loopfilter.c

/*
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
 *
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
 */

#include <math.h>

#include "./aom_config.h"
#include "./aom_dsp_rtcd.h"
#include "av1/common/loopfilter.h"
#include "av1/common/onyxc_int.h"
#include "av1/common/reconinter.h"
#include "aom_dsp/aom_dsp_common.h"
#include "aom_mem/aom_mem.h"
#include "aom_ports/mem.h"

#include "av1/common/seg_common.h"

// 64 bit masks for left transform size. Each 1 represents a position where
// we should apply a loop filter across the left border of an 8x8 block
// boundary.
//
// In the case of TX_16X16->  ( in low order byte first we end up with
// a mask that looks like this
//
//    10101010
//    10101010
//    10101010
//    10101010
//    10101010
//    10101010
//    10101010
//    10101010
//
// A loopfilter should be applied to every other 8x8 horizontally.
static const uint64_t left_64x64_txform_mask[TX_SIZES] = {
#if CONFIG_CB4X4
  0xffffffffffffffffULL,  // TX_2X2
#endif
  0xffffffffffffffffULL,  // TX_4X4
  0xffffffffffffffffULL,  // TX_8x8
  0x5555555555555555ULL,  // TX_16x16
  0x1111111111111111ULL,  // TX_32x32
#if CONFIG_TX64X64
  0x0101010101010101ULL,  // TX_64x64
#endif                    // CONFIG_TX64X64
};

// 64 bit masks for above transform size. Each 1 represents a position where
// we should apply a loop filter across the top border of an 8x8 block
// boundary.
//
// In the case of TX_32x32 ->  ( in low order byte first we end up with
// a mask that looks like this
//
//    11111111
//    00000000
//    00000000
//    00000000
//    11111111
//    00000000
//    00000000
//    00000000
//
// A loopfilter should be applied to every other 4 the row vertically.
static const uint64_t above_64x64_txform_mask[TX_SIZES] = {
#if CONFIG_CB4X4
  0xffffffffffffffffULL,  // TX_4X4
#endif
  0xffffffffffffffffULL,  // TX_4X4
  0xffffffffffffffffULL,  // TX_8x8
  0x00ff00ff00ff00ffULL,  // TX_16x16
  0x000000ff000000ffULL,  // TX_32x32
#if CONFIG_TX64X64
  0x00000000000000ffULL,  // TX_64x64
#endif                    // CONFIG_TX64X64
};

// 64 bit masks for prediction sizes (left). Each 1 represents a position
// where left border of an 8x8 block. These are aligned to the right most
// appropriate bit, and then shifted into place.
//
// In the case of TX_16x32 ->  ( low order byte first ) we end up with
// a mask that looks like this :
//
//  10000000
//  10000000
//  10000000
//  10000000
//  00000000
//  00000000
//  00000000
//  00000000
static const uint64_t left_prediction_mask[BLOCK_SIZES] = {
#if CONFIG_CB4X4
  0x0000000000000001ULL,  // BLOCK_2X2,
  0x0000000000000001ULL,  // BLOCK_2X4,
  0x0000000000000001ULL,  // BLOCK_4X2,
#endif
  0x0000000000000001ULL,  // BLOCK_4X4,
  0x0000000000000001ULL,  // BLOCK_4X8,
  0x0000000000000001ULL,  // BLOCK_8X4,
  0x0000000000000001ULL,  // BLOCK_8X8,
  0x0000000000000101ULL,  // BLOCK_8X16,
  0x0000000000000001ULL,  // BLOCK_16X8,
  0x0000000000000101ULL,  // BLOCK_16X16,
  0x0000000001010101ULL,  // BLOCK_16X32,
  0x0000000000000101ULL,  // BLOCK_32X16,
  0x0000000001010101ULL,  // BLOCK_32X32,
  0x0101010101010101ULL,  // BLOCK_32X64,
  0x0000000001010101ULL,  // BLOCK_64X32,
  0x0101010101010101ULL,  // BLOCK_64X64
};

// 64 bit mask to shift and set for each prediction size.
static const uint64_t above_prediction_mask[BLOCK_SIZES] = {
#if CONFIG_CB4X4
  0x0000000000000001ULL,  // BLOCK_2X2
  0x0000000000000001ULL,  // BLOCK_2X4
  0x0000000000000001ULL,  // BLOCK_4X2
#endif
  0x0000000000000001ULL,  // BLOCK_4X4
  0x0000000000000001ULL,  // BLOCK_4X8
  0x0000000000000001ULL,  // BLOCK_8X4
  0x0000000000000001ULL,  // BLOCK_8X8
  0x0000000000000001ULL,  // BLOCK_8X16,
  0x0000000000000003ULL,  // BLOCK_16X8
  0x0000000000000003ULL,  // BLOCK_16X16
  0x0000000000000003ULL,  // BLOCK_16X32,
  0x000000000000000fULL,  // BLOCK_32X16,
  0x000000000000000fULL,  // BLOCK_32X32,
  0x000000000000000fULL,  // BLOCK_32X64,
  0x00000000000000ffULL,  // BLOCK_64X32,
  0x00000000000000ffULL,  // BLOCK_64X64
};
// 64 bit mask to shift and set for each prediction size. A bit is set for
// each 8x8 block that would be in the left most block of the given block
// size in the 64x64 block.
static const uint64_t size_mask[BLOCK_SIZES] = {
#if CONFIG_CB4X4
  0x0000000000000001ULL,  // BLOCK_2X2
  0x0000000000000001ULL,  // BLOCK_2X4
  0x0000000000000001ULL,  // BLOCK_4X2
#endif
  0x0000000000000001ULL,  // BLOCK_4X4
  0x0000000000000001ULL,  // BLOCK_4X8
  0x0000000000000001ULL,  // BLOCK_8X4
  0x0000000000000001ULL,  // BLOCK_8X8
  0x0000000000000101ULL,  // BLOCK_8X16,
  0x0000000000000003ULL,  // BLOCK_16X8
  0x0000000000000303ULL,  // BLOCK_16X16
  0x0000000003030303ULL,  // BLOCK_16X32,
  0x0000000000000f0fULL,  // BLOCK_32X16,
  0x000000000f0f0f0fULL,  // BLOCK_32X32,
  0x0f0f0f0f0f0f0f0fULL,  // BLOCK_32X64,
  0x00000000ffffffffULL,  // BLOCK_64X32,
  0xffffffffffffffffULL,  // BLOCK_64X64
};

// These are used for masking the left and above borders.
static const uint64_t left_border = 0x1111111111111111ULL;
static const uint64_t above_border = 0x000000ff000000ffULL;

// 16 bit masks for uv transform sizes.
static const uint16_t left_64x64_txform_mask_uv[TX_SIZES] = {
#if CONFIG_CB4X4
  0xffff,  // TX_2X2
#endif
  0xffff,  // TX_4X4
  0xffff,  // TX_8x8
  0x5555,  // TX_16x16
  0x1111,  // TX_32x32
#if CONFIG_TX64X64
  0x0101,  // TX_64x64, never used
#endif     // CONFIG_TX64X64
};

static const uint16_t above_64x64_txform_mask_uv[TX_SIZES] = {
#if CONFIG_CB4X4
  0xffff,  // TX_2X2
#endif
  0xffff,  // TX_4X4
  0xffff,  // TX_8x8
  0x0f0f,  // TX_16x16
  0x000f,  // TX_32x32
#if CONFIG_TX64X64
  0x0003,  // TX_64x64, never used
#endif     // CONFIG_TX64X64
};

// 16 bit left mask to shift and set for each uv prediction size.
static const uint16_t left_prediction_mask_uv[BLOCK_SIZES] = {
#if CONFIG_CB4X4
  0x0001,  // BLOCK_2X2,
  0x0001,  // BLOCK_2X4,
  0x0001,  // BLOCK_4X2,
#endif
  0x0001,  // BLOCK_4X4,
  0x0001,  // BLOCK_4X8,
  0x0001,  // BLOCK_8X4,
  0x0001,  // BLOCK_8X8,
  0x0001,  // BLOCK_8X16,
  0x0001,  // BLOCK_16X8,
  0x0001,  // BLOCK_16X16,
  0x0011,  // BLOCK_16X32,
  0x0001,  // BLOCK_32X16,
  0x0011,  // BLOCK_32X32,
  0x1111,  // BLOCK_32X64
  0x0011,  // BLOCK_64X32,
  0x1111,  // BLOCK_64X64
};
// 16 bit above mask to shift and set for uv each prediction size.
static const uint16_t above_prediction_mask_uv[BLOCK_SIZES] = {
#if CONFIG_CB4X4
  0x0001,  // BLOCK_2X2
  0x0001,  // BLOCK_2X4
  0x0001,  // BLOCK_4X2
#endif
  0x0001,  // BLOCK_4X4
  0x0001,  // BLOCK_4X8
  0x0001,  // BLOCK_8X4
  0x0001,  // BLOCK_8X8
  0x0001,  // BLOCK_8X16,
  0x0001,  // BLOCK_16X8
  0x0001,  // BLOCK_16X16
  0x0001,  // BLOCK_16X32,
  0x0003,  // BLOCK_32X16,
  0x0003,  // BLOCK_32X32,
  0x0003,  // BLOCK_32X64,
  0x000f,  // BLOCK_64X32,
  0x000f,  // BLOCK_64X64
};

// 64 bit mask to shift and set for each uv prediction size
static const uint16_t size_mask_uv[BLOCK_SIZES] = {
#if CONFIG_CB4X4
  0x0001,  // BLOCK_2X2
  0x0001,  // BLOCK_2X4
  0x0001,  // BLOCK_4X2
#endif
  0x0001,  // BLOCK_4X4
  0x0001,  // BLOCK_4X8
  0x0001,  // BLOCK_8X4
  0x0001,  // BLOCK_8X8
  0x0001,  // BLOCK_8X16,
  0x0001,  // BLOCK_16X8
  0x0001,  // BLOCK_16X16
  0x0011,  // BLOCK_16X32,
  0x0003,  // BLOCK_32X16,
  0x0033,  // BLOCK_32X32,
  0x3333,  // BLOCK_32X64,
  0x00ff,  // BLOCK_64X32,
  0xffff,  // BLOCK_64X64
};
static const uint16_t left_border_uv = 0x1111;
static const uint16_t above_border_uv = 0x000f;

static const int mode_lf_lut[MB_MODE_COUNT] = {
  0, 0, 0, 0, 0, 0, 0, 0, 0, 0,  // INTRA_MODES
  1, 1, 0, 1                     // INTER_MODES (ZEROMV == 0)
#if CONFIG_EXT_INTER
  ,
  1,                            // NEWFROMNEARMV mode
  1, 1, 1, 1, 1, 1, 1, 1, 0, 1  // INTER_COMPOUND_MODES (ZERO_ZEROMV == 0)
#endif                          // CONFIG_EXT_INTER
};

static void update_sharpness(loop_filter_info_n *lfi, int sharpness_lvl) {
  int lvl;

  // For each possible value for the loop filter fill out limits
  for (lvl = 0; lvl <= MAX_LOOP_FILTER; lvl++) {
    // Set loop filter parameters that control sharpness.
    int block_inside_limit = lvl >> ((sharpness_lvl > 0) + (sharpness_lvl > 4));

    if (sharpness_lvl > 0) {
      if (block_inside_limit > (9 - sharpness_lvl))
        block_inside_limit = (9 - sharpness_lvl);
    }

    if (block_inside_limit < 1) block_inside_limit = 1;

    memset(lfi->lfthr[lvl].lim, block_inside_limit, SIMD_WIDTH);
    memset(lfi->lfthr[lvl].mblim, (2 * (lvl + 2) + block_inside_limit),
           SIMD_WIDTH);
  }
}

static uint8_t get_filter_level(const loop_filter_info_n *lfi_n,
                                const MB_MODE_INFO *mbmi) {
#if CONFIG_SUPERTX
  const int segment_id = AOMMIN(mbmi->segment_id, mbmi->segment_id_supertx);
  assert(
      IMPLIES(supertx_enabled(mbmi), mbmi->segment_id_supertx != MAX_SEGMENTS));
  assert(IMPLIES(supertx_enabled(mbmi),
                 mbmi->segment_id_supertx <= mbmi->segment_id));
#else
  const int segment_id = mbmi->segment_id;
#endif  // CONFIG_SUPERTX
  return lfi_n->lvl[segment_id][mbmi->ref_frame[0]][mode_lf_lut[mbmi->mode]];
}

void av1_loop_filter_init(AV1_COMMON *cm) {
  loop_filter_info_n *lfi = &cm->lf_info;
  struct loopfilter *lf = &cm->lf;
  int lvl;

  // init limits for given sharpness
  update_sharpness(lfi, lf->sharpness_level);
  lf->last_sharpness_level = lf->sharpness_level;

  // init hev threshold const vectors
  for (lvl = 0; lvl <= MAX_LOOP_FILTER; lvl++)
    memset(lfi->lfthr[lvl].hev_thr, (lvl >> 4), SIMD_WIDTH);
}

void av1_loop_filter_frame_init(AV1_COMMON *cm, int default_filt_lvl) {
  int seg_id;
  // n_shift is the multiplier for lf_deltas
  // the multiplier is 1 for when filter_lvl is between 0 and 31;
  // 2 when filter_lvl is between 32 and 63
  const int scale = 1 << (default_filt_lvl >> 5);
  loop_filter_info_n *const lfi = &cm->lf_info;
  struct loopfilter *const lf = &cm->lf;
  const struct segmentation *const seg = &cm->seg;

  // update limits if sharpness has changed
  if (lf->last_sharpness_level != lf->sharpness_level) {
    update_sharpness(lfi, lf->sharpness_level);
    lf->last_sharpness_level = lf->sharpness_level;
  }

  for (seg_id = 0; seg_id < MAX_SEGMENTS; seg_id++) {
    int lvl_seg = default_filt_lvl;
    if (segfeature_active(seg, seg_id, SEG_LVL_ALT_LF)) {
      const int data = get_segdata(seg, seg_id, SEG_LVL_ALT_LF);
      lvl_seg = clamp(
          seg->abs_delta == SEGMENT_ABSDATA ? data : default_filt_lvl + data, 0,
          MAX_LOOP_FILTER);
    }

    if (!lf->mode_ref_delta_enabled) {
      // we could get rid of this if we assume that deltas are set to
      // zero when not in use; encoder always uses deltas
      memset(lfi->lvl[seg_id], lvl_seg, sizeof(lfi->lvl[seg_id]));
    } else {
      int ref, mode;
      const int intra_lvl = lvl_seg + lf->ref_deltas[INTRA_FRAME] * scale;
      lfi->lvl[seg_id][INTRA_FRAME][0] = clamp(intra_lvl, 0, MAX_LOOP_FILTER);

      for (ref = LAST_FRAME; ref < TOTAL_REFS_PER_FRAME; ++ref) {
        for (mode = 0; mode < MAX_MODE_LF_DELTAS; ++mode) {
          const int inter_lvl = lvl_seg + lf->ref_deltas[ref] * scale +
                                lf->mode_deltas[mode] * scale;
          lfi->lvl[seg_id][ref][mode] = clamp(inter_lvl, 0, MAX_LOOP_FILTER);
        }
      }
    }
  }
}

static void filter_selectively_vert_row2(int subsampling_factor, uint8_t *s,
                                         int pitch, unsigned int mask_16x16_l,
                                         unsigned int mask_8x8_l,
                                         unsigned int mask_4x4_l,
                                         unsigned int mask_4x4_int_l,
                                         const loop_filter_info_n *lfi_n,
                                         const uint8_t *lfl) {
  const int mask_shift = subsampling_factor ? 4 : 8;
  const int mask_cutoff = subsampling_factor ? 0xf : 0xff;
  const int lfl_forward = subsampling_factor ? 4 : 8;

  unsigned int mask_16x16_0 = mask_16x16_l & mask_cutoff;
  unsigned int mask_8x8_0 = mask_8x8_l & mask_cutoff;
  unsigned int mask_4x4_0 = mask_4x4_l & mask_cutoff;
  unsigned int mask_4x4_int_0 = mask_4x4_int_l & mask_cutoff;
  unsigned int mask_16x16_1 = (mask_16x16_l >> mask_shift) & mask_cutoff;
  unsigned int mask_8x8_1 = (mask_8x8_l >> mask_shift) & mask_cutoff;
  unsigned int mask_4x4_1 = (mask_4x4_l >> mask_shift) & mask_cutoff;
  unsigned int mask_4x4_int_1 = (mask_4x4_int_l >> mask_shift) & mask_cutoff;
  unsigned int mask;

  for (mask = mask_16x16_0 | mask_8x8_0 | mask_4x4_0 | mask_4x4_int_0 |
              mask_16x16_1 | mask_8x8_1 | mask_4x4_1 | mask_4x4_int_1;
       mask; mask >>= 1) {
    const loop_filter_thresh *lfi0 = lfi_n->lfthr + *lfl;
    const loop_filter_thresh *lfi1 = lfi_n->lfthr + *(lfl + lfl_forward);

    if (mask & 1) {
      if ((mask_16x16_0 | mask_16x16_1) & 1) {
        if ((mask_16x16_0 & mask_16x16_1) & 1) {
          aom_lpf_vertical_16_dual(s, pitch, lfi0->mblim, lfi0->lim,
                                   lfi0->hev_thr);
        } else if (mask_16x16_0 & 1) {
          aom_lpf_vertical_16(s, pitch, lfi0->mblim, lfi0->lim, lfi0->hev_thr);
        } else {
          aom_lpf_vertical_16(s + 8 * pitch, pitch, lfi1->mblim, lfi1->lim,
                              lfi1->hev_thr);
        }
      }

      if ((mask_8x8_0 | mask_8x8_1) & 1) {
        if ((mask_8x8_0 & mask_8x8_1) & 1) {
          aom_lpf_vertical_8_dual(s, pitch, lfi0->mblim, lfi0->lim,
                                  lfi0->hev_thr, lfi1->mblim, lfi1->lim,
                                  lfi1->hev_thr);
        } else if (mask_8x8_0 & 1) {
          aom_lpf_vertical_8(s, pitch, lfi0->mblim, lfi0->lim, lfi0->hev_thr);
        } else {
          aom_lpf_vertical_8(s + 8 * pitch, pitch, lfi1->mblim, lfi1->lim,
                             lfi1->hev_thr);
        }
      }

      if ((mask_4x4_0 | mask_4x4_1) & 1) {
        if ((mask_4x4_0 & mask_4x4_1) & 1) {
          aom_lpf_vertical_4_dual(s, pitch, lfi0->mblim, lfi0->lim,
                                  lfi0->hev_thr, lfi1->mblim, lfi1->lim,
                                  lfi1->hev_thr);
        } else if (mask_4x4_0 & 1) {
          aom_lpf_vertical_4(s, pitch, lfi0->mblim, lfi0->lim, lfi0->hev_thr);
        } else {
          aom_lpf_vertical_4(s + 8 * pitch, pitch, lfi1->mblim, lfi1->lim,
                             lfi1->hev_thr);
        }
      }

      if ((mask_4x4_int_0 | mask_4x4_int_1) & 1) {
        if ((mask_4x4_int_0 & mask_4x4_int_1) & 1) {
          aom_lpf_vertical_4_dual(s + 4, pitch, lfi0->mblim, lfi0->lim,
                                  lfi0->hev_thr, lfi1->mblim, lfi1->lim,
                                  lfi1->hev_thr);
        } else if (mask_4x4_int_0 & 1) {
          aom_lpf_vertical_4(s + 4, pitch, lfi0->mblim, lfi0->lim,
                             lfi0->hev_thr);
        } else {
          aom_lpf_vertical_4(s + 8 * pitch + 4, pitch, lfi1->mblim, lfi1->lim,
                             lfi1->hev_thr);
        }
      }
    }

    s += 8;
    lfl += 1;
    mask_16x16_0 >>= 1;
    mask_8x8_0 >>= 1;
    mask_4x4_0 >>= 1;
    mask_4x4_int_0 >>= 1;
    mask_16x16_1 >>= 1;
    mask_8x8_1 >>= 1;
    mask_4x4_1 >>= 1;
    mask_4x4_int_1 >>= 1;
  }
}

#if CONFIG_AOM_HIGHBITDEPTH
static void highbd_filter_selectively_vert_row2(
    int subsampling_factor, uint16_t *s, int pitch, unsigned int mask_16x16_l,
    unsigned int mask_8x8_l, unsigned int mask_4x4_l,
    unsigned int mask_4x4_int_l, const loop_filter_info_n *lfi_n,
    const uint8_t *lfl, int bd) {
  const int mask_shift = subsampling_factor ? 4 : 8;
  const int mask_cutoff = subsampling_factor ? 0xf : 0xff;
  const int lfl_forward = subsampling_factor ? 4 : 8;

  unsigned int mask_16x16_0 = mask_16x16_l & mask_cutoff;
  unsigned int mask_8x8_0 = mask_8x8_l & mask_cutoff;
  unsigned int mask_4x4_0 = mask_4x4_l & mask_cutoff;
  unsigned int mask_4x4_int_0 = mask_4x4_int_l & mask_cutoff;
  unsigned int mask_16x16_1 = (mask_16x16_l >> mask_shift) & mask_cutoff;
  unsigned int mask_8x8_1 = (mask_8x8_l >> mask_shift) & mask_cutoff;
  unsigned int mask_4x4_1 = (mask_4x4_l >> mask_shift) & mask_cutoff;
  unsigned int mask_4x4_int_1 = (mask_4x4_int_l >> mask_shift) & mask_cutoff;
  unsigned int mask;

  for (mask = mask_16x16_0 | mask_8x8_0 | mask_4x4_0 | mask_4x4_int_0 |
              mask_16x16_1 | mask_8x8_1 | mask_4x4_1 | mask_4x4_int_1;
       mask; mask >>= 1) {
    const loop_filter_thresh *lfi0 = lfi_n->lfthr + *lfl;
    const loop_filter_thresh *lfi1 = lfi_n->lfthr + *(lfl + lfl_forward);

    if (mask & 1) {
      if ((mask_16x16_0 | mask_16x16_1) & 1) {
        if ((mask_16x16_0 & mask_16x16_1) & 1) {
          aom_highbd_lpf_vertical_16_dual(s, pitch, lfi0->mblim, lfi0->lim,
                                          lfi0->hev_thr, bd);
        } else if (mask_16x16_0 & 1) {
          aom_highbd_lpf_vertical_16(s, pitch, lfi0->mblim, lfi0->lim,
                                     lfi0->hev_thr, bd);
        } else {
          aom_highbd_lpf_vertical_16(s + 8 * pitch, pitch, lfi1->mblim,
                                     lfi1->lim, lfi1->hev_thr, bd);
        }
      }

      if ((mask_8x8_0 | mask_8x8_1) & 1) {
        if ((mask_8x8_0 & mask_8x8_1) & 1) {
          aom_highbd_lpf_vertical_8_dual(s, pitch, lfi0->mblim, lfi0->lim,
                                         lfi0->hev_thr, lfi1->mblim, lfi1->lim,
                                         lfi1->hev_thr, bd);
        } else if (mask_8x8_0 & 1) {
          aom_highbd_lpf_vertical_8(s, pitch, lfi0->mblim, lfi0->lim,
                                    lfi0->hev_thr, bd);
        } else {
          aom_highbd_lpf_vertical_8(s + 8 * pitch, pitch, lfi1->mblim,
                                    lfi1->lim, lfi1->hev_thr, bd);
        }
      }

      if ((mask_4x4_0 | mask_4x4_1) & 1) {
        if ((mask_4x4_0 & mask_4x4_1) & 1) {
          aom_highbd_lpf_vertical_4_dual(s, pitch, lfi0->mblim, lfi0->lim,
                                         lfi0->hev_thr, lfi1->mblim, lfi1->lim,
                                         lfi1->hev_thr, bd);
        } else if (mask_4x4_0 & 1) {
          aom_highbd_lpf_vertical_4(s, pitch, lfi0->mblim, lfi0->lim,
                                    lfi0->hev_thr, bd);
        } else {
          aom_highbd_lpf_vertical_4(s + 8 * pitch, pitch, lfi1->mblim,
                                    lfi1->lim, lfi1->hev_thr, bd);
        }
      }

      if ((mask_4x4_int_0 | mask_4x4_int_1) & 1) {
        if ((mask_4x4_int_0 & mask_4x4_int_1) & 1) {
          aom_highbd_lpf_vertical_4_dual(s + 4, pitch, lfi0->mblim, lfi0->lim,
                                         lfi0->hev_thr, lfi1->mblim, lfi1->lim,
                                         lfi1->hev_thr, bd);
        } else if (mask_4x4_int_0 & 1) {
          aom_highbd_lpf_vertical_4(s + 4, pitch, lfi0->mblim, lfi0->lim,
                                    lfi0->hev_thr, bd);
        } else {
          aom_highbd_lpf_vertical_4(s + 8 * pitch + 4, pitch, lfi1->mblim,
                                    lfi1->lim, lfi1->hev_thr, bd);
        }
      }
    }

    s += 8;
    lfl += 1;
    mask_16x16_0 >>= 1;
    mask_8x8_0 >>= 1;
    mask_4x4_0 >>= 1;
    mask_4x4_int_0 >>= 1;
    mask_16x16_1 >>= 1;
    mask_8x8_1 >>= 1;
    mask_4x4_1 >>= 1;
    mask_4x4_int_1 >>= 1;
  }
}
#endif  // CONFIG_AOM_HIGHBITDEPTH

static void filter_selectively_horiz(
    uint8_t *s, int pitch, unsigned int mask_16x16, unsigned int mask_8x8,
    unsigned int mask_4x4, unsigned int mask_4x4_int,
    const loop_filter_info_n *lfi_n, const uint8_t *lfl) {
  unsigned int mask;
  int count;

  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int; mask;
       mask >>= count) {
    const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;

    count = 1;
    if (mask & 1) {
      if (mask_16x16 & 1) {
        if ((mask_16x16 & 3) == 3) {
          aom_lpf_horizontal_edge_16(s, pitch, lfi->mblim, lfi->lim,
                                     lfi->hev_thr);
          count = 2;
        } else {
          aom_lpf_horizontal_edge_8(s, pitch, lfi->mblim, lfi->lim,
                                    lfi->hev_thr);
        }
      } else if (mask_8x8 & 1) {
        if ((mask_8x8 & 3) == 3) {
          // Next block's thresholds.
          const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);

          aom_lpf_horizontal_8_dual(s, pitch, lfi->mblim, lfi->lim,
                                    lfi->hev_thr, lfin->mblim, lfin->lim,
                                    lfin->hev_thr);

          if ((mask_4x4_int & 3) == 3) {
            aom_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
                                      lfi->lim, lfi->hev_thr, lfin->mblim,
                                      lfin->lim, lfin->hev_thr);
          } else {
            if (mask_4x4_int & 1)
              aom_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                                   lfi->hev_thr);
            else if (mask_4x4_int & 2)
              aom_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
                                   lfin->lim, lfin->hev_thr);
          }
          count = 2;
        } else {
          aom_lpf_horizontal_8(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);

          if (mask_4x4_int & 1)
            aom_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                                 lfi->hev_thr);
        }
      } else if (mask_4x4 & 1) {
        if ((mask_4x4 & 3) == 3) {
          // Next block's thresholds.
          const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);

          aom_lpf_horizontal_4_dual(s, pitch, lfi->mblim, lfi->lim,
                                    lfi->hev_thr, lfin->mblim, lfin->lim,
                                    lfin->hev_thr);
          if ((mask_4x4_int & 3) == 3) {
            aom_lpf_horizontal_4_dual(s + 4 * pitch, pitch, lfi->mblim,
                                      lfi->lim, lfi->hev_thr, lfin->mblim,
                                      lfin->lim, lfin->hev_thr);
          } else {
            if (mask_4x4_int & 1)
              aom_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                                   lfi->hev_thr);
            else if (mask_4x4_int & 2)
              aom_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
                                   lfin->lim, lfin->hev_thr);
          }
          count = 2;
        } else {
          aom_lpf_horizontal_4(s, pitch, lfi->mblim, lfi->lim, lfi->hev_thr);

          if (mask_4x4_int & 1)
            aom_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                                 lfi->hev_thr);
        }
      } else if (mask_4x4_int & 1) {
        aom_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                             lfi->hev_thr);
      }
    }
    s += 8 * count;
    lfl += count;
    mask_16x16 >>= count;
    mask_8x8 >>= count;
    mask_4x4 >>= count;
    mask_4x4_int >>= count;
  }
}

#if CONFIG_AOM_HIGHBITDEPTH
static void highbd_filter_selectively_horiz(
    uint16_t *s, int pitch, unsigned int mask_16x16, unsigned int mask_8x8,
    unsigned int mask_4x4, unsigned int mask_4x4_int,
    const loop_filter_info_n *lfi_n, const uint8_t *lfl, int bd) {
  unsigned int mask;
  int count;

  for (mask = mask_16x16 | mask_8x8 | mask_4x4 | mask_4x4_int; mask;
       mask >>= count) {
    const loop_filter_thresh *lfi = lfi_n->lfthr + *lfl;

    count = 1;
    if (mask & 1) {
      if (mask_16x16 & 1) {
        if ((mask_16x16 & 3) == 3) {
          aom_highbd_lpf_horizontal_edge_16(s, pitch, lfi->mblim, lfi->lim,
                                            lfi->hev_thr, bd);
          count = 2;
        } else {
          aom_highbd_lpf_horizontal_edge_8(s, pitch, lfi->mblim, lfi->lim,
                                           lfi->hev_thr, bd);
        }
      } else if (mask_8x8 & 1) {
        if ((mask_8x8 & 3) == 3) {
          // Next block's thresholds.
          const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);

          aom_highbd_lpf_horizontal_8_dual(s, pitch, lfi->mblim, lfi->lim,
                                           lfi->hev_thr, lfin->mblim, lfin->lim,
                                           lfin->hev_thr, bd);

          if ((mask_4x4_int & 3) == 3) {
            aom_highbd_lpf_horizontal_4_dual(
                s + 4 * pitch, pitch, lfi->mblim, lfi->lim, lfi->hev_thr,
                lfin->mblim, lfin->lim, lfin->hev_thr, bd);
          } else {
            if (mask_4x4_int & 1) {
              aom_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
                                          lfi->lim, lfi->hev_thr, bd);
            } else if (mask_4x4_int & 2) {
              aom_highbd_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
                                          lfin->lim, lfin->hev_thr, bd);
            }
          }
          count = 2;
        } else {
          aom_highbd_lpf_horizontal_8(s, pitch, lfi->mblim, lfi->lim,
                                      lfi->hev_thr, bd);

          if (mask_4x4_int & 1) {
            aom_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
                                        lfi->lim, lfi->hev_thr, bd);
          }
        }
      } else if (mask_4x4 & 1) {
        if ((mask_4x4 & 3) == 3) {
          // Next block's thresholds.
          const loop_filter_thresh *lfin = lfi_n->lfthr + *(lfl + 1);

          aom_highbd_lpf_horizontal_4_dual(s, pitch, lfi->mblim, lfi->lim,
                                           lfi->hev_thr, lfin->mblim, lfin->lim,
                                           lfin->hev_thr, bd);
          if ((mask_4x4_int & 3) == 3) {
            aom_highbd_lpf_horizontal_4_dual(
                s + 4 * pitch, pitch, lfi->mblim, lfi->lim, lfi->hev_thr,
                lfin->mblim, lfin->lim, lfin->hev_thr, bd);
          } else {
            if (mask_4x4_int & 1) {
              aom_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
                                          lfi->lim, lfi->hev_thr, bd);
            } else if (mask_4x4_int & 2) {
              aom_highbd_lpf_horizontal_4(s + 8 + 4 * pitch, pitch, lfin->mblim,
                                          lfin->lim, lfin->hev_thr, bd);
            }
          }
          count = 2;
        } else {
          aom_highbd_lpf_horizontal_4(s, pitch, lfi->mblim, lfi->lim,
                                      lfi->hev_thr, bd);

          if (mask_4x4_int & 1) {
            aom_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim,
                                        lfi->lim, lfi->hev_thr, bd);
          }
        }
      } else if (mask_4x4_int & 1) {
        aom_highbd_lpf_horizontal_4(s + 4 * pitch, pitch, lfi->mblim, lfi->lim,
                                    lfi->hev_thr, bd);
      }
    }
    s += 8 * count;
    lfl += count;
    mask_16x16 >>= count;
    mask_8x8 >>= count;
    mask_4x4 >>= count;
    mask_4x4_int >>= count;
  }
}
#endif  // CONFIG_AOM_HIGHBITDEPTH

// This function ors into the current lfm structure, where to do loop
// filters for the specific mi we are looking at. It uses information
// including the block_size_type (32x16, 32x32, etc.), the transform size,
// whether there were any coefficients encoded, and the loop filter strength
// block we are currently looking at. Shift is used to position the
// 1's we produce.
// TODO(JBB) Need another function for different resolution color..
static void build_masks(const loop_filter_info_n *const lfi_n,
                        const MODE_INFO *mi, const int shift_y,
                        const int shift_uv, LOOP_FILTER_MASK *lfm) {
  const MB_MODE_INFO *mbmi = &mi->mbmi;
  const BLOCK_SIZE block_size = mbmi->sb_type;
  // TODO(debargha): Check if masks can be setup correctly when
  // rectangular transfroms are used with the EXT_TX expt.
  const TX_SIZE tx_size_y = txsize_sqr_up_map[mbmi->tx_size];
  const TX_SIZE tx_size_uv =
      txsize_sqr_up_map[uv_txsize_lookup[block_size][mbmi->tx_size][1][1]];
  const int filter_level = get_filter_level(lfi_n, mbmi);
  uint64_t *const left_y = &lfm->left_y[tx_size_y];
  uint64_t *const above_y = &lfm->above_y[tx_size_y];
  uint64_t *const int_4x4_y = &lfm->int_4x4_y;
  uint16_t *const left_uv = &lfm->left_uv[tx_size_uv];
  uint16_t *const above_uv = &lfm->above_uv[tx_size_uv];
  uint16_t *const int_4x4_uv = &lfm->left_int_4x4_uv;
  int i;

  // If filter level is 0 we don't loop filter.
  if (!filter_level) {
    return;
  } else {
    const int w = num_8x8_blocks_wide_lookup[block_size];
    const int h = num_8x8_blocks_high_lookup[block_size];
    const int row = (shift_y >> MAX_MIB_SIZE_LOG2);
    const int col = shift_y - (row << MAX_MIB_SIZE_LOG2);

    for (i = 0; i < h; i++) memset(&lfm->lfl_y[row + i][col], filter_level, w);
  }

  // These set 1 in the current block size for the block size edges.
  // For instance if the block size is 32x16, we'll set:
  //    above =   1111
  //              0000
  //    and
  //    left  =   1000
  //          =   1000
  // NOTE : In this example the low bit is left most ( 1000 ) is stored as
  //        1,  not 8...
  //
  // U and V set things on a 16 bit scale.
  //
  *above_y |= above_prediction_mask[block_size] << shift_y;
  *above_uv |= above_prediction_mask_uv[block_size] << shift_uv;
  *left_y |= left_prediction_mask[block_size] << shift_y;
  *left_uv |= left_prediction_mask_uv[block_size] << shift_uv;

  // If the block has no coefficients and is not intra we skip applying
  // the loop filter on block edges.
  if (mbmi->skip && is_inter_block(mbmi)) return;

  // Here we are adding a mask for the transform size. The transform
  // size mask is set to be correct for a 64x64 prediction block size. We
  // mask to match the size of the block we are working on and then shift it
  // into place..
  *above_y |= (size_mask[block_size] & above_64x64_txform_mask[tx_size_y])
              << shift_y;
  *above_uv |=
      (size_mask_uv[block_size] & above_64x64_txform_mask_uv[tx_size_uv])
      << shift_uv;

  *left_y |= (size_mask[block_size] & left_64x64_txform_mask[tx_size_y])
             << shift_y;
  *left_uv |= (size_mask_uv[block_size] & left_64x64_txform_mask_uv[tx_size_uv])
              << shift_uv;

  // Here we are trying to determine what to do with the internal 4x4 block
  // boundaries.  These differ from the 4x4 boundaries on the outside edge of
  // an 8x8 in that the internal ones can be skipped and don't depend on
  // the prediction block size.
  if (tx_size_y == TX_4X4)
    *int_4x4_y |= (size_mask[block_size] & 0xffffffffffffffffULL) << shift_y;

  if (tx_size_uv == TX_4X4)
    *int_4x4_uv |= (size_mask_uv[block_size] & 0xffff) << shift_uv;
}

// This function does the same thing as the one above with the exception that
// it only affects the y masks. It exists because for blocks < 16x16 in size,
// we only update u and v masks on the first block.
static void build_y_mask(const loop_filter_info_n *const lfi_n,
                         const MODE_INFO *mi, const int shift_y,
#if CONFIG_SUPERTX
                         int supertx_enabled,
#endif  // CONFIG_SUPERTX
                         LOOP_FILTER_MASK *lfm) {
  const MB_MODE_INFO *mbmi = &mi->mbmi;
  const TX_SIZE tx_size_y = txsize_sqr_up_map[mbmi->tx_size];
#if CONFIG_SUPERTX
  const BLOCK_SIZE block_size =
      supertx_enabled ? (BLOCK_SIZE)(3 * tx_size_y) : mbmi->sb_type;
#else
  const BLOCK_SIZE block_size = mbmi->sb_type;
#endif
  const int filter_level = get_filter_level(lfi_n, mbmi);
  uint64_t *const left_y = &lfm->left_y[tx_size_y];
  uint64_t *const above_y = &lfm->above_y[tx_size_y];
  uint64_t *const int_4x4_y = &lfm->int_4x4_y;
  int i;

  if (!filter_level) {
    return;
  } else {
    const int w = num_8x8_blocks_wide_lookup[block_size];
    const int h = num_8x8_blocks_high_lookup[block_size];
    const int row = (shift_y >> MAX_MIB_SIZE_LOG2);
    const int col = shift_y - (row << MAX_MIB_SIZE_LOG2);

    for (i = 0; i < h; i++) memset(&lfm->lfl_y[row + i][col], filter_level, w);
  }

  *above_y |= above_prediction_mask[block_size] << shift_y;
  *left_y |= left_prediction_mask[block_size] << shift_y;

  if (mbmi->skip && is_inter_block(mbmi)) return;

  *above_y |= (size_mask[block_size] & above_64x64_txform_mask[tx_size_y])
              << shift_y;

  *left_y |= (size_mask[block_size] & left_64x64_txform_mask[tx_size_y])
             << shift_y;

  if (tx_size_y == TX_4X4)
    *int_4x4_y |= (size_mask[block_size] & 0xffffffffffffffffULL) << shift_y;
}

// This function sets up the bit masks for the entire 64x64 region represented
// by mi_row, mi_col.
// TODO(JBB): This function only works for yv12.
void av1_setup_mask(AV1_COMMON *const cm, const int mi_row, const int mi_col,
                    MODE_INFO **mi, const int mode_info_stride,
                    LOOP_FILTER_MASK *lfm) {
  int idx_32, idx_16, idx_8;
  const loop_filter_info_n *const lfi_n = &cm->lf_info;
  MODE_INFO **mip = mi;
  MODE_INFO **mip2 = mi;

  // These are offsets to the next mi in the 64x64 block. It is what gets
  // added to the mi ptr as we go through each loop. It helps us to avoid
  // setting up special row and column counters for each index. The last step
  // brings us out back to the starting position.
  const int offset_32[] = { 4, (mode_info_stride << 2) - 4, 4,
                            -(mode_info_stride << 2) - 4 };
  const int offset_16[] = { 2, (mode_info_stride << 1) - 2, 2,
                            -(mode_info_stride << 1) - 2 };
  const int offset[] = { 1, mode_info_stride - 1, 1, -mode_info_stride - 1 };

  // Following variables represent shifts to position the current block
  // mask over the appropriate block. A shift of 36 to the left will move
  // the bits for the final 32 by 32 block in the 64x64 up 4 rows and left
  // 4 rows to the appropriate spot.
  const int shift_32_y[] = { 0, 4, 32, 36 };
  const int shift_16_y[] = { 0, 2, 16, 18 };
  const int shift_8_y[] = { 0, 1, 8, 9 };
  const int shift_32_uv[] = { 0, 2, 8, 10 };
  const int shift_16_uv[] = { 0, 1, 4, 5 };
  int i;
  const int max_rows = AOMMIN(cm->mi_rows - mi_row, MAX_MIB_SIZE);
  const int max_cols = AOMMIN(cm->mi_cols - mi_col, MAX_MIB_SIZE);
#if CONFIG_EXT_PARTITION
  assert(0 && "Not yet updated");
#endif  // CONFIG_EXT_PARTITION

  av1_zero(*lfm);
  assert(mip[0] != NULL);

  // TODO(jimbankoski): Try moving most of the following code into decode
  // loop and storing lfm in the mbmi structure so that we don't have to go
  // through the recursive loop structure multiple times.
  switch (mip[0]->mbmi.sb_type) {
    case BLOCK_64X64: build_masks(lfi_n, mip[0], 0, 0, lfm); break;
    case BLOCK_64X32: build_masks(lfi_n, mip[0], 0, 0, lfm);
#if CONFIG_SUPERTX && CONFIG_TX64X64
      if (supertx_enabled(&mip[0]->mbmi)) break;
#endif  // CONFIG_SUPERTX && CONFIG_TX64X64
      mip2 = mip + mode_info_stride * 4;
      if (4 >= max_rows) break;
      build_masks(lfi_n, mip2[0], 32, 8, lfm);
      break;
    case BLOCK_32X64: build_masks(lfi_n, mip[0], 0, 0, lfm);
#if CONFIG_SUPERTX && CONFIG_TX64X64
      if (supertx_enabled(&mip[0]->mbmi)) break;
#endif  // CONFIG_SUPERTX && CONFIG_TX64X64
      mip2 = mip + 4;
      if (4 >= max_cols) break;
      build_masks(lfi_n, mip2[0], 4, 2, lfm);
      break;
    default:
#if CONFIG_SUPERTX && CONFIG_TX64X64
      if (mip[0]->mbmi.tx_size == TX_64X64) {
        build_masks(lfi_n, mip[0], 0, 0, lfm);
      } else {
#endif  // CONFIG_SUPERTX && CONFIG_TX64X64
        for (idx_32 = 0; idx_32 < 4; mip += offset_32[idx_32], ++idx_32) {
          const int shift_y_32 = shift_32_y[idx_32];
          const int shift_uv_32 = shift_32_uv[idx_32];
          const int mi_32_col_offset = ((idx_32 & 1) << 2);
          const int mi_32_row_offset = ((idx_32 >> 1) << 2);
          if (mi_32_col_offset >= max_cols || mi_32_row_offset >= max_rows)
            continue;
          switch (mip[0]->mbmi.sb_type) {
            case BLOCK_32X32:
              build_masks(lfi_n, mip[0], shift_y_32, shift_uv_32, lfm);
              break;
            case BLOCK_32X16:
              build_masks(lfi_n, mip[0], shift_y_32, shift_uv_32, lfm);
#if CONFIG_SUPERTX
              if (supertx_enabled(&mip[0]->mbmi)) break;
#endif
              if (mi_32_row_offset + 2 >= max_rows) continue;
              mip2 = mip + mode_info_stride * 2;
              build_masks(lfi_n, mip2[0], shift_y_32 + 16, shift_uv_32 + 4,
                          lfm);
              break;
            case BLOCK_16X32:
              build_masks(lfi_n, mip[0], shift_y_32, shift_uv_32, lfm);
#if CONFIG_SUPERTX
              if (supertx_enabled(&mip[0]->mbmi)) break;
#endif
              if (mi_32_col_offset + 2 >= max_cols) continue;
              mip2 = mip + 2;
              build_masks(lfi_n, mip2[0], shift_y_32 + 2, shift_uv_32 + 1, lfm);
              break;
            default:
#if CONFIG_SUPERTX
              if (mip[0]->mbmi.tx_size == TX_32X32) {
                build_masks(lfi_n, mip[0], shift_y_32, shift_uv_32, lfm);
                break;
              }
#endif
              for (idx_16 = 0; idx_16 < 4; mip += offset_16[idx_16], ++idx_16) {
                const int shift_y_32_16 = shift_y_32 + shift_16_y[idx_16];
                const int shift_uv_32_16 = shift_uv_32 + shift_16_uv[idx_16];
                const int mi_16_col_offset =
                    mi_32_col_offset + ((idx_16 & 1) << 1);
                const int mi_16_row_offset =
                    mi_32_row_offset + ((idx_16 >> 1) << 1);

                if (mi_16_col_offset >= max_cols ||
                    mi_16_row_offset >= max_rows)