tor-browser

The Tor Browser
git clone https://git.dasho.dev/tor-browser.git
Log | Files | Refs | README | LICENSE

vp9recon.c (30342B)


      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 "libavutil/avassert.h"
     25 #include "libavutil/frame.h"
     26 #include "libavutil/mem_internal.h"
     27 
     28 #include "progressframe.h"
     29 #include "videodsp.h"
     30 #include "vp9data.h"
     31 #include "vp9dec.h"
     32 
     33 static av_always_inline int check_intra_mode(VP9TileData *td, int mode, uint8_t **a,
     34                                             uint8_t *dst_edge, ptrdiff_t stride_edge,
     35                                             uint8_t *dst_inner, ptrdiff_t stride_inner,
     36                                             uint8_t *l, int col, int x, int w,
     37                                             int row, int y, enum TxfmMode tx,
     38                                             int p, int ss_h, int ss_v, int bytesperpixel)
     39 {
     40    const VP9Context *s = td->s;
     41    int have_top = row > 0 || y > 0;
     42    int have_left = col > td->tile_col_start || x > 0;
     43    int have_right = x < w - 1;
     44    int bpp = s->s.h.bpp;
     45    static const uint8_t mode_conv[10][2 /* have_left */][2 /* have_top */] = {
     46        [VERT_PRED]            = { { DC_127_PRED,          VERT_PRED            },
     47                                   { DC_127_PRED,          VERT_PRED            } },
     48        [HOR_PRED]             = { { DC_129_PRED,          DC_129_PRED          },
     49                                   { HOR_PRED,             HOR_PRED             } },
     50        [DC_PRED]              = { { DC_128_PRED,          TOP_DC_PRED          },
     51                                   { LEFT_DC_PRED,         DC_PRED              } },
     52        [DIAG_DOWN_LEFT_PRED]  = { { DC_127_PRED,          DIAG_DOWN_LEFT_PRED  },
     53                                   { DC_127_PRED,          DIAG_DOWN_LEFT_PRED  } },
     54        [DIAG_DOWN_RIGHT_PRED] = { { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED },
     55                                   { DIAG_DOWN_RIGHT_PRED, DIAG_DOWN_RIGHT_PRED } },
     56        [VERT_RIGHT_PRED]      = { { VERT_RIGHT_PRED,      VERT_RIGHT_PRED      },
     57                                   { VERT_RIGHT_PRED,      VERT_RIGHT_PRED      } },
     58        [HOR_DOWN_PRED]        = { { HOR_DOWN_PRED,        HOR_DOWN_PRED        },
     59                                   { HOR_DOWN_PRED,        HOR_DOWN_PRED        } },
     60        [VERT_LEFT_PRED]       = { { DC_127_PRED,          VERT_LEFT_PRED       },
     61                                   { DC_127_PRED,          VERT_LEFT_PRED       } },
     62        [HOR_UP_PRED]          = { { DC_129_PRED,          DC_129_PRED          },
     63                                   { HOR_UP_PRED,          HOR_UP_PRED          } },
     64        [TM_VP8_PRED]          = { { DC_129_PRED,          VERT_PRED            },
     65                                   { HOR_PRED,             TM_VP8_PRED          } },
     66    };
     67    static const struct {
     68        uint8_t needs_left:1;
     69        uint8_t needs_top:1;
     70        uint8_t needs_topleft:1;
     71        uint8_t needs_topright:1;
     72        uint8_t invert_left:1;
     73    } edges[N_INTRA_PRED_MODES] = {
     74        [VERT_PRED]            = { .needs_top  = 1 },
     75        [HOR_PRED]             = { .needs_left = 1 },
     76        [DC_PRED]              = { .needs_top  = 1, .needs_left = 1 },
     77        [DIAG_DOWN_LEFT_PRED]  = { .needs_top  = 1, .needs_topright = 1 },
     78        [DIAG_DOWN_RIGHT_PRED] = { .needs_left = 1, .needs_top = 1,
     79                                   .needs_topleft = 1 },
     80        [VERT_RIGHT_PRED]      = { .needs_left = 1, .needs_top = 1,
     81                                   .needs_topleft = 1 },
     82        [HOR_DOWN_PRED]        = { .needs_left = 1, .needs_top = 1,
     83                                   .needs_topleft = 1 },
     84        [VERT_LEFT_PRED]       = { .needs_top  = 1, .needs_topright = 1 },
     85        [HOR_UP_PRED]          = { .needs_left = 1, .invert_left = 1 },
     86        [TM_VP8_PRED]          = { .needs_left = 1, .needs_top = 1,
     87                                   .needs_topleft = 1 },
     88        [LEFT_DC_PRED]         = { .needs_left = 1 },
     89        [TOP_DC_PRED]          = { .needs_top  = 1 },
     90        [DC_128_PRED]          = { 0 },
     91        [DC_127_PRED]          = { 0 },
     92        [DC_129_PRED]          = { 0 }
     93    };
     94 
     95    av_assert2(mode >= 0 && mode < 10);
     96    mode = mode_conv[mode][have_left][have_top];
     97    if (edges[mode].needs_top) {
     98        uint8_t *top, *topleft;
     99        int n_px_need = 4 << tx, n_px_have = (((s->cols - col) << !ss_h) - x) * 4;
    100        int n_px_need_tr = 0;
    101 
    102        if (tx == TX_4X4 && edges[mode].needs_topright && have_right)
    103            n_px_need_tr = 4;
    104 
    105        // if top of sb64-row, use s->intra_pred_data[] instead of
    106        // dst[-stride] for intra prediction (it contains pre- instead of
    107        // post-loopfilter data)
    108        if (have_top) {
    109            top = !(row & 7) && !y ?
    110                s->intra_pred_data[p] + (col * (8 >> ss_h) + x * 4) * bytesperpixel :
    111                y == 0 ? &dst_edge[-stride_edge] : &dst_inner[-stride_inner];
    112            if (have_left)
    113                topleft = !(row & 7) && !y ?
    114                    s->intra_pred_data[p] + (col * (8 >> ss_h) + x * 4) * bytesperpixel :
    115                    y == 0 || x == 0 ? &dst_edge[-stride_edge] :
    116                    &dst_inner[-stride_inner];
    117        }
    118 
    119        if (have_top &&
    120            (!edges[mode].needs_topleft || (have_left && top == topleft)) &&
    121            (tx != TX_4X4 || !edges[mode].needs_topright || have_right) &&
    122            n_px_need + n_px_need_tr <= n_px_have) {
    123            *a = top;
    124        } else {
    125            if (have_top) {
    126                if (n_px_need <= n_px_have) {
    127                    memcpy(*a, top, n_px_need * bytesperpixel);
    128                } else {
    129 #define memset_bpp(c, i1, v, i2, num) do { \
    130    if (bytesperpixel == 1) { \
    131        memset(&(c)[(i1)], (v)[(i2)], (num)); \
    132    } else { \
    133        int n, val = AV_RN16A(&(v)[(i2) * 2]); \
    134        for (n = 0; n < (num); n++) { \
    135            AV_WN16A(&(c)[((i1) + n) * 2], val); \
    136        } \
    137    } \
    138 } while (0)
    139                    memcpy(*a, top, n_px_have * bytesperpixel);
    140                    memset_bpp(*a, n_px_have, (*a), n_px_have - 1, n_px_need - n_px_have);
    141                }
    142            } else {
    143 #define memset_val(c, val, num) do { \
    144    if (bytesperpixel == 1) { \
    145        memset((c), (val), (num)); \
    146    } else { \
    147        int n; \
    148        for (n = 0; n < (num); n++) { \
    149            AV_WN16A(&(c)[n * 2], (val)); \
    150        } \
    151    } \
    152 } while (0)
    153                memset_val(*a, (128 << (bpp - 8)) - 1, n_px_need);
    154            }
    155            if (edges[mode].needs_topleft) {
    156                if (have_left && have_top) {
    157 #define assign_bpp(c, i1, v, i2) do { \
    158    if (bytesperpixel == 1) { \
    159        (c)[(i1)] = (v)[(i2)]; \
    160    } else { \
    161        AV_COPY16(&(c)[(i1) * 2], &(v)[(i2) * 2]); \
    162    } \
    163 } while (0)
    164                    assign_bpp(*a, -1, topleft, -1);
    165                } else {
    166 #define assign_val(c, i, v) do { \
    167    if (bytesperpixel == 1) { \
    168        (c)[(i)] = (v); \
    169    } else { \
    170        AV_WN16A(&(c)[(i) * 2], (v)); \
    171    } \
    172 } while (0)
    173                    assign_val((*a), -1, (128 << (bpp - 8)) + (have_top ? +1 : -1));
    174                }
    175            }
    176            if (tx == TX_4X4 && edges[mode].needs_topright) {
    177                if (have_top && have_right &&
    178                    n_px_need + n_px_need_tr <= n_px_have) {
    179                    memcpy(&(*a)[4 * bytesperpixel], &top[4 * bytesperpixel], 4 * bytesperpixel);
    180                } else {
    181                    memset_bpp(*a, 4, *a, 3, 4);
    182                }
    183            }
    184        }
    185    }
    186    if (edges[mode].needs_left) {
    187        if (have_left) {
    188            int n_px_need = 4 << tx, i, n_px_have = (((s->rows - row) << !ss_v) - y) * 4;
    189            uint8_t *dst = x == 0 ? dst_edge : dst_inner;
    190            ptrdiff_t stride = x == 0 ? stride_edge : stride_inner;
    191 
    192            if (edges[mode].invert_left) {
    193                if (n_px_need <= n_px_have) {
    194                    for (i = 0; i < n_px_need; i++)
    195                        assign_bpp(l, i, &dst[i * stride], -1);
    196                } else {
    197                    for (i = 0; i < n_px_have; i++)
    198                        assign_bpp(l, i, &dst[i * stride], -1);
    199                    memset_bpp(l, n_px_have, l, n_px_have - 1, n_px_need - n_px_have);
    200                }
    201            } else {
    202                if (n_px_need <= n_px_have) {
    203                    for (i = 0; i < n_px_need; i++)
    204                        assign_bpp(l, n_px_need - 1 - i, &dst[i * stride], -1);
    205                } else {
    206                    for (i = 0; i < n_px_have; i++)
    207                        assign_bpp(l, n_px_need - 1 - i, &dst[i * stride], -1);
    208                    memset_bpp(l, 0, l, n_px_need - n_px_have, n_px_need - n_px_have);
    209                }
    210            }
    211        } else {
    212            memset_val(l, (128 << (bpp - 8)) + 1, 4 << tx);
    213        }
    214    }
    215 
    216    return mode;
    217 }
    218 
    219 static av_always_inline void intra_recon(VP9TileData *td, ptrdiff_t y_off,
    220                                         ptrdiff_t uv_off, int bytesperpixel)
    221 {
    222    const VP9Context *s = td->s;
    223    VP9Block *b = td->b;
    224    int row = td->row, col = td->col;
    225    int w4 = ff_vp9_bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;
    226    int h4 = ff_vp9_bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);
    227    int end_x = FFMIN(2 * (s->cols - col), w4);
    228    int end_y = FFMIN(2 * (s->rows - row), h4);
    229    int tx = 4 * s->s.h.lossless + b->tx, uvtx = b->uvtx + 4 * s->s.h.lossless;
    230    int uvstep1d = 1 << b->uvtx, p;
    231    uint8_t *dst = td->dst[0], *dst_r = s->s.frames[CUR_FRAME].tf.f->data[0] + y_off;
    232    LOCAL_ALIGNED_32(uint8_t, a_buf, [96]);
    233    LOCAL_ALIGNED_32(uint8_t, l, [64]);
    234 
    235    for (n = 0, y = 0; y < end_y; y += step1d) {
    236        uint8_t *ptr = dst, *ptr_r = dst_r;
    237        for (x = 0; x < end_x; x += step1d, ptr += 4 * step1d * bytesperpixel,
    238                               ptr_r += 4 * step1d * bytesperpixel, n += step) {
    239            int mode = b->mode[b->bs > BS_8x8 && b->tx == TX_4X4 ?
    240                               y * 2 + x : 0];
    241            uint8_t *a = &a_buf[32];
    242            enum TxfmType txtp = ff_vp9_intra_txfm_type[mode];
    243            int eob = b->skip ? 0 : b->tx > TX_8X8 ? AV_RN16A(&td->eob[n]) : td->eob[n];
    244 
    245            mode = check_intra_mode(td, mode, &a, ptr_r,
    246                                    s->s.frames[CUR_FRAME].tf.f->linesize[0],
    247                                    ptr, td->y_stride, l,
    248                                    col, x, w4, row, y, b->tx, 0, 0, 0, bytesperpixel);
    249            s->dsp.intra_pred[b->tx][mode](ptr, td->y_stride, l, a);
    250            if (eob)
    251                s->dsp.itxfm_add[tx][txtp](ptr, td->y_stride,
    252                                           td->block + 16 * n * bytesperpixel, eob);
    253        }
    254        dst_r += 4 * step1d * s->s.frames[CUR_FRAME].tf.f->linesize[0];
    255        dst   += 4 * step1d * td->y_stride;
    256    }
    257 
    258    // U/V
    259    w4    >>= s->ss_h;
    260    end_x >>= s->ss_h;
    261    end_y >>= s->ss_v;
    262    step = 1 << (b->uvtx * 2);
    263    for (p = 0; p < 2; p++) {
    264        dst   = td->dst[1 + p];
    265        dst_r = s->s.frames[CUR_FRAME].tf.f->data[1 + p] + uv_off;
    266        for (n = 0, y = 0; y < end_y; y += uvstep1d) {
    267            uint8_t *ptr = dst, *ptr_r = dst_r;
    268            for (x = 0; x < end_x; x += uvstep1d, ptr += 4 * uvstep1d * bytesperpixel,
    269                                   ptr_r += 4 * uvstep1d * bytesperpixel, n += step) {
    270                int mode = b->uvmode;
    271                uint8_t *a = &a_buf[32];
    272                int eob = b->skip ? 0 : b->uvtx > TX_8X8 ? AV_RN16A(&td->uveob[p][n]) : td->uveob[p][n];
    273 
    274                mode = check_intra_mode(td, mode, &a, ptr_r,
    275                                        s->s.frames[CUR_FRAME].tf.f->linesize[1],
    276                                        ptr, td->uv_stride, l, col, x, w4, row, y,
    277                                        b->uvtx, p + 1, s->ss_h, s->ss_v, bytesperpixel);
    278                s->dsp.intra_pred[b->uvtx][mode](ptr, td->uv_stride, l, a);
    279                if (eob)
    280                    s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, td->uv_stride,
    281                                                    td->uvblock[p] + 16 * n * bytesperpixel, eob);
    282            }
    283            dst_r += 4 * uvstep1d * s->s.frames[CUR_FRAME].tf.f->linesize[1];
    284            dst   += 4 * uvstep1d * td->uv_stride;
    285        }
    286    }
    287 }
    288 
    289 void ff_vp9_intra_recon_8bpp(VP9TileData *td, ptrdiff_t y_off, ptrdiff_t uv_off)
    290 {
    291    intra_recon(td, y_off, uv_off, 1);
    292 }
    293 
    294 void ff_vp9_intra_recon_16bpp(VP9TileData *td, ptrdiff_t y_off, ptrdiff_t uv_off)
    295 {
    296    intra_recon(td, y_off, uv_off, 2);
    297 }
    298 
    299 static av_always_inline void mc_luma_unscaled(VP9TileData *td, const vp9_mc_func (*mc)[2],
    300                                              uint8_t *dst, ptrdiff_t dst_stride,
    301                                              const uint8_t *ref, ptrdiff_t ref_stride,
    302                                              const ProgressFrame *ref_frame,
    303                                              ptrdiff_t y, ptrdiff_t x, const VP9mv *mv,
    304                                              int bw, int bh, int w, int h, int bytesperpixel)
    305 {
    306    const VP9Context *s = td->s;
    307    int mx = mv->x, my = mv->y, th;
    308 
    309    y += my >> 3;
    310    x += mx >> 3;
    311    ref += y * ref_stride + x * bytesperpixel;
    312    mx &= 7;
    313    my &= 7;
    314    // FIXME bilinear filter only needs 0/1 pixels, not 3/4
    315    // we use +7 because the last 7 pixels of each sbrow can be changed in
    316    // the longest loopfilter of the next sbrow
    317    th = (y + bh + 4 * !!my + 7) >> 6;
    318    ff_progress_frame_await(ref_frame, FFMAX(th, 0));
    319    // The arm/aarch64 _hv filters read one more row than what actually is
    320    // needed, so switch to emulated edge one pixel sooner vertically
    321    // (!!my * 5) than horizontally (!!mx * 4).
    322    // The arm/aarch64 _h filters read one more pixel than what actually is
    323    // needed, so switch to emulated edge if that would read beyond the bottom
    324    // right block.
    325    if (x < !!mx * 3 || y < !!my * 3 ||
    326        ((ARCH_AARCH64 || ARCH_ARM) && (x + !!mx * 5 > w - bw) && (y + !!my * 5 + 1 > h - bh)) ||
    327        x + !!mx * 4 > w - bw || y + !!my * 5 > h - bh) {
    328        s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
    329                                 ref - !!my * 3 * ref_stride - !!mx * 3 * bytesperpixel,
    330                                 160, ref_stride,
    331                                 bw + !!mx * 7, bh + !!my * 7,
    332                                 x - !!mx * 3, y - !!my * 3, w, h);
    333        ref = td->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;
    334        ref_stride = 160;
    335    }
    336    mc[!!mx][!!my](dst, dst_stride, ref, ref_stride, bh, mx << 1, my << 1);
    337 }
    338 
    339 static av_always_inline void mc_chroma_unscaled(VP9TileData *td, const vp9_mc_func (*mc)[2],
    340                                                uint8_t *dst_u, uint8_t *dst_v,
    341                                                ptrdiff_t dst_stride,
    342                                                const uint8_t *ref_u, ptrdiff_t src_stride_u,
    343                                                const uint8_t *ref_v, ptrdiff_t src_stride_v,
    344                                                const ProgressFrame *ref_frame,
    345                                                ptrdiff_t y, ptrdiff_t x, const VP9mv *mv,
    346                                                int bw, int bh, int w, int h, int bytesperpixel)
    347 {
    348    const VP9Context *s = td->s;
    349    int mx = mv->x * (1 << !s->ss_h), my = mv->y * (1 << !s->ss_v), th;
    350 
    351    y += my >> 4;
    352    x += mx >> 4;
    353    ref_u += y * src_stride_u + x * bytesperpixel;
    354    ref_v += y * src_stride_v + x * bytesperpixel;
    355    mx &= 15;
    356    my &= 15;
    357    // FIXME bilinear filter only needs 0/1 pixels, not 3/4
    358    // we use +7 because the last 7 pixels of each sbrow can be changed in
    359    // the longest loopfilter of the next sbrow
    360    th = (y + bh + 4 * !!my + 7) >> (6 - s->ss_v);
    361    ff_progress_frame_await(ref_frame, FFMAX(th, 0));
    362    // The arm/aarch64 _hv filters read one more row than what actually is
    363    // needed, so switch to emulated edge one pixel sooner vertically
    364    // (!!my * 5) than horizontally (!!mx * 4).
    365    // The arm/aarch64 _h filters read one more pixel than what actually is
    366    // needed, so switch to emulated edge if that would read beyond the bottom
    367    // right block.
    368    if (x < !!mx * 3 || y < !!my * 3 ||
    369        ((ARCH_AARCH64 || ARCH_ARM) && (x + !!mx * 5 > w - bw) && (y + !!my * 5 + 1 > h - bh)) ||
    370        x + !!mx * 4 > w - bw || y + !!my * 5 > h - bh) {
    371        s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
    372                                 ref_u - !!my * 3 * src_stride_u - !!mx * 3 * bytesperpixel,
    373                                 160, src_stride_u,
    374                                 bw + !!mx * 7, bh + !!my * 7,
    375                                 x - !!mx * 3, y - !!my * 3, w, h);
    376        ref_u = td->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;
    377        mc[!!mx][!!my](dst_u, dst_stride, ref_u, 160, bh, mx, my);
    378 
    379        s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
    380                                 ref_v - !!my * 3 * src_stride_v - !!mx * 3 * bytesperpixel,
    381                                 160, src_stride_v,
    382                                 bw + !!mx * 7, bh + !!my * 7,
    383                                 x - !!mx * 3, y - !!my * 3, w, h);
    384        ref_v = td->edge_emu_buffer + !!my * 3 * 160 + !!mx * 3 * bytesperpixel;
    385        mc[!!mx][!!my](dst_v, dst_stride, ref_v, 160, bh, mx, my);
    386    } else {
    387        mc[!!mx][!!my](dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my);
    388        mc[!!mx][!!my](dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my);
    389    }
    390 }
    391 
    392 #define mc_luma_dir(td, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, \
    393                    px, py, pw, ph, bw, bh, w, h, i) \
    394    mc_luma_unscaled(td, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \
    395                     mv, bw, bh, w, h, bytesperpixel)
    396 #define mc_chroma_dir(td, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
    397                      row, col, mv, px, py, pw, ph, bw, bh, w, h, i) \
    398    mc_chroma_unscaled(td, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
    399                       row, col, mv, bw, bh, w, h, bytesperpixel)
    400 #define SCALED 0
    401 #define FN(x) x##_8bpp
    402 #define BYTES_PER_PIXEL 1
    403 #include "vp9_mc_template.c"
    404 #undef FN
    405 #undef BYTES_PER_PIXEL
    406 #define FN(x) x##_16bpp
    407 #define BYTES_PER_PIXEL 2
    408 #include "vp9_mc_template.c"
    409 #undef mc_luma_dir
    410 #undef mc_chroma_dir
    411 #undef FN
    412 #undef BYTES_PER_PIXEL
    413 #undef SCALED
    414 
    415 static av_always_inline void mc_luma_scaled(VP9TileData *td, vp9_scaled_mc_func smc,
    416                                            const vp9_mc_func (*mc)[2],
    417                                            uint8_t *dst, ptrdiff_t dst_stride,
    418                                            const uint8_t *ref, ptrdiff_t ref_stride,
    419                                            const ProgressFrame *ref_frame,
    420                                            ptrdiff_t y, ptrdiff_t x, const VP9mv *in_mv,
    421                                            int px, int py, int pw, int ph,
    422                                            int bw, int bh, int w, int h, int bytesperpixel,
    423                                            const uint16_t *scale, const uint8_t *step)
    424 {
    425    const VP9Context *s = td->s;
    426    if (s->s.frames[CUR_FRAME].tf.f->width == ref_frame->f->width &&
    427        s->s.frames[CUR_FRAME].tf.f->height == ref_frame->f->height) {
    428        mc_luma_unscaled(td, mc, dst, dst_stride, ref, ref_stride, ref_frame,
    429                         y, x, in_mv, bw, bh, w, h, bytesperpixel);
    430    } else {
    431 #define scale_mv(n, dim) (((int64_t)(n) * scale[dim]) >> 14)
    432    int mx, my;
    433    int refbw_m1, refbh_m1;
    434    int th;
    435    VP9mv mv;
    436 
    437    mv.x = av_clip(in_mv->x, -(x + pw - px + 4) * 8, (s->cols * 8 - x + px + 3) * 8);
    438    mv.y = av_clip(in_mv->y, -(y + ph - py + 4) * 8, (s->rows * 8 - y + py + 3) * 8);
    439    // BUG libvpx seems to scale the two components separately. This introduces
    440    // rounding errors but we have to reproduce them to be exactly compatible
    441    // with the output from libvpx...
    442    mx = scale_mv(mv.x * 2, 0) + scale_mv(x * 16, 0);
    443    my = scale_mv(mv.y * 2, 1) + scale_mv(y * 16, 1);
    444 
    445    y = my >> 4;
    446    x = mx >> 4;
    447    ref += y * ref_stride + x * bytesperpixel;
    448    mx &= 15;
    449    my &= 15;
    450    refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;
    451    refbh_m1 = ((bh - 1) * step[1] + my) >> 4;
    452    // FIXME bilinear filter only needs 0/1 pixels, not 3/4
    453    // we use +7 because the last 7 pixels of each sbrow can be changed in
    454    // the longest loopfilter of the next sbrow
    455    th = (y + refbh_m1 + 4 + 7) >> 6;
    456    ff_progress_frame_await(ref_frame, FFMAX(th, 0));
    457    // The arm/aarch64 _hv filters read one more row than what actually is
    458    // needed, so switch to emulated edge one pixel sooner vertically
    459    // (y + 5 >= h - refbh_m1) than horizontally (x + 4 >= w - refbw_m1).
    460    if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 5 >= h - refbh_m1) {
    461        s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
    462                                 ref - 3 * ref_stride - 3 * bytesperpixel,
    463                                 288, ref_stride,
    464                                 refbw_m1 + 8, refbh_m1 + 8,
    465                                 x - 3, y - 3, w, h);
    466        ref = td->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;
    467        ref_stride = 288;
    468    }
    469    smc(dst, dst_stride, ref, ref_stride, bh, mx, my, step[0], step[1]);
    470    }
    471 }
    472 
    473 static av_always_inline void mc_chroma_scaled(VP9TileData *td, vp9_scaled_mc_func smc,
    474                                              const vp9_mc_func (*mc)[2],
    475                                              uint8_t *dst_u, uint8_t *dst_v,
    476                                              ptrdiff_t dst_stride,
    477                                              const uint8_t *ref_u, ptrdiff_t src_stride_u,
    478                                              const uint8_t *ref_v, ptrdiff_t src_stride_v,
    479                                              const ProgressFrame *ref_frame,
    480                                              ptrdiff_t y, ptrdiff_t x, const VP9mv *in_mv,
    481                                              int px, int py, int pw, int ph,
    482                                              int bw, int bh, int w, int h, int bytesperpixel,
    483                                              const uint16_t *scale, const uint8_t *step)
    484 {
    485    const VP9Context *s = td->s;
    486    if (s->s.frames[CUR_FRAME].tf.f->width == ref_frame->f->width &&
    487        s->s.frames[CUR_FRAME].tf.f->height == ref_frame->f->height) {
    488        mc_chroma_unscaled(td, mc, dst_u, dst_v, dst_stride, ref_u, src_stride_u,
    489                           ref_v, src_stride_v, ref_frame,
    490                           y, x, in_mv, bw, bh, w, h, bytesperpixel);
    491    } else {
    492    int mx, my;
    493    int refbw_m1, refbh_m1;
    494    int th;
    495    VP9mv mv;
    496 
    497    if (s->ss_h) {
    498        // BUG https://code.google.com/p/webm/issues/detail?id=820
    499        mv.x = av_clip(in_mv->x, -(x + pw - px + 4) * 16, (s->cols * 4 - x + px + 3) * 16);
    500        mx = scale_mv(mv.x, 0) + (scale_mv(x * 16, 0) & ~15) + (scale_mv(x * 32, 0) & 15);
    501    } else {
    502        mv.x = av_clip(in_mv->x, -(x + pw - px + 4) * 8, (s->cols * 8 - x + px + 3) * 8);
    503        mx = scale_mv(mv.x * 2, 0) + scale_mv(x * 16, 0);
    504    }
    505    if (s->ss_v) {
    506        // BUG https://code.google.com/p/webm/issues/detail?id=820
    507        mv.y = av_clip(in_mv->y, -(y + ph - py + 4) * 16, (s->rows * 4 - y + py + 3) * 16);
    508        my = scale_mv(mv.y, 1) + (scale_mv(y * 16, 1) & ~15) + (scale_mv(y * 32, 1) & 15);
    509    } else {
    510        mv.y = av_clip(in_mv->y, -(y + ph - py + 4) * 8, (s->rows * 8 - y + py + 3) * 8);
    511        my = scale_mv(mv.y * 2, 1) + scale_mv(y * 16, 1);
    512    }
    513 #undef scale_mv
    514    y = my >> 4;
    515    x = mx >> 4;
    516    ref_u += y * src_stride_u + x * bytesperpixel;
    517    ref_v += y * src_stride_v + x * bytesperpixel;
    518    mx &= 15;
    519    my &= 15;
    520    refbw_m1 = ((bw - 1) * step[0] + mx) >> 4;
    521    refbh_m1 = ((bh - 1) * step[1] + my) >> 4;
    522    // FIXME bilinear filter only needs 0/1 pixels, not 3/4
    523    // we use +7 because the last 7 pixels of each sbrow can be changed in
    524    // the longest loopfilter of the next sbrow
    525    th = (y + refbh_m1 + 4 + 7) >> (6 - s->ss_v);
    526    ff_progress_frame_await(ref_frame, FFMAX(th, 0));
    527    // The arm/aarch64 _hv filters read one more row than what actually is
    528    // needed, so switch to emulated edge one pixel sooner vertically
    529    // (y + 5 >= h - refbh_m1) than horizontally (x + 4 >= w - refbw_m1).
    530    if (x < 3 || y < 3 || x + 4 >= w - refbw_m1 || y + 5 >= h - refbh_m1) {
    531        s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
    532                                 ref_u - 3 * src_stride_u - 3 * bytesperpixel,
    533                                 288, src_stride_u,
    534                                 refbw_m1 + 8, refbh_m1 + 8,
    535                                 x - 3, y - 3, w, h);
    536        ref_u = td->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;
    537        smc(dst_u, dst_stride, ref_u, 288, bh, mx, my, step[0], step[1]);
    538 
    539        s->vdsp.emulated_edge_mc(td->edge_emu_buffer,
    540                                 ref_v - 3 * src_stride_v - 3 * bytesperpixel,
    541                                 288, src_stride_v,
    542                                 refbw_m1 + 8, refbh_m1 + 8,
    543                                 x - 3, y - 3, w, h);
    544        ref_v = td->edge_emu_buffer + 3 * 288 + 3 * bytesperpixel;
    545        smc(dst_v, dst_stride, ref_v, 288, bh, mx, my, step[0], step[1]);
    546    } else {
    547        smc(dst_u, dst_stride, ref_u, src_stride_u, bh, mx, my, step[0], step[1]);
    548        smc(dst_v, dst_stride, ref_v, src_stride_v, bh, mx, my, step[0], step[1]);
    549    }
    550    }
    551 }
    552 
    553 #define mc_luma_dir(td, mc, dst, dst_ls, src, src_ls, tref, row, col, mv, \
    554                    px, py, pw, ph, bw, bh, w, h, i) \
    555    mc_luma_scaled(td, s->dsp.s##mc, s->dsp.mc, dst, dst_ls, src, src_ls, tref, row, col, \
    556                   mv, px, py, pw, ph, bw, bh, w, h, bytesperpixel, \
    557                   s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])
    558 #define mc_chroma_dir(td, mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
    559                      row, col, mv, px, py, pw, ph, bw, bh, w, h, i) \
    560    mc_chroma_scaled(td, s->dsp.s##mc, s->dsp.mc, dstu, dstv, dst_ls, srcu, srcu_ls, srcv, srcv_ls, tref, \
    561                     row, col, mv, px, py, pw, ph, bw, bh, w, h, bytesperpixel, \
    562                     s->mvscale[b->ref[i]], s->mvstep[b->ref[i]])
    563 #define SCALED 1
    564 #define FN(x) x##_scaled_8bpp
    565 #define BYTES_PER_PIXEL 1
    566 #include "vp9_mc_template.c"
    567 #undef FN
    568 #undef BYTES_PER_PIXEL
    569 #define FN(x) x##_scaled_16bpp
    570 #define BYTES_PER_PIXEL 2
    571 #include "vp9_mc_template.c"
    572 #undef mc_luma_dir
    573 #undef mc_chroma_dir
    574 #undef FN
    575 #undef BYTES_PER_PIXEL
    576 #undef SCALED
    577 
    578 static av_always_inline void inter_recon(VP9TileData *td, int bytesperpixel)
    579 {
    580    const VP9Context *s = td->s;
    581    VP9Block *b = td->b;
    582    int row = td->row, col = td->col;
    583 
    584    if (s->mvscale[b->ref[0]][0] == REF_INVALID_SCALE ||
    585        (b->comp && s->mvscale[b->ref[1]][0] == REF_INVALID_SCALE)) {
    586        if (!s->td->error_info) {
    587            s->td->error_info = AVERROR_INVALIDDATA;
    588            av_log(NULL, AV_LOG_ERROR, "Bitstream not supported, "
    589                                       "reference frame has invalid dimensions\n");
    590        }
    591        return;
    592    }
    593 
    594    if (s->mvscale[b->ref[0]][0] || (b->comp && s->mvscale[b->ref[1]][0])) {
    595        if (bytesperpixel == 1) {
    596            inter_pred_scaled_8bpp(td);
    597        } else {
    598            inter_pred_scaled_16bpp(td);
    599        }
    600    } else {
    601        if (bytesperpixel == 1) {
    602            inter_pred_8bpp(td);
    603        } else {
    604            inter_pred_16bpp(td);
    605        }
    606    }
    607 
    608    if (!b->skip) {
    609        /* mostly copied intra_recon() */
    610 
    611        int w4 = ff_vp9_bwh_tab[1][b->bs][0] << 1, step1d = 1 << b->tx, n;
    612        int h4 = ff_vp9_bwh_tab[1][b->bs][1] << 1, x, y, step = 1 << (b->tx * 2);
    613        int end_x = FFMIN(2 * (s->cols - col), w4);
    614        int end_y = FFMIN(2 * (s->rows - row), h4);
    615        int tx = 4 * s->s.h.lossless + b->tx, uvtx = b->uvtx + 4 * s->s.h.lossless;
    616        int uvstep1d = 1 << b->uvtx, p;
    617        uint8_t *dst = td->dst[0];
    618 
    619        // y itxfm add
    620        for (n = 0, y = 0; y < end_y; y += step1d) {
    621            uint8_t *ptr = dst;
    622            for (x = 0; x < end_x; x += step1d,
    623                 ptr += 4 * step1d * bytesperpixel, n += step) {
    624                int eob = b->tx > TX_8X8 ? AV_RN16A(&td->eob[n]) : td->eob[n];
    625 
    626                if (eob)
    627                    s->dsp.itxfm_add[tx][DCT_DCT](ptr, td->y_stride,
    628                                                  td->block + 16 * n * bytesperpixel, eob);
    629            }
    630            dst += 4 * td->y_stride * step1d;
    631        }
    632 
    633        // uv itxfm add
    634        end_x >>= s->ss_h;
    635        end_y >>= s->ss_v;
    636        step = 1 << (b->uvtx * 2);
    637        for (p = 0; p < 2; p++) {
    638            dst = td->dst[p + 1];
    639            for (n = 0, y = 0; y < end_y; y += uvstep1d) {
    640                uint8_t *ptr = dst;
    641                for (x = 0; x < end_x; x += uvstep1d,
    642                     ptr += 4 * uvstep1d * bytesperpixel, n += step) {
    643                    int eob = b->uvtx > TX_8X8 ? AV_RN16A(&td->uveob[p][n]) : td->uveob[p][n];
    644 
    645                    if (eob)
    646                        s->dsp.itxfm_add[uvtx][DCT_DCT](ptr, td->uv_stride,
    647                                                        td->uvblock[p] + 16 * n * bytesperpixel, eob);
    648                }
    649                dst += 4 * uvstep1d * td->uv_stride;
    650            }
    651        }
    652    }
    653 }
    654 
    655 void ff_vp9_inter_recon_8bpp(VP9TileData *td)
    656 {
    657    inter_recon(td, 1);
    658 }
    659 
    660 void ff_vp9_inter_recon_16bpp(VP9TileData *td)
    661 {
    662    inter_recon(td, 2);
    663 }