reconinter.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 <assert.h>

#include "./aom_scale_rtcd.h"
#include "./aom_dsp_rtcd.h"
#include "./aom_config.h"

#include "aom/aom_integer.h"
#include "aom_dsp/blend.h"

#include "av1/common/blockd.h"
#include "av1/common/reconinter.h"
#include "av1/common/reconintra.h"
#if CONFIG_MOTION_VAR
#include "av1/common/onyxc_int.h"
#endif  // CONFIG_MOTION_VAR
#if CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION
#include "av1/common/warped_motion.h"
#endif  // CONFIG_GLOBAL_MOTION || CONFIG_WARPED_MOTION

#if CONFIG_EXT_INTER

#define NSMOOTHERS 1
static int get_masked_weight(int m, int smoothness) {
#define SMOOTHER_LEN 32
  static const uint8_t smoothfn[NSMOOTHERS][2 * SMOOTHER_LEN + 1] = { {
      0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,  0,
      0,  0,  0,  0,  0,  0,  0,  0,  0,  1,  2,  4,  7,  13, 21, 32, 43,
      51, 57, 60, 62, 63, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
      64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
  } };
  if (m < -SMOOTHER_LEN)
    return 0;
  else if (m > SMOOTHER_LEN)
    return (1 << WEDGE_WEIGHT_BITS);
  else
    return smoothfn[smoothness][m + SMOOTHER_LEN];
}

// [smoother][negative][direction]
DECLARE_ALIGNED(16, static uint8_t,
                wedge_mask_obl[NSMOOTHERS][2][WEDGE_DIRECTIONS]
                              [MASK_MASTER_SIZE * MASK_MASTER_SIZE]);

DECLARE_ALIGNED(16, static uint8_t,
                wedge_signflip_lookup[BLOCK_SIZES][MAX_WEDGE_TYPES]);

// 3 * MAX_WEDGE_SQUARE is an easy to compute and fairly tight upper bound
// on the sum of all mask sizes up to an including MAX_WEDGE_SQUARE.
DECLARE_ALIGNED(16, static uint8_t,
                wedge_mask_buf[2 * MAX_WEDGE_TYPES * 3 * MAX_WEDGE_SQUARE]);

static wedge_masks_type wedge_masks[BLOCK_SIZES][2];

// Some unused wedge codebooks left temporarily to facilitate experiments.
// To be removed when setteld.
static wedge_code_type wedge_codebook_8_hgtw[8] = {
  { WEDGE_OBLIQUE27, 4, 4 },  { WEDGE_OBLIQUE63, 4, 4 },
  { WEDGE_OBLIQUE117, 4, 4 }, { WEDGE_OBLIQUE153, 4, 4 },
  { WEDGE_OBLIQUE27, 4, 2 },  { WEDGE_OBLIQUE27, 4, 6 },
  { WEDGE_OBLIQUE153, 4, 2 }, { WEDGE_OBLIQUE153, 4, 6 },
};

static wedge_code_type wedge_codebook_8_hltw[8] = {
  { WEDGE_OBLIQUE27, 4, 4 },  { WEDGE_OBLIQUE63, 4, 4 },
  { WEDGE_OBLIQUE117, 4, 4 }, { WEDGE_OBLIQUE153, 4, 4 },
  { WEDGE_OBLIQUE63, 2, 4 },  { WEDGE_OBLIQUE63, 6, 4 },
  { WEDGE_OBLIQUE117, 2, 4 }, { WEDGE_OBLIQUE117, 6, 4 },
};

static wedge_code_type wedge_codebook_8_heqw[8] = {
  { WEDGE_OBLIQUE27, 4, 4 },  { WEDGE_OBLIQUE63, 4, 4 },
  { WEDGE_OBLIQUE117, 4, 4 }, { WEDGE_OBLIQUE153, 4, 4 },
  { WEDGE_HORIZONTAL, 4, 2 }, { WEDGE_HORIZONTAL, 4, 6 },
  { WEDGE_VERTICAL, 2, 4 },   { WEDGE_VERTICAL, 6, 4 },
};

#if !USE_LARGE_WEDGE_CODEBOOK
static const wedge_code_type wedge_codebook_16_hgtw[16] = {
  { WEDGE_OBLIQUE27, 4, 4 },  { WEDGE_OBLIQUE63, 4, 4 },
  { WEDGE_OBLIQUE117, 4, 4 }, { WEDGE_OBLIQUE153, 4, 4 },
  { WEDGE_HORIZONTAL, 4, 2 }, { WEDGE_HORIZONTAL, 4, 4 },
  { WEDGE_HORIZONTAL, 4, 6 }, { WEDGE_VERTICAL, 4, 4 },
  { WEDGE_OBLIQUE27, 4, 2 },  { WEDGE_OBLIQUE27, 4, 6 },
  { WEDGE_OBLIQUE153, 4, 2 }, { WEDGE_OBLIQUE153, 4, 6 },
  { WEDGE_OBLIQUE63, 2, 4 },  { WEDGE_OBLIQUE63, 6, 4 },
  { WEDGE_OBLIQUE117, 2, 4 }, { WEDGE_OBLIQUE117, 6, 4 },
};

static const wedge_code_type wedge_codebook_16_hltw[16] = {
  { WEDGE_OBLIQUE27, 4, 4 },  { WEDGE_OBLIQUE63, 4, 4 },
  { WEDGE_OBLIQUE117, 4, 4 }, { WEDGE_OBLIQUE153, 4, 4 },
  { WEDGE_VERTICAL, 2, 4 },   { WEDGE_VERTICAL, 4, 4 },
  { WEDGE_VERTICAL, 6, 4 },   { WEDGE_HORIZONTAL, 4, 4 },
  { WEDGE_OBLIQUE27, 4, 2 },  { WEDGE_OBLIQUE27, 4, 6 },
  { WEDGE_OBLIQUE153, 4, 2 }, { WEDGE_OBLIQUE153, 4, 6 },
  { WEDGE_OBLIQUE63, 2, 4 },  { WEDGE_OBLIQUE63, 6, 4 },
  { WEDGE_OBLIQUE117, 2, 4 }, { WEDGE_OBLIQUE117, 6, 4 },
};

static const wedge_code_type wedge_codebook_16_heqw[16] = {
  { WEDGE_OBLIQUE27, 4, 4 },  { WEDGE_OBLIQUE63, 4, 4 },
  { WEDGE_OBLIQUE117, 4, 4 }, { WEDGE_OBLIQUE153, 4, 4 },
  { WEDGE_HORIZONTAL, 4, 2 }, { WEDGE_HORIZONTAL, 4, 6 },
  { WEDGE_VERTICAL, 2, 4 },   { WEDGE_VERTICAL, 6, 4 },
  { WEDGE_OBLIQUE27, 4, 2 },  { WEDGE_OBLIQUE27, 4, 6 },
  { WEDGE_OBLIQUE153, 4, 2 }, { WEDGE_OBLIQUE153, 4, 6 },
  { WEDGE_OBLIQUE63, 2, 4 },  { WEDGE_OBLIQUE63, 6, 4 },
  { WEDGE_OBLIQUE117, 2, 4 }, { WEDGE_OBLIQUE117, 6, 4 },
};

const wedge_params_type wedge_params_lookup[BLOCK_SIZES] = {
  { 0, NULL, NULL, 0, NULL },
  { 0, NULL, NULL, 0, NULL },
  { 0, NULL, NULL, 0, NULL },
  { 4, wedge_codebook_16_heqw, wedge_signflip_lookup[3], 0, wedge_masks[3] },
  { 4, wedge_codebook_16_hgtw, wedge_signflip_lookup[4], 0, wedge_masks[4] },
  { 4, wedge_codebook_16_hltw, wedge_signflip_lookup[5], 0, wedge_masks[5] },
  { 4, wedge_codebook_16_heqw, wedge_signflip_lookup[6], 0, wedge_masks[6] },
  { 4, wedge_codebook_16_hgtw, wedge_signflip_lookup[7], 0, wedge_masks[7] },
  { 4, wedge_codebook_16_hltw, wedge_signflip_lookup[8], 0, wedge_masks[8] },
  { 4, wedge_codebook_16_heqw, wedge_signflip_lookup[9], 0, wedge_masks[9] },
  { 0, wedge_codebook_8_hgtw, wedge_signflip_lookup[10], 0, wedge_masks[10] },
  { 0, wedge_codebook_8_hltw, wedge_signflip_lookup[11], 0, wedge_masks[11] },
  { 0, wedge_codebook_8_heqw, wedge_signflip_lookup[12], 0, wedge_masks[12] },
#if CONFIG_EXT_PARTITION
  { 0, NULL, NULL, 0, NULL },
  { 0, NULL, NULL, 0, NULL },
  { 0, NULL, NULL, 0, NULL },
#endif  // CONFIG_EXT_PARTITION
};

#else

static const wedge_code_type wedge_codebook_32_hgtw[32] = {
  { WEDGE_OBLIQUE27, 4, 4 },  { WEDGE_OBLIQUE63, 4, 4 },
  { WEDGE_OBLIQUE117, 4, 4 }, { WEDGE_OBLIQUE153, 4, 4 },
  { WEDGE_HORIZONTAL, 4, 2 }, { WEDGE_HORIZONTAL, 4, 4 },
  { WEDGE_HORIZONTAL, 4, 6 }, { WEDGE_VERTICAL, 4, 4 },
  { WEDGE_OBLIQUE27, 4, 1 },  { WEDGE_OBLIQUE27, 4, 2 },
  { WEDGE_OBLIQUE27, 4, 3 },  { WEDGE_OBLIQUE27, 4, 5 },
  { WEDGE_OBLIQUE27, 4, 6 },  { WEDGE_OBLIQUE27, 4, 7 },
  { WEDGE_OBLIQUE153, 4, 1 }, { WEDGE_OBLIQUE153, 4, 2 },
  { WEDGE_OBLIQUE153, 4, 3 }, { WEDGE_OBLIQUE153, 4, 5 },
  { WEDGE_OBLIQUE153, 4, 6 }, { WEDGE_OBLIQUE153, 4, 7 },
  { WEDGE_OBLIQUE63, 1, 4 },  { WEDGE_OBLIQUE63, 2, 4 },
  { WEDGE_OBLIQUE63, 3, 4 },  { WEDGE_OBLIQUE63, 5, 4 },
  { WEDGE_OBLIQUE63, 6, 4 },  { WEDGE_OBLIQUE63, 7, 4 },
  { WEDGE_OBLIQUE117, 1, 4 }, { WEDGE_OBLIQUE117, 2, 4 },
  { WEDGE_OBLIQUE117, 3, 4 }, { WEDGE_OBLIQUE117, 5, 4 },
  { WEDGE_OBLIQUE117, 6, 4 }, { WEDGE_OBLIQUE117, 7, 4 },
};

static const wedge_code_type wedge_codebook_32_hltw[32] = {
  { WEDGE_OBLIQUE27, 4, 4 },  { WEDGE_OBLIQUE63, 4, 4 },
  { WEDGE_OBLIQUE117, 4, 4 }, { WEDGE_OBLIQUE153, 4, 4 },
  { WEDGE_VERTICAL, 2, 4 },   { WEDGE_VERTICAL, 4, 4 },
  { WEDGE_VERTICAL, 6, 4 },   { WEDGE_HORIZONTAL, 4, 4 },
  { WEDGE_OBLIQUE27, 4, 1 },  { WEDGE_OBLIQUE27, 4, 2 },
  { WEDGE_OBLIQUE27, 4, 3 },  { WEDGE_OBLIQUE27, 4, 5 },
  { WEDGE_OBLIQUE27, 4, 6 },  { WEDGE_OBLIQUE27, 4, 7 },
  { WEDGE_OBLIQUE153, 4, 1 }, { WEDGE_OBLIQUE153, 4, 2 },
  { WEDGE_OBLIQUE153, 4, 3 }, { WEDGE_OBLIQUE153, 4, 5 },
  { WEDGE_OBLIQUE153, 4, 6 }, { WEDGE_OBLIQUE153, 4, 7 },
  { WEDGE_OBLIQUE63, 1, 4 },  { WEDGE_OBLIQUE63, 2, 4 },
  { WEDGE_OBLIQUE63, 3, 4 },  { WEDGE_OBLIQUE63, 5, 4 },
  { WEDGE_OBLIQUE63, 6, 4 },  { WEDGE_OBLIQUE63, 7, 4 },
  { WEDGE_OBLIQUE117, 1, 4 }, { WEDGE_OBLIQUE117, 2, 4 },
  { WEDGE_OBLIQUE117, 3, 4 }, { WEDGE_OBLIQUE117, 5, 4 },
  { WEDGE_OBLIQUE117, 6, 4 }, { WEDGE_OBLIQUE117, 7, 4 },
};

static const wedge_code_type wedge_codebook_32_heqw[32] = {
  { WEDGE_OBLIQUE27, 4, 4 },  { WEDGE_OBLIQUE63, 4, 4 },
  { WEDGE_OBLIQUE117, 4, 4 }, { WEDGE_OBLIQUE153, 4, 4 },
  { WEDGE_HORIZONTAL, 4, 2 }, { WEDGE_HORIZONTAL, 4, 6 },
  { WEDGE_VERTICAL, 2, 4 },   { WEDGE_VERTICAL, 6, 4 },
  { WEDGE_OBLIQUE27, 4, 1 },  { WEDGE_OBLIQUE27, 4, 2 },
  { WEDGE_OBLIQUE27, 4, 3 },  { WEDGE_OBLIQUE27, 4, 5 },
  { WEDGE_OBLIQUE27, 4, 6 },  { WEDGE_OBLIQUE27, 4, 7 },
  { WEDGE_OBLIQUE153, 4, 1 }, { WEDGE_OBLIQUE153, 4, 2 },
  { WEDGE_OBLIQUE153, 4, 3 }, { WEDGE_OBLIQUE153, 4, 5 },
  { WEDGE_OBLIQUE153, 4, 6 }, { WEDGE_OBLIQUE153, 4, 7 },
  { WEDGE_OBLIQUE63, 1, 4 },  { WEDGE_OBLIQUE63, 2, 4 },
  { WEDGE_OBLIQUE63, 3, 4 },  { WEDGE_OBLIQUE63, 5, 4 },
  { WEDGE_OBLIQUE63, 6, 4 },  { WEDGE_OBLIQUE63, 7, 4 },
  { WEDGE_OBLIQUE117, 1, 4 }, { WEDGE_OBLIQUE117, 2, 4 },
  { WEDGE_OBLIQUE117, 3, 4 }, { WEDGE_OBLIQUE117, 5, 4 },
  { WEDGE_OBLIQUE117, 6, 4 }, { WEDGE_OBLIQUE117, 7, 4 },
};

const wedge_params_type wedge_params_lookup[BLOCK_SIZES] = {
  { 0, NULL, NULL, 0, NULL },
  { 0, NULL, NULL, 0, NULL },
  { 0, NULL, NULL, 0, NULL },
  { 5, wedge_codebook_32_heqw, wedge_signflip_lookup[3], 0, wedge_masks[3] },
  { 5, wedge_codebook_32_hgtw, wedge_signflip_lookup[4], 0, wedge_masks[4] },
  { 5, wedge_codebook_32_hltw, wedge_signflip_lookup[5], 0, wedge_masks[5] },
  { 5, wedge_codebook_32_heqw, wedge_signflip_lookup[6], 0, wedge_masks[6] },
  { 5, wedge_codebook_32_hgtw, wedge_signflip_lookup[7], 0, wedge_masks[7] },
  { 5, wedge_codebook_32_hltw, wedge_signflip_lookup[8], 0, wedge_masks[8] },
  { 5, wedge_codebook_32_heqw, wedge_signflip_lookup[9], 0, wedge_masks[9] },
  { 0, wedge_codebook_8_hgtw, wedge_signflip_lookup[10], 0, wedge_masks[10] },
  { 0, wedge_codebook_8_hltw, wedge_signflip_lookup[11], 0, wedge_masks[11] },
  { 0, wedge_codebook_8_heqw, wedge_signflip_lookup[12], 0, wedge_masks[12] },
#if CONFIG_EXT_PARTITION
  { 0, NULL, NULL, 0, NULL },
  { 0, NULL, NULL, 0, NULL },
  { 0, NULL, NULL, 0, NULL },
#endif  // CONFIG_EXT_PARTITION
};
#endif  // USE_LARGE_WEDGE_CODEBOOK

static const uint8_t *get_wedge_mask_inplace(int wedge_index, int neg,
                                             BLOCK_SIZE sb_type) {
  const uint8_t *master;
  const int bh = 4 << b_height_log2_lookup[sb_type];
  const int bw = 4 << b_width_log2_lookup[sb_type];
  const wedge_code_type *a =
      wedge_params_lookup[sb_type].codebook + wedge_index;
  const int smoother = wedge_params_lookup[sb_type].smoother;
  int woff, hoff;
  const uint8_t wsignflip = wedge_params_lookup[sb_type].signflip[wedge_index];

  assert(wedge_index >= 0 &&
         wedge_index < (1 << get_wedge_bits_lookup(sb_type)));
  woff = (a->x_offset * bw) >> 3;
  hoff = (a->y_offset * bh) >> 3;
  master = wedge_mask_obl[smoother][neg ^ wsignflip][a->direction] +
           MASK_MASTER_STRIDE * (MASK_MASTER_SIZE / 2 - hoff) +
           MASK_MASTER_SIZE / 2 - woff;
  return master;
}

const uint8_t *av1_get_soft_mask(int wedge_index, int wedge_sign,
                                 BLOCK_SIZE sb_type, int offset_x,
                                 int offset_y) {
  const uint8_t *mask =
      get_wedge_mask_inplace(wedge_index, wedge_sign, sb_type);
  if (mask) mask -= (offset_x + offset_y * MASK_MASTER_STRIDE);
  return mask;
}

// get a mask according to the compound type
// TODO(sarahparker) this needs to be extended for other experiments and
// is currently only intended for ext_inter alone
#if CONFIG_EXT_INTER
const uint8_t *av1_get_compound_type_mask(
    const INTERINTER_COMPOUND_DATA *const comp_data, BLOCK_SIZE sb_type,
    int invert) {
  assert(is_masked_compound_type(comp_data->type));
  switch (comp_data->type) {
    case COMPOUND_WEDGE:
      return av1_get_contiguous_soft_mask(
          comp_data->wedge_index,
          invert ? !comp_data->wedge_sign : comp_data->wedge_sign, sb_type);
    default: assert(0); return NULL;
  }
}
#endif  // CONFIG_EXT_INTER

static void init_wedge_master_masks() {
  int i, j, s;
  const int w = MASK_MASTER_SIZE;
  const int h = MASK_MASTER_SIZE;
  const int stride = MASK_MASTER_STRIDE;
  const int a[2] = { 2, 1 };
  const double asqrt = sqrt(a[0] * a[0] + a[1] * a[1]);
  for (s = 0; s < NSMOOTHERS; s++) {
    for (i = 0; i < h; ++i)
      for (j = 0; j < w; ++j) {
        int x = (2 * j + 1 - w);
        int y = (2 * i + 1 - h);
        int m = (int)rint((a[0] * x + a[1] * y) / asqrt);
        wedge_mask_obl[s][1][WEDGE_OBLIQUE63][i * stride + j] =
            wedge_mask_obl[s][1][WEDGE_OBLIQUE27][j * stride + i] =
                get_masked_weight(m, s);
        wedge_mask_obl[s][1][WEDGE_OBLIQUE117][i * stride + w - 1 - j] =
            wedge_mask_obl[s][1][WEDGE_OBLIQUE153][(w - 1 - j) * stride + i] =
                (1 << WEDGE_WEIGHT_BITS) - get_masked_weight(m, s);
        wedge_mask_obl[s][0][WEDGE_OBLIQUE63][i * stride + j] =
            wedge_mask_obl[s][0][WEDGE_OBLIQUE27][j * stride + i] =
                (1 << WEDGE_WEIGHT_BITS) - get_masked_weight(m, s);
        wedge_mask_obl[s][0][WEDGE_OBLIQUE117][i * stride + w - 1 - j] =
            wedge_mask_obl[s][0][WEDGE_OBLIQUE153][(w - 1 - j) * stride + i] =
                get_masked_weight(m, s);
        wedge_mask_obl[s][1][WEDGE_VERTICAL][i * stride + j] =
            wedge_mask_obl[s][1][WEDGE_HORIZONTAL][j * stride + i] =
                get_masked_weight(x, s);
        wedge_mask_obl[s][0][WEDGE_VERTICAL][i * stride + j] =
            wedge_mask_obl[s][0][WEDGE_HORIZONTAL][j * stride + i] =
                (1 << WEDGE_WEIGHT_BITS) - get_masked_weight(x, s);
      }
  }
}

// If the signs for the wedges for various blocksizes are
// inconsistent flip the sign flag. Do it only once for every
// wedge codebook.
static void init_wedge_signs() {
  BLOCK_SIZE sb_type;
  memset(wedge_signflip_lookup, 0, sizeof(wedge_signflip_lookup));
  for (sb_type = BLOCK_4X4; sb_type < BLOCK_SIZES; ++sb_type) {
    const int bw = block_size_wide[sb_type];
    const int bh = block_size_high[sb_type];
    const wedge_params_type wedge_params = wedge_params_lookup[sb_type];
    const int wbits = wedge_params.bits;
    const int wtypes = 1 << wbits;
    int i, w;
    if (wbits == 0) continue;
    for (w = 0; w < wtypes; ++w) {
      const uint8_t *mask = get_wedge_mask_inplace(w, 0, sb_type);
      int sum = 0;
      for (i = 0; i < bw; ++i) sum += mask[i];
      for (i = 0; i < bh; ++i) sum += mask[i * MASK_MASTER_STRIDE];
      sum = (sum + (bw + bh) / 2) / (bw + bh);
      wedge_params.signflip[w] = (sum < 32);
    }
  }
}

static void init_wedge_masks() {
  uint8_t *dst = wedge_mask_buf;
  BLOCK_SIZE bsize;
  memset(wedge_masks, 0, sizeof(wedge_masks));
  for (bsize = BLOCK_4X4; bsize < BLOCK_SIZES; ++bsize) {
    const uint8_t *mask;
    const int bw = block_size_wide[bsize];
    const int bh = block_size_high[bsize];
    const wedge_params_type *wedge_params = &wedge_params_lookup[bsize];
    const int wbits = wedge_params->bits;
    const int wtypes = 1 << wbits;
    int w;
    if (wbits == 0) continue;
    for (w = 0; w < wtypes; ++w) {
      mask = get_wedge_mask_inplace(w, 0, bsize);
      aom_convolve_copy(mask, MASK_MASTER_STRIDE, dst, bw, NULL, 0, NULL, 0, bw,
                        bh);
      wedge_params->masks[0][w] = dst;
      dst += bw * bh;

      mask = get_wedge_mask_inplace(w, 1, bsize);
      aom_convolve_copy(mask, MASK_MASTER_STRIDE, dst, bw, NULL, 0, NULL, 0, bw,
                        bh);
      wedge_params->masks[1][w] = dst;
      dst += bw * bh;
    }
    assert(sizeof(wedge_mask_buf) >= (size_t)(dst - wedge_mask_buf));
  }
}

// Equation of line: f(x, y) = a[0]*(x - a[2]*w/8) + a[1]*(y - a[3]*h/8) = 0
void av1_init_wedge_masks() {
  init_wedge_master_masks();
  init_wedge_signs();
  init_wedge_masks();
}

#if CONFIG_SUPERTX
static void build_masked_compound_wedge_extend(
    uint8_t *dst, int dst_stride, const uint8_t *src0, int src0_stride,
    const uint8_t *src1, int src1_stride, int wedge_index, int wedge_sign,
    BLOCK_SIZE sb_type, int wedge_offset_x, int wedge_offset_y, int h, int w) {
  const int subh = (2 << b_height_log2_lookup[sb_type]) == h;
  const int subw = (2 << b_width_log2_lookup[sb_type]) == w;
  const uint8_t *mask = av1_get_soft_mask(wedge_index, wedge_sign, sb_type,
                                          wedge_offset_x, wedge_offset_y);
  aom_blend_a64_mask(dst, dst_stride, src0, src0_stride, src1, src1_stride,
                     mask, MASK_MASTER_STRIDE, h, w, subh, subw);
}

#if CONFIG_AOM_HIGHBITDEPTH
static void build_masked_compound_wedge_extend_highbd(
    uint8_t *dst_8, int dst_stride, const uint8_t *src0_8, int src0_stride,
    const uint8_t *src1_8, int src1_stride, int wedge_index, int wedge_sign,
    BLOCK_SIZE sb_type, int wedge_offset_x, int wedge_offset_y, int h, int w,
    int bd) {
  const int subh = (2 << b_height_log2_lookup[sb_type]) == h;
  const int subw = (2 << b_width_log2_lookup[sb_type]) == w;
  const uint8_t *mask = av1_get_soft_mask(wedge_index, wedge_sign, sb_type,
                                          wedge_offset_x, wedge_offset_y);
  aom_highbd_blend_a64_mask(dst_8, dst_stride, src0_8, src0_stride, src1_8,
                            src1_stride, mask, MASK_MASTER_STRIDE, h, w, subh,
                            subw, bd);
}
#endif  // CONFIG_AOM_HIGHBITDEPTH
#endif  // CONFIG_SUPERTX

static void build_masked_compound(
    uint8_t *dst, int dst_stride, const uint8_t *src0, int src0_stride,
    const uint8_t *src1, int src1_stride,
    const INTERINTER_COMPOUND_DATA *const comp_data, BLOCK_SIZE sb_type, int h,
    int w) {
  // Derive subsampling from h and w passed in. May be refactored to
  // pass in subsampling factors directly.
  const int subh = (2 << b_height_log2_lookup[sb_type]) == h;
  const int subw = (2 << b_width_log2_lookup[sb_type]) == w;
  const uint8_t *mask = av1_get_compound_type_mask(comp_data, sb_type, 0);
  aom_blend_a64_mask(dst, dst_stride, src0, src0_stride, src1, src1_stride,
                     mask, block_size_wide[sb_type], h, w, subh, subw);
}

#if CONFIG_AOM_HIGHBITDEPTH
static void build_masked_compound_wedge_highbd(
    uint8_t *dst_8, int dst_stride, const uint8_t *src0_8, int src0_stride,
    const uint8_t *src1_8, int src1_stride, int wedge_index, int wedge_sign,
    BLOCK_SIZE sb_type, int h, int w, int bd) {
  // Derive subsampling from h and w passed in. May be refactored to
  // pass in subsampling factors directly.
  const int subh = (2 << b_height_log2_lookup[sb_type]) == h;
  const int subw = (2 << b_width_log2_lookup[sb_type]) == w;
  const uint8_t *mask =
      av1_get_contiguous_soft_mask(wedge_index, wedge_sign, sb_type);
  aom_highbd_blend_a64_mask(dst_8, dst_stride, src0_8, src0_stride, src1_8,
                            src1_stride, mask, block_size_wide[sb_type], h, w,
                            subh, subw, bd);
}
#endif  // CONFIG_AOM_HIGHBITDEPTH

void av1_make_masked_inter_predictor(const uint8_t *pre, int pre_stride,
                                     uint8_t *dst, int dst_stride,
                                     const int subpel_x, const int subpel_y,
                                     const struct scale_factors *sf, int w,
                                     int h,
#if CONFIG_DUAL_FILTER
                                     const InterpFilter *interp_filter,
#else
                                     const InterpFilter interp_filter,
#endif
                                     int xs, int ys,
#if CONFIG_SUPERTX
                                     int wedge_offset_x, int wedge_offset_y,
#endif  // CONFIG_SUPERTX
                                     const MACROBLOCKD *xd) {
  const MODE_INFO *mi = xd->mi[0];
  const INTERINTER_COMPOUND_DATA *const comp_data =
      &mi->mbmi.interinter_compound_data;
// The prediction filter types used here should be those for
// the second reference block.
#if CONFIG_DUAL_FILTER
  InterpFilter tmp_ipf[4] = {
    interp_filter[2], interp_filter[3], interp_filter[2], interp_filter[3],
  };
#else
  InterpFilter tmp_ipf = interp_filter;
#endif  // CONFIG_DUAL_FILTER
#if CONFIG_AOM_HIGHBITDEPTH
  DECLARE_ALIGNED(16, uint8_t, tmp_dst_[2 * MAX_SB_SQUARE]);
  uint8_t *tmp_dst = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
                         ? CONVERT_TO_BYTEPTR(tmp_dst_)
                         : tmp_dst_;
  av1_make_inter_predictor(pre, pre_stride, tmp_dst, MAX_SB_SIZE, subpel_x,
                           subpel_y, sf, w, h, 0, tmp_ipf, xs, ys, xd);
#if CONFIG_SUPERTX
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
    build_masked_compound_wedge_extend_highbd(
        dst, dst_stride, dst, dst_stride, tmp_dst, MAX_SB_SIZE,
        comp_data->wedge_index, comp_data->wedge_sign, mi->mbmi.sb_type,
        wedge_offset_x, wedge_offset_y, h, w, xd->bd);
  else
    build_masked_compound_wedge_extend(
        dst, dst_stride, dst, dst_stride, tmp_dst, MAX_SB_SIZE,
        comp_data->wedge_index, comp_data->wedge_sign, mi->mbmi.sb_type,
        wedge_offset_x, wedge_offset_y, h, w);
#else
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
    build_masked_compound_wedge_highbd(
        dst, dst_stride, dst, dst_stride, tmp_dst, MAX_SB_SIZE,
        comp_data->wedge_index, comp_data->wedge_sign, mi->mbmi.sb_type, h, w,
        xd->bd);
  else
    build_masked_compound(dst, dst_stride, dst, dst_stride, tmp_dst,
                          MAX_SB_SIZE, comp_data, mi->mbmi.sb_type, h, w);
#endif  // CONFIG_SUPERTX
#else   // CONFIG_AOM_HIGHBITDEPTH
  DECLARE_ALIGNED(16, uint8_t, tmp_dst[MAX_SB_SQUARE]);
  av1_make_inter_predictor(pre, pre_stride, tmp_dst, MAX_SB_SIZE, subpel_x,
                           subpel_y, sf, w, h, 0, tmp_ipf, xs, ys, xd);
#if CONFIG_SUPERTX
  build_masked_compound_wedge_extend(dst, dst_stride, dst, dst_stride, tmp_dst,
                                     MAX_SB_SIZE, comp_data->wedge_index,
                                     comp_data->wedge_sign, mi->mbmi.sb_type,
                                     wedge_offset_x, wedge_offset_y, h, w);
#else
  build_masked_compound(dst, dst_stride, dst, dst_stride, tmp_dst, MAX_SB_SIZE,
                        comp_data, mi->mbmi.sb_type, h, w);
#endif  // CONFIG_SUPERTX
#endif  // CONFIG_AOM_HIGHBITDEPTH
}
#endif  // CONFIG_EXT_INTER

#if CONFIG_AOM_HIGHBITDEPTH
void av1_highbd_build_inter_predictor(
    const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride,
    const MV *src_mv, const struct scale_factors *sf, int w, int h, int ref,
#if CONFIG_DUAL_FILTER
    const InterpFilter *interp_filter,
#else
    const InterpFilter interp_filter,
#endif
    enum mv_precision precision, int x, int y, int bd) {
  const int is_q4 = precision == MV_PRECISION_Q4;
  const MV mv_q4 = { is_q4 ? src_mv->row : src_mv->row * 2,
                     is_q4 ? src_mv->col : src_mv->col * 2 };
  MV32 mv = av1_scale_mv(&mv_q4, x, y, sf);
  const int subpel_x = mv.col & SUBPEL_MASK;
  const int subpel_y = mv.row & SUBPEL_MASK;

  src += (mv.row >> SUBPEL_BITS) * src_stride + (mv.col >> SUBPEL_BITS);

  highbd_inter_predictor(src, src_stride, dst, dst_stride, subpel_x, subpel_y,
                         sf, w, h, ref, interp_filter, sf->x_step_q4,
                         sf->y_step_q4, bd);
}
#endif  // CONFIG_AOM_HIGHBITDEPTH

void av1_build_inter_predictor(const uint8_t *src, int src_stride, uint8_t *dst,
                               int dst_stride, const MV *src_mv,
                               const struct scale_factors *sf, int w, int h,
                               int ref,
#if CONFIG_DUAL_FILTER
                               const InterpFilter *interp_filter,
#else
                               const InterpFilter interp_filter,
#endif
                               enum mv_precision precision, int x, int y) {
  const int is_q4 = precision == MV_PRECISION_Q4;
  const MV mv_q4 = { is_q4 ? src_mv->row : src_mv->row * 2,
                     is_q4 ? src_mv->col : src_mv->col * 2 };
  MV32 mv = av1_scale_mv(&mv_q4, x, y, sf);
  const int subpel_x = mv.col & SUBPEL_MASK;
  const int subpel_y = mv.row & SUBPEL_MASK;

  src += (mv.row >> SUBPEL_BITS) * src_stride + (mv.col >> SUBPEL_BITS);

  inter_predictor(src, src_stride, dst, dst_stride, subpel_x, subpel_y, sf, w,
                  h, ref, interp_filter, sf->x_step_q4, sf->y_step_q4);
}

void build_inter_predictors(MACROBLOCKD *xd, int plane,
#if CONFIG_MOTION_VAR
                            int mi_col_offset, int mi_row_offset,
#endif  // CONFIG_MOTION_VAR
                            int block, int bw, int bh, int x, int y, int w,
                            int h,
#if CONFIG_SUPERTX && CONFIG_EXT_INTER
                            int wedge_offset_x, int wedge_offset_y,
#endif  // CONFIG_SUPERTX && CONFIG_EXT_INTER
                            int mi_x, int mi_y) {
  struct macroblockd_plane *const pd = &xd->plane[plane];
#if CONFIG_MOTION_VAR
  const MODE_INFO *mi = xd->mi[mi_col_offset + xd->mi_stride * mi_row_offset];
  const int build_for_obmc = !(mi_col_offset == 0 && mi_row_offset == 0);
#else
  const MODE_INFO *mi = xd->mi[0];
#endif  // CONFIG_MOTION_VAR
  const int is_compound = has_second_ref(&mi->mbmi);
  int ref;
#if CONFIG_GLOBAL_MOTION
  WarpedMotionParams *gm[2];
  int is_global[2];
  for (ref = 0; ref < 1 + is_compound; ++ref) {
    gm[ref] = &xd->global_motion[mi->mbmi.ref_frame[ref]];
    is_global[ref] =
        (get_y_mode(mi, block) == ZEROMV && gm[ref]->wmtype > TRANSLATION);
  }
  // TODO(sarahparker) remove these once gm works with all experiments
  (void)gm;
  (void)is_global;
#endif  // CONFIG_GLOBAL_MOTION

// TODO(sarahparker) enable the use of DUAL_FILTER in warped motion functions
// in order to allow GLOBAL_MOTION and DUAL_FILTER to work together
#if CONFIG_SUB8X8_MC
#if CONFIG_MOTION_VAR
  if (mi->mbmi.sb_type < BLOCK_8X8 && plane > 0 && !build_for_obmc) {
#else
  if (mi->mbmi.sb_type < BLOCK_8X8 && plane > 0) {
#endif  // CONFIG_MOTION_VAR
    // block size in log2
    const int b4_wl = b_width_log2_lookup[mi->mbmi.sb_type];
    const int b4_hl = b_height_log2_lookup[mi->mbmi.sb_type];
    const int b8_sl = b_width_log2_lookup[BLOCK_8X8];

    // block size
    const int b4_w = 1 << b4_wl;
    const int b4_h = 1 << b4_hl;
    const int b8_s = 1 << b8_sl;
    int idx, idy;

    const int x_base = x;
    const int y_base = y;

    // processing unit size
    const int x_step = w >> (b8_sl - b4_wl);
    const int y_step = h >> (b8_sl - b4_hl);

    for (idy = 0; idy < b8_s; idy += b4_h) {
      for (idx = 0; idx < b8_s; idx += b4_w) {
        const int chr_idx = (idy * 2) + idx;
        for (ref = 0; ref < 1 + is_compound; ++ref) {
          const struct scale_factors *const sf = &xd->block_refs[ref]->sf;
          struct buf_2d *const pre_buf = &pd->pre[ref];
          struct buf_2d *const dst_buf = &pd->dst;
          uint8_t *dst = dst_buf->buf;
          const MV mv = mi->bmi[chr_idx].as_mv[ref].as_mv;
          const MV mv_q4 = clamp_mv_to_umv_border_sb(
              xd, &mv, bw, bh, pd->subsampling_x, pd->subsampling_y);
          uint8_t *pre;
          MV32 scaled_mv;
          int xs, ys, subpel_x, subpel_y;
          const int is_scaled = av1_is_scaled(sf);

          x = x_base + idx * x_step;
          y = y_base + idy * y_step;

          dst += dst_buf->stride * y + x;

          if (is_scaled) {
            pre =
                pre_buf->buf + scaled_buffer_offset(x, y, pre_buf->stride, sf);
            scaled_mv = av1_scale_mv(&mv_q4, mi_x + x, mi_y + y, sf);
            xs = sf->x_step_q4;
            ys = sf->y_step_q4;
          } else {
            pre = pre_buf->buf + y * pre_buf->stride + x;
            scaled_mv.row = mv_q4.row;
            scaled_mv.col = mv_q4.col;
            xs = ys = 16;
          }

          subpel_x = scaled_mv.col & SUBPEL_MASK;
          subpel_y = scaled_mv.row & SUBPEL_MASK;
          pre += (scaled_mv.row >> SUBPEL_BITS) * pre_buf->stride +
                 (scaled_mv.col >> SUBPEL_BITS);

#if CONFIG_EXT_INTER
          if (ref &&
              is_masked_compound_type(mi->mbmi.interinter_compound_data.type))
            av1_make_masked_inter_predictor(
                pre, pre_buf->stride, dst, dst_buf->stride, subpel_x, subpel_y,
                sf, w, h, mi->mbmi.interp_filter, xs, ys,
#if CONFIG_SUPERTX
                wedge_offset_x, wedge_offset_y,
#endif  // CONFIG_SUPERTX
                xd);
          else
#endif  // CONFIG_EXT_INTER
            av1_make_inter_predictor(pre, pre_buf->stride, dst, dst_buf->stride,
                                     subpel_x, subpel_y, sf, x_step, y_step,
                                     ref, mi->mbmi.interp_filter, xs, ys, xd);
        }
      }
    }
    return;
  }
#endif

  for (ref = 0; ref < 1 + is_compound; ++ref) {
    const struct scale_factors *const sf = &xd->block_refs[ref]->sf;
    struct buf_2d *const pre_buf = &pd->pre[ref];
    struct buf_2d *const dst_buf = &pd->dst;
    uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
    const MV mv = mi->mbmi.sb_type < BLOCK_8X8
#if CONFIG_MOTION_VAR
                      ? (build_for_obmc ? mi->bmi[block].as_mv[ref].as_mv
                                        : average_split_mvs(pd, mi, ref, block))
#else
                      ? average_split_mvs(pd, mi, ref, block)
#endif  // CONFIG_MOTION_VAR
                      : mi->mbmi.mv[ref].as_mv;

    // TODO(jkoleszar): This clamping is done in the incorrect place for the
    // scaling case. It needs to be done on the scaled MV, not the pre-scaling
    // MV. Note however that it performs the subsampling aware scaling so
    // that the result is always q4.
    // mv_precision precision is MV_PRECISION_Q4.
    const MV mv_q4 = clamp_mv_to_umv_border_sb(
        xd, &mv, bw, bh, pd->subsampling_x, pd->subsampling_y);

    uint8_t *pre;
    MV32 scaled_mv;
    int xs, ys, subpel_x, subpel_y;
    const int is_scaled = av1_is_scaled(sf);

    if (is_scaled) {
      pre = pre_buf->buf + scaled_buffer_offset(x, y, pre_buf->stride, sf);
      scaled_mv = av1_scale_mv(&mv_q4, mi_x + x, mi_y + y, sf);
      xs = sf->x_step_q4;
      ys = sf->y_step_q4;
    } else {
      pre = pre_buf->buf + (y * pre_buf->stride + x);
      scaled_mv.row = mv_q4.row;
      scaled_mv.col = mv_q4.col;
      xs = ys = 16;
    }

    subpel_x = scaled_mv.col & SUBPEL_MASK;
    subpel_y = scaled_mv.row & SUBPEL_MASK;
    pre += (scaled_mv.row >> SUBPEL_BITS) * pre_buf->stride +
           (scaled_mv.col >> SUBPEL_BITS);

#if CONFIG_EXT_INTER
    if (ref && is_masked_compound_type(mi->mbmi.interinter_compound_data.type))
      av1_make_masked_inter_predictor(pre, pre_buf->stride, dst,
                                      dst_buf->stride, subpel_x, subpel_y, sf,
                                      w, h, mi->mbmi.interp_filter, xs, ys,
#if CONFIG_SUPERTX
                                      wedge_offset_x, wedge_offset_y,
#endif  // CONFIG_SUPERTX
                                      xd);
    else
#else  // CONFIG_EXT_INTER
#if CONFIG_GLOBAL_MOTION
    if (is_global[ref])
      av1_warp_plane(gm[ref],
#if CONFIG_AOM_HIGHBITDEPTH
                     xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH, xd->bd,
#endif  // CONFIG_AOM_HIGHBITDEPTH
                     pre_buf->buf0, pre_buf->width, pre_buf->height,
                     pre_buf->stride, dst, (mi_x >> pd->subsampling_x) + x,
                     (mi_y >> pd->subsampling_y) + y, w, h, dst_buf->stride,
                     pd->subsampling_x, pd->subsampling_y, xs, ys, ref);
    else
#endif  // CONFIG_GLOBAL_MOTION
#endif  // CONFIG_EXT_INTER
      av1_make_inter_predictor(pre, pre_buf->stride, dst, dst_buf->stride,
                               subpel_x, subpel_y, sf, w, h, ref,
                               mi->mbmi.interp_filter, xs, ys, xd);
  }
}

void av1_build_inter_predictor_sub8x8(MACROBLOCKD *xd, int plane, int i, int ir,
                                      int ic, int mi_row, int mi_col) {
  struct macroblockd_plane *const pd = &xd->plane[plane];
  MODE_INFO *const mi = xd->mi[0];
  const BLOCK_SIZE plane_bsize = get_plane_block_size(mi->mbmi.sb_type, pd);
  const int width = block_size_wide[plane_bsize];
  const int height = block_size_high[plane_bsize];
  uint8_t *const dst = &pd->dst.buf[(ir * pd->dst.stride + ic) << 2];
  int ref;
  const int is_compound = has_second_ref(&mi->mbmi);

  for (ref = 0; ref < 1 + is_compound; ++ref) {
    const uint8_t *pre =
        &pd->pre[ref].buf[(ir * pd->pre[ref].stride + ic) << 2];
#if CONFIG_AOM_HIGHBITDEPTH
    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
      av1_highbd_build_inter_predictor(
          pre, pd->pre[ref].stride, dst, pd->dst.stride,
          &mi->bmi[i].as_mv[ref].as_mv, &xd->block_refs[ref]->sf, width, height,
          ref, mi->mbmi.interp_filter, MV_PRECISION_Q3,
          mi_col * MI_SIZE + 4 * ic, mi_row * MI_SIZE + 4 * ir, xd->bd);
    } else {
      av1_build_inter_predictor(
          pre, pd->pre[ref].stride, dst, pd->dst.stride,
          &mi->bmi[i].as_mv[ref].as_mv, &xd->block_refs[ref]->sf, width, height,
          ref, mi->mbmi.interp_filter, MV_PRECISION_Q3,
          mi_col * MI_SIZE + 4 * ic, mi_row * MI_SIZE + 4 * ir);
    }
#else
    av1_build_inter_predictor(
        pre, pd->pre[ref].stride, dst, pd->dst.stride,
        &mi->bmi[i].as_mv[ref].as_mv, &xd->block_refs[ref]->sf, width, height,
        ref, mi->mbmi.interp_filter, MV_PRECISION_Q3, mi_col * MI_SIZE + 4 * ic,
        mi_row * MI_SIZE + 4 * ir);
#endif  // CONFIG_AOM_HIGHBITDEPTH
  }
}

static void build_inter_predictors_for_planes(MACROBLOCKD *xd, BLOCK_SIZE bsize,
                                              int mi_row, int mi_col,
                                              int plane_from, int plane_to) {
  int plane;
  const int mi_x = mi_col * MI_SIZE;
  const int mi_y = mi_row * MI_SIZE;
  for (plane = plane_from; plane <= plane_to; ++plane) {
    const struct macroblockd_plane *pd = &xd->plane[plane];
    const int bw = block_size_wide[bsize] >> pd->subsampling_x;
    const int bh = block_size_high[bsize] >> pd->subsampling_y;

    if (xd->mi[0]->mbmi.sb_type < BLOCK_8X8) {
      const PARTITION_TYPE bp = bsize - xd->mi[0]->mbmi.sb_type;
      const int have_vsplit = bp != PARTITION_HORZ;
      const int have_hsplit = bp != PARTITION_VERT;
      const int num_4x4_w = 2 >> ((!have_vsplit) | pd->subsampling_x);
      const int num_4x4_h = 2 >> ((!have_hsplit) | pd->subsampling_y);
      const int pw = 8 >> (have_vsplit | pd->subsampling_x);
      const int ph = 8 >> (have_hsplit | pd->subsampling_y);
      int x, y;
      assert(bp != PARTITION_NONE && bp < PARTITION_TYPES);
      assert(bsize == BLOCK_8X8);
      assert(pw * num_4x4_w == bw && ph * num_4x4_h == bh);
      for (y = 0; y < num_4x4_h; ++y)
        for (x = 0; x < num_4x4_w; ++x)
          build_inter_predictors(xd, plane,
#if CONFIG_MOTION_VAR
                                 0, 0,
#endif  // CONFIG_MOTION_VAR
                                 y * 2 + x, bw, bh, 4 * x, 4 * y, pw, ph,
#if CONFIG_SUPERTX && CONFIG_EXT_INTER
                                 0, 0,
#endif  // CONFIG_SUPERTX && CONFIG_EXT_INTER
                                 mi_x, mi_y);
    } else {
      build_inter_predictors(xd, plane,
#if CONFIG_MOTION_VAR
                             0, 0,
#endif  // CONFIG_MOTION_VAR
                             0, bw, bh, 0, 0, bw, bh,
#if CONFIG_SUPERTX && CONFIG_EXT_INTER
                             0, 0,
#endif  // CONFIG_SUPERTX && CONFIG_EXT_INTER
                             mi_x, mi_y);
    }
  }
}

void av1_build_inter_predictors_sby(MACROBLOCKD *xd, int mi_row, int mi_col,
                                    BUFFER_SET *ctx, BLOCK_SIZE bsize) {
  build_inter_predictors_for_planes(xd, bsize, mi_row, mi_col, 0, 0);
#if CONFIG_EXT_INTER
  if (is_interintra_pred(&xd->mi[0]->mbmi)) {
    BUFFER_SET default_ctx = { { xd->plane[0].dst.buf, NULL, NULL },
                               { xd->plane[0].dst.stride, 0, 0 } };
    if (!ctx) ctx = &default_ctx;
    av1_build_interintra_predictors_sby(xd, xd->plane[0].dst.buf,
                                        xd->plane[0].dst.stride, ctx, bsize);
  }
#else
  (void)ctx;
#endif  // CONFIG_EXT_INTER
}

void av1_build_inter_predictors_sbp(MACROBLOCKD *xd, int mi_row, int mi_col,
                                    BUFFER_SET *ctx, BLOCK_SIZE bsize,
                                    int plane) {
  build_inter_predictors_for_planes(xd, bsize, mi_row, mi_col, plane, plane);
#if CONFIG_EXT_INTER
  if (is_interintra_pred(&xd->mi[