bl == BL_8X8) {

  2909.         bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  2910.         decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2911.     } else if (col + hbs < s->cols) { // FIXME why not <=?

  2912.         if (row + hbs < s->rows) { // FIXME why not <=?

  2913.             bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  2914.             switch (bp) {

  2915.             case PARTITION_NONE:

  2916.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2917.                 break;

  2918.             case PARTITION_H:

  2919.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2920.                 yoff  += hbs * 8 * y_stride;

  2921.                 uvoff += hbs * 4 * uv_stride;

  2922.                 decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, bl, bp);

  2923.                 break;

  2924.             case PARTITION_V:

  2925.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2926.                 yoff  += hbs * 8;

  2927.                 uvoff += hbs * 4;

  2928.                 decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, bl, bp);

  2929.                 break;

  2930.             case PARTITION_SPLIT:

  2931.                 decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2932.                 decode_sb(ctx, row, col + hbs, lflvl,

  2933.                           yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2934.                 yoff  += hbs * 8 * y_stride;

  2935.                 uvoff += hbs * 4 * uv_stride;

  2936.                 decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  2937.                 decode_sb(ctx, row + hbs, col + hbs, lflvl,

  2938.                           yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2939.                 break;

  2940.             default:

  2941.                 av_assert0(0);

  2942.             }

  2943.         } else if (vp56_rac_get_prob_branchy(&s->c, p[1])) {

  2944.             bp = PARTITION_SPLIT;

  2945.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2946.             decode_sb(ctx, row, col + hbs, lflvl,

  2947.                       yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2948.         } else {

  2949.             bp = PARTITION_H;

  2950.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2951.         }

  2952.     } else if (row + hbs < s->rows) { // FIXME why not <=?

  2953.         if (vp56_rac_get_prob_branchy(&s->c, p[2])) {

  2954.             bp = PARTITION_SPLIT;

  2955.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2956.             yoff  += hbs * 8 * y_stride;

  2957.             uvoff += hbs * 4 * uv_stride;

  2958.             decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  2959.         } else {

  2960.             bp = PARTITION_V;

  2961.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2962.         }

  2963.     } else {

  2964.         bp = PARTITION_SPLIT;

  2965.         decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2966.     }

  2967.     s->counts.partition[bl][c][bp]++;

  2968. }

  2969. 

  2970. static void decode_sb_mem(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  2971.                           ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  2972. {

  2973.     VP9Context *s = ctx->priv_data;

  2974.     VP9Block *b = s->b;

  2975.     ptrdiff_t hbs = 4 >> bl;

  2976.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2977.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  2978. 

  2979.     if (bl == BL_8X8) {

  2980.         av_assert2(b->bl == BL_8X8);

  2981.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  2982.     } else if (s->b->bl == bl) {

  2983.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  2984.         if (b->bp == PARTITION_H && row + hbs < s->rows) {

  2985.             yoff  += hbs * 8 * y_stride;

  2986.             uvoff += hbs * 4 * uv_stride;

  2987.             decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, b->bl, b->bp);

  2988.         } else if (b->bp == PARTITION_V && col + hbs < s->cols) {

  2989.             yoff  += hbs * 8;

  2990.             uvoff += hbs * 4;

  2991.             decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, b->bl, b->bp);

  2992.         }

  2993.     } else {

  2994.         decode_sb_mem(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2995.         if (col + hbs < s->cols) { // FIXME why not <=?

  2996.             if (row + hbs < s->rows) {

  2997.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff + 8 * hbs,

  2998.                               uvoff + 4 * hbs, bl + 1);

  2999.                 yoff  += hbs * 8 * y_stride;

  3000.                 uvoff += hbs * 4 * uv_stride;

  3001.                 decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3002.                 decode_sb_mem(ctx, row + hbs, col + hbs, lflvl,

  3003.                                     yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  3004.             } else {

  3005.                 yoff  += hbs * 8;

  3006.                 uvoff += hbs * 4;

  3007.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff, uvoff, bl + 1);

  3008.             }

  3009.         } else if (row + hbs < s->rows) {

  3010.             yoff  += hbs * 8 * y_stride;

  3011.             uvoff += hbs * 4 * uv_stride;

  3012.             decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3013.         }

  3014.     }

  3015. }

  3016. 

  3017. static void loopfilter_sb(AVCodecContext *ctx, struct VP9Filter *lflvl,

  3018.                           int row, int col, ptrdiff_t yoff, ptrdiff_t uvoff)

  3019. {

  3020.     VP9Context *s = ctx->priv_data;

  3021.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  3022.     uint8_t *dst = f->data[0] + yoff, *lvl = lflvl->level;

  3023.     ptrdiff_t ls_y = f->linesize[0], ls_uv = f->linesize[1];

  3024.     int y, x, p;

  3025. 

  3026.     // FIXME in how far can we interleave the v/h loopfilter calls? E.g.

  3027.     // if you think of them as acting on a 8x8 block max, we can interleave

  3028.     // each v/h within the single x loop, but that only works if we work on

  3029.     // 8 pixel blocks, and we won't always do that (we want at least 16px

  3030.     // to use SSE2 optimizations, perhaps 32 for AVX2)

  3031. 

  3032.     // filter edges between columns, Y plane (e.g. block1 | block2)

  3033.     for (y = 0; y < 8; y += 2, dst += 16 * ls_y, lvl += 16) {

  3034.         uint8_t *ptr = dst, *l = lvl, *hmask1 = lflvl->mask[0][0][y];

  3035.         uint8_t *hmask2 = lflvl->mask[0][0][y + 1];

  3036.         unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2], hm13 = hmask1[3];

  3037.         unsigned hm2 = hmask2[1] | hmask2[2], hm23 = hmask2[3];

  3038.         unsigned hm = hm1 | hm2 | hm13 | hm23;

  3039. 

  3040.         for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 8, l++) {

  3041.             if (hm1 & x) {

  3042.                 int L = *l, H = L >> 4;

  3043.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3044. 

  3045.                 if (col || x > 1) {

  3046.                     if (hmask1[0] & x) {

  3047.                         if (hmask2[0] & x) {

  3048.                             av_assert2(l[8] == L);

  3049.                             s->dsp.loop_filter_16[0](ptr, ls_y, E, I, H);

  3050.                         } else {

  3051.                             s->dsp.loop_filter_8[2][0](ptr, ls_y, E, I, H);

  3052.                         }

  3053.                     } else if (hm2 & x) {

  3054.                         L = l[8];

  3055.                         H |= (L >> 4) << 8;

  3056.                         E |= s->filter.mblim_lut[L] << 8;

  3057.                         I |= s->filter.lim_lut[L] << 8;

  3058.                         s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3059.                                                [!!(hmask2[1] & x)]

  3060.                                                [0](ptr, ls_y, E, I, H);

  3061.                     } else {

  3062.                         s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3063.                                             [0](ptr, ls_y, E, I, H);

  3064.                     }

  3065.                 }

  3066.             } else if (hm2 & x) {

  3067.                 int L = l[8], H = L >> 4;

  3068.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3069. 

  3070.                 if (col || x > 1) {

  3071.                     s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3072.                                         [0](ptr + 8 * ls_y, ls_y, E, I, H);

  3073.                 }

  3074.             }

  3075.             if (hm13 & x) {

  3076.                 int L = *l, H = L >> 4;

  3077.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3078. 

  3079.                 if (hm23 & x) {

  3080.                     L = l[8];

  3081.                     H |= (L >> 4) << 8;

  3082.                     E |= s->filter.mblim_lut[L] << 8;

  3083.                     I |= s->filter.lim_lut[L] << 8;

  3084.                     s->dsp.loop_filter_mix2[0][0][0](ptr + 4, ls_y, E, I, H);

  3085.                 } else {

  3086.                     s->dsp.loop_filter_8[0][0](ptr + 4, ls_y, E, I, H);

  3087.                 }

  3088.             } else if (hm23 & x) {

  3089.                 int L = l[8], H = L >> 4;

  3090.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3091. 

  3092.                 s->dsp.loop_filter_8[0][0](ptr + 8 * ls_y + 4, ls_y, E, I, H);

  3093.             }

  3094.         }

  3095.     }

  3096. 

  3097.     //                                          block1

  3098.     // filter edges between rows, Y plane (e.g. ------)

  3099.     //                                          block2

  3100.     dst = f->data[0] + yoff;

  3101.     lvl = lflvl->level;

  3102.     for (y = 0; y < 8; y++, dst += 8 * ls_y, lvl += 8) {

  3103.         uint8_t *ptr = dst, *l = lvl, *vmask = lflvl->mask[0][1][y];

  3104.         unsigned vm = vmask[0] | vmask[1] | vmask[2], vm3 = vmask[3];

  3105. 

  3106.         for (x = 1; vm & ~(x - 1); x <<= 2, ptr += 16, l += 2) {

  3107.             if (row || y) {

  3108.                 if (vm & x) {

  3109.                     int L = *l, H = L >> 4;

  3110.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3111. 

  3112.                     if (vmask[0] & x) {

  3113.                         if (vmask[0] & (x << 1)) {

  3114.                             av_assert2(l[1] == L);

  3115.                             s->dsp.loop_filter_16[1](ptr, ls_y, E, I, H);

  3116.                         } else {

  3117.                             s->dsp.loop_filter_8[2][1](ptr, ls_y, E, I, H);

  3118.                         }

  3119.                     } else if (vm & (x << 1)) {

  3120.                         L = l[1];

  3121.                         H |= (L >> 4) << 8;

  3122.                         E |= s->filter.mblim_lut[L] << 8;

  3123.                         I |= s->filter.lim_lut[L] << 8;

  3124.                         s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3125.                                                [!!(vmask[1] & (x << 1))]

  3126.                                                [1](ptr, ls_y, E, I, H);

  3127.                     } else {

  3128.                         s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3129.                                             [1](ptr, ls_y, E, I, H);

  3130.                     }

  3131.                 } else if (vm & (x << 1)) {

  3132.                     int L = l[1], H = L >> 4;

  3133.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3134. 

  3135.                     s->dsp.loop_filter_8[!!(vmask[1] & (x << 1))]

  3136.                                         [1](ptr + 8, ls_y, E, I, H);

  3137.                 }

  3138.             }

  3139.             if (vm3 & x) {

  3140.                 int L = *l, H = L >> 4;

  3141.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3142. 

  3143.                 if (vm3 & (x << 1)) {

  3144.                     L = l[1];

  3145.                     H |= (L >> 4) << 8;

  3146.                     E |= s->filter.mblim_lut[L] << 8;

  3147.                     I |= s->filter.lim_lut[L] << 8;

  3148.                     s->dsp.loop_filter_mix2[0][0][1](ptr + ls_y * 4, ls_y, E, I, H);

  3149.                 } else {

  3150.                     s->dsp.loop_filter_8[0][1](ptr + ls_y * 4, ls_y, E, I, H);

  3151.                 }

  3152.             } else if (vm3 & (x << 1)) {

  3153.                 int L = l[1], H = L >> 4;

  3154.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3155. 

  3156.                 s->dsp.loop_filter_8[0][1](ptr + ls_y * 4 + 8, ls_y, E, I, H);

  3157.             }

  3158.         }

  3159.     }

  3160. 

  3161.     // same principle but for U/V planes

  3162.     for (p = 0; p < 2; p++) {

  3163.         lvl = lflvl->level;

  3164.         dst = f->data[1 + p] + uvoff;

  3165.         for (y = 0; y < 8; y += 4, dst += 16 * ls_uv, lvl += 32) {

  3166.             uint8_t *ptr = dst, *l = lvl, *hmask1 = lflvl->mask[1][0][y];

  3167.             uint8_t *hmask2 = lflvl->mask[1][0][y + 2];

  3168.             unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2];

  3169.             unsigned hm2 = hmask2[1] | hmask2[2], hm = hm1 | hm2;

  3170. 

  3171.             for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 4) {

  3172.                 if (col || x > 1) {

  3173.                     if (hm1 & x) {

  3174.                         int L = *l, H = L >> 4;

  3175.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3176. 

  3177.                         if (hmask1[0] & x) {

  3178.                             if (hmask2[0] & x) {

  3179.                                 av_assert2(l[16] == L);

  3180.                                 s->dsp.loop_filter_16[0](ptr, ls_uv, E, I, H);

  3181.                             } else {

  3182.                                 s->dsp.loop_filter_8[2][0](ptr, ls_uv, E, I, H);

  3183.                             }

  3184.                         } else if (hm2 & x) {

  3185.                             L = l[16];

  3186.                             H |= (L >> 4) << 8;

  3187.                             E |= s->filter.mblim_lut[L] << 8;

  3188.                             I |= s->filter.lim_lut[L] << 8;

  3189.                             s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3190.                                                    [!!(hmask2[1] & x)]

  3191.                                                    [0](ptr, ls_uv, E, I, H);

  3192.                         } else {

  3193.                             s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3194.                                                 [0](ptr, ls_uv, E, I, H);

  3195.                         }

  3196.                     } else if (hm2 & x) {

  3197.                         int L = l[16], H = L >> 4;

  3198.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3199. 

  3200.                         s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3201.                                             [0](ptr + 8 * ls_uv, ls_uv, E, I, H);

  3202.                     }

  3203.                 }

  3204.                 if (x & 0xAA)

  3205.                     l += 2;

  3206.             }

  3207.         }

  3208.         lvl = lflvl->level;

  3209.         dst = f->data[1 + p] + uvoff;

  3210.         for (y = 0; y < 8; y++, dst += 4 * ls_uv) {

  3211.             uint8_t *ptr = dst, *l = lvl, *vmask = lflvl->mask[1][1][y];

  3212.             unsigned vm = vmask[0] | vmask[1] | vmask[2];

  3213. 

  3214.             for (x = 1; vm & ~(x - 1); x <<= 4, ptr += 16, l += 4) {

  3215.                 if (row || y) {

  3216.                     if (vm & x) {

  3217.                         int L = *l, H = L >> 4;

  3218.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3219. 

  3220.                         if (vmask[0] & x) {

  3221.                             if (vmask[0] & (x << 2)) {

  3222.                                 av_assert2(l[2] == L);

  3223.                                 s->dsp.loop_filter_16[1](ptr, ls_uv, E, I, H);

  3224.                             } else {

  3225.                                 s->dsp.loop_filter_8[2][1](ptr, ls_uv, E, I, H);

  3226.                             }

  3227.                         } else if (vm & (x << 2)) {

  3228.                             L = l[2];

  3229.                             H |= (L >> 4) << 8;

  3230.                             E |= s->filter.mblim_lut[L] << 8;

  3231.                             I |= s->filter.lim_lut[L] << 8;

  3232.                             s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3233.                                                    [!!(vmask[1] & (x << 2))]

  3234.                                                    [1](ptr, ls_uv, E, I, H);

  3235.                         } else {

  3236.                             s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3237.                                                 [1](ptr, ls_uv, E, I, H);

  3238.                         }

  3239.                     } else if (vm & (x << 2)) {

  3240.                         int L = l[2], H = L >> 4;

  3241.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3242. 

  3243.                         s->dsp.loop_filter_8[!!(vmask[1] & (x << 2))]

  3244.                                             [1](ptr + 8, ls_uv, E, I, H);

  3245.                     }

  3246.                 }

  3247.             }

  3248.             if (y & 1)

  3249.                 lvl += 16;

  3250.         }

  3251.     }

  3252. }

  3253. 

  3254. static void set_tile_offset(int *start, int *end, int idx, int log2_n, int n)

  3255. {

  3256.     int sb_start = ( idx      * n) >> log2_n;

  3257.     int sb_end   = ((idx + 1) * n) >> log2_n;

  3258.     *start = FFMIN(sb_start, n) << 3;

  3259.     *end   = FFMIN(sb_end,   n) << 3;

  3260. }

  3261. 

  3262. static av_always_inline void adapt_prob(uint8_t *p, unsigned ct0, unsigned ct1,

  3263.                                         int max_count, int update_factor)

  3264. {

  3265.     unsigned ct = ct0 + ct1, p2, p1;

  3266. 

  3267.     if (!ct)

  3268.         return;

  3269. 

  3270.     p1 = *p;

  3271.     p2 = ((ct0 << 8) + (ct >> 1)) / ct;

  3272.     p2 = av_clip(p2, 1, 255);

  3273.     ct = FFMIN(ct, max_count);

  3274.     update_factor = FASTDIV(update_factor * ct, max_count);

  3275. 

  3276.     // (p1 * (256 - update_factor) + p2 * update_factor + 128) >> 8

  3277.     *p = p1 + (((p2 - p1) * update_factor + 128) >> 8);

  3278. }

  3279. 

  3280. static void adapt_probs(VP9Context *s)

  3281. {

  3282.     int i, j, k, l, m;

  3283.     prob_context *p = &s->prob_ctx[s->framectxid].p;

  3284.     int uf = (s->keyframe || s->intraonly || !s->last_keyframe) ? 112 : 128;

  3285. 

  3286.     // coefficients

  3287.     for (i = 0; i < 4; i++)

  3288.         for (j = 0; j < 2; j++)

  3289.             for (k = 0; k < 2; k++)

  3290.                 for (l = 0; l < 6; l++)

  3291.                     for (m = 0; m < 6; m++) {

  3292.                         uint8_t *pp = s->prob_ctx[s->framectxid].coef[i][j][k][l][m];

  3293.                         unsigned *e = s->counts.eob[i][j][k][l][m];

  3294.                         unsigned *c = s->counts.coef[i][j][k][l][m];

  3295. 

  3296.                         if (l == 0 && m >= 3) // dc only has 3 pt

  3297.                             break;

  3298. 

  3299.                         adapt_prob(&pp[0], e[0], e[1], 24, uf);

  3300.                         adapt_prob(&pp[1], c[0], c[1] + c[2], 24, uf);

  3301.                         adapt_prob(&pp[2], c[1], c[2], 24, uf);

  3302.                     }

  3303. 

  3304.     if (s->keyframe || s->intraonly) {

  3305.         memcpy(p->skip,  s->prob.p.skip,  sizeof(p->skip));

  3306.         memcpy(p->tx32p, s->prob.p.tx32p, sizeof(p->tx32p));

  3307.         memcpy(p->tx16p, s->prob.p.tx16p, sizeof(p->tx16p));

  3308.         memcpy(p->tx8p,  s->prob.p.tx8p,  sizeof(p->tx8p));

  3309.         return;

  3310.     }

  3311. 

  3312.     // skip flag

  3313.     for (i = 0; i < 3; i++)

  3314.         adapt_prob(&p->skip[i], s->counts.skip[i][0], s->counts.skip[i][1], 20, 128);

  3315. 

  3316.     // intra/inter flag

  3317.     for (i = 0; i < 4; i++)

  3318.         adapt_prob(&p->intra[i], s->counts.intra[i][0], s->counts.intra[i][1], 20, 128);

  3319. 

  3320.     // comppred flag

  3321.     if (s->comppredmode == PRED_SWITCHABLE) {

  3322.       for (i = 0; i < 5; i++)

  3323.           adapt_prob(&p->comp[i], s->counts.comp[i][0], s->counts.comp[i][1], 20, 128);

  3324.     }

  3325. 

  3326.     // reference frames

  3327.     if (s->comppredmode != PRED_SINGLEREF) {

  3328.       for (i = 0; i < 5; i++)

  3329.           adapt_prob(&p->comp_ref[i], s->counts.comp_ref[i][0],

  3330.                      s->counts.comp_ref[i][1], 20, 128);

  3331.     }

  3332. 

  3333.     if (s->comppredmode != PRED_COMPREF) {

  3334.       for (i = 0; i < 5; i++) {

  3335.           uint8_t *pp = p->single_ref[i];

  3336.           unsigned (*c)[2] = s->counts.single_ref[i];

  3337. 

  3338.           adapt_prob(&pp[0], c[0][0], c[0][1], 20, 128);

  3339.           adapt_prob(&pp[1], c[1][0], c[1][1], 20, 128);

  3340.       }

  3341.     }

  3342. 

  3343.     // block partitioning

  3344.     for (i = 0; i < 4; i++)

  3345.         for (j = 0; j < 4; j++) {

  3346.             uint8_t *pp = p->partition[i][j];

  3347.             unsigned *c = s->counts.partition[i][j];

  3348. 

  3349.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3350.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3351.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3352.         }

  3353. 

  3354.     // tx size

  3355.     if (s->txfmmode == TX_SWITCHABLE) {

  3356.       for (i = 0; i < 2; i++) {

  3357.           unsigned *c16 = s->counts.tx16p[i], *c32 = s->counts.tx32p[i];

  3358. 

  3359.           adapt_prob(&p->tx8p[i], s->counts.tx8p[i][0], s->counts.tx8p[i][1], 20, 128);

  3360.           adapt_prob(&p->tx16p[i][0], c16[0], c16[1] + c16[2], 20, 128);

  3361.           adapt_prob(&p->tx16p[i][1], c16[1], c16[2], 20, 128);

  3362.           adapt_prob(&p->tx32p[i][0], c32[0], c32[1] + c32[2] + c32[3], 20, 128);

  3363.           adapt_prob(&p->tx32p[i][1], c32[1], c32[2] + c32[3], 20, 128);

  3364.           adapt_prob(&p->tx32p[i][2], c32[2], c32[3], 20, 128);

  3365.       }

  3366.     }

  3367. 

  3368.     // interpolation filter

  3369.     if (s->filtermode == FILTER_SWITCHABLE) {

  3370.         for (i = 0; i < 4; i++) {

  3371.             uint8_t *pp = p->filter[i];

  3372.             unsigned *c = s->counts.filter[i];

  3373. 

  3374.             adapt_prob(&pp[0], c[0], c[1] + c[2], 20, 128);

  3375.             adapt_prob(&pp[1], c[1], c[2], 20, 128);

  3376.         }

  3377.     }

  3378. 

  3379.     // inter modes

  3380.     for (i = 0; i < 7; i++) {

  3381.         uint8_t *pp = p->mv_mode[i];

  3382.         unsigned *c = s->counts.mv_mode[i];

  3383. 

  3384.         adapt_prob(&pp[0], c[2], c[1] + c[0] + c[3], 20, 128);

  3385.         adapt_prob(&pp[1], c[0], c[1] + c[3], 20, 128);

  3386.         adapt_prob(&pp[2], c[1], c[3], 20, 128);

  3387.     }

  3388. 

  3389.     // mv joints

  3390.     {

  3391.         uint8_t *pp = p->mv_joint;

  3392.         unsigned *c = s->counts.mv_joint;

  3393. 

  3394.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3395.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3396.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3397.     }

  3398. 

  3399.     // mv components

  3400.     for (i = 0; i < 2; i++) {

  3401.         uint8_t *pp;

  3402.         unsigned *c, (*c2)[2], sum;

  3403. 

  3404.         adapt_prob(&p->mv_comp[i].sign, s->counts.mv_comp[i].sign[0],

  3405.                    s->counts.mv_comp[i].sign[1], 20, 128);

  3406. 

  3407.         pp = p->mv_comp[i].classes;

  3408.         c = s->counts.mv_comp[i].classes;

  3409.         sum = c[1] + c[2] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9] + c[10];

  3410.         adapt_prob(&pp[0], c[0], sum, 20, 128);

  3411.         sum -= c[1];

  3412.         adapt_prob(&pp[1], c[1], sum, 20, 128);

  3413.         sum -= c[2] + c[3];

  3414.         adapt_prob(&pp[2], c[2] + c[3], sum, 20, 128);

  3415.         adapt_prob(&pp[3], c[2], c[3], 20, 128);

  3416.         sum -= c[4] + c[5];

  3417.         adapt_prob(&pp[4], c[4] + c[5], sum, 20, 128);

  3418.         adapt_prob(&pp[5], c[4], c[5], 20, 128);

  3419.         sum -= c[6];

  3420.         adapt_prob(&pp[6], c[6], sum, 20, 128);

  3421.         adapt_prob(&pp[7], c[7] + c[8], c[9] + c[10], 20, 128);

  3422.         adapt_prob(&pp[8], c[7], c[8], 20, 128);

  3423.         adapt_prob(&pp[9], c[9], c[10], 20, 128);

  3424. 

  3425.         adapt_prob(&p->mv_comp[i].class0, s->counts.mv_comp[i].class0[0],

  3426.                    s->counts.mv_comp[i].class0[1], 20, 128);

  3427.         pp = p->mv_comp[i].bits;

  3428.         c2 = s->counts.mv_comp[i].bits;

  3429.         for (j = 0; j < 10; j++)

  3430.             adapt_prob(&pp[j], c2[j][0], c2[j][1], 20, 128);

  3431. 

  3432.         for (j = 0; j < 2; j++) {

  3433.             pp = p->mv_comp[i].class0_fp[j];

  3434.             c = s->counts.mv_comp[i].class0_fp[j];

  3435.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3436.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3437.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3438.         }

  3439.         pp = p->mv_comp[i].fp;

  3440.         c = s->counts.mv_comp[i].fp;

  3441.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3442.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3443.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3444. 

  3445.         if (s->highprecisionmvs) {

  3446.             adapt_prob(&p->mv_comp[i].class0_hp, s->counts.mv_comp[i].class0_hp[0],

  3447.                        s->counts.mv_comp[i].class0_hp[1], 20, 128);

  3448.             adapt_prob(&p->mv_comp[i].hp, s->counts.mv_comp[i].hp[0],

  3449.                        s->counts.mv_comp[i].hp[1], 20, 128);

  3450.         }

  3451.     }

  3452. 

  3453.     // y intra modes

  3454.     for (i = 0; i < 4; i++) {

  3455.         uint8_t *pp = p->y_mode[i];

  3456.         unsigned *c = s->counts.y_mode[i], sum, s2;

  3457. 

  3458.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3459.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3460.         sum -= c[TM_VP8_PRED];

  3461.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3462.         sum -= c[VERT_PRED];

  3463.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3464.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3465.         sum -= s2;

  3466.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3467.         s2 -= c[HOR_PRED];

  3468.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3469.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3470.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3471.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3472.         sum -= c[VERT_LEFT_PRED];

  3473.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3474.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3475.     }

  3476. 

  3477.     // uv intra modes

  3478.     for (i = 0; i < 10; i++) {

  3479.         uint8_t *pp = p->uv_mode[i];

  3480.         unsigned *c = s->counts.uv_mode[i], sum, s2;

  3481. 

  3482.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3483.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3484.         sum -= c[TM_VP8_PRED];

  3485.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3486.         sum -= c[VERT_PRED];

  3487.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3488.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3489.         sum -= s2;

  3490.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3491.         s2 -= c[HOR_PRED];

  3492.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3493.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3494.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3495.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3496.         sum -= c[VERT_LEFT_PRED];

  3497.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3498.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3499.     }

  3500. }

  3501. 

  3502. static void free_buffers(VP9Context *s)

  3503. {

  3504.     av_freep(&s->intra_pred_data[0]);

  3505.     av_freep(&s->b_base);

  3506.     av_freep(&s->block_base);

  3507. }

  3508. 

  3509. static av_cold int vp9_decode_free(AVCodecContext *ctx)

  3510. {

  3511.     VP9Context *s = ctx->priv_data;

  3512.     int i;

  3513. 

  3514.     for (i = 0; i < 2; i++) {

  3515.         if (s->frames[i].tf.f->data[0])

  3516.             vp9_unref_frame(ctx, &s->frames[i]);

  3517.         av_frame_free(&s->frames[i].tf.f);

  3518.     }

  3519.     for (i = 0; i < 8; i++) {

  3520.         if (s->refs[i].f->data[0])

  3521.             ff_thread_release_buffer(ctx, &s->refs[i]);

  3522.         av_frame_free(&s->refs[i].f);

  3523.         if (s->next_refs[i].f->data[0])

  3524.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3525.         av_frame_free(&s->next_refs[i].f);

  3526.     }

  3527.     free_buffers(s);

  3528.     av_freep(&s->c_b);

  3529.     s->c_b_size = 0;

  3530. 

  3531.     return 0;

  3532. }

  3533. 

  3534. 

  3535. static int vp9_decode_frame(AVCodecContext *ctx, void *frame,

  3536.                             int *got_frame, AVPacket *pkt)

  3537. {

  3538.     const uint8_t *data = pkt->data;

  3539.     int size = pkt->size;

  3540.     VP9Context *s = ctx->priv_data;

  3541.     int res, tile_row, tile_col, i, ref, row, col;

  3542.     ptrdiff_t yoff, uvoff, ls_y, ls_uv;

  3543.     AVFrame *f;

  3544. 

  3545.     if ((res = decode_frame_header(ctx, data, size, &ref)) < 0) {

  3546.         return res;

  3547.     } else if (res == 0) {

  3548.         if (!s->refs[ref].f->data[0]) {

  3549.             av_log(ctx, AV_LOG_ERROR, "Requested reference %d not available\n", ref);

  3550.             return AVERROR_INVALIDDATA;

  3551.         }

  3552.         if ((res = av_frame_ref(frame, s->refs[ref].f)) < 0)

  3553.             return res;

  3554.         *got_frame = 1;

  3555.         return 0;

  3556.     }

  3557.     data += res;

  3558.     size -= res;

  3559. 

  3560.     if (s->frames[LAST_FRAME].tf.f->data[0])

  3561.         vp9_unref_frame(ctx, &s->frames[LAST_FRAME]);

  3562.     if (!s->keyframe && s->frames[CUR_FRAME].tf.f->data[0] &&

  3563.         (res = vp9_ref_frame(ctx, &s->frames[LAST_FRAME], &s->frames[CUR_FRAME])) < 0)

  3564.         return res;

  3565.     if (s->frames[CUR_FRAME].tf.f->data[0])

  3566.         vp9_unref_frame(ctx, &s->frames[CUR_FRAME]);

  3567.     if ((res = vp9_alloc_frame(ctx, &s->frames[CUR_FRAME])) < 0)

  3568.         return res;

  3569.     f = s->frames[CUR_FRAME].tf.f;

  3570.     f->key_frame = s->keyframe;

  3571.     f->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;

  3572.     ls_y = f->linesize[0];

  3573.     ls_uv =f->linesize[1];

  3574. 

  3575.     // ref frame setup

  3576.     for (i = 0; i < 8; i++) {

  3577.         if (s->next_refs[i].f->data[0])

  3578.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3579.         if (s->refreshrefmask & (1 << i)) {

  3580.             res = ff_thread_ref_frame(&s->next_refs[i], &s->frames[CUR_FRAME].tf);

  3581.         } else {

  3582.             res = ff_thread_ref_frame(&s->next_refs[i], &s->refs[i]);

  3583.         }

  3584.         if (res < 0)

  3585.             return res;

  3586.     }

  3587. 

  3588.     // main tile decode loop

  3589.     memset(s->above_partition_ctx, 0, s->cols);

  3590.     memset(s->above_skip_ctx, 0, s->cols);

  3591.     if (s->keyframe || s->intraonly) {

  3592.         memset(s->above_mode_ctx, DC_PRED, s->cols * 2);

  3593.     } else {

  3594.         memset(s->above_mode_ctx, NEARESTMV, s->cols);

  3595.     }

  3596.     memset(s->above_y_nnz_ctx, 0, s->sb_cols * 16);

  3597.     memset(s->above_uv_nnz_ctx[0], 0, s->sb_cols * 8);

  3598.     memset(s->above_uv_nnz_ctx[1], 0, s->sb_cols * 8);

  3599.     memset(s->above_segpred_ctx, 0, s->cols);

  3600.     s->pass = s->uses_2pass =

  3601.         ctx->active_thread_type == FF_THREAD_FRAME && s->refreshctx && !s->parallelmode;

  3602.     if ((res = update_block_buffers(ctx)) < 0) {

  3603.         av_log(ctx, AV_LOG_ERROR,

  3604.                "Failed to allocate block buffers\n");

  3605.         return res;

  3606.     }

  3607.     if (s->refreshctx && s->parallelmode) {

  3608.         int j, k, l, m;

  3609. 

  3610.         for (i = 0; i < 4; i++) {

  3611.             for (j = 0; j < 2; j++)

  3612.                 for (k = 0; k < 2; k++)

  3613.                     for (l = 0; l < 6; l++)

  3614.                         for (m = 0; m < 6; m++)

  3615.                             memcpy(s->prob_ctx[s->framectxid].coef[i][j][k][l][m],

  3616.                                    s->prob.coef[i][j][k][l][m], 3);

  3617.             if (s->txfmmode == i)

  3618.                 break;

  3619.         }

  3620.         s->prob_ctx[s->framectxid].p = s->prob.p;

  3621.         ff_thread_finish_setup(ctx);

  3622.     }

  3623. 

  3624.     do {

  3625.         yoff = uvoff = 0;

  3626.         s->b = s->b_base;

  3627.         s->block = s->block_base;

  3628.         s->uvblock[0] = s->uvblock_base[0];

  3629.         s->uvblock[1] = s->uvblock_base[1];

  3630.         s->eob = s->eob_base;

  3631.         s->uveob[0] = s->uveob_base[0];

  3632.         s->uveob[1] = s->uveob_base[1];

  3633. 

  3634.         for (tile_row = 0; tile_row < s->tiling.tile_rows; tile_row++) {

  3635.             set_tile_offset(&s->tiling.tile_row_start, &s->tiling.tile_row_end,

  3636.                             tile_row, s->tiling.log2_tile_rows, s->sb_rows);

  3637.             if (s->pass != 2) {

  3638.                 for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  3639.                     unsigned tile_size;

  3640. 

  3641.                     if (tile_col == s->tiling.tile_cols - 1 &&

  3642.                         tile_row == s->tiling.tile_rows - 1) {

  3643.                         tile_size = size;

  3644.                     } else {

  3645.                         tile_size = AV_RB32(data);

  3646.                         data += 4;

  3647.                         size -= 4;

  3648.                     }

  3649.                     if (tile_size > size) {

  3650.                         ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  3651.                         return AVERROR_INVALIDDATA;

  3652.                     }

  3653.                     ff_vp56_init_range_decoder(&s->c_b[tile_col], data, tile_size);

  3654.                     if (vp56_rac_get_prob_branchy(&s->c_b[tile_col], 128)) { // marker bit

  3655.                         ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  3656.                         return AVERROR_INVALIDDATA;

  3657.                     }

  3658.                     data += tile_size;

  3659.                     size -= tile_size;

  3660.                 }

  3661.             }

  3662. 

  3663.             for (row = s->tiling.tile_row_start; row < s->tiling.tile_row_end;

  3664.                  row += 8, yoff += ls_y * 64, uvoff += ls_uv * 32) {

  3665.                 struct VP9Filter *lflvl_ptr = s->lflvl;

  3666.                 ptrdiff_t yoff2 = yoff, uvoff2 = uvoff;

  3667. 

  3668.                 for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  3669.                     set_tile_offset(&s->tiling.tile_col_start, &s->tiling.tile_col_end,

  3670.                                     tile_col, s->tiling.log2_tile_cols, s->sb_cols);

  3671. 

  3672.                     if (s->pass != 2) {

  3673.                         memset(s->left_partition_ctx, 0, 8);

  3674.                         memset(s->left_skip_ctx, 0, 8);

  3675.                         if (s->keyframe || s->intraonly) {

  3676.                             memset(s->left_mode_ctx, DC_PRED, 16);

  3677.                         } else {

  3678.                             memset(s->left_mode_ctx, NEARESTMV, 8);

  3679.                         }

  3680.                         memset(s->left_y_nnz_ctx, 0, 16);

  3681.                         memset(s->left_uv_nnz_ctx, 0, 16);

  3682.                         memset(s->left_segpred_ctx, 0, 8);

  3683. 

  3684.                         memcpy(&s->c, &s->c_b[tile_col], sizeof(s->c));

  3685.                     }

  3686. 

  3687.                     for (col = s->tiling.tile_col_start;

  3688.                          col < s->tiling.tile_col_end;

  3689.                          col += 8, yoff2 += 64, uvoff2 += 32, lflvl_ptr++) {

  3690.                         // FIXME integrate with lf code (i.e. zero after each

  3691.                         // use, similar to invtxfm coefficients, or similar)

  3692.                         if (s->pass != 1) {

  3693.                             memset(lflvl_ptr->mask, 0, sizeof(lflvl_ptr->mask));

  3694.                         }

  3695. 

  3696.                         if (s->pass == 2) {

  3697.                             decode_sb_mem(ctx, row, col, lflvl_ptr,

  3698.                                           yoff2, uvoff2, BL_64X64);

  3699.                         } else {

  3700.                             decode_sb(ctx, row, col, lflvl_ptr,

  3701.                                       yoff2, uvoff2, BL_64X64);

  3702.                         }

  3703.                     }

  3704.                     if (s->pass != 2) {

  3705.                         memcpy(&s->c_b[tile_col], &s->c, sizeof(s->c));

  3706.                     }

  3707.                 }

  3708. 

  3709.                 if (s->pass == 1) {

  3710.                     continue;

  3711.                 }

  3712. 

  3713.                 // backup pre-loopfilter reconstruction data for intra

  3714.                 // prediction of next row of sb64s

  3715.                 if (row + 8 < s->rows) {

  3716.                     memcpy(s->intra_pred_data[0],

  3717.                            f->data[0] + yoff + 63 * ls_y,

  3718.                            8 * s->cols);

  3719.                     memcpy(s->intra_pred_data[1],

  3720.                            f->data[1] + uvoff + 31 * ls_uv,

  3721.                            4 * s->cols);

  3722.                     memcpy(s->intra_pred_data[2],

  3723.                            f->data[2] + uvoff + 31 * ls_uv,

  3724.                            4 * s->cols);

  3725.                 }

  3726. 

  3727.                 // loopfilter one row

  3728.                 if (s->filter.level) {

  3729.                     yoff2 = yoff;

  3730.                     uvoff2 = uvoff;

  3731.                     lflvl_ptr = s->lflvl;

  3732.                     for (col = 0; col < s->cols;

  3733.                          col += 8, yoff2 += 64, uvoff2 += 32, lflvl_ptr++) {

  3734.                         loopfilter_sb(ctx, lflvl_ptr, row, col, yoff2, uvoff2);

  3735.                     }

  3736.                 }

  3737. 

  3738.                 // FIXME maybe we can make this more finegrained by running the

  3739.                 // loopfilter per-block instead of after each sbrow

  3740.                 // In fact that would also make intra pred left preparation easier?

  3741.                 ff_thread_report_progress(&s->frames[CUR_FRAME].tf, row >> 3, 0);

  3742.             }

  3743.         }

  3744. 

  3745.         if (s->pass < 2 && s->refreshctx && !s->parallelmode) {

  3746.             adapt_probs(s);

  3747.             ff_thread_finish_setup(ctx);

  3748.         }

  3749.     } while (s->pass++ == 1);

  3750.     ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  3751. 

  3752.     // ref frame setup

  3753.     for (i = 0; i < 8; i++) {

  3754.         if (s->refs[i].f->data[0])

  3755.             ff_thread_release_buffer(ctx, &s->refs[i]);

  3756.         ff_thread_ref_frame(&s->refs[i], &s->next_refs[i]);

  3757.     }

  3758. 

  3759.     if (!s->invisible) {

  3760.         if ((res = av_frame_ref(frame, s->frames[CUR_FRAME].tf.f)) < 0)

  3761.             return res;

  3762.         *got_frame = 1;

  3763.     }

  3764. 

  3765.     return 0;

  3766. }

  3767. 

  3768. static void vp9_decode_flush(AVCodecContext *ctx)

  3769. {

  3770.     VP9Context *s = ctx->priv_data;

  3771.     int i;

  3772. 

  3773.     for (i = 0; i < 2; i++)

  3774.         vp9_unref_frame(ctx, &s->frames[i]);

  3775.     for (i = 0; i < 8; i++)

  3776.         ff_thread_release_buffer(ctx, &s->refs[i]);

  3777. }

  3778. 

  3779. static int init_frames(AVCodecContext *ctx)

  3780. {

  3781.     VP9Context *s = ctx->priv_data;

  3782.     int i;

  3783. 

  3784.     for (i = 0; i < 2; i++) {

  3785.         s->frames[i].tf.f = av_frame_alloc();

  3786.         if (!s->frames[i].tf.f) {

  3787.             vp9_decode_free(ctx);

  3788.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  3789.             return AVERROR(ENOMEM);

  3790.         }

  3791.     }

  3792.     for (i = 0; i < 8; i++) {

  3793.         s->refs[i].f = av_frame_alloc();

  3794.         s->next_refs[i].f = av_frame_alloc();

  3795.         if (!s->refs[i].f || !s->next_refs[i].f) {

  3796.             vp9_decode_free(ctx);

  3797.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  3798.             return AVERROR(ENOMEM);

  3799.         }

  3800.     }

  3801. 

  3802.     return 0;

  3803. }

  3804. 

  3805. static av_cold int vp9_decode_init(AVCodecContext *ctx)

  3806. {

  3807.     VP9Context *s = ctx->priv_data;

  3808. 

  3809.     ctx->internal->allocate_progress = 1;

  3810.     ctx->pix_fmt = AV_PIX_FMT_YUV420P;

  3811.     ff_vp9dsp_init(&s->dsp);

  3812.     ff_videodsp_init(&s->vdsp, 8);

  3813.     s->filter.sharpness = -1;

  3814. 

  3815.     return init_frames(ctx);

  3816. }

  3817. 

  3818. static av_cold int vp9_decode_init_thread_copy(AVCodecContext *avctx)

  3819. {

  3820.     return init_frames(avctx);

  3821. }

  3822. 

  3823. static int vp9_decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src)

  3824. {

  3825.     int i, res;

  3826.     VP9Context *s = dst->priv_data, *ssrc = src->priv_data;

  3827. 

  3828.     // detect size changes in other threads

  3829.     if (s->intra_pred_data[0] &&

  3830.         (!ssrc->intra_pred_data[0] || s->cols != ssrc->cols || s->rows != ssrc->rows)) {

  3831.         free_buffers(s);

  3832.     }

  3833. 

  3834.     for (i = 0; i < 2; i++) {

  3835.         if (s->frames[i].tf.f->data[0])

  3836.             vp9_unref_frame(dst, &s->frames[i]);

  3837.         if (ssrc->frames[i].tf.f->data[0]) {

  3838.             if ((res = vp9_ref_frame(dst, &s->frames[i], &ssrc->frames[i])) < 0)

  3839.                 return res;

  3840.         }

  3841.     }

  3842.     for (i = 0; i < 8; i++) {

  3843.         if (s->refs[i].f->data[0])

  3844.             ff_thread_release_buffer(dst, &s->refs[i]);

  3845.         if (ssrc->next_refs[i].f->data[0]) {

  3846.             if ((res = ff_thread_ref_frame(&s->refs[i], &ssrc->next_refs[i])) < 0)

  3847.                 return res;

  3848.         }

  3849.     }

  3850. 

  3851.     s->invisible = ssrc->invisible;

  3852.     s->keyframe = ssrc->keyframe;

  3853.     s->uses_2pass = ssrc->uses_2pass;

  3854.     memcpy(&s->prob_ctx, &ssrc->prob_ctx, sizeof(s->prob_ctx));

  3855.     memcpy(&s->lf_delta, &ssrc->lf_delta, sizeof(s->lf_delta));

  3856.     if (ssrc->segmentation.enabled) {

  3857.         memcpy(&s->segmentation.feat, &ssrc->segmentation.feat,

  3858.                sizeof(s->segmentation.feat));

  3859.     }

  3860. 

  3861.     return 0;

  3862. }

  3863. 

  3864. AVCodec ff_vp9_decoder = {

  3865.     .name                  = "vp9",

  3866.     .long_name             = NULL_IF_CONFIG_SMALL("Google VP9"),

  3867.     .type                  = AVMEDIA_TYPE_VIDEO,

  3868.     .id                    = AV_CODEC_ID_VP9,

  3869.     .priv_data_size        = sizeof(VP9Context),

  3870.     .init                  = vp9_decode_init,

  3871.     .close                 = vp9_decode_free,

  3872.     .decode                = vp9_decode_frame,

  3873.     .capabilities          = CODEC_CAP_DR1 | CODEC_CAP_FRAME_THREADS,

  3874.     .flush                 = vp9_decode_flush,

  3875.     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(vp9_decode_init_thread_copy),

  3876.     .update_thread_context = ONLY_IF_THREADS_ENABLED(vp9_decode_update_thread_context),

  3877. };