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.
77087 Commits
32 Branches
646MB
Tree: cd1b7e2bd7
Branches Tags
${ item.name }
Create branch ${ searchTerm }
from 'cd1b7e2bd7'
${ noResults }
FFmpeg/libavcodec/vp9.c

4296 lines
174KB
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. #include "libavutil/pixdesc.h"

  35. 

  36. #define VP9_SYNCCODE 0x498342

  37. 

  38. struct VP9Filter {

  39.     uint8_t level[8 * 8];

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

  41.                               [8 /* rows */][4 /* 0=16, 1=8, 2=4, 3=inner4 */];

  42. };

  43. 

  44. typedef struct VP9Block {

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

  46.     enum FilterMode filter;

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

  48.     enum BlockSize bs;

  49.     enum TxfmMode tx, uvtx;

  50.     enum BlockLevel bl;

  51.     enum BlockPartition bp;

  52. } VP9Block;

  53. 

  54. typedef struct VP9Context {

  55.     VP9SharedContext s;

  56. 

  57.     VP9DSPContext dsp;

  58.     VideoDSPContext vdsp;

  59.     GetBitContext gb;

  60.     VP56RangeCoder c;

  61.     VP56RangeCoder *c_b;

  62.     unsigned c_b_size;

  63.     VP9Block *b_base, *b;

  64.     int pass;

  65.     int row, row7, col, col7;

  66.     uint8_t *dst[3];

  67.     ptrdiff_t y_stride, uv_stride;

  68. 

  69.     uint8_t ss_h, ss_v;

  70.     uint8_t last_bpp, bpp, bpp_index, bytesperpixel;

  71.     uint8_t last_keyframe;

  72.     enum AVPixelFormat pix_fmt, last_fmt;

  73.     ThreadFrame next_refs[8];

  74. 

  75.     struct {

  76.         uint8_t lim_lut[64];

  77.         uint8_t mblim_lut[64];

  78.     } filter_lut;

  79.     unsigned tile_row_start, tile_row_end, tile_col_start, tile_col_end;

  80.     unsigned sb_cols, sb_rows, rows, cols;

  81.     struct {

  82.         prob_context p;

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

  84.     } prob_ctx[4];

  85.     struct {

  86.         prob_context p;

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

  88.     } prob;

  89.     struct {

  90.         unsigned y_mode[4][10];

  91.         unsigned uv_mode[10][10];

  92.         unsigned filter[4][3];

  93.         unsigned mv_mode[7][4];

  94.         unsigned intra[4][2];

  95.         unsigned comp[5][2];

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

  97.         unsigned comp_ref[5][2];

  98.         unsigned tx32p[2][4];

  99.         unsigned tx16p[2][3];

  100.         unsigned tx8p[2][2];

  101.         unsigned skip[3][2];

  102.         unsigned mv_joint[4];

  103.         struct {

  104.             unsigned sign[2];

  105.             unsigned classes[11];

  106.             unsigned class0[2];

  107.             unsigned bits[10][2];

  108.             unsigned class0_fp[2][4];

  109.             unsigned fp[4];

  110.             unsigned class0_hp[2];

  111.             unsigned hp[2];

  112.         } mv_comp[2];

  113.         unsigned partition[4][4][4];

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

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

  116.     } counts;

  117. 

  118.     // contextual (left/above) cache

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

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

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

  122.     DECLARE_ALIGNED(16, uint8_t, left_uv_nnz_ctx)[2][16];

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

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

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

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

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

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

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

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

  131.     uint8_t *above_partition_ctx;

  132.     uint8_t *above_mode_ctx;

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

  134.     uint8_t *above_y_nnz_ctx;

  135.     uint8_t *above_uv_nnz_ctx[2];

  136.     uint8_t *above_skip_ctx; // 1bit

  137.     uint8_t *above_txfm_ctx; // 2bit

  138.     uint8_t *above_segpred_ctx; // 1bit

  139.     uint8_t *above_intra_ctx; // 1bit

  140.     uint8_t *above_comp_ctx; // 1bit

  141.     uint8_t *above_ref_ctx; // 2bit

  142.     uint8_t *above_filter_ctx;

  143.     VP56mv (*above_mv_ctx)[2];

  144. 

  145.     // whole-frame cache

  146.     uint8_t *intra_pred_data[3];

  147.     struct VP9Filter *lflvl;

  148.     DECLARE_ALIGNED(32, uint8_t, edge_emu_buffer)[135 * 144 * 2];

  149. 

  150.     // block reconstruction intermediates

  151.     int block_alloc_using_2pass;

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

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

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

  155.     DECLARE_ALIGNED(32, uint8_t, tmp_y)[64 * 64 * 2];

  156.     DECLARE_ALIGNED(32, uint8_t, tmp_uv)[2][64 * 64 * 2];

  157.     uint16_t mvscale[3][2];

  158.     uint8_t mvstep[3][2];

  159. } VP9Context;

  160. 

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

  162.     {

  163.         { 16, 16 }, { 16, 8 }, { 8, 16 }, { 8, 8 }, { 8, 4 }, { 4, 8 },

  164.         { 4, 4 }, { 4, 2 }, { 2, 4 }, { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 },

  165.     }, {

  166.         { 8, 8 }, { 8, 4 }, { 4, 8 }, { 4, 4 }, { 4, 2 }, { 2, 4 },

  167.         { 2, 2 }, { 2, 1 }, { 1, 2 }, { 1, 1 }, { 1, 1 }, { 1, 1 }, { 1, 1 },

  168.     }

  169. };

  170. 

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

  172. {

  173.     VP9Context *s = ctx->priv_data;

  174.     int ret, sz;

  175. 

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

  177.         return ret;

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

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

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

  181.         return AVERROR(ENOMEM);

  182.     }

  183. 

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

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

  186. 

  187.     return 0;

  188. }

  189. 

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

  191. {

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

  193.     av_buffer_unref(&f->extradata);

  194.     f->segmentation_map = NULL;

  195. }

  196. 

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

  198. {

  199.     int res;

  200. 

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

  202.         return res;

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

  204.         vp9_unref_frame(ctx, dst);

  205.         return AVERROR(ENOMEM);

  206.     }

  207. 

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

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

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

  211. 

  212.     return 0;

  213. }

  214. 

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

  216. {

  217.     VP9Context *s = ctx->priv_data;

  218.     uint8_t *p;

  219.     int bytesperpixel = s->bytesperpixel, res;

  220. 

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

  222. 

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

  224.         return 0;

  225. 

  226.     if ((res = ff_set_dimensions(ctx, w, h)) < 0)

  227.         return res;

  228.     s->last_fmt  = ctx->pix_fmt = s->pix_fmt;

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

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

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

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

  233. 

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

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

  236.     // FIXME we slightly over-allocate here for subsampled chroma, but a little

  237.     // bit of padding shouldn't affect performance...

  238.     p = av_malloc(s->sb_cols * (128 + 192 * bytesperpixel +

  239.                                 sizeof(*s->lflvl) + 16 * sizeof(*s->above_mv_ctx)));

  240.     if (!p)

  241.         return AVERROR(ENOMEM);

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

  243.     assign(s->intra_pred_data[1],  uint8_t *,             64 * bytesperpixel);

  244.     assign(s->intra_pred_data[2],  uint8_t *,             64 * bytesperpixel);

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

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

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

  248.     assign(s->above_uv_nnz_ctx[0], uint8_t *,             16);

  249.     assign(s->above_uv_nnz_ctx[1], uint8_t *,             16);

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

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

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

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

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

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

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

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

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

  259. #undef assign

  260. 

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

  262.     av_freep(&s->b_base);

  263.     av_freep(&s->block_base);

  264. 

  265.     if (s->bpp != s->last_bpp) {

  266.         ff_vp9dsp_init(&s->dsp, s->bpp, ctx->flags & AV_CODEC_FLAG_BITEXACT);

  267.         ff_videodsp_init(&s->vdsp, s->bpp);

  268.         s->last_bpp = s->bpp;

  269.     }

  270. 

  271.     return 0;

  272. }

  273. 

  274. static int update_block_buffers(AVCodecContext *ctx)

  275. {

  276.     VP9Context *s = ctx->priv_data;

  277.     int chroma_blocks, chroma_eobs, bytesperpixel = s->bytesperpixel;

  278. 

  279.     if (s->b_base && s->block_base && s->block_alloc_using_2pass == s->s.frames[CUR_FRAME].uses_2pass)

  280.         return 0;

  281. 

  282.     av_free(s->b_base);

  283.     av_free(s->block_base);

  284.     chroma_blocks = 64 * 64 >> (s->ss_h + s->ss_v);

  285.     chroma_eobs   = 16 * 16 >> (s->ss_h + s->ss_v);

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

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

  288. 

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

  290.         s->block_base = av_mallocz(((64 * 64 + 2 * chroma_blocks) * bytesperpixel * sizeof(int16_t) +

  291.                                     16 * 16 + 2 * chroma_eobs) * sbs);

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

  293.             return AVERROR(ENOMEM);

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

  295.         s->uvblock_base[1] = s->uvblock_base[0] + sbs * chroma_blocks * bytesperpixel;

  296.         s->eob_base = (uint8_t *) (s->uvblock_base[1] + sbs * chroma_blocks * bytesperpixel);

  297.         s->uveob_base[0] = s->eob_base + 16 * 16 * sbs;

  298.         s->uveob_base[1] = s->uveob_base[0] + chroma_eobs * sbs;

  299.     } else {

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

  301.         s->block_base = av_mallocz((64 * 64 + 2 * chroma_blocks) * bytesperpixel * sizeof(int16_t) +

  302.                                    16 * 16 + 2 * chroma_eobs);

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

  304.             return AVERROR(ENOMEM);

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

  306.         s->uvblock_base[1] = s->uvblock_base[0] + chroma_blocks * bytesperpixel;

  307.         s->eob_base = (uint8_t *) (s->uvblock_base[1] + chroma_blocks * bytesperpixel);

  308.         s->uveob_base[0] = s->eob_base + 16 * 16;

  309.         s->uveob_base[1] = s->uveob_base[0] + chroma_eobs;

  310.     }

  311.     s->block_alloc_using_2pass = s->s.frames[CUR_FRAME].uses_2pass;

  312. 

  313.     return 0;

  314. }

  315. 

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

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

  318. {

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

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

  321. }

  322. 

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

  324. {

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

  326. }

  327. 

  328. // differential forward probability updates

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

  330. {

  331.     static const int inv_map_table[255] = {

  332.           7,  20,  33,  46,  59,  72,  85,  98, 111, 124, 137, 150, 163, 176,

  333.         189, 202, 215, 228, 241, 254,   1,   2,   3,   4,   5,   6,   8,   9,

  334.          10,  11,  12,  13,  14,  15,  16,  17,  18,  19,  21,  22,  23,  24,

  335.          25,  26,  27,  28,  29,  30,  31,  32,  34,  35,  36,  37,  38,  39,

  336.          40,  41,  42,  43,  44,  45,  47,  48,  49,  50,  51,  52,  53,  54,

  337.          55,  56,  57,  58,  60,  61,  62,  63,  64,  65,  66,  67,  68,  69,

  338.          70,  71,  73,  74,  75,  76,  77,  78,  79,  80,  81,  82,  83,  84,

  339.          86,  87,  88,  89,  90,  91,  92,  93,  94,  95,  96,  97,  99, 100,

  340.         101, 102, 103, 104, 105, 106, 107, 108, 109, 110, 112, 113, 114, 115,

  341.         116, 117, 118, 119, 120, 121, 122, 123, 125, 126, 127, 128, 129, 130,

  342.         131, 132, 133, 134, 135, 136, 138, 139, 140, 141, 142, 143, 144, 145,

  343.         146, 147, 148, 149, 151, 152, 153, 154, 155, 156, 157, 158, 159, 160,

  344.         161, 162, 164, 165, 166, 167, 168, 169, 170, 171, 172, 173, 174, 175,

  345.         177, 178, 179, 180, 181, 182, 183, 184, 185, 186, 187, 188, 190, 191,

  346.         192, 193, 194, 195, 196, 197, 198, 199, 200, 201, 203, 204, 205, 206,

  347.         207, 208, 209, 210, 211, 212, 213, 214, 216, 217, 218, 219, 220, 221,

  348.         222, 223, 224, 225, 226, 227, 229, 230, 231, 232, 233, 234, 235, 236,

  349.         237, 238, 239, 240, 242, 243, 244, 245, 246, 247, 248, 249, 250, 251,

  350.         252, 253, 253,

  351.     };

  352.     int d;

  353. 

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

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

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

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

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

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

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

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

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

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

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

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

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

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

  368. 

  369.     if (!vp8_rac_get(c)) {

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

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

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

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

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

  375.     } else {

  376.         d = vp8_rac_get_uint(c, 7);

  377.         if (d >= 65)

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

  379.         d += 64;

  380.         av_assert2(d < FF_ARRAY_ELEMS(inv_map_table));

  381.     }

  382. 

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

  384.                     255 - inv_recenter_nonneg(inv_map_table[d], 255 - p);

  385. }

  386. 

  387. static int read_colorspace_details(AVCodecContext *ctx)

  388. {

  389.     static const enum AVColorSpace colorspaces[8] = {

  390.         AVCOL_SPC_UNSPECIFIED, AVCOL_SPC_BT470BG, AVCOL_SPC_BT709, AVCOL_SPC_SMPTE170M,

  391.         AVCOL_SPC_SMPTE240M, AVCOL_SPC_BT2020_NCL, AVCOL_SPC_RESERVED, AVCOL_SPC_RGB,

  392.     };

  393.     VP9Context *s = ctx->priv_data;

  394.     int bits = ctx->profile <= 1 ? 0 : 1 + get_bits1(&s->gb); // 0:8, 1:10, 2:12

  395. 

  396.     s->bpp_index = bits;

  397.     s->bpp = 8 + bits * 2;

  398.     s->bytesperpixel = (7 + s->bpp) >> 3;

  399.     ctx->colorspace = colorspaces[get_bits(&s->gb, 3)];

  400.     if (ctx->colorspace == AVCOL_SPC_RGB) { // RGB = profile 1

  401.         static const enum AVPixelFormat pix_fmt_rgb[3] = {

  402.             AV_PIX_FMT_GBRP, AV_PIX_FMT_GBRP10, AV_PIX_FMT_GBRP12

  403.         };

  404.         s->ss_h = s->ss_v = 0;

  405.         ctx->color_range = AVCOL_RANGE_JPEG;

  406.         s->pix_fmt = pix_fmt_rgb[bits];

  407.         if (ctx->profile & 1) {

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

  409.                 av_log(ctx, AV_LOG_ERROR, "Reserved bit set in RGB\n");

  410.                 return AVERROR_INVALIDDATA;

  411.             }

  412.         } else {

  413.             av_log(ctx, AV_LOG_ERROR, "RGB not supported in profile %d\n",

  414.                    ctx->profile);

  415.             return AVERROR_INVALIDDATA;

  416.         }

  417.     } else {

  418.         static const enum AVPixelFormat pix_fmt_for_ss[3][2 /* v */][2 /* h */] = {

  419.             { { AV_PIX_FMT_YUV444P, AV_PIX_FMT_YUV422P },

  420.               { AV_PIX_FMT_YUV440P, AV_PIX_FMT_YUV420P } },

  421.             { { AV_PIX_FMT_YUV444P10, AV_PIX_FMT_YUV422P10 },

  422.               { AV_PIX_FMT_YUV440P10, AV_PIX_FMT_YUV420P10 } },

  423.             { { AV_PIX_FMT_YUV444P12, AV_PIX_FMT_YUV422P12 },

  424.               { AV_PIX_FMT_YUV440P12, AV_PIX_FMT_YUV420P12 } }

  425.         };

  426.         ctx->color_range = get_bits1(&s->gb) ? AVCOL_RANGE_JPEG : AVCOL_RANGE_MPEG;

  427.         if (ctx->profile & 1) {

  428.             s->ss_h = get_bits1(&s->gb);

  429.             s->ss_v = get_bits1(&s->gb);

  430.             s->pix_fmt = pix_fmt_for_ss[bits][s->ss_v][s->ss_h];

  431.             if (s->pix_fmt == AV_PIX_FMT_YUV420P) {

  432.                 av_log(ctx, AV_LOG_ERROR, "YUV 4:2:0 not supported in profile %d\n",

  433.                        ctx->profile);

  434.                 return AVERROR_INVALIDDATA;

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

  436.                 av_log(ctx, AV_LOG_ERROR, "Profile %d color details reserved bit set\n",

  437.                        ctx->profile);

  438.                 return AVERROR_INVALIDDATA;

  439.             }

  440.         } else {

  441.             s->ss_h = s->ss_v = 1;

  442.             s->pix_fmt = pix_fmt_for_ss[bits][1][1];

  443.         }

  444.     }

  445. 

  446.     return 0;

  447. }

  448. 

  449. static int decode_frame_header(AVCodecContext *ctx,

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

  451. {

  452.     VP9Context *s = ctx->priv_data;

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

  454.     int last_invisible;

  455.     const uint8_t *data2;

  456. 

  457.     /* general header */

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

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

  460.         return res;

  461.     }

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

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

  464.         return AVERROR_INVALIDDATA;

  465.     }

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

  467.     ctx->profile |= get_bits1(&s->gb) << 1;

  468.     if (ctx->profile == 3) ctx->profile += get_bits1(&s->gb);

  469.     if (ctx->profile > 3) {

  470.         av_log(ctx, AV_LOG_ERROR, "Profile %d is not yet supported\n", ctx->profile);

  471.         return AVERROR_INVALIDDATA;

  472.     }

  473.     s->s.h.profile = ctx->profile;

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

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

  476.         return 0;

  477.     }

  478.     s->last_keyframe  = s->s.h.keyframe;

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

  480.     last_invisible    = s->s.h.invisible;

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

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

  483.     s->s.h.use_last_frame_mvs = !s->s.h.errorres && !last_invisible;

  484.     if (s->s.h.keyframe) {

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

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

  487.             return AVERROR_INVALIDDATA;

  488.         }

  489.         if ((res = read_colorspace_details(ctx)) < 0)

  490.             return res;

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

  492.         s->s.h.refreshrefmask = 0xff;

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

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

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

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

  497.     } else {

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

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

  500.         if (s->s.h.intraonly) {

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

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

  503.                 return AVERROR_INVALIDDATA;

  504.             }

  505.             if (ctx->profile >= 1) {

  506.                 if ((res = read_colorspace_details(ctx)) < 0)

  507.                     return res;

  508.             } else {

  509.                 s->ss_h = s->ss_v = 1;

  510.                 s->bpp = 8;

  511.                 s->bpp_index = 0;

  512.                 s->bytesperpixel = 1;

  513.                 s->pix_fmt = AV_PIX_FMT_YUV420P;

  514.                 ctx->colorspace = AVCOL_SPC_BT470BG;

  515.                 ctx->color_range = AVCOL_RANGE_JPEG;

  516.             }

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

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

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

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

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

  522.         } else {

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

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

  525.             s->s.h.signbias[0]    = get_bits1(&s->gb) && !s->s.h.errorres;

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

  527.             s->s.h.signbias[1]    = get_bits1(&s->gb) && !s->s.h.errorres;

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

  529.             s->s.h.signbias[2]    = get_bits1(&s->gb) && !s->s.h.errorres;

  530.             if (!s->s.refs[s->s.h.refidx[0]].f->buf[0] ||

  531.                 !s->s.refs[s->s.h.refidx[1]].f->buf[0] ||

  532.                 !s->s.refs[s->s.h.refidx[2]].f->buf[0]) {

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

  534.                 return AVERROR_INVALIDDATA;

  535.             }

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

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

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

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

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

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

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

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

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

  545.             } else {

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

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

  548.             }

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

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

  551.             // the _last_ frame

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

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

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

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

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

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

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

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

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

  561.             if (s->s.h.allowcompinter) {

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

  563.                     s->s.h.fixcompref    = 2;

  564.                     s->s.h.varcompref[0] = 0;

  565.                     s->s.h.varcompref[1] = 1;

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

  567.                     s->s.h.fixcompref    = 1;

  568.                     s->s.h.varcompref[0] = 0;

  569.                     s->s.h.varcompref[1] = 2;

  570.                 } else {

  571.                     s->s.h.fixcompref    = 0;

  572.                     s->s.h.varcompref[0] = 1;

  573.                     s->s.h.varcompref[1] = 2;

  574.                 }

  575.             }

  576. 

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

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

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

  580. 

  581.                 if (ref->format != s->pix_fmt) {

  582.                     av_log(ctx, AV_LOG_ERROR,

  583.                            "Ref pixfmt (%s) did not match current frame (%s)",

  584.                            av_get_pix_fmt_name(ref->format),

  585.                            av_get_pix_fmt_name(s->pix_fmt));

  586.                     return AVERROR_INVALIDDATA;

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

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

  589.                 } else {

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

  591.                         av_log(ctx, AV_LOG_ERROR,

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

  593.                                refw, refh, w, h);

  594.                         return AVERROR_INVALIDDATA;

  595.                     }

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

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

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

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

  600.                 }

  601.             }

  602.         }

  603.     }

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

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

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

  607. 

  608.     /* loopfilter header data */

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

  610.         // reset loopfilter defaults

  611.         s->s.h.lf_delta.ref[0] = 1;

  612.         s->s.h.lf_delta.ref[1] = 0;

  613.         s->s.h.lf_delta.ref[2] = -1;

  614.         s->s.h.lf_delta.ref[3] = -1;

  615.         s->s.h.lf_delta.mode[0] = 0;

  616.         s->s.h.lf_delta.mode[1] = 0;

  617.         memset(s->s.h.segmentation.feat, 0, sizeof(s->s.h.segmentation.feat));

  618.     }

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

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

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

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

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

  624.         memset(s->filter_lut.lim_lut, 0, sizeof(s->filter_lut.lim_lut));

  625.     s->s.h.filter.sharpness = sharp;

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

  627.         if ((s->s.h.lf_delta.updated = get_bits1(&s->gb))) {

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

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

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

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

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

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

  634.         }

  635.     }

  636. 

  637.     /* quantization header data */

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

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

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

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

  642.     s->s.h.lossless    = s->s.h.yac_qi == 0 && s->s.h.ydc_qdelta == 0 &&

  643.                        s->s.h.uvdc_qdelta == 0 && s->s.h.uvac_qdelta == 0;

  644.     if (s->s.h.lossless)

  645.         ctx->properties |= FF_CODEC_PROPERTY_LOSSLESS;

  646. 

  647.     /* segmentation header info */

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

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

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

  651.                 s->s.h.segmentation.prob[i] = get_bits1(&s->gb) ?

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

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

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

  655.                     s->s.h.segmentation.pred_prob[i] = get_bits1(&s->gb) ?

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

  657.             }

  658.         }

  659. 

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

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

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

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

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

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

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

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

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

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

  670.             }

  671.         }

  672.     }

  673. 

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

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

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

  677. 

  678.         if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[i].q_enabled) {

  679.             if (s->s.h.segmentation.absolute_vals)

  680.                 qyac = av_clip_uintp2(s->s.h.segmentation.feat[i].q_val, 8);

  681.             else

  682.                 qyac = av_clip_uintp2(s->s.h.yac_qi + s->s.h.segmentation.feat[i].q_val, 8);

  683.         } else {

  684.             qyac  = s->s.h.yac_qi;

  685.         }

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

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

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

  689.         qyac  = av_clip_uintp2(qyac, 8);

  690. 

  691.         s->s.h.segmentation.feat[i].qmul[0][0] = vp9_dc_qlookup[s->bpp_index][qydc];

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

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

  694.         s->s.h.segmentation.feat[i].qmul[1][1] = vp9_ac_qlookup[s->bpp_index][quvac];

  695. 

  696.         sh = s->s.h.filter.level >= 32;

  697.         if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[i].lf_enabled) {

  698.             if (s->s.h.segmentation.absolute_vals)

  699.                 lflvl = av_clip_uintp2(s->s.h.segmentation.feat[i].lf_val, 6);

  700.             else

  701.                 lflvl = av_clip_uintp2(s->s.h.filter.level + s->s.h.segmentation.feat[i].lf_val, 6);

  702.         } else {

  703.             lflvl  = s->s.h.filter.level;

  704.         }

  705.         if (s->s.h.lf_delta.enabled) {

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

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

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

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

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

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

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

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

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

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

  716.             }

  717.         } else {

  718.             memset(s->s.h.segmentation.feat[i].lflvl, lflvl,

  719.                    sizeof(s->s.h.segmentation.feat[i].lflvl));

  720.         }

  721.     }

  722. 

  723.     /* tiling info */

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

  725.         av_log(ctx, AV_LOG_ERROR, "Failed to initialize decoder for %dx%d @ %d\n",

  726.                w, h, s->pix_fmt);

  727.         return res;

  728.     }

  729.     for (s->s.h.tiling.log2_tile_cols = 0;

  730.          s->sb_cols > (64 << s->s.h.tiling.log2_tile_cols);

  731.          s->s.h.tiling.log2_tile_cols++) ;

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

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

  734.     while (max > s->s.h.tiling.log2_tile_cols) {

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

  736.             s->s.h.tiling.log2_tile_cols++;

  737.         else

  738.             break;

  739.     }

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

  741.     s->s.h.tiling.tile_rows = 1 << s->s.h.tiling.log2_tile_rows;

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

  743.         s->s.h.tiling.tile_cols = 1 << s->s.h.tiling.log2_tile_cols;

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

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

  746.         if (!s->c_b) {

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

  748.             return AVERROR(ENOMEM);

  749.         }

  750.     }

  751. 

  752.     if (s->s.h.keyframe || s->s.h.errorres || (s->s.h.intraonly && s->s.h.resetctx == 3)) {

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

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

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

  756.                sizeof(vp9_default_coef_probs));

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

  758.                sizeof(vp9_default_coef_probs));

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

  760.                sizeof(vp9_default_coef_probs));

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

  762.                sizeof(vp9_default_coef_probs));

  763.     } else if (s->s.h.intraonly && s->s.h.resetctx == 2) {

  764.         s->prob_ctx[c].p = vp9_default_probs;

  765.         memcpy(s->prob_ctx[c].coef, vp9_default_coef_probs,

  766.                sizeof(vp9_default_coef_probs));

  767.     }

  768. 

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

  770.     s->s.h.compressed_header_size = size2 = get_bits(&s->gb, 16);

  771.     s->s.h.uncompressed_header_size = (get_bits_count(&s->gb) + 7) / 8;

  772. 

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

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

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

  776.         return AVERROR_INVALIDDATA;

  777.     }

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

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

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

  781.         return AVERROR_INVALIDDATA;

  782.     }

  783. 

  784.     if (s->s.h.keyframe || s->s.h.intraonly) {

  785.         memset(s->counts.coef, 0, sizeof(s->counts.coef));

  786.         memset(s->counts.eob,  0, sizeof(s->counts.eob));

  787.     } else {

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

  789.     }

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

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

  792.     // fw update)?

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

  794. 

  795.     // txfm updates

  796.     if (s->s.h.lossless) {

  797.         s->s.h.txfmmode = TX_4X4;

  798.     } else {

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

  800.         if (s->s.h.txfmmode == 3)

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

  802. 

  803.         if (s->s.h.txfmmode == TX_SWITCHABLE) {

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

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

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

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

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

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

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

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

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

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

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

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

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

  817.         }

  818.     }

  819. 

  820.     // coef updates

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

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

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

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

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

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

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

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

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

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

  831.                                 break;

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

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

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

  835.                                 } else {

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

  837.                                 }

  838.                             }

  839.                             p[3] = 0;

  840.                         }

  841.         } else {

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

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

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

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

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

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

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

  849.                                 break;

  850.                             memcpy(p, r, 3);

  851.                             p[3] = 0;

  852.                         }

  853.         }

  854.         if (s->s.h.txfmmode == i)

  855.             break;

  856.     }

  857. 

  858.     // mode updates

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

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

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

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

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

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

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

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

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

  868. 

  869.         if (s->s.h.filtermode == FILTER_SWITCHABLE)

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

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

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

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

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

  875. 

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

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

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

  879. 

  880.         if (s->s.h.allowcompinter) {

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

  882.             if (s->s.h.comppredmode)

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

  884.             if (s->s.h.comppredmode == PRED_SWITCHABLE)

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

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

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

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

  889.         } else {

  890.             s->s.h.comppredmode = PRED_SINGLEREF;

  891.         }

  892. 

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

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

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

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

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

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

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

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

  901.             }

  902.         }

  903. 

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

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

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

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

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

  909.         }

  910. 

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

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

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

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

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

  916. 

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

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

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

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

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

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

  923. 

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

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

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

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

  928. 

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

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

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

  932. 

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

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

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

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

  937. 

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

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

  940. 

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

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

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

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

  945.         }

  946. 

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

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

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

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

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

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

  953. 

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

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

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

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

  958.         }

  959. 

  960.         if (s->s.h.highprecisionmvs) {

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

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

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

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

  965. 

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

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

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

  969.             }

  970.         }

  971.     }

  972. 

  973.     return (data2 - data) + size2;

  974. }

  975. 

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

  977.                                       VP9Context *s)

  978. {

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

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

  981. }

  982. 

  983. static void find_ref_mvs(VP9Context *s,

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

  985. {

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

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

  988.                       { -1, -1 }, {  0, -1 }, { -1,  0 }, {  6, -1 }},

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

  990.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, {  2, -1 }},

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

  992.                       { -1, -1 }, { -3,  0 }, {  0, -3 }, { -1,  2 }},

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

  994.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  996.                       { -1,  1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

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

  998.                       {  1, -1 }, { -3,  0 }, {  0, -3 }, { -3, -3 }},

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

  1000.                       { -1, -1 }, {  0, -3 }, { -3,  0 }, { -3, -3 }},

  1001.         [BS_16x8]  = {{  0, -1 }, { -1,  0 }, {  1, -1 }, { -1, -1 },

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

  1003.         [BS_8x16]  = {{ -1,  0 }, {  0, -1 }, { -1,  1 }, { -1, -1 },

  1004.                       { -2,  0 }, {  0, -2 }, { -1, -2 }, { -2, -1 }},

  1005.         [BS_8x8]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1006.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1007.         [BS_8x4]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1008.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1009.         [BS_4x8]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1010.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1011.         [BS_4x4]   = {{  0, -1 }, { -1,  0 }, { -1, -1 }, {  0, -2 },

  1012.                       { -2,  0 }, { -1, -2 }, { -2, -1 }, { -2, -2 }},

  1013.     };

  1014.     VP9Block *b = s->b;

  1015.     int row = s->row, col = s->col, row7 = s->row7;

  1016.     const int8_t (*p)[2] = mv_ref_blk_off[b->bs];

  1017. #define INVALID_MV 0x80008000U

  1018.     uint32_t mem = INVALID_MV, mem_sub8x8 = INVALID_MV;

  1019.     int i;

  1020. 

  1021. #define RETURN_DIRECT_MV(mv) \

  1022.     do { \

  1023.         uint32_t m = AV_RN32A(&mv); \

  1024.         if (!idx) { \

  1025.             AV_WN32A(pmv, m); \

  1026.             return; \

  1027.         } else if (mem == INVALID_MV) { \

  1028.             mem = m; \

  1029.         } else if (m != mem) { \

  1030.             AV_WN32A(pmv, m); \

  1031.             return; \

  1032.         } \

  1033.     } while (0)

  1034. 

  1035.     if (sb >= 0) {

  1036.         if (sb == 2 || sb == 1) {

  1037.             RETURN_DIRECT_MV(b->mv[0][z]);

  1038.         } else if (sb == 3) {

  1039.             RETURN_DIRECT_MV(b->mv[2][z]);

  1040.             RETURN_DIRECT_MV(b->mv[1][z]);

  1041.             RETURN_DIRECT_MV(b->mv[0][z]);

  1042.         }

  1043. 

  1044. #define RETURN_MV(mv) \

  1045.     do { \

  1046.         if (sb > 0) { \

  1047.             VP56mv tmp; \

  1048.             uint32_t m; \

  1049.             av_assert2(idx == 1); \

  1050.             av_assert2(mem != INVALID_MV); \

  1051.             if (mem_sub8x8 == INVALID_MV) { \

  1052.                 clamp_mv(&tmp, &mv, s); \

  1053.                 m = AV_RN32A(&tmp); \

  1054.                 if (m != mem) { \

  1055.                     AV_WN32A(pmv, m); \

  1056.                     return; \

  1057.                 } \

  1058.                 mem_sub8x8 = AV_RN32A(&mv); \

  1059.             } else if (mem_sub8x8 != AV_RN32A(&mv)) { \

  1060.                 clamp_mv(&tmp, &mv, s); \

  1061.                 m = AV_RN32A(&tmp); \

  1062.                 if (m != mem) { \

  1063.                     AV_WN32A(pmv, m); \

  1064.                 } else { \

  1065.                     /* BUG I'm pretty sure this isn't the intention */ \

  1066.                     AV_WN32A(pmv, 0); \

  1067.                 } \

  1068.                 return; \

  1069.             } \

  1070.         } else { \

  1071.             uint32_t m = AV_RN32A(&mv); \

  1072.             if (!idx) { \

  1073.                 clamp_mv(pmv, &mv, s); \

  1074.                 return; \

  1075.             } else if (mem == INVALID_MV) { \

  1076.                 mem = m; \

  1077.             } else if (m != mem) { \

  1078.                 clamp_mv(pmv, &mv, s); \

  1079.                 return; \

  1080.             } \

  1081.         } \

  1082.     } while (0)

  1083. 

  1084.         if (row > 0) {

  1085.             struct VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[(row - 1) * s->sb_cols * 8 + col];

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

  1087.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][0]);

  1088.             } else if (mv->ref[1] == ref) {

  1089.                 RETURN_MV(s->above_mv_ctx[2 * col + (sb & 1)][1]);

  1090.             }

  1091.         }

  1092.         if (col > s->tile_col_start) {

  1093.             struct VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[row * s->sb_cols * 8 + col - 1];

  1094.             if (mv->ref[0] == ref) {

  1095.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][0]);

  1096.             } else if (mv->ref[1] == ref) {

  1097.                 RETURN_MV(s->left_mv_ctx[2 * row7 + (sb >> 1)][1]);

  1098.             }

  1099.         }

  1100.         i = 2;

  1101.     } else {

  1102.         i = 0;

  1103.     }

  1104. 

  1105.     // previously coded MVs in this neighbourhood, using same reference frame

  1106.     for (; i < 8; i++) {

  1107.         int c = p[i][0] + col, r = p[i][1] + row;

  1108. 

  1109.         if (c >= s->tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1110.             struct VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1111. 

  1112.             if (mv->ref[0] == ref) {

  1113.                 RETURN_MV(mv->mv[0]);

  1114.             } else if (mv->ref[1] == ref) {

  1115.                 RETURN_MV(mv->mv[1]);

  1116.             }

  1117.         }

  1118.     }

  1119. 

  1120.     // MV at this position in previous frame, using same reference frame

  1121.     if (s->s.h.use_last_frame_mvs) {

  1122.         struct VP9mvrefPair *mv = &s->s.frames[REF_FRAME_MVPAIR].mv[row * s->sb_cols * 8 + col];

  1123. 

  1124.         if (!s->s.frames[REF_FRAME_MVPAIR].uses_2pass)

  1125.             ff_thread_await_progress(&s->s.frames[REF_FRAME_MVPAIR].tf, row >> 3, 0);

  1126.         if (mv->ref[0] == ref) {

  1127.             RETURN_MV(mv->mv[0]);

  1128.         } else if (mv->ref[1] == ref) {

  1129.             RETURN_MV(mv->mv[1]);

  1130.         }

  1131.     }

  1132. 

  1133. #define RETURN_SCALE_MV(mv, scale) \

  1134.     do { \

  1135.         if (scale) { \

  1136.             VP56mv mv_temp = { -mv.x, -mv.y }; \

  1137.             RETURN_MV(mv_temp); \

  1138.         } else { \

  1139.             RETURN_MV(mv); \

  1140.         } \

  1141.     } while (0)

  1142. 

  1143.     // previously coded MVs in this neighbourhood, using different reference frame

  1144.     for (i = 0; i < 8; i++) {

  1145.         int c = p[i][0] + col, r = p[i][1] + row;

  1146. 

  1147.         if (c >= s->tile_col_start && c < s->cols && r >= 0 && r < s->rows) {

  1148.             struct VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[r * s->sb_cols * 8 + c];

  1149. 

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

  1151.                 RETURN_SCALE_MV(mv->mv[0], s->s.h.signbias[mv->ref[0]] != s->s.h.signbias[ref]);

  1152.             }

  1153.             if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1154.                 // BUG - libvpx has this condition regardless of whether

  1155.                 // we used the first ref MV and pre-scaling

  1156.                 AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1157.                 RETURN_SCALE_MV(mv->mv[1], s->s.h.signbias[mv->ref[1]] != s->s.h.signbias[ref]);

  1158.             }

  1159.         }

  1160.     }

  1161. 

  1162.     // MV at this position in previous frame, using different reference frame

  1163.     if (s->s.h.use_last_frame_mvs) {

  1164.         struct VP9mvrefPair *mv = &s->s.frames[REF_FRAME_MVPAIR].mv[row * s->sb_cols * 8 + col];

  1165. 

  1166.         // no need to await_progress, because we already did that above

  1167.         if (mv->ref[0] != ref && mv->ref[0] >= 0) {

  1168.             RETURN_SCALE_MV(mv->mv[0], s->s.h.signbias[mv->ref[0]] != s->s.h.signbias[ref]);

  1169.         }

  1170.         if (mv->ref[1] != ref && mv->ref[1] >= 0 &&

  1171.             // BUG - libvpx has this condition regardless of whether

  1172.             // we used the first ref MV and pre-scaling

  1173.             AV_RN32A(&mv->mv[0]) != AV_RN32A(&mv->mv[1])) {

  1174.             RETURN_SCALE_MV(mv->mv[1], s->s.h.signbias[mv->ref[1]] != s->s.h.signbias[ref]);

  1175.         }

  1176.     }

  1177. 

  1178.     AV_ZERO32(pmv);

  1179.     clamp_mv(pmv, pmv, s);

  1180. #undef INVALID_MV

  1181. #undef RETURN_MV

  1182. #undef RETURN_SCALE_MV

  1183. }

  1184. 

  1185. static av_always_inline int read_mv_component(VP9Context *s, int idx, int hp)

  1186. {

  1187.     int bit, sign = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].sign);

  1188.     int n, c = vp8_rac_get_tree(&s->c, vp9_mv_class_tree,

  1189.                                 s->prob.p.mv_comp[idx].classes);

  1190. 

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

  1192.     s->counts.mv_comp[idx].classes[c]++;

  1193.     if (c) {

  1194.         int m;

  1195. 

  1196.         for (n = 0, m = 0; m < c; m++) {

  1197.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].bits[m]);

  1198.             n |= bit << m;

  1199.             s->counts.mv_comp[idx].bits[m][bit]++;

  1200.         }

  1201.         n <<= 3;

  1202.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree, s->prob.p.mv_comp[idx].fp);

  1203.         n |= bit << 1;

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

  1205.         if (hp) {

  1206.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].hp);

  1207.             s->counts.mv_comp[idx].hp[bit]++;

  1208.             n |= bit;

  1209.         } else {

  1210.             n |= 1;

  1211.             // bug in libvpx - we count for bw entropy purposes even if the

  1212.             // bit wasn't coded

  1213.             s->counts.mv_comp[idx].hp[1]++;

  1214.         }

  1215.         n += 8 << c;

  1216.     } else {

  1217.         n = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0);

  1218.         s->counts.mv_comp[idx].class0[n]++;

  1219.         bit = vp8_rac_get_tree(&s->c, vp9_mv_fp_tree,

  1220.                                s->prob.p.mv_comp[idx].class0_fp[n]);

  1221.         s->counts.mv_comp[idx].class0_fp[n][bit]++;

  1222.         n = (n << 3) | (bit << 1);

  1223.         if (hp) {

  1224.             bit = vp56_rac_get_prob(&s->c, s->prob.p.mv_comp[idx].class0_hp);

  1225.             s->counts.mv_comp[idx].class0_hp[bit]++;

  1226.             n |= bit;

  1227.         } else {

  1228.             n |= 1;

  1229.             // bug in libvpx - we count for bw entropy purposes even if the

  1230.             // bit wasn't coded

  1231.             s->counts.mv_comp[idx].class0_hp[1]++;

  1232.         }

  1233.     }

  1234. 

  1235.     return sign ? -(n + 1) : (n + 1);

  1236. }

  1237. 

  1238. static void fill_mv(VP9Context *s,

  1239.                     VP56mv *mv, int mode, int sb)

  1240. {

  1241.     VP9Block *b = s->b;

  1242. 

  1243.     if (mode == ZEROMV) {

  1244.         AV_ZERO64(mv);

  1245.     } else {

  1246.         int hp;

  1247. 

  1248.         // FIXME cache this value and reuse for other subblocks

  1249.         find_ref_mvs(s, &mv[0], b->ref[0], 0, mode == NEARMV,

  1250.                      mode == NEWMV ? -1 : sb);

  1251.         // FIXME maybe move this code into find_ref_mvs()

  1252.         if ((mode == NEWMV || sb == -1) &&

  1253.             !(hp = s->s.h.highprecisionmvs && abs(mv[0].x) < 64 && abs(mv[0].y) < 64)) {

  1254.             if (mv[0].y & 1) {

  1255.                 if (mv[0].y < 0)

  1256.                     mv[0].y++;

  1257.                 else

  1258.                     mv[0].y--;

  1259.             }

  1260.             if (mv[0].x & 1) {

  1261.                 if (mv[0].x < 0)

  1262.                     mv[0].x++;

  1263.                 else

  1264.                     mv[0].x--;

  1265.             }

  1266.         }

  1267.         if (mode == NEWMV) {

  1268.             enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1269.                                               s->prob.p.mv_joint);

  1270. 

  1271.             s->counts.mv_joint[j]++;

  1272.             if (j >= MV_JOINT_V)

  1273.                 mv[0].y += read_mv_component(s, 0, hp);

  1274.             if (j & 1)

  1275.                 mv[0].x += read_mv_component(s, 1, hp);

  1276.         }

  1277. 

  1278.         if (b->comp) {

  1279.             // FIXME cache this value and reuse for other subblocks

  1280.             find_ref_mvs(s, &mv[1], b->ref[1], 1, mode == NEARMV,

  1281.                          mode == NEWMV ? -1 : sb);

  1282.             if ((mode == NEWMV || sb == -1) &&

  1283.                 !(hp = s->s.h.highprecisionmvs && abs(mv[1].x) < 64 && abs(mv[1].y) < 64)) {

  1284.                 if (mv[1].y & 1) {

  1285.                     if (mv[1].y < 0)

  1286.                         mv[1].y++;

  1287.                     else

  1288.                         mv[1].y--;

  1289.                 }

  1290.                 if (mv[1].x & 1) {

  1291.                     if (mv[1].x < 0)

  1292.                         mv[1].x++;

  1293.                     else

  1294.                         mv[1].x--;

  1295.                 }

  1296.             }

  1297.             if (mode == NEWMV) {

  1298.                 enum MVJoint j = vp8_rac_get_tree(&s->c, vp9_mv_joint_tree,

  1299.                                                   s->prob.p.mv_joint);

  1300. 

  1301.                 s->counts.mv_joint[j]++;

  1302.                 if (j >= MV_JOINT_V)

  1303.                     mv[1].y += read_mv_component(s, 0, hp);

  1304.                 if (j & 1)

  1305.                     mv[1].x += read_mv_component(s, 1, hp);

  1306.             }

  1307.         }

  1308.     }

  1309. }

  1310. 

  1311. static av_always_inline void setctx_2d(uint8_t *ptr, int w, int h,

  1312.                                        ptrdiff_t stride, int v)

  1313. {

  1314.     switch (w) {

  1315.     case 1:

  1316.         do {

  1317.             *ptr = v;

  1318.             ptr += stride;

  1319.         } while (--h);

  1320.         break;

  1321.     case 2: {

  1322.         int v16 = v * 0x0101;

  1323.         do {

  1324.             AV_WN16A(ptr, v16);

  1325.             ptr += stride;

  1326.         } while (--h);

  1327.         break;

  1328.     }

  1329.     case 4: {

  1330.         uint32_t v32 = v * 0x01010101;

  1331.         do {

  1332.             AV_WN32A(ptr, v32);

  1333.             ptr += stride;

  1334.         } while (--h);

  1335.         break;

  1336.     }

  1337.     case 8: {

  1338. #if HAVE_FAST_64BIT

  1339.         uint64_t v64 = v * 0x0101010101010101ULL;

  1340.         do {

  1341.             AV_WN64A(ptr, v64);

  1342.             ptr += stride;

  1343.         } while (--h);

  1344. #else

  1345.         uint32_t v32 = v * 0x01010101;

  1346.         do {

  1347.             AV_WN32A(ptr,     v32);

  1348.             AV_WN32A(ptr + 4, v32);

  1349.             ptr += stride;

  1350.         } while (--h);

  1351. #endif

  1352.         break;

  1353.     }

  1354.     }

  1355. }

  1356. 

  1357. static void decode_mode(AVCodecContext *ctx)

  1358. {

  1359.     static const uint8_t left_ctx[N_BS_SIZES] = {

  1360.         0x0, 0x8, 0x0, 0x8, 0xc, 0x8, 0xc, 0xe, 0xc, 0xe, 0xf, 0xe, 0xf

  1361.     };

  1362.     static const uint8_t above_ctx[N_BS_SIZES] = {

  1363.         0x0, 0x0, 0x8, 0x8, 0x8, 0xc, 0xc, 0xc, 0xe, 0xe, 0xe, 0xf, 0xf

  1364.     };

  1365.     static const uint8_t max_tx_for_bl_bp[N_BS_SIZES] = {

  1366.         TX_32X32, TX_32X32, TX_32X32, TX_32X32, TX_16X16, TX_16X16,

  1367.         TX_16X16, TX_8X8, TX_8X8, TX_8X8, TX_4X4, TX_4X4, TX_4X4

  1368.     };

  1369.     VP9Context *s = ctx->priv_data;

  1370.     VP9Block *b = s->b;

  1371.     int row = s->row, col = s->col, row7 = s->row7;

  1372.     enum TxfmMode max_tx = max_tx_for_bl_bp[b->bs];

  1373.     int bw4 = bwh_tab[1][b->bs][0], w4 = FFMIN(s->cols - col, bw4);

  1374.     int bh4 = bwh_tab[1][b->bs][1], h4 = FFMIN(s->rows - row, bh4), y;

  1375.     int have_a = row > 0, have_l = col > s->tile_col_start;

  1376.     int vref, filter_id;

  1377. 

  1378.     if (!s->s.h.segmentation.enabled) {

  1379.         b->seg_id = 0;

  1380.     } else if (s->s.h.keyframe || s->s.h.intraonly) {

  1381.         b->seg_id = !s->s.h.segmentation.update_map ? 0 :

  1382.                     vp8_rac_get_tree(&s->c, vp9_segmentation_tree, s->s.h.segmentation.prob);

  1383.     } else if (!s->s.h.segmentation.update_map ||

  1384.                (s->s.h.segmentation.temporal &&

  1385.                 vp56_rac_get_prob_branchy(&s->c,

  1386.                     s->s.h.segmentation.pred_prob[s->above_segpred_ctx[col] +

  1387.                                     s->left_segpred_ctx[row7]]))) {

  1388.         if (!s->s.h.errorres && s->s.frames[REF_FRAME_SEGMAP].segmentation_map) {

  1389.             int pred = 8, x;

  1390.             uint8_t *refsegmap = s->s.frames[REF_FRAME_SEGMAP].segmentation_map;

  1391. 

  1392.             if (!s->s.frames[REF_FRAME_SEGMAP].uses_2pass)

  1393.                 ff_thread_await_progress(&s->s.frames[REF_FRAME_SEGMAP].tf, row >> 3, 0);

  1394.             for (y = 0; y < h4; y++) {

  1395.                 int idx_base = (y + row) * 8 * s->sb_cols + col;

  1396.                 for (x = 0; x < w4; x++)

  1397.                     pred = FFMIN(pred, refsegmap[idx_base + x]);

  1398.             }

  1399.             av_assert1(pred < 8);

  1400.             b->seg_id = pred;

  1401.         } else {

  1402.             b->seg_id = 0;

  1403.         }

  1404. 

  1405.         memset(&s->above_segpred_ctx[col], 1, w4);

  1406.         memset(&s->left_segpred_ctx[row7], 1, h4);

  1407.     } else {

  1408.         b->seg_id = vp8_rac_get_tree(&s->c, vp9_segmentation_tree,

  1409.                                      s->s.h.segmentation.prob);

  1410. 

  1411.         memset(&s->above_segpred_ctx[col], 0, w4);

  1412.         memset(&s->left_segpred_ctx[row7], 0, h4);

  1413.     }

  1414.     if (s->s.h.segmentation.enabled &&

  1415.         (s->s.h.segmentation.update_map || s->s.h.keyframe || s->s.h.intraonly)) {

  1416.         setctx_2d(&s->s.frames[CUR_FRAME].segmentation_map[row * 8 * s->sb_cols + col],

  1417.                   bw4, bh4, 8 * s->sb_cols, b->seg_id);

  1418.     }

  1419. 

  1420.     b->skip = s->s.h.segmentation.enabled &&

  1421.         s->s.h.segmentation.feat[b->seg_id].skip_enabled;

  1422.     if (!b->skip) {

  1423.         int c = s->left_skip_ctx[row7] + s->above_skip_ctx[col];

  1424.         b->skip = vp56_rac_get_prob(&s->c, s->prob.p.skip[c]);

  1425.         s->counts.skip[c][b->skip]++;

  1426.     }

  1427. 

  1428.     if (s->s.h.keyframe || s->s.h.intraonly) {

  1429.         b->intra = 1;

  1430.     } else if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[b->seg_id].ref_enabled) {

  1431.         b->intra = !s->s.h.segmentation.feat[b->seg_id].ref_val;

  1432.     } else {

  1433.         int c, bit;

  1434. 

  1435.         if (have_a && have_l) {

  1436.             c = s->above_intra_ctx[col] + s->left_intra_ctx[row7];

  1437.             c += (c == 2);

  1438.         } else {

  1439.             c = have_a ? 2 * s->above_intra_ctx[col] :

  1440.                 have_l ? 2 * s->left_intra_ctx[row7] : 0;

  1441.         }

  1442.         bit = vp56_rac_get_prob(&s->c, s->prob.p.intra[c]);

  1443.         s->counts.intra[c][bit]++;

  1444.         b->intra = !bit;

  1445.     }

  1446. 

  1447.     if ((b->intra || !b->skip) && s->s.h.txfmmode == TX_SWITCHABLE) {

  1448.         int c;

  1449.         if (have_a) {

  1450.             if (have_l) {

  1451.                 c = (s->above_skip_ctx[col] ? max_tx :

  1452.                      s->above_txfm_ctx[col]) +

  1453.                     (s->left_skip_ctx[row7] ? max_tx :

  1454.                      s->left_txfm_ctx[row7]) > max_tx;

  1455.             } else {

  1456.                 c = s->above_skip_ctx[col] ? 1 :

  1457.                     (s->above_txfm_ctx[col] * 2 > max_tx);

  1458.             }

  1459.         } else if (have_l) {

  1460.             c = s->left_skip_ctx[row7] ? 1 :

  1461.                 (s->left_txfm_ctx[row7] * 2 > max_tx);

  1462.         } else {

  1463.             c = 1;

  1464.         }

  1465.         switch (max_tx) {

  1466.         case TX_32X32:

  1467.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][0]);

  1468.             if (b->tx) {

  1469.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][1]);

  1470.                 if (b->tx == 2)

  1471.                     b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx32p[c][2]);

  1472.             }

  1473.             s->counts.tx32p[c][b->tx]++;

  1474.             break;

  1475.         case TX_16X16:

  1476.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][0]);

  1477.             if (b->tx)

  1478.                 b->tx += vp56_rac_get_prob(&s->c, s->prob.p.tx16p[c][1]);

  1479.             s->counts.tx16p[c][b->tx]++;

  1480.             break;

  1481.         case TX_8X8:

  1482.             b->tx = vp56_rac_get_prob(&s->c, s->prob.p.tx8p[c]);

  1483.             s->counts.tx8p[c][b->tx]++;

  1484.             break;

  1485.         case TX_4X4:

  1486.             b->tx = TX_4X4;

  1487.             break;

  1488.         }

  1489.     } else {

  1490.         b->tx = FFMIN(max_tx, s->s.h.txfmmode);

  1491.     }

  1492. 

  1493.     if (s->s.h.keyframe || s->s.h.intraonly) {

  1494.         uint8_t *a = &s->above_mode_ctx[col * 2];

  1495.         uint8_t *l = &s->left_mode_ctx[(row7) << 1];

  1496. 

  1497.         b->comp = 0;

  1498.         if (b->bs > BS_8x8) {

  1499.             // FIXME the memory storage intermediates here aren't really

  1500.             // necessary, they're just there to make the code slightly

  1501.             // simpler for now

  1502.             b->mode[0] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1503.                                     vp9_default_kf_ymode_probs[a[0]][l[0]]);

  1504.             if (b->bs != BS_8x4) {

  1505.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1506.                                  vp9_default_kf_ymode_probs[a[1]][b->mode[0]]);

  1507.                 l[0] = a[1] = b->mode[1];

  1508.             } else {

  1509.                 l[0] = a[1] = b->mode[1] = b->mode[0];

  1510.             }

  1511.             if (b->bs != BS_4x8) {

  1512.                 b->mode[2] = a[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1513.                                         vp9_default_kf_ymode_probs[a[0]][l[1]]);

  1514.                 if (b->bs != BS_8x4) {

  1515.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1516.                                   vp9_default_kf_ymode_probs[a[1]][b->mode[2]]);

  1517.                     l[1] = a[1] = b->mode[3];

  1518.                 } else {

  1519.                     l[1] = a[1] = b->mode[3] = b->mode[2];

  1520.                 }

  1521.             } else {

  1522.                 b->mode[2] = b->mode[0];

  1523.                 l[1] = a[1] = b->mode[3] = b->mode[1];

  1524.             }

  1525.         } else {

  1526.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1527.                                           vp9_default_kf_ymode_probs[*a][*l]);

  1528.             b->mode[3] = b->mode[2] = b->mode[1] = b->mode[0];

  1529.             // FIXME this can probably be optimized

  1530.             memset(a, b->mode[0], bwh_tab[0][b->bs][0]);

  1531.             memset(l, b->mode[0], bwh_tab[0][b->bs][1]);

  1532.         }

  1533.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1534.                                      vp9_default_kf_uvmode_probs[b->mode[3]]);

  1535.     } else if (b->intra) {

  1536.         b->comp = 0;

  1537.         if (b->bs > BS_8x8) {

  1538.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1539.                                           s->prob.p.y_mode[0]);

  1540.             s->counts.y_mode[0][b->mode[0]]++;

  1541.             if (b->bs != BS_8x4) {

  1542.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1543.                                               s->prob.p.y_mode[0]);

  1544.                 s->counts.y_mode[0][b->mode[1]]++;

  1545.             } else {

  1546.                 b->mode[1] = b->mode[0];

  1547.             }

  1548.             if (b->bs != BS_4x8) {

  1549.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1550.                                               s->prob.p.y_mode[0]);

  1551.                 s->counts.y_mode[0][b->mode[2]]++;

  1552.                 if (b->bs != BS_8x4) {

  1553.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1554.                                                   s->prob.p.y_mode[0]);

  1555.                     s->counts.y_mode[0][b->mode[3]]++;

  1556.                 } else {

  1557.                     b->mode[3] = b->mode[2];

  1558.                 }

  1559.             } else {

  1560.                 b->mode[2] = b->mode[0];

  1561.                 b->mode[3] = b->mode[1];

  1562.             }

  1563.         } else {

  1564.             static const uint8_t size_group[10] = {

  1565.                 3, 3, 3, 3, 2, 2, 2, 1, 1, 1

  1566.             };

  1567.             int sz = size_group[b->bs];

  1568. 

  1569.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1570.                                           s->prob.p.y_mode[sz]);

  1571.             b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1572.             s->counts.y_mode[sz][b->mode[3]]++;

  1573.         }

  1574.         b->uvmode = vp8_rac_get_tree(&s->c, vp9_intramode_tree,

  1575.                                      s->prob.p.uv_mode[b->mode[3]]);

  1576.         s->counts.uv_mode[b->mode[3]][b->uvmode]++;

  1577.     } else {

  1578.         static const uint8_t inter_mode_ctx_lut[14][14] = {

  1579.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1580.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1581.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1582.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1583.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1584.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1585.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1586.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1587.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1588.             { 6, 6, 6, 6, 6, 6, 6, 6, 6, 6, 5, 5, 5, 5 },

  1589.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1590.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 2, 2, 1, 3 },

  1591.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 1, 1, 0, 3 },

  1592.             { 5, 5, 5, 5, 5, 5, 5, 5, 5, 5, 3, 3, 3, 4 },

  1593.         };

  1594. 

  1595.         if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[b->seg_id].ref_enabled) {

  1596.             av_assert2(s->s.h.segmentation.feat[b->seg_id].ref_val != 0);

  1597.             b->comp = 0;

  1598.             b->ref[0] = s->s.h.segmentation.feat[b->seg_id].ref_val - 1;

  1599.         } else {

  1600.             // read comp_pred flag

  1601.             if (s->s.h.comppredmode != PRED_SWITCHABLE) {

  1602.                 b->comp = s->s.h.comppredmode == PRED_COMPREF;

  1603.             } else {

  1604.                 int c;

  1605. 

  1606.                 // FIXME add intra as ref=0xff (or -1) to make these easier?

  1607.                 if (have_a) {

  1608.                     if (have_l) {

  1609.                         if (s->above_comp_ctx[col] && s->left_comp_ctx[row7]) {

  1610.                             c = 4;

  1611.                         } else if (s->above_comp_ctx[col]) {

  1612.                             c = 2 + (s->left_intra_ctx[row7] ||

  1613.                                      s->left_ref_ctx[row7] == s->s.h.fixcompref);

  1614.                         } else if (s->left_comp_ctx[row7]) {

  1615.                             c = 2 + (s->above_intra_ctx[col] ||

  1616.                                      s->above_ref_ctx[col] == s->s.h.fixcompref);

  1617.                         } else {

  1618.                             c = (!s->above_intra_ctx[col] &&

  1619.                                  s->above_ref_ctx[col] == s->s.h.fixcompref) ^

  1620.                             (!s->left_intra_ctx[row7] &&

  1621.                              s->left_ref_ctx[row & 7] == s->s.h.fixcompref);

  1622.                         }

  1623.                     } else {

  1624.                         c = s->above_comp_ctx[col] ? 3 :

  1625.                         (!s->above_intra_ctx[col] && s->above_ref_ctx[col] == s->s.h.fixcompref);

  1626.                     }

  1627.                 } else if (have_l) {

  1628.                     c = s->left_comp_ctx[row7] ? 3 :

  1629.                     (!s->left_intra_ctx[row7] && s->left_ref_ctx[row7] == s->s.h.fixcompref);

  1630.                 } else {

  1631.                     c = 1;

  1632.                 }

  1633.                 b->comp = vp56_rac_get_prob(&s->c, s->prob.p.comp[c]);

  1634.                 s->counts.comp[c][b->comp]++;

  1635.             }

  1636. 

  1637.             // read actual references

  1638.             // FIXME probably cache a few variables here to prevent repetitive

  1639.             // memory accesses below

  1640.             if (b->comp) /* two references */ {

  1641.                 int fix_idx = s->s.h.signbias[s->s.h.fixcompref], var_idx = !fix_idx, c, bit;

  1642. 

  1643.                 b->ref[fix_idx] = s->s.h.fixcompref;

  1644.                 // FIXME can this codeblob be replaced by some sort of LUT?

  1645.                 if (have_a) {

  1646.                     if (have_l) {

  1647.                         if (s->above_intra_ctx[col]) {

  1648.                             if (s->left_intra_ctx[row7]) {

  1649.                                 c = 2;

  1650.                             } else {

  1651.                                 c = 1 + 2 * (s->left_ref_ctx[row7] != s->s.h.varcompref[1]);

  1652.                             }

  1653.                         } else if (s->left_intra_ctx[row7]) {

  1654.                             c = 1 + 2 * (s->above_ref_ctx[col] != s->s.h.varcompref[1]);

  1655.                         } else {

  1656.                             int refl = s->left_ref_ctx[row7], refa = s->above_ref_ctx[col];

  1657. 

  1658.                             if (refl == refa && refa == s->s.h.varcompref[1]) {

  1659.                                 c = 0;

  1660.                             } else if (!s->left_comp_ctx[row7] && !s->above_comp_ctx[col]) {

  1661.                                 if ((refa == s->s.h.fixcompref && refl == s->s.h.varcompref[0]) ||

  1662.                                     (refl == s->s.h.fixcompref && refa == s->s.h.varcompref[0])) {

  1663.                                     c = 4;

  1664.                                 } else {

  1665.                                     c = (refa == refl) ? 3 : 1;

  1666.                                 }

  1667.                             } else if (!s->left_comp_ctx[row7]) {

  1668.                                 if (refa == s->s.h.varcompref[1] && refl != s->s.h.varcompref[1]) {

  1669.                                     c = 1;

  1670.                                 } else {

  1671.                                     c = (refl == s->s.h.varcompref[1] &&

  1672.                                          refa != s->s.h.varcompref[1]) ? 2 : 4;

  1673.                                 }

  1674.                             } else if (!s->above_comp_ctx[col]) {

  1675.                                 if (refl == s->s.h.varcompref[1] && refa != s->s.h.varcompref[1]) {

  1676.                                     c = 1;

  1677.                                 } else {

  1678.                                     c = (refa == s->s.h.varcompref[1] &&

  1679.                                          refl != s->s.h.varcompref[1]) ? 2 : 4;

  1680.                                 }

  1681.                             } else {

  1682.                                 c = (refl == refa) ? 4 : 2;

  1683.                             }

  1684.                         }

  1685.                     } else {

  1686.                         if (s->above_intra_ctx[col]) {

  1687.                             c = 2;

  1688.                         } else if (s->above_comp_ctx[col]) {

  1689.                             c = 4 * (s->above_ref_ctx[col] != s->s.h.varcompref[1]);

  1690.                         } else {

  1691.                             c = 3 * (s->above_ref_ctx[col] != s->s.h.varcompref[1]);

  1692.                         }

  1693.                     }

  1694.                 } else if (have_l) {

  1695.                     if (s->left_intra_ctx[row7]) {

  1696.                         c = 2;

  1697.                     } else if (s->left_comp_ctx[row7]) {

  1698.                         c = 4 * (s->left_ref_ctx[row7] != s->s.h.varcompref[1]);

  1699.                     } else {

  1700.                         c = 3 * (s->left_ref_ctx[row7] != s->s.h.varcompref[1]);

  1701.                     }

  1702.                 } else {

  1703.                     c = 2;

  1704.                 }

  1705.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.comp_ref[c]);

  1706.                 b->ref[var_idx] = s->s.h.varcompref[bit];

  1707.                 s->counts.comp_ref[c][bit]++;

  1708.             } else /* single reference */ {

  1709.                 int bit, c;

  1710. 

  1711.                 if (have_a && !s->above_intra_ctx[col]) {

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

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

  1714.                             if (s->above_comp_ctx[col]) {

  1715.                                 c = 1 + (!s->s.h.fixcompref || !s->left_ref_ctx[row7] ||

  1716.                                          !s->above_ref_ctx[col]);

  1717.                             } else {

  1718.                                 c = (3 * !s->above_ref_ctx[col]) +

  1719.                                     (!s->s.h.fixcompref || !s->left_ref_ctx[row7]);

  1720.                             }

  1721.                         } else if (s->above_comp_ctx[col]) {

  1722.                             c = (3 * !s->left_ref_ctx[row7]) +

  1723.                                 (!s->s.h.fixcompref || !s->above_ref_ctx[col]);

  1724.                         } else {

  1725.                             c = 2 * !s->left_ref_ctx[row7] + 2 * !s->above_ref_ctx[col];

  1726.                         }

  1727.                     } else if (s->above_intra_ctx[col]) {

  1728.                         c = 2;

  1729.                     } else if (s->above_comp_ctx[col]) {

  1730.                         c = 1 + (!s->s.h.fixcompref || !s->above_ref_ctx[col]);

  1731.                     } else {

  1732.                         c = 4 * (!s->above_ref_ctx[col]);

  1733.                     }

  1734.                 } else if (have_l && !s->left_intra_ctx[row7]) {

  1735.                     if (s->left_intra_ctx[row7]) {

  1736.                         c = 2;

  1737.                     } else if (s->left_comp_ctx[row7]) {

  1738.                         c = 1 + (!s->s.h.fixcompref || !s->left_ref_ctx[row7]);

  1739.                     } else {

  1740.                         c = 4 * (!s->left_ref_ctx[row7]);

  1741.                     }

  1742.                 } else {

  1743.                     c = 2;

  1744.                 }

  1745.                 bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][0]);

  1746.                 s->counts.single_ref[c][0][bit]++;

  1747.                 if (!bit) {

  1748.                     b->ref[0] = 0;

  1749.                 } else {

  1750.                     // FIXME can this codeblob be replaced by some sort of LUT?

  1751.                     if (have_a) {

  1752.                         if (have_l) {

  1753.                             if (s->left_intra_ctx[row7]) {

  1754.                                 if (s->above_intra_ctx[col]) {

  1755.                                     c = 2;

  1756.                                 } else if (s->above_comp_ctx[col]) {

  1757.                                     c = 1 + 2 * (s->s.h.fixcompref == 1 ||

  1758.                                                  s->above_ref_ctx[col] == 1);

  1759.                                 } else if (!s->above_ref_ctx[col]) {

  1760.                                     c = 3;

  1761.                                 } else {

  1762.                                     c = 4 * (s->above_ref_ctx[col] == 1);

  1763.                                 }

  1764.                             } else if (s->above_intra_ctx[col]) {

  1765.                                 if (s->left_intra_ctx[row7]) {

  1766.                                     c = 2;

  1767.                                 } else if (s->left_comp_ctx[row7]) {

  1768.                                     c = 1 + 2 * (s->s.h.fixcompref == 1 ||

  1769.                                                  s->left_ref_ctx[row7] == 1);

  1770.                                 } else if (!s->left_ref_ctx[row7]) {

  1771.                                     c = 3;

  1772.                                 } else {

  1773.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1774.                                 }

  1775.                             } else if (s->above_comp_ctx[col]) {

  1776.                                 if (s->left_comp_ctx[row7]) {

  1777.                                     if (s->left_ref_ctx[row7] == s->above_ref_ctx[col]) {

  1778.                                         c = 3 * (s->s.h.fixcompref == 1 ||

  1779.                                                  s->left_ref_ctx[row7] == 1);

  1780.                                     } else {

  1781.                                         c = 2;

  1782.                                     }

  1783.                                 } else if (!s->left_ref_ctx[row7]) {

  1784.                                     c = 1 + 2 * (s->s.h.fixcompref == 1 ||

  1785.                                                  s->above_ref_ctx[col] == 1);

  1786.                                 } else {

  1787.                                     c = 3 * (s->left_ref_ctx[row7] == 1) +

  1788.                                     (s->s.h.fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1789.                                 }

  1790.                             } else if (s->left_comp_ctx[row7]) {

  1791.                                 if (!s->above_ref_ctx[col]) {

  1792.                                     c = 1 + 2 * (s->s.h.fixcompref == 1 ||

  1793.                                                  s->left_ref_ctx[row7] == 1);

  1794.                                 } else {

  1795.                                     c = 3 * (s->above_ref_ctx[col] == 1) +

  1796.                                     (s->s.h.fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1797.                                 }

  1798.                             } else if (!s->above_ref_ctx[col]) {

  1799.                                 if (!s->left_ref_ctx[row7]) {

  1800.                                     c = 3;

  1801.                                 } else {

  1802.                                     c = 4 * (s->left_ref_ctx[row7] == 1);

  1803.                                 }

  1804.                             } else if (!s->left_ref_ctx[row7]) {

  1805.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1806.                             } else {

  1807.                                 c = 2 * (s->left_ref_ctx[row7] == 1) +

  1808.                                 2 * (s->above_ref_ctx[col] == 1);

  1809.                             }

  1810.                         } else {

  1811.                             if (s->above_intra_ctx[col] ||

  1812.                                 (!s->above_comp_ctx[col] && !s->above_ref_ctx[col])) {

  1813.                                 c = 2;

  1814.                             } else if (s->above_comp_ctx[col]) {

  1815.                                 c = 3 * (s->s.h.fixcompref == 1 || s->above_ref_ctx[col] == 1);

  1816.                             } else {

  1817.                                 c = 4 * (s->above_ref_ctx[col] == 1);

  1818.                             }

  1819.                         }

  1820.                     } else if (have_l) {

  1821.                         if (s->left_intra_ctx[row7] ||

  1822.                             (!s->left_comp_ctx[row7] && !s->left_ref_ctx[row7])) {

  1823.                             c = 2;

  1824.                         } else if (s->left_comp_ctx[row7]) {

  1825.                             c = 3 * (s->s.h.fixcompref == 1 || s->left_ref_ctx[row7] == 1);

  1826.                         } else {

  1827.                             c = 4 * (s->left_ref_ctx[row7] == 1);

  1828.                         }

  1829.                     } else {

  1830.                         c = 2;

  1831.                     }

  1832.                     bit = vp56_rac_get_prob(&s->c, s->prob.p.single_ref[c][1]);

  1833.                     s->counts.single_ref[c][1][bit]++;

  1834.                     b->ref[0] = 1 + bit;

  1835.                 }

  1836.             }

  1837.         }

  1838. 

  1839.         if (b->bs <= BS_8x8) {

  1840.             if (s->s.h.segmentation.enabled && s->s.h.segmentation.feat[b->seg_id].skip_enabled) {

  1841.                 b->mode[0] = b->mode[1] = b->mode[2] = b->mode[3] = ZEROMV;

  1842.             } else {

  1843.                 static const uint8_t off[10] = {

  1844.                     3, 0, 0, 1, 0, 0, 0, 0, 0, 0

  1845.                 };

  1846. 

  1847.                 // FIXME this needs to use the LUT tables from find_ref_mvs

  1848.                 // because not all are -1,0/0,-1

  1849.                 int c = inter_mode_ctx_lut[s->above_mode_ctx[col + off[b->bs]]]

  1850.                                           [s->left_mode_ctx[row7 + off[b->bs]]];

  1851. 

  1852.                 b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1853.                                               s->prob.p.mv_mode[c]);

  1854.                 b->mode[1] = b->mode[2] = b->mode[3] = b->mode[0];

  1855.                 s->counts.mv_mode[c][b->mode[0] - 10]++;

  1856.             }

  1857.         }

  1858. 

  1859.         if (s->s.h.filtermode == FILTER_SWITCHABLE) {

  1860.             int c;

  1861. 

  1862.             if (have_a && s->above_mode_ctx[col] >= NEARESTMV) {

  1863.                 if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1864.                     c = s->above_filter_ctx[col] == s->left_filter_ctx[row7] ?

  1865.                         s->left_filter_ctx[row7] : 3;

  1866.                 } else {

  1867.                     c = s->above_filter_ctx[col];

  1868.                 }

  1869.             } else if (have_l && s->left_mode_ctx[row7] >= NEARESTMV) {

  1870.                 c = s->left_filter_ctx[row7];

  1871.             } else {

  1872.                 c = 3;

  1873.             }

  1874. 

  1875.             filter_id = vp8_rac_get_tree(&s->c, vp9_filter_tree,

  1876.                                          s->prob.p.filter[c]);

  1877.             s->counts.filter[c][filter_id]++;

  1878.             b->filter = vp9_filter_lut[filter_id];

  1879.         } else {

  1880.             b->filter = s->s.h.filtermode;

  1881.         }

  1882. 

  1883.         if (b->bs > BS_8x8) {

  1884.             int c = inter_mode_ctx_lut[s->above_mode_ctx[col]][s->left_mode_ctx[row7]];

  1885. 

  1886.             b->mode[0] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1887.                                           s->prob.p.mv_mode[c]);

  1888.             s->counts.mv_mode[c][b->mode[0] - 10]++;

  1889.             fill_mv(s, b->mv[0], b->mode[0], 0);

  1890. 

  1891.             if (b->bs != BS_8x4) {

  1892.                 b->mode[1] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1893.                                               s->prob.p.mv_mode[c]);

  1894.                 s->counts.mv_mode[c][b->mode[1] - 10]++;

  1895.                 fill_mv(s, b->mv[1], b->mode[1], 1);

  1896.             } else {

  1897.                 b->mode[1] = b->mode[0];

  1898.                 AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1899.                 AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1900.             }

  1901. 

  1902.             if (b->bs != BS_4x8) {

  1903.                 b->mode[2] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1904.                                               s->prob.p.mv_mode[c]);

  1905.                 s->counts.mv_mode[c][b->mode[2] - 10]++;

  1906.                 fill_mv(s, b->mv[2], b->mode[2], 2);

  1907. 

  1908.                 if (b->bs != BS_8x4) {

  1909.                     b->mode[3] = vp8_rac_get_tree(&s->c, vp9_inter_mode_tree,

  1910.                                                   s->prob.p.mv_mode[c]);

  1911.                     s->counts.mv_mode[c][b->mode[3] - 10]++;

  1912.                     fill_mv(s, b->mv[3], b->mode[3], 3);

  1913.                 } else {

  1914.                     b->mode[3] = b->mode[2];

  1915.                     AV_COPY32(&b->mv[3][0], &b->mv[2][0]);

  1916.                     AV_COPY32(&b->mv[3][1], &b->mv[2][1]);

  1917.                 }

  1918.             } else {

  1919.                 b->mode[2] = b->mode[0];

  1920.                 AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1921.                 AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1922.                 b->mode[3] = b->mode[1];

  1923.                 AV_COPY32(&b->mv[3][0], &b->mv[1][0]);

  1924.                 AV_COPY32(&b->mv[3][1], &b->mv[1][1]);

  1925.             }

  1926.         } else {

  1927.             fill_mv(s, b->mv[0], b->mode[0], -1);

  1928.             AV_COPY32(&b->mv[1][0], &b->mv[0][0]);

  1929.             AV_COPY32(&b->mv[2][0], &b->mv[0][0]);

  1930.             AV_COPY32(&b->mv[3][0], &b->mv[0][0]);

  1931.             AV_COPY32(&b->mv[1][1], &b->mv[0][1]);

  1932.             AV_COPY32(&b->mv[2][1], &b->mv[0][1]);

  1933.             AV_COPY32(&b->mv[3][1], &b->mv[0][1]);

  1934.         }

  1935. 

  1936.         vref = b->ref[b->comp ? s->s.h.signbias[s->s.h.varcompref[0]] : 0];

  1937.     }

  1938. 

  1939. #if HAVE_FAST_64BIT

  1940. #define SPLAT_CTX(var, val, n) \

  1941.     switch (n) { \

  1942.     case 1:  var = val;                                    break; \

  1943.     case 2:  AV_WN16A(&var, val *             0x0101);     break; \

  1944.     case 4:  AV_WN32A(&var, val *         0x01010101);     break; \

  1945.     case 8:  AV_WN64A(&var, val * 0x0101010101010101ULL);  break; \

  1946.     case 16: { \

  1947.         uint64_t v64 = val * 0x0101010101010101ULL; \

  1948.         AV_WN64A(              &var,     v64); \

  1949.         AV_WN64A(&((uint8_t *) &var)[8], v64); \

  1950.         break; \

  1951.     } \

  1952.     }

  1953. #else

  1954. #define SPLAT_CTX(var, val, n) \

  1955.     switch (n) { \

  1956.     case 1:  var = val;                         break; \

  1957.     case 2:  AV_WN16A(&var, val *     0x0101);  break; \

  1958.     case 4:  AV_WN32A(&var, val * 0x01010101);  break; \

  1959.     case 8: { \

  1960.         uint32_t v32 = val * 0x01010101; \

  1961.         AV_WN32A(              &var,     v32); \

  1962.         AV_WN32A(&((uint8_t *) &var)[4], v32); \

  1963.         break; \

  1964.     } \

  1965.     case 16: { \

  1966.         uint32_t v32 = val * 0x01010101; \

  1967.         AV_WN32A(              &var,      v32); \

  1968.         AV_WN32A(&((uint8_t *) &var)[4],  v32); \

  1969.         AV_WN32A(&((uint8_t *) &var)[8],  v32); \

  1970.         AV_WN32A(&((uint8_t *) &var)[12], v32); \

  1971.         break; \

  1972.     } \

  1973.     }

  1974. #endif

  1975. 

  1976.     switch (bwh_tab[1][b->bs][0]) {

  1977. #define SET_CTXS(dir, off, n) \

  1978.     do { \

  1979.         SPLAT_CTX(s->dir##_skip_ctx[off],      b->skip,          n); \

  1980.         SPLAT_CTX(s->dir##_txfm_ctx[off],      b->tx,            n); \

  1981.         SPLAT_CTX(s->dir##_partition_ctx[off], dir##_ctx[b->bs], n); \

  1982.         if (!s->s.h.keyframe && !s->s.h.intraonly) { \

  1983.             SPLAT_CTX(s->dir##_intra_ctx[off], b->intra,   n); \

  1984.             SPLAT_CTX(s->dir##_comp_ctx[off],  b->comp,    n); \

  1985.             SPLAT_CTX(s->dir##_mode_ctx[off],  b->mode[3], n); \

  1986.             if (!b->intra) { \

  1987.                 SPLAT_CTX(s->dir##_ref_ctx[off], vref, n); \

  1988.                 if (s->s.h.filtermode == FILTER_SWITCHABLE) { \

  1989.                     SPLAT_CTX(s->dir##_filter_ctx[off], filter_id, n); \

  1990.                 } \

  1991.             } \

  1992.         } \

  1993.     } while (0)

  1994.     case 1: SET_CTXS(above, col, 1); break;

  1995.     case 2: SET_CTXS(above, col, 2); break;

  1996.     case 4: SET_CTXS(above, col, 4); break;

  1997.     case 8: SET_CTXS(above, col, 8); break;

  1998.     }

  1999.     switch (bwh_tab[1][b->bs][1]) {

  2000.     case 1: SET_CTXS(left, row7, 1); break;

  2001.     case 2: SET_CTXS(left, row7, 2); break;

  2002.     case 4: SET_CTXS(left, row7, 4); break;

  2003.     case 8: SET_CTXS(left, row7, 8); break;

  2004.     }

  2005. #undef SPLAT_CTX

  2006. #undef SET_CTXS

  2007. 

  2008.     if (!s->s.h.keyframe && !s->s.h.intraonly) {

  2009.         if (b->bs > BS_8x8) {

  2010.             int mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  2011. 

  2012.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][0], &b->mv[1][0]);

  2013.             AV_COPY32(&s->left_mv_ctx[row7 * 2 + 0][1], &b->mv[1][1]);

  2014.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][0], mv0);

  2015.             AV_WN32A(&s->left_mv_ctx[row7 * 2 + 1][1], mv1);

  2016.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][0], &b->mv[2][0]);

  2017.             AV_COPY32(&s->above_mv_ctx[col * 2 + 0][1], &b->mv[2][1]);

  2018.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][0], mv0);

  2019.             AV_WN32A(&s->above_mv_ctx[col * 2 + 1][1], mv1);

  2020.         } else {

  2021.             int n, mv0 = AV_RN32A(&b->mv[3][0]), mv1 = AV_RN32A(&b->mv[3][1]);

  2022. 

  2023.             for (n = 0; n < w4 * 2; n++) {

  2024.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][0], mv0);

  2025.                 AV_WN32A(&s->above_mv_ctx[col * 2 + n][1], mv1);

  2026.             }

  2027.             for (n = 0; n < h4 * 2; n++) {

  2028.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][0], mv0);

  2029.                 AV_WN32A(&s->left_mv_ctx[row7 * 2 + n][1], mv1);

  2030.             }

  2031.         }

  2032.     }

  2033. 

  2034.     // FIXME kinda ugly

  2035.     for (y = 0; y < h4; y++) {

  2036.         int x, o = (row + y) * s->sb_cols * 8 + col;

  2037.         struct VP9mvrefPair *mv = &s->s.frames[CUR_FRAME].mv[o];

  2038. 

  2039.         if (b->intra) {

  2040.             for (x = 0; x < w4; x++) {

  2041.                 mv[x].ref[0] =

  2042.                 mv[x].ref[1] = -1;

  2043.             }

  2044.         } else if (b->comp) {

  2045.             for (x = 0; x < w4; x++) {

  2046.                 mv[x].ref[0] = b->ref[0];

  2047.                 mv[x].ref[1] = b->ref[1];

  2048.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  2049.                 AV_COPY32(&mv[x].mv[1], &b->mv[3][1]);

  2050.             }

  2051.         } else {

  2052.             for (x = 0; x < w4; x++) {

  2053.                 mv[x].ref[0] = b->ref[0];

  2054.                 mv[x].ref[1] = -1;

  2055.                 AV_COPY32(&mv[x].mv[0], &b->mv[3][0]);

  2056.             }

  2057.         }

  2058.     }

  2059. }

  2060. 

  2061. // FIXME merge cnt/eob arguments?

  2062. static av_always_inline int

  2063. decode_coeffs_b_generic(VP56RangeCoder *c, int16_t *coef, int n_coeffs,

  2064.                         int is_tx32x32, int is8bitsperpixel, int bpp, unsigned (*cnt)[6][3],

  2065.                         unsigned (*eob)[6][2], uint8_t (*p)[6][11],

  2066.                         int nnz, const int16_t *scan, const int16_t (*nb)[2],

  2067.                         const int16_t *band_counts, const int16_t *qmul)

  2068. {

  2069.     int i = 0, band = 0, band_left = band_counts[band];

  2070.     uint8_t *tp = p[0][nnz];

  2071.     uint8_t cache[1024];

  2072. 

  2073.     do {

  2074.         int val, rc;

  2075. 

  2076.         val = vp56_rac_get_prob_branchy(c, tp[0]); // eob

  2077.         eob[band][nnz][val]++;

  2078.         if (!val)

  2079.             break;

  2080. 

  2081.     skip_eob:

  2082.         if (!vp56_rac_get_prob_branchy(c, tp[1])) { // zero

  2083.             cnt[band][nnz][0]++;

  2084.             if (!--band_left)

  2085.                 band_left = band_counts[++band];

  2086.             cache[scan[i]] = 0;

  2087.             nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  2088.             tp = p[band][nnz];

  2089.             if (++i == n_coeffs)

  2090.                 break; //invalid input; blocks should end with EOB

  2091.             goto skip_eob;

  2092.         }

  2093. 

  2094.         rc = scan[i];

  2095.         if (!vp56_rac_get_prob_branchy(c, tp[2])) { // one

  2096.             cnt[band][nnz][1]++;

  2097.             val = 1;

  2098.             cache[rc] = 1;

  2099.         } else {

  2100.             // fill in p[3-10] (model fill) - only once per frame for each pos

  2101.             if (!tp[3])

  2102.                 memcpy(&tp[3], vp9_model_pareto8[tp[2]], 8);

  2103. 

  2104.             cnt[band][nnz][2]++;

  2105.             if (!vp56_rac_get_prob_branchy(c, tp[3])) { // 2, 3, 4

  2106.                 if (!vp56_rac_get_prob_branchy(c, tp[4])) {

  2107.                     cache[rc] = val = 2;

  2108.                 } else {

  2109.                     val = 3 + vp56_rac_get_prob(c, tp[5]);

  2110.                     cache[rc] = 3;

  2111.                 }

  2112.             } else if (!vp56_rac_get_prob_branchy(c, tp[6])) { // cat1/2

  2113.                 cache[rc] = 4;

  2114.                 if (!vp56_rac_get_prob_branchy(c, tp[7])) {

  2115.                     val = 5 + vp56_rac_get_prob(c, 159);

  2116.                 } else {

  2117.                     val  = 7 + (vp56_rac_get_prob(c, 165) << 1);

  2118.                     val +=      vp56_rac_get_prob(c, 145);

  2119.                 }

  2120.             } else { // cat 3-6

  2121.                 cache[rc] = 5;

  2122.                 if (!vp56_rac_get_prob_branchy(c, tp[8])) {

  2123.                     if (!vp56_rac_get_prob_branchy(c, tp[9])) {

  2124.                         val  = 11 + (vp56_rac_get_prob(c, 173) << 2);

  2125.                         val +=      (vp56_rac_get_prob(c, 148) << 1);

  2126.                         val +=       vp56_rac_get_prob(c, 140);

  2127.                     } else {

  2128.                         val  = 19 + (vp56_rac_get_prob(c, 176) << 3);

  2129.                         val +=      (vp56_rac_get_prob(c, 155) << 2);

  2130.                         val +=      (vp56_rac_get_prob(c, 140) << 1);

  2131.                         val +=       vp56_rac_get_prob(c, 135);

  2132.                     }

  2133.                 } else if (!vp56_rac_get_prob_branchy(c, tp[10])) {

  2134.                     val  = 35 + (vp56_rac_get_prob(c, 180) << 4);

  2135.                     val +=      (vp56_rac_get_prob(c, 157) << 3);

  2136.                     val +=      (vp56_rac_get_prob(c, 141) << 2);

  2137.                     val +=      (vp56_rac_get_prob(c, 134) << 1);

  2138.                     val +=       vp56_rac_get_prob(c, 130);

  2139.                 } else {

  2140.                     val = 67;

  2141.                     if (!is8bitsperpixel) {

  2142.                         if (bpp == 12) {

  2143.                             val += vp56_rac_get_prob(c, 255) << 17;

  2144.                             val += vp56_rac_get_prob(c, 255) << 16;

  2145.                         }

  2146.                         val +=  (vp56_rac_get_prob(c, 255) << 15);

  2147.                         val +=  (vp56_rac_get_prob(c, 255) << 14);

  2148.                     }

  2149.                     val +=      (vp56_rac_get_prob(c, 254) << 13);

  2150.                     val +=      (vp56_rac_get_prob(c, 254) << 12);

  2151.                     val +=      (vp56_rac_get_prob(c, 254) << 11);

  2152.                     val +=      (vp56_rac_get_prob(c, 252) << 10);

  2153.                     val +=      (vp56_rac_get_prob(c, 249) << 9);

  2154.                     val +=      (vp56_rac_get_prob(c, 243) << 8);

  2155.                     val +=      (vp56_rac_get_prob(c, 230) << 7);

  2156.                     val +=      (vp56_rac_get_prob(c, 196) << 6);

  2157.                     val +=      (vp56_rac_get_prob(c, 177) << 5);

  2158.                     val +=      (vp56_rac_get_prob(c, 153) << 4);

  2159.                     val +=      (vp56_rac_get_prob(c, 140) << 3);

  2160.                     val +=      (vp56_rac_get_prob(c, 133) << 2);

  2161.                     val +=      (vp56_rac_get_prob(c, 130) << 1);

  2162.                     val +=       vp56_rac_get_prob(c, 129);

  2163.                 }

  2164.             }

  2165.         }

  2166. #define STORE_COEF(c, i, v) do { \

  2167.     if (is8bitsperpixel) { \

  2168.         c[i] = v; \

  2169.     } else { \

  2170.         AV_WN32A(&c[i * 2], v); \

  2171.     } \

  2172. } while (0)

  2173.         if (!--band_left)

  2174.             band_left = band_counts[++band];

  2175.         if (is_tx32x32)

  2176.             STORE_COEF(coef, rc, ((vp8_rac_get(c) ? -val : val) * qmul[!!i]) / 2);

  2177.         else

  2178.             STORE_COEF(coef, rc, (vp8_rac_get(c) ? -val : val) * qmul[!!i]);

  2179.         nnz = (1 + cache[nb[i][0]] + cache[nb[i][1]]) >> 1;

  2180.         tp = p[band][nnz];

  2181.     } while (++i < n_coeffs);

  2182. 

  2183.     return i;

  2184. }

  2185. 

  2186. static int decode_coeffs_b_8bpp(VP9Context *s, int16_t *coef, int n_coeffs,

  2187.                                 unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  2188.                                 uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  2189.                                 const int16_t (*nb)[2], const int16_t *band_counts,

  2190.                                 const int16_t *qmul)

  2191. {

  2192.     return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 0, 1, 8, cnt, eob, p,

  2193.                                    nnz, scan, nb, band_counts, qmul);

  2194. }

  2195. 

  2196. static int decode_coeffs_b32_8bpp(VP9Context *s, int16_t *coef, int n_coeffs,

  2197.                                   unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  2198.                                   uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  2199.                                   const int16_t (*nb)[2], const int16_t *band_counts,

  2200.                                   const int16_t *qmul)

  2201. {

  2202.     return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 1, 1, 8, cnt, eob, p,

  2203.                                    nnz, scan, nb, band_counts, qmul);

  2204. }

  2205. 

  2206. static int decode_coeffs_b_16bpp(VP9Context *s, int16_t *coef, int n_coeffs,

  2207.                                  unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  2208.                                  uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  2209.                                  const int16_t (*nb)[2], const int16_t *band_counts,

  2210.                                  const int16_t *qmul)

  2211. {

  2212.     return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 0, 0, s->bpp, cnt, eob, p,

  2213.                                    nnz, scan, nb, band_counts, qmul);

  2214. }

  2215. 

  2216. static int decode_coeffs_b32_16bpp(VP9Context *s, int16_t *coef, int n_coeffs,

  2217.                                    unsigned (*cnt)[6][3], unsigned (*eob)[6][2],

  2218.                                    uint8_t (*p)[6][11], int nnz, const int16_t *scan,

  2219.                                    const int16_t (*nb)[2], const int16_t *band_counts,

  2220.                                    const int16_t *qmul)

  2221. {

  2222.     return decode_coeffs_b_generic(&s->c, coef, n_coeffs, 1, 0, s->bpp, cnt, eob, p,

  2223.                                    nnz, scan, nb, band_counts, qmul);

  2224. }

  2225. 

  2226. static av_always_inline int decode_coeffs(AVCodecContext *ctx, int is8bitsperpixel)

  2227. {

  2228.     VP9Context *s = ctx->priv_data;

  2229.     VP9Block *b = s->b;

  2230.     int row = s->row, col = s->col;

  2231.     uint8_t (*p)[6][11] = s->prob.coef[b->tx][0 /* y */][!b->intra];

  2232.     unsigned (*c)[6][3] = s->counts.coef[b->tx][0 /* y */][!b->intra];

  2233.     unsigned (*e)[6][2] = s->counts.eob[b->tx][0 /* y */][!b->intra];

  2234.     int w4 = bwh_tab[1][b->bs][0] << 1, h4 = bwh_tab[1][b->bs][1] << 1;

  2235.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2236.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2237.     int n, pl, x, y, res;

  2238.     int16_t (*qmul)[2] = s->s.h.segmentation.feat[b->seg_id].qmul;

  2239.     int tx = 4 * s->s.h.lossless + b->tx;

  2240.     const int16_t * const *yscans = vp9_scans[tx];

  2241.     const int16_t (* const *ynbs)[2] = vp9_scans_nb[tx];

  2242.     const int16_t *uvscan = vp9_scans[b->uvtx][DCT_DCT];

  2243.     const int16_t (*uvnb)[2] = vp9_scans_nb[b->uvtx][DCT_DCT];

  2244.     uint8_t *a = &s->above_y_nnz_ctx[col * 2];

  2245.     uint8_t *l = &s->left_y_nnz_ctx[(row & 7) << 1];

  2246.     static const int16_t band_counts[4][8] = {

  2247.         { 1, 2, 3, 4,  3,   16 - 13 },

  2248.         { 1, 2, 3, 4, 11,   64 - 21 },

  2249.         { 1, 2, 3, 4, 11,  256 - 21 },

  2250.         { 1, 2, 3, 4, 11, 1024 - 21 },

  2251.     };

  2252.     const int16_t *y_band_counts = band_counts[b->tx];

  2253.     const int16_t *uv_band_counts = band_counts[b->uvtx];

  2254.     int bytesperpixel = is8bitsperpixel ? 1 : 2;

  2255.     int total_coeff = 0;

  2256. 

  2257. #define MERGE(la, end, step, rd) \

  2258.     for (n = 0; n < end; n += step) \

  2259.         la[n] = !!rd(&la[n])

  2260. #define MERGE_CTX(step, rd) \

  2261.     do { \

  2262.         MERGE(l, end_y, step, rd); \

  2263.         MERGE(a, end_x, step, rd); \

  2264.     } while (0)

  2265. 

  2266. #define DECODE_Y_COEF_LOOP(step, mode_index, v) \

  2267.     for (n = 0, y = 0; y < end_y; y += step) { \

  2268.         for (x = 0; x < end_x; x += step, n += step * step) { \

  2269.             enum TxfmType txtp = vp9_intra_txfm_type[b->mode[mode_index]]; \

  2270.             res = (is8bitsperpixel ? decode_coeffs_b##v##_8bpp : decode_coeffs_b##v##_16bpp) \

  2271.                                     (s, s->block + 16 * n * bytesperpixel, 16 * step * step, \

  2272.                                      c, e, p, a[x] + l[y], yscans[txtp], \

  2273.                                      ynbs[txtp], y_band_counts, qmul[0]); \

  2274.             a[x] = l[y] = !!res; \

  2275.             total_coeff |= !!res; \

  2276.             if (step >= 4) { \

  2277.                 AV_WN16A(&s->eob[n], res); \

  2278.             } else { \

  2279.                 s->eob[n] = res; \

  2280.             } \

  2281.         } \

  2282.     }

  2283. 

  2284. #define SPLAT(la, end, step, cond) \

  2285.     if (step == 2) { \

  2286.         for (n = 1; n < end; n += step) \

  2287.             la[n] = la[n - 1]; \

  2288.     } else if (step == 4) { \

  2289.         if (cond) { \

  2290.             for (n = 0; n < end; n += step) \

  2291.                 AV_WN32A(&la[n], la[n] * 0x01010101); \

  2292.         } else { \

  2293.             for (n = 0; n < end; n += step) \

  2294.                 memset(&la[n + 1], la[n], FFMIN(end - n - 1, 3)); \

  2295.         } \

  2296.     } else /* step == 8 */ { \

  2297.         if (cond) { \

  2298.             if (HAVE_FAST_64BIT) { \

  2299.                 for (n = 0; n < end; n += step) \

  2300.                     AV_WN64A(&la[n], la[n] * 0x0101010101010101ULL); \

  2301.             } else { \

  2302.                 for (n = 0; n < end; n += step) { \

  2303.                     uint32_t v32 = la[n] * 0x01010101; \

  2304.                     AV_WN32A(&la[n],     v32); \

  2305.                     AV_WN32A(&la[n + 4], v32); \

  2306.                 } \

  2307.             } \

  2308.         } else { \

  2309.             for (n = 0; n < end; n += step) \

  2310.                 memset(&la[n + 1], la[n], FFMIN(end - n - 1, 7)); \

  2311.         } \

  2312.     }

  2313. #define SPLAT_CTX(step) \

  2314.     do { \

  2315.         SPLAT(a, end_x, step, end_x == w4); \

  2316.         SPLAT(l, end_y, step, end_y == h4); \

  2317.     } while (0)

  2318. 

  2319.     /* y tokens */

  2320.     switch (b->tx) {

  2321.     case TX_4X4:

  2322.         DECODE_Y_COEF_LOOP(1, b->bs > BS_8x8 ? n : 0,);

  2323.         break;

  2324.     case TX_8X8:

  2325.         MERGE_CTX(2, AV_RN16A);

  2326.         DECODE_Y_COEF_LOOP(2, 0,);

  2327.         SPLAT_CTX(2);

  2328.         break;

  2329.     case TX_16X16:

  2330.         MERGE_CTX(4, AV_RN32A);

  2331.         DECODE_Y_COEF_LOOP(4, 0,);

  2332.         SPLAT_CTX(4);

  2333.         break;

  2334.     case TX_32X32:

  2335.         MERGE_CTX(8, AV_RN64A);

  2336.         DECODE_Y_COEF_LOOP(8, 0, 32);

  2337.         SPLAT_CTX(8);

  2338.         break;

  2339.     }

  2340. 

  2341. #define DECODE_UV_COEF_LOOP(step, v) \

  2342.     for (n = 0, y = 0; y < end_y; y += step) { \

  2343.         for (x = 0; x < end_x; x += step, n += step * step) { \

  2344.             res = (is8bitsperpixel ? decode_coeffs_b##v##_8bpp : decode_coeffs_b##v##_16bpp) \

  2345.                                     (s, s->uvblock[pl] + 16 * n * bytesperpixel, \

  2346.                                      16 * step * step, c, e, p, a[x] + l[y], \

  2347.                                      uvscan, uvnb, uv_band_counts, qmul[1]); \

  2348.             a[x] = l[y] = !!res; \

  2349.             total_coeff |= !!res; \

  2350.             if (step >= 4) { \

  2351.                 AV_WN16A(&s->uveob[pl][n], res); \

  2352.             } else { \

  2353.                 s->uveob[pl][n] = res; \

  2354.             } \

  2355.         } \

  2356.     }

  2357. 

  2358.     p = s->prob.coef[b->uvtx][1 /* uv */][!b->intra];

  2359.     c = s->counts.coef[b->uvtx][1 /* uv */][!b->intra];

  2360.     e = s->counts.eob[b->uvtx][1 /* uv */][!b->intra];

  2361.     w4 >>= s->ss_h;

  2362.     end_x >>= s->ss_h;

  2363.     h4 >>= s->ss_v;

  2364.     end_y >>= s->ss_v;

  2365.     for (pl = 0; pl < 2; pl++) {

  2366.         a = &s->above_uv_nnz_ctx[pl][col << !s->ss_h];

  2367.         l = &s->left_uv_nnz_ctx[pl][(row & 7) << !s->ss_v];

  2368.         switch (b->uvtx) {

  2369.         case TX_4X4:

  2370.             DECODE_UV_COEF_LOOP(1,);

  2371.             break;

  2372.         case TX_8X8:

  2373.             MERGE_CTX(2, AV_RN16A);

  2374.             DECODE_UV_COEF_LOOP(2,);

  2375.             SPLAT_CTX(2);

  2376.             break;

  2377.         case TX_16X16:

  2378.             MERGE_CTX(4, AV_RN32A);

  2379.             DECODE_UV_COEF_LOOP(4,);

  2380.             SPLAT_CTX(4);

  2381.             break;

  2382.         case TX_32X32:

  2383.             MERGE_CTX(8, AV_RN64A);

  2384.             DECODE_UV_COEF_LOOP(8, 32);

  2385.             SPLAT_CTX(8);

  2386.             break;

  2387.         }

  2388.     }

  2389. 

  2390.     return total_coeff;

  2391. }

  2392. 

  2393. static int decode_coeffs_8bpp(AVCodecContext *ctx)

  2394. {

  2395.     return decode_coeffs(ctx, 1);

  2396. }

  2397. 

  2398. static int decode_coeffs_16bpp(AVCodecContext *ctx)

  2399. {

  2400.     return decode_coeffs(ctx, 0);

  2401. }

  2402. 

  2403. static av_always_inline int check_intra_mode(VP9Context *s, int mode, uint8_t **a,

  2404.                                              uint8_t *dst_edge, ptrdiff_t stride_edge,

  2405.                                              uint8_t *dst_inner, ptrdiff_t stride_inner,

  2406.                                              uint8_t *l, int col, int x, int w,

  2407.                                              int row, int y, enum TxfmMode tx,

  2408.                                              int p, int ss_h, int ss_v, int bytesperpixel)

  2409. {

  2410.     int have_top = row > 0 || y > 0;

  2411.     int have_left = col > s->tile_col_start || x > 0;

  2412.     int have_right = x < w - 1;

  2413.     int bpp = s->bpp;

  2414.     static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {

  2415.         [VERT_PRED]            = { { DC_127_PRED,          VERT_PRED },

  2416.                                    { DC_127_PRED,          VERT_PRED } },

  2417.         [HOR_PRED]             = { { DC_129_PRED,          DC_129_PRED },

  2418.                                    { HOR_PRED,             HOR_PRED } },

  2419.         [DC_PRED]              = { { DC_128_PRED,          TOP_DC_PRED },

  2420.                                    { LEFT_DC_PRED,         DC_PRED } },

  2421.         [DIAG_DOWN_LEFT_PRED]  = { { DC_127_PRED,          DIAG_DOWN_LEFT_PRED },

  2422.                                    { DC_127_PRED,          DIAG_DOWN_LEFT_PRED } },

  2423.         [DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },

  2424.                                    { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },

  2425.         [VERT_RIGHT_PRED]      = { { VERT_RIGHT_PRED,      VERT_RIGHT_PRED },

  2426.                                    { VERT_RIGHT_PRED,      VERT_RIGHT_PRED } },

  2427.         [HOR_DOWN_PRED]        = { { HOR_DOWN_PRED,        HOR_DOWN_PRED },

  2428.                                    { HOR_DOWN_PRED,        HOR_DOWN_PRED } },

  2429.         [VERT_LEFT_PRED]       = { { DC_127_PRED,          VERT_LEFT_PRED },

  2430.                                    { DC_127_PRED,          VERT_LEFT_PRED } },

  2431.         [HOR_UP_PRED]          = { { DC_129_PRED,          DC_129_PRED },

  2432.                                    { HOR_UP_PRED,          HOR_UP_PRED } },

  2433.         [TM_VP8_PRED]          = { { DC_129_PRED,          VERT_PRED },

  2434.                                    { HOR_PRED,             TM_VP8_PRED } },

  2435.     };

  2436.     static const struct {

  2437.         uint8_t needs_left:1;

  2438.         uint8_t needs_top:1;

  2439.         uint8_t needs_topleft:1;

  2440.         uint8_t needs_topright:1;

  2441.         uint8_t invert_left:1;

  2442.     } edges[N_INTRA_PRED_MODES] = {

  2443.         [VERT_PRED]            = { .needs_top  = 1 },

  2444.         [HOR_PRED]             = { .needs_left = 1 },

  2445.         [DC_PRED]              = { .needs_top  = 1, .needs_left = 1 },

  2446.         [DIAG_DOWN_LEFT_PRED]  = { .needs_top  = 1, .needs_topright = 1 },

  2447.         [DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2448.         [VERT_RIGHT_PRED]      = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2449.         [HOR_DOWN_PRED]        = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2450.         [VERT_LEFT_PRED]       = { .needs_top  = 1, .needs_topright = 1 },

  2451.         [HOR_UP_PRED]          = { .needs_left = 1, .invert_left = 1 },

  2452.         [TM_VP8_PRED]          = { .needs_left = 1, .needs_top = 1, .needs_topleft = 1 },

  2453.         [LEFT_DC_PRED]         = { .needs_left = 1 },

  2454.         [TOP_DC_PRED]          = { .needs_top  = 1 },

  2455.         [DC_128_PRED]          = { 0 },

  2456.         [DC_127_PRED]          = { 0 },

  2457.         [DC_129_PRED]          = { 0 }

  2458.     };

  2459. 

  2460.     av_assert2(mode >= 0 && mode < 10);

  2461.     mode = mode_conv[mode][have_left][have_top];

  2462.     if (edges[mode].needs_top) {

  2463.         uint8_t *top, *topleft;

  2464.         int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !ss_h) - x) * 4;

  2465.         int n_px_need_tr = 0;

  2466. 

  2467.         if (tx == TX_4X4 && edges[mode].needs_topright && have_right)

  2468.             n_px_need_tr = 4;

  2469. 

  2470.         // if top of sb64-row, use s->intra_pred_data[] instead of

  2471.         // dst[-stride] for intra prediction (it contains pre- instead of

  2472.         // post-loopfilter data)

  2473.         if (have_top) {

  2474.             top = !(row & 7) && !y ?

  2475.                 s->intra_pred_data[p] + (col * (8 >> ss_h) + x * 4) * bytesperpixel :

  2476.                 y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];

  2477.             if (have_left)

  2478.                 topleft = !(row & 7) && !y ?

  2479.                     s->intra_pred_data[p] + (col * (8 >> ss_h) + x * 4) * bytesperpixel :

  2480.                     y == 0 || x == 0 ? &dst_edge[-stride_edge] :

  2481.                     &dst_inner[-stride_inner];

  2482.         }

  2483. 

  2484.         if (have_top &&

  2485.             (!edges[mode].needs_topleft || (have_left && top == topleft)) &&

  2486.             (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&

  2487.             n_px_need + n_px_need_tr <= n_px_have) {

  2488.             *a = top;

  2489.         } else {

  2490.             if (have_top) {

  2491.                 if (n_px_need <= n_px_have) {

  2492.                     memcpy(*a, top, n_px_need * bytesperpixel);

  2493.                 } else {

  2494. #define memset_bpp(c, i1, v, i2, num) do { \

  2495.     if (bytesperpixel == 1) { \

  2496.         memset(&(c)[(i1)], (v)[(i2)], (num)); \

  2497.     } else { \

  2498.         int n, val = AV_RN16A(&(v)[(i2) * 2]); \

  2499.         for (n = 0; n < (num); n++) { \

  2500.             AV_WN16A(&(c)[((i1) + n) * 2], val); \

  2501.         } \

  2502.     } \

  2503. } while (0)

  2504.                     memcpy(*a, top, n_px_have * bytesperpixel);

  2505.                     memset_bpp(*a, n_px_have, (*a), n_px_have - 1, n_px_need - n_px_have);

  2506.                 }

  2507.             } else {

  2508. #define memset_val(c, val, num) do { \

  2509.     if (bytesperpixel == 1) { \

  2510.         memset((c), (val), (num)); \

  2511.     } else { \

  2512.         int n; \

  2513.         for (n = 0; n < (num); n++) { \

  2514.             AV_WN16A(&(c)[n * 2], (val)); \

  2515.         } \

  2516.     } \

  2517. } while (0)

  2518.                 memset_val(*a, (128 << (bpp - 8)) - 1, n_px_need);

  2519.             }

  2520.             if (edges[mode].needs_topleft) {

  2521.                 if (have_left && have_top) {

  2522. #define assign_bpp(c, i1, v, i2) do { \

  2523.     if (bytesperpixel == 1) { \

  2524.         (c)[(i1)] = (v)[(i2)]; \

  2525.     } else { \

  2526.         AV_COPY16(&(c)[(i1) * 2], &(v)[(i2) * 2]); \

  2527.     } \

  2528. } while (0)

  2529.                     assign_bpp(*a, -1, topleft, -1);

  2530.                 } else {

  2531. #define assign_val(c, i, v) do { \

  2532.     if (bytesperpixel == 1) { \

  2533.         (c)[(i)] = (v); \

  2534.     } else { \

  2535.         AV_WN16A(&(c)[(i) * 2], (v)); \

  2536.     } \

  2537. } while (0)

  2538.                     assign_val((*a), -1, (128 << (bpp - 8)) + (have_top ? +1 : -1));

  2539.                 }

  2540.             }

  2541.             if (tx == TX_4X4 && edges[mode].needs_topright) {

  2542.                 if (have_top && have_right &&

  2543.                     n_px_need + n_px_need_tr <= n_px_have) {

  2544.                     memcpy(&(*a)[4 * bytesperpixel], &top[4 * bytesperpixel], 4 * bytesperpixel);

  2545.                 } else {

  2546.                     memset_bpp(*a, 4, *a, 3, 4);

  2547.                 }

  2548.             }

  2549.         }

  2550.     }

  2551.     if (edges[mode].needs_left) {

  2552.         if (have_left) {

  2553.             int n_px_need = 4 << tx, i, n_px_have = (((s->rows - row) << !ss_v) - y) * 4;

  2554.             uint8_t *dst = x == 0 ? dst_edge : dst_inner;

  2555.             ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;

  2556. 

  2557.             if (edges[mode].invert_left) {

  2558.                 if (n_px_need <= n_px_have) {

  2559.                     for (i = 0; i < n_px_need; i++)

  2560.                         assign_bpp(l, i, &dst[i * stride], -1);

  2561.                 } else {

  2562.                     for (i = 0; i < n_px_have; i++)

  2563.                         assign_bpp(l, i, &dst[i * stride], -1);

  2564.                     memset_bpp(l, n_px_have, l, n_px_have - 1, n_px_need - n_px_have);

  2565.                 }

  2566.             } else {

  2567.                 if (n_px_need <= n_px_have) {

  2568.                     for (i = 0; i < n_px_need; i++)

  2569.                         assign_bpp(l, n_px_need - 1 - i, &dst[i * stride], -1);

  2570.                 } else {

  2571.                     for (i = 0; i < n_px_have; i++)

  2572.                         assign_bpp(l, n_px_need - 1 - i, &dst[i * stride], -1);

  2573.                     memset_bpp(l, 0, l, n_px_need - n_px_have, n_px_need - n_px_have);

  2574.                 }

  2575.             }

  2576.         } else {

  2577.             memset_val(l, (128 << (bpp - 8)) + 1, 4 << tx);

  2578.         }

  2579.     }

  2580. 

  2581.     return mode;

  2582. }

  2583. 

  2584. static av_always_inline void intra_recon(AVCodecContext *ctx, ptrdiff_t y_off,

  2585.                                          ptrdiff_t uv_off, int bytesperpixel)

  2586. {

  2587.     VP9Context *s = ctx->priv_data;

  2588.     VP9Block *b = s->b;

  2589.     int row = s->row, col = s->col;

  2590.     int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2591.     int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2592.     int end_x = FFMIN(2 * (s->cols - col), w4);

  2593.     int end_y = FFMIN(2 * (s->rows - row), h4);

  2594.     int tx = 4 * s->s.h.lossless + b->tx, uvtx = b->uvtx + 4 * s->s.h.lossless;

  2595.     int uvstep1d = 1 << b->uvtx, p;

  2596.     uint8_t *dst = s->dst[0], *dst_r = s->s.frames[CUR_FRAME].tf.f->data[0] + y_off;

  2597.     LOCAL_ALIGNED_32(uint8_t, a_buf, [96]);

  2598.     LOCAL_ALIGNED_32(uint8_t, l, [64]);

  2599. 

  2600.     for (n = 0, y = 0; y < end_y; y += step1d) {

  2601.         uint8_t *ptr = dst, *ptr_r = dst_r;

  2602.         for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d * bytesperpixel,

  2603.                                ptr_r += 4 * step1d * bytesperpixel, n += step) {

  2604.             int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?

  2605.                                y * 2 + x : 0];

  2606.             uint8_t *a = &a_buf[32];

  2607.             enum TxfmType txtp = vp9_intra_txfm_type[mode];

  2608.             int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2609. 

  2610.             mode = check_intra_mode(s, mode, &a, ptr_r,

  2611.                                     s->s.frames[CUR_FRAME].tf.f->linesize[0],

  2612.                                     ptr, s->y_stride, l,

  2613.                                     col, x, w4, row, y, b->tx, 0, 0, 0, bytesperpixel);

  2614.             s->dsp.intra_pred[b->tx][mode](ptr, s->y_stride, l, a);

  2615.             if (eob)

  2616.                 s->dsp.itxfm_add[tx][txtp](ptr, s->y_stride,

  2617.                                            s->block + 16 * n * bytesperpixel, eob);

  2618.         }

  2619.         dst_r += 4 * step1d * s->s.frames[CUR_FRAME].tf.f->linesize[0];

  2620.         dst   += 4 * step1d * s->y_stride;

  2621.     }

  2622. 

  2623.     // U/V

  2624.     w4 >>= s->ss_h;

  2625.     end_x >>= s->ss_h;

  2626.     end_y >>= s->ss_v;

  2627.     step = 1 << (b->uvtx * 2);

  2628.     for (p = 0; p < 2; p++) {

  2629.         dst   = s->dst[1 + p];

  2630.         dst_r = s->s.frames[CUR_FRAME].tf.f->data[1 + p] + uv_off;

  2631.         for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2632.             uint8_t *ptr = dst, *ptr_r = dst_r;

  2633.             for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d * bytesperpixel,

  2634.                                    ptr_r += 4 * uvstep1d * bytesperpixel, n += step) {

  2635.                 int mode = b->uvmode;

  2636.                 uint8_t *a = &a_buf[32];

  2637.                 int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2638. 

  2639.                 mode = check_intra_mode(s, mode, &a, ptr_r,

  2640.                                         s->s.frames[CUR_FRAME].tf.f->linesize[1],

  2641.                                         ptr, s->uv_stride, l, col, x, w4, row, y,

  2642.                                         b->uvtx, p + 1, s->ss_h, s->ss_v, bytesperpixel);

  2643.                 s->dsp.intra_pred[b->uvtx][mode](ptr, s->uv_stride, l, a);

  2644.                 if (eob)

  2645.                     s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2646.                                                     s->uvblock[p] + 16 * n * bytesperpixel, eob);

  2647.             }

  2648.             dst_r += 4 * uvstep1d * s->s.frames[CUR_FRAME].tf.f->linesize[1];

  2649.             dst   += 4 * uvstep1d * s->uv_stride;

  2650.         }

  2651.     }

  2652. }

  2653. 

  2654. static void intra_recon_8bpp(AVCodecContext *ctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  2655. {

  2656.     intra_recon(ctx, y_off, uv_off, 1);

  2657. }

  2658. 

  2659. static void intra_recon_16bpp(AVCodecContext *ctx, ptrdiff_t y_off, ptrdiff_t uv_off)

  2660. {

  2661.     intra_recon(ctx, y_off, uv_off, 2);

  2662. }

  2663. 

  2664. static av_always_inline void mc_luma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2],

  2665.                                               uint8_t *dst, ptrdiff_t dst_stride,

  2666.                                               const uint8_t *ref, ptrdiff_t ref_stride,

  2667.                                               ThreadFrame *ref_frame,

  2668.                                               ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2669.                                               int bw, int bh, int w, int h, int bytesperpixel)

  2670. {

  2671.     int mx = mv->x, my = mv->y, th;

  2672. 

  2673.     y += my >> 3;

  2674.     x += mx >> 3;

  2675.     ref += y * ref_stride + x * bytesperpixel;

  2676.     mx &= 7;

  2677.     my &= 7;

  2678.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2679.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2680.     // the longest loopfilter of the next sbrow

  2681.     th = (y + bh + 4 * !!my + 7) >> 6;

  2682.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2683.     if (x < !!mx * 3 || y < !!my * 3 ||

  2684.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2685.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2686.                                  ref - !!my * 3 * ref_stride - !!mx * 3 * bytesperpixel,

  2687.                                  160, ref_stride,

  2688.                                  bw + !!mx * 7, bh + !!my * 7,

  2689.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2690.         ref = s->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;

  2691.         ref_stride = 160;

  2692.     }

  2693.     mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);

  2694. }

  2695. 

  2696. static av_always_inline void mc_chroma_unscaled(VP9Context *s, vp9_mc_func (*mc)[2],

  2697.                                                 uint8_t *dst_u, uint8_t *dst_v,

  2698.                                                 ptrdiff_t dst_stride,

  2699.                                                 const uint8_t *ref_u, ptrdiff_t src_stride_u,

  2700.                                                 const uint8_t *ref_v, ptrdiff_t src_stride_v,

  2701.                                                 ThreadFrame *ref_frame,

  2702.                                                 ptrdiff_t y, ptrdiff_t x, const VP56mv *mv,

  2703.                                                 int bw, int bh, int w, int h, int bytesperpixel)

  2704. {

  2705.     int mx = mv->x << !s->ss_h, my = mv->y << !s->ss_v, th;

  2706. 

  2707.     y += my >> 4;

  2708.     x += mx >> 4;

  2709.     ref_u += y * src_stride_u + x * bytesperpixel;

  2710.     ref_v += y * src_stride_v + x * bytesperpixel;

  2711.     mx &= 15;

  2712.     my &= 15;

  2713.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2714.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2715.     // the longest loopfilter of the next sbrow

  2716.     th = (y + bh + 4 * !!my + 7) >> (6 - s->ss_v);

  2717.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2718.     if (x < !!mx * 3 || y < !!my * 3 ||

  2719.         x + !!mx * 4 > w - bw || y + !!my * 4 > h - bh) {

  2720.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2721.                                  ref_u - !!my * 3 * src_stride_u - !!mx * 3 * bytesperpixel,

  2722.                                  160, src_stride_u,

  2723.                                  bw + !!mx * 7, bh + !!my * 7,

  2724.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2725.         ref_u = s->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;

  2726.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, 160, bh, mx, my);

  2727. 

  2728.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2729.                                  ref_v - !!my * 3 * src_stride_v - !!mx * 3 * bytesperpixel,

  2730.                                  160, src_stride_v,

  2731.                                  bw + !!mx * 7, bh + !!my * 7,

  2732.                                  x - !!mx * 3, y - !!my * 3, w, h);

  2733.         ref_v = s->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;

  2734.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, 160, bh, mx, my);

  2735.     } else {

  2736.         mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);

  2737.         mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);

  2738.     }

  2739. }

  2740. 

  2741. #define mc_luma_dir(s, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, \

  2742.                     px, py, pw, ph, bw, bh, w, h, i) \

  2743.     mc_luma_unscaled(s, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \

  2744.                      mv, bw, bh, w, h, bytesperpixel)

  2745. #define mc_chroma_dir(s, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2746.                       row, col, mv, px, py, pw, ph, bw, bh, w, h, i) \

  2747.     mc_chroma_unscaled(s, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2748.                        row, col, mv, bw, bh, w, h, bytesperpixel)

  2749. #define SCALED 0

  2750. #define FN(x) x##_8bpp

  2751. #define BYTES_PER_PIXEL 1

  2752. #include "vp9_mc_template.c"

  2753. #undef FN

  2754. #undef BYTES_PER_PIXEL

  2755. #define FN(x) x##_16bpp

  2756. #define BYTES_PER_PIXEL 2

  2757. #include "vp9_mc_template.c"

  2758. #undef mc_luma_dir

  2759. #undef mc_chroma_dir

  2760. #undef FN

  2761. #undef BYTES_PER_PIXEL

  2762. #undef SCALED

  2763. 

  2764. static av_always_inline void mc_luma_scaled(VP9Context *s, vp9_scaled_mc_func smc,

  2765.                                             vp9_mc_func (*mc)[2],

  2766.                                             uint8_t *dst, ptrdiff_t dst_stride,

  2767.                                             const uint8_t *ref, ptrdiff_t ref_stride,

  2768.                                             ThreadFrame *ref_frame,

  2769.                                             ptrdiff_t y, ptrdiff_t x, const VP56mv *in_mv,

  2770.                                             int px, int py, int pw, int ph,

  2771.                                             int bw, int bh, int w, int h, int bytesperpixel,

  2772.                                             const uint16_t *scale, const uint8_t *step)

  2773. {

  2774.     if (s->s.frames[CUR_FRAME].tf.f->width == ref_frame->f->width &&

  2775.         s->s.frames[CUR_FRAME].tf.f->height == ref_frame->f->height) {

  2776.         mc_luma_unscaled(s, mc, dst, dst_stride, ref, ref_stride, ref_frame,

  2777.                          y, x, in_mv, bw, bh, w, h, bytesperpixel);

  2778.     } else {

  2779. #define scale_mv(n, dim) (((int64_t)(n) * scale[dim]) >> 14)

  2780.     int mx, my;

  2781.     int refbw_m1, refbh_m1;

  2782.     int th;

  2783.     VP56mv mv;

  2784. 

  2785.     mv.x = av_clip(in_mv->x, -(x + pw - px + 4) << 3, (s->cols * 8 - x + px + 3) << 3);

  2786.     mv.y = av_clip(in_mv->y, -(y + ph - py + 4) << 3, (s->rows * 8 - y + py + 3) << 3);

  2787.     // BUG libvpx seems to scale the two components separately. This introduces

  2788.     // rounding errors but we have to reproduce them to be exactly compatible

  2789.     // with the output from libvpx...

  2790.     mx = scale_mv(mv.x * 2, 0) + scale_mv(x * 16, 0);

  2791.     my = scale_mv(mv.y * 2, 1) + scale_mv(y * 16, 1);

  2792. 

  2793.     y = my >> 4;

  2794.     x = mx >> 4;

  2795.     ref += y * ref_stride + x * bytesperpixel;

  2796.     mx &= 15;

  2797.     my &= 15;

  2798.     refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;

  2799.     refbh_m1 = ((bh - 1) * step[1] + my) >> 4;

  2800.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2801.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2802.     // the longest loopfilter of the next sbrow

  2803.     th = (y + refbh_m1 + 4 + 7) >> 6;

  2804.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2805.     if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 4 >= h - refbh_m1) {

  2806.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2807.                                  ref - 3 * ref_stride - 3 * bytesperpixel,

  2808.                                  288, ref_stride,

  2809.                                  refbw_m1 + 8, refbh_m1 + 8,

  2810.                                  x - 3, y - 3, w, h);

  2811.         ref = s->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;

  2812.         ref_stride = 288;

  2813.     }

  2814.     smc(dst, dst_stride, ref, ref_stride, bh, mx, my, step[0], step[1]);

  2815.     }

  2816. }

  2817. 

  2818. static av_always_inline void mc_chroma_scaled(VP9Context *s, vp9_scaled_mc_func smc,

  2819.                                               vp9_mc_func (*mc)[2],

  2820.                                               uint8_t *dst_u, uint8_t *dst_v,

  2821.                                               ptrdiff_t dst_stride,

  2822.                                               const uint8_t *ref_u, ptrdiff_t src_stride_u,

  2823.                                               const uint8_t *ref_v, ptrdiff_t src_stride_v,

  2824.                                               ThreadFrame *ref_frame,

  2825.                                               ptrdiff_t y, ptrdiff_t x, const VP56mv *in_mv,

  2826.                                               int px, int py, int pw, int ph,

  2827.                                               int bw, int bh, int w, int h, int bytesperpixel,

  2828.                                               const uint16_t *scale, const uint8_t *step)

  2829. {

  2830.     if (s->s.frames[CUR_FRAME].tf.f->width == ref_frame->f->width &&

  2831.         s->s.frames[CUR_FRAME].tf.f->height == ref_frame->f->height) {

  2832.         mc_chroma_unscaled(s, mc, dst_u, dst_v, dst_stride, ref_u, src_stride_u,

  2833.                            ref_v, src_stride_v, ref_frame,

  2834.                            y, x, in_mv, bw, bh, w, h, bytesperpixel);

  2835.     } else {

  2836.     int mx, my;

  2837.     int refbw_m1, refbh_m1;

  2838.     int th;

  2839.     VP56mv mv;

  2840. 

  2841.     if (s->ss_h) {

  2842.         // BUG https://code.google.com/p/webm/issues/detail?id=820

  2843.         mv.x = av_clip(in_mv->x, -(x + pw - px + 4) << 4, (s->cols * 4 - x + px + 3) << 4);

  2844.         mx = scale_mv(mv.x, 0) + (scale_mv(x * 16, 0) & ~15) + (scale_mv(x * 32, 0) & 15);

  2845.     } else {

  2846.         mv.x = av_clip(in_mv->x, -(x + pw - px + 4) << 3, (s->cols * 8 - x + px + 3) << 3);

  2847.         mx = scale_mv(mv.x << 1, 0) + scale_mv(x * 16, 0);

  2848.     }

  2849.     if (s->ss_v) {

  2850.         // BUG https://code.google.com/p/webm/issues/detail?id=820

  2851.         mv.y = av_clip(in_mv->y, -(y + ph - py + 4) << 4, (s->rows * 4 - y + py + 3) << 4);

  2852.         my = scale_mv(mv.y, 1) + (scale_mv(y * 16, 1) & ~15) + (scale_mv(y * 32, 1) & 15);

  2853.     } else {

  2854.         mv.y = av_clip(in_mv->y, -(y + ph - py + 4) << 3, (s->rows * 8 - y + py + 3) << 3);

  2855.         my = scale_mv(mv.y << 1, 1) + scale_mv(y * 16, 1);

  2856.     }

  2857. #undef scale_mv

  2858.     y = my >> 4;

  2859.     x = mx >> 4;

  2860.     ref_u += y * src_stride_u + x * bytesperpixel;

  2861.     ref_v += y * src_stride_v + x * bytesperpixel;

  2862.     mx &= 15;

  2863.     my &= 15;

  2864.     refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;

  2865.     refbh_m1 = ((bh - 1) * step[1] + my) >> 4;

  2866.     // FIXME bilinear filter only needs 0/1 pixels, not 3/4

  2867.     // we use +7 because the last 7 pixels of each sbrow can be changed in

  2868.     // the longest loopfilter of the next sbrow

  2869.     th = (y + refbh_m1 + 4 + 7) >> (6 - s->ss_v);

  2870.     ff_thread_await_progress(ref_frame, FFMAX(th, 0), 0);

  2871.     if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 4 >= h - refbh_m1) {

  2872.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2873.                                  ref_u - 3 * src_stride_u - 3 * bytesperpixel,

  2874.                                  288, src_stride_u,

  2875.                                  refbw_m1 + 8, refbh_m1 + 8,

  2876.                                  x - 3, y - 3, w, h);

  2877.         ref_u = s->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;

  2878.         smc(dst_u, dst_stride, ref_u, 288, bh, mx, my, step[0], step[1]);

  2879. 

  2880.         s->vdsp.emulated_edge_mc(s->edge_emu_buffer,

  2881.                                  ref_v - 3 * src_stride_v - 3 * bytesperpixel,

  2882.                                  288, src_stride_v,

  2883.                                  refbw_m1 + 8, refbh_m1 + 8,

  2884.                                  x - 3, y - 3, w, h);

  2885.         ref_v = s->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;

  2886.         smc(dst_v, dst_stride, ref_v, 288, bh, mx, my, step[0], step[1]);

  2887.     } else {

  2888.         smc(dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my, step[0], step[1]);

  2889.         smc(dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my, step[0], step[1]);

  2890.     }

  2891.     }

  2892. }

  2893. 

  2894. #define mc_luma_dir(s, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, \

  2895.                     px, py, pw, ph, bw, bh, w, h, i) \

  2896.     mc_luma_scaled(s, s->dsp.s##mc, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \

  2897.                    mv, px, py, pw, ph, bw, bh, w, h, bytesperpixel, \

  2898.                    s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])

  2899. #define mc_chroma_dir(s, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2900.                       row, col, mv, px, py, pw, ph, bw, bh, w, h, i) \

  2901.     mc_chroma_scaled(s, s->dsp.s##mc, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \

  2902.                      row, col, mv, px, py, pw, ph, bw, bh, w, h, bytesperpixel, \

  2903.                      s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])

  2904. #define SCALED 1

  2905. #define FN(x) x##_scaled_8bpp

  2906. #define BYTES_PER_PIXEL 1

  2907. #include "vp9_mc_template.c"

  2908. #undef FN

  2909. #undef BYTES_PER_PIXEL

  2910. #define FN(x) x##_scaled_16bpp

  2911. #define BYTES_PER_PIXEL 2

  2912. #include "vp9_mc_template.c"

  2913. #undef mc_luma_dir

  2914. #undef mc_chroma_dir

  2915. #undef FN

  2916. #undef BYTES_PER_PIXEL

  2917. #undef SCALED

  2918. 

  2919. static av_always_inline void inter_recon(AVCodecContext *ctx, int bytesperpixel)

  2920. {

  2921.     VP9Context *s = ctx->priv_data;

  2922.     VP9Block *b = s->b;

  2923.     int row = s->row, col = s->col;

  2924. 

  2925.     if (s->mvscale[b->ref[0]][0] || (b->comp && s->mvscale[b->ref[1]][0])) {

  2926.         if (bytesperpixel == 1) {

  2927.             inter_pred_scaled_8bpp(ctx);

  2928.         } else {

  2929.             inter_pred_scaled_16bpp(ctx);

  2930.         }

  2931.     } else {

  2932.         if (bytesperpixel == 1) {

  2933.             inter_pred_8bpp(ctx);

  2934.         } else {

  2935.             inter_pred_16bpp(ctx);

  2936.         }

  2937.     }

  2938.     if (!b->skip) {

  2939.         /* mostly copied intra_recon() */

  2940. 

  2941.         int w4 = bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;

  2942.         int h4 = bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);

  2943.         int end_x = FFMIN(2 * (s->cols - col), w4);

  2944.         int end_y = FFMIN(2 * (s->rows - row), h4);

  2945.         int tx = 4 * s->s.h.lossless + b->tx, uvtx = b->uvtx + 4 * s->s.h.lossless;

  2946.         int uvstep1d = 1 << b->uvtx, p;

  2947.         uint8_t *dst = s->dst[0];

  2948. 

  2949.         // y itxfm add

  2950.         for (n = 0, y = 0; y < end_y; y += step1d) {

  2951.             uint8_t *ptr = dst;

  2952.             for (x = 0; x < end_x; x += step1d,

  2953.                  ptr += 4 * step1d * bytesperpixel, n += step) {

  2954.                 int eob = b->tx > TX_8X8 ? AV_RN16A(&s->eob[n]) : s->eob[n];

  2955. 

  2956.                 if (eob)

  2957.                     s->dsp.itxfm_add[tx][DCT_DCT](ptr, s->y_stride,

  2958.                                                   s->block + 16 * n * bytesperpixel, eob);

  2959.             }

  2960.             dst += 4 * s->y_stride * step1d;

  2961.         }

  2962. 

  2963.         // uv itxfm add

  2964.         end_x >>= s->ss_h;

  2965.         end_y >>= s->ss_v;

  2966.         step = 1 << (b->uvtx * 2);

  2967.         for (p = 0; p < 2; p++) {

  2968.             dst = s->dst[p + 1];

  2969.             for (n = 0, y = 0; y < end_y; y += uvstep1d) {

  2970.                 uint8_t *ptr = dst;

  2971.                 for (x = 0; x < end_x; x += uvstep1d,

  2972.                      ptr += 4 * uvstep1d * bytesperpixel, n += step) {

  2973.                     int eob = b->uvtx > TX_8X8 ? AV_RN16A(&s->uveob[p][n]) : s->uveob[p][n];

  2974. 

  2975.                     if (eob)

  2976.                         s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, s->uv_stride,

  2977.                                                         s->uvblock[p] + 16 * n * bytesperpixel, eob);

  2978.                 }

  2979.                 dst += 4 * uvstep1d * s->uv_stride;

  2980.             }

  2981.         }

  2982.     }

  2983. }

  2984. 

  2985. static void inter_recon_8bpp(AVCodecContext *ctx)

  2986. {

  2987.     inter_recon(ctx, 1);

  2988. }

  2989. 

  2990. static void inter_recon_16bpp(AVCodecContext *ctx)

  2991. {

  2992.     inter_recon(ctx, 2);

  2993. }

  2994. 

  2995. static av_always_inline void mask_edges(uint8_t (*mask)[8][4], int ss_h, int ss_v,

  2996.                                         int row_and_7, int col_and_7,

  2997.                                         int w, int h, int col_end, int row_end,

  2998.                                         enum TxfmMode tx, int skip_inter)

  2999. {

  3000.     static const unsigned wide_filter_col_mask[2] = { 0x11, 0x01 };

  3001.     static const unsigned wide_filter_row_mask[2] = { 0x03, 0x07 };

  3002. 

  3003.     // FIXME I'm pretty sure all loops can be replaced by a single LUT if

  3004.     // we make VP9Filter.mask uint64_t (i.e. row/col all single variable)

  3005.     // and make the LUT 5-indexed (bl, bp, is_uv, tx and row/col), and then

  3006.     // use row_and_7/col_and_7 as shifts (1*col_and_7+8*row_and_7)

  3007. 

  3008.     // the intended behaviour of the vp9 loopfilter is to work on 8-pixel

  3009.     // edges. This means that for UV, we work on two subsampled blocks at

  3010.     // a time, and we only use the topleft block's mode information to set

  3011.     // things like block strength. Thus, for any block size smaller than

  3012.     // 16x16, ignore the odd portion of the block.

  3013.     if (tx == TX_4X4 && (ss_v | ss_h)) {

  3014.         if (h == ss_v) {

  3015.             if (row_and_7 & 1)

  3016.                 return;

  3017.             if (!row_end)

  3018.                 h += 1;

  3019.         }

  3020.         if (w == ss_h) {

  3021.             if (col_and_7 & 1)

  3022.                 return;

  3023.             if (!col_end)

  3024.                 w += 1;

  3025.         }

  3026.     }

  3027. 

  3028.     if (tx == TX_4X4 && !skip_inter) {

  3029.         int t = 1 << col_and_7, m_col = (t << w) - t, y;

  3030.         // on 32-px edges, use the 8-px wide loopfilter; else, use 4-px wide

  3031.         int m_row_8 = m_col & wide_filter_col_mask[ss_h], m_row_4 = m_col - m_row_8;

  3032. 

  3033.         for (y = row_and_7; y < h + row_and_7; y++) {

  3034.             int col_mask_id = 2 - !(y & wide_filter_row_mask[ss_v]);

  3035. 

  3036.             mask[0][y][1] |= m_row_8;

  3037.             mask[0][y][2] |= m_row_4;

  3038.             // for odd lines, if the odd col is not being filtered,

  3039.             // skip odd row also:

  3040.             // .---. <-- a

  3041.             // |   |

  3042.             // |___| <-- b

  3043.             // ^   ^

  3044.             // c   d

  3045.             //

  3046.             // if a/c are even row/col and b/d are odd, and d is skipped,

  3047.             // e.g. right edge of size-66x66.webm, then skip b also (bug)

  3048.             if ((ss_h & ss_v) && (col_end & 1) && (y & 1)) {

  3049.                 mask[1][y][col_mask_id] |= (t << (w - 1)) - t;

  3050.             } else {

  3051.                 mask[1][y][col_mask_id] |= m_col;

  3052.             }

  3053.             if (!ss_h)

  3054.                 mask[0][y][3] |= m_col;

  3055.             if (!ss_v) {

  3056.                 if (ss_h && (col_end & 1))

  3057.                     mask[1][y][3] |= (t << (w - 1)) - t;

  3058.                 else

  3059.                     mask[1][y][3] |= m_col;

  3060.             }

  3061.         }

  3062.     } else {

  3063.         int y, t = 1 << col_and_7, m_col = (t << w) - t;

  3064. 

  3065.         if (!skip_inter) {

  3066.             int mask_id = (tx == TX_8X8);

  3067.             static const unsigned masks[4] = { 0xff, 0x55, 0x11, 0x01 };

  3068.             int l2 = tx + ss_h - 1, step1d;

  3069.             int m_row = m_col & masks[l2];

  3070. 

  3071.             // at odd UV col/row edges tx16/tx32 loopfilter edges, force

  3072.             // 8wd loopfilter to prevent going off the visible edge.

  3073.             if (ss_h && tx > TX_8X8 && (w ^ (w - 1)) == 1) {

  3074.                 int m_row_16 = ((t << (w - 1)) - t) & masks[l2];

  3075.                 int m_row_8 = m_row - m_row_16;

  3076. 

  3077.                 for (y = row_and_7; y < h + row_and_7; y++) {

  3078.                     mask[0][y][0] |= m_row_16;

  3079.                     mask[0][y][1] |= m_row_8;

  3080.                 }

  3081.             } else {

  3082.                 for (y = row_and_7; y < h + row_and_7; y++)

  3083.                     mask[0][y][mask_id] |= m_row;

  3084.             }

  3085. 

  3086.             l2 = tx + ss_v - 1;

  3087.             step1d = 1 << l2;

  3088.             if (ss_v && tx > TX_8X8 && (h ^ (h - 1)) == 1) {

  3089.                 for (y = row_and_7; y < h + row_and_7 - 1; y += step1d)

  3090.                     mask[1][y][0] |= m_col;

  3091.                 if (y - row_and_7 == h - 1)

  3092.                     mask[1][y][1] |= m_col;

  3093.             } else {

  3094.                 for (y = row_and_7; y < h + row_and_7; y += step1d)

  3095.                     mask[1][y][mask_id] |= m_col;

  3096.             }

  3097.         } else if (tx != TX_4X4) {

  3098.             int mask_id;

  3099. 

  3100.             mask_id = (tx == TX_8X8) || (h == ss_v);

  3101.             mask[1][row_and_7][mask_id] |= m_col;

  3102.             mask_id = (tx == TX_8X8) || (w == ss_h);

  3103.             for (y = row_and_7; y < h + row_and_7; y++)

  3104.                 mask[0][y][mask_id] |= t;

  3105.         } else {

  3106.             int t8 = t & wide_filter_col_mask[ss_h], t4 = t - t8;

  3107. 

  3108.             for (y = row_and_7; y < h + row_and_7; y++) {

  3109.                 mask[0][y][2] |= t4;

  3110.                 mask[0][y][1] |= t8;

  3111.             }

  3112.             mask[1][row_and_7][2 - !(row_and_7 & wide_filter_row_mask[ss_v])] |= m_col;

  3113.         }

  3114.     }

  3115. }

  3116. 

  3117. static void decode_b(AVCodecContext *ctx, int row, int col,

  3118.                      struct VP9Filter *lflvl, ptrdiff_t yoff, ptrdiff_t uvoff,

  3119.                      enum BlockLevel bl, enum BlockPartition bp)

  3120. {

  3121.     VP9Context *s = ctx->priv_data;

  3122.     VP9Block *b = s->b;

  3123.     enum BlockSize bs = bl * 3 + bp;

  3124.     int bytesperpixel = s->bytesperpixel;

  3125.     int w4 = bwh_tab[1][bs][0], h4 = bwh_tab[1][bs][1], lvl;

  3126.     int emu[2];

  3127.     AVFrame *f = s->s.frames[CUR_FRAME].tf.f;

  3128. 

  3129.     s->row = row;

  3130.     s->row7 = row & 7;

  3131.     s->col = col;

  3132.     s->col7 = col & 7;

  3133.     s->min_mv.x = -(128 + col * 64);

  3134.     s->min_mv.y = -(128 + row * 64);

  3135.     s->max_mv.x = 128 + (s->cols - col - w4) * 64;

  3136.     s->max_mv.y = 128 + (s->rows - row - h4) * 64;

  3137.     if (s->pass < 2) {

  3138.         b->bs = bs;

  3139.         b->bl = bl;

  3140.         b->bp = bp;

  3141.         decode_mode(ctx);

  3142.         b->uvtx = b->tx - ((s->ss_h && w4 * 2 == (1 << b->tx)) ||

  3143.                            (s->ss_v && h4 * 2 == (1 << b->tx)));

  3144. 

  3145.         if (!b->skip) {

  3146.             int has_coeffs;

  3147. 

  3148.             if (bytesperpixel == 1) {

  3149.                 has_coeffs = decode_coeffs_8bpp(ctx);

  3150.             } else {

  3151.                 has_coeffs = decode_coeffs_16bpp(ctx);

  3152.             }

  3153.             if (!has_coeffs && b->bs <= BS_8x8 && !b->intra) {

  3154.                 b->skip = 1;

  3155.                 memset(&s->above_skip_ctx[col], 1, w4);

  3156.                 memset(&s->left_skip_ctx[s->row7], 1, h4);

  3157.             }

  3158.         } else {

  3159.             int row7 = s->row7;

  3160. 

  3161. #define SPLAT_ZERO_CTX(v, n) \

  3162.     switch (n) { \

  3163.     case 1:  v = 0;          break; \

  3164.     case 2:  AV_ZERO16(&v);  break; \

  3165.     case 4:  AV_ZERO32(&v);  break; \

  3166.     case 8:  AV_ZERO64(&v);  break; \

  3167.     case 16: AV_ZERO128(&v); break; \

  3168.     }

  3169. #define SPLAT_ZERO_YUV(dir, var, off, n, dir2) \

  3170.     do { \

  3171.         SPLAT_ZERO_CTX(s->dir##_y_##var[off * 2], n * 2); \

  3172.         if (s->ss_##dir2) { \

  3173.             SPLAT_ZERO_CTX(s->dir##_uv_##var[0][off], n); \

  3174.             SPLAT_ZERO_CTX(s->dir##_uv_##var[1][off], n); \

  3175.         } else { \

  3176.             SPLAT_ZERO_CTX(s->dir##_uv_##var[0][off * 2], n * 2); \

  3177.             SPLAT_ZERO_CTX(s->dir##_uv_##var[1][off * 2], n * 2); \

  3178.         } \

  3179.     } while (0)

  3180. 

  3181.             switch (w4) {

  3182.             case 1: SPLAT_ZERO_YUV(above, nnz_ctx, col, 1, h); break;

  3183.             case 2: SPLAT_ZERO_YUV(above, nnz_ctx, col, 2, h); break;

  3184.             case 4: SPLAT_ZERO_YUV(above, nnz_ctx, col, 4, h); break;

  3185.             case 8: SPLAT_ZERO_YUV(above, nnz_ctx, col, 8, h); break;

  3186.             }

  3187.             switch (h4) {

  3188.             case 1: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 1, v); break;

  3189.             case 2: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 2, v); break;

  3190.             case 4: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 4, v); break;

  3191.             case 8: SPLAT_ZERO_YUV(left, nnz_ctx, row7, 8, v); break;

  3192.             }

  3193.         }

  3194. 

  3195.         if (s->pass == 1) {

  3196.             s->b++;

  3197.             s->block += w4 * h4 * 64 * bytesperpixel;

  3198.             s->uvblock[0] += w4 * h4 * 64 * bytesperpixel >> (s->ss_h + s->ss_v);

  3199.             s->uvblock[1] += w4 * h4 * 64 * bytesperpixel >> (s->ss_h + s->ss_v);

  3200.             s->eob += 4 * w4 * h4;

  3201.             s->uveob[0] += 4 * w4 * h4 >> (s->ss_h + s->ss_v);

  3202.             s->uveob[1] += 4 * w4 * h4 >> (s->ss_h + s->ss_v);

  3203. 

  3204.             return;

  3205.         }

  3206.     }

  3207. 

  3208.     // emulated overhangs if the stride of the target buffer can't hold. This

  3209.     // makes it possible to support emu-edge and so on even if we have large block

  3210.     // overhangs

  3211.     emu[0] = (col + w4) * 8 * bytesperpixel > f->linesize[0] ||

  3212.              (row + h4) > s->rows;

  3213.     emu[1] = ((col + w4) * 8 >> s->ss_h) * bytesperpixel > f->linesize[1] ||

  3214.              (row + h4) > s->rows;

  3215.     if (emu[0]) {

  3216.         s->dst[0] = s->tmp_y;

  3217.         s->y_stride = 128;

  3218.     } else {

  3219.         s->dst[0] = f->data[0] + yoff;

  3220.         s->y_stride = f->linesize[0];

  3221.     }

  3222.     if (emu[1]) {

  3223.         s->dst[1] = s->tmp_uv[0];

  3224.         s->dst[2] = s->tmp_uv[1];

  3225.         s->uv_stride = 128;

  3226.     } else {

  3227.         s->dst[1] = f->data[1] + uvoff;

  3228.         s->dst[2] = f->data[2] + uvoff;

  3229.         s->uv_stride = f->linesize[1];

  3230.     }

  3231.     if (b->intra) {

  3232.         if (s->bpp > 8) {

  3233.             intra_recon_16bpp(ctx, yoff, uvoff);

  3234.         } else {

  3235.             intra_recon_8bpp(ctx, yoff, uvoff);

  3236.         }

  3237.     } else {

  3238.         if (s->bpp > 8) {

  3239.             inter_recon_16bpp(ctx);

  3240.         } else {

  3241.             inter_recon_8bpp(ctx);

  3242.         }

  3243.     }

  3244.     if (emu[0]) {

  3245.         int w = FFMIN(s->cols - col, w4) * 8, h = FFMIN(s->rows - row, h4) * 8, n, o = 0;

  3246. 

  3247.         for (n = 0; o < w; n++) {

  3248.             int bw = 64 >> n;

  3249. 

  3250.             av_assert2(n <= 4);

  3251.             if (w & bw) {

  3252.                 s->dsp.mc[n][0][0][0][0](f->data[0] + yoff + o * bytesperpixel, f->linesize[0],

  3253.                                          s->tmp_y + o * bytesperpixel, 128, h, 0, 0);

  3254.                 o += bw;

  3255.             }

  3256.         }

  3257.     }

  3258.     if (emu[1]) {

  3259.         int w = FFMIN(s->cols - col, w4) * 8 >> s->ss_h;

  3260.         int h = FFMIN(s->rows - row, h4) * 8 >> s->ss_v, n, o = 0;

  3261. 

  3262.         for (n = s->ss_h; o < w; n++) {

  3263.             int bw = 64 >> n;

  3264. 

  3265.             av_assert2(n <= 4);

  3266.             if (w & bw) {

  3267.                 s->dsp.mc[n][0][0][0][0](f->data[1] + uvoff + o * bytesperpixel, f->linesize[1],

  3268.                                          s->tmp_uv[0] + o * bytesperpixel, 128, h, 0, 0);

  3269.                 s->dsp.mc[n][0][0][0][0](f->data[2] + uvoff + o * bytesperpixel, f->linesize[2],

  3270.                                          s->tmp_uv[1] + o * bytesperpixel, 128, h, 0, 0);

  3271.                 o += bw;

  3272.             }

  3273.         }

  3274.     }

  3275. 

  3276.     // pick filter level and find edges to apply filter to

  3277.     if (s->s.h.filter.level &&

  3278.         (lvl = s->s.h.segmentation.feat[b->seg_id].lflvl[b->intra ? 0 : b->ref[0] + 1]

  3279.                                                       [b->mode[3] != ZEROMV]) > 0) {

  3280.         int x_end = FFMIN(s->cols - col, w4), y_end = FFMIN(s->rows - row, h4);

  3281.         int skip_inter = !b->intra && b->skip, col7 = s->col7, row7 = s->row7;

  3282. 

  3283.         setctx_2d(&lflvl->level[row7 * 8 + col7], w4, h4, 8, lvl);

  3284.         mask_edges(lflvl->mask[0], 0, 0, row7, col7, x_end, y_end, 0, 0, b->tx, skip_inter);

  3285.         if (s->ss_h || s->ss_v)

  3286.             mask_edges(lflvl->mask[1], s->ss_h, s->ss_v, row7, col7, x_end, y_end,

  3287.                        s->cols & 1 && col + w4 >= s->cols ? s->cols & 7 : 0,

  3288.                        s->rows & 1 && row + h4 >= s->rows ? s->rows & 7 : 0,

  3289.                        b->uvtx, skip_inter);

  3290. 

  3291.         if (!s->filter_lut.lim_lut[lvl]) {

  3292.             int sharp = s->s.h.filter.sharpness;

  3293.             int limit = lvl;

  3294. 

  3295.             if (sharp > 0) {

  3296.                 limit >>= (sharp + 3) >> 2;

  3297.                 limit = FFMIN(limit, 9 - sharp);

  3298.             }

  3299.             limit = FFMAX(limit, 1);

  3300. 

  3301.             s->filter_lut.lim_lut[lvl] = limit;

  3302.             s->filter_lut.mblim_lut[lvl] = 2 * (lvl + 2) + limit;

  3303.         }

  3304.     }

  3305. 

  3306.     if (s->pass == 2) {

  3307.         s->b++;

  3308.         s->block += w4 * h4 * 64 * bytesperpixel;

  3309.         s->uvblock[0] += w4 * h4 * 64 * bytesperpixel >> (s->ss_v + s->ss_h);

  3310.         s->uvblock[1] += w4 * h4 * 64 * bytesperpixel >> (s->ss_v + s->ss_h);

  3311.         s->eob += 4 * w4 * h4;

  3312.         s->uveob[0] += 4 * w4 * h4 >> (s->ss_v + s->ss_h);

  3313.         s->uveob[1] += 4 * w4 * h4 >> (s->ss_v + s->ss_h);

  3314.     }

  3315. }

  3316. 

  3317. static void decode_sb(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  3318.                       ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  3319. {

  3320.     VP9Context *s = ctx->priv_data;

  3321.     int c = ((s->above_partition_ctx[col] >> (3 - bl)) & 1) |

  3322.             (((s->left_partition_ctx[row & 0x7] >> (3 - bl)) & 1) << 1);

  3323.     const uint8_t *p = s->s.h.keyframe || s->s.h.intraonly ? vp9_default_kf_partition_probs[bl][c] :

  3324.                                                      s->prob.p.partition[bl][c];

  3325.     enum BlockPartition bp;

  3326.     ptrdiff_t hbs = 4 >> bl;

  3327.     AVFrame *f = s->s.frames[CUR_FRAME].tf.f;

  3328.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  3329.     int bytesperpixel = s->bytesperpixel;

  3330. 

  3331.     if (bl == BL_8X8) {

  3332.         bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  3333.         decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3334.     } else if (col + hbs < s->cols) { // FIXME why not <=?

  3335.         if (row + hbs < s->rows) { // FIXME why not <=?

  3336.             bp = vp8_rac_get_tree(&s->c, vp9_partition_tree, p);

  3337.             switch (bp) {

  3338.             case PARTITION_NONE:

  3339.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3340.                 break;

  3341.             case PARTITION_H:

  3342.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3343.                 yoff  += hbs * 8 * y_stride;

  3344.                 uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3345.                 decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, bl, bp);

  3346.                 break;

  3347.             case PARTITION_V:

  3348.                 decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3349.                 yoff  += hbs * 8 * bytesperpixel;

  3350.                 uvoff += hbs * 8 * bytesperpixel >> s->ss_h;

  3351.                 decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, bl, bp);

  3352.                 break;

  3353.             case PARTITION_SPLIT:

  3354.                 decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3355.                 decode_sb(ctx, row, col + hbs, lflvl,

  3356.                           yoff + 8 * hbs * bytesperpixel,

  3357.                           uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);

  3358.                 yoff  += hbs * 8 * y_stride;

  3359.                 uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3360.                 decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3361.                 decode_sb(ctx, row + hbs, col + hbs, lflvl,

  3362.                           yoff + 8 * hbs * bytesperpixel,

  3363.                           uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);

  3364.                 break;

  3365.             default:

  3366.                 av_assert0(0);

  3367.             }

  3368.         } else if (vp56_rac_get_prob_branchy(&s->c, p[1])) {

  3369.             bp = PARTITION_SPLIT;

  3370.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3371.             decode_sb(ctx, row, col + hbs, lflvl,

  3372.                       yoff + 8 * hbs * bytesperpixel,

  3373.                       uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);

  3374.         } else {

  3375.             bp = PARTITION_H;

  3376.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3377.         }

  3378.     } else if (row + hbs < s->rows) { // FIXME why not <=?

  3379.         if (vp56_rac_get_prob_branchy(&s->c, p[2])) {

  3380.             bp = PARTITION_SPLIT;

  3381.             decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3382.             yoff  += hbs * 8 * y_stride;

  3383.             uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3384.             decode_sb(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3385.         } else {

  3386.             bp = PARTITION_V;

  3387.             decode_b(ctx, row, col, lflvl, yoff, uvoff, bl, bp);

  3388.         }

  3389.     } else {

  3390.         bp = PARTITION_SPLIT;

  3391.         decode_sb(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3392.     }

  3393.     s->counts.partition[bl][c][bp]++;

  3394. }

  3395. 

  3396. static void decode_sb_mem(AVCodecContext *ctx, int row, int col, struct VP9Filter *lflvl,

  3397.                           ptrdiff_t yoff, ptrdiff_t uvoff, enum BlockLevel bl)

  3398. {

  3399.     VP9Context *s = ctx->priv_data;

  3400.     VP9Block *b = s->b;

  3401.     ptrdiff_t hbs = 4 >> bl;

  3402.     AVFrame *f = s->s.frames[CUR_FRAME].tf.f;

  3403.     ptrdiff_t y_stride = f->linesize[0], uv_stride = f->linesize[1];

  3404.     int bytesperpixel = s->bytesperpixel;

  3405. 

  3406.     if (bl == BL_8X8) {

  3407.         av_assert2(b->bl == BL_8X8);

  3408.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  3409.     } else if (s->b->bl == bl) {

  3410.         decode_b(ctx, row, col, lflvl, yoff, uvoff, b->bl, b->bp);

  3411.         if (b->bp == PARTITION_H && row + hbs < s->rows) {

  3412.             yoff  += hbs * 8 * y_stride;

  3413.             uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3414.             decode_b(ctx, row + hbs, col, lflvl, yoff, uvoff, b->bl, b->bp);

  3415.         } else if (b->bp == PARTITION_V && col + hbs < s->cols) {

  3416.             yoff  += hbs * 8 * bytesperpixel;

  3417.             uvoff += hbs * 8 * bytesperpixel >> s->ss_h;

  3418.             decode_b(ctx, row, col + hbs, lflvl, yoff, uvoff, b->bl, b->bp);

  3419.         }

  3420.     } else {

  3421.         decode_sb_mem(ctx, row, col, lflvl, yoff, uvoff, bl + 1);

  3422.         if (col + hbs < s->cols) { // FIXME why not <=?

  3423.             if (row + hbs < s->rows) {

  3424.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff + 8 * hbs * bytesperpixel,

  3425.                               uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);

  3426.                 yoff  += hbs * 8 * y_stride;

  3427.                 uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3428.                 decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3429.                 decode_sb_mem(ctx, row + hbs, col + hbs, lflvl,

  3430.                               yoff + 8 * hbs * bytesperpixel,

  3431.                               uvoff + (8 * hbs * bytesperpixel >> s->ss_h), bl + 1);

  3432.             } else {

  3433.                 yoff  += hbs * 8 * bytesperpixel;

  3434.                 uvoff += hbs * 8 * bytesperpixel >> s->ss_h;

  3435.                 decode_sb_mem(ctx, row, col + hbs, lflvl, yoff, uvoff, bl + 1);

  3436.             }

  3437.         } else if (row + hbs < s->rows) {

  3438.             yoff  += hbs * 8 * y_stride;

  3439.             uvoff += hbs * 8 * uv_stride >> s->ss_v;

  3440.             decode_sb_mem(ctx, row + hbs, col, lflvl, yoff, uvoff, bl + 1);

  3441.         }

  3442.     }

  3443. }

  3444. 

  3445. static av_always_inline void filter_plane_cols(VP9Context *s, int col, int ss_h, int ss_v,

  3446.                                                uint8_t *lvl, uint8_t (*mask)[4],

  3447.                                                uint8_t *dst, ptrdiff_t ls)

  3448. {

  3449.     int y, x, bytesperpixel = s->bytesperpixel;

  3450. 

  3451.     // filter edges between columns (e.g. block1 | block2)

  3452.     for (y = 0; y < 8; y += 2 << ss_v, dst += 16 * ls, lvl += 16 << ss_v) {

  3453.         uint8_t *ptr = dst, *l = lvl, *hmask1 = mask[y], *hmask2 = mask[y + 1 + ss_v];

  3454.         unsigned hm1 = hmask1[0] | hmask1[1] | hmask1[2], hm13 = hmask1[3];

  3455.         unsigned hm2 = hmask2[1] | hmask2[2], hm23 = hmask2[3];

  3456.         unsigned hm = hm1 | hm2 | hm13 | hm23;

  3457. 

  3458.         for (x = 1; hm & ~(x - 1); x <<= 1, ptr += 8 * bytesperpixel >> ss_h) {

  3459.             if (col || x > 1) {

  3460.                 if (hm1 & x) {

  3461.                     int L = *l, H = L >> 4;

  3462.                     int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

  3463. 

  3464.                     if (hmask1[0] & x) {

  3465.                         if (hmask2[0] & x) {

  3466.                             av_assert2(l[8 << ss_v] == L);

  3467.                             s->dsp.loop_filter_16[0](ptr, ls, E, I, H);

  3468.                         } else {

  3469.                             s->dsp.loop_filter_8[2][0](ptr, ls, E, I, H);

  3470.                         }

  3471.                     } else if (hm2 & x) {

  3472.                         L = l[8 << ss_v];

  3473.                         H |= (L >> 4) << 8;

  3474.                         E |= s->filter_lut.mblim_lut[L] << 8;

  3475.                         I |= s->filter_lut.lim_lut[L] << 8;

  3476.                         s->dsp.loop_filter_mix2[!!(hmask1[1] & x)]

  3477.                                                [!!(hmask2[1] & x)]

  3478.                                                [0](ptr, ls, E, I, H);

  3479.                     } else {

  3480.                         s->dsp.loop_filter_8[!!(hmask1[1] & x)]

  3481.                                             [0](ptr, ls, E, I, H);

  3482.                     }

  3483.                 } else if (hm2 & x) {

  3484.                     int L = l[8 << ss_v], H = L >> 4;

  3485.                     int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

  3486. 

  3487.                     s->dsp.loop_filter_8[!!(hmask2[1] & x)]

  3488.                                         [0](ptr + 8 * ls, ls, E, I, H);

  3489.                 }

  3490.             }

  3491.             if (ss_h) {

  3492.                 if (x & 0xAA)

  3493.                     l += 2;

  3494.             } else {

  3495.                 if (hm13 & x) {

  3496.                     int L = *l, H = L >> 4;

  3497.                     int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

  3498. 

  3499.                     if (hm23 & x) {

  3500.                         L = l[8 << ss_v];

  3501.                         H |= (L >> 4) << 8;

  3502.                         E |= s->filter_lut.mblim_lut[L] << 8;

  3503.                         I |= s->filter_lut.lim_lut[L] << 8;

  3504.                         s->dsp.loop_filter_mix2[0][0][0](ptr + 4 * bytesperpixel, ls, E, I, H);

  3505.                     } else {

  3506.                         s->dsp.loop_filter_8[0][0](ptr + 4 * bytesperpixel, ls, E, I, H);

  3507.                     }

  3508.                 } else if (hm23 & x) {

  3509.                     int L = l[8 << ss_v], H = L >> 4;

  3510.                     int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

  3511. 

  3512.                     s->dsp.loop_filter_8[0][0](ptr + 8 * ls + 4 * bytesperpixel, ls, E, I, H);

  3513.                 }

  3514.                 l++;

  3515.             }

  3516.         }

  3517.     }

  3518. }

  3519. 

  3520. static av_always_inline void filter_plane_rows(VP9Context *s, int row, int ss_h, int ss_v,

  3521.                                                uint8_t *lvl, uint8_t (*mask)[4],

  3522.                                                uint8_t *dst, ptrdiff_t ls)

  3523. {

  3524.     int y, x, bytesperpixel = s->bytesperpixel;

  3525. 

  3526.     //                                 block1

  3527.     // filter edges between rows (e.g. ------)

  3528.     //                                 block2

  3529.     for (y = 0; y < 8; y++, dst += 8 * ls >> ss_v) {

  3530.         uint8_t *ptr = dst, *l = lvl, *vmask = mask[y];

  3531.         unsigned vm = vmask[0] | vmask[1] | vmask[2], vm3 = vmask[3];

  3532. 

  3533.         for (x = 1; vm & ~(x - 1); x <<= (2 << ss_h), ptr += 16 * bytesperpixel, l += 2 << ss_h) {

  3534.             if (row || y) {

  3535.                 if (vm & x) {

  3536.                     int L = *l, H = L >> 4;

  3537.                     int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

  3538. 

  3539.                     if (vmask[0] & x) {

  3540.                         if (vmask[0] & (x << (1 + ss_h))) {

  3541.                             av_assert2(l[1 + ss_h] == L);

  3542.                             s->dsp.loop_filter_16[1](ptr, ls, E, I, H);

  3543.                         } else {

  3544.                             s->dsp.loop_filter_8[2][1](ptr, ls, E, I, H);

  3545.                         }

  3546.                     } else if (vm & (x << (1 + ss_h))) {

  3547.                         L = l[1 + ss_h];

  3548.                         H |= (L >> 4) << 8;

  3549.                         E |= s->filter_lut.mblim_lut[L] << 8;

  3550.                         I |= s->filter_lut.lim_lut[L] << 8;

  3551.                         s->dsp.loop_filter_mix2[!!(vmask[1] &  x)]

  3552.                                                [!!(vmask[1] & (x << (1 + ss_h)))]

  3553.                                                [1](ptr, ls, E, I, H);

  3554.                     } else {

  3555.                         s->dsp.loop_filter_8[!!(vmask[1] & x)]

  3556.                                             [1](ptr, ls, E, I, H);

  3557.                     }

  3558.                 } else if (vm & (x << (1 + ss_h))) {

  3559.                     int L = l[1 + ss_h], H = L >> 4;

  3560.                     int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

  3561. 

  3562.                     s->dsp.loop_filter_8[!!(vmask[1] & (x << (1 + ss_h)))]

  3563.                                         [1](ptr + 8 * bytesperpixel, ls, E, I, H);

  3564.                 }

  3565.             }

  3566.             if (!ss_v) {

  3567.                 if (vm3 & x) {

  3568.                     int L = *l, H = L >> 4;

  3569.                     int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

  3570. 

  3571.                     if (vm3 & (x << (1 + ss_h))) {

  3572.                         L = l[1 + ss_h];

  3573.                         H |= (L >> 4) << 8;

  3574.                         E |= s->filter_lut.mblim_lut[L] << 8;

  3575.                         I |= s->filter_lut.lim_lut[L] << 8;

  3576.                         s->dsp.loop_filter_mix2[0][0][1](ptr + ls * 4, ls, E, I, H);

  3577.                     } else {

  3578.                         s->dsp.loop_filter_8[0][1](ptr + ls * 4, ls, E, I, H);

  3579.                     }

  3580.                 } else if (vm3 & (x << (1 + ss_h))) {

  3581.                     int L = l[1 + ss_h], H = L >> 4;

  3582.                     int E = s->filter_lut.mblim_lut[L], I = s->filter_lut.lim_lut[L];

  3583. 

  3584.                     s->dsp.loop_filter_8[0][1](ptr + ls * 4 + 8 * bytesperpixel, ls, E, I, H);

  3585.                 }

  3586.             }

  3587.         }

  3588.         if (ss_v) {

  3589.             if (y & 1)

  3590.                 lvl += 16;

  3591.         } else {

  3592.             lvl += 8;

  3593.         }

  3594.     }

  3595. }

  3596. 

  3597. static void loopfilter_sb(AVCodecContext *ctx, struct VP9Filter *lflvl,

  3598.                           int row, int col, ptrdiff_t yoff, ptrdiff_t uvoff)

  3599. {

  3600.     VP9Context *s = ctx->priv_data;

  3601.     AVFrame *f = s->s.frames[CUR_FRAME].tf.f;

  3602.     uint8_t *dst = f->data[0] + yoff;

  3603.     ptrdiff_t ls_y = f->linesize[0], ls_uv = f->linesize[1];

  3604.     uint8_t (*uv_masks)[8][4] = lflvl->mask[s->ss_h | s->ss_v];

  3605.     int p;

  3606. 

  3607.     // FIXME in how far can we interleave the v/h loopfilter calls? E.g.

  3608.     // if you think of them as acting on a 8x8 block max, we can interleave

  3609.     // each v/h within the single x loop, but that only works if we work on

  3610.     // 8 pixel blocks, and we won't always do that (we want at least 16px

  3611.     // to use SSE2 optimizations, perhaps 32 for AVX2)

  3612. 

  3613.     filter_plane_cols(s, col, 0, 0, lflvl->level, lflvl->mask[0][0], dst, ls_y);

  3614.     filter_plane_rows(s, row, 0, 0, lflvl->level, lflvl->mask[0][1], dst, ls_y);

  3615. 

  3616.     for (p = 0; p < 2; p++) {

  3617.         dst = f->data[1 + p] + uvoff;

  3618.         filter_plane_cols(s, col, s->ss_h, s->ss_v, lflvl->level, uv_masks[0], dst, ls_uv);

  3619.         filter_plane_rows(s, row, s->ss_h, s->ss_v, lflvl->level, uv_masks[1], dst, ls_uv);

  3620.     }

  3621. }

  3622. 

  3623. static void set_tile_offset(int *start, int *end, int idx, int log2_n, int n)

  3624. {

  3625.     int sb_start = ( idx      * n) >> log2_n;

  3626.     int sb_end   = ((idx + 1) * n) >> log2_n;

  3627.     *start = FFMIN(sb_start, n) << 3;

  3628.     *end   = FFMIN(sb_end,   n) << 3;

  3629. }

  3630. 

  3631. static av_always_inline void adapt_prob(uint8_t *p, unsigned ct0, unsigned ct1,

  3632.                                         int max_count, int update_factor)

  3633. {

  3634.     unsigned ct = ct0 + ct1, p2, p1;

  3635. 

  3636.     if (!ct)

  3637.         return;

  3638. 

  3639.     p1 = *p;

  3640.     p2 = ((ct0 << 8) + (ct >> 1)) / ct;

  3641.     p2 = av_clip(p2, 1, 255);

  3642.     ct = FFMIN(ct, max_count);

  3643.     update_factor = FASTDIV(update_factor * ct, max_count);

  3644. 

  3645.     // (p1 * (256 - update_factor) + p2 * update_factor + 128) >> 8

  3646.     *p = p1 + (((p2 - p1) * update_factor + 128) >> 8);

  3647. }

  3648. 

  3649. static void adapt_probs(VP9Context *s)

  3650. {

  3651.     int i, j, k, l, m;

  3652.     prob_context *p = &s->prob_ctx[s->s.h.framectxid].p;

  3653.     int uf = (s->s.h.keyframe || s->s.h.intraonly || !s->last_keyframe) ? 112 : 128;

  3654. 

  3655.     // coefficients

  3656.     for (i = 0; i < 4; i++)

  3657.         for (j = 0; j < 2; j++)

  3658.             for (k = 0; k < 2; k++)

  3659.                 for (l = 0; l < 6; l++)

  3660.                     for (m = 0; m < 6; m++) {

  3661.                         uint8_t *pp = s->prob_ctx[s->s.h.framectxid].coef[i][j][k][l][m];

  3662.                         unsigned *e = s->counts.eob[i][j][k][l][m];

  3663.                         unsigned *c = s->counts.coef[i][j][k][l][m];

  3664. 

  3665.                         if (l == 0 && m >= 3) // dc only has 3 pt

  3666.                             break;

  3667. 

  3668.                         adapt_prob(&pp[0], e[0], e[1], 24, uf);

  3669.                         adapt_prob(&pp[1], c[0], c[1] + c[2], 24, uf);

  3670.                         adapt_prob(&pp[2], c[1], c[2], 24, uf);

  3671.                     }

  3672. 

  3673.     if (s->s.h.keyframe || s->s.h.intraonly) {

  3674.         memcpy(p->skip,  s->prob.p.skip,  sizeof(p->skip));

  3675.         memcpy(p->tx32p, s->prob.p.tx32p, sizeof(p->tx32p));

  3676.         memcpy(p->tx16p, s->prob.p.tx16p, sizeof(p->tx16p));

  3677.         memcpy(p->tx8p,  s->prob.p.tx8p,  sizeof(p->tx8p));

  3678.         return;

  3679.     }

  3680. 

  3681.     // skip flag

  3682.     for (i = 0; i < 3; i++)

  3683.         adapt_prob(&p->skip[i], s->counts.skip[i][0], s->counts.skip[i][1], 20, 128);

  3684. 

  3685.     // intra/inter flag

  3686.     for (i = 0; i < 4; i++)

  3687.         adapt_prob(&p->intra[i], s->counts.intra[i][0], s->counts.intra[i][1], 20, 128);

  3688. 

  3689.     // comppred flag

  3690.     if (s->s.h.comppredmode == PRED_SWITCHABLE) {

  3691.       for (i = 0; i < 5; i++)

  3692.           adapt_prob(&p->comp[i], s->counts.comp[i][0], s->counts.comp[i][1], 20, 128);

  3693.     }

  3694. 

  3695.     // reference frames

  3696.     if (s->s.h.comppredmode != PRED_SINGLEREF) {

  3697.       for (i = 0; i < 5; i++)

  3698.           adapt_prob(&p->comp_ref[i], s->counts.comp_ref[i][0],

  3699.                      s->counts.comp_ref[i][1], 20, 128);

  3700.     }

  3701. 

  3702.     if (s->s.h.comppredmode != PRED_COMPREF) {

  3703.       for (i = 0; i < 5; i++) {

  3704.           uint8_t *pp = p->single_ref[i];

  3705.           unsigned (*c)[2] = s->counts.single_ref[i];

  3706. 

  3707.           adapt_prob(&pp[0], c[0][0], c[0][1], 20, 128);

  3708.           adapt_prob(&pp[1], c[1][0], c[1][1], 20, 128);

  3709.       }

  3710.     }

  3711. 

  3712.     // block partitioning

  3713.     for (i = 0; i < 4; i++)

  3714.         for (j = 0; j < 4; j++) {

  3715.             uint8_t *pp = p->partition[i][j];

  3716.             unsigned *c = s->counts.partition[i][j];

  3717. 

  3718.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3719.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3720.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3721.         }

  3722. 

  3723.     // tx size

  3724.     if (s->s.h.txfmmode == TX_SWITCHABLE) {

  3725.       for (i = 0; i < 2; i++) {

  3726.           unsigned *c16 = s->counts.tx16p[i], *c32 = s->counts.tx32p[i];

  3727. 

  3728.           adapt_prob(&p->tx8p[i], s->counts.tx8p[i][0], s->counts.tx8p[i][1], 20, 128);

  3729.           adapt_prob(&p->tx16p[i][0], c16[0], c16[1] + c16[2], 20, 128);

  3730.           adapt_prob(&p->tx16p[i][1], c16[1], c16[2], 20, 128);

  3731.           adapt_prob(&p->tx32p[i][0], c32[0], c32[1] + c32[2] + c32[3], 20, 128);

  3732.           adapt_prob(&p->tx32p[i][1], c32[1], c32[2] + c32[3], 20, 128);

  3733.           adapt_prob(&p->tx32p[i][2], c32[2], c32[3], 20, 128);

  3734.       }

  3735.     }

  3736. 

  3737.     // interpolation filter

  3738.     if (s->s.h.filtermode == FILTER_SWITCHABLE) {

  3739.         for (i = 0; i < 4; i++) {

  3740.             uint8_t *pp = p->filter[i];

  3741.             unsigned *c = s->counts.filter[i];

  3742. 

  3743.             adapt_prob(&pp[0], c[0], c[1] + c[2], 20, 128);

  3744.             adapt_prob(&pp[1], c[1], c[2], 20, 128);

  3745.         }

  3746.     }

  3747. 

  3748.     // inter modes

  3749.     for (i = 0; i < 7; i++) {

  3750.         uint8_t *pp = p->mv_mode[i];

  3751.         unsigned *c = s->counts.mv_mode[i];

  3752. 

  3753.         adapt_prob(&pp[0], c[2], c[1] + c[0] + c[3], 20, 128);

  3754.         adapt_prob(&pp[1], c[0], c[1] + c[3], 20, 128);

  3755.         adapt_prob(&pp[2], c[1], c[3], 20, 128);

  3756.     }

  3757. 

  3758.     // mv joints

  3759.     {

  3760.         uint8_t *pp = p->mv_joint;

  3761.         unsigned *c = s->counts.mv_joint;

  3762. 

  3763.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3764.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3765.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3766.     }

  3767. 

  3768.     // mv components

  3769.     for (i = 0; i < 2; i++) {

  3770.         uint8_t *pp;

  3771.         unsigned *c, (*c2)[2], sum;

  3772. 

  3773.         adapt_prob(&p->mv_comp[i].sign, s->counts.mv_comp[i].sign[0],

  3774.                    s->counts.mv_comp[i].sign[1], 20, 128);

  3775. 

  3776.         pp = p->mv_comp[i].classes;

  3777.         c = s->counts.mv_comp[i].classes;

  3778.         sum = c[1] + c[2] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9] + c[10];

  3779.         adapt_prob(&pp[0], c[0], sum, 20, 128);

  3780.         sum -= c[1];

  3781.         adapt_prob(&pp[1], c[1], sum, 20, 128);

  3782.         sum -= c[2] + c[3];

  3783.         adapt_prob(&pp[2], c[2] + c[3], sum, 20, 128);

  3784.         adapt_prob(&pp[3], c[2], c[3], 20, 128);

  3785.         sum -= c[4] + c[5];

  3786.         adapt_prob(&pp[4], c[4] + c[5], sum, 20, 128);

  3787.         adapt_prob(&pp[5], c[4], c[5], 20, 128);

  3788.         sum -= c[6];

  3789.         adapt_prob(&pp[6], c[6], sum, 20, 128);

  3790.         adapt_prob(&pp[7], c[7] + c[8], c[9] + c[10], 20, 128);

  3791.         adapt_prob(&pp[8], c[7], c[8], 20, 128);

  3792.         adapt_prob(&pp[9], c[9], c[10], 20, 128);

  3793. 

  3794.         adapt_prob(&p->mv_comp[i].class0, s->counts.mv_comp[i].class0[0],

  3795.                    s->counts.mv_comp[i].class0[1], 20, 128);

  3796.         pp = p->mv_comp[i].bits;

  3797.         c2 = s->counts.mv_comp[i].bits;

  3798.         for (j = 0; j < 10; j++)

  3799.             adapt_prob(&pp[j], c2[j][0], c2[j][1], 20, 128);

  3800. 

  3801.         for (j = 0; j < 2; j++) {

  3802.             pp = p->mv_comp[i].class0_fp[j];

  3803.             c = s->counts.mv_comp[i].class0_fp[j];

  3804.             adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3805.             adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3806.             adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3807.         }

  3808.         pp = p->mv_comp[i].fp;

  3809.         c = s->counts.mv_comp[i].fp;

  3810.         adapt_prob(&pp[0], c[0], c[1] + c[2] + c[3], 20, 128);

  3811.         adapt_prob(&pp[1], c[1], c[2] + c[3], 20, 128);

  3812.         adapt_prob(&pp[2], c[2], c[3], 20, 128);

  3813. 

  3814.         if (s->s.h.highprecisionmvs) {

  3815.             adapt_prob(&p->mv_comp[i].class0_hp, s->counts.mv_comp[i].class0_hp[0],

  3816.                        s->counts.mv_comp[i].class0_hp[1], 20, 128);

  3817.             adapt_prob(&p->mv_comp[i].hp, s->counts.mv_comp[i].hp[0],

  3818.                        s->counts.mv_comp[i].hp[1], 20, 128);

  3819.         }

  3820.     }

  3821. 

  3822.     // y intra modes

  3823.     for (i = 0; i < 4; i++) {

  3824.         uint8_t *pp = p->y_mode[i];

  3825.         unsigned *c = s->counts.y_mode[i], sum, s2;

  3826. 

  3827.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3828.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3829.         sum -= c[TM_VP8_PRED];

  3830.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3831.         sum -= c[VERT_PRED];

  3832.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3833.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3834.         sum -= s2;

  3835.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3836.         s2 -= c[HOR_PRED];

  3837.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3838.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3839.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3840.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3841.         sum -= c[VERT_LEFT_PRED];

  3842.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3843.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3844.     }

  3845. 

  3846.     // uv intra modes

  3847.     for (i = 0; i < 10; i++) {

  3848.         uint8_t *pp = p->uv_mode[i];

  3849.         unsigned *c = s->counts.uv_mode[i], sum, s2;

  3850. 

  3851.         sum = c[0] + c[1] + c[3] + c[4] + c[5] + c[6] + c[7] + c[8] + c[9];

  3852.         adapt_prob(&pp[0], c[DC_PRED], sum, 20, 128);

  3853.         sum -= c[TM_VP8_PRED];

  3854.         adapt_prob(&pp[1], c[TM_VP8_PRED], sum, 20, 128);

  3855.         sum -= c[VERT_PRED];

  3856.         adapt_prob(&pp[2], c[VERT_PRED], sum, 20, 128);

  3857.         s2 = c[HOR_PRED] + c[DIAG_DOWN_RIGHT_PRED] + c[VERT_RIGHT_PRED];

  3858.         sum -= s2;

  3859.         adapt_prob(&pp[3], s2, sum, 20, 128);

  3860.         s2 -= c[HOR_PRED];

  3861.         adapt_prob(&pp[4], c[HOR_PRED], s2, 20, 128);

  3862.         adapt_prob(&pp[5], c[DIAG_DOWN_RIGHT_PRED], c[VERT_RIGHT_PRED], 20, 128);

  3863.         sum -= c[DIAG_DOWN_LEFT_PRED];

  3864.         adapt_prob(&pp[6], c[DIAG_DOWN_LEFT_PRED], sum, 20, 128);

  3865.         sum -= c[VERT_LEFT_PRED];

  3866.         adapt_prob(&pp[7], c[VERT_LEFT_PRED], sum, 20, 128);

  3867.         adapt_prob(&pp[8], c[HOR_DOWN_PRED], c[HOR_UP_PRED], 20, 128);

  3868.     }

  3869. }

  3870. 

  3871. static void free_buffers(VP9Context *s)

  3872. {

  3873.     av_freep(&s->intra_pred_data[0]);

  3874.     av_freep(&s->b_base);

  3875.     av_freep(&s->block_base);

  3876. }

  3877. 

  3878. static av_cold int vp9_decode_free(AVCodecContext *ctx)

  3879. {

  3880.     VP9Context *s = ctx->priv_data;

  3881.     int i;

  3882. 

  3883.     for (i = 0; i < 3; i++) {

  3884.         if (s->s.frames[i].tf.f->buf[0])

  3885.             vp9_unref_frame(ctx, &s->s.frames[i]);

  3886.         av_frame_free(&s->s.frames[i].tf.f);

  3887.     }

  3888.     for (i = 0; i < 8; i++) {

  3889.         if (s->s.refs[i].f->buf[0])

  3890.             ff_thread_release_buffer(ctx, &s->s.refs[i]);

  3891.         av_frame_free(&s->s.refs[i].f);

  3892.         if (s->next_refs[i].f->buf[0])

  3893.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3894.         av_frame_free(&s->next_refs[i].f);

  3895.     }

  3896.     free_buffers(s);

  3897.     av_freep(&s->c_b);

  3898.     s->c_b_size = 0;

  3899. 

  3900.     return 0;

  3901. }

  3902. 

  3903. 

  3904. static int vp9_decode_frame(AVCodecContext *ctx, void *frame,

  3905.                             int *got_frame, AVPacket *pkt)

  3906. {

  3907.     const uint8_t *data = pkt->data;

  3908.     int size = pkt->size;

  3909.     VP9Context *s = ctx->priv_data;

  3910.     int res, tile_row, tile_col, i, ref, row, col;

  3911.     int retain_segmap_ref = s->s.frames[REF_FRAME_SEGMAP].segmentation_map &&

  3912.                             (!s->s.h.segmentation.enabled || !s->s.h.segmentation.update_map);

  3913.     ptrdiff_t yoff, uvoff, ls_y, ls_uv;

  3914.     AVFrame *f;

  3915.     int bytesperpixel;

  3916. 

  3917.     if ((res = decode_frame_header(ctx, data, size, &ref)) < 0) {

  3918.         return res;

  3919.     } else if (res == 0) {

  3920.         if (!s->s.refs[ref].f->buf[0]) {

  3921.             av_log(ctx, AV_LOG_ERROR, "Requested reference %d not available\n", ref);

  3922.             return AVERROR_INVALIDDATA;

  3923.         }

  3924.         if ((res = av_frame_ref(frame, s->s.refs[ref].f)) < 0)

  3925.             return res;

  3926.         ((AVFrame *)frame)->pkt_pts = pkt->pts;

  3927.         ((AVFrame *)frame)->pkt_dts = pkt->dts;

  3928.         for (i = 0; i < 8; i++) {

  3929.             if (s->next_refs[i].f->buf[0])

  3930.                 ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3931.             if (s->s.refs[i].f->buf[0] &&

  3932.                 (res = ff_thread_ref_frame(&s->next_refs[i], &s->s.refs[i])) < 0)

  3933.                 return res;

  3934.         }

  3935.         *got_frame = 1;

  3936.         return pkt->size;

  3937.     }

  3938.     data += res;

  3939.     size -= res;

  3940. 

  3941.     if (!retain_segmap_ref || s->s.h.keyframe || s->s.h.intraonly) {

  3942.         if (s->s.frames[REF_FRAME_SEGMAP].tf.f->buf[0])

  3943.             vp9_unref_frame(ctx, &s->s.frames[REF_FRAME_SEGMAP]);

  3944.         if (!s->s.h.keyframe && !s->s.h.intraonly && !s->s.h.errorres && s->s.frames[CUR_FRAME].tf.f->buf[0] &&

  3945.             (res = vp9_ref_frame(ctx, &s->s.frames[REF_FRAME_SEGMAP], &s->s.frames[CUR_FRAME])) < 0)

  3946.             return res;

  3947.     }

  3948.     if (s->s.frames[REF_FRAME_MVPAIR].tf.f->buf[0])

  3949.         vp9_unref_frame(ctx, &s->s.frames[REF_FRAME_MVPAIR]);

  3950.     if (!s->s.h.intraonly && !s->s.h.keyframe && !s->s.h.errorres && s->s.frames[CUR_FRAME].tf.f->buf[0] &&

  3951.         (res = vp9_ref_frame(ctx, &s->s.frames[REF_FRAME_MVPAIR], &s->s.frames[CUR_FRAME])) < 0)

  3952.         return res;

  3953.     if (s->s.frames[CUR_FRAME].tf.f->buf[0])

  3954.         vp9_unref_frame(ctx, &s->s.frames[CUR_FRAME]);

  3955.     if ((res = vp9_alloc_frame(ctx, &s->s.frames[CUR_FRAME])) < 0)

  3956.         return res;

  3957.     f = s->s.frames[CUR_FRAME].tf.f;

  3958.     f->key_frame = s->s.h.keyframe;

  3959.     f->pict_type = (s->s.h.keyframe || s->s.h.intraonly) ? AV_PICTURE_TYPE_I : AV_PICTURE_TYPE_P;

  3960.     ls_y = f->linesize[0];

  3961.     ls_uv =f->linesize[1];

  3962. 

  3963.     if (s->s.frames[REF_FRAME_SEGMAP].tf.f->buf[0] &&

  3964.         (s->s.frames[REF_FRAME_MVPAIR].tf.f->width  != s->s.frames[CUR_FRAME].tf.f->width ||

  3965.          s->s.frames[REF_FRAME_MVPAIR].tf.f->height != s->s.frames[CUR_FRAME].tf.f->height)) {

  3966.         vp9_unref_frame(ctx, &s->s.frames[REF_FRAME_SEGMAP]);

  3967.     }

  3968. 

  3969.     // ref frame setup

  3970.     for (i = 0; i < 8; i++) {

  3971.         if (s->next_refs[i].f->buf[0])

  3972.             ff_thread_release_buffer(ctx, &s->next_refs[i]);

  3973.         if (s->s.h.refreshrefmask & (1 << i)) {

  3974.             res = ff_thread_ref_frame(&s->next_refs[i], &s->s.frames[CUR_FRAME].tf);

  3975.         } else if (s->s.refs[i].f->buf[0]) {

  3976.             res = ff_thread_ref_frame(&s->next_refs[i], &s->s.refs[i]);

  3977.         }

  3978.         if (res < 0)

  3979.             return res;

  3980.     }

  3981. 

  3982.     // main tile decode loop

  3983.     bytesperpixel = s->bytesperpixel;

  3984.     memset(s->above_partition_ctx, 0, s->cols);

  3985.     memset(s->above_skip_ctx, 0, s->cols);

  3986.     if (s->s.h.keyframe || s->s.h.intraonly) {

  3987.         memset(s->above_mode_ctx, DC_PRED, s->cols * 2);

  3988.     } else {

  3989.         memset(s->above_mode_ctx, NEARESTMV, s->cols);

  3990.     }

  3991.     memset(s->above_y_nnz_ctx, 0, s->sb_cols * 16);

  3992.     memset(s->above_uv_nnz_ctx[0], 0, s->sb_cols * 16 >> s->ss_h);

  3993.     memset(s->above_uv_nnz_ctx[1], 0, s->sb_cols * 16 >> s->ss_h);

  3994.     memset(s->above_segpred_ctx, 0, s->cols);

  3995.     s->pass = s->s.frames[CUR_FRAME].uses_2pass =

  3996.         ctx->active_thread_type == FF_THREAD_FRAME && s->s.h.refreshctx && !s->s.h.parallelmode;

  3997.     if ((res = update_block_buffers(ctx)) < 0) {

  3998.         av_log(ctx, AV_LOG_ERROR,

  3999.                "Failed to allocate block buffers\n");

  4000.         return res;

  4001.     }

  4002.     if (s->s.h.refreshctx && s->s.h.parallelmode) {

  4003.         int j, k, l, m;

  4004. 

  4005.         for (i = 0; i < 4; i++) {

  4006.             for (j = 0; j < 2; j++)

  4007.                 for (k = 0; k < 2; k++)

  4008.                     for (l = 0; l < 6; l++)

  4009.                         for (m = 0; m < 6; m++)

  4010.                             memcpy(s->prob_ctx[s->s.h.framectxid].coef[i][j][k][l][m],

  4011.                                    s->prob.coef[i][j][k][l][m], 3);

  4012.             if (s->s.h.txfmmode == i)

  4013.                 break;

  4014.         }

  4015.         s->prob_ctx[s->s.h.framectxid].p = s->prob.p;

  4016.         ff_thread_finish_setup(ctx);

  4017.     } else if (!s->s.h.refreshctx) {

  4018.         ff_thread_finish_setup(ctx);

  4019.     }

  4020. 

  4021.     do {

  4022.         yoff = uvoff = 0;

  4023.         s->b = s->b_base;

  4024.         s->block = s->block_base;

  4025.         s->uvblock[0] = s->uvblock_base[0];

  4026.         s->uvblock[1] = s->uvblock_base[1];

  4027.         s->eob = s->eob_base;

  4028.         s->uveob[0] = s->uveob_base[0];

  4029.         s->uveob[1] = s->uveob_base[1];

  4030. 

  4031.         for (tile_row = 0; tile_row < s->s.h.tiling.tile_rows; tile_row++) {

  4032.             set_tile_offset(&s->tile_row_start, &s->tile_row_end,

  4033.                             tile_row, s->s.h.tiling.log2_tile_rows, s->sb_rows);

  4034.             if (s->pass != 2) {

  4035.                 for (tile_col = 0; tile_col < s->s.h.tiling.tile_cols; tile_col++) {

  4036.                     int64_t tile_size;

  4037. 

  4038.                     if (tile_col == s->s.h.tiling.tile_cols - 1 &&

  4039.                         tile_row == s->s.h.tiling.tile_rows - 1) {

  4040.                         tile_size = size;

  4041.                     } else {

  4042.                         tile_size = AV_RB32(data);

  4043.                         data += 4;

  4044.                         size -= 4;

  4045.                     }

  4046.                     if (tile_size > size) {

  4047.                         ff_thread_report_progress(&s->s.frames[CUR_FRAME].tf, INT_MAX, 0);

  4048.                         return AVERROR_INVALIDDATA;

  4049.                     }

  4050.                     ff_vp56_init_range_decoder(&s->c_b[tile_col], data, tile_size);

  4051.                     if (vp56_rac_get_prob_branchy(&s->c_b[tile_col], 128)) { // marker bit

  4052.                         ff_thread_report_progress(&s->s.frames[CUR_FRAME].tf, INT_MAX, 0);

  4053.                         return AVERROR_INVALIDDATA;

  4054.                     }

  4055.                     data += tile_size;

  4056.                     size -= tile_size;

  4057.                 }

  4058.             }

  4059. 

  4060.             for (row = s->tile_row_start; row < s->tile_row_end;

  4061.                  row += 8, yoff += ls_y * 64, uvoff += ls_uv * 64 >> s->ss_v) {

  4062.                 struct VP9Filter *lflvl_ptr = s->lflvl;

  4063.                 ptrdiff_t yoff2 = yoff, uvoff2 = uvoff;

  4064. 

  4065.                 for (tile_col = 0; tile_col < s->s.h.tiling.tile_cols; tile_col++) {

  4066.                     set_tile_offset(&s->tile_col_start, &s->tile_col_end,

  4067.                                     tile_col, s->s.h.tiling.log2_tile_cols, s->sb_cols);

  4068. 

  4069.                     if (s->pass != 2) {

  4070.                         memset(s->left_partition_ctx, 0, 8);

  4071.                         memset(s->left_skip_ctx, 0, 8);

  4072.                         if (s->s.h.keyframe || s->s.h.intraonly) {

  4073.                             memset(s->left_mode_ctx, DC_PRED, 16);

  4074.                         } else {

  4075.                             memset(s->left_mode_ctx, NEARESTMV, 8);

  4076.                         }

  4077.                         memset(s->left_y_nnz_ctx, 0, 16);

  4078.                         memset(s->left_uv_nnz_ctx, 0, 32);

  4079.                         memset(s->left_segpred_ctx, 0, 8);

  4080. 

  4081.                         memcpy(&s->c, &s->c_b[tile_col], sizeof(s->c));

  4082.                     }

  4083. 

  4084.                     for (col = s->tile_col_start;

  4085.                          col < s->tile_col_end;

  4086.                          col += 8, yoff2 += 64 * bytesperpixel,

  4087.                          uvoff2 += 64 * bytesperpixel >> s->ss_h, lflvl_ptr++) {

  4088.                         // FIXME integrate with lf code (i.e. zero after each

  4089.                         // use, similar to invtxfm coefficients, or similar)

  4090.                         if (s->pass != 1) {

  4091.                             memset(lflvl_ptr->mask, 0, sizeof(lflvl_ptr->mask));

  4092.                         }

  4093. 

  4094.                         if (s->pass == 2) {

  4095.                             decode_sb_mem(ctx, row, col, lflvl_ptr,

  4096.                                           yoff2, uvoff2, BL_64X64);

  4097.                         } else {

  4098.                             decode_sb(ctx, row, col, lflvl_ptr,

  4099.                                       yoff2, uvoff2, BL_64X64);

  4100.                         }

  4101.                     }

  4102.                     if (s->pass != 2) {

  4103.                         memcpy(&s->c_b[tile_col], &s->c, sizeof(s->c));

  4104.                     }

  4105.                 }

  4106. 

  4107.                 if (s->pass == 1) {

  4108.                     continue;

  4109.                 }

  4110. 

  4111.                 // backup pre-loopfilter reconstruction data for intra

  4112.                 // prediction of next row of sb64s

  4113.                 if (row + 8 < s->rows) {

  4114.                     memcpy(s->intra_pred_data[0],

  4115.                            f->data[0] + yoff + 63 * ls_y,

  4116.                            8 * s->cols * bytesperpixel);

  4117.                     memcpy(s->intra_pred_data[1],

  4118.                            f->data[1] + uvoff + ((64 >> s->ss_v) - 1) * ls_uv,

  4119.                            8 * s->cols * bytesperpixel >> s->ss_h);

  4120.                     memcpy(s->intra_pred_data[2],

  4121.                            f->data[2] + uvoff + ((64 >> s->ss_v) - 1) * ls_uv,

  4122.                            8 * s->cols * bytesperpixel >> s->ss_h);

  4123.                 }

  4124. 

  4125.                 // loopfilter one row

  4126.                 if (s->s.h.filter.level) {

  4127.                     yoff2 = yoff;

  4128.                     uvoff2 = uvoff;

  4129.                     lflvl_ptr = s->lflvl;

  4130.                     for (col = 0; col < s->cols;

  4131.                          col += 8, yoff2 += 64 * bytesperpixel,

  4132.                          uvoff2 += 64 * bytesperpixel >> s->ss_h, lflvl_ptr++) {

  4133.                         loopfilter_sb(ctx, lflvl_ptr, row, col, yoff2, uvoff2);

  4134.                     }

  4135.                 }

  4136. 

  4137.                 // FIXME maybe we can make this more finegrained by running the

  4138.                 // loopfilter per-block instead of after each sbrow

  4139.                 // In fact that would also make intra pred left preparation easier?

  4140.                 ff_thread_report_progress(&s->s.frames[CUR_FRAME].tf, row >> 3, 0);

  4141.             }

  4142.         }

  4143. 

  4144.         if (s->pass < 2 && s->s.h.refreshctx && !s->s.h.parallelmode) {

  4145.             adapt_probs(s);

  4146.             ff_thread_finish_setup(ctx);

  4147.         }

  4148.     } while (s->pass++ == 1);

  4149.     ff_thread_report_progress(&s->s.frames[CUR_FRAME].tf, INT_MAX, 0);

  4150. 

  4151.     // ref frame setup

  4152.     for (i = 0; i < 8; i++) {

  4153.         if (s->s.refs[i].f->buf[0])

  4154.             ff_thread_release_buffer(ctx, &s->s.refs[i]);

  4155.         if (s->next_refs[i].f->buf[0] &&

  4156.             (res = ff_thread_ref_frame(&s->s.refs[i], &s->next_refs[i])) < 0)

  4157.             return res;

  4158.     }

  4159. 

  4160.     if (!s->s.h.invisible) {

  4161.         if ((res = av_frame_ref(frame, s->s.frames[CUR_FRAME].tf.f)) < 0)

  4162.             return res;

  4163.         *got_frame = 1;

  4164.     }

  4165. 

  4166.     return pkt->size;

  4167. }

  4168. 

  4169. static void vp9_decode_flush(AVCodecContext *ctx)

  4170. {

  4171.     VP9Context *s = ctx->priv_data;

  4172.     int i;

  4173. 

  4174.     for (i = 0; i < 3; i++)

  4175.         vp9_unref_frame(ctx, &s->s.frames[i]);

  4176.     for (i = 0; i < 8; i++)

  4177.         ff_thread_release_buffer(ctx, &s->s.refs[i]);

  4178. }

  4179. 

  4180. static int init_frames(AVCodecContext *ctx)

  4181. {

  4182.     VP9Context *s = ctx->priv_data;

  4183.     int i;

  4184. 

  4185.     for (i = 0; i < 3; i++) {

  4186.         s->s.frames[i].tf.f = av_frame_alloc();

  4187.         if (!s->s.frames[i].tf.f) {

  4188.             vp9_decode_free(ctx);

  4189.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  4190.             return AVERROR(ENOMEM);

  4191.         }

  4192.     }

  4193.     for (i = 0; i < 8; i++) {

  4194.         s->s.refs[i].f = av_frame_alloc();

  4195.         s->next_refs[i].f = av_frame_alloc();

  4196.         if (!s->s.refs[i].f || !s->next_refs[i].f) {

  4197.             vp9_decode_free(ctx);

  4198.             av_log(ctx, AV_LOG_ERROR, "Failed to allocate frame buffer %d\n", i);

  4199.             return AVERROR(ENOMEM);

  4200.         }

  4201.     }

  4202. 

  4203.     return 0;

  4204. }

  4205. 

  4206. static av_cold int vp9_decode_init(AVCodecContext *ctx)

  4207. {

  4208.     VP9Context *s = ctx->priv_data;

  4209. 

  4210.     ctx->internal->allocate_progress = 1;

  4211.     s->last_bpp = 0;

  4212.     s->s.h.filter.sharpness = -1;

  4213. 

  4214.     return init_frames(ctx);

  4215. }

  4216. 

  4217. #if HAVE_THREADS

  4218. static av_cold int vp9_decode_init_thread_copy(AVCodecContext *avctx)

  4219. {

  4220.     return init_frames(avctx);

  4221. }

  4222. 

  4223. static int vp9_decode_update_thread_context(AVCodecContext *dst, const AVCodecContext *src)

  4224. {

  4225.     int i, res;

  4226.     VP9Context *s = dst->priv_data, *ssrc = src->priv_data;

  4227. 

  4228.     // detect size changes in other threads

  4229.     if (s->intra_pred_data[0] &&

  4230.         (!ssrc->intra_pred_data[0] || s->cols != ssrc->cols ||

  4231.          s->rows != ssrc->rows || s->bpp != ssrc->bpp || s->pix_fmt != ssrc->pix_fmt)) {

  4232.         free_buffers(s);

  4233.     }

  4234. 

  4235.     for (i = 0; i < 3; i++) {

  4236.         if (s->s.frames[i].tf.f->buf[0])

  4237.             vp9_unref_frame(dst, &s->s.frames[i]);

  4238.         if (ssrc->s.frames[i].tf.f->buf[0]) {

  4239.             if ((res = vp9_ref_frame(dst, &s->s.frames[i], &ssrc->s.frames[i])) < 0)

  4240.                 return res;

  4241.         }

  4242.     }

  4243.     for (i = 0; i < 8; i++) {

  4244.         if (s->s.refs[i].f->buf[0])

  4245.             ff_thread_release_buffer(dst, &s->s.refs[i]);

  4246.         if (ssrc->next_refs[i].f->buf[0]) {

  4247.             if ((res = ff_thread_ref_frame(&s->s.refs[i], &ssrc->next_refs[i])) < 0)

  4248.                 return res;

  4249.         }

  4250.     }

  4251. 

  4252.     s->s.h.invisible = ssrc->s.h.invisible;

  4253.     s->s.h.keyframe = ssrc->s.h.keyframe;

  4254.     s->s.h.intraonly = ssrc->s.h.intraonly;

  4255.     s->ss_v = ssrc->ss_v;

  4256.     s->ss_h = ssrc->ss_h;

  4257.     s->s.h.segmentation.enabled = ssrc->s.h.segmentation.enabled;

  4258.     s->s.h.segmentation.update_map = ssrc->s.h.segmentation.update_map;

  4259.     s->s.h.segmentation.absolute_vals = ssrc->s.h.segmentation.absolute_vals;

  4260.     s->bytesperpixel = ssrc->bytesperpixel;

  4261.     s->bpp = ssrc->bpp;

  4262.     s->bpp_index = ssrc->bpp_index;

  4263.     s->pix_fmt = ssrc->pix_fmt;

  4264.     memcpy(&s->prob_ctx, &ssrc->prob_ctx, sizeof(s->prob_ctx));

  4265.     memcpy(&s->s.h.lf_delta, &ssrc->s.h.lf_delta, sizeof(s->s.h.lf_delta));

  4266.     memcpy(&s->s.h.segmentation.feat, &ssrc->s.h.segmentation.feat,

  4267.            sizeof(s->s.h.segmentation.feat));

  4268. 

  4269.     return 0;

  4270. }

  4271. #endif

  4272. 

  4273. static const AVProfile profiles[] = {

  4274.     { FF_PROFILE_VP9_0, "Profile 0" },

  4275.     { FF_PROFILE_VP9_1, "Profile 1" },

  4276.     { FF_PROFILE_VP9_2, "Profile 2" },

  4277.     { FF_PROFILE_VP9_3, "Profile 3" },

  4278.     { FF_PROFILE_UNKNOWN },

  4279. };

  4280. 

  4281. AVCodec ff_vp9_decoder = {

  4282.     .name                  = "vp9",

  4283.     .long_name             = NULL_IF_CONFIG_SMALL("Google VP9"),

  4284.     .type                  = AVMEDIA_TYPE_VIDEO,

  4285.     .id                    = AV_CODEC_ID_VP9,

  4286.     .priv_data_size        = sizeof(VP9Context),

  4287.     .init                  = vp9_decode_init,

  4288.     .close                 = vp9_decode_free,

  4289.     .decode                = vp9_decode_frame,

  4290.     .capabilities          = AV_CODEC_CAP_DR1 | AV_CODEC_CAP_FRAME_THREADS,

  4291.     .flush                 = vp9_decode_flush,

  4292.     .init_thread_copy      = ONLY_IF_THREADS_ENABLED(vp9_decode_init_thread_copy),

  4293.     .update_thread_context = ONLY_IF_THREADS_ENABLED(vp9_decode_update_thread_context),

  4294.     .profiles              = NULL_IF_CONFIG_SMALL(profiles),

  4295. };