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.
58680 Commits
32 Branches
283MB
Tree: 46955ae430
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from '46955ae430'
${ noResults }
FFmpeg/libavcodec/vp9.c

3636 lines
144KB
Raw Blame History 

  1. /*

  2.  * VP9 compatible video decoder

  3.  *

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

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

  6.  *

  7.  * This file is part of FFmpeg.

  8.  *

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

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

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

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

  13.  *

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

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

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

  17.  * Lesser General Public License for more details.

  18.  *

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

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

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

  22.  */

  23. 

  24. #include "avcodec.h"

  25. #include "get_bits.h"

  26. #include "internal.h"

  27. #include "thread.h"

  28. #include "videodsp.h"

  29. #include "vp56.h"

  30. #include "vp9.h"

  31. #include "vp9data.h"

  32. #include "vp9dsp.h"

  33. #include "libavutil/avassert.h"

  34. 

  35. #define VP9_SYNCCODE 0x498342

  36. 

  37. enum CompPredMode {

  38.     PRED_SINGLEREF,

  39.     PRED_COMPREF,

  40.     PRED_SWITCHABLE,

  41. };

  42. 

  43. enum BlockLevel {

  44.     BL_64X64,

  45.     BL_32X32,

  46.     BL_16X16,

  47.     BL_8X8,

  48. };

  49. 

  50. enum BlockSize {

  51.     BS_64x64,

  52.     BS_64x32,

  53.     BS_32x64,

  54.     BS_32x32,

  55.     BS_32x16,

  56.     BS_16x32,

  57.     BS_16x16,

  58.     BS_16x8,

  59.     BS_8x16,

  60.     BS_8x8,

  61.     BS_8x4,

  62.     BS_4x8,

  63.     BS_4x4,

  64.     N_BS_SIZES,

  65. };

  66. 

  67. struct VP9mvrefPair {

  68.     VP56mv mv[2];

  69.     int8_t ref[2];

  70. };

  71. 

  72. typedef struct VP9Frame {

  73.     ThreadFrame tf;

  74.     AVBufferRef *extradata;

  75.     uint8_t *segmentation_map;

  76.     struct VP9mvrefPair *mv;

  77. } VP9Frame;

  78. 

  79. struct VP9Filter {

  80.     uint8_t level[8 * 8];

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

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

  83. };

  84. 

  85. typedef struct VP9Block {

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

  87.     enum FilterMode filter;

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

  89.     enum BlockSize bs;

  90.     enum TxfmMode tx, uvtx;

  91. 

  92. } VP9Block;

  93. 

  94. typedef struct VP9Context {

  95.     VP9DSPContext dsp;

  96.     VideoDSPContext vdsp;

  97.     GetBitContext gb;

  98.     VP56RangeCoder c;

  99.     VP56RangeCoder *c_b;

  100.     unsigned c_b_size;

  101.     VP9Block *b_base, *b;

  102.     int row, row7, col, col7;

  103.     uint8_t *dst[3];

  104.     ptrdiff_t y_stride, uv_stride;

  105. 

  106.     // bitstream header

  107.     uint8_t profile;

  108.     uint8_t keyframe, last_keyframe;

  109.     uint8_t invisible;

  110.     uint8_t use_last_frame_mvs;

  111.     uint8_t errorres;

  112.     uint8_t colorspace;

  113.     uint8_t fullrange;

  114.     uint8_t intraonly;

  115.     uint8_t resetctx;

  116.     uint8_t refreshrefmask;

  117.     uint8_t highprecisionmvs;

  118.     enum FilterMode filtermode;

  119.     uint8_t allowcompinter;

  120.     uint8_t fixcompref;

  121.     uint8_t refreshctx;

  122.     uint8_t parallelmode;

  123.     uint8_t framectxid;

  124.     uint8_t refidx[3];

  125.     uint8_t signbias[3];

  126.     uint8_t varcompref[2];

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

  128. #define CUR_FRAME 0

  129. #define LAST_FRAME 1

  130.     VP9Frame frames[2];

  131. 

  132.     struct {

  133.         uint8_t level;

  134.         int8_t sharpness;

  135.         uint8_t lim_lut[64];

  136.         uint8_t mblim_lut[64];

  137.     } filter;

  138.     struct {

  139.         uint8_t enabled;

  140.         int8_t mode[2];

  141.         int8_t ref[4];

  142.     } lf_delta;

  143.     uint8_t yac_qi;

  144.     int8_t ydc_qdelta, uvdc_qdelta, uvac_qdelta;

  145.     uint8_t lossless;

  146.     struct {

  147.         uint8_t enabled;

  148.         uint8_t temporal;

  149.         uint8_t absolute_vals;

  150.         uint8_t update_map;

  151.         struct {

  152.             uint8_t q_enabled;

  153.             uint8_t lf_enabled;

  154.             uint8_t ref_enabled;

  155.             uint8_t skip_enabled;

  156.             uint8_t ref_val;

  157.             int16_t q_val;

  158.             int8_t lf_val;

  159.             int16_t qmul[2][2];

  160.             uint8_t lflvl[4][2];

  161.         } feat[8];

  162.     } segmentation;

  163.     struct {

  164.         unsigned log2_tile_cols, log2_tile_rows;

  165.         unsigned tile_cols, tile_rows;

  166.         unsigned tile_row_start, tile_row_end, tile_col_start, tile_col_end;

  167.     } tiling;

  168.     unsigned sb_cols, sb_rows, rows, cols;

  169.     struct {

  170.         prob_context p;

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

  172.     } prob_ctx[4];

  173.     struct {

  174.         prob_context p;

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

  176.         uint8_t seg[7];

  177.         uint8_t segpred[3];

  178.     } prob;

  179.     struct {

  180.         unsigned y_mode[4][10];

  181.         unsigned uv_mode[10][10];

  182.         unsigned filter[4][3];

  183.         unsigned mv_mode[7][4];

  184.         unsigned intra[4][2];

  185.         unsigned comp[5][2];

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

  187.         unsigned comp_ref[5][2];

  188.         unsigned tx32p[2][4];

  189.         unsigned tx16p[2][3];

  190.         unsigned tx8p[2][2];

  191.         unsigned skip[3][2];

  192.         unsigned mv_joint[4];

  193.         struct {

  194.             unsigned sign[2];

  195.             unsigned classes[11];

  196.             unsigned class0[2];

  197.             unsigned bits[10][2];

  198.             unsigned class0_fp[2][4];

  199.             unsigned fp[4];

  200.             unsigned class0_hp[2];

  201.             unsigned hp[2];

  202.         } mv_comp[2];

  203.         unsigned partition[4][4][4];

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

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

  206.     } counts;

  207.     enum TxfmMode txfmmode;

  208.     enum CompPredMode comppredmode;

  209. 

  210.     // contextual (left/above) cache

  211.     uint8_t left_partition_ctx[8], *above_partition_ctx;

  212.     uint8_t left_mode_ctx[16], *above_mode_ctx;

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

  214.     uint8_t left_y_nnz_ctx[16], *above_y_nnz_ctx;

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

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

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

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

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

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

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

  222.     uint8_t left_filter_ctx[8], *above_filter_ctx;

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

  224. 

  225.     // whole-frame cache

  226.     uint8_t *intra_pred_data[3];

  227.     struct VP9Filter *lflvl;

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

  229. 

  230.     // block reconstruction intermediates

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

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

  233.     VP56mv min_mv, max_mv;

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

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

  236. } VP9Context;

  237. 

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

  239.     {

  240.         { 16, 16 }, { 16, 8 }, { 8, 16 }, { 8, 8 }, { 8, 4 }, { 4, 8 },

  241.         { 4, 4 }, { 4, 2 }, { 2, 4 }, { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 },

  242.     }, {

  243.         { 8, 8 }, { 8, 4 }, { 4, 8 }, { 4, 4 }, { 4, 2 }, { 2, 4 },

  244.         { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 }, { 1, 1 }, { 1, 1 }, { 1, 1 },

  245.     }

  246. };

  247. 

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

  249. {

  250.     VP9Context *s = ctx->priv_data;

  251.     int ret, sz;

  252. 

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

  254.         return ret;

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

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

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

  258.         return AVERROR(ENOMEM);

  259.     }

  260. 

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

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

  263. 

  264.     return 0;

  265. }

  266. 

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

  268. {

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

  270.     av_buffer_unref(&f->extradata);

  271. }

  272. 

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

  274. {

  275.     int res;

  276. 

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

  278.         return res;

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

  280.         vp9_unref_frame(ctx, dst);

  281.         return AVERROR(ENOMEM);

  282.     }

  283. 

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

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

  286. 

  287.     return 0;

  288. }

  289. 

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

  291. {

  292.     VP9Context *s = ctx->priv_data;

  293.     uint8_t *p;

  294. 

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

  296. 

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

  298.         return 0;

  299. 

  300.     ctx->width  = w;

  301.     ctx->height = h;

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

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

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

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

  306. 

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

  308.     av_freep(&s->above_partition_ctx);

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

  310.     if (!p)

  311.         return AVERROR(ENOMEM);

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  329. #undef assign

  330. 

  331.     av_free(s->b_base);

  332.     av_free(s->block_base);

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

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

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

  336.         return AVERROR(ENOMEM);

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

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

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

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

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

  342. 

  343.     return 0;

  344. }

  345. 

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

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

  348. {

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

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

  351. }

  352. 

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

  354. {

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

  356. }

  357. 

  358. // differential forward probability updates

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

  360. {

  361.     static const int inv_map_table[254] = {

  362.           7,  20,  33,  46,  59,  72,  85,  98, 111, 124, 137, 150, 163, 176,

  363.         189, 202, 215, 228, 241, 254,   1,   2,   3,   4,   5,   6,   8,   9,

  364.          10,  11,  12,  13,  14,  15,  16,  17,  18,  19,  21,  22,  23,  24,

  365.          25,  26,  27,  28,  29,  30,  31,  32,  34,  35,  36,  37,  38,  39,

  366.          40,  41,  42,  43,  44,  45,  47,  48,  49,  50,  51,  52,  53,  54,

  367.          55,  56,  57,  58,  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,

  368.          70,  71,  73,  74,  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,

  369.          86,  87,  88,  89,  90,  91,  92,  93,  94,  95,  96,  97,  99, 100,

  370.         101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 112, 113, 114, 115,

  371.         116, 117, 118, 119, 120, 121, 122, 123, 125, 126, 127, 128, 129, 130,

  372.         131, 132, 133, 134, 135, 136, 138, 139, 140, 141, 142, 143, 144, 145,

  373.         146, 147, 148, 149, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160,

  374.         161, 162, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175,

  375.         177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 190, 191,

  376.         192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 204, 205, 206,

  377.         207, 208, 209, 210, 211, 212, 213, 214, 216, 217, 218, 219, 220, 221,

  378.         222, 223, 224, 225, 226, 227, 229, 230, 231, 232, 233, 234, 235, 236,

  379.         237, 238, 239, 240, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251,

  380.         252, 253,

  381.     };

  382.     int d;

  383. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  398. 

  399.     if (!vp8_rac_get(c)) {

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

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

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

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

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

  405.     } else {

  406.         d = vp8_rac_get_uint(c, 7);

  407.         if (d >= 65)

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

  409.         d += 64;

  410.     }

  411. 

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

  413.                     255 - inv_recenter_nonneg(inv_map_table[d], 255 - p);

  414. }

  415. 

  416. static int decode_frame_header(AVCodecContext *ctx,

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

  418. {

  419.     VP9Context *s = ctx->priv_data;

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

  421.     int last_invisible;

  422.     const uint8_t *data2;

  423. 

  424.     /* general header */

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

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

  427.         return res;

  428.     }

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

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

  431.         return AVERROR_INVALIDDATA;

  432.     }

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

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

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

  436.         return AVERROR_INVALIDDATA;

  437.     }

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

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

  440.         return 0;

  441.     }

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

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

  444.     last_invisible    = s->invisible;

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

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

  447.     // FIXME disable this upon resolution change

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

  449.     if (s->keyframe) {

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

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

  452.             return AVERROR_INVALIDDATA;

  453.         }

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

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

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

  457.             return AVERROR_INVALIDDATA;

  458.         }

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

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

  461.         s->refreshrefmask = 0xff;

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

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

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

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

  466.     } else {

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

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

  469.         if (s->intraonly) {

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

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

  472.                 return AVERROR_INVALIDDATA;

  473.             }

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

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

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

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

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

  479.         } else {

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

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

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

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

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

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

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

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

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

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

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

  491.                 return AVERROR_INVALIDDATA;

  492.             }

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

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

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

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

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

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

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

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

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

  502.             } else {

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

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

  505.             }

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

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

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

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

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

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

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

  513.             if (s->allowcompinter) {

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

  515.                     s->fixcompref    = 2;

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

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

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

  519.                     s->fixcompref    = 1;

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

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

  522.                 } else {

  523.                     s->fixcompref    = 0;

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

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

  526.                 }

  527.             }

  528.         }

  529.     }

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

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

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

  533. 

  534.     /* loopfilter header data */

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

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

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

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

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

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

  541.     s->filter.sharpness = sharp;

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

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

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

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

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

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

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

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

  550.         }

  551.     } else {

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

  553.     }

  554. 

  555.     /* quantization header data */

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

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

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

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

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

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

  562. 

  563.     /* segmentation header info */

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

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

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

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

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

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

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

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

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

  573.         }

  574. 

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

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

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

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

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

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

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

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

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

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

  585.             }

  586.         }

  587.     } else {

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

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

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

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

  592.     }

  593. 

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

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

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

  597. 

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

  599.             if (s->segmentation.absolute_vals)

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

  601.             else

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

  603.         } else {

  604.             qyac  = s->yac_qi;

  605.         }

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

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

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

  609.         qyac  = av_clip_uintp2(qyac, 8);

  610. 

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

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

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

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

  615. 

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

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

  618.             if (s->segmentation.absolute_vals)

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

  620.             else

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

  622.         } else {

  623.             lflvl  = s->filter.level;

  624.         }

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

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

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

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

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

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

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

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

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

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

  635.         }

  636.     }

  637. 

  638.     /* tiling info */

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

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

  641.         return res;

  642.     }

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

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

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

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

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

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

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

  650.             s->tiling.log2_tile_cols++;

  651.         else

  652.             break;

  653.     }

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

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

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

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

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

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

  660.         if (!s->c_b) {

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

  662.             return AVERROR(ENOMEM);

  663.         }

  664.     }

  665. 

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

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

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

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

  670.                sizeof(vp9_default_coef_probs));

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

  672.                sizeof(vp9_default_coef_probs));

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

  674.                sizeof(vp9_default_coef_probs));

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

  676.                sizeof(vp9_default_coef_probs));

  677.     }

  678. 

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

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

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

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

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

  684.         return AVERROR_INVALIDDATA;

  685.     }

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

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

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

  689.         return AVERROR_INVALIDDATA;

  690.     }

  691. 

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

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

  694.     } else {

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

  696.     }

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

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

  699.     // fw update)?

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

  701. 

  702.     // txfm updates

  703.     if (s->lossless) {

  704.         s->txfmmode = TX_4X4;

  705.     } else {

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

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

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

  709. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  724.         }

  725.     }

  726. 

  727.     // coef updates

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

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

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

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

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

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

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

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

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

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

  738.                                 break;

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

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

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

  742.                                 } else {

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

  744.                                 }

  745.                             }

  746.                             p[3] = 0;

  747.                         }

  748.         } else {

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

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

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

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

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

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

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

  756.                                 break;

  757.                             memcpy(p, r, 3);

  758.                             p[3] = 0;

  759.                         }

  760.         }

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

  762.             break;

  763.     }

  764. 

  765.     // mode updates

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

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

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

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

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

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

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

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

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

  775. 

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

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

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

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

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

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

  782. 

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

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

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

  786. 

  787.         if (s->allowcompinter) {

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

  789.             if (s->comppredmode)

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

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

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

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

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

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

  796.         } else {

  797.             s->comppredmode = PRED_SINGLEREF;

  798.         }

  799. 

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

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

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

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

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

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

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

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

  808.             }

  809.         }

  810. 

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

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

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

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

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

  816.         }

  817. 

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

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

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

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

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

  823. 

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

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

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

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

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

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

  830. 

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

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

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

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

  835. 

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

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

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

  839. 

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

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

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

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

  844. 

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

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

  847. 

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

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

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

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

  852.         }

  853. 

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

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

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

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

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

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

  860. 

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

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

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

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

  865.         }

  866. 

  867.         if (s->highprecisionmvs) {

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

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

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

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

  872. 

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

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

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

  876.             }

  877.         }

  878.     }

  879. 

  880.     return (data2 - data) + size2;

  881. }

  882. 

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

  884.                                       VP9Context *s)

  885. {

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

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

  888. }

  889. 

  890. static void find_ref_mvs(VP9Context *s,

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

  892. {

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

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

  895.                       { -1, -1 }, {  0, -1 }, { -1,  0 }, {  6, -1 }},

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

  897.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, {  2, -1 }},

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

  899.                       { -1, -1 }, { -3,  0 }, {  0, -3 }, { -1,  2 }},

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

  901.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  903.                       { -1,  1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  905.                       {  1, -1 }, { -3,  0 }, {  0, -3 }, { -3, -3 }},

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

  907.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  909.                       {  0, -2 }, { -2,  0 }, { -2, -1 }, { -1, -2 }},

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

  911.                       { -2,  0 }, {  0, -2 }, { -1, -2 }, { -2, -1 }},

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

  913.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

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

  915.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

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

  917.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

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

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

  920.     };

  921.     VP9Block *b = s->b;

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

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

  924. #define INVALID_MV 0x80008000U

  925.     uint32_t mem = INVALID_MV;

  926.     int i;

  927. 

  928. #define RETURN_DIRECT_MV(mv) \

  929.     do { \

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

  931.         if (!idx) { \

  932.             AV_WN32A(pmv, m); \

  933.             return; \

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

  935.             mem = m; \

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

  937.             AV_WN32A(pmv, m); \

  938.             return; \

  939.         } \

  940.     } while (0)

  941. 

  942.     if (sb >= 0) {

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

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

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

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

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

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

  949.         }

  950. 

  951. #define RETURN_MV(mv) \

  952.     do { \

  953.         if (sb > 0) { \

  954.             VP56mv tmp; \

  955.             uint32_t m; \

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

  957.             m = AV_RN32A(&tmp); \

  958.             if (!idx) { \

  959.                 AV_WN32A(pmv, m); \

  960.                 return; \

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

  962.                 mem = m; \

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

  964.                 AV_WN32A(pmv, m); \

  965.                 return; \

  966.             } \

  967.         } else { \

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

  969.             if (!idx) { \

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

  971.                 return; \

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

  973.                 mem = m; \

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

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

  976.                 return; \

  977.             } \

  978.         } \

  979.     } while (0)

  980. 

  981.         if (row > 0) {

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

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

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

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

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

  987.             }

  988.         }

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

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

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

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

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

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

  995.             }

  996.         }

  997.         i = 2;

  998.     } else {

  999.         i = 0;

  1000.     }

  1001. 

  1002.     // previously coded MVs in this neighbourhood, using same reference frame

  1003.     for (; i < 8; i++) {

  1004.         int c = p[i][0] + col, r = p[i][1] + row;

  1005. 

  1006.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1007.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1008. 

  1009.             if (mv->ref[0] == ref) {

  1010.                 RETURN_MV(mv->mv[0]);

  1011.             } else if (mv->ref[1] == ref) {

  1012.                 RETURN_MV(mv->mv[1]);

  1013.             }

  1014.         }

  1015.     }

  1016. 

  1017.     // MV at this position in previous frame, using same reference frame

  1018.     if (s->use_last_frame_mvs) {

  1019.         struct VP9mvrefPair *mv = &s->frames[LAST_FRAME].mv[row * s->sb_cols * 8 + col];

  1020. 

  1021.         if (mv->ref[0] == ref) {

  1022.             RETURN_MV(mv->mv[0]);

  1023.         } else if (mv->ref[1] == ref) {

  1024.             RETURN_MV(mv->mv[1]);

  1025.         }

  1026.     }

  1027. 

  1028. #define RETURN_SCALE_MV(mv, scale) \

  1029.     do { \

  1030.         if (scale) { \

  1031.             VP56mv mv_temp = { -mv.x, -mv.y }; \

  1032.             RETURN_MV(mv_temp); \

  1033.         } else { \

  1034.             RETURN_MV(mv); \

  1035.         } \

  1036.     } while (0)

  1037. 

  1038.     // previously coded MVs in this neighbourhood, using different reference frame

  1039.     for (i = 0; i < 8; i++) {

  1040.         int c = p[i][0] + col, r = p[i][1] + row;

  1041. 

  1042.         if (c >= s->tiling.tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1043.             struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1044. 

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

  1046.                 RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1047.             }

  1048.             if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1049.                 // BUG - libvpx has this condition regardless of whether

  1050.                 // we used the first ref MV and pre-scaling

  1051.                 AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1052.                 RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1053.             }

  1054.         }

  1055.     }

  1056. 

  1057.     // MV at this position in previous frame, using different reference frame

  1058.     if (s->use_last_frame_mvs) {

  1059.         struct VP9mvrefPair *mv = &s->frames[LAST_FRAME].mv[row * s->sb_cols * 8 + col];

  1060. 

  1061.         if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1062.             RETURN_SCALE_MV(mv->mv[0], s->signbias[mv->ref[0]] != s->signbias[ref]);

  1063.         }

  1064.         if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1065.             // BUG - libvpx has this condition regardless of whether

  1066.             // we used the first ref MV and pre-scaling

  1067.             AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1068.             RETURN_SCALE_MV(mv->mv[1], s->signbias[mv->ref[1]] != s->signbias[ref]);

  1069.         }

  1070.     }

  1071. 

  1072.     AV_ZERO32(pmv);

  1073. #undef INVALID_MV

  1074. #undef RETURN_MV

  1075. #undef RETURN_SCALE_MV

  1076. }

  1077. 

  1078. static av_always_inline int read_mv_component(VP9Context *s, int idx, int hp)

  1079. {

  1080.     int bit, sign = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].sign);

  1081.     int n, c = vp8_rac_get_tree(&s->c, vp9_mv_class_tree,

  1082.                                 s->prob.p.mv_comp[idx].classes);

  1083. 

  1084.     s->counts.mv_comp[idx].sign[sign]++;

  1085.     s->counts.mv_comp[idx].classes[c]++;

  1086.     if (c) {

  1087.         int m;

  1088. 

  1089.         for (n = 0, m = 0; m < c; m++) {

  1090.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].bits[m]);

  1091.             n |= bit << m;

  1092.             s->counts.mv_comp[idx].bits[m][bit]++;

  1093.         }

  1094.         n <<= 3;

  1095.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree, s->prob.p.mv_comp[idx].fp);

  1096.         n |= bit << 1;

  1097.         s->counts.mv_comp[idx].fp[bit]++;

  1098.         if (hp) {

  1099.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].hp);

  1100.             s->counts.mv_comp[idx].hp[bit]++;

  1101.             n |= bit;

  1102.         } else {

  1103.             n |= 1;

  1104.             // bug in libvpx - we count for bw entropy purposes even if the

  1105.             // bit wasn't coded

  1106.             s->counts.mv_comp[idx].hp[1]++;

  1107.         }

  1108.         n += 8 << c;

  1109.     } else {

  1110.         n = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0);

  1111.         s->counts.mv_comp[idx].class0[n]++;

  1112.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree,

  1113.                                s->prob.p.mv_comp[idx].class0_fp[n]);

  1114.         s->counts.mv_comp[idx].class0_fp[n][bit]++;

  1115.         n = (n << 3) | (bit << 1);

  1116.         if (hp) {

  1117.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0_hp);

  1118.             s->counts.mv_comp[idx].class0_hp[bit]++;

  1119.             n |= bit;

  1120.         } else {

  1121.             n |= 1;

  1122.             // bug in libvpx - we count for bw entropy purposes even if the

  1123.             // bit wasn't coded

  1124.             s->counts.mv_comp[idx].class0_hp[1]++;

  1125.         }

  1126.     }

  1127. 

  1128.     return sign ? -(n + 1) : (n + 1);

  1129. }

  1130. 

  1131. static void fill_mv(VP9Context *s,

  1132.                     VP56mv *mv, int mode, int sb)

  1133. {

  1134.     VP9Block *b = s->b;

  1135. 

  1136.     if (mode == ZEROMV) {

  1137.         memset(mv, 0, sizeof(*mv) * 2);

  1138.     } else {

  1139.         int hp;

  1140. 

  1141.         // FIXME cache this value and reuse for other subblocks

  1142.         find_ref_mvs(s, &mv[0], b->ref[0], 0, mode == NEARMV,

  1143.                      mode == NEWMV ? -1 : sb);

  1144.         // FIXME maybe move this code into find_ref_mvs()

  1145.         if ((mode == NEWMV || sb == -1) &&

  1146.             !(hp = s->highprecisionmvs && abs(mv[0].x) < 64 && abs(mv[0].y) < 64)) {

  1147.             if (mv[0].y & 1) {

  1148.                 if (mv[0].y < 0)

  1149.                     mv[0].y++;

  1150.                 else

  1151.                     mv[0].y--;

  1152.             }

  1153.             if (mv[0].x & 1) {

  1154.                 if (mv[0].x < 0)

  1155.                     mv[0].x++;

  1156.                 else

  1157.                     mv[0].x--;

  1158.             }

  1159.         }

  1160.         if (mode == NEWMV) {

  1161.             enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1162.                                               s->prob.p.mv_joint);

  1163. 

  1164.             s->counts.mv_joint[j]++;

  1165.             if (j >= MV_JOINT_V)

  1166.                 mv[0].y += read_mv_component(s, 0, hp);

  1167.             if (j & 1)

  1168.                 mv[0].x += read_mv_component(s, 1, hp);

  1169.         }

  1170. 

  1171.         if (b->comp) {

  1172.             // FIXME cache this value and reuse for other subblocks

  1173.             find_ref_mvs(s, &mv[1], b->ref[1], 1, mode == NEARMV,

  1174.                          mode == NEWMV ? -1 : sb);

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

  1176.                 !(hp = s->highprecisionmvs && abs(mv[1].x) < 64 && abs(mv[1].y) < 64)) {

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

  1178.                     if (mv[1].y < 0)

  1179.                         mv[1].y++;

  1180.                     else

  1181.                         mv[1].y--;

  1182.                 }

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

  1184.                     if (mv[1].x < 0)

  1185.                         mv[1].x++;

  1186.                     else

  1187.                         mv[1].x--;

  1188.                 }

  1189.             }

  1190.             if (mode == NEWMV) {

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

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

  1193. 

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

  1195.                 if (j >= MV_JOINT_V)

  1196.                     mv[1].y += read_mv_component(s, 0, hp);

  1197.                 if (j & 1)

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

  1199.             }

  1200.         }

  1201.     }

  1202. }

  1203. 

  1204. static void decode_mode(AVCodecContext *ctx)

  1205. {

  1206.     static const uint8_t left_ctx[N_BS_SIZES] = {

  1207.         0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf

  1208.     };

  1209.     static const uint8_t above_ctx[N_BS_SIZES] = {

  1210.         0x0, 0x0, 0x8, 0x8, 0x8, 0xc, 0xc, 0xc, 0xe, 0xe, 0xe, 0xf, 0xf

  1211.     };

  1212.     static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {

  1213.         TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,

  1214.         TX_16X16, TX_8X8, TX_8X8, TX_8X8, TX_4X4, TX_4X4, TX_4X4

  1215.     };

  1216.     VP9Context *s = ctx->priv_data;

  1217.     VP9Block *b = s->b;

  1218.     int row = s->row, col = s->col, row7 = s->row7;

  1219.     enum TxfmMode max_tx = max_tx_for_bl_bp[b->bs];

  1220.     int w4 = FFMIN(s->cols - col, bwh_tab[1][b->bs][0]);

  1221.     int h4 = FFMIN(s->rows - row, bwh_tab[1][b->bs][1]), y;

  1222.     int have_a = row > 0, have_l = col > s->tiling.tile_col_start;

  1223. 

  1224.     if (!s->segmentation.enabled) {

  1225.         b->seg_id = 0;

  1226.     } else if (s->keyframe || s->intraonly) {

  1227.         b->seg_id = s->segmentation.update_map ?

  1228.             vp8_rac_get_tree(&s->c, vp9_segmentation_tree, s->prob.seg) : 0;

  1229.     } else if (!s->segmentation.update_map ||

  1230.                (s->segmentation.temporal &&

  1231.                 vp56_rac_get_prob_branchy(&s->c,

  1232.                     s->prob.segpred[s->above_segpred_ctx[col] +

  1233.                                     s->left_segpred_ctx[row7]]))) {

  1234.         int pred = 8, x;

  1235.         uint8_t *refsegmap = s->frames[LAST_FRAME].segmentation_map;

  1236. 

  1237.         for (y = 0; y < h4; y++)

  1238.             for (x = 0; x < w4; x++)

  1239.                 pred = FFMIN(pred, refsegmap[(y + row) * 8 * s->sb_cols + x + col]);

  1240.         av_assert1(pred < 8);

  1241.         b->seg_id = pred;

  1242. 

  1243.         memset(&s->above_segpred_ctx[col], 1, w4);

  1244.         memset(&s->left_segpred_ctx[row7], 1, h4);

  1245.     } else {

  1246.         b->seg_id = vp8_rac_get_tree(&s->c, vp9_segmentation_tree,

  1247.                                      s->prob.seg);

  1248. 

  1249.         memset(&s->above_segpred_ctx[col], 0, w4);

  1250.         memset(&s->left_segpred_ctx[row7], 0, h4);

  1251.     }

  1252.     if ((s->segmentation.enabled && s->segmentation.update_map) || s->keyframe) {

  1253.         uint8_t *segmap = s->frames[CUR_FRAME].segmentation_map;

  1254. 

  1255.         for (y = 0; y < h4; y++)

  1256.             memset(&segmap[(y + row) * 8 * s->sb_cols + col], b->seg_id, w4);

  1257.     }

  1258. 

  1259.     b->skip = s->segmentation.enabled &&

  1260.         s->segmentation.feat[b->seg_id].skip_enabled;

  1261.     if (!b->skip) {

  1262.         int c = s->left_skip_ctx[row7] + s->above_skip_ctx[col];

  1263.         b->skip = vp56_rac_get_prob(&s->c, s->prob.p.skip[c]);

  1264.         s->counts.skip[c][b->skip]++;

  1265.     }

  1266. 

  1267.     if (s->keyframe || s->intraonly) {

  1268.         b->intra = 1;

  1269.     } else if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1270.         b->intra = !s->segmentation.feat[b->seg_id].ref_val;

  1271.     } else {

  1272.         int c, bit;

  1273. 

  1274.         if (have_a && have_l) {

  1275.             c = s->above_intra_ctx[col] + s->left_intra_ctx[row7];

  1276.             c += (c == 2);

  1277.         } else {

  1278.             c = have_a ? 2 * s->above_intra_ctx[col] :

  1279.                 have_l ? 2 * s->left_intra_ctx[row7] : 0;

  1280.         }

  1281.         bit = vp56_rac_get_prob(&s->c, s->prob.p.intra[c]);

  1282.         s->counts.intra[c][bit]++;

  1283.         b->intra = !bit;

  1284.     }

  1285. 

  1286.     if ((b->intra || !b->skip) && s->txfmmode == TX_SWITCHABLE) {

  1287.         int c;

  1288.         if (have_a) {

  1289.             if (have_l) {

  1290.                 c = (s->above_skip_ctx[col] ? max_tx :

  1291.                      s->above_txfm_ctx[col]) +

  1292.                     (s->left_skip_ctx[row7] ? max_tx :

  1293.                      s->left_txfm_ctx[row7]) > max_tx;

  1294.             } else {

  1295.                 c = s->above_skip_ctx[col] ? 1 :

  1296.                     (s->above_txfm_ctx[col] * 2 > max_tx);

  1297.             }

  1298.         } else if (have_l) {

  1299.             c = s->left_skip_ctx[row7] ? 1 :

  1300.                 (s->left_txfm_ctx[row7] * 2 > max_tx);

  1301.         } else {

  1302.             c = 1;

  1303.         }

  1304.         switch (max_tx) {

  1305.         case TX_32X32:

  1306.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][0]);

  1307.             if (b->tx) {

  1308.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][1]);

  1309.                 if (b->tx == 2)

  1310.                     b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][2]);

  1311.             }

  1312.             s->counts.tx32p[c][b->tx]++;

  1313.             break;

  1314.         case TX_16X16:

  1315.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][0]);

  1316.             if (b->tx)

  1317.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][1]);

  1318.             s->counts.tx16p[c][b->tx]++;

  1319.             break;

  1320.         case TX_8X8:

  1321.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx8p[c]);

  1322.             s->counts.tx8p[c][b->tx]++;

  1323.             break;

  1324.         case TX_4X4:

  1325.             b->tx = TX_4X4;

  1326.             break;

  1327.         }

  1328.     } else {

  1329.         b->tx = FFMIN(max_tx, s->txfmmode);

  1330.     }

  1331. 

  1332.     if (s->keyframe || s->intraonly) {

  1333.         uint8_t *a = &s->above_mode_ctx[col * 2];

  1334.         uint8_t *l = &s->left_mode_ctx[(row7) << 1];

  1335. 

  1336.         b->comp = 0;

  1337.         if (b->bs > BS_8x8) {

  1338.             // FIXME the memory storage intermediates here aren't really

  1339.             // necessary, they're just there to make the code slightly

  1340.             // simpler for now

  1341.             b->mode[0] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1342.                                     vp9_default_kf_ymode_probs[a[0]][l[0]]);

  1343.             if (b->bs != BS_8x4) {

  1344.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1345.                                  vp9_default_kf_ymode_probs[a[1]][b->mode[0]]);

  1346.                 l[0] = a[1] = b->mode[1];

  1347.             } else {

  1348.                 l[0] = a[1] = b->mode[1] = b->mode[0];

  1349.             }

  1350.             if (b->bs != BS_4x8) {

  1351.                 b->mode[2] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1352.                                         vp9_default_kf_ymode_probs[a[0]][l[1]]);

  1353.                 if (b->bs != BS_8x4) {

  1354.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1355.                                   vp9_default_kf_ymode_probs[a[1]][b->mode[2]]);

  1356.                     l[1] = a[1] = b->mode[3];

  1357.                 } else {

  1358.                     l[1] = a[1] = b->mode[3] = b->mode[2];

  1359.                 }

  1360.             } else {

  1361.                 b->mode[2] = b->mode[0];

  1362.                 l[1] = a[1] = b->mode[3] = b->mode[1];

  1363.             }

  1364.         } else {

  1365.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1366.                                           vp9_default_kf_ymode_probs[*a][*l]);

  1367.             b->mode[3] = b->mode[2] = b->mode[1] = b->mode[0];

  1368.             // FIXME this can probably be optimized

  1369.             memset(a, b->mode[0], bwh_tab[0][b->bs][0]);

  1370.             memset(l, b->mode[0], bwh_tab[0][b->bs][1]);

  1371.         }

  1372.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1373.                                      vp9_default_kf_uvmode_probs[b->mode[3]]);

  1374.     } else if (b->intra) {

  1375.         b->comp = 0;

  1376.         if (b->bs > BS_8x8) {

  1377.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1378.                                           s->prob.p.y_mode[0]);

  1379.             s->counts.y_mode[0][b->mode[0]]++;

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

  1381.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1382.                                               s->prob.p.y_mode[0]);

  1383.                 s->counts.y_mode[0][b->mode[1]]++;

  1384.             } else {

  1385.                 b->mode[1] = b->mode[0];

  1386.             }

  1387.             if (b->bs != BS_4x8) {

  1388.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1389.                                               s->prob.p.y_mode[0]);

  1390.                 s->counts.y_mode[0][b->mode[2]]++;

  1391.                 if (b->bs != BS_8x4) {

  1392.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1393.                                                   s->prob.p.y_mode[0]);

  1394.                     s->counts.y_mode[0][b->mode[3]]++;

  1395.                 } else {

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

  1397.                 }

  1398.             } else {

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

  1400.                 b->mode[3] = b->mode[1];

  1401.             }

  1402.         } else {

  1403.             static const uint8_t size_group[10] = {

  1404.                 3, 3, 3, 3, 2, 2, 2, 1, 1, 1

  1405.             };

  1406.             int sz = size_group[b->bs];

  1407. 

  1408.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

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

  1410.             b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1411.             s->counts.y_mode[sz][b->mode[3]]++;

  1412.         }

  1413.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1414.                                      s->prob.p.uv_mode[b->mode[3]]);

  1415.         s->counts.uv_mode[b->mode[3]][b->uvmode]++;

  1416.     } else {

  1417.         static const uint8_t inter_mode_ctx_lut[14][14] = {

  1418.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1419.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1420.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1421.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1422.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1423.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1424.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1425.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1426.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1427.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1428.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1429.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1430.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },

  1431.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },

  1432.         };

  1433. 

  1434.         if (s->segmentation.feat[b->seg_id].ref_enabled) {

  1435.             av_assert2(s->segmentation.feat[b->seg_id].ref_val != 0);

  1436.             b->comp = 0;

  1437.             b->ref[0] = s->segmentation.feat[b->seg_id].ref_val - 1;

  1438.         } else {

  1439.             // read comp_pred flag

  1440.             if (s->comppredmode != PRED_SWITCHABLE) {

  1441.                 b->comp = s->comppredmode == PRED_COMPREF;

  1442.             } else {

  1443.                 int c;

  1444. 

  1445.                 // FIXME add intra as ref=0xff (or -1) to make these easier?

  1446.                 if (have_a) {

  1447.                     if (have_l) {

  1448.                         if (s->above_comp_ctx[col] && s->left_comp_ctx[row7]) {

  1449.                             c = 4;

  1450.                         } else if (s->above_comp_ctx[col]) {

  1451.                             c = 2 + (s->left_intra_ctx[row7] ||

  1452.                                      s->left_ref_ctx[row7] == s->fixcompref);

  1453.                         } else if (s->left_comp_ctx[row7]) {

  1454.                             c = 2 + (s->above_intra_ctx[col] ||

  1455.                                      s->above_ref_ctx[col] == s->fixcompref);

  1456.                         } else {

  1457.                             c = (!s->above_intra_ctx[col] &&

  1458.                                  s->above_ref_ctx[col] == s->fixcompref) ^

  1459.                             (!s->left_intra_ctx[row7] &&

  1460.                              s->left_ref_ctx[row & 7] == s->fixcompref);

  1461.                         }

  1462.                     } else {

  1463.                         c = s->above_comp_ctx[col] ? 3 :

  1464.                         (!s->above_intra_ctx[col] && s->above_ref_ctx[col] == s->fixcompref);

  1465.                     }

  1466.                 } else if (have_l) {

  1467.                     c = s->left_comp_ctx[row7] ? 3 :

  1468.                     (!s->left_intra_ctx[row7] && s->left_ref_ctx[row7] == s->fixcompref);

  1469.                 } else {

  1470.                     c = 1;

  1471.                 }

  1472.                 b->comp = vp56_rac_get_prob(&s->c, s->prob.p.comp[c]);

  1473.                 s->counts.comp[c][b->comp]++;

  1474.             }

  1475. 

  1476.             // read actual references

  1477.             // FIXME probably cache a few variables here to prevent repetitive

  1478.             // memory accesses below

  1479.             if (b->comp) /* two references */ {

  1480.                 int fix_idx = s->signbias[s->fixcompref], var_idx = !fix_idx, c, bit;

  1481. 

  1482.                 b->ref[fix_idx] = s->fixcompref;

  1483.                 // FIXME can this codeblob be replaced by some sort of LUT?

  1484.                 if (have_a) {

  1485.                     if (have_l) {

  1486.                         if (s->above_intra_ctx[col]) {

  1487.                             if (s->left_intra_ctx[row7]) {

  1488.                                 c = 2;

  1489.                             } else {

  1490.                                 c = 1 + 2 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1491.                             }

  1492.                         } else if (s->left_intra_ctx[row7]) {

  1493.                             c = 1 + 2 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1494.                         } else {

  1495.                             int refl = s->left_ref_ctx[row7], refa = s->above_ref_ctx[col];

  1496. 

  1497.                             if (refl == refa && refa == s->varcompref[1]) {

  1498.                                 c = 0;

  1499.                             } else if (!s->left_comp_ctx[row7] && !s->above_comp_ctx[col]) {

  1500.                                 if ((refa == s->fixcompref && refl == s->varcompref[0]) ||

  1501.                                     (refl == s->fixcompref && refa == s->varcompref[0])) {

  1502.                                     c = 4;

  1503.                                 } else {

  1504.                                     c = (refa == refl) ? 3 : 1;

  1505.                                 }

  1506.                             } else if (!s->left_comp_ctx[row7]) {

  1507.                                 if (refa == s->varcompref[1] && refl != s->varcompref[1]) {

  1508.                                     c = 1;

  1509.                                 } else {

  1510.                                     c = (refl == s->varcompref[1] &&

  1511.                                          refa != s->varcompref[1]) ? 2 : 4;

  1512.                                 }

  1513.                             } else if (!s->above_comp_ctx[col]) {

  1514.                                 if (refl == s->varcompref[1] && refa != s->varcompref[1]) {

  1515.                                     c = 1;

  1516.                                 } else {

  1517.                                     c = (refa == s->varcompref[1] &&

  1518.                                          refl != s->varcompref[1]) ? 2 : 4;

  1519.                                 }

  1520.                             } else {

  1521.                                 c = (refl == refa) ? 4 : 2;

  1522.                             }

  1523.                         }

  1524.                     } else {

  1525.                         if (s->above_intra_ctx[col]) {

  1526.                             c = 2;

  1527.                         } else if (s->above_comp_ctx[col]) {

  1528.                             c = 4 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1529.                         } else {

  1530.                             c = 3 * (s->above_ref_ctx[col] != s->varcompref[1]);

  1531.                         }

  1532.                     }

  1533.                 } else if (have_l) {

  1534.                     if (s->left_intra_ctx[row7]) {

  1535.                         c = 2;

  1536.                     } else if (s->left_comp_ctx[row7]) {

  1537.                         c = 4 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1538.                     } else {

  1539.                         c = 3 * (s->left_ref_ctx[row7] != s->varcompref[1]);

  1540.                     }

  1541.                 } else {

  1542.                     c = 2;

  1543.                 }

  1544.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.comp_ref[c]);

  1545.                 b->ref[var_idx] = s->varcompref[bit];

  1546.                 s->counts.comp_ref[c][bit]++;

  1547.             } else /* single reference */ {

  1548.                 int bit, c;

  1549. 

  1550.                 if (have_a && !s->above_intra_ctx[col]) {

  1551.                     if (have_l && !s->left_intra_ctx[row7]) {

  1552.                         if (s->left_comp_ctx[row7]) {

  1553.                             if (s->above_comp_ctx[col]) {

  1554.                                 c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7] ||

  1555.                                          !s->above_ref_ctx[col]);

  1556.                             } else {

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

  1558.                                     (!s->fixcompref || !s->left_ref_ctx[row7]);

  1559.                             }

  1560.                         } else if (s->above_comp_ctx[col]) {

  1561.                             c = (3 * !s->left_ref_ctx[row7]) +

  1562.                                 (!s->fixcompref || !s->above_ref_ctx[col]);

  1563.                         } else {

  1564.                             c = 2 * !s->left_ref_ctx[row7] + 2 * !s->above_ref_ctx[col];

  1565.                         }

  1566.                     } else if (s->above_intra_ctx[col]) {

  1567.                         c = 2;

  1568.                     } else if (s->above_comp_ctx[col]) {

  1569.                         c = 1 + (!s->fixcompref || !s->above_ref_ctx[col]);

  1570.                     } else {

  1571.                         c = 4 * (!s->above_ref_ctx[col]);

  1572.                     }

  1573.                 } else if (have_l && !s->left_intra_ctx[row7]) {

  1574.                     if (s->left_intra_ctx[row7]) {

  1575.                         c = 2;

  1576.                     } else if (s->left_comp_ctx[row7]) {

  1577.                         c = 1 + (!s->fixcompref || !s->left_ref_ctx[row7]);

  1578.                     } else {

  1579.                         c = 4 * (!s->left_ref_ctx[row7]);

  1580.                     }

  1581.                 } else {

  1582.                     c = 2;

  1583.                 }

  1584.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][0]);

  1585.                 s->counts.single_ref[c][0][bit]++;

  1586.                 if (!bit) {

  1587.                     b->ref[0] = 0;

  1588.                 } else {

  1589.                     // FIXME can this codeblob be replaced by some sort of LUT?

  1590.                     if (have_a) {

  1591.                         if (have_l) {

  1592.                             if (s->left_intra_ctx[row7]) {

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

  1594.                                     c = 2;

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

  1596.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1597.                                                  s->above_ref_ctx[col] == 1);

  1598.                                 } else if (!s->above_ref_ctx[col]) {

  1599.                                     c = 3;

  1600.                                 } else {

  1601.                                     c = 4 * (s->above_ref_ctx[col] == 1);

  1602.                                 }

  1603.                             } else if (s->above_intra_ctx[col]) {

  1604.                                 if (s->left_intra_ctx[row7]) {

  1605.                                     c = 2;

  1606.                                 } else if (s->left_comp_ctx[row7]) {

  1607.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1608.                                                  s->left_ref_ctx[row7] == 1);

  1609.                                 } else if (!s->left_ref_ctx[row7]) {

  1610.                                     c = 3;

  1611.                                 } else {

  1612.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1613.                                 }

  1614.                             } else if (s->above_comp_ctx[col]) {

  1615.                                 if (s->left_comp_ctx[row7]) {

  1616.                                     if (s->left_ref_ctx[row7] == s->above_ref_ctx[col]) {

  1617.                                         c = 3 * (s->fixcompref == 1 ||

  1618.                                                  s->left_ref_ctx[row7] == 1);

  1619.                                     } else {

  1620.                                         c = 2;

  1621.                                     }

  1622.                                 } else if (!s->left_ref_ctx[row7]) {

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

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

  1625.                                 } else {

  1626.                                     c = 3 * (s->left_ref_ctx[row7] == 1) +

  1627.                                     (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1628.                                 }

  1629.                             } else if (s->left_comp_ctx[row7]) {

  1630.                                 if (!s->above_ref_ctx[col]) {

  1631.                                     c = 1 + 2 * (s->fixcompref == 1 ||

  1632.                                                  s->left_ref_ctx[row7] == 1);

  1633.                                 } else {

  1634.                                     c = 3 * (s->above_ref_ctx[col] == 1) +

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

  1636.                                 }

  1637.                             } else if (!s->above_ref_ctx[col]) {

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

  1639.                                     c = 3;

  1640.                                 } else {

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

  1642.                                 }

  1643.                             } else if (!s->left_ref_ctx[row7]) {

  1644.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1645.                             } else {

  1646.                                 c = 2 * (s->left_ref_ctx[row7] == 1) +

  1647.                                 2 * (s->above_ref_ctx[col] == 1);

  1648.                             }

  1649.                         } else {

  1650.                             if (s->above_intra_ctx[col] ||

  1651.                                 (!s->above_comp_ctx[col] && !s->above_ref_ctx[col])) {

  1652.                                 c = 2;

  1653.                             } else if (s->above_comp_ctx[col]) {

  1654.                                 c = 3 * (s->fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1655.                             } else {

  1656.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1657.                             }

  1658.                         }

  1659.                     } else if (have_l) {

  1660.                         if (s->left_intra_ctx[row7] ||

  1661.                             (!s->left_comp_ctx[row7] && !s->left_ref_ctx[row7])) {

  1662.                             c = 2;

  1663.                         } else if (s->left_comp_ctx[row7]) {

  1664.                             c = 3 * (s->fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1665.                         } else {

  1666.                             c = 4 * (s->left_ref_ctx[row7] == 1);

  1667.                         }

  1668.                     } else {

  1669.                         c = 2;

  1670.                     }

  1671.                     bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][1]);

  1672.                     s->counts.single_ref[c][1][bit]++;

  1673.                     b->ref[0] = 1 + bit;

  1674.                 }

  1675.             }

  1676.         }

  1677. 

  1678.         if (b->bs <= BS_8x8) {

  1679.             if (s->segmentation.feat[b->seg_id].skip_enabled) {

  1680.                 b->mode[0] = b->mode[1] = b->mode[2] = b->mode[3] = ZEROMV;

  1681.             } else {

  1682.                 static const uint8_t off[10] = {

  1683.                     3, 0, 0, 1, 0, 0, 0, 0, 0, 0

  1684.                 };

  1685. 

  1686.                 // FIXME this needs to use the LUT tables from find_ref_mvs

  1687.                 // because not all are -1,0/0,-1

  1688.                 int c = inter_mode_ctx_lut[s->above_mode_ctx[col + off[b->bs]]]

  1689.                                           [s->left_mode_ctx[row7 + off[b->bs]]];

  1690. 

  1691.                 b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1692.                                               s->prob.p.mv_mode[c]);

  1693.                 b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1694.                 s->counts.mv_mode[c][b->mode[0] - 10]++;

  1695.             }

  1696.         }

  1697. 

  1698.         if (s->filtermode == FILTER_SWITCHABLE) {

  1699.             int c;

  1700. 

  1701.             if (have_a && s->above_mode_ctx[col] >= NEARESTMV) {

  1702.                 if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1703.                     c = s->above_filter_ctx[col] == s->left_filter_ctx[row7] ?

  1704.                         s->left_filter_ctx[row7] : 3;

  1705.                 } else {

  1706.                     c = s->above_filter_ctx[col];

  1707.                 }

  1708.             } else if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1709.                 c = s->left_filter_ctx[row7];

  1710.             } else {

  1711.                 c = 3;

  1712.             }

  1713. 

  1714.             b->filter = vp8_rac_get_tree(&s->c, vp9_filter_tree,

  1715.                                          s->prob.p.filter[c]);

  1716.             s->counts.filter[c][b->filter]++;

  1717.         } else {

  1718.             b->filter = s->filtermode;

  1719.         }

  1720. 

  1721.         if (b->bs > BS_8x8) {

  1722.             int c = inter_mode_ctx_lut[s->above_mode_ctx[col]][s->left_mode_ctx[row7]];

  1723. 

  1724.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1725.                                           s->prob.p.mv_mode[c]);

  1726.             s->counts.mv_mode[c][b->mode[0] - 10]++;

  1727.             fill_mv(s, b->mv[0], b->mode[0], 0);

  1728. 

  1729.             if (b->bs != BS_8x4) {

  1730.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1731.                                               s->prob.p.mv_mode[c]);

  1732.                 s->counts.mv_mode[c][b->mode[1] - 10]++;

  1733.                 fill_mv(s, b->mv[1], b->mode[1], 1);

  1734.             } else {

  1735.                 b->mode[1] = b->mode[0];

  1736.                 AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1737.                 AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1738.             }

  1739. 

  1740.             if (b->bs != BS_4x8) {

  1741.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1742.                                               s->prob.p.mv_mode[c]);

  1743.                 s->counts.mv_mode[c][b->mode[2] - 10]++;

  1744.                 fill_mv(s, b->mv[2], b->mode[2], 2);

  1745. 

  1746.                 if (b->bs != BS_8x4) {

  1747.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1748.                                                   s->prob.p.mv_mode[c]);

  1749.                     s->counts.mv_mode[c][b->mode[3] - 10]++;

  1750.                     fill_mv(s, b->mv[3], b->mode[3], 3);

  1751.                 } else {

  1752.                     b->mode[3] = b->mode[2];

  1753.                     AV_COPY32(&b->mv[3][0], &b->mv[2][0]);

  1754.                     AV_COPY32(&b->mv[3][1], &b->mv[2][1]);

  1755.                 }

  1756.             } else {

  1757.                 b->mode[2] = b->mode[0];

  1758.                 AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1759.                 AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1760.                 b->mode[3] = b->mode[1];

  1761.                 AV_COPY32(&b->mv[3][0], &b->mv[1][0]);

  1762.                 AV_COPY32(&b->mv[3][1], &b->mv[1][1]);

  1763.             }

  1764.         } else {

  1765.             fill_mv(s, b->mv[0], b->mode[0], -1);

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

  1767.             AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1768.             AV_COPY32(&b->mv[3][0], &b->mv[0][0]);

  1769.             AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1770.             AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1771.             AV_COPY32(&b->mv[3][1], &b->mv[0][1]);

  1772.         }

  1773.     }

  1774. 

  1775.     // FIXME this can probably be optimized

  1776.     memset(&s->above_skip_ctx[col], b->skip, w4);

  1777.     memset(&s->left_skip_ctx[row7], b->skip, h4);

  1778.     memset(&s->above_txfm_ctx[col], b->tx, w4);

  1779.     memset(&s->left_txfm_ctx[row7], b->tx, h4);

  1780.     memset(&s->above_partition_ctx[col], above_ctx[b->bs], w4);

  1781.     memset(&s->left_partition_ctx[row7], left_ctx[b->bs], h4);

  1782.     if (!s->keyframe && !s->intraonly) {

  1783.         memset(&s->above_intra_ctx[col], b->intra, w4);

  1784.         memset(&s->left_intra_ctx[row7], b->intra, h4);

  1785.         memset(&s->above_comp_ctx[col], b->comp, w4);

  1786.         memset(&s->left_comp_ctx[row7], b->comp, h4);

  1787.         memset(&s->above_mode_ctx[col], b->mode[3], w4);

  1788.         memset(&s->left_mode_ctx[row7], b->mode[3], h4);

  1789.         if (s->filtermode == FILTER_SWITCHABLE && !b->intra ) {

  1790.             memset(&s->above_filter_ctx[col], b->filter, w4);

  1791.             memset(&s->left_filter_ctx[row7], b->filter, h4);

  1792.             b->filter = vp9_filter_lut[b->filter];

  1793.         }

  1794.         if (b->bs > BS_8x8) {

  1795.             int mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  1796. 

  1797.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][0], &b->mv[1][0]);

  1798.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][1], &b->mv[1][1]);

  1799.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][0], mv0);

  1800.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][1], mv1);

  1801.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][0], &b->mv[2][0]);

  1802.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][1], &b->mv[2][1]);

  1803.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][0], mv0);

  1804.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][1], mv1);

  1805.         } else {

  1806.             int n, mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  1807. 

  1808.             for (n = 0; n < w4 * 2; n++) {

  1809.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][0], mv0);

  1810.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][1], mv1);

  1811.             }

  1812.             for (n = 0; n < h4 * 2; n++) {

  1813.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][0], mv0);

  1814.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][1], mv1);

  1815.             }

  1816.         }

  1817. 

  1818.         if (!b->intra) { // FIXME write 0xff or -1 if intra, so we can use this

  1819.                          // as a direct check in above branches

  1820.             int vref = b->ref[b->comp ? s->signbias[s->varcompref[0]] : 0];

  1821. 

  1822.             memset(&s->above_ref_ctx[col], vref, w4);

  1823.             memset(&s->left_ref_ctx[row7], vref, h4);

  1824.         }

  1825.     }

  1826. 

  1827.     // FIXME kinda ugly

  1828.     for (y = 0; y < h4; y++) {

  1829.         int x, o = (row + y) * s->sb_cols * 8 + col;

  1830.         struct VP9mvrefPair *mv = &s->frames[CUR_FRAME].mv[o];

  1831. 

  1832.         if (b->intra) {

  1833.             for (x = 0; x < w4; x++) {

  1834.                 mv[x].ref[0] =

  1835.                 mv[x].ref[1] = -1;

  1836.             }

  1837.         } else if (b->comp) {

  1838.             for (x = 0; x < w4; x++) {

  1839.                 mv[x].ref[0] = b->ref[0];

  1840.                 mv[x].ref[1] = b->ref[1];

  1841.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  1842.                 AV_COPY32(&mv[x].mv[1], &b->mv[3][1]);

  1843.             }

  1844.         } else {

  1845.             for (x = 0; x < w4; x++) {

  1846.                 mv[x].ref[0] = b->ref[0];

  1847.                 mv[x].ref[1] = -1;

  1848.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  1849.             }

  1850.         }

  1851.     }

  1852. }

  1853. 

  1854. // FIXME remove tx argument, and merge cnt/eob arguments?

  1855. static int decode_coeffs_b(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  1856.                            enum TxfmMode tx, unsigned (*cnt)[6][3],

  1857.                            unsigned (*eob)[6][2], uint8_t (*p)[6][11],

  1858.                            int nnz, const int16_t *scan, const int16_t (*nb)[2],

  1859.                            const int16_t *band_counts, const int16_t *qmul)

  1860. {

  1861.     int i = 0, band = 0, band_left = band_counts[band];

  1862.     uint8_t *tp = p[0][nnz];

  1863.     uint8_t cache[1024];

  1864. 

  1865.     do {

  1866.         int val, rc;

  1867. 

  1868.         val = vp56_rac_get_prob_branchy(c, tp[0]); // eob

  1869.         eob[band][nnz][val]++;

  1870.         if (!val)

  1871.             break;

  1872. 

  1873.     skip_eob:

  1874.         if (!vp56_rac_get_prob_branchy(c, tp[1])) { // zero

  1875.             cnt[band][nnz][0]++;

  1876.             if (!--band_left)

  1877.                 band_left = band_counts[++band];

  1878.             cache[scan[i]] = 0;

  1879.             nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  1880.             tp = p[band][nnz];

  1881.             if (++i == n_coeffs)

  1882.                 break; //invalid input; blocks should end with EOB

  1883.             goto skip_eob;

  1884.         }

  1885. 

  1886.         rc = scan[i];

  1887.         if (!vp56_rac_get_prob_branchy(c, tp[2])) { // one

  1888.             cnt[band][nnz][1]++;

  1889.             val = 1;

  1890.             cache[rc] = 1;

  1891.         } else {

  1892.             // fill in p[3-10] (model fill) - only once per frame for each pos

  1893.             if (!tp[3])

  1894.                 memcpy(&tp[3], vp9_model_pareto8[tp[2]], 8);

  1895. 

  1896.             cnt[band][nnz][2]++;

  1897.             if (!vp56_rac_get_prob_branchy(c, tp[3])) { // 2, 3, 4

  1898.                 if (!vp56_rac_get_prob_branchy(c, tp[4])) {

  1899.                     cache[rc] = val = 2;

  1900.                 } else {

  1901.                     val = 3 + vp56_rac_get_prob(c, tp[5]);

  1902.                     cache[rc] = 3;

  1903.                 }

  1904.             } else if (!vp56_rac_get_prob_branchy(c, tp[6])) { // cat1/2

  1905.                 cache[rc] = 4;

  1906.                 if (!vp56_rac_get_prob_branchy(c, tp[7])) {

  1907.                     val = 5 + vp56_rac_get_prob(c, 159);

  1908.                 } else {

  1909.                     val  = 7 + (vp56_rac_get_prob(c, 165) << 1);

  1910.                     val +=      vp56_rac_get_prob(c, 145);

  1911.                 }

  1912.             } else { // cat 3-6

  1913.                 cache[rc] = 5;

  1914.                 if (!vp56_rac_get_prob_branchy(c, tp[8])) {

  1915.                     if (!vp56_rac_get_prob_branchy(c, tp[9])) {

  1916.                         val  = 11 + (vp56_rac_get_prob(c, 173) << 2);

  1917.                         val +=      (vp56_rac_get_prob(c, 148) << 1);

  1918.                         val +=       vp56_rac_get_prob(c, 140);

  1919.                     } else {

  1920.                         val  = 19 + (vp56_rac_get_prob(c, 176) << 3);

  1921.                         val +=      (vp56_rac_get_prob(c, 155) << 2);

  1922.                         val +=      (vp56_rac_get_prob(c, 140) << 1);

  1923.                         val +=       vp56_rac_get_prob(c, 135);

  1924.                     }

  1925.                 } else if (!vp56_rac_get_prob_branchy(c, tp[10])) {

  1926.                     val  = 35 + (vp56_rac_get_prob(c, 180) << 4);

  1927.                     val +=      (vp56_rac_get_prob(c, 157) << 3);

  1928.                     val +=      (vp56_rac_get_prob(c, 141) << 2);

  1929.                     val +=      (vp56_rac_get_prob(c, 134) << 1);

  1930.                     val +=       vp56_rac_get_prob(c, 130);

  1931.                 } else {

  1932.                     val  = 67 + (vp56_rac_get_prob(c, 254) << 13);

  1933.                     val +=      (vp56_rac_get_prob(c, 254) << 12);

  1934.                     val +=      (vp56_rac_get_prob(c, 254) << 11);

  1935.                     val +=      (vp56_rac_get_prob(c, 252) << 10);

  1936.                     val +=      (vp56_rac_get_prob(c, 249) << 9);

  1937.                     val +=      (vp56_rac_get_prob(c, 243) << 8);

  1938.                     val +=      (vp56_rac_get_prob(c, 230) << 7);

  1939.                     val +=      (vp56_rac_get_prob(c, 196) << 6);

  1940.                     val +=      (vp56_rac_get_prob(c, 177) << 5);

  1941.                     val +=      (vp56_rac_get_prob(c, 153) << 4);

  1942.                     val +=      (vp56_rac_get_prob(c, 140) << 3);

  1943.                     val +=      (vp56_rac_get_prob(c, 133) << 2);

  1944.                     val +=      (vp56_rac_get_prob(c, 130) << 1);

  1945.                     val +=       vp56_rac_get_prob(c, 129);

  1946.                 }

  1947.             }

  1948.         }

  1949.         if (!--band_left)

  1950.             band_left = band_counts[++band];

  1951.         if (tx == TX_32X32) // FIXME slow

  1952.             coef[rc] = ((vp8_rac_get(c) ? -val : val) * qmul[!!i]) / 2;

  1953.         else

  1954.             coef[rc] = (vp8_rac_get(c) ? -val : val) * qmul[!!i];

  1955.         nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  1956.         tp = p[band][nnz];

  1957.     } while (++i < n_coeffs);

  1958. 

  1959.     return i;

  1960. }

  1961. 

  1962. static int decode_coeffs(AVCodecContext *ctx)

  1963. {

  1964.     VP9Context *s = ctx->priv_data;

  1965.     VP9Block *b = s->b;

  1966.     int row = s->row, col = s->col;

  1967.     uint8_t (*p)[6][11] = s->prob.coef[b->tx][0 /* y */][!b->intra];

  1968.     unsigned (*c)[6][3] = s->counts.coef[b->tx][0 /* y */][!b->intra];

  1969.     unsigned (*e)[6][2] = s->counts.eob[b->tx][0 /* y */][!b->intra];

  1970.     int w4 = bwh_tab[1][b->bs][0] << 1, h4 = bwh_tab[1][b->bs][1] << 1;

  1971.     int end_x = FFMIN(2 * (s->cols - col), w4);

  1972.     int end_y = FFMIN(2 * (s->rows - row), h4);

  1973.     int n, pl, x, y, step1d = 1 << b->tx, step = 1 << (b->tx * 2);

  1974.     int uvstep1d = 1 << b->uvtx, uvstep = 1 << (b->uvtx * 2), res;

  1975.     int16_t (*qmul)[2] = s->segmentation.feat[b->seg_id].qmul;

  1976.     int tx = 4 * s->lossless + b->tx;

  1977.     const int16_t * const *yscans = vp9_scans[tx];

  1978.     const int16_t (* const *ynbs)[2] = vp9_scans_nb[tx];

  1979.     const int16_t *uvscan = vp9_scans[b->uvtx][DCT_DCT];

  1980.     const int16_t (*uvnb)[2] = vp9_scans_nb[b->uvtx][DCT_DCT];

  1981.     uint8_t *a = &s->above_y_nnz_ctx[col * 2];

  1982.     uint8_t *l = &s->left_y_nnz_ctx[(row & 7) << 1];

  1983.     static const int16_t band_counts[4][8] = {

  1984.         { 1, 2, 3, 4,  3,   16 - 13 },

  1985.         { 1, 2, 3, 4, 11,   64 - 21 },

  1986.         { 1, 2, 3, 4, 11,  256 - 21 },

  1987.         { 1, 2, 3, 4, 11, 1024 - 21 },

  1988.     };

  1989.     const int16_t *y_band_counts = band_counts[b->tx];

  1990.     const int16_t *uv_band_counts = band_counts[b->uvtx];

  1991. 

  1992.     /* y tokens */

  1993.     if (b->tx > TX_4X4) { // FIXME slow

  1994.         for (y = 0; y < end_y; y += step1d)

  1995.             for (x = 1; x < step1d; x++)

  1996.                 l[y] |= l[y + x];

  1997.         for (x = 0; x < end_x; x += step1d)

  1998.             for (y = 1; y < step1d; y++)

  1999.                 a[x] |= a[x + y];

  2000.     }

  2001.     for (n = 0, y = 0; y < end_y; y += step1d) {

  2002.         for (x = 0; x < end_x; x += step1d, n += step) {

  2003.             enum TxfmType txtp = vp9_intra_txfm_type[b->mode[b->tx == TX_4X4 &&

  2004.                                                              b->bs > BS_8x8 ?

  2005.                                                              n : 0]];

  2006.             int nnz = a[x] + l[y];

  2007.             if ((res = decode_coeffs_b(&s->c, s->block + 16 * n, 16 * step,

  2008.                                        b->tx, c, e, p, nnz, yscans[txtp],

  2009.                                        ynbs[txtp], y_band_counts, qmul[0])) < 0)

  2010.                 return res;

  2011.             a[x] = l[y] = !!res;

  2012.             if (b->tx > TX_8X8) {

  2013.                 AV_WN16A(&s->eob[n], res);

  2014.             } else {

  2015.                 s->eob[n] = res;

  2016.             }

  2017.         }

  2018.     }

  2019.     if (b->tx > TX_4X4) { // FIXME slow

  2020.         for (y = 0; y < end_y; y += step1d)

  2021.             memset(&l[y + 1], l[y], FFMIN(end_y - y - 1, step1d - 1));

  2022.         for (x = 0; x < end_x; x += step1d)

  2023.             memset(&a[x + 1], a[x], FFMIN(end_x - x - 1, step1d - 1));

  2024.     }

  2025. 

  2026.     p = s->prob.coef[b->uvtx][1 /* uv */][!b->intra];

  2027.     c = s->counts.coef[b->uvtx][1 /* uv */][!b->intra];

  2028.     e = s->counts.eob[b->uvtx][1 /* uv */][!b->intra];

  2029.     w4 >>= 1;

  2030.     h4 >>= 1;

  2031.     end_x >>= 1;

  2032.     end_y >>= 1;

  2033.     for (pl = 0; pl < 2; pl++) {

  2034.         a = &s->above_uv_nnz_ctx[pl][col];

  2035.         l = &s->left_uv_nnz_ctx[pl][row & 7];

  2036.         if (b->uvtx > TX_4X4) { // FIXME slow

  2037.             for (y = 0; y < end_y; y += uvstep1d)

  2038.                 for (x = 1; x < uvstep1d; x++)

  2039.                     l[y] |= l[y + x];

  2040.             for (x = 0; x < end_x; x += uvstep1d)

  2041.                 for (y = 1; y < uvstep1d; y++)

  2042.                     a[x] |= a[x + y];

  2043.         }

  2044.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2045.             for (x = 0; x < end_x; x += uvstep1d, n += uvstep) {

  2046.                 int nnz = a[x] + l[y];

  2047.                 if ((res = decode_coeffs_b(&s->c, s->uvblock[pl] + 16 * n,

  2048.                                            16 * uvstep, b->uvtx, c, e, p, nnz,

  2049.                                            uvscan, uvnb, uv_band_counts,

  2050.                                            qmul[1])) < 0)

  2051.                     return res;

  2052.                 a[x] = l[y] = !!res;

  2053.                 if (b->uvtx > TX_8X8) {

  2054.                     AV_WN16A(&s->uveob[pl][n], res);

  2055.                 } else {

  2056.                     s->uveob[pl][n] = res;

  2057.                 }

  2058.             }

  2059.         }

  2060.         if (b->uvtx > TX_4X4) { // FIXME slow

  2061.             for (y = 0; y < end_y; y += uvstep1d)

  2062.                 memset(&l[y + 1], l[y], FFMIN(end_y - y - 1, uvstep1d - 1));

  2063.             for (x = 0; x < end_x; x += uvstep1d)

  2064.                 memset(&a[x + 1], a[x], FFMIN(end_x - x - 1, uvstep1d - 1));

  2065.         }

  2066.     }

  2067. 

  2068.     return 0;

  2069. }

  2070. 

  2071. static av_always_inline int check_intra_mode(VP9Context *s, int mode, uint8_t **a,

  2072.                                              uint8_t *dst_edge, ptrdiff_t stride_edge,

  2073.                                              uint8_t *dst_inner, ptrdiff_t stride_inner,

  2074.                                              uint8_t *l, int col, int x, int w,

  2075.                                              int row, int y, enum TxfmMode tx,

  2076.                                              int p)

  2077. {

  2078.     int have_top = row > 0 || y > 0;

  2079.     int have_left = col > s->tiling.tile_col_start || x > 0;

  2080.     int have_right = x < w - 1;

  2081.     static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {

  2082.         [VERT_PRED]            = { { DC_127_PRED,          VERT_PRED },

  2083.                                    { DC_127_PRED,          VERT_PRED } },

  2084.         [HOR_PRED]             = { { DC_129_PRED,          DC_129_PRED },

  2085.                                    { HOR_PRED,             HOR_PRED } },

  2086.         [DC_PRED]              = { { DC_128_PRED,          TOP_DC_PRED },

  2087.                                    { LEFT_DC_PRED,         DC_PRED } },

  2088.         [DIAG_DOWN_LEFT_PRED]  = { { DC_127_PRED,          DIAG_DOWN_LEFT_PRED },

  2089.                                    { DC_127_PRED,          DIAG_DOWN_LEFT_PRED } },

  2090.         [DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },

  2091.                                    { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },

  2092.         [VERT_RIGHT_PRED]      = { { VERT_RIGHT_PRED,      VERT_RIGHT_PRED },

  2093.                                    { VERT_RIGHT_PRED,      VERT_RIGHT_PRED } },

  2094.         [HOR_DOWN_PRED]        = { { HOR_DOWN_PRED,        HOR_DOWN_PRED },

  2095.                                    { HOR_DOWN_PRED,        HOR_DOWN_PRED } },

  2096.         [VERT_LEFT_PRED]       = { { DC_127_PRED,          VERT_LEFT_PRED },

  2097.                                    { DC_127_PRED,          VERT_LEFT_PRED } },

  2098.         [HOR_UP_PRED]          = { { DC_129_PRED,          DC_129_PRED },

  2099.                                    { HOR_UP_PRED,          HOR_UP_PRED } },

  2100.         [TM_VP8_PRED]          = { { DC_129_PRED,          VERT_PRED },

  2101.                                    { HOR_PRED,             TM_VP8_PRED } },

  2102.     };

  2103.     static const struct {

  2104.         uint8_t needs_left:1;

  2105.         uint8_t needs_top:1;

  2106.         uint8_t needs_topleft:1;

  2107.         uint8_t needs_topright:1;

  2108.     } edges[N_INTRA_PRED_MODES] = {

  2109.         [VERT_PRED]            = { .needs_top  = 1 },

  2110.         [HOR_PRED]             = { .needs_left = 1 },

  2111.         [DC_PRED]              = { .needs_top  = 1, .needs_left = 1 },

  2112.         [DIAG_DOWN_LEFT_PRED]  = { .needs_top  = 1, .needs_topright = 1 },

  2113.         [DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2114.         [VERT_RIGHT_PRED]      = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2115.         [HOR_DOWN_PRED]        = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2116.         [VERT_LEFT_PRED]       = { .needs_top  = 1, .needs_topright = 1 },

  2117.         [HOR_UP_PRED]          = { .needs_left = 1 },

  2118.         [TM_VP8_PRED]          = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2119.         [LEFT_DC_PRED]         = { .needs_left = 1 },

  2120.         [TOP_DC_PRED]          = { .needs_top  = 1 },

  2121.         [DC_128_PRED]          = { 0 },

  2122.         [DC_127_PRED]          = { 0 },

  2123.         [DC_129_PRED]          = { 0 }

  2124.     };

  2125. 

  2126.     av_assert2(mode >= 0 && mode < 10);

  2127.     mode = mode_conv[mode][have_left][have_top];

  2128.     if (edges[mode].needs_top) {

  2129.         uint8_t *top, *topleft;

  2130.         int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !p) - x) * 4;

  2131.         int n_px_need_tr = 0;

  2132. 

  2133.         if (tx == TX_4X4 && edges[mode].needs_topright && have_right)

  2134.             n_px_need_tr = 4;

  2135. 

  2136.         // if top of sb64-row, use s->intra_pred_data[] instead of

  2137.         // dst[-stride] for intra prediction (it contains pre- instead of

  2138.         // post-loopfilter data)

  2139.         if (have_top) {

  2140.             top = !(row & 7) && !y ?

  2141.                 s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  2142.                 y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];

  2143.             if (have_left)

  2144.                 topleft = !(row & 7) && !y ?

  2145.                     s->intra_pred_data[p] + col * (8 >> !!p) + x * 4 :

  2146.                     y == 0 || x == 0 ? &dst_edge[-stride_edge] :

  2147.                     &dst_inner[-stride_inner];

  2148.         }

  2149. 

  2150.         if (have_top &&

  2151.             (!edges[mode].needs_topleft || (have_left && top == topleft)) &&

  2152.             (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&

  2153.             n_px_need + n_px_need_tr <= n_px_have) {

  2154.             *a = top;

  2155.         } else {

  2156.             if (have_top) {

  2157.                 if (n_px_need <= n_px_have) {

  2158.                     memcpy(*a, top, n_px_need);

  2159.                 } else {

  2160.                     memcpy(*a, top, n_px_have);

  2161.                     memset(&(*a)[n_px_have], (*a)[n_px_have - 1],

  2162.                            n_px_need - n_px_have);

  2163.                 }

  2164.             } else {

  2165.                 memset(*a, 127, n_px_need);

  2166.             }

  2167.             if (edges[mode].needs_topleft) {

  2168.                 if (have_left && have_top) {

  2169.                     (*a)[-1] = topleft[-1];

  2170.                 } else {

  2171.                     (*a)[-1] = have_top ? 129 : 127;

  2172.                 }

  2173.             }

  2174.             if (tx == TX_4X4 && edges[mode].needs_topright) {

  2175.                 if (have_top && have_right &&

  2176.                     n_px_need + n_px_need_tr <= n_px_have) {

  2177.                     memcpy(&(*a)[4], &top[4], 4);

  2178.                 } else {

  2179.                     memset(&(*a)[4], (*a)[3], 4);

  2180.                 }

  2181.             }

  2182.         }

  2183.     }

  2184.     if (edges[mode].needs_left) {

  2185.         if (have_left) {

  2186.             int n_px_need = 4 << tx, i, n_px_have = (((s->rows - row) << !p) - y) * 4;

  2187.             uint8_t *dst = x == 0 ? dst_edge : dst_inner;

  2188.             ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;

  2189. 

  2190.             if (n_px_need <= n_px_have) {

  2191.                 for (i = 0; i < n_px_need; i++)

  2192.                     l[i] = dst[i * stride - 1];

  2193.             } else {

  2194.                 for (i = 0; i < n_px_have; i++)

  2195.                     l[i] = dst[i * stride - 1];

  2196.                 memset(&l[i], l[i - 1], n_px_need - n_px_have);

  2197.             }

  2198.         } else {

  2199.             memset(l, 129, 4 << tx);

  2200.         }

  2201.     }

  2202. 

  2203.     return mode;

  2204. }

  2205. 

  2206. static void intra_recon(AVCodecContext *ctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  2207. {

  2208.     VP9Context *s = ctx->priv_data;

  2209.     VP9Block *b = s->b;

  2210.     int row = s->row, col = s->col;

  2211.     int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2212.     int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2213.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2214.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2215.     int tx = 4 * s->lossless + b->tx, uvtx = b->uvtx + 4 * s->lossless;

  2216.     int uvstep1d = 1 << b->uvtx, p;

  2217.     uint8_t *dst = s->dst[0], *dst_r = s->frames[CUR_FRAME].tf.f->data[0] + y_off;

  2218. 

  2219.     for (n = 0, y = 0; y < end_y; y += step1d) {

  2220.         uint8_t *ptr = dst, *ptr_r = dst_r;

  2221.         for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d,

  2222.                                ptr_r += 4 * step1d, n += step) {

  2223.             int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?

  2224.                                y * 2 + x : 0];

  2225.             LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);

  2226.             uint8_t *a = &a_buf[16], l[32];

  2227.             enum TxfmType txtp = vp9_intra_txfm_type[mode];

  2228.             int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2229. 

  2230.             mode = check_intra_mode(s, mode, &a, ptr_r,

  2231.                                     s->frames[CUR_FRAME].tf.f->linesize[0],

  2232.                                     ptr, s->y_stride, l,

  2233.                                     col, x, w4, row, y, b->tx, 0);

  2234.             s->dsp.intra_pred[b->tx][mode](ptr, s->y_stride, l, a);

  2235.             if (eob)

  2236.                 s->dsp.itxfm_add[tx][txtp](ptr, s->y_stride,

  2237.                                            s->block + 16 * n, eob);

  2238.         }

  2239.         dst_r += 4 * step1d * s->frames[CUR_FRAME].tf.f->linesize[0];

  2240.         dst   += 4 * step1d * s->y_stride;

  2241.     }

  2242. 

  2243.     // U/V

  2244.     h4 >>= 1;

  2245.     w4 >>= 1;

  2246.     end_x >>= 1;

  2247.     end_y >>= 1;

  2248.     step = 1 << (b->uvtx * 2);

  2249.     for (p = 0; p < 2; p++) {

  2250.         dst   = s->dst[1 + p];

  2251.         dst_r = s->frames[CUR_FRAME].tf.f->data[1 + p] + uv_off;

  2252.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2253.             uint8_t *ptr = dst, *ptr_r = dst_r;

  2254.             for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d,

  2255.                                    ptr_r += 4 * uvstep1d, n += step) {

  2256.                 int mode = b->uvmode;

  2257.                 LOCAL_ALIGNED_16(uint8_t, a_buf, [48]);

  2258.                 uint8_t *a = &a_buf[16], l[32];

  2259.                 int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2260. 

  2261.                 mode = check_intra_mode(s, mode, &a, ptr_r,

  2262.                                         s->frames[CUR_FRAME].tf.f->linesize[1],

  2263.                                         ptr, s->uv_stride, l,

  2264.                                         col, x, w4, row, y, b->uvtx, p + 1);

  2265.                 s->dsp.intra_pred[b->uvtx][mode](ptr, s->uv_stride, l, a);

  2266.                 if (eob)

  2267.                     s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2268.                                                     s->uvblock[p] + 16 * n, eob);

  2269.             }

  2270.             dst_r += 4 * uvstep1d * s->frames[CUR_FRAME].tf.f->linesize[1];

  2271.             dst   += 4 * uvstep1d * s->uv_stride;

  2272.         }

  2273.     }

  2274. }

  2275. 

  2276. static av_always_inline void mc_luma_dir(VP9Context *s, vp9_mc_func (*mc)[2],

  2277.                                          uint8_t *dst, ptrdiff_t dst_stride,

  2278.                                          const uint8_t *ref, ptrdiff_t ref_stride,

  2279.                                          ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2280.                                          int bw, int bh, int w, int h)

  2281. {

  2282.     int mx = mv->x, my = mv->y;

  2283. 

  2284.     y += my >> 3;

  2285.     x += mx >> 3;

  2286.     ref += y * ref_stride + x;

  2287.     mx &= 7;

  2288.     my &= 7;

  2289.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2290.     if (x < !!mx * 3 || y < !!my * 3 ||

  2291.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2292.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2293.                                  ref - !!my * 3 * ref_stride - !!mx * 3,

  2294.                                  80, ref_stride,

  2295.                                  bw + !!mx * 7, bh + !!my * 7,

  2296.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2297.         ref = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2298.         ref_stride = 80;

  2299.     }

  2300.     mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);

  2301. }

  2302. 

  2303. static av_always_inline void mc_chroma_dir(VP9Context *s, vp9_mc_func (*mc)[2],

  2304.                                            uint8_t *dst_u, uint8_t *dst_v,

  2305.                                            ptrdiff_t dst_stride,

  2306.                                            const uint8_t *ref_u, ptrdiff_t src_stride_u,

  2307.                                            const uint8_t *ref_v, ptrdiff_t src_stride_v,

  2308.                                            ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2309.                                            int bw, int bh, int w, int h)

  2310. {

  2311.     int mx = mv->x, my = mv->y;

  2312. 

  2313.     y += my >> 4;

  2314.     x += mx >> 4;

  2315.     ref_u += y * src_stride_u + x;

  2316.     ref_v += y * src_stride_v + x;

  2317.     mx &= 15;

  2318.     my &= 15;

  2319.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2320.     if (x < !!mx * 3 || y < !!my * 3 ||

  2321.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2322.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2323.                                  ref_u - !!my * 3 * src_stride_u - !!mx * 3,

  2324.                                  80, src_stride_u,

  2325.                                  bw + !!mx * 7, bh + !!my * 7,

  2326.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2327.         ref_u = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2328.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, 80, bh, mx, my);

  2329. 

  2330.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2331.                                  ref_v - !!my * 3 * src_stride_v - !!mx * 3,

  2332.                                  80, src_stride_v,

  2333.                                  bw + !!mx * 7, bh + !!my * 7,

  2334.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2335.         ref_v = s->edge_emu_buffer + !!my * 3 * 80 + !!mx * 3;

  2336.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, 80, bh, mx, my);

  2337.     } else {

  2338.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);

  2339.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);

  2340.     }

  2341. }

  2342. 

  2343. static void inter_recon(AVCodecContext *ctx)

  2344. {

  2345.     static const uint8_t bwlog_tab[2][N_BS_SIZES] = {

  2346.         { 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3, 4, 4 },

  2347.         { 1, 1, 2, 2, 2, 3, 3, 3, 4, 4, 4, 4, 4 },

  2348.     };

  2349.     VP9Context *s = ctx->priv_data;

  2350.     VP9Block *b = s->b;

  2351.     int row = s->row, col = s->col;

  2352.     ThreadFrame *tref1 = &s->refs[s->refidx[b->ref[0]]];

  2353.     AVFrame *ref1 = tref1->f;

  2354.     ThreadFrame *tref2 = b->comp ? &s->refs[s->refidx[b->ref[1]]] : NULL;

  2355.     AVFrame *ref2 = b->comp ? tref2->f : NULL;

  2356.     int w = ctx->width, h = ctx->height;

  2357.     ptrdiff_t ls_y = s->y_stride, ls_uv = s->uv_stride;

  2358. 

  2359.     // y inter pred

  2360.     if (b->bs > BS_8x8) {

  2361.         if (b->bs == BS_8x4) {

  2362.             mc_luma_dir(s, s->dsp.mc[3][b->filter][0], s->dst[0], ls_y,

  2363.                         ref1->data[0], ref1->linesize[0],

  2364.                         row << 3, col << 3, &b->mv[0][0], 8, 4, w, h);

  2365.             mc_luma_dir(s, s->dsp.mc[3][b->filter][0],

  2366.                         s->dst[0] + 4 * ls_y, ls_y,

  2367.                         ref1->data[0], ref1->linesize[0],

  2368.                         (row << 3) + 4, col << 3, &b->mv[2][0], 8, 4, w, h);

  2369. 

  2370.             if (b->comp) {

  2371.                 mc_luma_dir(s, s->dsp.mc[3][b->filter][1], s->dst[0], ls_y,

  2372.                             ref2->data[0], ref2->linesize[0],

  2373.                             row << 3, col << 3, &b->mv[0][1], 8, 4, w, h);

  2374.                 mc_luma_dir(s, s->dsp.mc[3][b->filter][1],

  2375.                             s->dst[0] + 4 * ls_y, ls_y,

  2376.                             ref2->data[0], ref2->linesize[0],

  2377.                             (row << 3) + 4, col << 3, &b->mv[2][1], 8, 4, w, h);

  2378.             }

  2379.         } else if (b->bs == BS_4x8) {

  2380.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0], ls_y,

  2381.                         ref1->data[0], ref1->linesize[0],

  2382.                         row << 3, col << 3, &b->mv[0][0], 4, 8, w, h);

  2383.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0] + 4, ls_y,

  2384.                         ref1->data[0], ref1->linesize[0],

  2385.                         row << 3, (col << 3) + 4, &b->mv[1][0], 4, 8, w, h);

  2386. 

  2387.             if (b->comp) {

  2388.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0], ls_y,

  2389.                             ref2->data[0], ref2->linesize[0],

  2390.                             row << 3, col << 3, &b->mv[0][1], 4, 8, w, h);

  2391.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0] + 4, ls_y,

  2392.                             ref2->data[0], ref2->linesize[0],

  2393.                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 8, w, h);

  2394.             }

  2395.         } else {

  2396.             av_assert2(b->bs == BS_4x4);

  2397. 

  2398.             // FIXME if two horizontally adjacent blocks have the same MV,

  2399.             // do a w8 instead of a w4 call

  2400.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0], ls_y,

  2401.                         ref1->data[0], ref1->linesize[0],

  2402.                         row << 3, col << 3, &b->mv[0][0], 4, 4, w, h);

  2403.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0], s->dst[0] + 4, ls_y,

  2404.                         ref1->data[0], ref1->linesize[0],

  2405.                         row << 3, (col << 3) + 4, &b->mv[1][0], 4, 4, w, h);

  2406.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0],

  2407.                         s->dst[0] + 4 * ls_y, ls_y,

  2408.                         ref1->data[0], ref1->linesize[0],

  2409.                         (row << 3) + 4, col << 3, &b->mv[2][0], 4, 4, w, h);

  2410.             mc_luma_dir(s, s->dsp.mc[4][b->filter][0],

  2411.                         s->dst[0] + 4 * ls_y + 4, ls_y,

  2412.                         ref1->data[0], ref1->linesize[0],

  2413.                         (row << 3) + 4, (col << 3) + 4, &b->mv[3][0], 4, 4, w, h);

  2414. 

  2415.             if (b->comp) {

  2416.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0], ls_y,

  2417.                             ref2->data[0], ref2->linesize[0],

  2418.                             row << 3, col << 3, &b->mv[0][1], 4, 4, w, h);

  2419.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1], s->dst[0] + 4, ls_y,

  2420.                             ref2->data[0], ref2->linesize[0],

  2421.                             row << 3, (col << 3) + 4, &b->mv[1][1], 4, 4, w, h);

  2422.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1],

  2423.                             s->dst[0] + 4 * ls_y, ls_y,

  2424.                             ref2->data[0], ref2->linesize[0],

  2425.                             (row << 3) + 4, col << 3, &b->mv[2][1], 4, 4, w, h);

  2426.                 mc_luma_dir(s, s->dsp.mc[4][b->filter][1],

  2427.                             s->dst[0] + 4 * ls_y + 4, ls_y,

  2428.                             ref2->data[0], ref2->linesize[0],

  2429.                             (row << 3) + 4, (col << 3) + 4, &b->mv[3][1], 4, 4, w, h);

  2430.             }

  2431.         }

  2432.     } else {

  2433.         int bwl = bwlog_tab[0][b->bs];

  2434.         int bw = bwh_tab[0][b->bs][0] * 4, bh = bwh_tab[0][b->bs][1] * 4;

  2435. 

  2436.         mc_luma_dir(s, s->dsp.mc[bwl][b->filter][0], s->dst[0], ls_y,

  2437.                     ref1->data[0], ref1->linesize[0],

  2438.                     row << 3, col << 3, &b->mv[0][0],bw, bh, w, h);

  2439. 

  2440.         if (b->comp)

  2441.             mc_luma_dir(s, s->dsp.mc[bwl][b->filter][1], s->dst[0], ls_y,

  2442.                         ref2->data[0], ref2->linesize[0],

  2443.                         row << 3, col << 3, &b->mv[0][1], bw, bh, w, h);

  2444.     }

  2445. 

  2446.     // uv inter pred

  2447.     {

  2448.         int bwl = bwlog_tab[1][b->bs];

  2449.         int bw = bwh_tab[1][b->bs][0] * 4, bh = bwh_tab[1][b->bs][1] * 4;

  2450.         VP56mv mvuv;

  2451. 

  2452.         w = (w + 1) >> 1;

  2453.         h = (h + 1) >> 1;

  2454.         if (b->bs > BS_8x8) {

  2455.             mvuv.x = ROUNDED_DIV(b->mv[0][0].x + b->mv[1][0].x + b->mv[2][0].x + b->mv[3][0].x, 4);

  2456.             mvuv.y = ROUNDED_DIV(b->mv[0][0].y + b->mv[1][0].y + b->mv[2][0].y + b->mv[3][0].y, 4);

  2457.         } else {

  2458.             mvuv = b->mv[0][0];

  2459.         }

  2460. 

  2461.         mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][0],

  2462.                       s->dst[1], s->dst[2], ls_uv,

  2463.                       ref1->data[1], ref1->linesize[1],

  2464.                       ref1->data[2], ref1->linesize[2],

  2465.                       row << 2, col << 2, &mvuv, bw, bh, w, h);

  2466. 

  2467.         if (b->comp) {

  2468.             if (b->bs > BS_8x8) {

  2469.                 mvuv.x = ROUNDED_DIV(b->mv[0][1].x + b->mv[1][1].x + b->mv[2][1].x + b->mv[3][1].x, 4);

  2470.                 mvuv.y = ROUNDED_DIV(b->mv[0][1].y + b->mv[1][1].y + b->mv[2][1].y + b->mv[3][1].y, 4);

  2471.             } else {

  2472.                 mvuv = b->mv[0][1];

  2473.             }

  2474.             mc_chroma_dir(s, s->dsp.mc[bwl][b->filter][1],

  2475.                           s->dst[1], s->dst[2], ls_uv,

  2476.                           ref2->data[1], ref2->linesize[1],

  2477.                           ref2->data[2], ref2->linesize[2],

  2478.                           row << 2, col << 2, &mvuv, bw, bh, w, h);

  2479.         }

  2480.     }

  2481. 

  2482.     if (!b->skip) {

  2483.         /* mostly copied intra_reconn() */

  2484. 

  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];

  2492. 

  2493.         // y itxfm add

  2494.         for (n = 0, y = 0; y < end_y; y += step1d) {

  2495.             uint8_t *ptr = dst;

  2496.             for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d, n += step) {

  2497.                 int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2498. 

  2499.                 if (eob)

  2500.                     s->dsp.itxfm_add[tx][DCT_DCT](ptr, s->y_stride,

  2501.                                                   s->block + 16 * n, eob);

  2502.             }

  2503.             dst += 4 * s->y_stride * step1d;

  2504.         }

  2505. 

  2506.         // uv itxfm add

  2507.         h4 >>= 1;

  2508.         w4 >>= 1;

  2509.         end_x >>= 1;

  2510.         end_y >>= 1;

  2511.         step = 1 << (b->uvtx * 2);

  2512.         for (p = 0; p < 2; p++) {

  2513.             dst = s->dst[p + 1];

  2514.             for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2515.                 uint8_t *ptr = dst;

  2516.                 for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d, n += step) {

  2517.                     int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2518. 

  2519.                     if (eob)

  2520.                         s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2521.                                                         s->uvblock[p] + 16 * n, eob);

  2522.                 }

  2523.                 dst += 4 * uvstep1d * s->uv_stride;

  2524.             }

  2525.         }

  2526.     }

  2527. }

  2528. 

  2529. static av_always_inline void mask_edges(struct VP9Filter *lflvl, int is_uv,

  2530.                                         int row_and_7, int col_and_7,

  2531.                                         int w, int h, int col_end, int row_end,

  2532.                                         enum TxfmMode tx, int skip_inter)

  2533. {

  2534.     // FIXME I'm pretty sure all loops can be replaced by a single LUT if

  2535.     // we make VP9Filter.mask uint64_t (i.e. row/col all single variable)

  2536.     // and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then

  2537.     // use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)

  2538. 

  2539.     // the intended behaviour of the vp9 loopfilter is to work on 8-pixel

  2540.     // edges. This means that for UV, we work on two subsampled blocks at

  2541.     // a time, and we only use the topleft block's mode information to set

  2542.     // things like block strength. Thus, for any block size smaller than

  2543.     // 16x16, ignore the odd portion of the block.

  2544.     if (tx == TX_4X4 && is_uv) {

  2545.         if (h == 1) {

  2546.             if (row_and_7 & 1)

  2547.                 return;

  2548.             if (!row_end)

  2549.                 h += 1;

  2550.         }

  2551.         if (w == 1) {

  2552.             if (col_and_7 & 1)

  2553.                 return;

  2554.             if (!col_end)

  2555.                 w += 1;

  2556.         }

  2557.     }

  2558. 

  2559.     if (tx == TX_4X4 && !skip_inter) {

  2560.         int t = 1 << col_and_7, m_col = (t << w) - t, y;

  2561.         int m_col_odd = (t << (w - 1)) - t;

  2562. 

  2563.         // on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide

  2564.         if (is_uv) {

  2565.             int m_row_8 = m_col & 0x01, m_row_4 = m_col - m_row_8;

  2566. 

  2567.             for (y = row_and_7; y < h + row_and_7; y++) {

  2568.                 int col_mask_id = 2 - !(y & 7);

  2569. 

  2570.                 lflvl->mask[is_uv][0][y][1] |= m_row_8;

  2571.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  2572.                 // for odd lines, if the odd col is not being filtered,

  2573.                 // skip odd row also:

  2574.                 // .---. <-- a

  2575.                 // |   |

  2576.                 // |___| <-- b

  2577.                 // ^   ^

  2578.                 // c   d

  2579.                 //

  2580.                 // if a/c are even row/col and b/d are odd, and d is skipped,

  2581.                 // e.g. right edge of size-66x66.webm, then skip b also (bug)

  2582.                 if ((col_end & 1) && (y & 1)) {

  2583.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col_odd;

  2584.                 } else {

  2585.                     lflvl->mask[is_uv][1][y][col_mask_id] |= m_col;

  2586.                 }

  2587.             }

  2588.         } else {

  2589.             int m_row_8 = m_col & 0x11, m_row_4 = m_col - m_row_8;

  2590. 

  2591.             for (y = row_and_7; y < h + row_and_7; y++) {

  2592.                 int col_mask_id = 2 - !(y & 3);

  2593. 

  2594.                 lflvl->mask[is_uv][0][y][1] |= m_row_8; // row edge

  2595.                 lflvl->mask[is_uv][0][y][2] |= m_row_4;

  2596.                 lflvl->mask[is_uv][1][y][col_mask_id] |= m_col; // col edge

  2597.                 lflvl->mask[is_uv][0][y][3] |= m_col;

  2598.                 lflvl->mask[is_uv][1][y][3] |= m_col;

  2599.             }

  2600.         }

  2601.     } else {

  2602.         int y, t = 1 << col_and_7, m_col = (t << w) - t;

  2603. 

  2604.         if (!skip_inter) {

  2605.             int mask_id = (tx == TX_8X8);

  2606.             int l2 = tx + is_uv - 1, step1d = 1 << l2;

  2607.             static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };

  2608.             int m_row = m_col & masks[l2];

  2609. 

  2610.             // at odd UV col/row edges tx16/tx32 loopfilter edges, force

  2611.             // 8wd loopfilter to prevent going off the visible edge.

  2612.             if (is_uv && tx > TX_8X8 && (w ^ (w - 1)) == 1) {

  2613.                 int m_row_16 = ((t << (w - 1)) - t) & masks[l2];

  2614.                 int m_row_8 = m_row - m_row_16;

  2615. 

  2616.                 for (y = row_and_7; y < h + row_and_7; y++) {

  2617.                     lflvl->mask[is_uv][0][y][0] |= m_row_16;

  2618.                     lflvl->mask[is_uv][0][y][1] |= m_row_8;

  2619.                 }

  2620.             } else {

  2621.                 for (y = row_and_7; y < h + row_and_7; y++)

  2622.                     lflvl->mask[is_uv][0][y][mask_id] |= m_row;

  2623.             }

  2624. 

  2625.             if (is_uv && tx > TX_8X8 && (h ^ (h - 1)) == 1) {

  2626.                 for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)

  2627.                     lflvl->mask[is_uv][1][y][0] |= m_col;

  2628.                 if (y - row_and_7 == h - 1)

  2629.                     lflvl->mask[is_uv][1][y][1] |= m_col;

  2630.             } else {

  2631.                 for (y = row_and_7; y < h + row_and_7; y += step1d)

  2632.                     lflvl->mask[is_uv][1][y][mask_id] |= m_col;

  2633.             }

  2634.         } else if (tx != TX_4X4) {

  2635.             int mask_id;

  2636. 

  2637.             mask_id = (tx == TX_8X8) || (is_uv && h == 1);

  2638.             lflvl->mask[is_uv][1][row_and_7][mask_id] |= m_col;

  2639.             mask_id = (tx == TX_8X8) || (is_uv && w == 1);

  2640.             for (y = row_and_7; y < h + row_and_7; y++)

  2641.                 lflvl->mask[is_uv][0][y][mask_id] |= t;

  2642.         } else if (is_uv) {

  2643.             int t8 = t & 0x01, t4 = t - t8;

  2644. 

  2645.             for (y = row_and_7; y < h + row_and_7; y++) {

  2646.                 lflvl->mask[is_uv][0][y][2] |= t4;

  2647.                 lflvl->mask[is_uv][0][y][1] |= t8;

  2648.             }

  2649.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 7)] |= m_col;

  2650.         } else {

  2651.             int t8 = t & 0x11, t4 = t - t8;

  2652. 

  2653.             for (y = row_and_7; y < h + row_and_7; y++) {

  2654.                 lflvl->mask[is_uv][0][y][2] |= t4;

  2655.                 lflvl->mask[is_uv][0][y][1] |= t8;

  2656.             }

  2657.             lflvl->mask[is_uv][1][row_and_7][2 - !(row_and_7 & 3)] |= m_col;

  2658.         }

  2659.     }

  2660. }

  2661. 

  2662. static int decode_b(AVCodecContext *ctx, int row, int col,

  2663.                     struct VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,

  2664.                     enum BlockLevel bl, enum BlockPartition bp)

  2665. {

  2666.     VP9Context *s = ctx->priv_data;

  2667.     VP9Block *b = s->b;

  2668.     enum BlockSize bs = bl * 3 + bp;

  2669.     int res, y, w4 = bwh_tab[1][bs][0], h4 = bwh_tab[1][bs][1], lvl;

  2670.     int emu[2];

  2671.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2672. 

  2673.     s->row = row;

  2674.     s->row7 = row & 7;

  2675.     s->col = col;

  2676.     s->col7 = col & 7;

  2677.     s->min_mv.x = -(128 + col * 64);

  2678.     s->min_mv.y = -(128 + row * 64);

  2679.     s->max_mv.x = 128 + (s->cols - col - w4) * 64;

  2680.     s->max_mv.y = 128 + (s->rows - row - h4) * 64;

  2681.     b->bs = bs;

  2682.     decode_mode(ctx);

  2683.     b->uvtx = b->tx - (w4 * 2 == (1 << b->tx) || h4 * 2 == (1 << b->tx));

  2684. 

  2685.     if (!b->skip) {

  2686.         if ((res = decode_coeffs(ctx)) < 0)

  2687.             return res;

  2688.     } else {

  2689.         int pl;

  2690. 

  2691.         memset(&s->above_y_nnz_ctx[col * 2], 0, w4 * 2);

  2692.         memset(&s->left_y_nnz_ctx[(row & 7) << 1], 0, h4 * 2);

  2693.         for (pl = 0; pl < 2; pl++) {

  2694.             memset(&s->above_uv_nnz_ctx[pl][col], 0, w4);

  2695.             memset(&s->left_uv_nnz_ctx[pl][row & 7], 0, h4);

  2696.         }

  2697.     }

  2698. 

  2699.     // emulated overhangs if the stride of the target buffer can't hold. This

  2700.     // allows to support emu-edge and so on even if we have large block

  2701.     // overhangs

  2702.     emu[0] = (col + w4) * 8 > f->linesize[0] ||

  2703.              (row + h4) > s->rows + 2 * !(ctx->flags & CODEC_FLAG_EMU_EDGE);

  2704.     emu[1] = (col + w4) * 4 > f->linesize[1] ||

  2705.              (row + h4) > s->rows + 2 * !(ctx->flags & CODEC_FLAG_EMU_EDGE);

  2706.     if (emu[0]) {

  2707.         s->dst[0] = s->tmp_y;

  2708.         s->y_stride = 64;

  2709.     } else {

  2710.         s->dst[0] = f->data[0] + yoff;

  2711.         s->y_stride = f->linesize[0];

  2712.     }

  2713.     if (emu[1]) {

  2714.         s->dst[1] = s->tmp_uv[0];

  2715.         s->dst[2] = s->tmp_uv[1];

  2716.         s->uv_stride = 32;

  2717.     } else {

  2718.         s->dst[1] = f->data[1] + uvoff;

  2719.         s->dst[2] = f->data[2] + uvoff;

  2720.         s->uv_stride = f->linesize[1];

  2721.     }

  2722.     if (b->intra) {

  2723.         intra_recon(ctx, yoff, uvoff);

  2724.     } else {

  2725.         inter_recon(ctx);

  2726.     }

  2727.     if (emu[0]) {

  2728.         int w = FFMIN(s->cols - col, w4) * 8, h = FFMIN(s->rows - row, h4) * 8, n, o = 0;

  2729. 

  2730.         for (n = 0; o < w; n++) {

  2731.             int bw = 64 >> n;

  2732. 

  2733.             av_assert2(n <= 4);

  2734.             if (w & bw) {

  2735.                 s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o, f->linesize[0],

  2736.                                          s->tmp_y + o, 64, h, 0, 0);

  2737.                 o += bw;

  2738.             }

  2739.         }

  2740.     }

  2741.     if (emu[1]) {

  2742.         int w = FFMIN(s->cols - col, w4) * 4, h = FFMIN(s->rows - row, h4) * 4, n, o = 0;

  2743. 

  2744.         for (n = 1; o < w; n++) {

  2745.             int bw = 64 >> n;

  2746. 

  2747.             av_assert2(n <= 4);

  2748.             if (w & bw) {

  2749.                 s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o, f->linesize[1],

  2750.                                          s->tmp_uv[0] + o, 32, h, 0, 0);

  2751.                 s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o, f->linesize[2],

  2752.                                          s->tmp_uv[1] + o, 32, h, 0, 0);

  2753.                 o += bw;

  2754.             }

  2755.         }

  2756.     }

  2757. 

  2758.     // pick filter level and find edges to apply filter to

  2759.     if (s->filter.level &&

  2760.         (lvl = s->segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]

  2761.                                                     [b->mode[3] != ZEROMV]) > 0) {

  2762.         int x_end = FFMIN(s->cols - col, w4), y_end = FFMIN(s->rows - row, h4);

  2763.         int skip_inter = !b->intra && b->skip;

  2764. 

  2765.         for (y = 0; y < h4; y++)

  2766.             memset(&lflvl->level[((row & 7) + y) * 8 + (col & 7)], lvl, w4);

  2767.         mask_edges(lflvl, 0, row & 7, col & 7, x_end, y_end, 0, 0, b->tx, skip_inter);

  2768.         mask_edges(lflvl, 1, row & 7, col & 7, x_end, y_end,

  2769.                    s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,

  2770.                    s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,

  2771.                    b->uvtx, skip_inter);

  2772. 

  2773.         if (!s->filter.lim_lut[lvl]) {

  2774.             int sharp = s->filter.sharpness;

  2775.             int limit = lvl;

  2776. 

  2777.             if (sharp > 0) {

  2778.                 limit >>= (sharp + 3) >> 2;

  2779.                 limit = FFMIN(limit, 9 - sharp);

  2780.             }

  2781.             limit = FFMAX(limit, 1);

  2782. 

  2783.             s->filter.lim_lut[lvl] = limit;

  2784.             s->filter.mblim_lut[lvl] = 2 * (lvl + 2) + limit;

  2785.         }

  2786.     }

  2787. 

  2788.     return 0;

  2789. }

  2790. 

  2791. static int decode_sb(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  2792.                      ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  2793. {

  2794.     VP9Context *s = ctx->priv_data;

  2795.     int c = ((s->above_partition_ctx[col] >> (3 - bl)) & 1) |

  2796.             (((s->left_partition_ctx[row & 0x7] >> (3 - bl)) & 1) << 1), res;

  2797.     const uint8_t *p = s->keyframe ? vp9_default_kf_partition_probs[bl][c] :

  2798.                                      s->prob.p.partition[bl][c];

  2799.     enum BlockPartition bp;

  2800.     ptrdiff_t hbs = 4 >> bl;

  2801.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2802.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  2803. 

  2804.     if (bl == BL_8X8) {

  2805.         bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  2806.         res = decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2807.     } else if (col + hbs < s->cols) {

  2808.         if (row + hbs < s->rows) {

  2809.             bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  2810.             switch (bp) {

  2811.             case PARTITION_NONE:

  2812.                 res = decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2813.                 break;

  2814.             case PARTITION_H:

  2815.                 if (!(res = decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp))) {

  2816.                     yoff  += hbs * 8 * y_stride;

  2817.                     uvoff += hbs * 4 * uv_stride;

  2818.                     res = decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, bl, bp);

  2819.                 }

  2820.                 break;

  2821.             case PARTITION_V:

  2822.                 if (!(res = decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp))) {

  2823.                     yoff  += hbs * 8;

  2824.                     uvoff += hbs * 4;

  2825.                     res = decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, bl, bp);

  2826.                 }

  2827.                 break;

  2828.             case PARTITION_SPLIT:

  2829.                 if (!(res = decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1))) {

  2830.                     if (!(res = decode_sb(ctx, row, col + hbs, lflvl,

  2831.                                           yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1))) {

  2832.                         yoff  += hbs * 8 * y_stride;

  2833.                         uvoff += hbs * 4 * uv_stride;

  2834.                         if (!(res = decode_sb(ctx, row + hbs, col, lflvl,

  2835.                                               yoff, uvoff, bl + 1)))

  2836.                             res = decode_sb(ctx, row + hbs, col + hbs, lflvl,

  2837.                                             yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2838.                     }

  2839.                 }

  2840.                 break;

  2841.             default:

  2842.                 av_assert0(0);

  2843.             }

  2844.         } else if (vp56_rac_get_prob_branchy(&s->c, p[1])) {

  2845.             bp = PARTITION_SPLIT;

  2846.             if (!(res = decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1)))

  2847.                 res = decode_sb(ctx, row, col + hbs, lflvl,

  2848.                                 yoff + 8 * hbs, uvoff + 4 * hbs, bl + 1);

  2849.         } else {

  2850.             bp = PARTITION_H;

  2851.             res = decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2852.         }

  2853.     } else if (row + hbs < s->rows) {

  2854.         if (vp56_rac_get_prob_branchy(&s->c, p[2])) {

  2855.             bp = PARTITION_SPLIT;

  2856.             if (!(res = decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1))) {

  2857.                 yoff  += hbs * 8 * y_stride;

  2858.                 uvoff += hbs * 4 * uv_stride;

  2859.                 res = decode_sb(ctx, row + hbs, col, lflvl,

  2860.                                 yoff, uvoff, bl + 1);

  2861.             }

  2862.         } else {

  2863.             bp = PARTITION_V;

  2864.             res = decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  2865.         }

  2866.     } else {

  2867.         bp = PARTITION_SPLIT;

  2868.         res = decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  2869.     }

  2870.     s->counts.partition[bl][c][bp]++;

  2871. 

  2872.     return res;

  2873. }

  2874. 

  2875. static void loopfilter_sb(AVCodecContext *ctx, struct VP9Filter *lflvl,

  2876.                           int row, int col, ptrdiff_t yoff, ptrdiff_t uvoff)

  2877. {

  2878.     VP9Context *s = ctx->priv_data;

  2879.     AVFrame *f = s->frames[CUR_FRAME].tf.f;

  2880.     uint8_t *dst = f->data[0] + yoff, *lvl = lflvl->level;

  2881.     ptrdiff_t ls_y = f->linesize[0], ls_uv = f->linesize[1];

  2882.     int y, x, p;

  2883. 

  2884.     // FIXME in how far can we interleave the v/h loopfilter calls? E.g.

  2885.     // if you think of them as acting on a 8x8 block max, we can interleave

  2886.     // each v/h within the single x loop, but that only works if we work on

  2887.     // 8 pixel blocks, and we won't always do that (we want at least 16px

  2888.     // to use SSE2 optimizations, perhaps 32 for AVX2)

  2889. 

  2890.     // filter edges between columns, Y plane (e.g. block1 | block2)

  2891.     for (y = 0; y < 8; y += 2, dst += 16 * ls_y, lvl += 16) {

  2892.         uint8_t *ptr = dst, *l = lvl, *hmask1 = lflvl->mask[0][0][y];

  2893.         uint8_t *hmask2 = lflvl->mask[0][0][y + 1];

  2894.         unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2], hm13 = hmask1[3];

  2895.         unsigned hm2 = hmask2[1] | hmask2[2], hm23 = hmask2[3];

  2896.         unsigned hm = hm1 | hm2 | hm13 | hm23;

  2897. 

  2898.         for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 8, l++) {

  2899.             if (hm1 & x) {

  2900.                 int L = *l, H = L >> 4;

  2901.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  2902. 

  2903.                 if (col || x > 1) {

  2904.                     if (hmask1[0] & x) {

  2905.                         if (hmask2[0] & x) {

  2906.                             av_assert2(l[8] == L);

  2907.                             s->dsp.loop_filter_16[0](ptr, ls_y, E, I, H);

  2908.                         } else {

  2909.                             s->dsp.loop_filter_8[2][0](ptr, ls_y, E, I, H);

  2910.                         }

  2911.                     } else if (hm2 & x) {

  2912.                         L = l[8];

  2913.                         H |= (L >> 4) << 8;

  2914.                         E |= s->filter.mblim_lut[L] << 8;

  2915.                         I |= s->filter.lim_lut[L] << 8;

  2916.                         s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  2917.                                                [!!(hmask2[1] & x)]

  2918.                                                [0](ptr, ls_y, E, I, H);

  2919.                     } else {

  2920.                         s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  2921.                                             [0](ptr, ls_y, E, I, H);

  2922.                     }

  2923.                 }

  2924.             } else if (hm2 & x) {

  2925.                 int L = l[8], H = L >> 4;

  2926.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  2927. 

  2928.                 if (col || x > 1) {

  2929.                     s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  2930.                                         [0](ptr + 8 * ls_y, ls_y, E, I, H);

  2931.                 }

  2932.             }

  2933.             if (hm13 & x) {

  2934.                 int L = *l, H = L >> 4;

  2935.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  2936. 

  2937.                 if (hm23 & x) {

  2938.                     L = l[8];

  2939.                     H |= (L >> 4) << 8;

  2940.                     E |= s->filter.mblim_lut[L] << 8;

  2941.                     I |= s->filter.lim_lut[L] << 8;

  2942.                     s->dsp.loop_filter_mix2[0][0][0](ptr + 4, ls_y, E, I, H);

  2943.                 } else {

  2944.                     s->dsp.loop_filter_8[0][0](ptr + 4, ls_y, E, I, H);

  2945.                 }

  2946.             } else if (hm23 & x) {

  2947.                 int L = l[8], H = L >> 4;

  2948.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  2949. 

  2950.                 s->dsp.loop_filter_8[0][0](ptr + 8 * ls_y + 4, ls_y, E, I, H);

  2951.             }

  2952.         }

  2953.     }

  2954. 

  2955.     //                                          block1

  2956.     // filter edges between rows, Y plane (e.g. ------)

  2957.     //                                          block2

  2958.     dst = f->data[0] + yoff;

  2959.     lvl = lflvl->level;

  2960.     for (y = 0; y < 8; y++, dst += 8 * ls_y, lvl += 8) {

  2961.         uint8_t *ptr = dst, *l = lvl, *vmask = lflvl->mask[0][1][y];

  2962.         unsigned vm = vmask[0] | vmask[1] | vmask[2], vm3 = vmask[3];

  2963. 

  2964.         for (x = 1; vm & ~(x - 1); x <<= 2, ptr += 16, l += 2) {

  2965.             if (row || y) {

  2966.                 if (vm & x) {

  2967.                     int L = *l, H = L >> 4;

  2968.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  2969. 

  2970.                     if (vmask[0] & x) {

  2971.                         if (vmask[0] & (x << 1)) {

  2972.                             av_assert2(l[1] == L);

  2973.                             s->dsp.loop_filter_16[1](ptr, ls_y, E, I, H);

  2974.                         } else {

  2975.                             s->dsp.loop_filter_8[2][1](ptr, ls_y, E, I, H);

  2976.                         }

  2977.                     } else if (vm & (x << 1)) {

  2978.                         L = l[1];

  2979.                         H |= (L >> 4) << 8;

  2980.                         E |= s->filter.mblim_lut[L] << 8;

  2981.                         I |= s->filter.lim_lut[L] << 8;

  2982.                         s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  2983.                                                [!!(vmask[1] & (x << 1))]

  2984.                                                [1](ptr, ls_y, E, I, H);

  2985.                     } else {

  2986.                         s->dsp.loop_filter_8[!!(vmask[1] & x)]

  2987.                                             [1](ptr, ls_y, E, I, H);

  2988.                     }

  2989.                 } else if (vm & (x << 1)) {

  2990.                     int L = l[1], H = L >> 4;

  2991.                     int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  2992. 

  2993.                     s->dsp.loop_filter_8[!!(vmask[1] & (x << 1))]

  2994.                                         [1](ptr + 8, ls_y, E, I, H);

  2995.                 }

  2996.             }

  2997.             if (vm3 & x) {

  2998.                 int L = *l, H = L >> 4;

  2999.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3000. 

  3001.                 if (vm3 & (x << 1)) {

  3002.                     L = l[1];

  3003.                     H |= (L >> 4) << 8;

  3004.                     E |= s->filter.mblim_lut[L] << 8;

  3005.                     I |= s->filter.lim_lut[L] << 8;

  3006.                     s->dsp.loop_filter_mix2[0][0][1](ptr + ls_y * 4, ls_y, E, I, H);

  3007.                 } else {

  3008.                     s->dsp.loop_filter_8[0][1](ptr + ls_y * 4, ls_y, E, I, H);

  3009.                 }

  3010.             } else if (vm3 & (x << 1)) {

  3011.                 int L = l[1], H = L >> 4;

  3012.                 int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3013. 

  3014.                 s->dsp.loop_filter_8[0][1](ptr + ls_y * 4 + 8, ls_y, E, I, H);

  3015.             }

  3016.         }

  3017.     }

  3018. 

  3019.     // same principle but for U/V planes

  3020.     for (p = 0; p < 2; p++) {

  3021.         lvl = lflvl->level;

  3022.         dst = f->data[1 + p] + uvoff;

  3023.         for (y = 0; y < 8; y += 4, dst += 16 * ls_uv, lvl += 32) {

  3024.             uint8_t *ptr = dst, *l = lvl, *hmask1 = lflvl->mask[1][0][y];

  3025.             uint8_t *hmask2 = lflvl->mask[1][0][y + 2];

  3026.             unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2];

  3027.             unsigned hm2 = hmask2[1] | hmask2[2], hm = hm1 | hm2;

  3028. 

  3029.             for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 4) {

  3030.                 if (col || x > 1) {

  3031.                     if (hm1 & x) {

  3032.                         int L = *l, H = L >> 4;

  3033.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3034. 

  3035.                         if (hmask1[0] & x) {

  3036.                             if (hmask2[0] & x) {

  3037.                                 av_assert2(l[16] == L);

  3038.                                 s->dsp.loop_filter_16[0](ptr, ls_uv, E, I, H);

  3039.                             } else {

  3040.                                 s->dsp.loop_filter_8[2][0](ptr, ls_uv, E, I, H);

  3041.                             }

  3042.                         } else if (hm2 & x) {

  3043.                             L = l[16];

  3044.                             H |= (L >> 4) << 8;

  3045.                             E |= s->filter.mblim_lut[L] << 8;

  3046.                             I |= s->filter.lim_lut[L] << 8;

  3047.                             s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3048.                                                    [!!(hmask2[1] & x)]

  3049.                                                    [0](ptr, ls_uv, E, I, H);

  3050.                         } else {

  3051.                             s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3052.                                                 [0](ptr, ls_uv, E, I, H);

  3053.                         }

  3054.                     } else if (hm2 & x) {

  3055.                         int L = l[16], H = L >> 4;

  3056.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3057. 

  3058.                         s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3059.                                             [0](ptr + 8 * ls_uv, ls_uv, E, I, H);

  3060.                     }

  3061.                 }

  3062.                 if (x & 0xAA)

  3063.                     l += 2;

  3064.             }

  3065.         }

  3066.         lvl = lflvl->level;

  3067.         dst = f->data[1 + p] + uvoff;

  3068.         for (y = 0; y < 8; y++, dst += 4 * ls_uv) {

  3069.             uint8_t *ptr = dst, *l = lvl, *vmask = lflvl->mask[1][1][y];

  3070.             unsigned vm = vmask[0] | vmask[1] | vmask[2];

  3071. 

  3072.             for (x = 1; vm & ~(x - 1); x <<= 4, ptr += 16, l += 4) {

  3073.                 if (row || y) {

  3074.                     if (vm & x) {

  3075.                         int L = *l, H = L >> 4;

  3076.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3077. 

  3078.                         if (vmask[0] & x) {

  3079.                             if (vmask[0] & (x << 2)) {

  3080.                                 av_assert2(l[2] == L);

  3081.                                 s->dsp.loop_filter_16[1](ptr, ls_uv, E, I, H);

  3082.                             } else {

  3083.                                 s->dsp.loop_filter_8[2][1](ptr, ls_uv, E, I, H);

  3084.                             }

  3085.                         } else if (vm & (x << 2)) {

  3086.                             L = l[2];

  3087.                             H |= (L >> 4) << 8;

  3088.                             E |= s->filter.mblim_lut[L] << 8;

  3089.                             I |= s->filter.lim_lut[L] << 8;

  3090.                             s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3091.                                                    [!!(vmask[1] & (x << 2))]

  3092.                                                    [1](ptr, ls_uv, E, I, H);

  3093.                         } else {

  3094.                             s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3095.                                                 [1](ptr, ls_uv, E, I, H);

  3096.                         }

  3097.                     } else if (vm & (x << 2)) {

  3098.                         int L = l[2], H = L >> 4;

  3099.                         int E = s->filter.mblim_lut[L], I = s->filter.lim_lut[L];

  3100. 

  3101.                         s->dsp.loop_filter_8[!!(vmask[1] & (x << 2))]

  3102.                                             [1](ptr + 8, ls_uv, E, I, H);

  3103.                     }

  3104.                 }

  3105.             }

  3106.             if (y & 1)

  3107.                 lvl += 16;

  3108.         }

  3109.     }

  3110. }

  3111. 

  3112. static void set_tile_offset(int *start, int *end, int idx, int log2_n, int n)

  3113. {

  3114.     int sb_start = ( idx      * n) >> log2_n;

  3115.     int sb_end   = ((idx + 1) * n) >> log2_n;

  3116.     *start = FFMIN(sb_start, n) << 3;

  3117.     *end   = FFMIN(sb_end,   n) << 3;

  3118. }

  3119. 

  3120. static av_always_inline void adapt_prob(uint8_t *p, unsigned ct0, unsigned ct1,

  3121.                                         int max_count, int update_factor)

  3122. {

  3123.     unsigned ct = ct0 + ct1, p2, p1;

  3124. 

  3125.     if (!ct)

  3126.         return;

  3127. 

  3128.     p1 = *p;

  3129.     p2 = ((ct0 << 8) + (ct >> 1)) / ct;

  3130.     p2 = av_clip(p2, 1, 255);

  3131.     ct = FFMIN(ct, max_count);

  3132.     update_factor = FASTDIV(update_factor * ct, max_count);

  3133. 

  3134.     // (p1 * (256 - update_factor) + p2 * update_factor + 128) >> 8

  3135.     *p = p1 + (((p2 - p1) * update_factor + 128) >> 8);

  3136. }

  3137. 

  3138. static void adapt_probs(VP9Context *s)

  3139. {

  3140.     int i, j, k, l, m;

  3141.     prob_context *p = &s->prob_ctx[s->framectxid].p;

  3142.     int uf = (s->keyframe || s->intraonly || !s->last_keyframe) ? 112 : 128;

  3143. 

  3144.     // coefficients

  3145.     for (i = 0; i < 4; i++)

  3146.         for (j = 0; j < 2; j++)

  3147.             for (k = 0; k < 2; k++)

  3148.                 for (l = 0; l < 6; l++)

  3149.                     for (m = 0; m < 6; m++) {

  3150.                         uint8_t *pp = s->prob_ctx[s->framectxid].coef[i][j][k][l][m];

  3151.                         unsigned *e = s->counts.eob[i][j][k][l][m];

  3152.                         unsigned *c = s->counts.coef[i][j][k][l][m];

  3153. 

  3154.                         if (l == 0 && m >= 3) // dc only has 3 pt

  3155.                             break;

  3156. 

  3157.                         adapt_prob(&pp[0], e[0], e[1], 24, uf);

  3158.                         adapt_prob(&pp[1], c[0], c[1] + c[2], 24, uf);

  3159.                         adapt_prob(&pp[2], c[1], c[2], 24, uf);

  3160.                     }

  3161. 

  3162.     if (s->keyframe || s->intraonly) {

  3163.         memcpy(p->skip,  s->prob.p.skip,  sizeof(p->skip));

  3164.         memcpy(p->tx32p, s->prob.p.tx32p, sizeof(p->tx32p));

  3165.         memcpy(p->tx16p, s->prob.p.tx16p, sizeof(p->tx16p));

  3166.         memcpy(p->tx8p,  s->prob.p.tx8p,  sizeof(p->tx8p));

  3167.         return;

  3168.     }

  3169. 

  3170.     // skip flag

  3171.     for (i = 0; i < 3; i++)

  3172.         adapt_prob(&p->skip[i], s->counts.skip[i][0], s->counts.skip[i][1], 20, 128);

  3173. 

  3174.     // intra/inter flag

  3175.     for (i = 0; i < 4; i++)

  3176.         adapt_prob(&p->intra[i], s->counts.intra[i][0], s->counts.intra[i][1], 20, 128);

  3177. 

  3178.     // comppred flag

  3179.     if (s->comppredmode == PRED_SWITCHABLE) {

  3180.       for (i = 0; i < 5; i++)

  3181.           adapt_prob(&p->comp[i], s->counts.comp[i][0], s->counts.comp[i][1], 20, 128);

  3182.     }

  3183. 

  3184.     // reference frames

  3185.     if (s->comppredmode != PRED_SINGLEREF) {

  3186.       for (i = 0; i < 5; i++)

  3187.           adapt_prob(&p->comp_ref[i], s->counts.comp_ref[i][0],

  3188.                      s->counts.comp_ref[i][1], 20, 128);

  3189.     }

  3190. 

  3191.     if (s->comppredmode != PRED_COMPREF) {

  3192.       for (i = 0; i < 5; i++) {

  3193.           uint8_t *pp = p->single_ref[i];

  3194.           unsigned (*c)[2] = s->counts.single_ref[i];

  3195. 

  3196.           adapt_prob(&pp[0], c[0][0], c[0][1], 20, 128);

  3197.           adapt_prob(&pp[1], c[1][0], c[1][1], 20, 128);

  3198.       }

  3199.     }

  3200. 

  3201.     // block partitioning

  3202.     for (i = 0; i < 4; i++)

  3203.         for (j = 0; j < 4; j++) {

  3204.             uint8_t *pp = p->partition[i][j];

  3205.             unsigned *c = s->counts.partition[i][j];

  3206. 

  3207.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3208.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3209.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3210.         }

  3211. 

  3212.     // tx size

  3213.     if (s->txfmmode == TX_SWITCHABLE) {

  3214.       for (i = 0; i < 2; i++) {

  3215.           unsigned *c16 = s->counts.tx16p[i], *c32 = s->counts.tx32p[i];

  3216. 

  3217.           adapt_prob(&p->tx8p[i], s->counts.tx8p[i][0], s->counts.tx8p[i][1], 20, 128);

  3218.           adapt_prob(&p->tx16p[i][0], c16[0], c16[1] + c16[2], 20, 128);

  3219.           adapt_prob(&p->tx16p[i][1], c16[1], c16[2], 20, 128);

  3220.           adapt_prob(&p->tx32p[i][0], c32[0], c32[1] + c32[2] + c32[3], 20, 128);

  3221.           adapt_prob(&p->tx32p[i][1], c32[1], c32[2] + c32[3], 20, 128);

  3222.           adapt_prob(&p->tx32p[i][2], c32[2], c32[3], 20, 128);

  3223.       }

  3224.     }

  3225. 

  3226.     // interpolation filter

  3227.     if (s->filtermode == FILTER_SWITCHABLE) {

  3228.         for (i = 0; i < 4; i++) {

  3229.             uint8_t *pp = p->filter[i];

  3230.             unsigned *c = s->counts.filter[i];

  3231. 

  3232.             adapt_prob(&pp[0], c[0], c[1] + c[2], 20, 128);

  3233.             adapt_prob(&pp[1], c[1], c[2], 20, 128);

  3234.         }

  3235.     }

  3236. 

  3237.     // inter modes

  3238.     for (i = 0; i < 7; i++) {

  3239.         uint8_t *pp = p->mv_mode[i];

  3240.         unsigned *c = s->counts.mv_mode[i];

  3241. 

  3242.         adapt_prob(&pp[0], c[2], c[1] + c[0] + c[3], 20, 128);

  3243.         adapt_prob(&pp[1], c[0], c[1] + c[3], 20, 128);

  3244.         adapt_prob(&pp[2], c[1], c[3], 20, 128);

  3245.     }

  3246. 

  3247.     // mv joints

  3248.     {

  3249.         uint8_t *pp = p->mv_joint;

  3250.         unsigned *c = s->counts.mv_joint;

  3251. 

  3252.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3253.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3254.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3255.     }

  3256. 

  3257.     // mv components

  3258.     for (i = 0; i < 2; i++) {

  3259.         uint8_t *pp;

  3260.         unsigned *c, (*c2)[2], sum;

  3261. 

  3262.         adapt_prob(&p->mv_comp[i].sign, s->counts.mv_comp[i].sign[0],

  3263.                    s->counts.mv_comp[i].sign[1], 20, 128);

  3264. 

  3265.         pp = p->mv_comp[i].classes;

  3266.         c = s->counts.mv_comp[i].classes;

  3267.         sum = c[1] + c[2] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9] + c[10];

  3268.         adapt_prob(&pp[0], c[0], sum, 20, 128);

  3269.         sum -= c[1];

  3270.         adapt_prob(&pp[1], c[1], sum, 20, 128);

  3271.         sum -= c[2] + c[3];

  3272.         adapt_prob(&pp[2], c[2] + c[3], sum, 20, 128);

  3273.         adapt_prob(&pp[3], c[2], c[3], 20, 128);

  3274.         sum -= c[4] + c[5];

  3275.         adapt_prob(&pp[4], c[4] + c[5], sum, 20, 128);

  3276.         adapt_prob(&pp[5], c[4], c[5], 20, 128);

  3277.         sum -= c[6];

  3278.         adapt_prob(&pp[6], c[6], sum, 20, 128);

  3279.         adapt_prob(&pp[7], c[7] + c[8], c[9] + c[10], 20, 128);

  3280.         adapt_prob(&pp[8], c[7], c[8], 20, 128);

  3281.         adapt_prob(&pp[9], c[9], c[10], 20, 128);

  3282. 

  3283.         adapt_prob(&p->mv_comp[i].class0, s->counts.mv_comp[i].class0[0],

  3284.                    s->counts.mv_comp[i].class0[1], 20, 128);

  3285.         pp = p->mv_comp[i].bits;

  3286.         c2 = s->counts.mv_comp[i].bits;

  3287.         for (j = 0; j < 10; j++)

  3288.             adapt_prob(&pp[j], c2[j][0], c2[j][1], 20, 128);

  3289. 

  3290.         for (j = 0; j < 2; j++) {

  3291.             pp = p->mv_comp[i].class0_fp[j];

  3292.             c = s->counts.mv_comp[i].class0_fp[j];

  3293.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3294.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3295.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3296.         }

  3297.         pp = p->mv_comp[i].fp;

  3298.         c = s->counts.mv_comp[i].fp;

  3299.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3300.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3301.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3302. 

  3303.         if (s->highprecisionmvs) {

  3304.             adapt_prob(&p->mv_comp[i].class0_hp, s->counts.mv_comp[i].class0_hp[0],

  3305.                        s->counts.mv_comp[i].class0_hp[1], 20, 128);

  3306.             adapt_prob(&p->mv_comp[i].hp, s->counts.mv_comp[i].hp[0],

  3307.                        s->counts.mv_comp[i].hp[1], 20, 128);

  3308.         }

  3309.     }

  3310. 

  3311.     // y intra modes

  3312.     for (i = 0; i < 4; i++) {

  3313.         uint8_t *pp = p->y_mode[i];

  3314.         unsigned *c = s->counts.y_mode[i], sum, s2;

  3315. 

  3316.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3317.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3318.         sum -= c[TM_VP8_PRED];

  3319.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3320.         sum -= c[VERT_PRED];

  3321.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3322.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3323.         sum -= s2;

  3324.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3325.         s2 -= c[HOR_PRED];

  3326.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3327.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3328.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3329.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3330.         sum -= c[VERT_LEFT_PRED];

  3331.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3332.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3333.     }

  3334. 

  3335.     // uv intra modes

  3336.     for (i = 0; i < 10; i++) {

  3337.         uint8_t *pp = p->uv_mode[i];

  3338.         unsigned *c = s->counts.uv_mode[i], sum, s2;

  3339. 

  3340.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3341.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3342.         sum -= c[TM_VP8_PRED];

  3343.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3344.         sum -= c[VERT_PRED];

  3345.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3346.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3347.         sum -= s2;

  3348.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3349.         s2 -= c[HOR_PRED];

  3350.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3351.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3352.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3353.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3354.         sum -= c[VERT_LEFT_PRED];

  3355.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3356.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3357.     }

  3358. }

  3359. 

  3360. static av_cold int vp9_decode_free(AVCodecContext *ctx)

  3361. {

  3362.     VP9Context *s = ctx->priv_data;

  3363.     int i;

  3364. 

  3365.     for (i = 0; i < 2; i++) {

  3366.         if (s->frames[i].tf.f->data[0])

  3367.             vp9_unref_frame(ctx, &s->frames[i]);

  3368.         av_frame_free(&s->frames[i].tf.f);

  3369.     }

  3370.     for (i = 0; i < 8; i++) {

  3371.         if (s->refs[i].f->data[0])

  3372.             ff_thread_release_buffer(ctx, &s->refs[i]);

  3373.         av_frame_free(&s->refs[i].f);

  3374.         if (s->next_refs[i].f->data[0])

  3375.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3376.         av_frame_free(&s->next_refs[i].f);

  3377.     }

  3378.     av_freep(&s->above_partition_ctx);

  3379.     av_freep(&s->c_b);

  3380.     av_freep(&s->b_base);

  3381.     av_freep(&s->block_base);

  3382. 

  3383.     return 0;

  3384. }

  3385. 

  3386. 

  3387. static int vp9_decode_frame(AVCodecContext *ctx, AVFrame *frame,

  3388.                             int *got_frame, AVPacket *pkt)

  3389. {

  3390.     const uint8_t *data = pkt->data;

  3391.     int size = pkt->size;

  3392.     VP9Context *s = ctx->priv_data;

  3393.     int res, tile_row, tile_col, i, ref, row, col;

  3394.     ptrdiff_t yoff = 0, uvoff = 0, ls_y, ls_uv;

  3395.     AVFrame *f;

  3396. 

  3397.     if ((res = decode_frame_header(ctx, data, size, &ref)) < 0) {

  3398.         return res;

  3399.     } else if (res == 0) {

  3400.         if (!s->refs[ref].f->data[0]) {

  3401.             av_log(ctx, AV_LOG_ERROR, "Requested reference %d not available\n", ref);

  3402.             return AVERROR_INVALIDDATA;

  3403.         }

  3404.         if ((res = av_frame_ref(frame, s->refs[ref].f)) < 0)

  3405.             return res;

  3406.         *got_frame = 1;

  3407.         return 0;

  3408.     }

  3409.     data += res;

  3410.     size -= res;

  3411. 

  3412.     if (s->frames[LAST_FRAME].tf.f->data[0])

  3413.         vp9_unref_frame(ctx, &s->frames[LAST_FRAME]);

  3414.     if (!s->keyframe && s->frames[CUR_FRAME].tf.f->data[0] &&

  3415.         (res = vp9_ref_frame(ctx, &s->frames[LAST_FRAME], &s->frames[CUR_FRAME])) < 0)

  3416.         return res;

  3417.     if (s->frames[CUR_FRAME].tf.f->data[0])

  3418.         vp9_unref_frame(ctx, &s->frames[CUR_FRAME]);

  3419.     if ((res = vp9_alloc_frame(ctx, &s->frames[CUR_FRAME])) < 0)

  3420.         return res;

  3421.     f = s->frames[CUR_FRAME].tf.f;

  3422.     f->key_frame = s->keyframe;

  3423.     f->pict_type = s->keyframe ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;

  3424.     ls_y = f->linesize[0];

  3425.     ls_uv =f->linesize[1];

  3426. 

  3427.     // ref frame setup

  3428.     for (i = 0; i < 8; i++)

  3429.         if (s->refreshrefmask & (1 << i)) {

  3430.             if (s->next_refs[i].f->data[0])

  3431.                 ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3432.             if ((res = ff_thread_ref_frame(&s->next_refs[i], &s->frames[CUR_FRAME].tf)) < 0)

  3433.                 return res;

  3434.         }

  3435. 

  3436.     // main tile decode loop

  3437.     memset(s->above_partition_ctx, 0, s->cols);

  3438.     memset(s->above_skip_ctx, 0, s->cols);

  3439.     if (s->keyframe || s->intraonly) {

  3440.         memset(s->above_mode_ctx, DC_PRED, s->cols * 2);

  3441.     } else {

  3442.         memset(s->above_mode_ctx, NEARESTMV, s->cols);

  3443.     }

  3444.     memset(s->above_y_nnz_ctx, 0, s->sb_cols * 16);

  3445.     memset(s->above_uv_nnz_ctx[0], 0, s->sb_cols * 8);

  3446.     memset(s->above_uv_nnz_ctx[1], 0, s->sb_cols * 8);

  3447.     memset(s->above_segpred_ctx, 0, s->cols);

  3448.     s->b = s->b_base;

  3449.     s->block = s->block_base;

  3450.     s->uvblock[0] = s->uvblock_base[0];

  3451.     s->uvblock[1] = s->uvblock_base[1];

  3452.     s->eob = s->eob_base;

  3453.     s->uveob[0] = s->uveob_base[0];

  3454.     s->uveob[1] = s->uveob_base[1];

  3455.     for (tile_row = 0; tile_row < s->tiling.tile_rows; tile_row++) {

  3456.         set_tile_offset(&s->tiling.tile_row_start, &s->tiling.tile_row_end,

  3457.                         tile_row, s->tiling.log2_tile_rows, s->sb_rows);

  3458.         for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  3459.             unsigned tile_size;

  3460. 

  3461.             if (tile_col == s->tiling.tile_cols - 1 &&

  3462.                 tile_row == s->tiling.tile_rows - 1) {

  3463.                 tile_size = size;

  3464.             } else {

  3465.                 tile_size = AV_RB32(data);

  3466.                 data += 4;

  3467.                 size -= 4;

  3468.             }

  3469.             if (tile_size > size)

  3470.                 return AVERROR_INVALIDDATA;

  3471.             ff_vp56_init_range_decoder(&s->c_b[tile_col], data, tile_size);

  3472.             if (vp56_rac_get_prob_branchy(&s->c_b[tile_col], 128)) // marker bit

  3473.                 return AVERROR_INVALIDDATA;

  3474.             data += tile_size;

  3475.             size -= tile_size;

  3476.         }

  3477. 

  3478.         for (row = s->tiling.tile_row_start; row < s->tiling.tile_row_end;

  3479.              row += 8, yoff += ls_y * 64, uvoff += ls_uv * 32) {

  3480.             struct VP9Filter *lflvl_ptr = s->lflvl;

  3481.             ptrdiff_t yoff2 = yoff, uvoff2 = uvoff;

  3482. 

  3483.             for (tile_col = 0; tile_col < s->tiling.tile_cols; tile_col++) {

  3484.                 set_tile_offset(&s->tiling.tile_col_start, &s->tiling.tile_col_end,

  3485.                                 tile_col, s->tiling.log2_tile_cols, s->sb_cols);

  3486. 

  3487.                 memset(s->left_partition_ctx, 0, 8);

  3488.                 memset(s->left_skip_ctx, 0, 8);

  3489.                 if (s->keyframe || s->intraonly) {

  3490.                     memset(s->left_mode_ctx, DC_PRED, 16);

  3491.                 } else {

  3492.                     memset(s->left_mode_ctx, NEARESTMV, 8);

  3493.                 }

  3494.                 memset(s->left_y_nnz_ctx, 0, 16);

  3495.                 memset(s->left_uv_nnz_ctx, 0, 16);

  3496.                 memset(s->left_segpred_ctx, 0, 8);

  3497. 

  3498.                 memcpy(&s->c, &s->c_b[tile_col], sizeof(s->c));

  3499.                 for (col = s->tiling.tile_col_start;

  3500.                      col < s->tiling.tile_col_end;

  3501.                      col += 8, yoff2 += 64, uvoff2 += 32, lflvl_ptr++) {

  3502.                     // FIXME integrate with lf code (i.e. zero after each

  3503.                     // use, similar to invtxfm coefficients, or similar)

  3504.                     memset(lflvl_ptr->mask, 0, sizeof(lflvl_ptr->mask));

  3505. 

  3506.                     if ((res = decode_sb(ctx, row, col, lflvl_ptr,

  3507.                                          yoff2, uvoff2, BL_64X64)) < 0)

  3508.                         return res;

  3509.                 }

  3510.                 memcpy(&s->c_b[tile_col], &s->c, sizeof(s->c));

  3511.             }

  3512. 

  3513.             // backup pre-loopfilter reconstruction data for intra

  3514.             // prediction of next row of sb64s

  3515.             if (row + 8 < s->rows) {

  3516.                 memcpy(s->intra_pred_data[0],

  3517.                        f->data[0] + yoff + 63 * ls_y,

  3518.                        8 * s->cols);

  3519.                 memcpy(s->intra_pred_data[1],

  3520.                        f->data[1] + uvoff + 31 * ls_uv,

  3521.                        4 * s->cols);

  3522.                 memcpy(s->intra_pred_data[2],

  3523.                        f->data[2] + uvoff + 31 * ls_uv,

  3524.                        4 * s->cols);

  3525.             }

  3526. 

  3527.             // loopfilter one row

  3528.             if (s->filter.level) {

  3529.                 yoff2 = yoff;

  3530.                 uvoff2 = uvoff;

  3531.                 lflvl_ptr = s->lflvl;

  3532.                 for (col = 0; col < s->cols;

  3533.                      col += 8, yoff2 += 64, uvoff2 += 32, lflvl_ptr++) {

  3534.                     loopfilter_sb(ctx, lflvl_ptr, row, col, yoff2, uvoff2);

  3535.                 }

  3536.             }

  3537.         }

  3538.     }

  3539. 

  3540.     // bw adaptivity (or in case of parallel decoding mode, fw adaptivity

  3541.     // probability maintenance between frames)

  3542.     if (s->refreshctx) {

  3543.         if (s->parallelmode) {

  3544.             int j, k, l, m;

  3545. 

  3546.             for (i = 0; i < 4; i++)

  3547.                 for (j = 0; j < 2; j++)

  3548.                     for (k = 0; k < 2; k++)

  3549.                         for (l = 0; l < 6; l++)

  3550.                             for (m = 0; m < 6; m++)

  3551.                                 memcpy(s->prob_ctx[s->framectxid].coef[i][j][k][l][m],

  3552.                                        s->prob.coef[i][j][k][l][m], 3);

  3553.             s->prob_ctx[s->framectxid].p = s->prob.p;

  3554.         } else {

  3555.             adapt_probs(s);

  3556.         }

  3557.     }

  3558. 

  3559.     // ref frame setup

  3560.     for (i = 0; i < 8; i++) {

  3561.         if (s->refs[i].f->data[0])

  3562.             ff_thread_release_buffer(ctx, &s->refs[i]);

  3563.         ff_thread_ref_frame(&s->refs[i], &s->next_refs[i]);

  3564.     }

  3565. 

  3566.     if (!s->invisible) {

  3567.         if ((res = av_frame_ref(frame, s->frames[CUR_FRAME].tf.f)) < 0)

  3568.             return res;

  3569.         *got_frame = 1;

  3570.     }

  3571. 

  3572.     return 0;

  3573. }

  3574. 

  3575. static void vp9_decode_flush(AVCodecContext *ctx)

  3576. {

  3577.     VP9Context *s = ctx->priv_data;

  3578.     int i;

  3579. 

  3580.     for (i = 0; i < 2; i++)

  3581.         vp9_unref_frame(ctx, &s->frames[i]);

  3582.     for (i = 0; i < 8; i++)

  3583.         ff_thread_release_buffer(ctx, &s->refs[i]);

  3584. }

  3585. 

  3586. static int init_frames(AVCodecContext *ctx)

  3587. {

  3588.     VP9Context *s = ctx->priv_data;

  3589.     int i;

  3590. 

  3591.     for (i = 0; i < 2; i++) {

  3592.         s->frames[i].tf.f = av_frame_alloc();

  3593.         if (!s->frames[i].tf.f) {

  3594.             vp9_decode_free(ctx);

  3595.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  3596.             return AVERROR(ENOMEM);

  3597.         }

  3598.     }

  3599.     for (i = 0; i < 8; i++) {

  3600.         s->refs[i].f = av_frame_alloc();

  3601.         s->next_refs[i].f = av_frame_alloc();

  3602.         if (!s->refs[i].f || !s->next_refs[i].f) {

  3603.             vp9_decode_free(ctx);

  3604.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  3605.             return AVERROR(ENOMEM);

  3606.         }

  3607.     }

  3608. 

  3609.     return 0;

  3610. }

  3611. 

  3612. static av_cold int vp9_decode_init(AVCodecContext *ctx)

  3613. {

  3614.     VP9Context *s = ctx->priv_data;

  3615. 

  3616.     ctx->pix_fmt = AV_PIX_FMT_YUV420P;

  3617.     ff_vp9dsp_init(&s->dsp);

  3618.     ff_videodsp_init(&s->vdsp, 8);

  3619.     s->filter.sharpness = -1;

  3620. 

  3621.     return init_frames(ctx);

  3622. }

  3623. 

  3624. AVCodec ff_vp9_decoder = {

  3625.   .name                  = "vp9",

  3626.   .long_name             = NULL_IF_CONFIG_SMALL("Google VP9"),

  3627.   .type                  = AVMEDIA_TYPE_VIDEO,

  3628.   .id                    = AV_CODEC_ID_VP9,

  3629.   .priv_data_size        = sizeof(VP9Context),

  3630.   .init                  = vp9_decode_init,

  3631.   .close                 = vp9_decode_free,

  3632.   .capabilities          = CODEC_CAP_DR1,

  3633.   .flush                 = vp9_decode_flush,

  3634.     .decode                = vp9_decode_frame,

  3635. };