Home | History | Annotate | Download | only in common
      1 /*
      2  *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
      3  *
      4  *  Use of this source code is governed by a BSD-style license
      5  *  that can be found in the LICENSE file in the root of the source
      6  *  tree. An additional intellectual property rights grant can be found
      7  *  in the file PATENTS.  All contributing project authors may
      8  *  be found in the AUTHORS file in the root of the source tree.
      9  */
     10 
     11 #include "./vpx_config.h"
     12 #include "./vpx_dsp_rtcd.h"
     13 #include "vp9/common/vp9_loopfilter.h"
     14 #include "vp9/common/vp9_onyxc_int.h"
     15 #include "vp9/common/vp9_reconinter.h"
     16 #include "vpx_dsp/vpx_dsp_common.h"
     17 #include "vpx_mem/vpx_mem.h"
     18 #include "vpx_ports/mem.h"
     19 
     20 #include "vp9/common/vp9_seg_common.h"
     21 
     22 // 64 bit masks for left transform size. Each 1 represents a position where
     23 // we should apply a loop filter across the left border of an 8x8 block
     24 // boundary.
     25 //
     26 // In the case of TX_16X16->  ( in low order byte first we end up with
     27 // a mask that looks like this
     28 //
     29 //    10101010
     30 //    10101010
     31 //    10101010
     32 //    10101010
     33 //    10101010
     34 //    10101010
     35 //    10101010
     36 //    10101010
     37 //
     38 // A loopfilter should be applied to every other 8x8 horizontally.
     39 static const uint64_t left_64x64_txform_mask[TX_SIZES] = {
     40   0xffffffffffffffffULL,  // TX_4X4
     41   0xffffffffffffffffULL,  // TX_8x8
     42   0x5555555555555555ULL,  // TX_16x16
     43   0x1111111111111111ULL,  // TX_32x32
     44 };
     45 
     46 // 64 bit masks for above transform size. Each 1 represents a position where
     47 // we should apply a loop filter across the top border of an 8x8 block
     48 // boundary.
     49 //
     50 // In the case of TX_32x32 ->  ( in low order byte first we end up with
     51 // a mask that looks like this
     52 //
     53 //    11111111
     54 //    00000000
     55 //    00000000
     56 //    00000000
     57 //    11111111
     58 //    00000000
     59 //    00000000
     60 //    00000000
     61 //
     62 // A loopfilter should be applied to every other 4 the row vertically.
     63 static const uint64_t above_64x64_txform_mask[TX_SIZES] = {
     64   0xffffffffffffffffULL,  // TX_4X4
     65   0xffffffffffffffffULL,  // TX_8x8
     66   0x00ff00ff00ff00ffULL,  // TX_16x16
     67   0x000000ff000000ffULL,  // TX_32x32
     68 };
     69 
     70 // 64 bit masks for prediction sizes (left). Each 1 represents a position
     71 // where left border of an 8x8 block. These are aligned to the right most
     72 // appropriate bit, and then shifted into place.
     73 //
     74 // In the case of TX_16x32 ->  ( low order byte first ) we end up with
     75 // a mask that looks like this :
     76 //
     77 //  10000000
     78 //  10000000
     79 //  10000000
     80 //  10000000
     81 //  00000000
     82 //  00000000
     83 //  00000000
     84 //  00000000
     85 static const uint64_t left_prediction_mask[BLOCK_SIZES] = {
     86   0x0000000000000001ULL,  // BLOCK_4X4,
     87   0x0000000000000001ULL,  // BLOCK_4X8,
     88   0x0000000000000001ULL,  // BLOCK_8X4,
     89   0x0000000000000001ULL,  // BLOCK_8X8,
     90   0x0000000000000101ULL,  // BLOCK_8X16,
     91   0x0000000000000001ULL,  // BLOCK_16X8,
     92   0x0000000000000101ULL,  // BLOCK_16X16,
     93   0x0000000001010101ULL,  // BLOCK_16X32,
     94   0x0000000000000101ULL,  // BLOCK_32X16,
     95   0x0000000001010101ULL,  // BLOCK_32X32,
     96   0x0101010101010101ULL,  // BLOCK_32X64,
     97   0x0000000001010101ULL,  // BLOCK_64X32,
     98   0x0101010101010101ULL,  // BLOCK_64X64
     99 };
    100 
    101 // 64 bit mask to shift and set for each prediction size.
    102 static const uint64_t above_prediction_mask[BLOCK_SIZES] = {
    103   0x0000000000000001ULL,  // BLOCK_4X4
    104   0x0000000000000001ULL,  // BLOCK_4X8
    105   0x0000000000000001ULL,  // BLOCK_8X4
    106   0x0000000000000001ULL,  // BLOCK_8X8
    107   0x0000000000000001ULL,  // BLOCK_8X16,
    108   0x0000000000000003ULL,  // BLOCK_16X8
    109   0x0000000000000003ULL,  // BLOCK_16X16
    110   0x0000000000000003ULL,  // BLOCK_16X32,
    111   0x000000000000000fULL,  // BLOCK_32X16,
    112   0x000000000000000fULL,  // BLOCK_32X32,
    113   0x000000000000000fULL,  // BLOCK_32X64,
    114   0x00000000000000ffULL,  // BLOCK_64X32,
    115   0x00000000000000ffULL,  // BLOCK_64X64
    116 };
    117 // 64 bit mask to shift and set for each prediction size. A bit is set for
    118 // each 8x8 block that would be in the left most block of the given block
    119 // size in the 64x64 block.
    120 static const uint64_t size_mask[BLOCK_SIZES] = {
    121   0x0000000000000001ULL,  // BLOCK_4X4
    122   0x0000000000000001ULL,  // BLOCK_4X8
    123   0x0000000000000001ULL,  // BLOCK_8X4
    124   0x0000000000000001ULL,  // BLOCK_8X8
    125   0x0000000000000101ULL,  // BLOCK_8X16,
    126   0x0000000000000003ULL,  // BLOCK_16X8
    127   0x0000000000000303ULL,  // BLOCK_16X16
    128   0x0000000003030303ULL,  // BLOCK_16X32,
    129   0x0000000000000f0fULL,  // BLOCK_32X16,
    130   0x000000000f0f0f0fULL,  // BLOCK_32X32,
    131   0x0f0f0f0f0f0f0f0fULL,  // BLOCK_32X64,
    132   0x00000000ffffffffULL,  // BLOCK_64X32,
    133   0xffffffffffffffffULL,  // BLOCK_64X64
    134 };
    135 
    136 // These are used for masking the left and above borders.
    137 static const uint64_t left_border = 0x1111111111111111ULL;
    138 static const uint64_t above_border = 0x000000ff000000ffULL;
    139 
    140 // 16 bit masks for uv transform sizes.
    141 static const uint16_t left_64x64_txform_mask_uv[TX_SIZES] = {
    142   0xffff,  // TX_4X4
    143   0xffff,  // TX_8x8
    144   0x5555,  // TX_16x16
    145   0x1111,  // TX_32x32
    146 };
    147 
    148 static const uint16_t above_64x64_txform_mask_uv[TX_SIZES] = {
    149   0xffff,  // TX_4X4
    150   0xffff,  // TX_8x8
    151   0x0f0f,  // TX_16x16
    152   0x000f,  // TX_32x32
    153 };
    154 
    155 // 16 bit left mask to shift and set for each uv prediction size.
    156 static const uint16_t left_prediction_mask_uv[BLOCK_SIZES] = {
    157   0x0001,  // BLOCK_4X4,
    158   0x0001,  // BLOCK_4X8,
    159   0x0001,  // BLOCK_8X4,
    160   0x0001,  // BLOCK_8X8,
    161   0x0001,  // BLOCK_8X16,
    162   0x0001,  // BLOCK_16X8,
    163   0x0001,  // BLOCK_16X16,
    164   0x0011,  // BLOCK_16X32,
    165   0x0001,  // BLOCK_32X16,
    166   0x0011,  // BLOCK_32X32,
    167   0x1111,  // BLOCK_32X64
    168   0x0011,  // BLOCK_64X32,
    169   0x1111,  // BLOCK_64X64
    170 };
    171 // 16 bit above mask to shift and set for uv each prediction size.
    172 static const uint16_t above_prediction_mask_uv[BLOCK_SIZES] = {
    173   0x0001,  // BLOCK_4X4
    174   0x0001,  // BLOCK_4X8
    175   0x0001,  // BLOCK_8X4
    176   0x0001,  // BLOCK_8X8
    177   0x0001,  // BLOCK_8X16,
    178   0x0001,  // BLOCK_16X8
    179   0x0001,  // BLOCK_16X16
    180   0x0001,  // BLOCK_16X32,
    181   0x0003,  // BLOCK_32X16,
    182   0x0003,  // BLOCK_32X32,
    183   0x0003,  // BLOCK_32X64,
    184   0x000f,  // BLOCK_64X32,
    185   0x000f,  // BLOCK_64X64
    186 };
    187 
    188 // 64 bit mask to shift and set for each uv prediction size
    189 static const uint16_t size_mask_uv[BLOCK_SIZES] = {
    190   0x0001,  // BLOCK_4X4
    191   0x0001,  // BLOCK_4X8
    192   0x0001,  // BLOCK_8X4
    193   0x0001,  // BLOCK_8X8
    194   0x0001,  // BLOCK_8X16,
    195   0x0001,  // BLOCK_16X8
    196   0x0001,  // BLOCK_16X16
    197   0x0011,  // BLOCK_16X32,
    198   0x0003,  // BLOCK_32X16,
    199   0x0033,  // BLOCK_32X32,
    200   0x3333,  // BLOCK_32X64,
    201   0x00ff,  // BLOCK_64X32,
    202   0xffff,  // BLOCK_64X64
    203 };
    204 static const uint16_t left_border_uv = 0x1111;
    205 static const uint16_t above_border_uv = 0x000f;
    206 
    207 static const int mode_lf_lut[MB_MODE_COUNT] = {
    208   0, 0, 0, 0, 0, 0, 0, 0, 0, 0,  // INTRA_MODES
    209   1, 1, 0, 1                     // INTER_MODES (ZEROMV == 0)
    210 };
    211 
    212 static void update_sharpness(loop_filter_info_n *lfi, int sharpness_lvl) {
    213   int lvl;
    214 
    215   // For each possible value for the loop filter fill out limits
    216   for (lvl = 0; lvl <= MAX_LOOP_FILTER; lvl++) {
    217     // Set loop filter parameters that control sharpness.
    218     int block_inside_limit = lvl >> ((sharpness_lvl > 0) + (sharpness_lvl > 4));
    219 
    220     if (sharpness_lvl > 0) {
    221       if (block_inside_limit > (9 - sharpness_lvl))
    222         block_inside_limit = (9 - sharpness_lvl);
    223     }
    224 
    225     if (block_inside_limit < 1) block_inside_limit = 1;
    226 
    227     memset(lfi->lfthr[lvl].lim, block_inside_limit, SIMD_WIDTH);
    228     memset(lfi->lfthr[lvl].mblim, (2 * (lvl + 2) + block_inside_limit),
    229            SIMD_WIDTH);
    230   }
    231 }
    232 
    233 static uint8_t get_filter_level(const loop_filter_info_n *lfi_n,
    234                                 const MODE_INFO *mi) {
    235   return lfi_n->lvl[mi->segment_id][mi->ref_frame[0]][mode_lf_lut[mi->mode]];
    236 }
    237 
    238 void vp9_loop_filter_init(VP9_COMMON *cm) {
    239   loop_filter_info_n *lfi = &cm->lf_info;
    240   struct loopfilter *lf = &cm->lf;
    241   int lvl;
    242 
    243   // init limits for given sharpness
    244   update_sharpness(lfi, lf->sharpness_level);
    245   lf->last_sharpness_level = lf->sharpness_level;
    246 
    247   // init hev threshold const vectors
    248   for (lvl = 0; lvl <= MAX_LOOP_FILTER; lvl++)
    249     memset(lfi->lfthr[lvl].hev_thr, (lvl >> 4), SIMD_WIDTH);
    250 }
    251 
    252 void vp9_loop_filter_frame_init(VP9_COMMON *cm, int default_filt_lvl) {
    253   int seg_id;
    254   // n_shift is the multiplier for lf_deltas
    255   // the multiplier is 1 for when filter_lvl is between 0 and 31;
    256   // 2 when filter_lvl is between 32 and 63
    257   const int scale = 1 << (default_filt_lvl >> 5);
    258   loop_filter_info_n *const lfi = &cm->lf_info;
    259   struct loopfilter *const lf = &cm->lf;
    260   const struct segmentation *const seg = &cm->seg;
    261 
    262   // update limits if sharpness has changed
    263   if (lf->last_sharpness_level != lf->sharpness_level) {
    264     update_sharpness(lfi, lf->sharpness_level);
    265     lf->last_sharpness_level = lf->sharpness_level;
    266   }
    267 
    268   for (seg_id = 0; seg_id < MAX_SEGMENTS; seg_id++) {
    269     int lvl_seg = default_filt_lvl;
    270     if (segfeature_active(seg, seg_id, SEG_LVL_ALT_LF)) {
    271       const int data = get_segdata(seg, seg_id, SEG_LVL_ALT_LF);
    272       lvl_seg = clamp(
    273           seg->abs_delta == SEGMENT_ABSDATA ? data : default_filt_lvl + data, 0,
    274           MAX_LOOP_FILTER);
    275     }
    276 
    277     if (!lf->mode_ref_delta_enabled) {
    278       // we could get rid of this if we assume that deltas are set to
    279       // zero when not in use; encoder always uses deltas
    280       memset(lfi->lvl[seg_id], lvl_seg, sizeof(lfi->lvl[seg_id]));
    281     } else {
    282       int ref, mode;
    283       const int intra_lvl = lvl_seg + lf->ref_deltas[INTRA_FRAME] * scale;
    284       lfi->lvl[seg_id][INTRA_FRAME][0] = clamp(intra_lvl, 0, MAX_LOOP_FILTER);
    285 
    286       for (ref = LAST_FRAME; ref < MAX_REF_FRAMES; ++ref) {
    287         for (mode = 0; mode < MAX_MODE_LF_DELTAS; ++mode) {
    288           const int inter_lvl = lvl_seg + lf->ref_deltas[ref] * scale +
    289                                 lf->mode_deltas[mode] * scale;
    290           lfi->lvl[seg_id][ref][mode] = clamp(inter_lvl, 0, MAX_LOOP_FILTER);
    291         }
    292       }
    293     }
    294   }
    295 }
    296 
    297 static void filter_selectively_vert_row2(
    298     int subsampling_factor, uint8_t *s, int pitch, unsigned int mask_16x16,
    299     unsigned int mask_8x8, unsigned int mask_4x4, unsigned int mask_4x4_int,
    300     const loop_filter_thresh *lfthr, const uint8_t *lfl) {
    301   const int dual_mask_cutoff = subsampling_factor ? 0xff : 0xffff;
    302   const int lfl_forward = subsampling_factor ? 4 : 8;
    303   const unsigned int dual_one = 1 | (1 << lfl_forward);
    304   unsigned int mask;
    305   uint8_t *ss[2];
    306   ss[0] = s;
    307 
    308   for (mask =
    309            (mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int) & dual_mask_cutoff;
    310        mask; mask = (mask & ~dual_one) >> 1) {
    311     if (mask & dual_one) {
    312       const loop_filter_thresh *lfis[2];
    313       lfis[0] = lfthr + *lfl;
    314       lfis[1] = lfthr + *(lfl + lfl_forward);
    315       ss[1] = ss[0] + 8 * pitch;
    316 
    317       if (mask_16x16 & dual_one) {
    318         if ((mask_16x16 & dual_one) == dual_one) {
    319           vpx_lpf_vertical_16_dual(ss[0], pitch, lfis[0]->mblim, lfis[0]->lim,
    320                                    lfis[0]->hev_thr);
    321         } else {
    322           const loop_filter_thresh *lfi = lfis[!(mask_16x16 & 1)];
    323           vpx_lpf_vertical_16(ss[!(mask_16x16 & 1)], pitch, lfi->mblim,
    324                               lfi->lim, lfi->hev_thr);
    325         }
    326       }
    327 
    328       if (mask_8x8 & dual_one) {
    329         if ((mask_8x8 & dual_one) == dual_one) {
    330           vpx_lpf_vertical_8_dual(ss[0], pitch, lfis[0]->mblim, lfis[0]->lim,
    331                                   lfis[0]->hev_thr, lfis[1]->mblim,
    332                                   lfis[1]->lim, lfis[1]->hev_thr);
    333         } else {
    334           const loop_filter_thresh *lfi = lfis[!(mask_8x8 & 1)];
    335           vpx_lpf_vertical_8(ss[!(mask_8x8 & 1)], pitch, lfi->mblim, lfi->lim,
    336                              lfi->hev_thr);
    337         }
    338       }
    339 
    340       if (mask_4x4 & dual_one) {
    341         if ((mask_4x4 & dual_one) == dual_one) {
    342           vpx_lpf_vertical_4_dual(ss[0], pitch, lfis[0]->mblim, lfis[0]->lim,
    343                                   lfis[0]->hev_thr, lfis[1]->mblim,
    344                                   lfis[1]->lim, lfis[1]->hev_thr);
    345         } else {
    346           const loop_filter_thresh *lfi = lfis[!(mask_4x4 & 1)];
    347           vpx_lpf_vertical_4(ss[!(mask_4x4 & 1)], pitch, lfi->mblim, lfi->lim,
    348                              lfi->hev_thr);
    349         }
    350       }
    351 
    352       if (mask_4x4_int & dual_one) {
    353         if ((mask_4x4_int & dual_one) == dual_one) {
    354           vpx_lpf_vertical_4_dual(
    355               ss[0] + 4, pitch, lfis[0]->mblim, lfis[0]->lim, lfis[0]->hev_thr,
    356               lfis[1]->mblim, lfis[1]->lim, lfis[1]->hev_thr);
    357         } else {
    358           const loop_filter_thresh *lfi = lfis[!(mask_4x4_int & 1)];
    359           vpx_lpf_vertical_4(ss[!(mask_4x4_int & 1)] + 4, pitch, lfi->mblim,
    360                              lfi->lim, lfi->hev_thr);
    361         }
    362       }
    363     }
    364 
    365     ss[0] += 8;
    366     lfl += 1;
    367     mask_16x16 >>= 1;
    368     mask_8x8 >>= 1;
    369     mask_4x4 >>= 1;
    370     mask_4x4_int >>= 1;
    371   }
    372 }
    373 
    374 #if CONFIG_VP9_HIGHBITDEPTH
    375 static void highbd_filter_selectively_vert_row2(
    376     int subsampling_factor, uint16_t *s, int pitch, unsigned int mask_16x16,
    377     unsigned int mask_8x8, unsigned int mask_4x4, unsigned int mask_4x4_int,
    378     const loop_filter_thresh *lfthr, const uint8_t *lfl, int bd) {
    379   const int dual_mask_cutoff = subsampling_factor ? 0xff : 0xffff;
    380   const int lfl_forward = subsampling_factor ? 4 : 8;
    381   const unsigned int dual_one = 1 | (1 << lfl_forward);
    382   unsigned int mask;
    383   uint16_t *ss[2];
    384   ss[0] = s;
    385 
    386   for (mask =
    387            (mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int) & dual_mask_cutoff;
    388        mask; mask = (mask & ~dual_one) >> 1) {
    389     if (mask & dual_one) {
    390       const loop_filter_thresh *lfis[2];
    391       lfis[0] = lfthr + *lfl;
    392       lfis[1] = lfthr + *(lfl + lfl_forward);
    393       ss[1] = ss[0] + 8 * pitch;
    394 
    395       if (mask_16x16 & dual_one) {
    396         if ((mask_16x16 & dual_one) == dual_one) {
    397           vpx_highbd_lpf_vertical_16_dual(ss[0], pitch, lfis[0]->mblim,
    398                                           lfis[0]->lim, lfis[0]->hev_thr, bd);
    399         } else {
    400           const loop_filter_thresh *lfi = lfis[!(mask_16x16 & 1)];
    401           vpx_highbd_lpf_vertical_16(ss[!(mask_16x16 & 1)], pitch, lfi->mblim,
    402                                      lfi->lim, lfi->hev_thr, bd);
    403         }
    404       }
    405 
    406       if (mask_8x8 & dual_one) {
    407         if ((mask_8x8 & dual_one) == dual_one) {
    408           vpx_highbd_lpf_vertical_8_dual(
    409               ss[0], pitch, lfis[0]->mblim, lfis[0]->lim, lfis[0]->hev_thr,
    410               lfis[1]->mblim, lfis[1]->lim, lfis[1]->hev_thr, bd);
    411         } else {
    412           const loop_filter_thresh *lfi = lfis[!(mask_8x8 & 1)];
    413           vpx_highbd_lpf_vertical_8(ss[!(mask_8x8 & 1)], pitch, lfi->mblim,
    414                                     lfi->lim, lfi->hev_thr, bd);
    415         }
    416       }
    417 
    418       if (mask_4x4 & dual_one) {
    419         if ((mask_4x4 & dual_one) == dual_one) {
    420           vpx_highbd_lpf_vertical_4_dual(
    421               ss[0], pitch, lfis[0]->mblim, lfis[0]->lim, lfis[0]->hev_thr,
    422               lfis[1]->mblim, lfis[1]->lim, lfis[1]->hev_thr, bd);
    423         } else {
    424           const loop_filter_thresh *lfi = lfis[!(mask_4x4 & 1)];
    425           vpx_highbd_lpf_vertical_4(ss[!(mask_4x4 & 1)], pitch, lfi->mblim,
    426                                     lfi->lim, lfi->hev_thr, bd);
    427         }
    428       }
    429 
    430       if (mask_4x4_int & dual_one) {
    431         if ((mask_4x4_int & dual_one) == dual_one) {
    432           vpx_highbd_lpf_vertical_4_dual(
    433               ss[0] + 4, pitch, lfis[0]->mblim, lfis[0]->lim, lfis[0]->hev_thr,
    434               lfis[1]->mblim, lfis[1]->lim, lfis[1]->hev_thr, bd);
    435         } else {
    436           const loop_filter_thresh *lfi = lfis[!(mask_4x4_int & 1)];
    437           vpx_highbd_lpf_vertical_4(ss[!(mask_4x4_int & 1)] + 4, pitch,
    438                                     lfi->mblim, lfi->lim, lfi->hev_thr, bd);
    439         }
    440       }
    441     }
    442 
    443     ss[0] += 8;
    444     lfl += 1;
    445     mask_16x16 >>= 1;
    446     mask_8x8 >>= 1;
    447     mask_4x4 >>= 1;
    448     mask_4x4_int >>= 1;
    449   }
    450 }
    451 #endif  // CONFIG_VP9_HIGHBITDEPTH
    452 
    453 static void filter_selectively_horiz(
    454     uint8_t *s, int pitch, unsigned int mask_16x16, unsigned int mask_8x8,
    455     unsigned int mask_4x4, unsigned int mask_4x4_int,
    456     const loop_filter_thresh *lfthr, const uint8_t *lfl) {
    457   unsigned int mask;
    458   int count;
    459 
    460   for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int; mask;
    461        mask >>= count) {
    462     count = 1;
    463     if (mask & 1) {
    464       const loop_filter_thresh *lfi = lfthr + *lfl;
    465 
    466       if (mask_16x16 & 1) {
    467         if ((mask_16x16 & 3) == 3) {
    468           vpx_lpf_horizontal_16_dual(s, pitch, lfi->mblim, lfi->lim,
    469                                      lfi->hev_thr);
    470           count = 2;
    471         } else {
    472           vpx_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
    473         }
    474       } else if (mask_8x8 & 1) {
    475         if ((mask_8x8 & 3) == 3) {
    476           // Next block's thresholds.
    477           const loop_filter_thresh *lfin = lfthr + *(lfl + 1);
    478 
    479           vpx_lpf_horizontal_8_dual(s, pitch, lfi->mblim, lfi->lim,
    480                                     lfi->hev_thr, lfin->mblim, lfin->lim,
    481                                     lfin->hev_thr);
    482 
    483           if ((mask_4x4_int & 3) == 3) {
    484             vpx_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
    485                                       lfi->lim, lfi->hev_thr, lfin->mblim,
    486                                       lfin->lim, lfin->hev_thr);
    487           } else {
    488             if (mask_4x4_int & 1)
    489               vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
    490                                    lfi->hev_thr);
    491             else if (mask_4x4_int & 2)
    492               vpx_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
    493                                    lfin->lim, lfin->hev_thr);
    494           }
    495           count = 2;
    496         } else {
    497           vpx_lpf_horizontal_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
    498 
    499           if (mask_4x4_int & 1)
    500             vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
    501                                  lfi->hev_thr);
    502         }
    503       } else if (mask_4x4 & 1) {
    504         if ((mask_4x4 & 3) == 3) {
    505           // Next block's thresholds.
    506           const loop_filter_thresh *lfin = lfthr + *(lfl + 1);
    507 
    508           vpx_lpf_horizontal_4_dual(s, pitch, lfi->mblim, lfi->lim,
    509                                     lfi->hev_thr, lfin->mblim, lfin->lim,
    510                                     lfin->hev_thr);
    511           if ((mask_4x4_int & 3) == 3) {
    512             vpx_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
    513                                       lfi->lim, lfi->hev_thr, lfin->mblim,
    514                                       lfin->lim, lfin->hev_thr);
    515           } else {
    516             if (mask_4x4_int & 1)
    517               vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
    518                                    lfi->hev_thr);
    519             else if (mask_4x4_int & 2)
    520               vpx_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
    521                                    lfin->lim, lfin->hev_thr);
    522           }
    523           count = 2;
    524         } else {
    525           vpx_lpf_horizontal_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
    526 
    527           if (mask_4x4_int & 1)
    528             vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
    529                                  lfi->hev_thr);
    530         }
    531       } else {
    532         vpx_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
    533                              lfi->hev_thr);
    534       }
    535     }
    536     s += 8 * count;
    537     lfl += count;
    538     mask_16x16 >>= count;
    539     mask_8x8 >>= count;
    540     mask_4x4 >>= count;
    541     mask_4x4_int >>= count;
    542   }
    543 }
    544 
    545 #if CONFIG_VP9_HIGHBITDEPTH
    546 static void highbd_filter_selectively_horiz(
    547     uint16_t *s, int pitch, unsigned int mask_16x16, unsigned int mask_8x8,
    548     unsigned int mask_4x4, unsigned int mask_4x4_int,
    549     const loop_filter_thresh *lfthr, const uint8_t *lfl, int bd) {
    550   unsigned int mask;
    551   int count;
    552 
    553   for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int; mask;
    554        mask >>= count) {
    555     count = 1;
    556     if (mask & 1) {
    557       const loop_filter_thresh *lfi = lfthr + *lfl;
    558 
    559       if (mask_16x16 & 1) {
    560         if ((mask_16x16 & 3) == 3) {
    561           vpx_highbd_lpf_horizontal_16_dual(s, pitch, lfi->mblim, lfi->lim,
    562                                             lfi->hev_thr, bd);
    563           count = 2;
    564         } else {
    565           vpx_highbd_lpf_horizontal_16(s, pitch, lfi->mblim, lfi->lim,
    566                                        lfi->hev_thr, bd);
    567         }
    568       } else if (mask_8x8 & 1) {
    569         if ((mask_8x8 & 3) == 3) {
    570           // Next block's thresholds.
    571           const loop_filter_thresh *lfin = lfthr + *(lfl + 1);
    572 
    573           vpx_highbd_lpf_horizontal_8_dual(s, pitch, lfi->mblim, lfi->lim,
    574                                            lfi->hev_thr, lfin->mblim, lfin->lim,
    575                                            lfin->hev_thr, bd);
    576 
    577           if ((mask_4x4_int & 3) == 3) {
    578             vpx_highbd_lpf_horizontal_4_dual(
    579                 s + 4 * pitch, pitch, lfi->mblim, lfi->lim, lfi->hev_thr,
    580                 lfin->mblim, lfin->lim, lfin->hev_thr, bd);
    581           } else {
    582             if (mask_4x4_int & 1) {
    583               vpx_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
    584                                           lfi->lim, lfi->hev_thr, bd);
    585             } else if (mask_4x4_int & 2) {
    586               vpx_highbd_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
    587                                           lfin->lim, lfin->hev_thr, bd);
    588             }
    589           }
    590           count = 2;
    591         } else {
    592           vpx_highbd_lpf_horizontal_8(s, pitch, lfi->mblim, lfi->lim,
    593                                       lfi->hev_thr, bd);
    594 
    595           if (mask_4x4_int & 1) {
    596             vpx_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
    597                                         lfi->lim, lfi->hev_thr, bd);
    598           }
    599         }
    600       } else if (mask_4x4 & 1) {
    601         if ((mask_4x4 & 3) == 3) {
    602           // Next block's thresholds.
    603           const loop_filter_thresh *lfin = lfthr + *(lfl + 1);
    604 
    605           vpx_highbd_lpf_horizontal_4_dual(s, pitch, lfi->mblim, lfi->lim,
    606                                            lfi->hev_thr, lfin->mblim, lfin->lim,
    607                                            lfin->hev_thr, bd);
    608           if ((mask_4x4_int & 3) == 3) {
    609             vpx_highbd_lpf_horizontal_4_dual(
    610                 s + 4 * pitch, pitch, lfi->mblim, lfi->lim, lfi->hev_thr,
    611                 lfin->mblim, lfin->lim, lfin->hev_thr, bd);
    612           } else {
    613             if (mask_4x4_int & 1) {
    614               vpx_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
    615                                           lfi->lim, lfi->hev_thr, bd);
    616             } else if (mask_4x4_int & 2) {
    617               vpx_highbd_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
    618                                           lfin->lim, lfin->hev_thr, bd);
    619             }
    620           }
    621           count = 2;
    622         } else {
    623           vpx_highbd_lpf_horizontal_4(s, pitch, lfi->mblim, lfi->lim,
    624                                       lfi->hev_thr, bd);
    625 
    626           if (mask_4x4_int & 1) {
    627             vpx_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
    628                                         lfi->lim, lfi->hev_thr, bd);
    629           }
    630         }
    631       } else {
    632         vpx_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
    633                                     lfi->hev_thr, bd);
    634       }
    635     }
    636     s += 8 * count;
    637     lfl += count;
    638     mask_16x16 >>= count;
    639     mask_8x8 >>= count;
    640     mask_4x4 >>= count;
    641     mask_4x4_int >>= count;
    642   }
    643 }
    644 #endif  // CONFIG_VP9_HIGHBITDEPTH
    645 
    646 // This function ors into the current lfm structure, where to do loop
    647 // filters for the specific mi we are looking at. It uses information
    648 // including the block_size_type (32x16, 32x32, etc.), the transform size,
    649 // whether there were any coefficients encoded, and the loop filter strength
    650 // block we are currently looking at. Shift is used to position the
    651 // 1's we produce.
    652 static void build_masks(const loop_filter_info_n *const lfi_n,
    653                         const MODE_INFO *mi, const int shift_y,
    654                         const int shift_uv, LOOP_FILTER_MASK *lfm) {
    655   const BLOCK_SIZE block_size = mi->sb_type;
    656   const TX_SIZE tx_size_y = mi->tx_size;
    657   const TX_SIZE tx_size_uv = uv_txsize_lookup[block_size][tx_size_y][1][1];
    658   const int filter_level = get_filter_level(lfi_n, mi);
    659   uint64_t *const left_y = &lfm->left_y[tx_size_y];
    660   uint64_t *const above_y = &lfm->above_y[tx_size_y];
    661   uint64_t *const int_4x4_y = &lfm->int_4x4_y;
    662   uint16_t *const left_uv = &lfm->left_uv[tx_size_uv];
    663   uint16_t *const above_uv = &lfm->above_uv[tx_size_uv];
    664   uint16_t *const int_4x4_uv = &lfm->int_4x4_uv;
    665   int i;
    666 
    667   // If filter level is 0 we don't loop filter.
    668   if (!filter_level) {
    669     return;
    670   } else {
    671     const int w = num_8x8_blocks_wide_lookup[block_size];
    672     const int h = num_8x8_blocks_high_lookup[block_size];
    673     int index = shift_y;
    674     for (i = 0; i < h; i++) {
    675       memset(&lfm->lfl_y[index], filter_level, w);
    676       index += 8;
    677     }
    678   }
    679 
    680   // These set 1 in the current block size for the block size edges.
    681   // For instance if the block size is 32x16, we'll set:
    682   //    above =   1111
    683   //              0000
    684   //    and
    685   //    left  =   1000
    686   //          =   1000
    687   // NOTE : In this example the low bit is left most ( 1000 ) is stored as
    688   //        1,  not 8...
    689   //
    690   // U and V set things on a 16 bit scale.
    691   //
    692   *above_y |= above_prediction_mask[block_size] << shift_y;
    693   *above_uv |= above_prediction_mask_uv[block_size] << shift_uv;
    694   *left_y |= left_prediction_mask[block_size] << shift_y;
    695   *left_uv |= left_prediction_mask_uv[block_size] << shift_uv;
    696 
    697   // If the block has no coefficients and is not intra we skip applying
    698   // the loop filter on block edges.
    699   if (mi->skip && is_inter_block(mi)) return;
    700 
    701   // Here we are adding a mask for the transform size. The transform
    702   // size mask is set to be correct for a 64x64 prediction block size. We
    703   // mask to match the size of the block we are working on and then shift it
    704   // into place..
    705   *above_y |= (size_mask[block_size] & above_64x64_txform_mask[tx_size_y])
    706               << shift_y;
    707   *above_uv |=
    708       (size_mask_uv[block_size] & above_64x64_txform_mask_uv[tx_size_uv])
    709       << shift_uv;
    710 
    711   *left_y |= (size_mask[block_size] & left_64x64_txform_mask[tx_size_y])
    712              << shift_y;
    713   *left_uv |= (size_mask_uv[block_size] & left_64x64_txform_mask_uv[tx_size_uv])
    714               << shift_uv;
    715 
    716   // Here we are trying to determine what to do with the internal 4x4 block
    717   // boundaries.  These differ from the 4x4 boundaries on the outside edge of
    718   // an 8x8 in that the internal ones can be skipped and don't depend on
    719   // the prediction block size.
    720   if (tx_size_y == TX_4X4) *int_4x4_y |= size_mask[block_size] << shift_y;
    721 
    722   if (tx_size_uv == TX_4X4)
    723     *int_4x4_uv |= (size_mask_uv[block_size] & 0xffff) << shift_uv;
    724 }
    725 
    726 // This function does the same thing as the one above with the exception that
    727 // it only affects the y masks. It exists because for blocks < 16x16 in size,
    728 // we only update u and v masks on the first block.
    729 static void build_y_mask(const loop_filter_info_n *const lfi_n,
    730                          const MODE_INFO *mi, const int shift_y,
    731                          LOOP_FILTER_MASK *lfm) {
    732   const BLOCK_SIZE block_size = mi->sb_type;
    733   const TX_SIZE tx_size_y = mi->tx_size;
    734   const int filter_level = get_filter_level(lfi_n, mi);
    735   uint64_t *const left_y = &lfm->left_y[tx_size_y];
    736   uint64_t *const above_y = &lfm->above_y[tx_size_y];
    737   uint64_t *const int_4x4_y = &lfm->int_4x4_y;
    738   int i;
    739 
    740   if (!filter_level) {
    741     return;
    742   } else {
    743     const int w = num_8x8_blocks_wide_lookup[block_size];
    744     const int h = num_8x8_blocks_high_lookup[block_size];
    745     int index = shift_y;
    746     for (i = 0; i < h; i++) {
    747       memset(&lfm->lfl_y[index], filter_level, w);
    748       index += 8;
    749     }
    750   }
    751 
    752   *above_y |= above_prediction_mask[block_size] << shift_y;
    753   *left_y |= left_prediction_mask[block_size] << shift_y;
    754 
    755   if (mi->skip && is_inter_block(mi)) return;
    756 
    757   *above_y |= (size_mask[block_size] & above_64x64_txform_mask[tx_size_y])
    758               << shift_y;
    759 
    760   *left_y |= (size_mask[block_size] & left_64x64_txform_mask[tx_size_y])
    761              << shift_y;
    762 
    763   if (tx_size_y == TX_4X4) *int_4x4_y |= size_mask[block_size] << shift_y;
    764 }
    765 
    766 void vp9_adjust_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col,
    767                      LOOP_FILTER_MASK *lfm) {
    768   int i;
    769 
    770   // The largest loopfilter we have is 16x16 so we use the 16x16 mask
    771   // for 32x32 transforms also.
    772   lfm->left_y[TX_16X16] |= lfm->left_y[TX_32X32];
    773   lfm->above_y[TX_16X16] |= lfm->above_y[TX_32X32];
    774   lfm->left_uv[TX_16X16] |= lfm->left_uv[TX_32X32];
    775   lfm->above_uv[TX_16X16] |= lfm->above_uv[TX_32X32];
    776 
    777   // We do at least 8 tap filter on every 32x32 even if the transform size
    778   // is 4x4. So if the 4x4 is set on a border pixel add it to the 8x8 and
    779   // remove it from the 4x4.
    780   lfm->left_y[TX_8X8] |= lfm->left_y[TX_4X4] & left_border;
    781   lfm->left_y[TX_4X4] &= ~left_border;
    782   lfm->above_y[TX_8X8] |= lfm->above_y[TX_4X4] & above_border;
    783   lfm->above_y[TX_4X4] &= ~above_border;
    784   lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_4X4] & left_border_uv;
    785   lfm->left_uv[TX_4X4] &= ~left_border_uv;
    786   lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_4X4] & above_border_uv;
    787   lfm->above_uv[TX_4X4] &= ~above_border_uv;
    788 
    789   // We do some special edge handling.
    790   if (mi_row + MI_BLOCK_SIZE > cm->mi_rows) {
    791     const uint64_t rows = cm->mi_rows - mi_row;
    792 
    793     // Each pixel inside the border gets a 1,
    794     const uint64_t mask_y = (((uint64_t)1 << (rows << 3)) - 1);
    795     const uint16_t mask_uv = (((uint16_t)1 << (((rows + 1) >> 1) << 2)) - 1);
    796 
    797     // Remove values completely outside our border.
    798     for (i = 0; i < TX_32X32; i++) {
    799       lfm->left_y[i] &= mask_y;
    800       lfm->above_y[i] &= mask_y;
    801       lfm->left_uv[i] &= mask_uv;
    802       lfm->above_uv[i] &= mask_uv;
    803     }
    804     lfm->int_4x4_y &= mask_y;
    805     lfm->int_4x4_uv &= mask_uv;
    806 
    807     // We don't apply a wide loop filter on the last uv block row. If set
    808     // apply the shorter one instead.
    809     if (rows == 1) {
    810       lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16];
    811       lfm->above_uv[TX_16X16] = 0;
    812     }
    813     if (rows == 5) {
    814       lfm->above_uv[TX_8X8] |= lfm->above_uv[TX_16X16] & 0xff00;
    815       lfm->above_uv[TX_16X16] &= ~(lfm->above_uv[TX_16X16] & 0xff00);
    816     }
    817   }
    818 
    819   if (mi_col + MI_BLOCK_SIZE > cm->mi_cols) {
    820     const uint64_t columns = cm->mi_cols - mi_col;
    821 
    822     // Each pixel inside the border gets a 1, the multiply copies the border
    823     // to where we need it.
    824     const uint64_t mask_y = (((1 << columns) - 1)) * 0x0101010101010101ULL;
    825     const uint16_t mask_uv = ((1 << ((columns + 1) >> 1)) - 1) * 0x1111;
    826 
    827     // Internal edges are not applied on the last column of the image so
    828     // we mask 1 more for the internal edges
    829     const uint16_t mask_uv_int = ((1 << (columns >> 1)) - 1) * 0x1111;
    830 
    831     // Remove the bits outside the image edge.
    832     for (i = 0; i < TX_32X32; i++) {
    833       lfm->left_y[i] &= mask_y;
    834       lfm->above_y[i] &= mask_y;
    835       lfm->left_uv[i] &= mask_uv;
    836       lfm->above_uv[i] &= mask_uv;
    837     }
    838     lfm->int_4x4_y &= mask_y;
    839     lfm->int_4x4_uv &= mask_uv_int;
    840 
    841     // We don't apply a wide loop filter on the last uv column. If set
    842     // apply the shorter one instead.
    843     if (columns == 1) {
    844       lfm->left_uv[TX_8X8] |= lfm->left_uv[TX_16X16];
    845       lfm->left_uv[TX_16X16] = 0;
    846     }
    847     if (columns == 5) {
    848       lfm->left_uv[TX_8X8] |= (lfm->left_uv[TX_16X16] & 0xcccc);
    849       lfm->left_uv[TX_16X16] &= ~(lfm->left_uv[TX_16X16] & 0xcccc);
    850     }
    851   }
    852   // We don't apply a loop filter on the first column in the image, mask that
    853   // out.
    854   if (mi_col == 0) {
    855     for (i = 0; i < TX_32X32; i++) {
    856       lfm->left_y[i] &= 0xfefefefefefefefeULL;
    857       lfm->left_uv[i] &= 0xeeee;
    858     }
    859   }
    860 
    861   // Assert if we try to apply 2 different loop filters at the same position.
    862   assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_8X8]));
    863   assert(!(lfm->left_y[TX_16X16] & lfm->left_y[TX_4X4]));
    864   assert(!(lfm->left_y[TX_8X8] & lfm->left_y[TX_4X4]));
    865   assert(!(lfm->int_4x4_y & lfm->left_y[TX_16X16]));
    866   assert(!(lfm->left_uv[TX_16X16] & lfm->left_uv[TX_8X8]));
    867   assert(!(lfm->left_uv[TX_16X16] & lfm->left_uv[TX_4X4]));
    868   assert(!(lfm->left_uv[TX_8X8] & lfm->left_uv[TX_4X4]));
    869   assert(!(lfm->int_4x4_uv & lfm->left_uv[TX_16X16]));
    870   assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_8X8]));
    871   assert(!(lfm->above_y[TX_16X16] & lfm->above_y[TX_4X4]));
    872   assert(!(lfm->above_y[TX_8X8] & lfm->above_y[TX_4X4]));
    873   assert(!(lfm->int_4x4_y & lfm->above_y[TX_16X16]));
    874   assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_8X8]));
    875   assert(!(lfm->above_uv[TX_16X16] & lfm->above_uv[TX_4X4]));
    876   assert(!(lfm->above_uv[TX_8X8] & lfm->above_uv[TX_4X4]));
    877   assert(!(lfm->int_4x4_uv & lfm->above_uv[TX_16X16]));
    878 }
    879 
    880 // This function sets up the bit masks for the entire 64x64 region represented
    881 // by mi_row, mi_col.
    882 void vp9_setup_mask(VP9_COMMON *const cm, const int mi_row, const int mi_col,
    883                     MODE_INFO **mi8x8, const int mode_info_stride,
    884                     LOOP_FILTER_MASK *lfm) {
    885   int idx_32, idx_16, idx_8;
    886   const loop_filter_info_n *const lfi_n = &cm->lf_info;
    887   MODE_INFO **mip = mi8x8;
    888   MODE_INFO **mip2 = mi8x8;
    889 
    890   // These are offsets to the next mi in the 64x64 block. It is what gets
    891   // added to the mi ptr as we go through each loop. It helps us to avoid
    892   // setting up special row and column counters for each index. The last step
    893   // brings us out back to the starting position.
    894   const int offset_32[] = { 4, (mode_info_stride << 2) - 4, 4,
    895                             -(mode_info_stride << 2) - 4 };
    896   const int offset_16[] = { 2, (mode_info_stride << 1) - 2, 2,
    897                             -(mode_info_stride << 1) - 2 };
    898   const int offset[] = { 1, mode_info_stride - 1, 1, -mode_info_stride - 1 };
    899 
    900   // Following variables represent shifts to position the current block
    901   // mask over the appropriate block. A shift of 36 to the left will move
    902   // the bits for the final 32 by 32 block in the 64x64 up 4 rows and left
    903   // 4 rows to the appropriate spot.
    904   const int shift_32_y[] = { 0, 4, 32, 36 };
    905   const int shift_16_y[] = { 0, 2, 16, 18 };
    906   const int shift_8_y[] = { 0, 1, 8, 9 };
    907   const int shift_32_uv[] = { 0, 2, 8, 10 };
    908   const int shift_16_uv[] = { 0, 1, 4, 5 };
    909   const int max_rows =
    910       (mi_row + MI_BLOCK_SIZE > cm->mi_rows ? cm->mi_rows - mi_row
    911                                             : MI_BLOCK_SIZE);
    912   const int max_cols =
    913       (mi_col + MI_BLOCK_SIZE > cm->mi_cols ? cm->mi_cols - mi_col
    914                                             : MI_BLOCK_SIZE);
    915 
    916   vp9_zero(*lfm);
    917   assert(mip[0] != NULL);
    918 
    919   switch (mip[0]->sb_type) {
    920     case BLOCK_64X64: build_masks(lfi_n, mip[0], 0, 0, lfm); break;
    921     case BLOCK_64X32:
    922       build_masks(lfi_n, mip[0], 0, 0, lfm);
    923       mip2 = mip + mode_info_stride * 4;
    924       if (4 >= max_rows) break;
    925       build_masks(lfi_n, mip2[0], 32, 8, lfm);
    926       break;
    927     case BLOCK_32X64:
    928       build_masks(lfi_n, mip[0], 0, 0, lfm);
    929       mip2 = mip + 4;
    930       if (4 >= max_cols) break;
    931       build_masks(lfi_n, mip2[0], 4, 2, lfm);
    932       break;
    933     default:
    934       for (idx_32 = 0; idx_32 < 4; mip += offset_32[idx_32], ++idx_32) {
    935         const int shift_y = shift_32_y[idx_32];
    936         const int shift_uv = shift_32_uv[idx_32];
    937         const int mi_32_col_offset = ((idx_32 & 1) << 2);
    938         const int mi_32_row_offset = ((idx_32 >> 1) << 2);
    939         if (mi_32_col_offset >= max_cols || mi_32_row_offset >= max_rows)
    940           continue;
    941         switch (mip[0]->sb_type) {
    942           case BLOCK_32X32:
    943             build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
    944             break;
    945           case BLOCK_32X16:
    946             build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
    947             if (mi_32_row_offset + 2 >= max_rows) continue;
    948             mip2 = mip + mode_info_stride * 2;
    949             build_masks(lfi_n, mip2[0], shift_y + 16, shift_uv + 4, lfm);
    950             break;
    951           case BLOCK_16X32:
    952             build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
    953             if (mi_32_col_offset + 2 >= max_cols) continue;
    954             mip2 = mip + 2;
    955             build_masks(lfi_n, mip2[0], shift_y + 2, shift_uv + 1, lfm);
    956             break;
    957           default:
    958             for (idx_16 = 0; idx_16 < 4; mip += offset_16[idx_16], ++idx_16) {
    959               const int shift_y = shift_32_y[idx_32] + shift_16_y[idx_16];
    960               const int shift_uv = shift_32_uv[idx_32] + shift_16_uv[idx_16];
    961               const int mi_16_col_offset =
    962                   mi_32_col_offset + ((idx_16 & 1) << 1);
    963               const int mi_16_row_offset =
    964                   mi_32_row_offset + ((idx_16 >> 1) << 1);
    965 
    966               if (mi_16_col_offset >= max_cols || mi_16_row_offset >= max_rows)
    967                 continue;
    968 
    969               switch (mip[0]->sb_type) {
    970                 case BLOCK_16X16:
    971                   build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
    972                   break;
    973                 case BLOCK_16X8:
    974                   build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
    975                   if (mi_16_row_offset + 1 >= max_rows) continue;
    976                   mip2 = mip + mode_info_stride;
    977                   build_y_mask(lfi_n, mip2[0], shift_y + 8, lfm);
    978                   break;
    979                 case BLOCK_8X16:
    980                   build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
    981                   if (mi_16_col_offset + 1 >= max_cols) continue;
    982                   mip2 = mip + 1;
    983                   build_y_mask(lfi_n, mip2[0], shift_y + 1, lfm);
    984                   break;
    985                 default: {
    986                   const int shift_y =
    987                       shift_32_y[idx_32] + shift_16_y[idx_16] + shift_8_y[0];
    988                   build_masks(lfi_n, mip[0], shift_y, shift_uv, lfm);
    989                   mip += offset[0];
    990                   for (idx_8 = 1; idx_8 < 4; mip += offset[idx_8], ++idx_8) {
    991                     const int shift_y = shift_32_y[idx_32] +
    992                                         shift_16_y[idx_16] + shift_8_y[idx_8];
    993                     const int mi_8_col_offset =
    994                         mi_16_col_offset + ((idx_8 & 1));
    995                     const int mi_8_row_offset =
    996                         mi_16_row_offset + ((idx_8 >> 1));
    997 
    998                     if (mi_8_col_offset >= max_cols ||
    999                         mi_8_row_offset >= max_rows)
   1000                       continue;
   1001                     build_y_mask(lfi_n, mip[0], shift_y, lfm);
   1002                   }
   1003                   break;
   1004                 }
   1005               }
   1006             }
   1007             break;
   1008         }
   1009       }
   1010       break;
   1011   }
   1012 }
   1013 
   1014 static void filter_selectively_vert(
   1015     uint8_t *s, int pitch, unsigned int mask_16x16, unsigned int mask_8x8,
   1016     unsigned int mask_4x4, unsigned int mask_4x4_int,
   1017     const loop_filter_thresh *lfthr, const uint8_t *lfl) {
   1018   unsigned int mask;
   1019 
   1020   for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int; mask;
   1021        mask >>= 1) {
   1022     const loop_filter_thresh *lfi = lfthr + *lfl;
   1023 
   1024     if (mask & 1) {
   1025       if (mask_16x16 & 1) {
   1026         vpx_lpf_vertical_16(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
   1027       } else if (mask_8x8 & 1) {
   1028         vpx_lpf_vertical_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
   1029       } else if (mask_4x4 & 1) {
   1030         vpx_lpf_vertical_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
   1031       }
   1032     }
   1033     if (mask_4x4_int & 1)
   1034       vpx_lpf_vertical_4(s + 4, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);
   1035     s += 8;
   1036     lfl += 1;
   1037     mask_16x16 >>= 1;
   1038     mask_8x8 >>= 1;
   1039     mask_4x4 >>= 1;
   1040     mask_4x4_int >>= 1;
   1041   }
   1042 }
   1043 
   1044 #if CONFIG_VP9_HIGHBITDEPTH
   1045 static void highbd_filter_selectively_vert(
   1046     uint16_t *s, int pitch, unsigned int mask_16x16, unsigned int mask_8x8,
   1047     unsigned int mask_4x4, unsigned int mask_4x4_int,
   1048     const loop_filter_thresh *lfthr, const uint8_t *lfl, int bd) {
   1049   unsigned int mask;
   1050 
   1051   for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int; mask;
   1052        mask >>= 1) {
   1053     const loop_filter_thresh *lfi = lfthr + *lfl;
   1054 
   1055     if (mask & 1) {
   1056       if (mask_16x16 & 1) {
   1057         vpx_highbd_lpf_vertical_16(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr,
   1058                                    bd);
   1059       } else if (mask_8x8 & 1) {
   1060         vpx_highbd_lpf_vertical_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr,
   1061                                   bd);
   1062       } else if (mask_4x4 & 1) {
   1063         vpx_highbd_lpf_vertical_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr,
   1064                                   bd);
   1065       }
   1066     }
   1067     if (mask_4x4_int & 1)
   1068       vpx_highbd_lpf_vertical_4(s + 4, pitch, lfi->mblim, lfi->lim,
   1069                                 lfi->hev_thr, bd);
   1070     s += 8;
   1071     lfl += 1;
   1072     mask_16x16 >>= 1;
   1073     mask_8x8 >>= 1;
   1074     mask_4x4 >>= 1;
   1075     mask_4x4_int >>= 1;
   1076   }
   1077 }
   1078 #endif  // CONFIG_VP9_HIGHBITDEPTH
   1079 
   1080 void vp9_filter_block_plane_non420(VP9_COMMON *cm,
   1081                                    struct macroblockd_plane *plane,
   1082                                    MODE_INFO **mi_8x8, int mi_row, int mi_col) {
   1083   const int ss_x = plane->subsampling_x;
   1084   const int ss_y = plane->subsampling_y;
   1085   const int row_step = 1 << ss_y;
   1086   const int col_step = 1 << ss_x;
   1087   const int row_step_stride = cm->mi_stride * row_step;
   1088   struct buf_2d *const dst = &plane->dst;
   1089   uint8_t *const dst0 = dst->buf;
   1090   unsigned int mask_16x16[MI_BLOCK_SIZE];
   1091   unsigned int mask_8x8[MI_BLOCK_SIZE];
   1092   unsigned int mask_4x4[MI_BLOCK_SIZE];
   1093   unsigned int mask_4x4_int[MI_BLOCK_SIZE];
   1094   uint8_t lfl[MI_BLOCK_SIZE * MI_BLOCK_SIZE];
   1095   int r, c;
   1096 
   1097   vp9_zero(mask_16x16);
   1098   vp9_zero(mask_8x8);
   1099   vp9_zero(mask_4x4);
   1100   vp9_zero(mask_4x4_int);
   1101   vp9_zero(lfl);
   1102 
   1103   for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
   1104     unsigned int mask_16x16_c = 0;
   1105     unsigned int mask_8x8_c = 0;
   1106     unsigned int mask_4x4_c = 0;
   1107     unsigned int border_mask;
   1108 
   1109     // Determine the vertical edges that need filtering
   1110     for (c = 0; c < MI_BLOCK_SIZE && mi_col + c < cm->mi_cols; c += col_step) {
   1111       const MODE_INFO *mi = mi_8x8[c];
   1112       const BLOCK_SIZE sb_type = mi[0].sb_type;
   1113       const int skip_this = mi[0].skip && is_inter_block(mi);
   1114       // left edge of current unit is block/partition edge -> no skip
   1115       const int block_edge_left =
   1116           (num_4x4_blocks_wide_lookup[sb_type] > 1)
   1117               ? !(c & (num_8x8_blocks_wide_lookup[sb_type] - 1))
   1118               : 1;
   1119       const int skip_this_c = skip_this && !block_edge_left;
   1120       // top edge of current unit is block/partition edge -> no skip
   1121       const int block_edge_above =
   1122           (num_4x4_blocks_high_lookup[sb_type] > 1)
   1123               ? !(r & (num_8x8_blocks_high_lookup[sb_type] - 1))
   1124               : 1;
   1125       const int skip_this_r = skip_this && !block_edge_above;
   1126       const TX_SIZE tx_size = get_uv_tx_size(mi, plane);
   1127       const int skip_border_4x4_c = ss_x && mi_col + c == cm->mi_cols - 1;
   1128       const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
   1129 
   1130       // Filter level can vary per MI
   1131       if (!(lfl[(r << 3) + (c >> ss_x)] = get_filter_level(&cm->lf_info, mi)))
   1132         continue;
   1133 
   1134       // Build masks based on the transform size of each block
   1135       if (tx_size == TX_32X32) {
   1136         if (!skip_this_c && ((c >> ss_x) & 3) == 0) {
   1137           if (!skip_border_4x4_c)
   1138             mask_16x16_c |= 1 << (c >> ss_x);
   1139           else
   1140             mask_8x8_c |= 1 << (c >> ss_x);
   1141         }
   1142         if (!skip_this_r && ((r >> ss_y) & 3) == 0) {
   1143           if (!skip_border_4x4_r)
   1144             mask_16x16[r] |= 1 << (c >> ss_x);
   1145           else
   1146             mask_8x8[r] |= 1 << (c >> ss_x);
   1147         }
   1148       } else if (tx_size == TX_16X16) {
   1149         if (!skip_this_c && ((c >> ss_x) & 1) == 0) {
   1150           if (!skip_border_4x4_c)
   1151             mask_16x16_c |= 1 << (c >> ss_x);
   1152           else
   1153             mask_8x8_c |= 1 << (c >> ss_x);
   1154         }
   1155         if (!skip_this_r && ((r >> ss_y) & 1) == 0) {
   1156           if (!skip_border_4x4_r)
   1157             mask_16x16[r] |= 1 << (c >> ss_x);
   1158           else
   1159             mask_8x8[r] |= 1 << (c >> ss_x);
   1160         }
   1161       } else {
   1162         // force 8x8 filtering on 32x32 boundaries
   1163         if (!skip_this_c) {
   1164           if (tx_size == TX_8X8 || ((c >> ss_x) & 3) == 0)
   1165             mask_8x8_c |= 1 << (c >> ss_x);
   1166           else
   1167             mask_4x4_c |= 1 << (c >> ss_x);
   1168         }
   1169 
   1170         if (!skip_this_r) {
   1171           if (tx_size == TX_8X8 || ((r >> ss_y) & 3) == 0)
   1172             mask_8x8[r] |= 1 << (c >> ss_x);
   1173           else
   1174             mask_4x4[r] |= 1 << (c >> ss_x);
   1175         }
   1176 
   1177         if (!skip_this && tx_size < TX_8X8 && !skip_border_4x4_c)
   1178           mask_4x4_int[r] |= 1 << (c >> ss_x);
   1179       }
   1180     }
   1181 
   1182     // Disable filtering on the leftmost column
   1183     border_mask = ~(mi_col == 0 ? 1 : 0);
   1184 #if CONFIG_VP9_HIGHBITDEPTH
   1185     if (cm->use_highbitdepth) {
   1186       highbd_filter_selectively_vert(
   1187           CONVERT_TO_SHORTPTR(dst->buf), dst->stride,
   1188           mask_16x16_c & border_mask, mask_8x8_c & border_mask,
   1189           mask_4x4_c & border_mask, mask_4x4_int[r], cm->lf_info.lfthr,
   1190           &lfl[r << 3], (int)cm->bit_depth);
   1191     } else {
   1192 #endif  // CONFIG_VP9_HIGHBITDEPTH
   1193       filter_selectively_vert(dst->buf, dst->stride, mask_16x16_c & border_mask,
   1194                               mask_8x8_c & border_mask,
   1195                               mask_4x4_c & border_mask, mask_4x4_int[r],
   1196                               cm->lf_info.lfthr, &lfl[r << 3]);
   1197 #if CONFIG_VP9_HIGHBITDEPTH
   1198     }
   1199 #endif  // CONFIG_VP9_HIGHBITDEPTH
   1200     dst->buf += 8 * dst->stride;
   1201     mi_8x8 += row_step_stride;
   1202   }
   1203 
   1204   // Now do horizontal pass
   1205   dst->buf = dst0;
   1206   for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += row_step) {
   1207     const int skip_border_4x4_r = ss_y && mi_row + r == cm->mi_rows - 1;
   1208     const unsigned int mask_4x4_int_r = skip_border_4x4_r ? 0 : mask_4x4_int[r];
   1209 
   1210     unsigned int mask_16x16_r;
   1211     unsigned int mask_8x8_r;
   1212     unsigned int mask_4x4_r;
   1213 
   1214     if (mi_row + r == 0) {
   1215       mask_16x16_r = 0;
   1216       mask_8x8_r = 0;
   1217       mask_4x4_r = 0;
   1218     } else {
   1219       mask_16x16_r = mask_16x16[r];
   1220       mask_8x8_r = mask_8x8[r];
   1221       mask_4x4_r = mask_4x4[r];
   1222     }
   1223 #if CONFIG_VP9_HIGHBITDEPTH
   1224     if (cm->use_highbitdepth) {
   1225       highbd_filter_selectively_horiz(
   1226           CONVERT_TO_SHORTPTR(dst->buf), dst->stride, mask_16x16_r, mask_8x8_r,
   1227           mask_4x4_r, mask_4x4_int_r, cm->lf_info.lfthr, &lfl[r << 3],
   1228           (int)cm->bit_depth);
   1229     } else {
   1230 #endif  // CONFIG_VP9_HIGHBITDEPTH
   1231       filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
   1232                                mask_4x4_r, mask_4x4_int_r, cm->lf_info.lfthr,
   1233                                &lfl[r << 3]);
   1234 #if CONFIG_VP9_HIGHBITDEPTH
   1235     }
   1236 #endif  // CONFIG_VP9_HIGHBITDEPTH
   1237     dst->buf += 8 * dst->stride;
   1238   }
   1239 }
   1240 
   1241 void vp9_filter_block_plane_ss00(VP9_COMMON *const cm,
   1242                                  struct macroblockd_plane *const plane,
   1243                                  int mi_row, LOOP_FILTER_MASK *lfm) {
   1244   struct buf_2d *const dst = &plane->dst;
   1245   uint8_t *const dst0 = dst->buf;
   1246   int r;
   1247   uint64_t mask_16x16 = lfm->left_y[TX_16X16];
   1248   uint64_t mask_8x8 = lfm->left_y[TX_8X8];
   1249   uint64_t mask_4x4 = lfm->left_y[TX_4X4];
   1250   uint64_t mask_4x4_int = lfm->int_4x4_y;
   1251 
   1252   assert(plane->subsampling_x == 0 && plane->subsampling_y == 0);
   1253 
   1254   // Vertical pass: do 2 rows at one time
   1255   for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 2) {
   1256 #if CONFIG_VP9_HIGHBITDEPTH
   1257     if (cm->use_highbitdepth) {
   1258       // Disable filtering on the leftmost column.
   1259       highbd_filter_selectively_vert_row2(
   1260           plane->subsampling_x, CONVERT_TO_SHORTPTR(dst->buf), dst->stride,
   1261           (unsigned int)mask_16x16, (unsigned int)mask_8x8,
   1262           (unsigned int)mask_4x4, (unsigned int)mask_4x4_int, cm->lf_info.lfthr,
   1263           &lfm->lfl_y[r << 3], (int)cm->bit_depth);
   1264     } else {
   1265 #endif  // CONFIG_VP9_HIGHBITDEPTH
   1266       // Disable filtering on the leftmost column.
   1267       filter_selectively_vert_row2(
   1268           plane->subsampling_x, dst->buf, dst->stride, (unsigned int)mask_16x16,
   1269           (unsigned int)mask_8x8, (unsigned int)mask_4x4,
   1270           (unsigned int)mask_4x4_int, cm->lf_info.lfthr, &lfm->lfl_y[r << 3]);
   1271 #if CONFIG_VP9_HIGHBITDEPTH
   1272     }
   1273 #endif  // CONFIG_VP9_HIGHBITDEPTH
   1274     dst->buf += 16 * dst->stride;
   1275     mask_16x16 >>= 16;
   1276     mask_8x8 >>= 16;
   1277     mask_4x4 >>= 16;
   1278     mask_4x4_int >>= 16;
   1279   }
   1280 
   1281   // Horizontal pass
   1282   dst->buf = dst0;
   1283   mask_16x16 = lfm->above_y[TX_16X16];
   1284   mask_8x8 = lfm->above_y[TX_8X8];
   1285   mask_4x4 = lfm->above_y[TX_4X4];
   1286   mask_4x4_int = lfm->int_4x4_y;
   1287 
   1288   for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r++) {
   1289     unsigned int mask_16x16_r;
   1290     unsigned int mask_8x8_r;
   1291     unsigned int mask_4x4_r;
   1292 
   1293     if (mi_row + r == 0) {
   1294       mask_16x16_r = 0;
   1295       mask_8x8_r = 0;
   1296       mask_4x4_r = 0;
   1297     } else {
   1298       mask_16x16_r = mask_16x16 & 0xff;
   1299       mask_8x8_r = mask_8x8 & 0xff;
   1300       mask_4x4_r = mask_4x4 & 0xff;
   1301     }
   1302 
   1303 #if CONFIG_VP9_HIGHBITDEPTH
   1304     if (cm->use_highbitdepth) {
   1305       highbd_filter_selectively_horiz(
   1306           CONVERT_TO_SHORTPTR(dst->buf), dst->stride, mask_16x16_r, mask_8x8_r,
   1307           mask_4x4_r, mask_4x4_int & 0xff, cm->lf_info.lfthr,
   1308           &lfm->lfl_y[r << 3], (int)cm->bit_depth);
   1309     } else {
   1310 #endif  // CONFIG_VP9_HIGHBITDEPTH
   1311       filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
   1312                                mask_4x4_r, mask_4x4_int & 0xff,
   1313                                cm->lf_info.lfthr, &lfm->lfl_y[r << 3]);
   1314 #if CONFIG_VP9_HIGHBITDEPTH
   1315     }
   1316 #endif  // CONFIG_VP9_HIGHBITDEPTH
   1317 
   1318     dst->buf += 8 * dst->stride;
   1319     mask_16x16 >>= 8;
   1320     mask_8x8 >>= 8;
   1321     mask_4x4 >>= 8;
   1322     mask_4x4_int >>= 8;
   1323   }
   1324 }
   1325 
   1326 void vp9_filter_block_plane_ss11(VP9_COMMON *const cm,
   1327                                  struct macroblockd_plane *const plane,
   1328                                  int mi_row, LOOP_FILTER_MASK *lfm) {
   1329   struct buf_2d *const dst = &plane->dst;
   1330   uint8_t *const dst0 = dst->buf;
   1331   int r, c;
   1332   uint8_t lfl_uv[16];
   1333 
   1334   uint16_t mask_16x16 = lfm->left_uv[TX_16X16];
   1335   uint16_t mask_8x8 = lfm->left_uv[TX_8X8];
   1336   uint16_t mask_4x4 = lfm->left_uv[TX_4X4];
   1337   uint16_t mask_4x4_int = lfm->int_4x4_uv;
   1338 
   1339   vp9_zero(lfl_uv);
   1340 
   1341   assert(plane->subsampling_x == 1 && plane->subsampling_y == 1);
   1342 
   1343   // Vertical pass: do 2 rows at one time
   1344   for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 4) {
   1345     for (c = 0; c < (MI_BLOCK_SIZE >> 1); c++) {
   1346       lfl_uv[(r << 1) + c] = lfm->lfl_y[(r << 3) + (c << 1)];
   1347       lfl_uv[((r + 2) << 1) + c] = lfm->lfl_y[((r + 2) << 3) + (c << 1)];
   1348     }
   1349 
   1350 #if CONFIG_VP9_HIGHBITDEPTH
   1351     if (cm->use_highbitdepth) {
   1352       // Disable filtering on the leftmost column.
   1353       highbd_filter_selectively_vert_row2(
   1354           plane->subsampling_x, CONVERT_TO_SHORTPTR(dst->buf), dst->stride,
   1355           (unsigned int)mask_16x16, (unsigned int)mask_8x8,
   1356           (unsigned int)mask_4x4, (unsigned int)mask_4x4_int, cm->lf_info.lfthr,
   1357           &lfl_uv[r << 1], (int)cm->bit_depth);
   1358     } else {
   1359 #endif  // CONFIG_VP9_HIGHBITDEPTH
   1360       // Disable filtering on the leftmost column.
   1361       filter_selectively_vert_row2(
   1362           plane->subsampling_x, dst->buf, dst->stride, (unsigned int)mask_16x16,
   1363           (unsigned int)mask_8x8, (unsigned int)mask_4x4,
   1364           (unsigned int)mask_4x4_int, cm->lf_info.lfthr, &lfl_uv[r << 1]);
   1365 #if CONFIG_VP9_HIGHBITDEPTH
   1366     }
   1367 #endif  // CONFIG_VP9_HIGHBITDEPTH
   1368 
   1369     dst->buf += 16 * dst->stride;
   1370     mask_16x16 >>= 8;
   1371     mask_8x8 >>= 8;
   1372     mask_4x4 >>= 8;
   1373     mask_4x4_int >>= 8;
   1374   }
   1375 
   1376   // Horizontal pass
   1377   dst->buf = dst0;
   1378   mask_16x16 = lfm->above_uv[TX_16X16];
   1379   mask_8x8 = lfm->above_uv[TX_8X8];
   1380   mask_4x4 = lfm->above_uv[TX_4X4];
   1381   mask_4x4_int = lfm->int_4x4_uv;
   1382 
   1383   for (r = 0; r < MI_BLOCK_SIZE && mi_row + r < cm->mi_rows; r += 2) {
   1384     const int skip_border_4x4_r = mi_row + r == cm->mi_rows - 1;
   1385     const unsigned int mask_4x4_int_r =
   1386         skip_border_4x4_r ? 0 : (mask_4x4_int & 0xf);
   1387     unsigned int mask_16x16_r;
   1388     unsigned int mask_8x8_r;
   1389     unsigned int mask_4x4_r;
   1390 
   1391     if (mi_row + r == 0) {
   1392       mask_16x16_r = 0;
   1393       mask_8x8_r = 0;
   1394       mask_4x4_r = 0;
   1395     } else {
   1396       mask_16x16_r = mask_16x16 & 0xf;
   1397       mask_8x8_r = mask_8x8 & 0xf;
   1398       mask_4x4_r = mask_4x4 & 0xf;
   1399     }
   1400 
   1401 #if CONFIG_VP9_HIGHBITDEPTH
   1402     if (cm->use_highbitdepth) {
   1403       highbd_filter_selectively_horiz(
   1404           CONVERT_TO_SHORTPTR(dst->buf), dst->stride, mask_16x16_r, mask_8x8_r,
   1405           mask_4x4_r, mask_4x4_int_r, cm->lf_info.lfthr, &lfl_uv[r << 1],
   1406           (int)cm->bit_depth);
   1407     } else {
   1408 #endif  // CONFIG_VP9_HIGHBITDEPTH
   1409       filter_selectively_horiz(dst->buf, dst->stride, mask_16x16_r, mask_8x8_r,
   1410                                mask_4x4_r, mask_4x4_int_r, cm->lf_info.lfthr,
   1411                                &lfl_uv[r << 1]);
   1412 #if CONFIG_VP9_HIGHBITDEPTH
   1413     }
   1414 #endif  // CONFIG_VP9_HIGHBITDEPTH
   1415 
   1416     dst->buf += 8 * dst->stride;
   1417     mask_16x16 >>= 4;
   1418     mask_8x8 >>= 4;
   1419     mask_4x4 >>= 4;
   1420     mask_4x4_int >>= 4;
   1421   }
   1422 }
   1423 
   1424 static void loop_filter_rows(YV12_BUFFER_CONFIG *frame_buffer, VP9_COMMON *cm,
   1425                              struct macroblockd_plane planes[MAX_MB_PLANE],
   1426                              int start, int stop, int y_only) {
   1427   const int num_planes = y_only ? 1 : MAX_MB_PLANE;
   1428   enum lf_path path;
   1429   int mi_row, mi_col;
   1430 
   1431   if (y_only)
   1432     path = LF_PATH_444;
   1433   else if (planes[1].subsampling_y == 1 && planes[1].subsampling_x == 1)
   1434     path = LF_PATH_420;
   1435   else if (planes[1].subsampling_y == 0 && planes[1].subsampling_x == 0)
   1436     path = LF_PATH_444;
   1437   else
   1438     path = LF_PATH_SLOW;
   1439 
   1440   for (mi_row = start; mi_row < stop; mi_row += MI_BLOCK_SIZE) {
   1441     MODE_INFO **mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
   1442     LOOP_FILTER_MASK *lfm = get_lfm(&cm->lf, mi_row, 0);
   1443 
   1444     for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE, ++lfm) {
   1445       int plane;
   1446 
   1447       vp9_setup_dst_planes(planes, frame_buffer, mi_row, mi_col);
   1448 
   1449       // TODO(jimbankoski): For 444 only need to do y mask.
   1450       vp9_adjust_mask(cm, mi_row, mi_col, lfm);
   1451 
   1452       vp9_filter_block_plane_ss00(cm, &planes[0], mi_row, lfm);
   1453       for (plane = 1; plane < num_planes; ++plane) {
   1454         switch (path) {
   1455           case LF_PATH_420:
   1456             vp9_filter_block_plane_ss11(cm, &planes[plane], mi_row, lfm);
   1457             break;
   1458           case LF_PATH_444:
   1459             vp9_filter_block_plane_ss00(cm, &planes[plane], mi_row, lfm);
   1460             break;
   1461           case LF_PATH_SLOW:
   1462             vp9_filter_block_plane_non420(cm, &planes[plane], mi + mi_col,
   1463                                           mi_row, mi_col);
   1464             break;
   1465         }
   1466       }
   1467     }
   1468   }
   1469 }
   1470 
   1471 void vp9_loop_filter_frame(YV12_BUFFER_CONFIG *frame, VP9_COMMON *cm,
   1472                            MACROBLOCKD *xd, int frame_filter_level, int y_only,
   1473                            int partial_frame) {
   1474   int start_mi_row, end_mi_row, mi_rows_to_filter;
   1475   if (!frame_filter_level) return;
   1476   start_mi_row = 0;
   1477   mi_rows_to_filter = cm->mi_rows;
   1478   if (partial_frame && cm->mi_rows > 8) {
   1479     start_mi_row = cm->mi_rows >> 1;
   1480     start_mi_row &= 0xfffffff8;
   1481     mi_rows_to_filter = VPXMAX(cm->mi_rows / 8, 8);
   1482   }
   1483   end_mi_row = start_mi_row + mi_rows_to_filter;
   1484   loop_filter_rows(frame, cm, xd->plane, start_mi_row, end_mi_row, y_only);
   1485 }
   1486 
   1487 // Used by the encoder to build the loopfilter masks.
   1488 // TODO(slavarnway): Do the encoder the same way the decoder does it and
   1489 //                   build the masks in line as part of the encode process.
   1490 void vp9_build_mask_frame(VP9_COMMON *cm, int frame_filter_level,
   1491                           int partial_frame) {
   1492   int start_mi_row, end_mi_row, mi_rows_to_filter;
   1493   int mi_col, mi_row;
   1494   if (!frame_filter_level) return;
   1495   start_mi_row = 0;
   1496   mi_rows_to_filter = cm->mi_rows;
   1497   if (partial_frame && cm->mi_rows > 8) {
   1498     start_mi_row = cm->mi_rows >> 1;
   1499     start_mi_row &= 0xfffffff8;
   1500     mi_rows_to_filter = VPXMAX(cm->mi_rows / 8, 8);
   1501   }
   1502   end_mi_row = start_mi_row + mi_rows_to_filter;
   1503 
   1504   vp9_loop_filter_frame_init(cm, frame_filter_level);
   1505 
   1506   for (mi_row = start_mi_row; mi_row < end_mi_row; mi_row += MI_BLOCK_SIZE) {
   1507     MODE_INFO **mi = cm->mi_grid_visible + mi_row * cm->mi_stride;
   1508     for (mi_col = 0; mi_col < cm->mi_cols; mi_col += MI_BLOCK_SIZE) {
   1509       // vp9_setup_mask() zeros lfm
   1510       vp9_setup_mask(cm, mi_row, mi_col, mi + mi_col, cm->mi_stride,
   1511                      get_lfm(&cm->lf, mi_row, mi_col));
   1512     }
   1513   }
   1514 }
   1515 
   1516 // 8x8 blocks in a superblock.  A "1" represents the first block in a 16x16
   1517 // or greater area.
   1518 static const uint8_t first_block_in_16x16[8][8] = {
   1519   { 1, 0, 1, 0, 1, 0, 1, 0 }, { 0, 0, 0, 0, 0, 0, 0, 0 },
   1520   { 1, 0, 1, 0, 1, 0, 1, 0 }, { 0, 0, 0, 0, 0, 0, 0, 0 },
   1521   { 1, 0, 1, 0, 1, 0, 1, 0 }, { 0, 0, 0, 0, 0, 0, 0, 0 },
   1522   { 1, 0, 1, 0, 1, 0, 1, 0 }, { 0, 0, 0, 0, 0, 0, 0, 0 }
   1523 };
   1524 
   1525 // This function sets up the bit masks for a block represented
   1526 // by mi_row, mi_col in a 64x64 region.
   1527 // TODO(SJL): This function only works for yv12.
   1528 void vp9_build_mask(VP9_COMMON *cm, const MODE_INFO *mi, int mi_row, int mi_col,
   1529                     int bw, int bh) {
   1530   const BLOCK_SIZE block_size = mi->sb_type;
   1531   const TX_SIZE tx_size_y = mi->tx_size;
   1532   const loop_filter_info_n *const lfi_n = &cm->lf_info;
   1533   const int filter_level = get_filter_level(lfi_n, mi);
   1534   const TX_SIZE tx_size_uv = uv_txsize_lookup[block_size][tx_size_y][1][1];
   1535   LOOP_FILTER_MASK *const lfm = get_lfm(&cm->lf, mi_row, mi_col);
   1536   uint64_t *const left_y = &lfm->left_y[tx_size_y];
   1537   uint64_t *const above_y = &lfm->above_y[tx_size_y];
   1538   uint64_t *const int_4x4_y = &lfm->int_4x4_y;
   1539   uint16_t *const left_uv = &lfm->left_uv[tx_size_uv];
   1540   uint16_t *const above_uv = &lfm->above_uv[tx_size_uv];
   1541   uint16_t *const int_4x4_uv = &lfm->int_4x4_uv;
   1542   const int row_in_sb = (mi_row & 7);
   1543   const int col_in_sb = (mi_col & 7);
   1544   const int shift_y = col_in_sb + (row_in_sb << 3);
   1545   const int shift_uv = (col_in_sb >> 1) + ((row_in_sb >> 1) << 2);
   1546   const int build_uv = first_block_in_16x16[row_in_sb][col_in_sb];
   1547 
   1548   if (!filter_level) {
   1549     return;
   1550   } else {
   1551     int index = shift_y;
   1552     int i;
   1553     for (i = 0; i < bh; i++) {
   1554       memset(&lfm->lfl_y[index], filter_level, bw);
   1555       index += 8;
   1556     }
   1557   }
   1558 
   1559   // These set 1 in the current block size for the block size edges.
   1560   // For instance if the block size is 32x16, we'll set:
   1561   //    above =   1111
   1562   //              0000
   1563   //    and
   1564   //    left  =   1000
   1565   //          =   1000
   1566   // NOTE : In this example the low bit is left most ( 1000 ) is stored as
   1567   //        1,  not 8...
   1568   //
   1569   // U and V set things on a 16 bit scale.
   1570   //
   1571   *above_y |= above_prediction_mask[block_size] << shift_y;
   1572   *left_y |= left_prediction_mask[block_size] << shift_y;
   1573 
   1574   if (build_uv) {
   1575     *above_uv |= above_prediction_mask_uv[block_size] << shift_uv;
   1576     *left_uv |= left_prediction_mask_uv[block_size] << shift_uv;
   1577   }
   1578 
   1579   // If the block has no coefficients and is not intra we skip applying
   1580   // the loop filter on block edges.
   1581   if (mi->skip && is_inter_block(mi)) return;
   1582 
   1583   // Add a mask for the transform size. The transform size mask is set to
   1584   // be correct for a 64x64 prediction block size. Mask to match the size of
   1585   // the block we are working on and then shift it into place.
   1586   *above_y |= (size_mask[block_size] & above_64x64_txform_mask[tx_size_y])
   1587               << shift_y;
   1588   *left_y |= (size_mask[block_size] & left_64x64_txform_mask[tx_size_y])
   1589              << shift_y;
   1590 
   1591   if (build_uv) {
   1592     *above_uv |=
   1593         (size_mask_uv[block_size] & above_64x64_txform_mask_uv[tx_size_uv])
   1594         << shift_uv;
   1595 
   1596     *left_uv |=
   1597         (size_mask_uv[block_size] & left_64x64_txform_mask_uv[tx_size_uv])
   1598         << shift_uv;
   1599   }
   1600 
   1601   // Try to determine what to do with the internal 4x4 block boundaries.  These
   1602   // differ from the 4x4 boundaries on the outside edge of an 8x8 in that the
   1603   // internal ones can be skipped and don't depend on the prediction block size.
   1604   if (tx_size_y == TX_4X4) *int_4x4_y |= size_mask[block_size] << shift_y;
   1605 
   1606   if (build_uv && tx_size_uv == TX_4X4)
   1607     *int_4x4_uv |= (size_mask_uv[block_size] & 0xffff) << shift_uv;
   1608 }
   1609 
   1610 void vp9_loop_filter_data_reset(
   1611     LFWorkerData *lf_data, YV12_BUFFER_CONFIG *frame_buffer,
   1612     struct VP9Common *cm, const struct macroblockd_plane planes[MAX_MB_PLANE]) {
   1613   lf_data->frame_buffer = frame_buffer;
   1614   lf_data->cm = cm;
   1615   lf_data->start = 0;
   1616   lf_data->stop = 0;
   1617   lf_data->y_only = 0;
   1618   memcpy(lf_data->planes, planes, sizeof(lf_data->planes));
   1619 }
   1620 
   1621 void vp9_reset_lfm(VP9_COMMON *const cm) {
   1622   if (cm->lf.filter_level) {
   1623     memset(cm->lf.lfm, 0,
   1624            ((cm->mi_rows + (MI_BLOCK_SIZE - 1)) >> 3) * cm->lf.lfm_stride *
   1625                sizeof(*cm->lf.lfm));
   1626   }
   1627 }
   1628 
   1629 int vp9_loop_filter_worker(void *arg1, void *unused) {
   1630   LFWorkerData *const lf_data = (LFWorkerData *)arg1;
   1631   (void)unused;
   1632   loop_filter_rows(lf_data->frame_buffer, lf_data->cm, lf_data->planes,
   1633                    lf_data->start, lf_data->stop, lf_data->y_only);
   1634   return 1;
   1635 }
   1636