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.
71636 Commits
32 Branches
303MB
Tree: e8b4f6d6be
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'e8b4f6d6be'
${ noResults }
FFmpeg/libavcodec/vp9.c

4111 lines
161KB
Raw Blame History 

  1. /*

  2.  * VP9 compatible video decoder

  3.  *

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

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

  6.  *

  7.  * This file is part of FFmpeg.

  8.  *

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

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

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

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

  13.  *

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

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

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

  17.  * Lesser General Public License for more details.

  18.  *

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

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

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

  22.  */

  23. 

  24. #include "avcodec.h"

  25. #include "get_bits.h"

  26. #include "internal.h"

  27. #include "thread.h"

  28. #include "videodsp.h"

  29. #include "vp56.h"

  30. #include "vp9.h"

  31. #include "vp9data.h"

  32. #include "vp9dsp.h"

  33. #include "libavutil/avassert.h"

  34. 

  35. #define VP9_SYNCCODE 0x498342

  36. 

  37. enum CompPredMode {

  38.     PRED_SINGLEREF,

  39.     PRED_COMPREF,

  40.     PRED_SWITCHABLE,

  41. };

  42. 

  43. enum BlockLevel {

  44.     BL_64X64,

  45.     BL_32X32,

  46.     BL_16X16,

  47.     BL_8X8,

  48. };

  49. 

  50. enum BlockSize {

  51.     BS_64x64,

  52.     BS_64x32,

  53.     BS_32x64,

  54.     BS_32x32,

  55.     BS_32x16,

  56.     BS_16x32,

  57.     BS_16x16,

  58.     BS_16x8,

  59.     BS_8x16,

  60.     BS_8x8,

  61.     BS_8x4,

  62.     BS_4x8,

  63.     BS_4x4,

  64.     N_BS_SIZES,

  65. };

  66. 

  67. struct VP9mvrefPair {

  68.     VP56mv mv[2];

  69.     int8_t ref[2];

  70. };

  71. 

  72. typedef struct VP9Frame {

  73.     ThreadFrame tf;

  74.     AVBufferRef *extradata;

  75.     uint8_t *segmentation_map;

  76.     struct VP9mvrefPair *mv;

  77.     int uses_2pass;

  78. } VP9Frame;

  79. 

  80. struct VP9Filter {

  81.     uint8_t level[8 * 8];

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

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

  84. };

  85. 

  86. typedef struct VP9Block {

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

  88.     enum FilterMode filter;

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

  90.     enum BlockSize bs;

  91.     enum TxfmMode tx, uvtx;

  92.     enum BlockLevel bl;

  93.     enum BlockPartition bp;

  94. } VP9Block;

  95. 

  96. typedef struct VP9Context {

  97.     VP9DSPContext dsp;

  98.     VideoDSPContext vdsp;

  99.     GetBitContext gb;

  100.     VP56RangeCoder c;

  101.     VP56RangeCoder *c_b;

  102.     unsigned c_b_size;

  103.     VP9Block *b_base, *b;

  104.     int pass;

  105.     int row, row7, col, col7;

  106.     uint8_t *dst[3];

  107.     ptrdiff_t y_stride, uv_stride;

  108. 

  109.     // bitstream header

  110.     uint8_t profile;

  111.     uint8_t keyframe, last_keyframe;

  112.     uint8_t invisible;

  113.     uint8_t use_last_frame_mvs;

  114.     uint8_t errorres;

  115.     uint8_t colorspace;

  116.     uint8_t fullrange;

  117.     uint8_t intraonly;

  118.     uint8_t resetctx;

  119.     uint8_t refreshrefmask;

  120.     uint8_t highprecisionmvs;

  121.     enum FilterMode filtermode;

  122.     uint8_t allowcompinter;

  123.     uint8_t fixcompref;

  124.     uint8_t refreshctx;

  125.     uint8_t parallelmode;

  126.     uint8_t framectxid;

  127.     uint8_t refidx[3];

  128.     uint8_t signbias[3];

  129.     uint8_t varcompref[2];

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

  131. #define CUR_FRAME 0

  132. #define REF_FRAME_MVPAIR 1

  133. #define REF_FRAME_SEGMAP 2

  134.     VP9Frame frames[3];

  135. 

  136.     struct {

  137.         uint8_t level;

  138.         int8_t sharpness;

  139.         uint8_t lim_lut[64];

  140.         uint8_t mblim_lut[64];

  141.     } filter;

  142.     struct {

  143.         uint8_t enabled;

  144.         int8_t mode[2];

  145.         int8_t ref[4];

  146.     } lf_delta;

  147.     uint8_t yac_qi;

  148.     int8_t ydc_qdelta, uvdc_qdelta, uvac_qdelta;

  149.     uint8_t lossless;

  150. #define MAX_SEGMENT 8

  151.     struct {

  152.         uint8_t enabled;

  153.         uint8_t temporal;

  154.         uint8_t absolute_vals;

  155.         uint8_t update_map;

  156.         struct {

  157.             uint8_t q_enabled;

  158.             uint8_t lf_enabled;

  159.             uint8_t ref_enabled;

  160.             uint8_t skip_enabled;

  161.             uint8_t ref_val;

  162.             int16_t q_val;

  163.             int8_t lf_val;

  164.             int16_t qmul[2][2];

  165.             uint8_t lflvl[4][2];

  166.         } feat[MAX_SEGMENT];

  167.     } segmentation;

  168.     struct {

  169.         unsigned log2_tile_cols, log2_tile_rows;

  170.         unsigned tile_cols, tile_rows;

  171.         unsigned tile_row_start, tile_row_end, tile_col_start, tile_col_end;

  172.     } tiling;

  173.     unsigned sb_cols, sb_rows, rows, cols;

  174.     struct {

  175.         prob_context p;

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

  177.     } prob_ctx[4];

  178.     struct {

  179.         prob_context p;

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

  181.         uint8_t seg[7];

  182.         uint8_t segpred[3];

  183.     } prob;

  184.     struct {

  185.         unsigned y_mode[4][10];

  186.         unsigned uv_mode[10][10];

  187.         unsigned filter[4][3];

  188.         unsigned mv_mode[7][4];

  189.         unsigned intra[4][2];

  190.         unsigned comp[5][2];

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

  192.         unsigned comp_ref[5][2];

  193.         unsigned tx32p[2][4];

  194.         unsigned tx16p[2][3];

  195.         unsigned tx8p[2][2];

  196.         unsigned skip[3][2];

  197.         unsigned mv_joint[4];

  198.         struct {

  199.             unsigned sign[2];

  200.             unsigned classes[11];

  201.             unsigned class0[2];

  202.             unsigned bits[10][2];

  203.             unsigned class0_fp[2][4];

  204.             unsigned fp[4];

  205.             unsigned class0_hp[2];

  206.             unsigned hp[2];

  207.         } mv_comp[2];

  208.         unsigned partition[4][4][4];

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

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

  211.     } counts;

  212.     enum TxfmMode txfmmode;

  213.     enum CompPredMode comppredmode;

  214. 

  215.     // contextual (left/above) cache

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

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

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

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

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

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

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

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

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

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

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

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

  228.     uint8_t *above_partition_ctx;

  229.     uint8_t *above_mode_ctx;

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

  231.     uint8_t *above_y_nnz_ctx;

  232.     uint8_t *above_uv_nnz_ctx[2];

  233.     uint8_t *above_skip_ctx; // 1bit

  234.     uint8_t *above_txfm_ctx; // 2bit

  235.     uint8_t *above_segpred_ctx; // 1bit

  236.     uint8_t *above_intra_ctx; // 1bit

  237.     uint8_t *above_comp_ctx; // 1bit

  238.     uint8_t *above_ref_ctx; // 2bit

  239.     uint8_t *above_filter_ctx;

  240.     VP56mv (*above_mv_ctx)[2];

  241. 

  242.     // whole-frame cache

  243.     uint8_t *intra_pred_data[3];

  244.     struct VP9Filter *lflvl;

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

  246. 

  247.     // block reconstruction intermediates

  248.     int block_alloc_using_2pass;

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

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

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

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

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

  254.     uint16_t mvscale[3][2];

  255.     uint8_t mvstep[3][2];

  256. } VP9Context;

  257. 

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

  259.     {

  260.         { 16, 16 }, { 16, 8 }, { 8, 16 }, { 8, 8 }, { 8, 4 }, { 4, 8 },

  261.         { 4, 4 }, { 4, 2 }, { 2, 4 }, { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 },

  262.     }, {

  263.         { 8, 8 }, { 8, 4 }, { 4, 8 }, { 4, 4 }, { 4, 2 }, { 2, 4 },

  264.         { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 }, { 1, 1 }, { 1, 1 }, { 1, 1 },

  265.     }

  266. };

  267. 

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

  269. {

  270.     VP9Context *s = ctx->priv_data;

  271.     int ret, sz;

  272. 

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

  274.         return ret;

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

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

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

  278.         return AVERROR(ENOMEM);

  279.     }

  280. 

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

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

  283. 

  284.     return 0;

  285. }

  286. 

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

  288. {

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

  290.     av_buffer_unref(&f->extradata);

  291. }

  292. 

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

  294. {

  295.     int res;

  296. 

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

  298.         return res;

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

  300.         vp9_unref_frame(ctx, dst);

  301.         return AVERROR(ENOMEM);

  302.     }

  303. 

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

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

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

  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->frames[CUR_FRAME].uses_2pass)

  364.         return 0;

  365. 

  366.     av_free(s->b_base);

  367.     av_free(s->block_base);

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

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

  370. 

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

  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->frames[CUR_FRAME].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_keyframe  = s->keyframe;

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

  494.     last_invisible    = s->invisible;

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

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

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

  498.     if (s->keyframe) {

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

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

  501.             return AVERROR_INVALIDDATA;

  502.         }

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

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

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

  506.             return AVERROR_INVALIDDATA;

  507.         }

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

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

  510.         s->refreshrefmask = 0xff;

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

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

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

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

  515.     } else {

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

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

  518.         if (s->intraonly) {

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

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

  521.                 return AVERROR_INVALIDDATA;

  522.             }

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

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

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

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

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

  528.         } else {

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

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

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

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

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

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

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

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

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

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

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

  540.                 return AVERROR_INVALIDDATA;

  541.             }

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

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

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

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

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

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

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

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

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

  551.             } else {

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

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

  554.             }

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

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

  557.             // the _last_ frame

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

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

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

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

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

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

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

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

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

  567.             if (s->allowcompinter) {

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

  569.                     s->fixcompref    = 2;

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

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

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

  573.                     s->fixcompref    = 1;

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

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

  576.                 } else {

  577.                     s->fixcompref    = 0;

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

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

  580.                 }

  581.             }

  582. 

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

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

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

  586. 

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

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

  589.                 } else {

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

  591.                         av_log(ctx, AV_LOG_ERROR,

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

  593.                                refw, refh, w, h);

  594.                         return AVERROR_INVALIDDATA;

  595.                     }

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

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

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

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

  600.                 }

  601.             }

  602.         }

  603.     }

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

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

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

  607. 

  608.     /* loopfilter header data */

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

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

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

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

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

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

  615.     s->filter.sharpness = sharp;

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

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

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

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

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

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

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

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

  624.         }

  625.     } else {

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

  627.     }

  628. 

  629.     /* quantization header data */

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

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

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

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

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

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

  636. 

  637.     /* segmentation header info */

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

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

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

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

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

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

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

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

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

  647.             }

  648.         }

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

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

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

  652.             av_log(ctx, AV_LOG_ERROR,

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

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

  655.             return AVERROR_INVALIDDATA;

  656.         }

  657. 

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

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

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

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

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

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

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

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

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

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

  668.             }

  669.         }

  670.     } else {

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

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

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

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

  675.     }

  676. 

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

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

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

  680. 

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

  682.             if (s->segmentation.absolute_vals)

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

  684.             else

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

  686.         } else {

  687.             qyac  = s->yac_qi;

  688.         }

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

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

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

  692.         qyac  = av_clip_uintp2(qyac, 8);

  693. 

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

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

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

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

  698. 

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

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

  701.             if (s->segmentation.absolute_vals)

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

  703.             else

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

  705.         } else {

  706.             lflvl  = s->filter.level;

  707.         }

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

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

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

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

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

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

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

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

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

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

  718.         }

  719.     }

  720. 

  721.     /* tiling info */

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

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

  724.         return res;

  725.     }

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

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

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

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

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

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

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

  733.             s->tiling.log2_tile_cols++;

  734.         else

  735.             break;

  736.     }

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

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

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

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

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

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

  743.         if (!s->c_b) {

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

  745.             return AVERROR(ENOMEM);

  746.         }

  747.     }

  748. 

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

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

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

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

  753.                sizeof(vp9_default_coef_probs));

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

  755.                sizeof(vp9_default_coef_probs));

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

  757.                sizeof(vp9_default_coef_probs));

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

  759.                sizeof(vp9_default_coef_probs));

  760.     }

  761. 

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

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

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

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

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

  767.         return AVERROR_INVALIDDATA;

  768.     }

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

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

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

  772.         return AVERROR_INVALIDDATA;

  773.     }

  774. 

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

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

  777.     } else {

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

  779.     }

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

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

  782.     // fw update)?

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

  784. 

  785.     // txfm updates

  786.     if (s->lossless) {

  787.         s->txfmmode = TX_4X4;

  788.     } else {

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

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

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

  792. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  807.         }

  808.     }

  809. 

  810.     // coef updates

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

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

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

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

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

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

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

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

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

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

  821.                                 break;

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

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

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

  825.                                 } else {

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

  827.                                 }

  828.                             }

  829.                             p[3] = 0;

  830.                         }

  831.         } else {

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

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

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

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

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

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

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

  839.                                 break;

  840.                             memcpy(p, r, 3);

  841.                             p[3] = 0;

  842.                         }

  843.         }

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

  845.             break;

  846.     }

  847. 

  848.     // mode updates

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

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

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

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

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

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

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

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

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

  858. 

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

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

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

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

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

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

  865. 

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

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

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

  869. 

  870.         if (s->allowcompinter) {

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

  872.             if (s->comppredmode)

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

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

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

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

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

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

  879.         } else {

  880.             s->comppredmode = PRED_SINGLEREF;

  881.         }

  882. 

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

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

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

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

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

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

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

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

  891.             }

  892.         }

  893. 

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

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

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

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

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

  899.         }

  900. 

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

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

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

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

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

  906. 

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

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

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

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

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

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

  913. 

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

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

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

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

  918. 

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

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

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

  922. 

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

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

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

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

  927. 

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

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

  930. 

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

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

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

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

  935.         }

  936. 

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

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

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

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

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

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

  943. 

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

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

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

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

  948.         }

  949. 

  950.         if (s->highprecisionmvs) {

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

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

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

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

  955. 

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

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

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

  959.             }

  960.         }

  961.     }

  962. 

  963.     return (data2 - data) + size2;

  964. }

  965. 

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

  967.                                       VP9Context *s)

  968. {

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

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

  971. }

  972. 

  973. static void find_ref_mvs(VP9Context *s,

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

  975. {

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

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

  978.                       { -1, -1 }, {  0, -1 }, { -1,  0 }, {  6, -1 }},

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

  980.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, {  2, -1 }},

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

  982.                       { -1, -1 }, { -3,  0 }, {  0, -3 }, { -1,  2 }},

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

  984.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  986.                       { -1,  1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  988.                       {  1, -1 }, { -3,  0 }, {  0, -3 }, { -3, -3 }},

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

  990.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  992.                       {  0, -2 }, { -2,  0 }, { -2, -1 }, { -1, -2 }},

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

  994.                       { -2,  0 }, {  0, -2 }, { -1, -2 }, { -2, -1 }},

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

  996.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

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

  998.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

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

  1000.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1001.         [BS_4x4]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1002.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1003.     };

  1004.     VP9Block *b = s->b;

  1005.     int row = s->row, col = s->col, row7 = s->row7;

  1006.     const int8_t (*p)[2] = mv_ref_blk_off[b->bs];

  1007. #define INVALID_MV 0x80008000U

  1008.     uint32_t mem = INVALID_MV;

  1009.     int i;

  1010. 

  1011. #define RETURN_DIRECT_MV(mv) \

  1012.     do { \

  1013.         uint32_t m = AV_RN32A(&mv); \

  1014.         if (!idx) { \

  1015.             AV_WN32A(pmv, m); \

  1016.             return; \

  1017.         } else if (mem == INVALID_MV) { \

  1018.             mem = m; \

  1019.         } else if (m != mem) { \

  1020.             AV_WN32A(pmv, m); \

  1021.             return; \

  1022.         } \

  1023.     } while (0)

  1024. 

  1025.     if (sb >= 0) {

  1026.         if (sb == 2 || sb == 1) {

  1027.             RETURN_DIRECT_MV(b->mv[0][z]);

  1028.         } else if (sb == 3) {

  1029.             RETURN_DIRECT_MV(b->mv[2][z]);

  1030.             RETURN_DIRECT_MV(b->mv[1][z]);

  1031.             RETURN_DIRECT_MV(b->mv[0][z]);

  1032.         }

  1033. 

  1034. #define RETURN_MV(mv) \

  1035.     do { \

  1036.         if (sb > 0) { \

  1037.             VP56mv tmp; \

  1038.             uint32_t m; \

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

  1040.             m = AV_RN32A(&tmp); \

  1041.             if (!idx) { \

  1042.                 AV_WN32A(pmv, m); \

  1043.                 return; \

  1044.             } else if (mem == INVALID_MV) { \

  1045.                 mem = m; \

  1046.             } else if (m != mem) { \

  1047.                 AV_WN32A(pmv, m); \

  1048.                 return; \

  1049.             } \

  1050.         } else { \

  1051.             uint32_t m = AV_RN32A(&mv); \

  1052.             if (!idx) { \

  1053.                 clamp_mv(pmv, &mv, s); \

  1054.                 return; \

  1055.             } else if (mem == INVALID_MV) { \

  1056.                 mem = m; \

  1057.             } else if (m != mem) { \

  1058.                 clamp_mv(pmv, &mv, s); \

  1059.                 return; \

  1060.             } \

  1061.         } \

  1062.     } while (0)

  1063. 

  1064.         if (row > 0) {

  1065.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[(row - 1) * s->sb_cols * 8 + col];

  1066.             if (mv->ref[0] == ref) {

  1067.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][0]);

  1068.             } else if (mv->ref[1] == ref) {

  1069.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][1]);

  1070.             }

  1071.         }

  1072.         if (col > s->tiling.tile_col_start) {

  1073.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[row * s->sb_cols * 8 + col - 1];

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

  1075.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][0]);

  1076.             } else if (mv->ref[1] == ref) {

  1077.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][1]);

  1078.             }

  1079.         }

  1080.         i = 2;

  1081.     } else {

  1082.         i = 0;

  1083.     }

  1084. 

  1085.     // previously coded MVs in this neighbourhood, using same reference frame

  1086.     for (; i < 8; i++) {

  1087.         int c = p[i][0] + col, r = p[i][1] + row;

  1088. 

  1089.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1090.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1091. 

  1092.             if (mv->ref[0] == ref) {

  1093.                 RETURN_MV(mv->mv[0]);

  1094.             } else if (mv->ref[1] == ref) {

  1095.                 RETURN_MV(mv->mv[1]);

  1096.             }

  1097.         }

  1098.     }

  1099. 

  1100.     // MV at this position in previous frame, using same reference frame

  1101.     if (s->use_last_frame_mvs) {

  1102.         struct VP9mvrefPair *mv = &s->frames[REF_FRAME_MVPAIR].mv[row * s->sb_cols * 8 + col];

  1103. 

  1104.         if (!s->frames[REF_FRAME_MVPAIR].uses_2pass)

  1105.             ff_thread_await_progress(&s->frames[REF_FRAME_MVPAIR].tf, row >> 3, 0);

  1106.         if (mv->ref[0] == ref) {

  1107.             RETURN_MV(mv->mv[0]);

  1108.         } else if (mv->ref[1] == ref) {

  1109.             RETURN_MV(mv->mv[1]);

  1110.         }

  1111.     }

  1112. 

  1113. #define RETURN_SCALE_MV(mv, scale) \

  1114.     do { \

  1115.         if (scale) { \

  1116.             VP56mv mv_temp = { -mv.x, -mv.y }; \

  1117.             RETURN_MV(mv_temp); \

  1118.         } else { \

  1119.             RETURN_MV(mv); \

  1120.         } \

  1121.     } while (0)

  1122. 

  1123.     // previously coded MVs in this neighbourhood, using different reference frame

  1124.     for (i = 0; i < 8; i++) {

  1125.         int c = p[i][0] + col, r = p[i][1] + row;

  1126. 

  1127.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1128.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1129. 

  1130.             if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1131.                 RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1132.             }

  1133.             if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1134.                 // BUG - libvpx has this condition regardless of whether

  1135.                 // we used the first ref MV and pre-scaling

  1136.                 AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1137.                 RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1138.             }

  1139.         }

  1140.     }

  1141. 

  1142.     // MV at this position in previous frame, using different reference frame

  1143.     if (s->use_last_frame_mvs) {

  1144.         struct VP9mvrefPair *mv = &s->frames[REF_FRAME_MVPAIR].mv[row * s->sb_cols * 8 + col];

  1145. 

  1146.         // no need to await_progress, because we already did that above

  1147.         if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1148.             RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1149.         }

  1150.         if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1151.             // BUG - libvpx has this condition regardless of whether

  1152.             // we used the first ref MV and pre-scaling

  1153.             AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1154.             RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1155.         }

  1156.     }

  1157. 

  1158.     AV_ZERO32(pmv);

  1159. #undef INVALID_MV

  1160. #undef RETURN_MV

  1161. #undef RETURN_SCALE_MV

  1162. }

  1163. 

  1164. static av_always_inline int read_mv_component(VP9Context *s, int idx, int hp)

  1165. {

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

  1167.     int n, c = vp8_rac_get_tree(&s->c, vp9_mv_class_tree,

  1168.                                 s->prob.p.mv_comp[idx].classes);

  1169. 

  1170.     s->counts.mv_comp[idx].sign[sign]++;

  1171.     s->counts.mv_comp[idx].classes[c]++;

  1172.     if (c) {

  1173.         int m;

  1174. 

  1175.         for (n = 0, m = 0; m < c; m++) {

  1176.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].bits[m]);

  1177.             n |= bit << m;

  1178.             s->counts.mv_comp[idx].bits[m][bit]++;

  1179.         }

  1180.         n <<= 3;

  1181.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree, s->prob.p.mv_comp[idx].fp);

  1182.         n |= bit << 1;

  1183.         s->counts.mv_comp[idx].fp[bit]++;

  1184.         if (hp) {

  1185.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].hp);

  1186.             s->counts.mv_comp[idx].hp[bit]++;

  1187.             n |= bit;

  1188.         } else {

  1189.             n |= 1;

  1190.             // bug in libvpx - we count for bw entropy purposes even if the

  1191.             // bit wasn't coded

  1192.             s->counts.mv_comp[idx].hp[1]++;

  1193.         }

  1194.         n += 8 << c;

  1195.     } else {

  1196.         n = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0);

  1197.         s->counts.mv_comp[idx].class0[n]++;

  1198.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree,

  1199.                                s->prob.p.mv_comp[idx].class0_fp[n]);

  1200.         s->counts.mv_comp[idx].class0_fp[n][bit]++;

  1201.         n = (n << 3) | (bit << 1);

  1202.         if (hp) {

  1203.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0_hp);

  1204.             s->counts.mv_comp[idx].class0_hp[bit]++;

  1205.             n |= bit;

  1206.         } else {

  1207.             n |= 1;

  1208.             // bug in libvpx - we count for bw entropy purposes even if the

  1209.             // bit wasn't coded

  1210.             s->counts.mv_comp[idx].class0_hp[1]++;

  1211.         }

  1212.     }

  1213. 

  1214.     return sign ? -(n + 1) : (n + 1);

  1215. }

  1216. 

  1217. static void fill_mv(VP9Context *s,

  1218.                     VP56mv *mv, int mode, int sb)

  1219. {

  1220.     VP9Block *b = s->b;

  1221. 

  1222.     if (mode == ZEROMV) {

  1223.         AV_ZERO64(mv);

  1224.     } else {

  1225.         int hp;

  1226. 

  1227.         // FIXME cache this value and reuse for other subblocks

  1228.         find_ref_mvs(s, &mv[0], b->ref[0], 0, mode == NEARMV,

  1229.                      mode == NEWMV ? -1 : sb);

  1230.         // FIXME maybe move this code into find_ref_mvs()

  1231.         if ((mode == NEWMV || sb == -1) &&

  1232.             !(hp = s->highprecisionmvs && abs(mv[0].x) < 64 && abs(mv[0].y) < 64)) {

  1233.             if (mv[0].y & 1) {

  1234.                 if (mv[0].y < 0)

  1235.                     mv[0].y++;

  1236.                 else

  1237.                     mv[0].y--;

  1238.             }

  1239.             if (mv[0].x & 1) {

  1240.                 if (mv[0].x < 0)

  1241.                     mv[0].x++;

  1242.                 else

  1243.                     mv[0].x--;

  1244.             }

  1245.         }

  1246.         if (mode == NEWMV) {

  1247.             enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1248.                                               s->prob.p.mv_joint);

  1249. 

  1250.             s->counts.mv_joint[j]++;

  1251.             if (j >= MV_JOINT_V)

  1252.                 mv[0].y += read_mv_component(s, 0, hp);

  1253.             if (j & 1)

  1254.                 mv[0].x += read_mv_component(s, 1, hp);

  1255.         }

  1256. 

  1257.         if (b->comp) {

  1258.             // FIXME cache this value and reuse for other subblocks

  1259.             find_ref_mvs(s, &mv[1], b->ref[1], 1, mode == NEARMV,

  1260.                          mode == NEWMV ? -1 : sb);

  1261.             if ((mode == NEWMV || sb == -1) &&

  1262.                 !(hp = s->highprecisionmvs && abs(mv[1].x) < 64 && abs(mv[1].y) < 64)) {

  1263.                 if (mv[1].y & 1) {

  1264.                     if (mv[1].y < 0)

  1265.                         mv[1].y++;

  1266.                     else

  1267.                         mv[1].y--;

  1268.                 }

  1269.                 if (mv[1].x & 1) {

  1270.                     if (mv[1].x < 0)

  1271.                         mv[1].x++;

  1272.                     else

  1273.                         mv[1].x--;

  1274.                 }

  1275.             }

  1276.             if (mode == NEWMV) {

  1277.                 enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1278.                                                   s->prob.p.mv_joint);

  1279. 

  1280.                 s->counts.mv_joint[j]++;

  1281.                 if (j >= MV_JOINT_V)

  1282.                     mv[1].y += read_mv_component(s, 0, hp);

  1283.                 if (j & 1)

  1284.                     mv[1].x += read_mv_component(s, 1, hp);

  1285.             }

  1286.         }

  1287.     }

  1288. }

  1289. 

  1290. static av_always_inline void setctx_2d(uint8_t *ptr, int w, int h,

  1291.                                        ptrdiff_t stride, int v)

  1292. {

  1293.     switch (w) {

  1294.     case 1:

  1295.         do {

  1296.             *ptr = v;

  1297.             ptr += stride;

  1298.         } while (--h);

  1299.         break;

  1300.     case 2: {

  1301.         int v16 = v * 0x0101;

  1302.         do {

  1303.             AV_WN16A(ptr, v16);

  1304.             ptr += stride;

  1305.         } while (--h);

  1306.         break;

  1307.     }

  1308.     case 4: {

  1309.         uint32_t v32 = v * 0x01010101;

  1310.         do {

  1311.             AV_WN32A(ptr, v32);

  1312.             ptr += stride;

  1313.         } while (--h);

  1314.         break;

  1315.     }

  1316.     case 8: {

  1317. #if HAVE_FAST_64BIT

  1318.         uint64_t v64 = v * 0x0101010101010101ULL;

  1319.         do {

  1320.             AV_WN64A(ptr, v64);

  1321.             ptr += stride;

  1322.         } while (--h);

  1323. #else

  1324.         uint32_t v32 = v * 0x01010101;

  1325.         do {

  1326.             AV_WN32A(ptr,     v32);

  1327.             AV_WN32A(ptr + 4, v32);

  1328.             ptr += stride;

  1329.         } while (--h);

  1330. #endif

  1331.         break;

  1332.     }

  1333.     }

  1334. }

  1335. 

  1336. static void decode_mode(AVCodecContext *ctx)

  1337. {

  1338.     static const uint8_t left_ctx[N_BS_SIZES] = {

  1339.         0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf

  1340.     };

  1341.     static const uint8_t above_ctx[N_BS_SIZES] = {

  1342.         0x0, 0x0, 0x8, 0x8, 0x8, 0xc, 0xc, 0xc, 0xe, 0xe, 0xe, 0xf, 0xf

  1343.     };

  1344.     static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {

  1345.         TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,

  1346.         TX_16X16, TX_8X8, TX_8X8, TX_8X8, TX_4X4, TX_4X4, TX_4X4

  1347.     };

  1348.     VP9Context *s = ctx->priv_data;

  1349.     VP9Block *b = s->b;

  1350.     int row = s->row, col = s->col, row7 = s->row7;

  1351.     enum TxfmMode max_tx = max_tx_for_bl_bp[b->bs];

  1352.     int w4 = FFMIN(s->cols - col, bwh_tab[1][b->bs][0]);

  1353.     int h4 = FFMIN(s->rows - row, bwh_tab[1][b->bs][1]), y;

  1354.     int have_a = row > 0, have_l = col > s->tiling.tile_col_start;

  1355.     int vref, filter_id;

  1356. 

  1357.     if (!s->segmentation.enabled) {

  1358.         b->seg_id = 0;

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

  1360.         b->seg_id = vp8_rac_get_tree(&s->c, vp9_segmentation_tree, s->prob.seg);

  1361.     } else if (!s->segmentation.update_map ||

  1362.                (s->segmentation.temporal &&

  1363.                 vp56_rac_get_prob_branchy(&s->c,

  1364.                     s->prob.segpred[s->above_segpred_ctx[col] +

  1365.                                     s->left_segpred_ctx[row7]]))) {

  1366.         if (!s->errorres) {

  1367.             int pred = 8, x;

  1368.             uint8_t *refsegmap = s->frames[REF_FRAME_SEGMAP].segmentation_map;

  1369. 

  1370.             if (!s->frames[REF_FRAME_SEGMAP].uses_2pass)

  1371.                 ff_thread_await_progress(&s->frames[REF_FRAME_SEGMAP].tf, row >> 3, 0);

  1372.             for (y = 0; y < h4; y++) {

  1373.                 int idx_base = (y + row) * 8 * s->sb_cols + col;

  1374.                 for (x = 0; x < w4; x++)

  1375.                     pred = FFMIN(pred, refsegmap[idx_base + x]);

  1376.             }

  1377.             av_assert1(pred < 8);

  1378.             b->seg_id = pred;

  1379.         } else {

  1380.             b->seg_id = 0;

  1381.         }

  1382. 

  1383.         memset(&s->above_segpred_ctx[col], 1, w4);

  1384.         memset(&s->left_segpred_ctx[row7], 1, h4);

  1385.     } else {

  1386.         b->seg_id = vp8_rac_get_tree(&s->c, vp9_segmentation_tree,

  1387.                                      s->prob.seg);

  1388. 

  1389.         memset(&s->above_segpred_ctx[col], 0, w4);

  1390.         memset(&s->left_segpred_ctx[row7], 0, h4);

  1391.     }

  1392.     if (s->segmentation.enabled &&

  1393.         (s->segmentation.update_map || s->keyframe || s->intraonly)) {

  1394.         setctx_2d(&s->frames[CUR_FRAME].segmentation_map[row * 8 * s->sb_cols + col],

  1395.                   w4, h4, 8 * s->sb_cols, b->seg_id);

  1396.     }

  1397. 

  1398.     b->skip = s->segmentation.enabled &&

  1399.         s->segmentation.feat[b->seg_id].skip_enabled;

  1400.     if (!b->skip) {

  1401.         int c = s->left_skip_ctx[row7] + s->above_skip_ctx[col];

  1402.         b->skip = vp56_rac_get_prob(&s->c, s->prob.p.skip[c]);

  1403.         s->counts.skip[c][b->skip]++;

  1404.     }

  1405. 

  1406.     if (s->keyframe || s->intraonly) {

  1407.         b->intra = 1;

  1408.     } else if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1409.         b->intra = !s->segmentation.feat[b->seg_id].ref_val;

  1410.     } else {

  1411.         int c, bit;

  1412. 

  1413.         if (have_a && have_l) {

  1414.             c = s->above_intra_ctx[col] + s->left_intra_ctx[row7];

  1415.             c += (c == 2);

  1416.         } else {

  1417.             c = have_a ? 2 * s->above_intra_ctx[col] :

  1418.                 have_l ? 2 * s->left_intra_ctx[row7] : 0;

  1419.         }

  1420.         bit = vp56_rac_get_prob(&s->c, s->prob.p.intra[c]);

  1421.         s->counts.intra[c][bit]++;

  1422.         b->intra = !bit;

  1423.     }

  1424. 

  1425.     if ((b->intra || !b->skip) && s->txfmmode == TX_SWITCHABLE) {

  1426.         int c;

  1427.         if (have_a) {

  1428.             if (have_l) {

  1429.                 c = (s->above_skip_ctx[col] ? max_tx :

  1430.                      s->above_txfm_ctx[col]) +

  1431.                     (s->left_skip_ctx[row7] ? max_tx :

  1432.                      s->left_txfm_ctx[row7]) > max_tx;

  1433.             } else {

  1434.                 c = s->above_skip_ctx[col] ? 1 :

  1435.                     (s->above_txfm_ctx[col] * 2 > max_tx);

  1436.             }

  1437.         } else if (have_l) {

  1438.             c = s->left_skip_ctx[row7] ? 1 :

  1439.                 (s->left_txfm_ctx[row7] * 2 > max_tx);

  1440.         } else {

  1441.             c = 1;

  1442.         }

  1443.         switch (max_tx) {

  1444.         case TX_32X32:

  1445.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][0]);

  1446.             if (b->tx) {

  1447.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][1]);

  1448.                 if (b->tx == 2)

  1449.                     b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][2]);

  1450.             }

  1451.             s->counts.tx32p[c][b->tx]++;

  1452.             break;

  1453.         case TX_16X16:

  1454.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][0]);

  1455.             if (b->tx)

  1456.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][1]);

  1457.             s->counts.tx16p[c][b->tx]++;

  1458.             break;

  1459.         case TX_8X8:

  1460.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx8p[c]);

  1461.             s->counts.tx8p[c][b->tx]++;

  1462.             break;

  1463.         case TX_4X4:

  1464.             b->tx = TX_4X4;

  1465.             break;

  1466.         }

  1467.     } else {

  1468.         b->tx = FFMIN(max_tx, s->txfmmode);

  1469.     }

  1470. 

  1471.     if (s->keyframe || s->intraonly) {

  1472.         uint8_t *a = &s->above_mode_ctx[col * 2];

  1473.         uint8_t *l = &s->left_mode_ctx[(row7) << 1];

  1474. 

  1475.         b->comp = 0;

  1476.         if (b->bs > BS_8x8) {

  1477.             // FIXME the memory storage intermediates here aren't really

  1478.             // necessary, they're just there to make the code slightly

  1479.             // simpler for now

  1480.             b->mode[0] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1481.                                     vp9_default_kf_ymode_probs[a[0]][l[0]]);

  1482.             if (b->bs != BS_8x4) {

  1483.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1484.                                  vp9_default_kf_ymode_probs[a[1]][b->mode[0]]);

  1485.                 l[0] = a[1] = b->mode[1];

  1486.             } else {

  1487.                 l[0] = a[1] = b->mode[1] = b->mode[0];

  1488.             }

  1489.             if (b->bs != BS_4x8) {

  1490.                 b->mode[2] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1491.                                         vp9_default_kf_ymode_probs[a[0]][l[1]]);

  1492.                 if (b->bs != BS_8x4) {

  1493.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1494.                                   vp9_default_kf_ymode_probs[a[1]][b->mode[2]]);

  1495.                     l[1] = a[1] = b->mode[3];

  1496.                 } else {

  1497.                     l[1] = a[1] = b->mode[3] = b->mode[2];

  1498.                 }

  1499.             } else {

  1500.                 b->mode[2] = b->mode[0];

  1501.                 l[1] = a[1] = b->mode[3] = b->mode[1];

  1502.             }

  1503.         } else {

  1504.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1505.                                           vp9_default_kf_ymode_probs[*a][*l]);

  1506.             b->mode[3] = b->mode[2] = b->mode[1] = b->mode[0];

  1507.             // FIXME this can probably be optimized

  1508.             memset(a, b->mode[0], bwh_tab[0][b->bs][0]);

  1509.             memset(l, b->mode[0], bwh_tab[0][b->bs][1]);

  1510.         }

  1511.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1512.                                      vp9_default_kf_uvmode_probs[b->mode[3]]);

  1513.     } else if (b->intra) {

  1514.         b->comp = 0;

  1515.         if (b->bs > BS_8x8) {

  1516.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1517.                                           s->prob.p.y_mode[0]);

  1518.             s->counts.y_mode[0][b->mode[0]]++;

  1519.             if (b->bs != BS_8x4) {

  1520.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1521.                                               s->prob.p.y_mode[0]);

  1522.                 s->counts.y_mode[0][b->mode[1]]++;

  1523.             } else {

  1524.                 b->mode[1] = b->mode[0];

  1525.             }

  1526.             if (b->bs != BS_4x8) {

  1527.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1528.                                               s->prob.p.y_mode[0]);

  1529.                 s->counts.y_mode[0][b->mode[2]]++;

  1530.                 if (b->bs != BS_8x4) {

  1531.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1532.                                                   s->prob.p.y_mode[0]);

  1533.                     s->counts.y_mode[0][b->mode[3]]++;

  1534.                 } else {

  1535.                     b->mode[3] = b->mode[2];

  1536.                 }

  1537.             } else {

  1538.                 b->mode[2] = b->mode[0];

  1539.                 b->mode[3] = b->mode[1];

  1540.             }

  1541.         } else {

  1542.             static const uint8_t size_group[10] = {

  1543.                 3, 3, 3, 3, 2, 2, 2, 1, 1, 1

  1544.             };

  1545.             int sz = size_group[b->bs];

  1546. 

  1547.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1548.                                           s->prob.p.y_mode[sz]);

  1549.             b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1550.             s->counts.y_mode[sz][b->mode[3]]++;

  1551.         }

  1552.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1553.                                      s->prob.p.uv_mode[b->mode[3]]);

  1554.         s->counts.uv_mode[b->mode[3]][b->uvmode]++;

  1555.     } else {

  1556.         static const uint8_t inter_mode_ctx_lut[14][14] = {

  1557.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1558.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1559.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1560.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1561.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1562.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1563.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1564.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1565.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1566.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1567.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1568.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1569.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },

  1570.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },

  1571.         };

  1572. 

  1573.         if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1574.             av_assert2(s->segmentation.feat[b->seg_id].ref_val != 0);

  1575.             b->comp = 0;

  1576.             b->ref[0] = s->segmentation.feat[b->seg_id].ref_val - 1;

  1577.         } else {

  1578.             // read comp_pred flag

  1579.             if (s->comppredmode != PRED_SWITCHABLE) {

  1580.                 b->comp = s->comppredmode == PRED_COMPREF;

  1581.             } else {

  1582.                 int c;

  1583. 

  1584.                 // FIXME add intra as ref=0xff (or -1) to make these easier?

  1585.                 if (have_a) {

  1586.                     if (have_l) {

  1587.                         if (s->above_comp_ctx[col] && s->left_comp_ctx[row7]) {

  1588.                             c = 4;

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

  1590.                             c = 2 + (s->left_intra_ctx[row7] ||

  1591.                                      s->left_ref_ctx[row7] == s->fixcompref);

  1592.                         } else if (s->left_comp_ctx[row7]) {

  1593.                             c = 2 + (s->above_intra_ctx[col] ||

  1594.                                      s->above_ref_ctx[col] == s->fixcompref);

  1595.                         } else {

  1596.                             c = (!s->above_intra_ctx[col] &&

  1597.                                  s->above_ref_ctx[col] == s->fixcompref) ^

  1598.                             (!s->left_intra_ctx[row7] &&

  1599.                              s->left_ref_ctx[row & 7] == s->fixcompref);

  1600.                         }

  1601.                     } else {

  1602.                         c = s->above_comp_ctx[col] ? 3 :

  1603.                         (!s->above_intra_ctx[col] && s->above_ref_ctx[col] == s->fixcompref);

  1604.                     }

  1605.                 } else if (have_l) {

  1606.                     c = s->left_comp_ctx[row7] ? 3 :

  1607.                     (!s->left_intra_ctx[row7] && s->left_ref_ctx[row7] == s->fixcompref);

  1608.                 } else {

  1609.                     c = 1;

  1610.                 }

  1611.                 b->comp = vp56_rac_get_prob(&s->c, s->prob.p.comp[c]);

  1612.                 s->counts.comp[c][b->comp]++;

  1613.             }

  1614. 

  1615.             // read actual references

  1616.             // FIXME probably cache a few variables here to prevent repetitive

  1617.             // memory accesses below

  1618.             if (b->comp) /* two references */ {

  1619.                 int fix_idx = s->signbias[s->fixcompref], var_idx = !fix_idx, c, bit;

  1620. 

  1621.                 b->ref[fix_idx] = s->fixcompref;

  1622.                 // FIXME can this codeblob be replaced by some sort of LUT?

  1623.                 if (have_a) {

  1624.                     if (have_l) {

  1625.                         if (s->above_intra_ctx[col]) {

  1626.                             if (s->left_intra_ctx[row7]) {

  1627.                                 c = 2;

  1628.                             } else {

  1629.                                 c = 1 + 2 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1630.                             }

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

  1632.                             c = 1 + 2 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1633.                         } else {

  1634.                             int refl = s->left_ref_ctx[row7], refa = s->above_ref_ctx[col];

  1635. 

  1636.                             if (refl == refa && refa == s->varcompref[1]) {

  1637.                                 c = 0;

  1638.                             } else if (!s->left_comp_ctx[row7] && !s->above_comp_ctx[col]) {

  1639.                                 if ((refa == s->fixcompref && refl == s->varcompref[0]) ||

  1640.                                     (refl == s->fixcompref && refa == s->varcompref[0])) {

  1641.                                     c = 4;

  1642.                                 } else {

  1643.                                     c = (refa == refl) ? 3 : 1;

  1644.                                 }

  1645.                             } else if (!s->left_comp_ctx[row7]) {

  1646.                                 if (refa == s->varcompref[1] && refl != s->varcompref[1]) {

  1647.                                     c = 1;

  1648.                                 } else {

  1649.                                     c = (refl == s->varcompref[1] &&

  1650.                                          refa != s->varcompref[1]) ? 2 : 4;

  1651.                                 }

  1652.                             } else if (!s->above_comp_ctx[col]) {

  1653.                                 if (refl == s->varcompref[1] && refa != s->varcompref[1]) {

  1654.                                     c = 1;

  1655.                                 } else {

  1656.                                     c = (refa == s->varcompref[1] &&

  1657.                                          refl != s->varcompref[1]) ? 2 : 4;

  1658.                                 }

  1659.                             } else {

  1660.                                 c = (refl == refa) ? 4 : 2;

  1661.                             }

  1662.                         }

  1663.                     } else {

  1664.                         if (s->above_intra_ctx[col]) {

  1665.                             c = 2;

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

  1667.                             c = 4 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1668.                         } else {

  1669.                             c = 3 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1670.                         }

  1671.                     }

  1672.                 } else if (have_l) {

  1673.                     if (s->left_intra_ctx[row7]) {

  1674.                         c = 2;

  1675.                     } else if (s->left_comp_ctx[row7]) {

  1676.                         c = 4 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1677.                     } else {

  1678.                         c = 3 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1679.                     }

  1680.                 } else {

  1681.                     c = 2;

  1682.                 }

  1683.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.comp_ref[c]);

  1684.                 b->ref[var_idx] = s->varcompref[bit];

  1685.                 s->counts.comp_ref[c][bit]++;

  1686.             } else /* single reference */ {

  1687.                 int bit, c;

  1688. 

  1689.                 if (have_a && !s->above_intra_ctx[col]) {

  1690.                     if (have_l && !s->left_intra_ctx[row7]) {

  1691.                         if (s->left_comp_ctx[row7]) {

  1692.                             if (s->above_comp_ctx[col]) {

  1693.                                 c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7] ||

  1694.                                          !s->above_ref_ctx[col]);

  1695.                             } else {

  1696.                                 c = (3 * !s->above_ref_ctx[col]) +

  1697.                                     (!s->fixcompref || !s->left_ref_ctx[row7]);

  1698.                             }

  1699.                         } else if (s->above_comp_ctx[col]) {

  1700.                             c = (3 * !s->left_ref_ctx[row7]) +

  1701.                                 (!s->fixcompref || !s->above_ref_ctx[col]);

  1702.                         } else {

  1703.                             c = 2 * !s->left_ref_ctx[row7] + 2 * !s->above_ref_ctx[col];

  1704.                         }

  1705.                     } else if (s->above_intra_ctx[col]) {

  1706.                         c = 2;

  1707.                     } else if (s->above_comp_ctx[col]) {

  1708.                         c = 1 + (!s->fixcompref || !s->above_ref_ctx[col]);

  1709.                     } else {

  1710.                         c = 4 * (!s->above_ref_ctx[col]);

  1711.                     }

  1712.                 } else if (have_l && !s->left_intra_ctx[row7]) {

  1713.                     if (s->left_intra_ctx[row7]) {

  1714.                         c = 2;

  1715.                     } else if (s->left_comp_ctx[row7]) {

  1716.                         c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7]);

  1717.                     } else {

  1718.                         c = 4 * (!s->left_ref_ctx[row7]);

  1719.                     }

  1720.                 } else {

  1721.                     c = 2;

  1722.                 }

  1723.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][0]);

  1724.                 s->counts.single_ref[c][0][bit]++;

  1725.                 if (!bit) {

  1726.                     b->ref[0] = 0;

  1727.                 } else {

  1728.                     // FIXME can this codeblob be replaced by some sort of LUT?

  1729.                     if (have_a) {

  1730.                         if (have_l) {

  1731.                             if (s->left_intra_ctx[row7]) {

  1732.                                 if (s->above_intra_ctx[col]) {

  1733.                                     c = 2;

  1734.                                 } else if (s->above_comp_ctx[col]) {

  1735.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1736.                                                  s->above_ref_ctx[col] == 1);

  1737.                                 } else if (!s->above_ref_ctx[col]) {

  1738.                                     c = 3;

  1739.                                 } else {

  1740.                                     c = 4 * (s->above_ref_ctx[col] == 1);

  1741.                                 }

  1742.                             } else if (s->above_intra_ctx[col]) {

  1743.                                 if (s->left_intra_ctx[row7]) {

  1744.                                     c = 2;

  1745.                                 } else if (s->left_comp_ctx[row7]) {

  1746.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1747.                                                  s->left_ref_ctx[row7] == 1);

  1748.                                 } else if (!s->left_ref_ctx[row7]) {

  1749.                                     c = 3;

  1750.                                 } else {

  1751.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1752.                                 }

  1753.                             } else if (s->above_comp_ctx[col]) {

  1754.                                 if (s->left_comp_ctx[row7]) {

  1755.                                     if (s->left_ref_ctx[row7] == s->above_ref_ctx[col]) {

  1756.                                         c = 3 * (s->fixcompref == 1 ||

  1757.                                                  s->left_ref_ctx[row7] == 1);

  1758.                                     } else {

  1759.                                         c = 2;

  1760.                                     }

  1761.                                 } else if (!s->left_ref_ctx[row7]) {

  1762.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1763.                                                  s->above_ref_ctx[col] == 1);

  1764.                                 } else {

  1765.                                     c = 3 * (s->left_ref_ctx[row7] == 1) +

  1766.                                     (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1767.                                 }

  1768.                             } else if (s->left_comp_ctx[row7]) {

  1769.                                 if (!s->above_ref_ctx[col]) {

  1770.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1771.                                                  s->left_ref_ctx[row7] == 1);

  1772.                                 } else {

  1773.                                     c = 3 * (s->above_ref_ctx[col] == 1) +

  1774.                                     (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1775.                                 }

  1776.                             } else if (!s->above_ref_ctx[col]) {

  1777.                                 if (!s->left_ref_ctx[row7]) {

  1778.                                     c = 3;

  1779.                                 } else {

  1780.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1781.                                 }

  1782.                             } else if (!s->left_ref_ctx[row7]) {

  1783.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1784.                             } else {

  1785.                                 c = 2 * (s->left_ref_ctx[row7] == 1) +

  1786.                                 2 * (s->above_ref_ctx[col] == 1);

  1787.                             }

  1788.                         } else {

  1789.                             if (s->above_intra_ctx[col] ||

  1790.                                 (!s->above_comp_ctx[col] && !s->above_ref_ctx[col])) {

  1791.                                 c = 2;

  1792.                             } else if (s->above_comp_ctx[col]) {

  1793.                                 c = 3 * (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1794.                             } else {

  1795.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1796.                             }

  1797.                         }

  1798.                     } else if (have_l) {

  1799.                         if (s->left_intra_ctx[row7] ||

  1800.                             (!s->left_comp_ctx[row7] && !s->left_ref_ctx[row7])) {

  1801.                             c = 2;

  1802.                         } else if (s->left_comp_ctx[row7]) {

  1803.                             c = 3 * (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1804.                         } else {

  1805.                             c = 4 * (s->left_ref_ctx[row7] == 1);

  1806.                         }

  1807.                     } else {

  1808.                         c = 2;

  1809.                     }

  1810.                     bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][1]);

  1811.                     s->counts.single_ref[c][1][bit]++;

  1812.                     b->ref[0] = 1 + bit;

  1813.                 }

  1814.             }

  1815.         }

  1816. 

  1817.         if (b->bs <= BS_8x8) {

  1818.             if (s->segmentation.feat[b->seg_id].skip_enabled) {

  1819.                 b->mode[0] = b->mode[1] = b->mode[2] = b->mode[3] = ZEROMV;

  1820.             } else {

  1821.                 static const uint8_t off[10] = {

  1822.                     3, 0, 0, 1, 0, 0, 0, 0, 0, 0

  1823.                 };

  1824. 

  1825.                 // FIXME this needs to use the LUT tables from find_ref_mvs

  1826.                 // because not all are -1,0/0,-1

  1827.                 int c = inter_mode_ctx_lut[s->above_mode_ctx[col + off[b->bs]]]

  1828.                                           [s->left_mode_ctx[row7 + off[b->bs]]];

  1829. 

  1830.                 b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1831.                                               s->prob.p.mv_mode[c]);

  1832.                 b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1833.                 s->counts.mv_mode[c][b->mode[0] - 10]++;

  1834.             }

  1835.         }

  1836. 

  1837.         if (s->filtermode == FILTER_SWITCHABLE) {

  1838.             int c;

  1839. 

  1840.             if (have_a && s->above_mode_ctx[col] >= NEARESTMV) {

  1841.                 if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1842.                     c = s->above_filter_ctx[col] == s->left_filter_ctx[row7] ?

  1843.                         s->left_filter_ctx[row7] : 3;

  1844.                 } else {

  1845.                     c = s->above_filter_ctx[col];

  1846.                 }

  1847.             } else if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1848.                 c = s->left_filter_ctx[row7];

  1849.             } else {

  1850.                 c = 3;

  1851.             }

  1852. 

  1853.             filter_id = vp8_rac_get_tree(&s->c, vp9_filter_tree,

  1854.                                          s->prob.p.filter[c]);

  1855.             s->counts.filter[c][filter_id]++;

  1856.             b->filter = vp9_filter_lut[filter_id];

  1857.         } else {

  1858.             b->filter = s->filtermode;

  1859.         }

  1860. 

  1861.         if (b->bs > BS_8x8) {

  1862.             int c = inter_mode_ctx_lut[s->above_mode_ctx[col]][s->left_mode_ctx[row7]];

  1863. 

  1864.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1865.                                           s->prob.p.mv_mode[c]);

  1866.             s->counts.mv_mode[c][b->mode[0] - 10]++;

  1867.             fill_mv(s, b->mv[0], b->mode[0], 0);

  1868. 

  1869.             if (b->bs != BS_8x4) {

  1870.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1871.                                               s->prob.p.mv_mode[c]);

  1872.                 s->counts.mv_mode[c][b->mode[1] - 10]++;

  1873.                 fill_mv(s, b->mv[1], b->mode[1], 1);

  1874.             } else {

  1875.                 b->mode[1] = b->mode[0];

  1876.                 AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1877.                 AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1878.             }

  1879. 

  1880.             if (b->bs != BS_4x8) {

  1881.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1882.                                               s->prob.p.mv_mode[c]);

  1883.                 s->counts.mv_mode[c][b->mode[2] - 10]++;

  1884.                 fill_mv(s, b->mv[2], b->mode[2], 2);

  1885. 

  1886.                 if (b->bs != BS_8x4) {

  1887.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1888.                                                   s->prob.p.mv_mode[c]);

  1889.                     s->counts.mv_mode[c][b->mode[3] - 10]++;

  1890.                     fill_mv(s, b->mv[3], b->mode[3], 3);

  1891.                 } else {

  1892.                     b->mode[3] = b->mode[2];

  1893.                     AV_COPY32(&b->mv[3][0], &b->mv[2][0]);

  1894.                     AV_COPY32(&b->mv[3][1], &b->mv[2][1]);

  1895.                 }

  1896.             } else {

  1897.                 b->mode[2] = b->mode[0];

  1898.                 AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1899.                 AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1900.                 b->mode[3] = b->mode[1];

  1901.                 AV_COPY32(&b->mv[3][0], &b->mv[1][0]);

  1902.                 AV_COPY32(&b->mv[3][1], &b->mv[1][1]);

  1903.             }

  1904.         } else {

  1905.             fill_mv(s, b->mv[0], b->mode[0], -1);

  1906.             AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1907.             AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1908.             AV_COPY32(&b->mv[3][0], &b->mv[0][0]);

  1909.             AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1910.             AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1911.             AV_COPY32(&b->mv[3][1], &b->mv[0][1]);

  1912.         }

  1913. 

  1914.         vref = b->ref[b->comp ? s->signbias[s->varcompref[0]] : 0];

  1915.     }

  1916. 

  1917. #if HAVE_FAST_64BIT

  1918. #define SPLAT_CTX(var, val, n) \

  1919.     switch (n) { \

  1920.     case 1:  var = val;                                    break; \

  1921.     case 2:  AV_WN16A(&var, val *             0x0101);     break; \

  1922.     case 4:  AV_WN32A(&var, val *         0x01010101);     break; \

  1923.     case 8:  AV_WN64A(&var, val * 0x0101010101010101ULL);  break; \

  1924.     case 16: { \

  1925.         uint64_t v64 = val * 0x0101010101010101ULL; \

  1926.         AV_WN64A(              &var,     v64); \

  1927.         AV_WN64A(&((uint8_t *) &var)[8], v64); \

  1928.         break; \

  1929.     } \

  1930.     }

  1931. #else

  1932. #define SPLAT_CTX(var, val, n) \

  1933.     switch (n) { \

  1934.     case 1:  var = val;                         break; \

  1935.     case 2:  AV_WN16A(&var, val *     0x0101);  break; \

  1936.     case 4:  AV_WN32A(&var, val * 0x01010101);  break; \

  1937.     case 8: { \

  1938.         uint32_t v32 = val * 0x01010101; \

  1939.         AV_WN32A(              &var,     v32); \

  1940.         AV_WN32A(&((uint8_t *) &var)[4], v32); \

  1941.         break; \

  1942.     } \

  1943.     case 16: { \

  1944.         uint32_t v32 = val * 0x01010101; \

  1945.         AV_WN32A(              &var,      v32); \

  1946.         AV_WN32A(&((uint8_t *) &var)[4],  v32); \

  1947.         AV_WN32A(&((uint8_t *) &var)[8],  v32); \

  1948.         AV_WN32A(&((uint8_t *) &var)[12], v32); \

  1949.         break; \

  1950.     } \

  1951.     }

  1952. #endif

  1953. 

  1954.     switch (bwh_tab[1][b->bs][0]) {

  1955. #define SET_CTXS(dir, off, n) \

  1956.     do { \

  1957.         SPLAT_CTX(s->dir##_skip_ctx[off],      b->skip,          n); \

  1958.         SPLAT_CTX(s->dir##_txfm_ctx[off],      b->tx,            n); \

  1959.         SPLAT_CTX(s->dir##_partition_ctx[off], dir##_ctx[b->bs], n); \

  1960.         if (!s->keyframe && !s->intraonly) { \

  1961.             SPLAT_CTX(s->dir##_intra_ctx[off], b->intra,   n); \

  1962.             SPLAT_CTX(s->dir##_comp_ctx[off],  b->comp,    n); \

  1963.             SPLAT_CTX(s->dir##_mode_ctx[off],  b->mode[3], n); \

  1964.             if (!b->intra) { \

  1965.                 SPLAT_CTX(s->dir##_ref_ctx[off], vref, n); \

  1966.                 if (s->filtermode == FILTER_SWITCHABLE) { \

  1967.                     SPLAT_CTX(s->dir##_filter_ctx[off], filter_id, n); \

  1968.                 } \

  1969.             } \

  1970.         } \

  1971.     } while (0)

  1972.     case 1: SET_CTXS(above, col, 1); break;

  1973.     case 2: SET_CTXS(above, col, 2); break;

  1974.     case 4: SET_CTXS(above, col, 4); break;

  1975.     case 8: SET_CTXS(above, col, 8); break;

  1976.     }

  1977.     switch (bwh_tab[1][b->bs][1]) {

  1978.     case 1: SET_CTXS(left, row7, 1); break;

  1979.     case 2: SET_CTXS(left, row7, 2); break;

  1980.     case 4: SET_CTXS(left, row7, 4); break;

  1981.     case 8: SET_CTXS(left, row7, 8); break;

  1982.     }

  1983. #undef SPLAT_CTX

  1984. #undef SET_CTXS

  1985. 

  1986.     if (!s->keyframe && !s->intraonly) {

  1987.         if (b->bs > BS_8x8) {

  1988.             int mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  1989. 

  1990.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][0], &b->mv[1][0]);

  1991.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][1], &b->mv[1][1]);

  1992.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][0], mv0);

  1993.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][1], mv1);

  1994.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][0], &b->mv[2][0]);

  1995.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][1], &b->mv[2][1]);

  1996.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][0], mv0);

  1997.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][1], mv1);

  1998.         } else {

  1999.             int n, mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  2000. 

  2001.             for (n = 0; n < w4 * 2; n++) {

  2002.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][0], mv0);

  2003.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][1], mv1);

  2004.             }

  2005.             for (n = 0; n < h4 * 2; n++) {

  2006.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][0], mv0);

  2007.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][1], mv1);

  2008.             }

  2009.         }

  2010.     }

  2011. 

  2012.     // FIXME kinda ugly

  2013.     for (y = 0; y < h4; y++) {

  2014.         int x, o = (row + y) * s->sb_cols * 8 + col;

  2015.         struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[o];

  2016. 

  2017.         if (b->intra) {

  2018.             for (x = 0; x < w4; x++) {

  2019.                 mv[x].ref[0] =

  2020.                 mv[x].ref[1] = -1;

  2021.             }

  2022.         } else if (b->comp) {

  2023.             for (x = 0; x < w4; x++) {

  2024.                 mv[x].ref[0] = b->ref[0];

  2025.                 mv[x].ref[1] = b->ref[1];

  2026.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  2027.                 AV_COPY32(&mv[x].mv[1], &b->mv[3][1]);

  2028.             }

  2029.         } else {

  2030.             for (x = 0; x < w4; x++) {

  2031.                 mv[x].ref[0] = b->ref[0];

  2032.                 mv[x].ref[1] = -1;

  2033.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  2034.             }

  2035.         }

  2036.     }

  2037. }

  2038. 

  2039. // FIXME merge cnt/eob arguments?

  2040. static av_always_inline int

  2041. decode_coeffs_b_generic(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  2042.                         int is_tx32x32, unsigned (*cnt)[6][3],

  2043.                         unsigned (*eob)[6][2], uint8_t (*p)[6][11],

  2044.                         int nnz, const int16_t *scan, const int16_t (*nb)[2],

  2045.                         const int16_t *band_counts, const int16_t *qmul)

  2046. {

  2047.     int i = 0, band = 0, band_left = band_counts[band];

  2048.     uint8_t *tp = p[0][nnz];

  2049.     uint8_t cache[1024];

  2050. 

  2051.     do {

  2052.         int val, rc;

  2053. 

  2054.         val = vp56_rac_get_prob_branchy(c, tp[0]); // eob

  2055.         eob[band][nnz][val]++;

  2056.         if (!val)

  2057.             break;

  2058. 

  2059.     skip_eob:

  2060.         if (!vp56_rac_get_prob_branchy(c, tp[1])) { // zero

  2061.             cnt[band][nnz][0]++;

  2062.             if (!--band_left)

  2063.                 band_left = band_counts[++band];

  2064.             cache[scan[i]] = 0;

  2065.             nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  2066.             tp = p[band][nnz];

  2067.             if (++i == n_coeffs)

  2068.                 break; //invalid input; blocks should end with EOB

  2069.             goto skip_eob;

  2070.         }

  2071. 

  2072.         rc = scan[i];

  2073.         if (!vp56_rac_get_prob_branchy(c, tp[2])) { // one

  2074.             cnt[band][nnz][1]++;

  2075.             val = 1;

  2076.             cache[rc] = 1;

  2077.         } else {

  2078.             // fill in p[3-10] (model fill) - only once per frame for each pos

  2079.             if (!tp[3])

  2080.                 memcpy(&tp[3], vp9_model_pareto8[tp[2]], 8);

  2081. 

  2082.             cnt[band][nnz][2]++;

  2083.             if (!vp56_rac_get_prob_branchy(c, tp[3])) { // 2, 3, 4

  2084.                 if (!vp56_rac_get_prob_branchy(c, tp[4])) {

  2085.                     cache[rc] = val = 2;

  2086.                 } else {

  2087.                     val = 3 + vp56_rac_get_prob(c, tp[5]);

  2088.                     cache[rc] = 3;

  2089.                 }

  2090.             } else if (!vp56_rac_get_prob_branchy(c, tp[6])) { // cat1/2

  2091.                 cache[rc] = 4;

  2092.                 if (!vp56_rac_get_prob_branchy(c, tp[7])) {

  2093.                     val = 5 + vp56_rac_get_prob(c, 159);

  2094.                 } else {

  2095.                     val  = 7 + (vp56_rac_get_prob(c, 165) << 1);

  2096.                     val +=      vp56_rac_get_prob(c, 145);

  2097.                 }

  2098.             } else { // cat 3-6

  2099.                 cache[rc] = 5;

  2100.                 if (!vp56_rac_get_prob_branchy(c, tp[8])) {

  2101.                     if (!vp56_rac_get_prob_branchy(c, tp[9])) {

  2102.                         val  = 11 + (vp56_rac_get_prob(c, 173) << 2);

  2103.                         val +=      (vp56_rac_get_prob(c, 148) << 1);

  2104.                         val +=       vp56_rac_get_prob(c, 140);

  2105.                     } else {

  2106.                         val  = 19 + (vp56_rac_get_prob(c, 176) << 3);

  2107.                         val +=      (vp56_rac_get_prob(c, 155) << 2);

  2108.                         val +=      (vp56_rac_get_prob(c, 140) << 1);

  2109.                         val +=       vp56_rac_get_prob(c, 135);

  2110.                     }

  2111.                 } else if (!vp56_rac_get_prob_branchy(c, tp[10])) {

  2112.                     val  = 35 + (vp56_rac_get_prob(c, 180) << 4);

  2113.                     val +=      (vp56_rac_get_prob(c, 157) << 3);

  2114.                     val +=      (vp56_rac_get_prob(c, 141) << 2);

  2115.                     val +=      (vp56_rac_get_prob(c, 134) << 1);

  2116.                     val +=       vp56_rac_get_prob(c, 130);

  2117.                 } else {

  2118.                     val  = 67 + (vp56_rac_get_prob(c, 254) << 13);

  2119.                     val +=      (vp56_rac_get_prob(c, 254) << 12);

  2120.                     val +=      (vp56_rac_get_prob(c, 254) << 11);

  2121.                     val +=      (vp56_rac_get_prob(c, 252) << 10);

  2122.                     val +=      (vp56_rac_get_prob(c, 249) << 9);

  2123.                     val +=      (vp56_rac_get_prob(c, 243) << 8);

  2124.                     val +=      (vp56_rac_get_prob(c, 230) << 7);

  2125.                     val +=      (vp56_rac_get_prob(c, 196) << 6);

  2126.                     val +=      (vp56_rac_get_prob(c, 177) << 5);

  2127.                     val +=      (vp56_rac_get_prob(c, 153) << 4);

  2128.                     val +=      (vp56_rac_get_prob(c, 140) << 3);

  2129.                     val +=      (vp56_rac_get_prob(c, 133) << 2);

  2130.                     val +=      (vp56_rac_get_prob(c, 130) << 1);

  2131.                     val +=       vp56_rac_get_prob(c, 129);

  2132.                 }

  2133.             }

  2134.         }

  2135.         if (!--band_left)

  2136.             band_left = band_counts[++band];

  2137.         if (is_tx32x32)

  2138.             coef[rc] = ((vp8_rac_get(c) ? -val : val) * qmul[!!i]) / 2;

  2139.         else

  2140.             coef[rc] = (vp8_rac_get(c) ? -val : val) * qmul[!!i];

  2141.         nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  2142.         tp = p[band][nnz];

  2143.     } while (++i < n_coeffs);

  2144. 

  2145.     return i;

  2146. }

  2147. 

  2148. static int decode_coeffs_b(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  2149.                            unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  2150.                            uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  2151.                            const int16_t (*nb)[2], const int16_t *band_counts,

  2152.                            const int16_t *qmul)

  2153. {

  2154.     return decode_coeffs_b_generic(c, coef, n_coeffs, 0, cnt, eob, p,

  2155.                                    nnz, scan, nb, band_counts, qmul);

  2156. }

  2157. 

  2158. static int decode_coeffs_b32(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  2159.                              unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  2160.                              uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  2161.                              const int16_t (*nb)[2], const int16_t *band_counts,

  2162.                              const int16_t *qmul)

  2163. {

  2164.     return decode_coeffs_b_generic(c, coef, n_coeffs, 1, cnt, eob, p,

  2165.                                    nnz, scan, nb, band_counts, qmul);

  2166. }

  2167. 

  2168. static void decode_coeffs(AVCodecContext *ctx)

  2169. {

  2170.     VP9Context *s = ctx->priv_data;

  2171.     VP9Block *b = s->b;

  2172.     int row = s->row, col = s->col;

  2173.     uint8_t (*p)[6][11] = s->prob.coef[b->tx][0 /* y */][!b->intra];

  2174.     unsigned (*c)[6][3] = s->counts.coef[b->tx][0 /* y */][!b->intra];

  2175.     unsigned (*e)[6][2] = s->counts.eob[b->tx][0 /* y */][!b->intra];

  2176.     int w4 = bwh_tab[1][b->bs][0] << 1, h4 = bwh_tab[1][b->bs][1] << 1;

  2177.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2178.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2179.     int n, pl, x, y, res;

  2180.     int16_t (*qmul)[2] = s->segmentation.feat[b->seg_id].qmul;

  2181.     int tx = 4 * s->lossless + b->tx;

  2182.     const int16_t * const *yscans = vp9_scans[tx];

  2183.     const int16_t (* const *ynbs)[2] = vp9_scans_nb[tx];

  2184.     const int16_t *uvscan = vp9_scans[b->uvtx][DCT_DCT];

  2185.     const int16_t (*uvnb)[2] = vp9_scans_nb[b->uvtx][DCT_DCT];

  2186.     uint8_t *a = &s->above_y_nnz_ctx[col * 2];

  2187.     uint8_t *l = &s->left_y_nnz_ctx[(row & 7) << 1];

  2188.     static const int16_t band_counts[4][8] = {

  2189.         { 1, 2, 3, 4,  3,   16 - 13 },

  2190.         { 1, 2, 3, 4, 11,   64 - 21 },

  2191.         { 1, 2, 3, 4, 11,  256 - 21 },

  2192.         { 1, 2, 3, 4, 11, 1024 - 21 },

  2193.     };

  2194.     const int16_t *y_band_counts = band_counts[b->tx];

  2195.     const int16_t *uv_band_counts = band_counts[b->uvtx];

  2196. 

  2197. #define MERGE(la, end, step, rd) \

  2198.     for (n = 0; n < end; n += step) \

  2199.         la[n] = !!rd(&la[n])

  2200. #define MERGE_CTX(step, rd) \

  2201.     do { \

  2202.         MERGE(l, end_y, step, rd); \

  2203.         MERGE(a, end_x, step, rd); \

  2204.     } while (0)

  2205. 

  2206. #define DECODE_Y_COEF_LOOP(step, mode_index, v) \

  2207.     for (n = 0, y = 0; y < end_y; y += step) { \

  2208.         for (x = 0; x < end_x; x += step, n += step * step) { \

  2209.             enum TxfmType txtp = vp9_intra_txfm_type[b->mode[mode_index]]; \

  2210.             res = decode_coeffs_b##v(&s->c, s->block + 16 * n, 16 * step * step, \

  2211.                                      c, e, p, a[x] + l[y], yscans[txtp], \

  2212.                                      ynbs[txtp], y_band_counts, qmul[0]); \

  2213.             a[x] = l[y] = !!res; \

  2214.             if (step >= 4) { \

  2215.                 AV_WN16A(&s->eob[n], res); \

  2216.             } else { \

  2217.                 s->eob[n] = res; \

  2218.             } \

  2219.         } \

  2220.     }

  2221. 

  2222. #define SPLAT(la, end, step, cond) \

  2223.     if (step == 2) { \

  2224.         for (n = 1; n < end; n += step) \

  2225.             la[n] = la[n - 1]; \

  2226.     } else if (step == 4) { \

  2227.         if (cond) { \

  2228.             for (n = 0; n < end; n += step) \

  2229.                 AV_WN32A(&la[n], la[n] * 0x01010101); \

  2230.         } else { \

  2231.             for (n = 0; n < end; n += step) \

  2232.                 memset(&la[n + 1], la[n], FFMIN(end - n - 1, 3)); \

  2233.         } \

  2234.     } else /* step == 8 */ { \

  2235.         if (cond) { \

  2236.             if (HAVE_FAST_64BIT) { \

  2237.                 for (n = 0; n < end; n += step) \

  2238.                     AV_WN64A(&la[n], la[n] * 0x0101010101010101ULL); \

  2239.             } else { \

  2240.                 for (n = 0; n < end; n += step) { \

  2241.                     uint32_t v32 = la[n] * 0x01010101; \

  2242.                     AV_WN32A(&la[n],     v32); \

  2243.                     AV_WN32A(&la[n + 4], v32); \

  2244.                 } \

  2245.             } \

  2246.         } else { \

  2247.             for (n = 0; n < end; n += step) \

  2248.                 memset(&la[n + 1], la[n], FFMIN(end - n - 1, 7)); \

  2249.         } \

  2250.     }

  2251. #define SPLAT_CTX(step) \

  2252.     do { \

  2253.         SPLAT(a, end_x, step, end_x == w4); \

  2254.         SPLAT(l, end_y, step, end_y == h4); \

  2255.     } while (0)

  2256. 

  2257.     /* y tokens */

  2258.     switch (b->tx) {

  2259.     case TX_4X4:

  2260.         DECODE_Y_COEF_LOOP(1, b->bs > BS_8x8 ? n : 0,);

  2261.         break;

  2262.     case TX_8X8:

  2263.         MERGE_CTX(2, AV_RN16A);

  2264.         DECODE_Y_COEF_LOOP(2, 0,);

  2265.         SPLAT_CTX(2);

  2266.         break;

  2267.     case TX_16X16:

  2268.         MERGE_CTX(4, AV_RN32A);

  2269.         DECODE_Y_COEF_LOOP(4, 0,);

  2270.         SPLAT_CTX(4);

  2271.         break;

  2272.     case TX_32X32:

  2273.         MERGE_CTX(8, AV_RN64A);

  2274.         DECODE_Y_COEF_LOOP(8, 0, 32);

  2275.         SPLAT_CTX(8);

  2276.         break;

  2277.     }

  2278. 

  2279. #define DECODE_UV_COEF_LOOP(step) \

  2280.     for (n = 0, y = 0; y < end_y; y += step) { \

  2281.         for (x = 0; x < end_x; x += step, n += step * step) { \

  2282.             res = decode_coeffs_b(&s->c, s->uvblock[pl] + 16 * n, \

  2283.                                   16 * step * step, c, e, p, a[x] + l[y], \

  2284.                                   uvscan, uvnb, uv_band_counts, qmul[1]); \

  2285.             a[x] = l[y] = !!res; \

  2286.             if (step >= 4) { \

  2287.                 AV_WN16A(&s->uveob[pl][n], res); \

  2288.             } else { \

  2289.                 s->uveob[pl][n] = res; \

  2290.             } \

  2291.         } \

  2292.     }

  2293. 

  2294.     p = s->prob.coef[b->uvtx][1 /* uv */][!b->intra];

  2295.     c = s->counts.coef[b->uvtx][1 /* uv */][!b->intra];

  2296.     e = s->counts.eob[b->uvtx][1 /* uv */][!b->intra];

  2297.     w4 >>= 1;

  2298.     h4 >>= 1;

  2299.     end_x >>= 1;

  2300.     end_y >>= 1;

  2301.     for (pl = 0; pl < 2; pl++) {

  2302.         a = &s->above_uv_nnz_ctx[pl][col];

  2303.         l = &s->left_uv_nnz_ctx[pl][row & 7];

  2304.         switch (b->uvtx) {

  2305.         case TX_4X4:

  2306.             DECODE_UV_COEF_LOOP(1);

  2307.             break;

  2308.         case TX_8X8:

  2309.             MERGE_CTX(2, AV_RN16A);

  2310.             DECODE_UV_COEF_LOOP(2);

  2311.             SPLAT_CTX(2);

  2312.             break;

  2313.         case TX_16X16:

  2314.             MERGE_CTX(4, AV_RN32A);

  2315.             DECODE_UV_COEF_LOOP(4);

  2316.             SPLAT_CTX(4);

  2317.             break;

  2318.         case TX_32X32:

  2319.             MERGE_CTX(8, AV_RN64A);

  2320.             // a 64x64 (max) uv block can ever only contain 1 tx32x32 block

  2321.             // so there is no need to loop

  2322.             res = decode_coeffs_b32(&s->c, s->uvblock[pl],

  2323.                                     1024, c, e, p, a[0] + l[0],

  2324.                                     uvscan, uvnb, uv_band_counts, qmul[1]);

  2325.             a[0] = l[0] = !!res;

  2326.             AV_WN16A(&s->uveob[pl][0], res);

  2327.             SPLAT_CTX(8);

  2328.             break;

  2329.         }

  2330.     }

  2331. }

  2332. 

  2333. static av_always_inline int check_intra_mode(VP9Context *s, int mode, uint8_t **a,

  2334.                                              uint8_t *dst_edge, ptrdiff_t stride_edge,

  2335.                                              uint8_t *dst_inner, ptrdiff_t stride_inner,

  2336.                                              uint8_t *l, int col, int x, int w,

  2337.                                              int row, int y, enum TxfmMode tx,

  2338.                                              int p)

  2339. {

  2340.     int have_top = row > 0 || y > 0;

  2341.     int have_left = col > s->tiling.tile_col_start || x > 0;

  2342.     int have_right = x < w - 1;

  2343.     static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {

  2344.         [VERT_PRED]            = { { DC_127_PRED,          VERT_PRED },

  2345.                                    { DC_127_PRED,          VERT_PRED } },

  2346.         [HOR_PRED]             = { { DC_129_PRED,          DC_129_PRED },

  2347.                                    { HOR_PRED,             HOR_PRED } },

  2348.         [DC_PRED]              = { { DC_128_PRED,          TOP_DC_PRED },

  2349.                                    { LEFT_DC_PRED,         DC_PRED } },

  2350.         [DIAG_DOWN_LEFT_PRED]  = { { DC_127_PRED,          DIAG_DOWN_LEFT_PRED },

  2351.                                    { DC_127_PRED,          DIAG_DOWN_LEFT_PRED } },

  2352.         [DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },

  2353.                                    { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },

  2354.         [VERT_RIGHT_PRED]      = { { VERT_RIGHT_PRED,      VERT_RIGHT_PRED },

  2355.                                    { VERT_RIGHT_PRED,      VERT_RIGHT_PRED } },

  2356.         [HOR_DOWN_PRED]        = { { HOR_DOWN_PRED,        HOR_DOWN_PRED },

  2357.                                    { HOR_DOWN_PRED,        HOR_DOWN_PRED } },

  2358.         [VERT_LEFT_PRED]       = { { DC_127_PRED,          VERT_LEFT_PRED },

  2359.                                    { DC_127_PRED,          VERT_LEFT_PRED } },

  2360.         [HOR_UP_PRED]          = { { DC_129_PRED,          DC_129_PRED },

  2361.                                    { HOR_UP_PRED,          HOR_UP_PRED } },

  2362.         [TM_VP8_PRED]          = { { DC_129_PRED,          VERT_PRED },

  2363.                                    { HOR_PRED,             TM_VP8_PRED } },

  2364.     };

  2365.     static const struct {

  2366.         uint8_t needs_left:1;

  2367.         uint8_t needs_top:1;

  2368.         uint8_t needs_topleft:1;

  2369.         uint8_t needs_topright:1;

  2370.         uint8_t invert_left:1;

  2371.     } edges[N_INTRA_PRED_MODES] = {

  2372.         [VERT_PRED]            = { .needs_top  = 1 },

  2373.         [HOR_PRED]             = { .needs_left = 1 },

  2374.         [DC_PRED]              = { .needs_top  = 1, .needs_left = 1 },

  2375.         [DIAG_DOWN_LEFT_PRED]  = { .needs_top  = 1, .needs_topright = 1 },

  2376.         [DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2377.         [VERT_RIGHT_PRED]      = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2378.         [HOR_DOWN_PRED]        = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2379.         [VERT_LEFT_PRED]       = { .needs_top  = 1, .needs_topright = 1 },

  2380.         [HOR_UP_PRED]          = { .needs_left = 1, .invert_left = 1 },

  2381.         [TM_VP8_PRED]          = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2382.         [LEFT_DC_PRED]         = { .needs_left = 1 },

  2383.         [TOP_DC_PRED]          = { .needs_top  = 1 },

  2384.         [DC_128_PRED]          = { 0 },

  2385.         [DC_127_PRED]          = { 0 },

  2386.         [DC_129_PRED]          = { 0 }

  2387.     };

  2388. 

  2389.     av_assert2(mode >= 0 && mode < 10);

  2390.     mode = mode_conv[mode][have_left][have_top];

  2391.     if (edges[mode].needs_top) {

  2392.         uint8_t *top, *topleft;

  2393.         int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !p) - x) * 4;

  2394.         int n_px_need_tr = 0;

  2395. 

  2396.         if (tx == TX_4X4 && edges[mode].needs_topright && have_right)

  2397.             n_px_need_tr = 4;

  2398. 

  2399.         // if top of sb64-row, use s->intra_pred_data[] instead of

  2400.         // dst[-stride] for intra prediction (it contains pre- instead of

  2401.         // post-loopfilter data)

  2402.         if (have_top) {

  2403.             top = !(row & 7) && !y ?

  2404.                 s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  2405.                 y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];

  2406.             if (have_left)

  2407.                 topleft = !(row & 7) && !y ?

  2408.                     s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  2409.                     y == 0 || x == 0 ? &dst_edge[-stride_edge] :

  2410.                     &dst_inner[-stride_inner];

  2411.         }

  2412. 

  2413.         if (have_top &&

  2414.             (!edges[mode].needs_topleft || (have_left && top == topleft)) &&

  2415.             (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&

  2416.             n_px_need + n_px_need_tr <= n_px_have) {

  2417.             *a = top;

  2418.         } else {

  2419.             if (have_top) {

  2420.                 if (n_px_need <= n_px_have) {

  2421.                     memcpy(*a, top, n_px_need);

  2422.                 } else {

  2423.                     memcpy(*a, top, n_px_have);

  2424.                     memset(&(*a)[n_px_have], (*a)[n_px_have - 1],

  2425.                            n_px_need - n_px_have);

  2426.                 }

  2427.             } else {

  2428.                 memset(*a, 127, n_px_need);

  2429.             }

  2430.             if (edges[mode].needs_topleft) {

  2431.                 if (have_left && have_top) {

  2432.                     (*a)[-1] = topleft[-1];

  2433.                 } else {

  2434.                     (*a)[-1] = have_top ? 129 : 127;

  2435.                 }

  2436.             }

  2437.             if (tx == TX_4X4 && edges[mode].needs_topright) {

  2438.                 if (have_top && have_right &&

  2439.                     n_px_need + n_px_need_tr <= n_px_have) {

  2440.                     memcpy(&(*a)[4], &top[4], 4);

  2441.                 } else {

  2442.                     memset(&(*a)[4], (*a)[3], 4);

  2443.                 }

  2444.             }

  2445.         }

  2446.     }

  2447.     if (edges[mode].needs_left) {

  2448.         if (have_left) {

  2449.             int n_px_need = 4 << tx, i, n_px_have = (((s->rows - row) << !p) - y) * 4;

  2450.             uint8_t *dst = x == 0 ? dst_edge : dst_inner;

  2451.             ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;

  2452. 

  2453.             if (edges[mode].invert_left) {

  2454.                 if (n_px_need <= n_px_have) {

  2455.                     for (i = 0; i < n_px_need; i++)

  2456.                         l[i] = dst[i * stride - 1];

  2457.                 } else {

  2458.                     for (i = 0; i < n_px_have; i++)

  2459.                         l[i] = dst[i * stride - 1];

  2460.                     memset(&l[n_px_have], l[n_px_have - 1], n_px_need - n_px_have);

  2461.                 }

  2462.             } else {

  2463.                 if (n_px_need <= n_px_have) {

  2464.                     for (i = 0; i < n_px_need; i++)

  2465.                         l[n_px_need - 1 - i] = dst[i * stride - 1];

  2466.                 } else {

  2467.                     for (i = 0; i < n_px_have; i++)

  2468.                         l[n_px_need - 1 - i] = dst[i * stride - 1];

  2469.                     memset(l, l[n_px_need - n_px_have], n_px_need - n_px_have);

  2470.                 }

  2471.             }

  2472.         } else {

  2473.             memset(l, 129, 4 << tx);

  2474.         }

  2475.     }

  2476. 

  2477.     return mode;

  2478. }

  2479. 

  2480. static void intra_recon(AVCodecContext *ctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  2481. {

  2482.     VP9Context *s = ctx->priv_data;

  2483.     VP9Block *b = s->b;

  2484.     int row = s->row, col = s->col;

  2485.     int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2486.     int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2487.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2488.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2489.     int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2490.     int uvstep1d = 1 << b->uvtx, p;

  2491.     uint8_t *dst = s->dst[0], *dst_r = s->frames[CUR_FRAME].tf.f->data[0] + y_off;

  2492.     LOCAL_ALIGNED_32(uint8_t, a_buf, [64]);

  2493.     LOCAL_ALIGNED_32(uint8_t, l, [32]);

  2494. 

  2495.     for (n = 0, y = 0; y < end_y; y += step1d) {

  2496.         uint8_t *ptr = dst, *ptr_r = dst_r;

  2497.         for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d,

  2498.                                ptr_r += 4 * step1d, n += step) {

  2499.             int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?

  2500.                                y * 2 + x : 0];

  2501.             uint8_t *a = &a_buf[32];

  2502.             enum TxfmType txtp = vp9_intra_txfm_type[mode];

  2503.             int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2504. 

  2505.             mode = check_intra_mode(s, mode, &a, ptr_r,

  2506.                                     s->frames[CUR_FRAME].tf.f->linesize[0],

  2507.                                     ptr, s->y_stride, l,

  2508.                                     col, x, w4, row, y, b->tx, 0);

  2509.             s->dsp.intra_pred[b->tx][mode](ptr, s->y_stride, l, a);

  2510.             if (eob)

  2511.                 s->dsp.itxfm_add[tx][txtp](ptr, s->y_stride,

  2512.                                            s->block + 16 * n, eob);

  2513.         }

  2514.         dst_r += 4 * step1d * s->frames[CUR_FRAME].tf.f->linesize[0];

  2515.         dst   += 4 * step1d * s->y_stride;

  2516.     }

  2517. 

  2518.     // U/V

  2519.     w4 >>= 1;

  2520.     end_x >>= 1;

  2521.     end_y >>= 1;

  2522.     step = 1 << (b->uvtx * 2);

  2523.     for (p = 0; p < 2; p++) {

  2524.         dst   = s->dst[1 + p];

  2525.         dst_r = s->frames[CUR_FRAME].tf.f->data[1 + p] + uv_off;

  2526.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2527.             uint8_t *ptr = dst, *ptr_r = dst_r;

  2528.             for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d,

  2529.                                    ptr_r += 4 * uvstep1d, n += step) {

  2530.                 int mode = b->uvmode;

  2531.                 uint8_t *a = &a_buf[32];

  2532.                 int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2533. 

  2534.                 mode = check_intra_mode(s, mode, &a, ptr_r,

  2535.                                         s->frames[CUR_FRAME].tf.f->linesize[1],

  2536.                                         ptr, s->uv_stride, l,

  2537.                                         col, x, w4, row, y, b->uvtx, p + 1);

  2538.                 s->dsp.intra_pred[b->uvtx][mode](ptr, s->uv_stride, l, a);

  2539.                 if (eob)

  2540.                     s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2541.                                                     s->uvblock[p] + 16 * n, eob);

  2542.             }

  2543.             dst_r += 4 * uvstep1d * s->frames[CUR_FRAME].tf.f->linesize[1];

  2544.             dst   += 4 * uvstep1d * s->uv_stride;

  2545.         }

  2546.     }

  2547. }

  2548. 

  2549. static av_always_inline void mc_luma_scaled(VP9Context *s, vp9_scaled_mc_func smc,

  2550.                                             uint8_t *dst, ptrdiff_t dst_stride,

  2551.                                             const uint8_t *ref, ptrdiff_t ref_stride,

  2552.                                             ThreadFrame *ref_frame,

  2553.                                             ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2554.                                             int bw, int bh, int w, int h,

  2555.                                             const uint16_t *scale, const uint8_t *step)

  2556. {

  2557. #define scale_mv(n, dim) (((int64_t)n * scale[dim]) >> 14)

  2558.     // BUG libvpx seems to scale the two components separately. This introduces

  2559.     // rounding errors but we have to reproduce them to be exactly compatible

  2560.     // with the output from libvpx...

  2561.     int mx = scale_mv(mv->x * 2, 0) + scale_mv(x * 16, 0);

  2562.     int my = scale_mv(mv->y * 2, 1) + scale_mv(y * 16, 1);

  2563.     int refbw_m1, refbh_m1;

  2564.     int th;

  2565. 

  2566.     y = my >> 4;

  2567.     x = mx >> 4;

  2568.     ref += y * ref_stride + x;

  2569.     mx &= 15;

  2570.     my &= 15;

  2571.     refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;

  2572.     refbh_m1 = ((bh - 1) * step[1] + my) >> 4;

  2573.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2574.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2575.     // the longest loopfilter of the next sbrow

  2576.     th = (y + refbh_m1 + 4 + 7) >> 6;

  2577.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2578.     if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 4 >= h - refbh_m1) {

  2579.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2580.                                  ref - 3 * ref_stride - 3,

  2581.                                  144, ref_stride,

  2582.                                  refbw_m1 + 8, refbh_m1 + 8,

  2583.                                  x - 3, y - 3, w, h);

  2584.         ref = s->edge_emu_buffer + 3 * 144 + 3;

  2585.         ref_stride = 144;

  2586.     }

  2587.     smc(dst, dst_stride, ref, ref_stride, bh, mx, my, step[0], step[1]);

  2588. }

  2589. 

  2590. static av_always_inline void mc_chroma_scaled(VP9Context *s, vp9_scaled_mc_func smc,

  2591.                                               uint8_t *dst_u, uint8_t *dst_v,

  2592.                                               ptrdiff_t dst_stride,

  2593.                                               const uint8_t *ref_u, ptrdiff_t src_stride_u,

  2594.                                               const uint8_t *ref_v, ptrdiff_t src_stride_v,

  2595.                                               ThreadFrame *ref_frame,

  2596.                                               ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2597.                                               int bw, int bh, int w, int h,

  2598.                                               const uint16_t *scale, const uint8_t *step)

  2599. {

  2600.     // BUG https://code.google.com/p/webm/issues/detail?id=820

  2601.     int mx = scale_mv(mv->x, 0) + (scale_mv(x * 16, 0) & ~15) + (scale_mv(x * 32, 0) & 15);

  2602.     int my = scale_mv(mv->y, 1) + (scale_mv(y * 16, 1) & ~15) + (scale_mv(y * 32, 1) & 15);

  2603. #undef scale_mv

  2604.     int refbw_m1, refbh_m1;

  2605.     int th;

  2606. 

  2607.     y = my >> 4;

  2608.     x = mx >> 4;

  2609.     ref_u += y * src_stride_u + x;

  2610.     ref_v += y * src_stride_v + x;

  2611.     mx &= 15;

  2612.     my &= 15;

  2613.     refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;

  2614.     refbh_m1 = ((bh - 1) * step[1] + my) >> 4;

  2615.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2616.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2617.     // the longest loopfilter of the next sbrow

  2618.     th = (y + refbh_m1 + 4 + 7) >> 5;

  2619.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2620.     if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 4 >= h - refbh_m1) {

  2621.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2622.                                  ref_u - 3 * src_stride_u - 3,

  2623.                                  144, src_stride_u,

  2624.                                  refbw_m1 + 8, refbh_m1 + 8,

  2625.                                  x - 3, y - 3, w, h);

  2626.         ref_u = s->edge_emu_buffer + 3 * 144 + 3;

  2627.         smc(dst_u, dst_stride, ref_u, 144, bh, mx, my, step[0], step[1]);

  2628. 

  2629.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2630.                                  ref_v - 3 * src_stride_v - 3,

  2631.                                  144, src_stride_v,

  2632.                                  refbw_m1 + 8, refbh_m1 + 8,

  2633.                                  x - 3, y - 3, w, h);

  2634.         ref_v = s->edge_emu_buffer + 3 * 144 + 3;

  2635.         smc(dst_v, dst_stride, ref_v, 144, bh, mx, my, step[0], step[1]);

  2636.     } else {

  2637.         smc(dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my, step[0], step[1]);

  2638.         smc(dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my, step[0], step[1]);

  2639.     }

  2640. }

  2641. 

  2642. #define FN(x) x##_scaled

  2643. #define mc_luma_dir(s, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, bw, bh, w, h, i) \

  2644.     mc_luma_scaled(s, s->dsp.s##mc, dst, dst_ls, src, src_ls, tref, row, col, \

  2645.                    mv, bw, bh, w, h, s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])

  2646. #define mc_chroma_dir(s, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2647.                       row, col, mv, bw, bh, w, h, i) \

  2648.     mc_chroma_scaled(s, s->dsp.s##mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2649.                      row, col, mv, bw, bh, w, h, s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])

  2650. #include "vp9_mc_template.c"

  2651. #undef mc_luma_dir

  2652. #undef mc_chroma_dir

  2653. #undef FN

  2654. 

  2655. static av_always_inline void mc_luma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2],

  2656.                                               uint8_t *dst, ptrdiff_t dst_stride,

  2657.                                               const uint8_t *ref, ptrdiff_t ref_stride,

  2658.                                               ThreadFrame *ref_frame,

  2659.                                               ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2660.                                               int bw, int bh, int w, int h)

  2661. {

  2662.     int mx = mv->x, my = mv->y, th;

  2663. 

  2664.     y += my >> 3;

  2665.     x += mx >> 3;

  2666.     ref += y * ref_stride + x;

  2667.     mx &= 7;

  2668.     my &= 7;

  2669.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2670.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2671.     // the longest loopfilter of the next sbrow

  2672.     th = (y + bh + 4 * !!my + 7) >> 6;

  2673.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2674.     if (x < !!mx * 3 || y < !!my * 3 ||

  2675.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2676.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2677.                                  ref - !!my * 3 * ref_stride - !!mx * 3,

  2678.                                  80, ref_stride,

  2679.                                  bw + !!mx * 7, bh + !!my * 7,

  2680.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2681.         ref = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2682.         ref_stride = 80;

  2683.     }

  2684.     mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);

  2685. }

  2686. 

  2687. static av_always_inline void mc_chroma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2],

  2688.                                                 uint8_t *dst_u, uint8_t *dst_v,

  2689.                                                 ptrdiff_t dst_stride,

  2690.                                                 const uint8_t *ref_u, ptrdiff_t src_stride_u,

  2691.                                                 const uint8_t *ref_v, ptrdiff_t src_stride_v,

  2692.                                                 ThreadFrame *ref_frame,

  2693.                                                 ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2694.                                                 int bw, int bh, int w, int h)

  2695. {

  2696.     int mx = mv->x, my = mv->y, th;

  2697. 

  2698.     y += my >> 4;

  2699.     x += mx >> 4;

  2700.     ref_u += y * src_stride_u + x;

  2701.     ref_v += y * src_stride_v + x;

  2702.     mx &= 15;

  2703.     my &= 15;

  2704.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2705.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2706.     // the longest loopfilter of the next sbrow

  2707.     th = (y + bh + 4 * !!my + 7) >> 5;

  2708.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2709.     if (x < !!mx * 3 || y < !!my * 3 ||

  2710.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2711.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2712.                                  ref_u - !!my * 3 * src_stride_u - !!mx * 3,

  2713.                                  80, src_stride_u,

  2714.                                  bw + !!mx * 7, bh + !!my * 7,

  2715.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2716.         ref_u = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2717.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, 80, bh, mx, my);

  2718. 

  2719.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2720.                                  ref_v - !!my * 3 * src_stride_v - !!mx * 3,

  2721.                                  80, src_stride_v,

  2722.                                  bw + !!mx * 7, bh + !!my * 7,

  2723.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2724.         ref_v = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2725.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, 80, bh, mx, my);

  2726.     } else {

  2727.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);

  2728.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);

  2729.     }

  2730. }

  2731. 

  2732. #define FN(x) x

  2733. #define mc_luma_dir(s, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, bw, bh, w, h, i) \

  2734.     mc_luma_unscaled(s, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \

  2735.                      mv, bw, bh, w, h)

  2736. #define mc_chroma_dir(s, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2737.                       row, col, mv, bw, bh, w, h, i) \

  2738.     mc_chroma_unscaled(s, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2739.                        row, col, mv, bw, bh, w, h)

  2740. #include "vp9_mc_template.c"

  2741. #undef mc_luma_dir_dir

  2742. #undef mc_chroma_dir_dir

  2743. #undef FN

  2744. 

  2745. static void inter_recon(AVCodecContext *ctx)

  2746. {

  2747.     VP9Context *s = ctx->priv_data;

  2748.     VP9Block *b = s->b;

  2749.     int row = s->row, col = s->col;

  2750. 

  2751.     if (s->mvscale[b->ref[0]][0] || (b->comp && s->mvscale[b->ref[1]][0])) {

  2752.         inter_pred_scaled(ctx);

  2753.     } else {

  2754.         inter_pred(ctx);

  2755.     }

  2756.     if (!b->skip) {

  2757.         /* mostly copied intra_recon() */

  2758. 

  2759.         int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2760.         int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2761.         int end_x = FFMIN(2 * (s->cols - col), w4);

  2762.         int end_y = FFMIN(2 * (s->rows - row), h4);

  2763.         int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2764.         int uvstep1d = 1 << b->uvtx, p;

  2765.         uint8_t *dst = s->dst[0];

  2766. 

  2767.         // y itxfm add

  2768.         for (n = 0, y = 0; y < end_y; y += step1d) {

  2769.             uint8_t *ptr = dst;

  2770.             for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d, n += step) {

  2771.                 int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2772. 

  2773.                 if (eob)

  2774.                     s->dsp.itxfm_add[tx][DCT_DCT](ptr, s->y_stride,

  2775.                                                   s->block + 16 * n, eob);

  2776.             }

  2777.             dst += 4 * s->y_stride * step1d;

  2778.         }

  2779. 

  2780.         // uv itxfm add

  2781.         end_x >>= 1;

  2782.         end_y >>= 1;

  2783.         step = 1 << (b->uvtx * 2);

  2784.         for (p = 0; p < 2; p++) {

  2785.             dst = s->dst[p + 1];

  2786.             for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2787.                 uint8_t *ptr = dst;

  2788.                 for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d, n += step) {

  2789.                     int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2790. 

  2791.                     if (eob)

  2792.                         s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2793.                                                         s->uvblock[p] + 16 * n, eob);

  2794.                 }

  2795.                 dst += 4 * uvstep1d * s->uv_stride;

  2796.             }

  2797.         }

  2798.     }

  2799. }

  2800. 

  2801. static av_always_inline void mask_edges(struct VP9Filter *lflvl, int is_uv,

  2802.                                         int row_and_7, int col_and_7,

  2803.                                         int w, int h, int col_end, int row_end,

  2804.                                         enum TxfmMode tx, int skip_inter)

  2805. {

  2806.     // FIXME I'm pretty sure all loops can be replaced by a single LUT if

  2807.     // we make VP9Filter.mask uint64_t (i.e. row/col all single variable)

  2808.     // and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then

  2809.     // use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)

  2810. 

  2811.     // the intended behaviour of the vp9 loopfilter is to work on 8-pixel

  2812.     // edges. This means that for UV, we work on two subsampled blocks at

  2813.     // a time, and we only use the topleft block's mode information to set

  2814.     // things like block strength. Thus, for any block size smaller than

  2815.     // 16x16, ignore the odd portion of the block.

  2816.     if (tx == TX_4X4 && is_uv) {

  2817.         if (h == 1) {

  2818.             if (row_and_7 & 1)

  2819.                 return;

  2820.             if (!row_end)

  2821.                 h += 1;

  2822.         }

  2823.         if (w == 1) {

  2824.             if (col_and_7 & 1)

  2825.                 return;

  2826.             if (!col_end)

  2827.                 w += 1;

  2828.         }

  2829.     }

  2830. 

  2831.     if (tx == TX_4X4 && !skip_inter) {

  2832.         int t = 1 << col_and_7, m_col = (t << w) - t, y;

  2833.         int m_col_odd = (t << (w - 1)) - t;

  2834. 

  2835.         // on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide

  2836.         if (is_uv) {

  2837.             int m_row_8 = m_col & 0x01, m_row_4 = m_col - m_row_8;

  2838. 

  2839.             for (y = row_and_7; y < h + row_and_7; y++) {

  2840.                 int col_mask_id = 2 - !(y & 7);

  2841. 

  2842.                 lflvl->mask[is_uv][0][y][1] |= m_row_8;

  2843.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  2844.                 // for odd lines, if the odd col is not being filtered,

  2845.                 // skip odd row also:

  2846.                 // .---. <-- a

  2847.                 // |   |

  2848.                 // |___| <-- b

  2849.                 // ^   ^

  2850.                 // c   d

  2851.                 //

  2852.                 // if a/c are even row/col and b/d are odd, and d is skipped,

  2853.                 // e.g. right edge of size-66x66.webm, then skip b also (bug)

  2854.                 if ((col_end & 1) && (y & 1)) {

  2855.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col_odd;

  2856.                 } else {

  2857.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col;

  2858.                 }

  2859.             }

  2860.         } else {

  2861.             int m_row_8 = m_col & 0x11, m_row_4 = m_col - m_row_8;

  2862. 

  2863.             for (y = row_and_7; y < h + row_and_7; y++) {

  2864.                 int col_mask_id = 2 - !(y & 3);

  2865. 

  2866.                 lflvl->mask[is_uv][0][y][1] |= m_row_8; // row edge

  2867.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  2868.                 lflvl->mask[is_uv][1][y][col_mask_id] |= m_col; // col edge

  2869.                 lflvl->mask[is_uv][0][y][3] |= m_col;

  2870.                 lflvl->mask[is_uv][1][y][3] |= m_col;

  2871.             }

  2872.         }

  2873.     } else {

  2874.         int y, t = 1 << col_and_7, m_col = (t << w) - t;

  2875. 

  2876.         if (!skip_inter) {

  2877.             int mask_id = (tx == TX_8X8);

  2878.             int l2 = tx + is_uv - 1, step1d = 1 << l2;

  2879.             static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };

  2880.             int m_row = m_col & masks[l2];

  2881. 

  2882.             // at odd UV col/row edges tx16/tx32 loopfilter edges, force

  2883.             // 8wd loopfilter to prevent going off the visible edge.

  2884.             if (is_uv && tx > TX_8X8 && (w ^ (w - 1)) == 1) {

  2885.                 int m_row_16 = ((t << (w - 1)) - t) & masks[l2];

  2886.                 int m_row_8 = m_row - m_row_16;

  2887. 

  2888.                 for (y = row_and_7; y < h + row_and_7; y++) {

  2889.                     lflvl->mask[is_uv][0][y][0] |= m_row_16;

  2890.                     lflvl->mask[is_uv][0][y][1] |= m_row_8;

  2891.                 }

  2892.             } else {

  2893.                 for (y = row_and_7; y < h + row_and_7; y++)

  2894.                     lflvl->mask[is_uv][0][y][mask_id] |= m_row;

  2895.             }

  2896. 

  2897.             if (is_uv && tx > TX_8X8 && (h ^ (h - 1)) == 1) {

  2898.                 for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)

  2899.                     lflvl->mask[is_uv][1][y][0] |= m_col;

  2900.                 if (y - row_and_7 == h - 1)

  2901.                     lflvl->mask[is_uv][1][y][1] |= m_col;

  2902.             } else {

  2903.                 for (y = row_and_7; y < h + row_and_7; y += step1d)

  2904.                     lflvl->mask[is_uv][1][y][mask_id] |= m_col;

  2905.             }

  2906.         } else if (tx != TX_4X4) {

  2907.             int mask_id;

  2908. 

  2909.             mask_id = (tx == TX_8X8) || (is_uv && h == 1);

  2910.             lflvl->mask[is_uv][1][row_and_7][mask_id] |= m_col;

  2911.             mask_id = (tx == TX_8X8) || (is_uv && w == 1);

  2912.             for (y = row_and_7; y < h + row_and_7; y++)

  2913.                 lflvl->mask[is_uv][0][y][mask_id] |= t;

  2914.         } else if (is_uv) {

  2915.             int t8 = t & 0x01, t4 = t - t8;

  2916. 

  2917.             for (y = row_and_7; y < h + row_and_7; y++) {

  2918.                 lflvl->mask[is_uv][0][y][2] |= t4;

  2919.                 lflvl->mask[is_uv][0][y][1] |= t8;

  2920.             }

  2921.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 7)] |= m_col;

  2922.         } else {

  2923.             int t8 = t & 0x11, t4 = t - t8;

  2924. 

  2925.             for (y = row_and_7; y < h + row_and_7; y++) {

  2926.                 lflvl->mask[is_uv][0][y][2] |= t4;

  2927.                 lflvl->mask[is_uv][0][y][1] |= t8;

  2928.             }

  2929.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 3)] |= m_col;

  2930.         }

  2931.     }

  2932. }

  2933. 

  2934. static void decode_b(AVCodecContext *ctx, int row, int col,

  2935.                      struct VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,

  2936.                      enum BlockLevel bl, enum BlockPartition bp)

  2937. {

  2938.     VP9Context *s = ctx->priv_data;

  2939.     VP9Block *b = s->b;

  2940.     enum BlockSize bs = bl * 3 + bp;

  2941.     int w4 = bwh_tab[1][bs][0], h4 = bwh_tab[1][bs][1], lvl;

  2942.     int emu[2];

  2943.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2944. 

  2945.     s->row = row;

  2946.     s->row7 = row & 7;

  2947.     s->col = col;

  2948.     s->col7 = col & 7;

  2949.     s->min_mv.x = -(128 + col * 64);

  2950.     s->min_mv.y = -(128 + row * 64);

  2951.     s->max_mv.x = 128 + (s->cols - col - w4) * 64;

  2952.     s->max_mv.y = 128 + (s->rows - row - h4) * 64;

  2953.     if (s->pass < 2) {

  2954.         b->bs = bs;

  2955.         b->bl = bl;

  2956.         b->bp = bp;

  2957.         decode_mode(ctx);

  2958.         b->uvtx = b->tx - (w4 * 2 == (1 << b->tx) || h4 * 2 == (1 << b->tx));

  2959. 

  2960.         if (!b->skip) {

  2961.             decode_coeffs(ctx);

  2962.         } else {

  2963.             int row7 = s->row7;

  2964. 

  2965. #define SPLAT_ZERO_CTX(v, n) \

  2966.     switch (n) { \

  2967.     case 1:  v = 0;          break; \

  2968.     case 2:  AV_ZERO16(&v);  break; \

  2969.     case 4:  AV_ZERO32(&v);  break; \

  2970.     case 8:  AV_ZERO64(&v);  break; \

  2971.     case 16: AV_ZERO128(&v); break; \

  2972.     }

  2973. #define SPLAT_ZERO_YUV(dir, var, off, n) \

  2974.     do { \

  2975.         SPLAT_ZERO_CTX(s->dir##_y_##var[off * 2], n * 2); \

  2976.         SPLAT_ZERO_CTX(s->dir##_uv_##var[0][off], n); \

  2977.         SPLAT_ZERO_CTX(s->dir##_uv_##var[1][off], n); \

  2978.     } while (0)

  2979. 

  2980.             switch (w4) {

  2981.             case 1: SPLAT_ZERO_YUV(above, nnz_ctx, col, 1); break;

  2982.             case 2: SPLAT_ZERO_YUV(above, nnz_ctx, col, 2); break;

  2983.             case 4: SPLAT_ZERO_YUV(above, nnz_ctx, col, 4); break;

  2984.             case 8: SPLAT_ZERO_YUV(above, nnz_ctx, col, 8); break;

  2985.             }

  2986.             switch (h4) {

  2987.             case 1: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 1); break;

  2988.             case 2: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 2); break;

  2989.             case 4: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 4); break;

  2990.             case 8: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 8); break;

  2991.             }

  2992.         }

  2993.         if (s->pass == 1) {

  2994.             s->b++;

  2995.             s->block += w4 * h4 * 64;

  2996.             s->uvblock[0] += w4 * h4 * 16;

  2997.             s->uvblock[1] += w4 * h4 * 16;

  2998.             s->eob += 4 * w4 * h4;

  2999.             s->uveob[0] += w4 * h4;

  3000.             s->uveob[1] += w4 * h4;

  3001. 

  3002.             return;

  3003.         }

  3004.     }

  3005. 

  3006.     // emulated overhangs if the stride of the target buffer can't hold. This

  3007.     // allows to support emu-edge and so on even if we have large block

  3008.     // overhangs

  3009.     emu[0] = (col + w4) * 8 > f->linesize[0] ||

  3010.              (row + h4) > s->rows;

  3011.     emu[1] = (col + w4) * 4 > f->linesize[1] ||

  3012.              (row + h4) > s->rows;

  3013.     if (emu[0]) {

  3014.         s->dst[0] = s->tmp_y;

  3015.         s->y_stride = 64;

  3016.     } else {

  3017.         s->dst[0] = f->data[0] + yoff;

  3018.         s->y_stride = f->linesize[0];

  3019.     }

  3020.     if (emu[1]) {

  3021.         s->dst[1] = s->tmp_uv[0];

  3022.         s->dst[2] = s->tmp_uv[1];

  3023.         s->uv_stride = 32;

  3024.     } else {

  3025.         s->dst[1] = f->data[1] + uvoff;

  3026.         s->dst[2] = f->data[2] + uvoff;

  3027.         s->uv_stride = f->linesize[1];

  3028.     }

  3029.     if (b->intra) {

  3030.         intra_recon(ctx, yoff, uvoff);

  3031.     } else {

  3032.         inter_recon(ctx);

  3033.     }

  3034.     if (emu[0]) {

  3035.         int w = FFMIN(s->cols - col, w4) * 8, h = FFMIN(s->rows - row, h4) * 8, n, o = 0;

  3036. 

  3037.         for (n = 0; o < w; n++) {

  3038.             int bw = 64 >> n;

  3039. 

  3040.             av_assert2(n <= 4);

  3041.             if (w & bw) {

  3042.                 s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o, f->linesize[0],

  3043.                                          s->tmp_y + o, 64, h, 0, 0);

  3044.                 o += bw;

  3045.             }

  3046.         }

  3047.     }

  3048.     if (emu[1]) {

  3049.         int w = FFMIN(s->cols - col, w4) * 4, h = FFMIN(s->rows - row, h4) * 4, n, o = 0;

  3050. 

  3051.         for (n = 1; o < w; n++) {

  3052.             int bw = 64 >> n;

  3053. 

  3054.             av_assert2(n <= 4);

  3055.             if (w & bw) {

  3056.                 s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o, f->linesize[1],

  3057.                                          s->tmp_uv[0] + o, 32, h, 0, 0);

  3058.                 s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o, f->linesize[2],

  3059.                                          s->tmp_uv[1] + o, 32, h, 0, 0);

  3060.                 o += bw;

  3061.             }

  3062.         }

  3063.     }

  3064. 

  3065.     // pick filter level and find edges to apply filter to

  3066.     if (s->filter.level &&

  3067.         (lvl = s->segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]

  3068.                                                     [b->mode[3] != ZEROMV]) > 0) {

  3069.         int x_end = FFMIN(s->cols - col, w4), y_end = FFMIN(s->rows - row, h4);

  3070.         int skip_inter = !b->intra && b->skip, col7 = s->col7, row7 = s->row7;

  3071. 

  3072.         setctx_2d(&lflvl->level[row7 * 8 + col7], w4, h4, 8, lvl);

  3073.         mask_edges(lflvl, 0, row7, col7, x_end, y_end, 0, 0, b->tx, skip_inter);

  3074.         mask_edges(lflvl, 1, row7, col7, x_end, y_end,

  3075.                    s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,

  3076.                    s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,

  3077.                    b->uvtx, skip_inter);

  3078. 

  3079.         if (!s->filter.lim_lut[lvl]) {

  3080.             int sharp = s->filter.sharpness;

  3081.             int limit = lvl;

  3082. 

  3083.             if (sharp > 0) {

  3084.                 limit >>= (sharp + 3) >> 2;

  3085.                 limit = FFMIN(limit, 9 - sharp);

  3086.             }

  3087.             limit = FFMAX(limit, 1);

  3088. 

  3089.             s->filter.lim_lut[lvl] = limit;

  3090.             s->filter.mblim_lut[lvl] = 2 * (lvl + 2) + limit;

  3091.         }

  3092.     }

  3093. 

  3094.     if (s->pass == 2) {

  3095.         s->b++;

  3096.         s->block += w4 * h4 * 64;

  3097.         s->uvblock[0] += w4 * h4 * 16;

  3098.         s->uvblock[1] += w4 * h4 * 16;

  3099.         s->eob += 4 * w4 * h4;

  3100.         s->uveob[0] += w4 * h4;

  3101.         s->uveob[1] += w4 * h4;

  3102.     }

  3103. }

  3104. 

  3105. static void decode_sb(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  3106.                       ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  3107. {

  3108.     VP9Context *s = ctx->priv_data;

  3109.     int c = ((s->above_partition_ctx[col] >> (3 - bl)) & 1) |

  3110.             (((s->left_partition_ctx[row & 0x7] >> (3 - bl)) & 1) << 1);

  3111.     const uint8_t *p = s->keyframe ? vp9_default_kf_partition_probs[bl][c] :

  3112.                                      s->prob.p.partition[bl][c];

  3113.     enum BlockPartition bp;

  3114.     ptrdiff_t hbs = 4 >> bl;

  3115.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  3116.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  3117. 

  3118.     if (bl == BL_8X8) {

  3119.         bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  3120.         decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3121.     } else if (col + hbs < s->cols) { // FIXME why not <=?

  3122.         if (row + hbs < s->rows) { // FIXME why not <=?

  3123.             bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  3124.             switch (bp) {

  3125.             case PARTITION_NONE:

  3126.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3127.                 break;

  3128.             case PARTITION_H:

  3129.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3130.                 yoff  += hbs * 8 * y_stride;

  3131.                 uvoff += hbs * 4 * uv_stride;

  3132.                 decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, bl, bp);

  3133.                 break;

  3134.             case PARTITION_V:

  3135.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3136.                 yoff  += hbs * 8;

  3137.                 uvoff += hbs * 4;

  3138.                 decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, bl, bp);

  3139.                 break;

  3140.             case PARTITION_SPLIT:

  3141.                 decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3142.                 decode_sb(ctx, row, col + hbs, lflvl,

  3143.                           yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  3144.                 yoff  += hbs * 8 * y_stride;

  3145.                 uvoff += hbs * 4 * uv_stride;

  3146.                 decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3147.                 decode_sb(ctx, row + hbs, col + hbs, lflvl,

  3148.                           yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  3149.                 break;

  3150.             default:

  3151.                 av_assert0(0);

  3152.             }

  3153.         } else if (vp56_rac_get_prob_branchy(&s->c, p[1])) {

  3154.             bp = PARTITION_SPLIT;

  3155.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3156.             decode_sb(ctx, row, col + hbs, lflvl,

  3157.                       yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  3158.         } else {

  3159.             bp = PARTITION_H;

  3160.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3161.         }

  3162.     } else if (row + hbs < s->rows) { // FIXME why not <=?

  3163.         if (vp56_rac_get_prob_branchy(&s->c, p[2])) {

  3164.             bp = PARTITION_SPLIT;

  3165.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3166.             yoff  += hbs * 8 * y_stride;

  3167.             uvoff += hbs * 4 * uv_stride;

  3168.             decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3169.         } else {

  3170.             bp = PARTITION_V;

  3171.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3172.         }

  3173.     } else {

  3174.         bp = PARTITION_SPLIT;

  3175.         decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3176.     }

  3177.     s->counts.partition[bl][c][bp]++;

  3178. }

  3179. 

  3180. static void decode_sb_mem(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  3181.                           ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  3182. {

  3183.     VP9Context *s = ctx->priv_data;

  3184.     VP9Block *b = s->b;

  3185.     ptrdiff_t hbs = 4 >> bl;

  3186.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  3187.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  3188. 

  3189.     if (bl == BL_8X8) {

  3190.         av_assert2(b->bl == BL_8X8);

  3191.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  3192.     } else if (s->b->bl == bl) {

  3193.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  3194.         if (b->bp == PARTITION_H && row + hbs < s->rows) {

  3195.             yoff  += hbs * 8 * y_stride;

  3196.             uvoff += hbs * 4 * uv_stride;

  3197.             decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, b->bl, b->bp);

  3198.         } else if (b->bp == PARTITION_V && col + hbs < s->cols) {

  3199.             yoff  += hbs * 8;

  3200.             uvoff += hbs * 4;

  3201.             decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, b->bl, b->bp);

  3202.         }

  3203.     } else {

  3204.         decode_sb_mem(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3205.         if (col + hbs < s->cols) { // FIXME why not <=?

  3206.             if (row + hbs < s->rows) {

  3207.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff + 8 * hbs,

  3208.                               uvoff + 4 * hbs, bl + 1);

  3209.                 yoff  += hbs * 8 * y_stride;

  3210.                 uvoff += hbs * 4 * uv_stride;

  3211.                 decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3212.                 decode_sb_mem(ctx, row + hbs, col + hbs, lflvl,

  3213.                                     yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  3214.             } else {

  3215.                 yoff  += hbs * 8;

  3216.                 uvoff += hbs * 4;

  3217.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff, uvoff, bl + 1);

  3218.             }

  3219.         } else if (row + hbs < s->rows) {

  3220.             yoff  += hbs * 8 * y_stride;

  3221.             uvoff += hbs * 4 * uv_stride;

  3222.             decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3223.         }

  3224.     }

  3225. }

  3226. 

  3227. static void loopfilter_sb(AVCodecContext *ctx, struct VP9Filter *lflvl,

  3228.                           int row, int col, ptrdiff_t yoff, ptrdiff_t uvoff)

  3229. {

  3230.     VP9Context *s = ctx->priv_data;

  3231.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  3232.     uint8_t *dst = f->data[0] + yoff, *lvl = lflvl->level;

  3233.     ptrdiff_t ls_y = f->linesize[0], ls_uv = f->linesize[1];

  3234.     int y, x, p;

  3235. 

  3236.     // FIXME in how far can we interleave the v/h loopfilter calls? E.g.

  3237.     // if you think of them as acting on a 8x8 block max, we can interleave

  3238.     // each v/h within the single x loop, but that only works if we work on

  3239.     // 8 pixel blocks, and we won't always do that (we want at least 16px

  3240.     // to use SSE2 optimizations, perhaps 32 for AVX2)

  3241. 

  3242.     // filter edges between columns, Y plane (e.g. block1 | block2)

  3243.     for (y = 0; y < 8; y += 2, dst += 16 * ls_y, lvl += 16) {

  3244.         uint8_t *ptr = dst, *l = lvl, *hmask1 = lflvl->mask[0][0][y];

  3245.         uint8_t *hmask2 = lflvl->mask[0][0][y + 1];

  3246.         unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2], hm13 = hmask1[3];

  3247.         unsigned hm2 = hmask2[1] | hmask2[2], hm23 = hmask2[3];

  3248.         unsigned hm = hm1 | hm2 | hm13 | hm23;

  3249. 

  3250.         for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 8, l++) {

  3251.             if (hm1 & x) {

  3252.                 int L = *l, H = L >> 4;

  3253.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3254. 

  3255.                 if (col || x > 1) {

  3256.                     if (hmask1[0] & x) {

  3257.                         if (hmask2[0] & x) {

  3258.                             av_assert2(l[8] == L);

  3259.                             s->dsp.loop_filter_16[0](ptr, ls_y, E, I, H);

  3260.                         } else {

  3261.                             s->dsp.loop_filter_8[2][0](ptr, ls_y, E, I, H);

  3262.                         }

  3263.                     } else if (hm2 & x) {

  3264.                         L = l[8];

  3265.                         H |= (L >> 4) << 8;

  3266.                         E |= s->filter.mblim_lut[L] << 8;

  3267.                         I |= s->filter.lim_lut[L] << 8;

  3268.                         s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3269.                                                [!!(hmask2[1] & x)]

  3270.                                                [0](ptr, ls_y, E, I, H);

  3271.                     } else {

  3272.                         s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3273.                                             [0](ptr, ls_y, E, I, H);

  3274.                     }

  3275.                 }

  3276.             } else if (hm2 & x) {

  3277.                 int L = l[8], H = L >> 4;

  3278.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3279. 

  3280.                 if (col || x > 1) {

  3281.                     s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3282.                                         [0](ptr + 8 * ls_y, ls_y, E, I, H);

  3283.                 }

  3284.             }

  3285.             if (hm13 & x) {

  3286.                 int L = *l, H = L >> 4;

  3287.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3288. 

  3289.                 if (hm23 & x) {

  3290.                     L = l[8];

  3291.                     H |= (L >> 4) << 8;

  3292.                     E |= s->filter.mblim_lut[L] << 8;

  3293.                     I |= s->filter.lim_lut[L] << 8;

  3294.                     s->dsp.loop_filter_mix2[0][0][0](ptr + 4, ls_y, E, I, H);

  3295.                 } else {

  3296.                     s->dsp.loop_filter_8[0][0](ptr + 4, ls_y, E, I, H);

  3297.                 }

  3298.             } else if (hm23 & x) {

  3299.                 int L = l[8], H = L >> 4;

  3300.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3301. 

  3302.                 s->dsp.loop_filter_8[0][0](ptr + 8 * ls_y + 4, ls_y, E, I, H);

  3303.             }

  3304.         }

  3305.     }

  3306. 

  3307.     //                                          block1

  3308.     // filter edges between rows, Y plane (e.g. ------)

  3309.     //                                          block2

  3310.     dst = f->data[0] + yoff;

  3311.     lvl = lflvl->level;

  3312.     for (y = 0; y < 8; y++, dst += 8 * ls_y, lvl += 8) {

  3313.         uint8_t *ptr = dst, *l = lvl, *vmask = lflvl->mask[0][1][y];

  3314.         unsigned vm = vmask[0] | vmask[1] | vmask[2], vm3 = vmask[3];

  3315. 

  3316.         for (x = 1; vm & ~(x - 1); x <<= 2, ptr += 16, l += 2) {

  3317.             if (row || y) {

  3318.                 if (vm & x) {

  3319.                     int L = *l, H = L >> 4;

  3320.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3321. 

  3322.                     if (vmask[0] & x) {

  3323.                         if (vmask[0] & (x << 1)) {

  3324.                             av_assert2(l[1] == L);

  3325.                             s->dsp.loop_filter_16[1](ptr, ls_y, E, I, H);

  3326.                         } else {

  3327.                             s->dsp.loop_filter_8[2][1](ptr, ls_y, E, I, H);

  3328.                         }

  3329.                     } else if (vm & (x << 1)) {

  3330.                         L = l[1];

  3331.                         H |= (L >> 4) << 8;

  3332.                         E |= s->filter.mblim_lut[L] << 8;

  3333.                         I |= s->filter.lim_lut[L] << 8;

  3334.                         s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3335.                                                [!!(vmask[1] & (x << 1))]

  3336.                                                [1](ptr, ls_y, E, I, H);

  3337.                     } else {

  3338.                         s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3339.                                             [1](ptr, ls_y, E, I, H);

  3340.                     }

  3341.                 } else if (vm & (x << 1)) {

  3342.                     int L = l[1], H = L >> 4;

  3343.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3344. 

  3345.                     s->dsp.loop_filter_8[!!(vmask[1] & (x << 1))]

  3346.                                         [1](ptr + 8, ls_y, E, I, H);

  3347.                 }

  3348.             }

  3349.             if (vm3 & x) {

  3350.                 int L = *l, H = L >> 4;

  3351.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3352. 

  3353.                 if (vm3 & (x << 1)) {

  3354.                     L = l[1];

  3355.                     H |= (L >> 4) << 8;

  3356.                     E |= s->filter.mblim_lut[L] << 8;

  3357.                     I |= s->filter.lim_lut[L] << 8;

  3358.                     s->dsp.loop_filter_mix2[0][0][1](ptr + ls_y * 4, ls_y, E, I, H);

  3359.                 } else {

  3360.                     s->dsp.loop_filter_8[0][1](ptr + ls_y * 4, ls_y, E, I, H);

  3361.                 }

  3362.             } else if (vm3 & (x << 1)) {

  3363.                 int L = l[1], H = L >> 4;

  3364.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3365. 

  3366.                 s->dsp.loop_filter_8[0][1](ptr + ls_y * 4 + 8, ls_y, E, I, H);

  3367.             }

  3368.         }

  3369.     }

  3370. 

  3371.     // same principle but for U/V planes

  3372.     for (p = 0; p < 2; p++) {

  3373.         lvl = lflvl->level;

  3374.         dst = f->data[1 + p] + uvoff;

  3375.         for (y = 0; y < 8; y += 4, dst += 16 * ls_uv, lvl += 32) {

  3376.             uint8_t *ptr = dst, *l = lvl, *hmask1 = lflvl->mask[1][0][y];

  3377.             uint8_t *hmask2 = lflvl->mask[1][0][y + 2];

  3378.             unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2];

  3379.             unsigned hm2 = hmask2[1] | hmask2[2], hm = hm1 | hm2;

  3380. 

  3381.             for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 4) {

  3382.                 if (col || x > 1) {

  3383.                     if (hm1 & x) {

  3384.                         int L = *l, H = L >> 4;

  3385.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3386. 

  3387.                         if (hmask1[0] & x) {

  3388.                             if (hmask2[0] & x) {

  3389.                                 av_assert2(l[16] == L);

  3390.                                 s->dsp.loop_filter_16[0](ptr, ls_uv, E, I, H);

  3391.                             } else {

  3392.                                 s->dsp.loop_filter_8[2][0](ptr, ls_uv, E, I, H);

  3393.                             }

  3394.                         } else if (hm2 & x) {

  3395.                             L = l[16];

  3396.                             H |= (L >> 4) << 8;

  3397.                             E |= s->filter.mblim_lut[L] << 8;

  3398.                             I |= s->filter.lim_lut[L] << 8;

  3399.                             s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3400.                                                    [!!(hmask2[1] & x)]

  3401.                                                    [0](ptr, ls_uv, E, I, H);

  3402.                         } else {

  3403.                             s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3404.                                                 [0](ptr, ls_uv, E, I, H);

  3405.                         }

  3406.                     } else if (hm2 & x) {

  3407.                         int L = l[16], H = L >> 4;

  3408.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3409. 

  3410.                         s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3411.                                             [0](ptr + 8 * ls_uv, ls_uv, E, I, H);

  3412.                     }

  3413.                 }

  3414.                 if (x & 0xAA)

  3415.                     l += 2;

  3416.             }

  3417.         }

  3418.         lvl = lflvl->level;

  3419.         dst = f->data[1 + p] + uvoff;

  3420.         for (y = 0; y < 8; y++, dst += 4 * ls_uv) {

  3421.             uint8_t *ptr = dst, *l = lvl, *vmask = lflvl->mask[1][1][y];

  3422.             unsigned vm = vmask[0] | vmask[1] | vmask[2];

  3423. 

  3424.             for (x = 1; vm & ~(x - 1); x <<= 4, ptr += 16, l += 4) {

  3425.                 if (row || y) {

  3426.                     if (vm & x) {

  3427.                         int L = *l, H = L >> 4;

  3428.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3429. 

  3430.                         if (vmask[0] & x) {

  3431.                             if (vmask[0] & (x << 2)) {

  3432.                                 av_assert2(l[2] == L);

  3433.                                 s->dsp.loop_filter_16[1](ptr, ls_uv, E, I, H);

  3434.                             } else {

  3435.                                 s->dsp.loop_filter_8[2][1](ptr, ls_uv, E, I, H);

  3436.                             }

  3437.                         } else if (vm & (x << 2)) {

  3438.                             L = l[2];

  3439.                             H |= (L >> 4) << 8;

  3440.                             E |= s->filter.mblim_lut[L] << 8;

  3441.                             I |= s->filter.lim_lut[L] << 8;

  3442.                             s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3443.                                                    [!!(vmask[1] & (x << 2))]

  3444.                                                    [1](ptr, ls_uv, E, I, H);

  3445.                         } else {

  3446.                             s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3447.                                                 [1](ptr, ls_uv, E, I, H);

  3448.                         }

  3449.                     } else if (vm & (x << 2)) {

  3450.                         int L = l[2], H = L >> 4;

  3451.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3452. 

  3453.                         s->dsp.loop_filter_8[!!(vmask[1] & (x << 2))]

  3454.                                             [1](ptr + 8, ls_uv, E, I, H);

  3455.                     }

  3456.                 }

  3457.             }

  3458.             if (y & 1)

  3459.                 lvl += 16;

  3460.         }

  3461.     }

  3462. }

  3463. 

  3464. static void set_tile_offset(int *start, int *end, int idx, int log2_n, int n)

  3465. {

  3466.     int sb_start = ( idx      * n) >> log2_n;

  3467.     int sb_end   = ((idx + 1) * n) >> log2_n;

  3468.     *start = FFMIN(sb_start, n) << 3;

  3469.     *end   = FFMIN(sb_end,   n) << 3;

  3470. }

  3471. 

  3472. static av_always_inline void adapt_prob(uint8_t *p, unsigned ct0, unsigned ct1,

  3473.                                         int max_count, int update_factor)

  3474. {

  3475.     unsigned ct = ct0 + ct1, p2, p1;

  3476. 

  3477.     if (!ct)

  3478.         return;

  3479. 

  3480.     p1 = *p;

  3481.     p2 = ((ct0 << 8) + (ct >> 1)) / ct;

  3482.     p2 = av_clip(p2, 1, 255);

  3483.     ct = FFMIN(ct, max_count);

  3484.     update_factor = FASTDIV(update_factor * ct, max_count);

  3485. 

  3486.     // (p1 * (256 - update_factor) + p2 * update_factor + 128) >> 8

  3487.     *p = p1 + (((p2 - p1) * update_factor + 128) >> 8);

  3488. }

  3489. 

  3490. static void adapt_probs(VP9Context *s)

  3491. {

  3492.     int i, j, k, l, m;

  3493.     prob_context *p = &s->prob_ctx[s->framectxid].p;

  3494.     int uf = (s->keyframe || s->intraonly || !s->last_keyframe) ? 112 : 128;

  3495. 

  3496.     // coefficients

  3497.     for (i = 0; i < 4; i++)

  3498.         for (j = 0; j < 2; j++)

  3499.             for (k = 0; k < 2; k++)

  3500.                 for (l = 0; l < 6; l++)

  3501.                     for (m = 0; m < 6; m++) {

  3502.                         uint8_t *pp = s->prob_ctx[s->framectxid].coef[i][j][k][l][m];

  3503.                         unsigned *e = s->counts.eob[i][j][k][l][m];

  3504.                         unsigned *c = s->counts.coef[i][j][k][l][m];

  3505. 

  3506.                         if (l == 0 && m >= 3) // dc only has 3 pt

  3507.                             break;

  3508. 

  3509.                         adapt_prob(&pp[0], e[0], e[1], 24, uf);

  3510.                         adapt_prob(&pp[1], c[0], c[1] + c[2], 24, uf);

  3511.                         adapt_prob(&pp[2], c[1], c[2], 24, uf);

  3512.                     }

  3513. 

  3514.     if (s->keyframe || s->intraonly) {

  3515.         memcpy(p->skip,  s->prob.p.skip,  sizeof(p->skip));

  3516.         memcpy(p->tx32p, s->prob.p.tx32p, sizeof(p->tx32p));

  3517.         memcpy(p->tx16p, s->prob.p.tx16p, sizeof(p->tx16p));

  3518.         memcpy(p->tx8p,  s->prob.p.tx8p,  sizeof(p->tx8p));

  3519.         return;

  3520.     }

  3521. 

  3522.     // skip flag

  3523.     for (i = 0; i < 3; i++)

  3524.         adapt_prob(&p->skip[i], s->counts.skip[i][0], s->counts.skip[i][1], 20, 128);

  3525. 

  3526.     // intra/inter flag

  3527.     for (i = 0; i < 4; i++)

  3528.         adapt_prob(&p->intra[i], s->counts.intra[i][0], s->counts.intra[i][1], 20, 128);

  3529. 

  3530.     // comppred flag

  3531.     if (s->comppredmode == PRED_SWITCHABLE) {

  3532.       for (i = 0; i < 5; i++)

  3533.           adapt_prob(&p->comp[i], s->counts.comp[i][0], s->counts.comp[i][1], 20, 128);

  3534.     }

  3535. 

  3536.     // reference frames

  3537.     if (s->comppredmode != PRED_SINGLEREF) {

  3538.       for (i = 0; i < 5; i++)

  3539.           adapt_prob(&p->comp_ref[i], s->counts.comp_ref[i][0],

  3540.                      s->counts.comp_ref[i][1], 20, 128);

  3541.     }

  3542. 

  3543.     if (s->comppredmode != PRED_COMPREF) {

  3544.       for (i = 0; i < 5; i++) {

  3545.           uint8_t *pp = p->single_ref[i];

  3546.           unsigned (*c)[2] = s->counts.single_ref[i];

  3547. 

  3548.           adapt_prob(&pp[0], c[0][0], c[0][1], 20, 128);

  3549.           adapt_prob(&pp[1], c[1][0], c[1][1], 20, 128);

  3550.       }

  3551.     }

  3552. 

  3553.     // block partitioning

  3554.     for (i = 0; i < 4; i++)

  3555.         for (j = 0; j < 4; j++) {

  3556.             uint8_t *pp = p->partition[i][j];

  3557.             unsigned *c = s->counts.partition[i][j];

  3558. 

  3559.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3560.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3561.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3562.         }

  3563. 

  3564.     // tx size

  3565.     if (s->txfmmode == TX_SWITCHABLE) {

  3566.       for (i = 0; i < 2; i++) {

  3567.           unsigned *c16 = s->counts.tx16p[i], *c32 = s->counts.tx32p[i];

  3568. 

  3569.           adapt_prob(&p->tx8p[i], s->counts.tx8p[i][0], s->counts.tx8p[i][1], 20, 128);

  3570.           adapt_prob(&p->tx16p[i][0], c16[0], c16[1] + c16[2], 20, 128);

  3571.           adapt_prob(&p->tx16p[i][1], c16[1], c16[2], 20, 128);

  3572.           adapt_prob(&p->tx32p[i][0], c32[0], c32[1] + c32[2] + c32[3], 20, 128);

  3573.           adapt_prob(&p->tx32p[i][1], c32[1], c32[2] + c32[3], 20, 128);

  3574.           adapt_prob(&p->tx32p[i][2], c32[2], c32[3], 20, 128);

  3575.       }

  3576.     }

  3577. 

  3578.     // interpolation filter

  3579.     if (s->filtermode == FILTER_SWITCHABLE) {

  3580.         for (i = 0; i < 4; i++) {

  3581.             uint8_t *pp = p->filter[i];

  3582.             unsigned *c = s->counts.filter[i];

  3583. 

  3584.             adapt_prob(&pp[0], c[0], c[1] + c[2], 20, 128);

  3585.             adapt_prob(&pp[1], c[1], c[2], 20, 128);

  3586.         }

  3587.     }

  3588. 

  3589.     // inter modes

  3590.     for (i = 0; i < 7; i++) {

  3591.         uint8_t *pp = p->mv_mode[i];

  3592.         unsigned *c = s->counts.mv_mode[i];

  3593. 

  3594.         adapt_prob(&pp[0], c[2], c[1] + c[0] + c[3], 20, 128);

  3595.         adapt_prob(&pp[1], c[0], c[1] + c[3], 20, 128);

  3596.         adapt_prob(&pp[2], c[1], c[3], 20, 128);

  3597.     }

  3598. 

  3599.     // mv joints

  3600.     {

  3601.         uint8_t *pp = p->mv_joint;

  3602.         unsigned *c = s->counts.mv_joint;

  3603. 

  3604.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3605.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3606.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3607.     }

  3608. 

  3609.     // mv components

  3610.     for (i = 0; i < 2; i++) {

  3611.         uint8_t *pp;

  3612.         unsigned *c, (*c2)[2], sum;

  3613. 

  3614.         adapt_prob(&p->mv_comp[i].sign, s->counts.mv_comp[i].sign[0],

  3615.                    s->counts.mv_comp[i].sign[1], 20, 128);

  3616. 

  3617.         pp = p->mv_comp[i].classes;

  3618.         c = s->counts.mv_comp[i].classes;

  3619.         sum = c[1] + c[2] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9] + c[10];

  3620.         adapt_prob(&pp[0], c[0], sum, 20, 128);

  3621.         sum -= c[1];

  3622.         adapt_prob(&pp[1], c[1], sum, 20, 128);

  3623.         sum -= c[2] + c[3];

  3624.         adapt_prob(&pp[2], c[2] + c[3], sum, 20, 128);

  3625.         adapt_prob(&pp[3], c[2], c[3], 20, 128);

  3626.         sum -= c[4] + c[5];

  3627.         adapt_prob(&pp[4], c[4] + c[5], sum, 20, 128);

  3628.         adapt_prob(&pp[5], c[4], c[5], 20, 128);

  3629.         sum -= c[6];

  3630.         adapt_prob(&pp[6], c[6], sum, 20, 128);

  3631.         adapt_prob(&pp[7], c[7] + c[8], c[9] + c[10], 20, 128);

  3632.         adapt_prob(&pp[8], c[7], c[8], 20, 128);

  3633.         adapt_prob(&pp[9], c[9], c[10], 20, 128);

  3634. 

  3635.         adapt_prob(&p->mv_comp[i].class0, s->counts.mv_comp[i].class0[0],

  3636.                    s->counts.mv_comp[i].class0[1], 20, 128);

  3637.         pp = p->mv_comp[i].bits;

  3638.         c2 = s->counts.mv_comp[i].bits;

  3639.         for (j = 0; j < 10; j++)

  3640.             adapt_prob(&pp[j], c2[j][0], c2[j][1], 20, 128);

  3641. 

  3642.         for (j = 0; j < 2; j++) {

  3643.             pp = p->mv_comp[i].class0_fp[j];

  3644.             c = s->counts.mv_comp[i].class0_fp[j];

  3645.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3646.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3647.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3648.         }

  3649.         pp = p->mv_comp[i].fp;

  3650.         c = s->counts.mv_comp[i].fp;

  3651.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3652.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3653.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3654. 

  3655.         if (s->highprecisionmvs) {

  3656.             adapt_prob(&p->mv_comp[i].class0_hp, s->counts.mv_comp[i].class0_hp[0],

  3657.                        s->counts.mv_comp[i].class0_hp[1], 20, 128);

  3658.             adapt_prob(&p->mv_comp[i].hp, s->counts.mv_comp[i].hp[0],

  3659.                        s->counts.mv_comp[i].hp[1], 20, 128);

  3660.         }

  3661.     }

  3662. 

  3663.     // y intra modes

  3664.     for (i = 0; i < 4; i++) {

  3665.         uint8_t *pp = p->y_mode[i];

  3666.         unsigned *c = s->counts.y_mode[i], sum, s2;

  3667. 

  3668.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3669.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3670.         sum -= c[TM_VP8_PRED];

  3671.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3672.         sum -= c[VERT_PRED];

  3673.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3674.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3675.         sum -= s2;

  3676.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3677.         s2 -= c[HOR_PRED];

  3678.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3679.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3680.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3681.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3682.         sum -= c[VERT_LEFT_PRED];

  3683.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3684.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3685.     }

  3686. 

  3687.     // uv intra modes

  3688.     for (i = 0; i < 10; i++) {

  3689.         uint8_t *pp = p->uv_mode[i];

  3690.         unsigned *c = s->counts.uv_mode[i], sum, s2;

  3691. 

  3692.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3693.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3694.         sum -= c[TM_VP8_PRED];

  3695.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3696.         sum -= c[VERT_PRED];

  3697.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3698.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3699.         sum -= s2;

  3700.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3701.         s2 -= c[HOR_PRED];

  3702.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3703.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3704.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3705.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3706.         sum -= c[VERT_LEFT_PRED];

  3707.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3708.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3709.     }

  3710. }

  3711. 

  3712. static void free_buffers(VP9Context *s)

  3713. {

  3714.     av_freep(&s->intra_pred_data[0]);

  3715.     av_freep(&s->b_base);

  3716.     av_freep(&s->block_base);

  3717. }

  3718. 

  3719. static av_cold int vp9_decode_free(AVCodecContext *ctx)

  3720. {

  3721.     VP9Context *s = ctx->priv_data;

  3722.     int i;

  3723. 

  3724.     for (i = 0; i < 3; i++) {

  3725.         if (s->frames[i].tf.f->data[0])

  3726.             vp9_unref_frame(ctx, &s->frames[i]);

  3727.         av_frame_free(&s->frames[i].tf.f);

  3728.     }

  3729.     for (i = 0; i < 8; i++) {

  3730.         if (s->refs[i].f->data[0])

  3731.             ff_thread_release_buffer(ctx, &s->refs[i]);

  3732.         av_frame_free(&s->refs[i].f);

  3733.         if (s->next_refs[i].f->data[0])

  3734.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3735.         av_frame_free(&s->next_refs[i].f);

  3736.     }

  3737.     free_buffers(s);

  3738.     av_freep(&s->c_b);

  3739.     s->c_b_size = 0;

  3740. 

  3741.     return 0;

  3742. }

  3743. 

  3744. 

  3745. static int vp9_decode_frame(AVCodecContext *ctx, void *frame,

  3746.                             int *got_frame, AVPacket *pkt)

  3747. {

  3748.     const uint8_t *data = pkt->data;

  3749.     int size = pkt->size;

  3750.     VP9Context *s = ctx->priv_data;

  3751.     int res, tile_row, tile_col, i, ref, row, col;

  3752.     int retain_segmap_ref = s->segmentation.enabled && !s->segmentation.update_map;

  3753.     ptrdiff_t yoff, uvoff, ls_y, ls_uv;

  3754.     AVFrame *f;

  3755. 

  3756.     if ((res = decode_frame_header(ctx, data, size, &ref)) < 0) {

  3757.         return res;

  3758.     } else if (res == 0) {

  3759.         if (!s->refs[ref].f->data[0]) {

  3760.             av_log(ctx, AV_LOG_ERROR, "Requested reference %d not available\n", ref);

  3761.             return AVERROR_INVALIDDATA;

  3762.         }

  3763.         if ((res = av_frame_ref(frame, s->refs[ref].f)) < 0)

  3764.             return res;

  3765.         *got_frame = 1;

  3766.         return pkt->size;

  3767.     }

  3768.     data += res;

  3769.     size -= res;

  3770. 

  3771.     if (!retain_segmap_ref) {

  3772.         if (s->frames[REF_FRAME_SEGMAP].tf.f->data[0])

  3773.             vp9_unref_frame(ctx, &s->frames[REF_FRAME_SEGMAP]);

  3774.         if (!s->keyframe && !s->intraonly && !s->errorres && s->frames[CUR_FRAME].tf.f->data[0] &&

  3775.             (res = vp9_ref_frame(ctx, &s->frames[REF_FRAME_SEGMAP], &s->frames[CUR_FRAME])) < 0)

  3776.             return res;

  3777.     }

  3778.     if (s->frames[REF_FRAME_MVPAIR].tf.f->data[0])

  3779.         vp9_unref_frame(ctx, &s->frames[REF_FRAME_MVPAIR]);

  3780.     if (!s->intraonly && !s->keyframe && !s->errorres && s->frames[CUR_FRAME].tf.f->data[0] &&

  3781.         (res = vp9_ref_frame(ctx, &s->frames[REF_FRAME_MVPAIR], &s->frames[CUR_FRAME])) < 0)

  3782.         return res;

  3783.     if (s->frames[CUR_FRAME].tf.f->data[0])

  3784.         vp9_unref_frame(ctx, &s->frames[CUR_FRAME]);

  3785.     if ((res = vp9_alloc_frame(ctx, &s->frames[CUR_FRAME])) < 0)

  3786.         return res;

  3787.     f = s->frames[CUR_FRAME].tf.f;

  3788.     f->key_frame = s->keyframe;

  3789.     f->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;

  3790.     ls_y = f->linesize[0];

  3791.     ls_uv =f->linesize[1];

  3792. 

  3793.     // ref frame setup

  3794.     for (i = 0; i < 8; i++) {

  3795.         if (s->next_refs[i].f->data[0])

  3796.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3797.         if (s->refreshrefmask & (1 << i)) {

  3798.             res = ff_thread_ref_frame(&s->next_refs[i], &s->frames[CUR_FRAME].tf);

  3799.         } else {

  3800.             res = ff_thread_ref_frame(&s->next_refs[i], &s->refs[i]);

  3801.         }

  3802.         if (res < 0)

  3803.             return res;

  3804.     }

  3805. 

  3806.     if (s->fullrange)

  3807.         ctx->color_range = AVCOL_RANGE_JPEG;

  3808.     else

  3809.         ctx->color_range = AVCOL_RANGE_MPEG;

  3810. 

  3811.     switch (s->colorspace) {

  3812.     case 1: ctx->colorspace = AVCOL_SPC_BT470BG; break;

  3813.     case 2: ctx->colorspace = AVCOL_SPC_BT709; break;

  3814.     case 3: ctx->colorspace = AVCOL_SPC_SMPTE170M; break;

  3815.     case 4: ctx->colorspace = AVCOL_SPC_SMPTE240M; break;

  3816.     }

  3817. 

  3818.     // main tile decode loop

  3819.     memset(s->above_partition_ctx, 0, s->cols);

  3820.     memset(s->above_skip_ctx, 0, s->cols);

  3821.     if (s->keyframe || s->intraonly) {

  3822.         memset(s->above_mode_ctx, DC_PRED, s->cols * 2);

  3823.     } else {

  3824.         memset(s->above_mode_ctx, NEARESTMV, s->cols);

  3825.     }

  3826.     memset(s->above_y_nnz_ctx, 0, s->sb_cols * 16);

  3827.     memset(s->above_uv_nnz_ctx[0], 0, s->sb_cols * 8);

  3828.     memset(s->above_uv_nnz_ctx[1], 0, s->sb_cols * 8);

  3829.     memset(s->above_segpred_ctx, 0, s->cols);

  3830.     s->pass = s->frames[CUR_FRAME].uses_2pass =

  3831.         ctx->active_thread_type == FF_THREAD_FRAME && s->refreshctx && !s->parallelmode;

  3832.     if ((res = update_block_buffers(ctx)) < 0) {

  3833.         av_log(ctx, AV_LOG_ERROR,

  3834.                "Failed to allocate block buffers\n");

  3835.         return res;

  3836.     }

  3837.     if (s->refreshctx && s->parallelmode) {

  3838.         int j, k, l, m;

  3839. 

  3840.         for (i = 0; i < 4; i++) {

  3841.             for (j = 0; j < 2; j++)

  3842.                 for (k = 0; k < 2; k++)

  3843.                     for (l = 0; l < 6; l++)

  3844.                         for (m = 0; m < 6; m++)

  3845.                             memcpy(s->prob_ctx[s->framectxid].coef[i][j][k][l][m],

  3846.                                    s->prob.coef[i][j][k][l][m], 3);

  3847.             if (s->txfmmode == i)

  3848.                 break;

  3849.         }

  3850.         s->prob_ctx[s->framectxid].p = s->prob.p;

  3851.         ff_thread_finish_setup(ctx);

  3852.     } else if (!s->refreshctx) {

  3853.         ff_thread_finish_setup(ctx);

  3854.     }

  3855. 

  3856.     do {

  3857.         yoff = uvoff = 0;

  3858.         s->b = s->b_base;

  3859.         s->block = s->block_base;

  3860.         s->uvblock[0] = s->uvblock_base[0];

  3861.         s->uvblock[1] = s->uvblock_base[1];

  3862.         s->eob = s->eob_base;

  3863.         s->uveob[0] = s->uveob_base[0];

  3864.         s->uveob[1] = s->uveob_base[1];

  3865. 

  3866.         for (tile_row = 0; tile_row < s->tiling.tile_rows; tile_row++) {

  3867.             set_tile_offset(&s->tiling.tile_row_start, &s->tiling.tile_row_end,

  3868.                             tile_row, s->tiling.log2_tile_rows, s->sb_rows);

  3869.             if (s->pass != 2) {

  3870.                 for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  3871.                     unsigned tile_size;

  3872. 

  3873.                     if (tile_col == s->tiling.tile_cols - 1 &&

  3874.                         tile_row == s->tiling.tile_rows - 1) {

  3875.                         tile_size = size;

  3876.                     } else {

  3877.                         tile_size = AV_RB32(data);

  3878.                         data += 4;

  3879.                         size -= 4;

  3880.                     }

  3881.                     if (tile_size > size) {

  3882.                         ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  3883.                         return AVERROR_INVALIDDATA;

  3884.                     }

  3885.                     ff_vp56_init_range_decoder(&s->c_b[tile_col], data, tile_size);

  3886.                     if (vp56_rac_get_prob_branchy(&s->c_b[tile_col], 128)) { // marker bit

  3887.                         ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  3888.                         return AVERROR_INVALIDDATA;

  3889.                     }

  3890.                     data += tile_size;

  3891.                     size -= tile_size;

  3892.                 }

  3893.             }

  3894. 

  3895.             for (row = s->tiling.tile_row_start; row < s->tiling.tile_row_end;

  3896.                  row += 8, yoff += ls_y * 64, uvoff += ls_uv * 32) {

  3897.                 struct VP9Filter *lflvl_ptr = s->lflvl;

  3898.                 ptrdiff_t yoff2 = yoff, uvoff2 = uvoff;

  3899. 

  3900.                 for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  3901.                     set_tile_offset(&s->tiling.tile_col_start, &s->tiling.tile_col_end,

  3902.                                     tile_col, s->tiling.log2_tile_cols, s->sb_cols);

  3903. 

  3904.                     if (s->pass != 2) {

  3905.                         memset(s->left_partition_ctx, 0, 8);

  3906.                         memset(s->left_skip_ctx, 0, 8);

  3907.                         if (s->keyframe || s->intraonly) {

  3908.                             memset(s->left_mode_ctx, DC_PRED, 16);

  3909.                         } else {

  3910.                             memset(s->left_mode_ctx, NEARESTMV, 8);

  3911.                         }

  3912.                         memset(s->left_y_nnz_ctx, 0, 16);

  3913.                         memset(s->left_uv_nnz_ctx, 0, 16);

  3914.                         memset(s->left_segpred_ctx, 0, 8);

  3915. 

  3916.                         memcpy(&s->c, &s->c_b[tile_col], sizeof(s->c));

  3917.                     }

  3918. 

  3919.                     for (col = s->tiling.tile_col_start;

  3920.                          col < s->tiling.tile_col_end;

  3921.                          col += 8, yoff2 += 64, uvoff2 += 32, lflvl_ptr++) {

  3922.                         // FIXME integrate with lf code (i.e. zero after each

  3923.                         // use, similar to invtxfm coefficients, or similar)

  3924.                         if (s->pass != 1) {

  3925.                             memset(lflvl_ptr->mask, 0, sizeof(lflvl_ptr->mask));

  3926.                         }

  3927. 

  3928.                         if (s->pass == 2) {

  3929.                             decode_sb_mem(ctx, row, col, lflvl_ptr,

  3930.                                           yoff2, uvoff2, BL_64X64);

  3931.                         } else {

  3932.                             decode_sb(ctx, row, col, lflvl_ptr,

  3933.                                       yoff2, uvoff2, BL_64X64);

  3934.                         }

  3935.                     }

  3936.                     if (s->pass != 2) {

  3937.                         memcpy(&s->c_b[tile_col], &s->c, sizeof(s->c));

  3938.                     }

  3939.                 }

  3940. 

  3941.                 if (s->pass == 1) {

  3942.                     continue;

  3943.                 }

  3944. 

  3945.                 // backup pre-loopfilter reconstruction data for intra

  3946.                 // prediction of next row of sb64s

  3947.                 if (row + 8 < s->rows) {

  3948.                     memcpy(s->intra_pred_data[0],

  3949.                            f->data[0] + yoff + 63 * ls_y,

  3950.                            8 * s->cols);

  3951.                     memcpy(s->intra_pred_data[1],

  3952.                            f->data[1] + uvoff + 31 * ls_uv,

  3953.                            4 * s->cols);

  3954.                     memcpy(s->intra_pred_data[2],

  3955.                            f->data[2] + uvoff + 31 * ls_uv,

  3956.                            4 * s->cols);

  3957.                 }

  3958. 

  3959.                 // loopfilter one row

  3960.                 if (s->filter.level) {

  3961.                     yoff2 = yoff;

  3962.                     uvoff2 = uvoff;

  3963.                     lflvl_ptr = s->lflvl;

  3964.                     for (col = 0; col < s->cols;

  3965.                          col += 8, yoff2 += 64, uvoff2 += 32, lflvl_ptr++) {

  3966.                         loopfilter_sb(ctx, lflvl_ptr, row, col, yoff2, uvoff2);

  3967.                     }

  3968.                 }

  3969. 

  3970.                 // FIXME maybe we can make this more finegrained by running the

  3971.                 // loopfilter per-block instead of after each sbrow

  3972.                 // In fact that would also make intra pred left preparation easier?

  3973.                 ff_thread_report_progress(&s->frames[CUR_FRAME].tf, row >> 3, 0);

  3974.             }

  3975.         }

  3976. 

  3977.         if (s->pass < 2 && s->refreshctx && !s->parallelmode) {

  3978.             adapt_probs(s);

  3979.             ff_thread_finish_setup(ctx);

  3980.         }

  3981.     } while (s->pass++ == 1);

  3982.     ff_thread_report_progress(&s->frames[CUR_FRAME].tf, INT_MAX, 0);

  3983. 

  3984.     // ref frame setup

  3985.     for (i = 0; i < 8; i++) {

  3986.         if (s->refs[i].f->data[0])

  3987.             ff_thread_release_buffer(ctx, &s->refs[i]);

  3988.         ff_thread_ref_frame(&s->refs[i], &s->next_refs[i]);

  3989.     }

  3990. 

  3991.     if (!s->invisible) {

  3992.         if ((res = av_frame_ref(frame, s->frames[CUR_FRAME].tf.f)) < 0)

  3993.             return res;

  3994.         *got_frame = 1;

  3995.     }

  3996. 

  3997.     return pkt->size;

  3998. }

  3999. 

  4000. static void vp9_decode_flush(AVCodecContext *ctx)

  4001. {

  4002.     VP9Context *s = ctx->priv_data;

  4003.     int i;

  4004. 

  4005.     for (i = 0; i < 3; i++)

  4006.         vp9_unref_frame(ctx, &s->frames[i]);

  4007.     for (i = 0; i < 8; i++)

  4008.         ff_thread_release_buffer(ctx, &s->refs[i]);

  4009. }

  4010. 

  4011. static int init_frames(AVCodecContext *ctx)

  4012. {

  4013.     VP9Context *s = ctx->priv_data;

  4014.     int i;

  4015. 

  4016.     for (i = 0; i < 3; i++) {

  4017.         s->frames[i].tf.f = av_frame_alloc();

  4018.         if (!s->frames[i].tf.f) {

  4019.             vp9_decode_free(ctx);

  4020.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  4021.             return AVERROR(ENOMEM);

  4022.         }

  4023.     }

  4024.     for (i = 0; i < 8; i++) {

  4025.         s->refs[i].f = av_frame_alloc();

  4026.         s->next_refs[i].f = av_frame_alloc();

  4027.         if (!s->refs[i].f || !s->next_refs[i].f) {

  4028.             vp9_decode_free(ctx);

  4029.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  4030.             return AVERROR(ENOMEM);

  4031.         }

  4032.     }

  4033. 

  4034.     return 0;

  4035. }

  4036. 

  4037. static av_cold int vp9_decode_init(AVCodecContext *ctx)

  4038. {

  4039.     VP9Context *s = ctx->priv_data;

  4040. 

  4041.     ctx->internal->allocate_progress = 1;

  4042.     ctx->pix_fmt = AV_PIX_FMT_YUV420P;

  4043.     ff_vp9dsp_init(&s->dsp);

  4044.     ff_videodsp_init(&s->vdsp, 8);

  4045.     s->filter.sharpness = -1;

  4046. 

  4047.     return init_frames(ctx);

  4048. }

  4049. 

  4050. static av_cold int vp9_decode_init_thread_copy(AVCodecContext *avctx)

  4051. {

  4052.     return init_frames(avctx);

  4053. }

  4054. 

  4055. static int vp9_decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src)

  4056. {

  4057.     int i, res;

  4058.     VP9Context *s = dst->priv_data, *ssrc = src->priv_data;

  4059. 

  4060.     // detect size changes in other threads

  4061.     if (s->intra_pred_data[0] &&

  4062.         (!ssrc->intra_pred_data[0] || s->cols != ssrc->cols || s->rows != ssrc->rows)) {

  4063.         free_buffers(s);

  4064.     }

  4065. 

  4066.     for (i = 0; i < 3; i++) {

  4067.         if (s->frames[i].tf.f->data[0])

  4068.             vp9_unref_frame(dst, &s->frames[i]);

  4069.         if (ssrc->frames[i].tf.f->data[0]) {

  4070.             if ((res = vp9_ref_frame(dst, &s->frames[i], &ssrc->frames[i])) < 0)

  4071.                 return res;

  4072.         }

  4073.     }

  4074.     for (i = 0; i < 8; i++) {

  4075.         if (s->refs[i].f->data[0])

  4076.             ff_thread_release_buffer(dst, &s->refs[i]);

  4077.         if (ssrc->next_refs[i].f->data[0]) {

  4078.             if ((res = ff_thread_ref_frame(&s->refs[i], &ssrc->next_refs[i])) < 0)

  4079.                 return res;

  4080.         }

  4081.     }

  4082. 

  4083.     s->invisible = ssrc->invisible;

  4084.     s->keyframe = ssrc->keyframe;

  4085.     s->segmentation.enabled = ssrc->segmentation.enabled;

  4086.     s->segmentation.update_map = ssrc->segmentation.update_map;

  4087.     memcpy(&s->prob_ctx, &ssrc->prob_ctx, sizeof(s->prob_ctx));

  4088.     memcpy(&s->lf_delta, &ssrc->lf_delta, sizeof(s->lf_delta));

  4089.     if (ssrc->segmentation.enabled) {

  4090.         memcpy(&s->segmentation.feat, &ssrc->segmentation.feat,

  4091.                sizeof(s->segmentation.feat));

  4092.     }

  4093. 

  4094.     return 0;

  4095. }

  4096. 

  4097. AVCodec ff_vp9_decoder = {

  4098.     .name                  = "vp9",

  4099.     .long_name             = NULL_IF_CONFIG_SMALL("Google VP9"),

  4100.     .type                  = AVMEDIA_TYPE_VIDEO,

  4101.     .id                    = AV_CODEC_ID_VP9,

  4102.     .priv_data_size        = sizeof(VP9Context),

  4103.     .init                  = vp9_decode_init,

  4104.     .close                 = vp9_decode_free,

  4105.     .decode                = vp9_decode_frame,

  4106.     .capabilities          = CODEC_CAP_DR1 | CODEC_CAP_FRAME_THREADS,

  4107.     .flush                 = vp9_decode_flush,

  4108.     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(vp9_decode_init_thread_copy),

  4109.     .update_thread_context = ONLY_IF_THREADS_ENABLED(vp9_decode_update_thread_context),

  4110. };