/*
 * 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 <stdio.h>
#include <limits.h>

#include "config/aom_config.h"
#include "config/aom_dsp_rtcd.h"
#include "config/aom_scale_rtcd.h"
#if CONFIG_DERIVED_MV
#include "config/av1_rtcd.h"
#endif  // CONFIG_DERIVED_MV

#include "aom/aom_integer.h"
#include "aom_dsp/blend.h"
#if CONFIG_DERIVED_MV
#include "aom_dsp/variance.h"
#endif  // CONFIG_DERIVED_MV

#include "av1/common/blockd.h"
#include "av1/common/mvref_common.h"
#include "av1/common/reconinter.h"
#include "av1/common/reconintra.h"
#include "av1/common/onyxc_int.h"
#include "av1/common/obmc.h"

#if CONFIG_INTERINTRA_ML
#include "av1/common/interintra_ml.h"
#endif

#define USE_PRECOMPUTED_WEDGE_MASK 1
#define USE_PRECOMPUTED_WEDGE_SIGN 1

#if CONFIG_CTX_ADAPT_LOG_WEIGHT
#define LOG_K_SIZE 1
#define LOG_WEIGHT_0 43
#define LOG_WEIGHT_1 40
#define DIFFLOG_THR 3
static const double log_k[9] = { 0.1004,  -0.0234, 0.1004,  -0.0234, -0.3079,
                                 -0.0234, 0.1004,  -0.0234, 0.1004 };

double kernel_correlation(const CONV_BUF_TYPE *src, int stride, int h, int w,
                          int cr, int cc, const double *kernel, int ws,
                          ConvolveParams *conv_params) {
  double crr = 0;
  const int ks = 2 * ws + 1;
  const int bd = 8;
  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
  const int round_offset = (1 << (offset_bits - conv_params->round_1)) +
                           (1 << (offset_bits - conv_params->round_1 - 1));
  const int round_bits =
      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
  const uint16_t max = (1 << round_bits);
  const uint16_t min = round_offset;
  for (int i = -ws; i <= ws; ++i) {
    for (int j = -ws; j <= ws; ++j) {
      int r = (cr + i < 0) ? 0 : (cr + i >= h) ? h - 1 : cr + i;
      int c = (cc + j < 0) ? 0 : (cc + j >= w) ? w - 1 : cc + j;
      double s = (src[r * stride + c] - min) / ((double)max);
      crr += s * kernel[(i + 1) * ks + (j + 1)];
    }
  }
  return crr;
}

double kernel_correlation_uint8(const uint8_t *src, int stride, int h, int w,
                                int cr, int cc, const double *kernel, int ws) {
  double crr = 0;
  const int ks = 2 * ws + 1;
  for (int i = -ws; i <= ws; ++i) {
    for (int j = -ws; j <= ws; ++j) {
      int r = (cr + i < 0) ? 0 : (cr + i >= h) ? h - 1 : cr + i;
      int c = (cc + j < 0) ? 0 : (cc + j >= w) ? w - 1 : cc + j;
      crr += src[r * stride + c] * kernel[(i + ws) * ks + (j + ws)];
    }
  }
  return crr;
}

double *gen_correlation(const CONV_BUF_TYPE *src, int stride, int h, int w,
                        const double *kernel, int ws,
                        ConvolveParams *conv_params) {
  double *R;
  R = malloc(h * w * sizeof(*R));
  for (int i = 0; i < h; ++i) {
    for (int j = 0; j < w; ++j) {
      R[i * w + j] =
          kernel_correlation(src, stride, h, w, i, j, kernel, ws, conv_params);
    }
  }
  return R;
}

double *gen_correlation_uint8(const uint8_t *src, int stride, int h, int w,
                              const double *kernel, int ws) {
  double *R;
  R = malloc(h * w * sizeof(*R));
  for (int i = 0; i < h; ++i) {
    for (int j = 0; j < w; ++j) {
      R[i * w + j] =
          kernel_correlation_uint8(src, stride, h, w, i, j, kernel, ws);
    }
  }
  return R;
}
#endif  // CONFIG_CTX_ADAPT_LOG_WEIGHT

#if CONFIG_DIFFWTD_42
#define DIFFWTD_MASK_VAL 42
#define NORMAL_MASK DIFFWTD_42
#define INVERSE_MASK DIFFWTD_42_INV
#else
#define DIFFWTD_MASK_VAL 38
#define NORMAL_MASK DIFFWTD_38
#define INVERSE_MASK DIFFWTD_38_INV
#endif  // CONFIG_DIFFWTD_42

// This function will determine whether or not to create a warped
// prediction.
int av1_allow_warp(const MB_MODE_INFO *const mbmi,
                   const WarpTypesAllowed *const warp_types,
                   const WarpedMotionParams *const gm_params,
                   int build_for_obmc, const struct scale_factors *const sf,
                   WarpedMotionParams *final_warp_params) {
  // Note: As per the spec, we must test the fixed point scales here, which are
  // at a higher precision (1 << 14) than the xs and ys in subpel_params (that
  // have 1 << 10 precision).
  if (av1_is_scaled(sf)) return 0;

  if (final_warp_params != NULL) *final_warp_params = default_warp_params;

  if (build_for_obmc) return 0;

  if (warp_types->local_warp_allowed && !mbmi->wm_params.invalid) {
    if (final_warp_params != NULL)
      memcpy(final_warp_params, &mbmi->wm_params, sizeof(*final_warp_params));
    return 1;
  } else if (warp_types->global_warp_allowed && !gm_params->invalid) {
    if (final_warp_params != NULL)
      memcpy(final_warp_params, gm_params, sizeof(*final_warp_params));
    return 1;
  }

  return 0;
}

#if CONFIG_EXT_IBC_MODES
void av1_intrabc_allocate_sb(uint16_t **InputBlock, const uint16_t width,
                             const uint16_t height) {
  (*InputBlock) = (uint16_t *)aom_malloc(width * height * sizeof(uint16_t));
}

void av1_fetch_prediction_sb(const uint8_t *src, int src_stride,
                             uint16_t *InputBlock, const uint16_t width,
                             const uint16_t height) {
  uint16_t *pixelStartAddr = CONVERT_TO_SHORTPTR(src);
  uint16_t size = (width << 1);

  // Populate prediction data
  for (int rows = 0; rows < height; ++rows) {
    memcpy(InputBlock, pixelStartAddr, size);
    pixelStartAddr += src_stride;
    InputBlock += width;
  }
}

void av1_write_prediction_sb(const uint8_t *dst, int dst_stride,
                             uint16_t *InputBlock, const uint16_t width,
                             const uint16_t height) {
  uint16_t *pixelStartAddr = CONVERT_TO_SHORTPTR(dst);
  uint16_t size = (width << 1);

  // Write prediction data
  for (int rows = 0; rows < height; ++rows) {
    memcpy(pixelStartAddr, InputBlock, size);
    InputBlock += 128;
    pixelStartAddr += dst_stride;
  }
}

void av1_intrabc_rotate90_sb(uint16_t *DstBlock, uint16_t *SrcBlock,
                             const uint16_t width, const uint16_t height) {
  // Rotate Block by 90 degrees
  for (int rows = 0; rows < height; ++rows) {
    for (int cols = 0; cols < width; ++cols) {
      DstBlock[cols * 128 + (height - 1 - rows)] = SrcBlock[cols];
    }
    SrcBlock += width;
  }
}

void av1_intrabc_rotate180_sb(uint16_t *DstBlock, uint16_t *SrcBlock,
                              const uint16_t width, const uint16_t height) {
  // Rotate Block by 180 degrees
  for (int rows = 0; rows < height; ++rows) {
    for (int cols = 0; cols < width; ++cols) {
      DstBlock[(height - 1 - rows) * 128 + (width - 1 - cols)] = SrcBlock[cols];
    }
    SrcBlock += width;
  }
}

void av1_intrabc_rotate270_sb(uint16_t *DstBlock, uint16_t *SrcBlock,
                              const uint16_t width, const uint16_t height) {
  // Rotate Block by 270 degrees
  for (int rows = 0; rows < height; ++rows) {
    for (int cols = 0; cols < width; ++cols) {
      DstBlock[(width - 1 - cols) * 128 + rows] = SrcBlock[cols];
    }
    SrcBlock += width;
  }
}

void av1_intrabc_mirror0_sb(uint16_t *DstBlock, uint16_t *SrcBlock,
                            const uint16_t width, const uint16_t height) {
  uint16_t size = (width << 1);

  // Mirror Block across the 0 degree axis
  DstBlock += (height - 1) * 128;
  for (int rows = 0; rows < height; ++rows) {
    memcpy(DstBlock, SrcBlock, size);
    DstBlock -= 128;
    SrcBlock += width;
  }
}

void av1_intrabc_mirror45_sb(uint16_t *DstBlock, uint16_t *SrcBlock,
                             const uint16_t width, const uint16_t height) {
  // Mirror Block across the 45 degree axis
  for (int rows = 0; rows < height; ++rows) {
    for (int cols = 0; cols < width; ++cols) {
      DstBlock[(width - 1 - cols) * 128 + (height - 1 - rows)] = SrcBlock[cols];
    }
    SrcBlock += width;
  }
}

void av1_intrabc_mirror90_sb(uint16_t *DstBlock, uint16_t *SrcBlock,
                             const uint16_t width, const uint16_t height) {
  // Mirror Block across the 90 degree axis
  for (int rows = 0; rows < height; ++rows) {
    for (int cols = 0; cols < width; ++cols) {
      DstBlock[rows * 128 + (width - 1 - cols)] = SrcBlock[cols];
    }
    SrcBlock += width;
  }
}

void av1_intrabc_mirror135_sb(uint16_t *DstBlock, uint16_t *SrcBlock,
                              const uint16_t width, const uint16_t height) {
  // Mirror Block across the 135 degree axis
  for (int rows = 0; rows < height; ++rows) {
    for (int cols = 0; cols < width; ++cols) {
      DstBlock[cols * 128 + rows] = SrcBlock[cols];
    }
    SrcBlock += width;
  }
}
#endif  // CONFIG_EXT_IBC_MODES

#if CONFIG_OPTFLOW_REFINEMENT

// Apply distance based weighted compound average
#define OPFL_DISTWTD_AVG 1

// Precision of refined MV returned, 0 being integer pel.
#define MV_REFINE_PREC_BITS 4  // (1/16-pel)

// Delta to use for computing gradients in bits, with 0 referring to
// integer-pel. The actual delta value used from the 1/8-pel original MVs
// is 2^(3 - SUBPEL_GRAD_DELTA_BITS). The max value of this macro is 3.
// TODO(debargha@, kslu@): experiment with values 0, 1, 2
#define SUBPEL_GRAD_DELTA_BITS 3
// Note: grad_prec_bits param returned correspond to the precision
// of the gradient information in bits assuming gradient
// computed at unit pixel step normalization is 0 scale.
// Negative values indicate gradient returned at reduced precision, and
// positive values indicate gradient returned at higher precision.
int av1_compute_subpel_gradients_highbd(
    const AV1_COMMON *cm, MACROBLOCKD *xd, int plane, const MB_MODE_INFO *mi,
    int bw, int bh, int mi_x, int mi_y,
    CalcSubpelParamsFunc calc_subpel_params_func,
    const void *const calc_subpel_params_func_args, int ref,
    uint16_t *pred_dst16, int *grad_prec_bits, int16_t *x_grad,
    int16_t *y_grad) {
  assert(cm->seq_params.order_hint_info.enable_order_hint);
  *grad_prec_bits = INT_MAX;

  uint8_t *pred_dst = CONVERT_TO_BYTEPTR(pred_dst16);
  // Compute distance between the current frame and reference
  const int cur_frame_index = cm->cur_frame->order_hint;
  const RefCntBuffer *const ref_buf = get_ref_frame_buf(cm, mi->ref_frame[ref]);
  assert(ref_buf != NULL);
  const int ref_index = ref_buf->order_hint;
  // Find the distance in display order between the current frame and each
  // reference
  const int r_dist = get_relative_dist(&cm->seq_params.order_hint_info,
                                       cur_frame_index, ref_index);

  // Do references one at a time
  const int is_compound = 0;
  ConvolveParams conv_params = get_conv_params_no_round(
      0, plane, xd->tmp_conv_dst, MAX_SB_SIZE, is_compound, xd->bd);
  av1_dist_wtd_comp_weight_assign(
      cm, mi, 0, &conv_params.fwd_offset, &conv_params.bck_offset,
      &conv_params.use_dist_wtd_comp_avg, is_compound);

  struct macroblockd_plane *const pd = &xd->plane[plane];
  struct buf_2d *const dst_buf = &pd->dst;
  const int is_intrabc = is_intrabc_block(mi);
  uint16_t tmp_buf1[MAX_SB_SIZE * MAX_SB_SIZE] = { 0 };
  uint16_t tmp_buf2[MAX_SB_SIZE * MAX_SB_SIZE] = { 0 };
  uint8_t *tmp_buf1_8 = CONVERT_TO_BYTEPTR(tmp_buf1);
  uint8_t *tmp_buf2_8 = CONVERT_TO_BYTEPTR(tmp_buf2);

  int is_global[2] = { 0, 0 };
  const WarpedMotionParams *const wm = &xd->global_motion[mi->ref_frame[ref]];
  is_global[ref] = is_global_mv_block(mi, wm->wmtype);
  const WarpTypesAllowed warp_types = { is_global[ref],
                                        mi->motion_mode == WARPED_CAUSAL };
  const struct scale_factors *const sf =
      is_intrabc ? &cm->sf_identity : xd->block_ref_scale_factors[ref];
  WarpedMotionParams final_warp_params;
  const int do_warp =
      (bw >= 8 && bh >= 8 &&
       av1_allow_warp(mi, &warp_types, &xd->global_motion[mi->ref_frame[ref]],
                      0, sf, &final_warp_params));
  // TODO(sarahparker) make compatible with warped modes
  if (do_warp || !r_dist) return 0;

  const int ss_x = pd->subsampling_x;
  const int ss_y = pd->subsampling_y;
  const int mi_row = -xd->mb_to_top_edge >> (3 + MI_SIZE_LOG2);
  const int mi_col = -xd->mb_to_left_edge >> (3 + MI_SIZE_LOG2);
  int row_start = plane ? (mi->chroma_ref_info.mi_row_chroma_base - mi_row) : 0;
  int col_start = plane ? (mi->chroma_ref_info.mi_col_chroma_base - mi_col) : 0;
  const int pre_x = (mi_x + MI_SIZE * col_start) >> ss_x;
  const int pre_y = (mi_y + MI_SIZE * row_start) >> ss_y;

  struct buf_2d *const pre_buf = is_intrabc ? dst_buf : &pd->pre[ref];
  const MV mv_orig = mi->mv[ref].as_mv;
  MV mv_modified = mv_orig;

  uint8_t *pre;
  SubpelParams subpel_params;
  int src_stride;
  calc_subpel_params_func(
      xd, sf, &mv_orig, plane, pre_x, pre_y, 0, 0, pre_buf, bw, bh, &warp_types,
      ref, 0, calc_subpel_params_func_args, &pre, &subpel_params, &src_stride);

  // Original predictor
  assert(mi->interinter_comp.type == COMPOUND_AVERAGE);
  conv_params.do_average = 0;
  av1_make_inter_predictor(pre, src_stride, pred_dst, bw, &subpel_params, sf,
                           bw, bh, &conv_params, mi->interp_filters,
                           &warp_types, mi_x >> pd->subsampling_x,
                           mi_y >> pd->subsampling_y, plane, ref, mi, 0, xd,
                           cm->allow_warped_motion, 0 /* border */);

  // X gradient
  // Get predictor to the left
  mv_modified.col = mv_orig.col - (1 << (3 - SUBPEL_GRAD_DELTA_BITS));
  mv_modified.row = mv_orig.row;
  calc_subpel_params_func(xd, sf, &mv_modified, plane, pre_x, pre_y, 0, 0,
                          pre_buf, bw, bh, &warp_types, ref, 0,
                          calc_subpel_params_func_args, &pre, &subpel_params,
                          &src_stride);
  av1_make_inter_predictor(pre, src_stride, tmp_buf1_8, bw, &subpel_params, sf,
                           bw, bh, &conv_params, mi->interp_filters,
                           &warp_types, mi_x >> pd->subsampling_x,
                           mi_y >> pd->subsampling_y, plane, ref, mi, 0, xd,
                           cm->allow_warped_motion, 0 /* border */);
  // Get predictor to the right
  mv_modified.col = mv_orig.col + (1 << (3 - SUBPEL_GRAD_DELTA_BITS));
  mv_modified.row = mv_orig.row;
  calc_subpel_params_func(xd, sf, &mv_modified, plane, pre_x, pre_y, 0, 0,
                          pre_buf, bw, bh, &warp_types, ref, 0,
                          calc_subpel_params_func_args, &pre, &subpel_params,
                          &src_stride);
  av1_make_inter_predictor(pre, src_stride, tmp_buf2_8, bw, &subpel_params, sf,
                           bw, bh, &conv_params, mi->interp_filters,
                           &warp_types, mi_x >> pd->subsampling_x,
                           mi_y >> pd->subsampling_y, plane, ref, mi, 0, xd,
                           cm->allow_warped_motion, 0 /* border */);
  // Compute difference.
  // Note since the deltas are at +2^g/8 and -2^g/8 subpel locations
  // (g = 3 - SUBPEL_GRAD_DELTA_BITS), the actual unit pel gradient is
  // 4/2^g = 2^(2-g) times the difference. Therefore the gradient returned
  // is at reduced precision by 2-g bits. That explains the grad_prec_bits
  // return value of g-2 at the end of this function.
  for (int i = 0; i < bh; i++) {
    for (int j = 0; j < bw; j++) {
      x_grad[i * bw + j] = tmp_buf2[i * bw + j] - tmp_buf1[i * bw + j];
    }
  }

  // Y gradient
  // Get predictor below
  mv_modified.col = mv_orig.col;
  mv_modified.row = mv_orig.row - (1 << (3 - SUBPEL_GRAD_DELTA_BITS));
  calc_subpel_params_func(xd, sf, &mv_modified, plane, pre_x, pre_y, 0, 0,
                          pre_buf, bw, bh, &warp_types, ref, 0,
                          calc_subpel_params_func_args, &pre, &subpel_params,
                          &src_stride);
  av1_make_inter_predictor(pre, src_stride, tmp_buf1_8, bw, &subpel_params, sf,
                           bw, bh, &conv_params, mi->interp_filters,
                           &warp_types, mi_x >> pd->subsampling_x,
                           mi_y >> pd->subsampling_y, plane, ref, mi, 0, xd,
                           cm->allow_warped_motion, 0 /* border */);
  // Get predictor above
  mv_modified.col = mv_orig.col;
  mv_modified.row = mv_orig.row + (1 << (3 - SUBPEL_GRAD_DELTA_BITS));
  calc_subpel_params_func(xd, sf, &mv_modified, plane, pre_x, pre_y, 0, 0,
                          pre_buf, bw, bh, &warp_types, ref, 0,
                          calc_subpel_params_func_args, &pre, &subpel_params,
                          &src_stride);
  av1_make_inter_predictor(pre, src_stride, tmp_buf2_8, bw, &subpel_params, sf,
                           bw, bh, &conv_params, mi->interp_filters,
                           &warp_types, mi_x >> pd->subsampling_x,
                           mi_y >> pd->subsampling_y, plane, ref, mi, 0, xd,
                           cm->allow_warped_motion, 0 /* border */);
  // Compute difference.
  // Note since the deltas are at +2^g/8 and -2^g/8 subpel locations
  // (g = 3 - SUBPEL_GRAD_DELTA_BITS), the actual unit pel gradient is
  // 4/2^g = 2^(2-g) times the difference. Therefore the gradient returned
  // is at reduced precision by 2-g bits. That explains the grad_prec_bits
  // return value of g-2 at the end of this function.
  for (int i = 0; i < bh; i++) {
    for (int j = 0; j < bw; j++) {
      y_grad[i * bw + j] = tmp_buf2[i * bw + j] - tmp_buf1[i * bw + j];
    }
  }
  *grad_prec_bits = 3 - SUBPEL_GRAD_DELTA_BITS - 2;
  return r_dist;
}

// Note: grad_prec_bits param returned correspond to the precision
// of the gradient information in bits assuming gradient
// computed at unit pixel step normalization is 0 scale.
// Negative values indicate gradient returned at reduced precision, and
// positive values indicate gradient returned at higher precision.
int av1_compute_subpel_gradients_lowbd(
    const AV1_COMMON *cm, MACROBLOCKD *xd, int plane, const MB_MODE_INFO *mi,
    int bw, int bh, int mi_x, int mi_y,
    CalcSubpelParamsFunc calc_subpel_params_func,
    const void *const calc_subpel_params_func_args, int ref, uint8_t *pred_dst,
    int *grad_prec_bits, int16_t *x_grad, int16_t *y_grad) {
  assert(cm->seq_params.order_hint_info.enable_order_hint);
  *grad_prec_bits = INT_MAX;

  // Compute distance between the current frame and reference
  const int cur_frame_index = cm->cur_frame->order_hint;
  const RefCntBuffer *const ref_buf = get_ref_frame_buf(cm, mi->ref_frame[ref]);
  assert(ref_buf != NULL);
  const int ref_index = ref_buf->order_hint;
  // Find the distance in display order between the current frame and each
  // reference
  const int r_dist = get_relative_dist(&cm->seq_params.order_hint_info,
                                       cur_frame_index, ref_index);

  // Do references one at a time
  const int is_compound = 0;
  ConvolveParams conv_params = get_conv_params_no_round(
      0, plane, xd->tmp_conv_dst, MAX_SB_SIZE, is_compound, xd->bd);
  av1_dist_wtd_comp_weight_assign(
      cm, mi, 0, &conv_params.fwd_offset, &conv_params.bck_offset,
      &conv_params.use_dist_wtd_comp_avg, is_compound);

  struct macroblockd_plane *const pd = &xd->plane[plane];
  struct buf_2d *const dst_buf = &pd->dst;
  const int is_intrabc = is_intrabc_block(mi);
  uint8_t tmp_buf1[MAX_SB_SIZE * MAX_SB_SIZE] = { 0 };
  uint8_t tmp_buf2[MAX_SB_SIZE * MAX_SB_SIZE] = { 0 };

  int is_global[2] = { 0, 0 };
  const WarpedMotionParams *const wm = &xd->global_motion[mi->ref_frame[ref]];
  is_global[ref] = is_global_mv_block(mi, wm->wmtype);
  const WarpTypesAllowed warp_types = { is_global[ref],
                                        mi->motion_mode == WARPED_CAUSAL };
  const struct scale_factors *const sf =
      is_intrabc ? &cm->sf_identity : xd->block_ref_scale_factors[ref];
  WarpedMotionParams final_warp_params;
  const int do_warp =
      (bw >= 8 && bh >= 8 &&
       av1_allow_warp(mi, &warp_types, &xd->global_motion[mi->ref_frame[ref]],
                      0, sf, &final_warp_params));
  // TODO(sarahparker) make compatible with warped modes
  if (do_warp || !r_dist) return 0;

  const int ss_x = pd->subsampling_x;
  const int ss_y = pd->subsampling_y;
  const int mi_row = -xd->mb_to_top_edge >> (3 + MI_SIZE_LOG2);
  const int mi_col = -xd->mb_to_left_edge >> (3 + MI_SIZE_LOG2);
  int row_start = plane ? (mi->chroma_ref_info.mi_row_chroma_base - mi_row) : 0;
  int col_start = plane ? (mi->chroma_ref_info.mi_col_chroma_base - mi_col) : 0;
  const int pre_x = (mi_x + MI_SIZE * col_start) >> ss_x;
  const int pre_y = (mi_y + MI_SIZE * row_start) >> ss_y;

  struct buf_2d *const pre_buf = is_intrabc ? dst_buf : &pd->pre[ref];
  const MV mv_orig = mi->mv[ref].as_mv;
  MV mv_modified = mv_orig;

  uint8_t *pre;
  SubpelParams subpel_params;
  int src_stride;
  calc_subpel_params_func(
      xd, sf, &mv_orig, plane, pre_x, pre_y, 0, 0, pre_buf, bw, bh, &warp_types,
      ref, 0, calc_subpel_params_func_args, &pre, &subpel_params, &src_stride);

  // Original predictor
  assert(mi->interinter_comp.type == COMPOUND_AVERAGE);
  conv_params.do_average = 0;
  av1_make_inter_predictor(pre, src_stride, pred_dst, bw, &subpel_params, sf,
                           bw, bh, &conv_params, mi->interp_filters,
                           &warp_types, mi_x >> pd->subsampling_x,
                           mi_y >> pd->subsampling_y, plane, ref, mi, 0, xd,
                           cm->allow_warped_motion, 0 /* border */);

  // X gradient
  // Get predictor to the left
  mv_modified.col = mv_orig.col - (1 << (3 - SUBPEL_GRAD_DELTA_BITS));
  mv_modified.row = mv_orig.row;
  calc_subpel_params_func(xd, sf, &mv_modified, plane, pre_x, pre_y, 0, 0,
                          pre_buf, bw, bh, &warp_types, ref, 0,
                          calc_subpel_params_func_args, &pre, &subpel_params,
                          &src_stride);
  av1_make_inter_predictor(pre, src_stride, tmp_buf1, bw, &subpel_params, sf,
                           bw, bh, &conv_params, mi->interp_filters,
                           &warp_types, mi_x >> pd->subsampling_x,
                           mi_y >> pd->subsampling_y, plane, ref, mi, 0, xd,
                           cm->allow_warped_motion, 0 /* border */);
  // Get predictor to the right
  mv_modified.col = mv_orig.col + (1 << (3 - SUBPEL_GRAD_DELTA_BITS));
  mv_modified.row = mv_orig.row;
  calc_subpel_params_func(xd, sf, &mv_modified, plane, pre_x, pre_y, 0, 0,
                          pre_buf, bw, bh, &warp_types, ref, 0,
                          calc_subpel_params_func_args, &pre, &subpel_params,
                          &src_stride);
  av1_make_inter_predictor(pre, src_stride, tmp_buf2, bw, &subpel_params, sf,
                           bw, bh, &conv_params, mi->interp_filters,
                           &warp_types, mi_x >> pd->subsampling_x,
                           mi_y >> pd->subsampling_y, plane, ref, mi, 0, xd,
                           cm->allow_warped_motion, 0 /* border */);
  // Compute difference.
  // Note since the deltas are at +2^g/8 and -2^g/8 subpel locations
  // (g = 3 - SUBPEL_GRAD_DELTA_BITS), the actual unit pel gradient is
  // 4/2^g = 2^(2-g) times the difference. Therefore the gradient returned
  // is at reduced precision by 2-g bits. That explains the grad_prec_bits
  // return value of g-2 at the end of this function.
  for (int i = 0; i < bh; i++) {
    for (int j = 0; j < bw; j++) {
      x_grad[i * bw + j] =
          (int16_t)tmp_buf2[i * bw + j] - (int16_t)tmp_buf1[i * bw + j];
    }
  }

  // Y gradient
  // Get predictor below
  mv_modified.col = mv_orig.col;
  mv_modified.row = mv_orig.row - (1 << (3 - SUBPEL_GRAD_DELTA_BITS));
  calc_subpel_params_func(xd, sf, &mv_modified, plane, pre_x, pre_y, 0, 0,
                          pre_buf, bw, bh, &warp_types, ref, 0,
                          calc_subpel_params_func_args, &pre, &subpel_params,
                          &src_stride);
  av1_make_inter_predictor(pre, src_stride, tmp_buf1, bw, &subpel_params, sf,
                           bw, bh, &conv_params, mi->interp_filters,
                           &warp_types, mi_x >> pd->subsampling_x,
                           mi_y >> pd->subsampling_y, plane, ref, mi, 0, xd,
                           cm->allow_warped_motion, 0 /* border */);
  // Get predictor above
  mv_modified.col = mv_orig.col;
  mv_modified.row = mv_orig.row + (1 << (3 - SUBPEL_GRAD_DELTA_BITS));
  calc_subpel_params_func(xd, sf, &mv_modified, plane, pre_x, pre_y, 0, 0,
                          pre_buf, bw, bh, &warp_types, ref, 0,
                          calc_subpel_params_func_args, &pre, &subpel_params,
                          &src_stride);
  av1_make_inter_predictor(pre, src_stride, tmp_buf2, bw, &subpel_params, sf,
                           bw, bh, &conv_params, mi->interp_filters,
                           &warp_types, mi_x >> pd->subsampling_x,
                           mi_y >> pd->subsampling_y, plane, ref, mi, 0, xd,
                           cm->allow_warped_motion, 0 /* border */);
  // Compute difference.
  // Note since the deltas are at +2^g/8 and -2^g/8 subpel locations
  // (g = 3 - SUBPEL_GRAD_DELTA_BITS), the actual unit pel gradient is
  // 4/2^g = 2^(2-g) times the difference. Therefore the gradient returned
  // is at reduced precision by 2-g bits. That explains the grad_prec_bits
  // return value of g-2 at the end of this function.
  for (int i = 0; i < bh; i++) {
    for (int j = 0; j < bw; j++) {
      y_grad[i * bw + j] =
          (int16_t)tmp_buf2[i * bw + j] - (int16_t)tmp_buf1[i * bw + j];
    }
  }
  *grad_prec_bits = 3 - SUBPEL_GRAD_DELTA_BITS - 2;
  return r_dist;
}

// Optical flow based mv refinement computation function:
//
// p0, pstride0: predictor 0 and its stride
// p1, pstride1: predictor 1 and its stride
// gx0, gy0: x and y gradients for p0
// gx1, gy1: x and y gradients for p1
// gstride: stride for all the gradients assumed to be the same
// bw, bh: block dumensions
// d0: distance of p0 to current frame, where +ve value refers to
//     p0 before the current frame.
// d1: distance of p1 to current frame, where +ve value refers to
//     p1 after the current frame.
// max_prec_bits: maximum offset in bits
// vx0, vy0: output high resolution mv offset for p0
// vx1, vy1: output high resolution mv offset for p1

void av1_opfl_mv_refinement_lowbd(const uint8_t *p0, int pstride0,
                                  const uint8_t *p1, int pstride1,
                                  const int16_t *gx0, const int16_t *gy0,
                                  const int16_t *gx1, const int16_t *gy1,
                                  int gstride, int bw, int bh, int d0, int d1,
                                  int grad_prec_bits, int mv_prec_bits,
                                  int *vx0, int *vy0, int *vx1, int *vy1) {
  int64_t su2 = 0;
  int64_t suv = 0;
  int64_t sv2 = 0;
  int64_t suw = 0;
  int64_t svw = 0;
  for (int i = 0; i < bh; ++i) {
    for (int j = 0; j < bw; ++j) {
      const int u = d0 * gx0[i * gstride + j] - d1 * gx1[i * gstride + j];
      const int v = d0 * gy0[i * gstride + j] - d1 * gy1[i * gstride + j];
      const int w = d0 * (p0[i * pstride0 + j] - p1[i * pstride1 + j]);
      su2 += (u * u);
      suv += (u * v);
      sv2 += (v * v);
      suw += (u * w);
      svw += (v * w);
    }
  }
  int bits = mv_prec_bits + grad_prec_bits;
  const int64_t D = su2 * sv2 - suv * suv;
  const int64_t Px = (suv * svw - sv2 * suw) * (1 << bits);
  const int64_t Py = (suv * suw - su2 * svw) * (1 << bits);

  if (D == 0) return;
  *vx0 = (int)DIVIDE_AND_ROUND_SIGNED(Px, D);
  *vy0 = (int)DIVIDE_AND_ROUND_SIGNED(Py, D);
  const int tx1 = (*vx0) * d1;
  const int ty1 = (*vy0) * d1;
  *vx1 = (int)DIVIDE_AND_ROUND_SIGNED(tx1, d0);
  *vy1 = (int)DIVIDE_AND_ROUND_SIGNED(ty1, d0);
}

void av1_opfl_mv_refinement_highbd(const uint16_t *p0, int pstride0,
                                   const uint16_t *p1, int pstride1,
                                   const int16_t *gx0, const int16_t *gy0,
                                   const int16_t *gx1, const int16_t *gy1,
                                   int gstride, int bw, int bh, int d0, int d1,
                                   int grad_prec_bits, int mv_prec_bits,
                                   int *vx0, int *vy0, int *vx1, int *vy1) {
  int64_t su2 = 0;
  int64_t suv = 0;
  int64_t sv2 = 0;
  int64_t suw = 0;
  int64_t svw = 0;
  for (int i = 0; i < bh; ++i) {
    for (int j = 0; j < bw; ++j) {
      const int u = d0 * gx0[i * gstride + j] - d1 * gx1[i * gstride + j];
      const int v = d0 * gy0[i * gstride + j] - d1 * gy1[i * gstride + j];
      const int w = d0 * (p0[i * pstride0 + j] - p1[i * pstride1 + j]);
      su2 += (u * u);
      suv += (u * v);
      sv2 += (v * v);
      suw += (u * w);
      svw += (v * w);
    }
  }
  int bits = mv_prec_bits + grad_prec_bits;
  const int64_t D = su2 * sv2 - suv * suv;
  const int64_t Px = (suv * svw - sv2 * suw) * (1 << bits);
  const int64_t Py = (suv * suw - su2 * svw) * (1 << bits);

  if (D == 0) return;
  *vx0 = (int)DIVIDE_AND_ROUND_SIGNED(Px, D);
  *vy0 = (int)DIVIDE_AND_ROUND_SIGNED(Py, D);
  const int tx1 = (*vx0) * d1;
  const int ty1 = (*vy0) * d1;
  *vx1 = (int)DIVIDE_AND_ROUND_SIGNED(tx1, d0);
  *vy1 = (int)DIVIDE_AND_ROUND_SIGNED(ty1, d0);
}

void getsub_4D(int64_t *sub, int64_t *mat, int64_t *vec) {
  *(sub++) = *(mat) * *(mat + 5) - *(mat + 1) * *(mat + 4);
  *(sub++) = *(mat) * *(mat + 6) - *(mat + 2) * *(mat + 4);
  *(sub++) = *(mat) * *(mat + 7) - *(mat + 3) * *(mat + 4);
  *(sub++) = *(mat) * *(vec + 1) - *(vec) * *(mat + 4);
  *(sub++) = *(mat + 1) * *(mat + 6) - *(mat + 2) * *(mat + 5);
  *(sub++) = *(mat + 1) * *(mat + 7) - *(mat + 3) * *(mat + 5);
  *(sub++) = *(mat + 1) * *(vec + 1) - *(vec) * *(mat + 5);
  *(sub++) = *(mat + 2) * *(mat + 7) - *(mat + 3) * *(mat + 6);
  *(sub++) = *(mat + 2) * *(vec + 1) - *(vec) * *(mat + 6);
  *(sub) = *(mat + 3) * *(vec + 1) - *(vec) * *(mat + 7);
}

int solver_4D(int64_t *mat, int64_t *vec, int precbits, int64_t *sol) {
  int64_t a[10], b[10];  // values of 20 specific 2D subdeterminants

  getsub_4D(&a[0], mat, vec);
  getsub_4D(&b[0], mat + 8, vec + 2);
  int64_t D = (a[0] * b[7] + a[7] * b[0] + a[2] * b[4] + a[4] * b[2] -
               a[5] * b[1] - a[1] * b[5]);
  if (D == 0) return 0;

  sol[0] = (a[5] * b[8] + a[8] * b[5] - a[6] * b[7] - a[7] * b[6] -
            a[4] * b[9] - a[9] * b[4])
           << precbits;
  sol[1] = (a[1] * b[9] + a[9] * b[1] + a[3] * b[7] + a[7] * b[3] -
            a[2] * b[8] - a[8] * b[2])
           << precbits;
  sol[2] = (a[2] * b[6] + a[6] * b[2] - a[0] * b[9] - a[9] * b[0] -
            a[3] * b[5] - a[5] * b[3])
           << precbits;
  sol[3] = (a[0] * b[8] + a[8] * b[0] + a[3] * b[4] + a[4] * b[3] -
            a[6] * b[1] - a[1] * b[6])
           << precbits;

  sol[0] = DIVIDE_AND_ROUND_SIGNED(sol[0], D);
  sol[1] = DIVIDE_AND_ROUND_SIGNED(sol[1], D);
  sol[2] = DIVIDE_AND_ROUND_SIGNED(sol[2], D);
  sol[3] = DIVIDE_AND_ROUND_SIGNED(sol[3], D);
  return 1;
}

void av1_opfl_mv_refinement4_lowbd(const uint8_t *p0, int pstride0,
                                   const uint8_t *p1, int pstride1,
                                   const int16_t *gx0, const int16_t *gy0,
                                   const int16_t *gx1, const int16_t *gy1,
                                   int gstride, int bw, int bh, int d0, int d1,
                                   int grad_prec_bits, int mv_prec_bits,
                                   int *vx0, int *vy0, int *vx1, int *vy1) {
  int64_t A[4 * 4] = { 0 };
  int64_t B[4] = { 0 };
  int64_t X[4];
  for (int i = 0; i < bh; ++i) {
    for (int j = 0; j < bw; ++j) {
      int a[4];
      a[0] = d0 * gx0[i * gstride + j];
      a[1] = d0 * gy0[i * gstride + j];
      a[2] = (-d1) * gx1[i * gstride + j];
      a[3] = (-d1) * gy1[i * gstride + j];
      const int d = p1[i * pstride1 + j] - p0[i * pstride0 + j];
      for (int s = 0; s < 4; ++s) {
        for (int t = 0; t <= s; ++t) A[s * 4 + t] += (a[s] * a[t]);
        B[s] += (a[s] * d);
      }
    }
  }
  for (int s = 0; s < 4; ++s) {
    for (int t = s + 1; t < 4; ++t) A[s * 4 + t] = A[t * 4 + s];
  }

  int bits = mv_prec_bits + grad_prec_bits;

  if (!solver_4D(A, B, bits, X)) return;
  *vx0 = d0 * (int)X[0];
  *vy0 = d0 * (int)X[1];
  *vx1 = d1 * (int)X[2];
  *vy1 = d1 * (int)X[3];
}

void av1_opfl_mv_refinement4_highbd(const uint16_t *p0, int pstride0,
                                    const uint16_t *p1, int pstride1,
                                    const int16_t *gx0, const int16_t *gy0,
                                    const int16_t *gx1, const int16_t *gy1,
                                    int gstride, int bw, int bh, int d0, int d1,
                                    int grad_prec_bits, int mv_prec_bits,
                                    int *vx0, int *vy0, int *vx1, int *vy1) {
  int64_t A[4 * 4] = { 0 };
  int64_t B[4] = { 0 };
  int64_t X[4];
  for (int i = 0; i < bh; ++i) {
    for (int j = 0; j < bw; ++j) {
      int a[4];
      a[0] = d0 * gx0[i * gstride + j];
      a[1] = d0 * gy0[i * gstride + j];
      a[2] = (-d1) * gx1[i * gstride + j];
      a[3] = (-d1) * gy1[i * gstride + j];
      const int d = p1[i * pstride1 + j] - p0[i * pstride0 + j];
      for (int s = 0; s < 4; ++s) {
        for (int t = 0; t <= s; ++t) A[s * 4 + t] += (a[s] * a[t]);
        B[s] += (a[s] * d);
      }
    }
  }
  for (int s = 0; s < 4; ++s) {
    for (int t = s + 1; t < 4; ++t) A[s * 4 + t] = A[t * 4 + s];
  }

  int bits = mv_prec_bits + grad_prec_bits;
  if (!solver_4D(A, B, bits, X)) return;

  *vx0 = d0 * (int)X[0];
  *vy0 = d0 * (int)X[1];
  *vx1 = d1 * (int)X[2];
  *vy1 = d1 * (int)X[3];
}

#if USE_OF_NXN
// Function to compute optical flow offsets in nxn blocks
int opfl_mv_refinement_nxn_highbd(const uint16_t *p0, int pstride0,
                                  const uint16_t *p1, int pstride1,
                                  const int16_t *gx0, const int16_t *gy0,
                                  const int16_t *gx1, const int16_t *gy1,
                                  int gstride, int bw, int bh, int n, int d0,
                                  int d1, int grad_prec_bits, int mv_prec_bits,
                                  int *vx0, int *vy0, int *vx1, int *vy1) {
  assert(bw % n == 0 && bh % n == 0);
  int n_blocks = 0;
  for (int i = 0; i < bh; i += n) {
    for (int j = 0; j < bw; j += n) {
      av1_opfl_mv_refinement_highbd(
          p0 + (i * pstride0 + j), pstride0, p1 + (i * pstride1 + j), pstride1,
          gx0 + (i * gstride + j), gy0 + (i * gstride + j),
          gx1 + (i * gstride + j), gy1 + (i * gstride + j), gstride, n, n, d0,
          d1, grad_prec_bits, mv_prec_bits, vx0 + n_blocks, vy0 + n_blocks,
          vx1 + n_blocks, vy1 + n_blocks);
      n_blocks++;
    }
  }
  return n_blocks;
}
// Function to compute optical flow offsets in nxn blocks
int opfl_mv_refinement_nxn_lowbd(const uint8_t *p0, int pstride0,
                                 const uint8_t *p1, int pstride1,
                                 const int16_t *gx0, const int16_t *gy0,
                                 const int16_t *gx1, const int16_t *gy1,
                                 int gstride, int bw, int bh, int n, int d0,
                                 int d1, int grad_prec_bits, int mv_prec_bits,
                                 int *vx0, int *vy0, int *vx1, int *vy1) {
  assert(bw % n == 0 && bh % n == 0);
  int n_blocks = 0;
  for (int i = 0; i < bh; i += n) {
    for (int j = 0; j < bw; j += n) {
      av1_opfl_mv_refinement_lowbd(
          p0 + (i * pstride0 + j), pstride0, p1 + (i * pstride1 + j), pstride1,
          gx0 + (i * gstride + j), gy0 + (i * gstride + j),
          gx1 + (i * gstride + j), gy1 + (i * gstride + j), gstride, n, n, d0,
          d1, grad_prec_bits, mv_prec_bits, vx0 + n_blocks, vy0 + n_blocks,
          vx1 + n_blocks, vy1 + n_blocks);
      n_blocks++;
    }
  }
  return n_blocks;
}
#endif  // USE_OF_NXN

static int get_optflow_based_mv_highbd(
    const AV1_COMMON *cm, MACROBLOCKD *xd, int plane, const MB_MODE_INFO *mbmi,
    int_mv *mv_refined, int bw, int bh, int mi_x, int mi_y,
    CalcSubpelParamsFunc calc_subpel_params_func,
    const void *const calc_subpel_params_func_args) {
  // Arrays to hold optical flow offsets. If the experiment USE_OF_NXN is
  // off, these will only be length 1
  int vx0[N_OF_OFFSETS] = { 0 };
  int vx1[N_OF_OFFSETS] = { 0 };
  int vy0[N_OF_OFFSETS] = { 0 };
  int vy1[N_OF_OFFSETS] = { 0 };
  const int target_prec = MV_REFINE_PREC_BITS;
  // Convert output MV to 1/16th pel
  for (int mvi = 0; mvi < N_OF_OFFSETS; mvi++) {
    mv_refined[mvi * 2].as_mv.row *= 2;
    mv_refined[mvi * 2].as_mv.col *= 2;
    mv_refined[mvi * 2 + 1].as_mv.row *= 2;
    mv_refined[mvi * 2 + 1].as_mv.col *= 2;
  }

  // Allocate gradient and prediction buffers
  int16_t *g0 = aom_malloc(2 * MAX_SB_SIZE * MAX_SB_SIZE * sizeof(*g0));
  memset(g0, 0, 2 * MAX_SB_SIZE * MAX_SB_SIZE * sizeof(*g0));
  uint16_t *dst0 = aom_malloc(MAX_SB_SIZE * MAX_SB_SIZE * sizeof(*dst0));
  memset(dst0, 0, MAX_SB_SIZE * MAX_SB_SIZE * sizeof(*dst0));
  int16_t *g1 = aom_malloc(2 * MAX_SB_SIZE * MAX_SB_SIZE * sizeof(*g1));
  memset(g1, 0, 2 * MAX_SB_SIZE * MAX_SB_SIZE * sizeof(*g1));
  uint16_t *dst1 = aom_malloc(MAX_SB_SIZE * MAX_SB_SIZE * sizeof(*dst1));
  memset(dst1, 0, MAX_SB_SIZE * MAX_SB_SIZE * sizeof(*dst1));

  int16_t *gx0 = g0;
  int16_t *gy0 = g0 + (MAX_SB_SIZE * MAX_SB_SIZE);
  int16_t *gx1 = g1;
  int16_t *gy1 = g1 + (MAX_SB_SIZE * MAX_SB_SIZE);
  int n_blocks = 1;
  int grad_prec_bits;

  // Compute gradients and predictor for P0
  int d0 = av1_compute_subpel_gradients_highbd(
      cm, xd, plane, mbmi, bw, bh, mi_x, mi_y, calc_subpel_params_func,
      calc_subpel_params_func_args, 0, dst0, &grad_prec_bits, gx0, gy0);
  if (d0 == 0) goto exit_refinement;

  // Compute gradients and predictor for P1
  int d1 = av1_compute_subpel_gradients_highbd(
      cm, xd, plane, mbmi, bw, bh, mi_x, mi_y, calc_subpel_params_func,
      calc_subpel_params_func_args, 1, dst1, &grad_prec_bits, gx1, gy1);
  if (d1 == 0) goto exit_refinement;

#if USE_OF_NXN
  int n = (bh <= 16 && bw <= 16) ? OF_MIN_BSIZE : OF_BSIZE;
  n_blocks = opfl_mv_refinement_nxn_highbd(
      dst0, bw, dst1, bw, gx0, gy0, gx1, gy1, bw, bw, bh, n, d0, d1,
      grad_prec_bits, target_prec, vx0, vy0, vx1, vy1);
#else
  av1_opfl_mv_refinement_highbd(dst0, bw, dst1, bw, gx0, gy0, gx1, gy1, bw, bw,
                                bh, d0, d1, grad_prec_bits, target_prec, vx0,
                                vy0, vx1, vy1);
#endif

  for (int i = 0; i < n_blocks; i++) {
    mv_refined[i * 2].as_mv.row += vy0[i];
    mv_refined[i * 2].as_mv.col += vx0[i];
    mv_refined[i * 2 + 1].as_mv.row += vy1[i];
    mv_refined[i * 2 + 1].as_mv.col += vx1[i];
  }

exit_refinement:
  aom_free(g0);
  aom_free(dst0);
  aom_free(g1);
  aom_free(dst1);
  return target_prec;
}

static int get_optflow_based_mv_lowbd(
    const AV1_COMMON *cm, MACROBLOCKD *xd, int plane, const MB_MODE_INFO *mbmi,
    int_mv *mv_refined, int bw, int bh, int mi_x, int mi_y,
    CalcSubpelParamsFunc calc_subpel_params_func,
    const void *const calc_subpel_params_func_args) {
  // Arrays to hold optical flow offsets. If the experiment USE_OF_NXN is
  // off, these will only be length 1
  int vx0[N_OF_OFFSETS] = { 0 };
  int vx1[N_OF_OFFSETS] = { 0 };
  int vy0[N_OF_OFFSETS] = { 0 };
  int vy1[N_OF_OFFSETS] = { 0 };
  const int target_prec = MV_REFINE_PREC_BITS;
  // Convert output MV to 1/16th pel
  for (int mvi = 0; mvi < N_OF_OFFSETS; mvi++) {
    mv_refined[mvi * 2].as_mv.row *= 2;
    mv_refined[mvi * 2].as_mv.col *= 2;
    mv_refined[mvi * 2 + 1].as_mv.row *= 2;
    mv_refined[mvi * 2 + 1].as_mv.col *= 2;
  }

  // Allocate gradient and prediction buffers
  int16_t *g0 = aom_malloc(2 * MAX_SB_SIZE * MAX_SB_SIZE * sizeof(*g0));
  memset(g0, 0, 2 * MAX_SB_SIZE * MAX_SB_SIZE * sizeof(*g0));
  uint8_t *dst0 = aom_malloc(MAX_SB_SIZE * MAX_SB_SIZE * sizeof(*dst0));
  memset(dst0, 0, MAX_SB_SIZE * MAX_SB_SIZE * sizeof(*dst0));
  int16_t *g1 = aom_malloc(2 * MAX_SB_SIZE * MAX_SB_SIZE * sizeof(*g1));
  memset(g1, 0, 2 * MAX_SB_SIZE * MAX_SB_SIZE * sizeof(*g1));
  uint8_t *dst1 = aom_malloc(MAX_SB_SIZE * MAX_SB_SIZE * sizeof(*dst1));
  memset(dst1, 0, MAX_SB_SIZE * MAX_SB_SIZE * sizeof(*dst1));

  int16_t *gx0 = g0;
  int16_t *gy0 = g0 + (MAX_SB_SIZE * MAX_SB_SIZE);
  int16_t *gx1 = g1;
  int16_t *gy1 = g1 + (MAX_SB_SIZE * MAX_SB_SIZE);
  int n_blocks = 1;
  int grad_prec_bits;

  // Compute gradients and predictor for P0
  int d0 = av1_compute_subpel_gradients_lowbd(
      cm, xd, plane, mbmi, bw, bh, mi_x, mi_y, calc_subpel_params_func,
      calc_subpel_params_func_args, 0, dst0, &grad_prec_bits, gx0, gy0);
  if (d0 == 0) goto exit_refinement;

  // Compute gradients and predictor for P1
  int d1 = av1_compute_subpel_gradients_lowbd(
      cm, xd, plane, mbmi, bw, bh, mi_x, mi_y, calc_subpel_params_func,
      calc_subpel_params_func_args, 1, dst1, &grad_prec_bits, gx1, gy1);
  if (d1 == 0) goto exit_refinement;

#if USE_OF_NXN
  int n = (bh <= 16 && bw <= 16) ? OF_MIN_BSIZE : OF_BSIZE;
  n_blocks = opfl_mv_refinement_nxn_lowbd(
      dst0, bw, dst1, bw, gx0, gy0, gx1, gy1, bw, bw, bh, n, d0, d1,
      grad_prec_bits, target_prec, vx0, vy0, vx1, vy1);
#else
  av1_opfl_mv_refinement_lowbd(dst0, bw, dst1, bw, gx0, gy0, gx1, gy1, bw, bw,
                               bh, d0, d1, grad_prec_bits, target_prec, vx0,
                               vy0, vx1, vy1);
#endif

  for (int i = 0; i < n_blocks; i++) {
    mv_refined[i * 2].as_mv.row += vy0[i];
    mv_refined[i * 2].as_mv.col += vx0[i];
    mv_refined[i * 2 + 1].as_mv.row += vy1[i];
    mv_refined[i * 2 + 1].as_mv.col += vx1[i];
  }

exit_refinement:
  aom_free(g0);
  aom_free(dst0);
  aom_free(g1);
  aom_free(dst1);
  return target_prec;
}

// Refine MV using optical flow. The final output MV will be in 1/16
// precision.
int av1_get_optflow_based_mv(const AV1_COMMON *cm, MACROBLOCKD *xd, int plane,
                             const MB_MODE_INFO *mbmi, int_mv *mv_refined,
                             int bw, int bh, int mi_x, int mi_y,
                             CalcSubpelParamsFunc calc_subpel_params_func,
                             const void *const calc_subpel_params_func_args) {
  if (is_cur_buf_hbd(xd))
    return get_optflow_based_mv_highbd(cm, xd, plane, mbmi, mv_refined, bw, bh,
                                       mi_x, mi_y, calc_subpel_params_func,
                                       calc_subpel_params_func_args);
  return get_optflow_based_mv_lowbd(cm, xd, plane, mbmi, mv_refined, bw, bh,
                                    mi_x, mi_y, calc_subpel_params_func,
                                    calc_subpel_params_func_args);
}
#endif  // CONFIG_OPTFLOW_REFINEMENT

static void av1_make_inter_predictor_aux(
    const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride,
    const SubpelParams *subpel_params, const struct scale_factors *sf, int w,
    int h, int orig_w, int orig_h, ConvolveParams *conv_params,
    int_interpfilters interp_filters, const WarpTypesAllowed *warp_types,
    int p_col, int p_row, int plane, int ref, const MB_MODE_INFO *mi,
    int build_for_obmc, const MACROBLOCKD *xd, int can_use_previous);

#if CONFIG_OPTFLOW_REFINEMENT && USE_OF_NXN
// Makes the interpredictor for the region by dividing it up into nxn blocks
// and running the interpredictor code on each one.
void make_inter_pred_of_nxn(
    uint8_t *dst, int dst_stride, SubpelParams *subpel_params,
    const struct scale_factors *sf, int w, int h, ConvolveParams *conv_params,
    int_interpfilters interp_filters, const WarpTypesAllowed *warp_types,
    int p_col, int p_row, int plane, int ref, MB_MODE_INFO *mi, MACROBLOCKD *xd,
    int can_use_previous, int n, int_mv *mv_refined, int pre_x, int pre_y,
    struct buf_2d *const pre_buf, CalcSubpelParamsFunc calc_subpel_params_func,
    const void *const calc_subpel_params_func_args) {
  int n_blocks = 0;
  CONV_BUF_TYPE *orig_conv_dst = conv_params->dst;
  assert(w % n == 0);
  assert(h % n == 0);
  uint8_t *pre;
  int src_stride = 0;

  // Process whole nxn blocks.
  for (int j = 0; j <= h - n; j += n) {
    for (int i = 0; i <= w - n; i += n) {
      calc_subpel_params_func(xd, sf, &(mv_refined[n_blocks * 2 + ref].as_mv),
                              plane, pre_x, pre_y, i, j, pre_buf, n, n,
                              warp_types, ref, 1, calc_subpel_params_func_args,
                              &pre, subpel_params, &src_stride);
      av1_make_inter_predictor_aux(pre, src_stride, dst, dst_stride,
                                   subpel_params, sf, n, n, w, h, conv_params,
                                   interp_filters, warp_types, p_col, p_row,
                                   plane, ref, mi, 0, xd, can_use_previous);
      n_blocks++;
      dst += n;
      conv_params->dst += n;
      p_col += n;
    }
    dst -= w;
    conv_params->dst -= w;
    p_col -= w;

    dst += n * dst_stride;
    conv_params->dst += n * conv_params->dst_stride;
    p_row += n;
  }

  conv_params->dst = orig_conv_dst;
}
#endif  // CONFIG_OPTFLOW_REFINEMENT && USE_OF_NXN

// Makes the interpredictor for the region by dividing it up into 8x8 blocks
// and running the interpredictor code on each one.
void make_inter_pred_8x8(const uint8_t *src, int src_stride, uint8_t *dst,
                         int dst_stride, const SubpelParams *subpel_params,
                         const struct scale_factors *sf, int w, int h,
                         int orig_w, int orig_h, ConvolveParams *conv_params,
                         int_interpfilters interp_filters,
                         const WarpTypesAllowed *warp_types, int p_col,
                         int p_row, int plane, int ref, const MB_MODE_INFO *mi,
                         int build_for_obmc, const MACROBLOCKD *xd,
                         int can_use_previous) {
  CONV_BUF_TYPE *orig_conv_dst = conv_params->dst;
  assert(w % 2 == 0);
  assert(h % 2 == 0);
  assert(IMPLIES(w % 8 != 0, orig_w < 8));
  assert(IMPLIES(h % 8 != 0, orig_h < 8));

  // Process whole 8x8 blocks. If width / height are not multiples of 8,
  // a smaller transform is used for the remaining area.
  for (int j = 0; j <= h - 8; j += 8) {
    for (int i = 0; i <= w - 8; i += 8) {
      av1_make_inter_predictor_aux(
          src, src_stride, dst, dst_stride, subpel_params, sf, 8, 8, orig_w,
          orig_h, conv_params, interp_filters, warp_types, p_col, p_row, plane,
          ref, mi, build_for_obmc, xd, can_use_previous);
      src += 8;
      dst += 8;
      conv_params->dst += 8;
      p_col += 8;
    }
    if (w % 8 != 0) {
      int offset = w % 8;
      av1_make_inter_predictor_aux(
          src, src_stride, dst, dst_stride, subpel_params, sf, offset, 8,
          orig_w, orig_h, conv_params, interp_filters, warp_types, p_col, p_row,
          plane, ref, mi, build_for_obmc, xd, can_use_previous);
      src += offset;
      dst += offset;
      conv_params->dst += offset;
      p_col += offset;
    }
    src -= w;
    dst -= w;
    conv_params->dst -= w;
    p_col -= w;

    src += 8 * src_stride;
    dst += 8 * dst_stride;
    conv_params->dst += 8 * conv_params->dst_stride;
    p_row += 8;
  }

  if (h % 8 == 0) {
    conv_params->dst = orig_conv_dst;
    return;
  }
  // There may be a small region left along the bottom. Compute using
  // smaller blocks.
  int h_offset = h % 8;
  for (int i = 0; i <= w - 8; i += 8) {
    av1_make_inter_predictor_aux(
        src, src_stride, dst, dst_stride, subpel_params, sf, 8, h_offset,
        orig_w, orig_h, conv_params, interp_filters, warp_types, p_col, p_row,
        plane, ref, mi, build_for_obmc, xd, can_use_previous);
    src += 8;
    dst += 8;
    conv_params->dst += 8;
    p_col += 8;
  }
  if (w % 8 != 0) {
    av1_make_inter_predictor_aux(
        src, src_stride, dst, dst_stride, subpel_params, sf, h_offset, w % 8,
        orig_w, orig_h, conv_params, interp_filters, warp_types, p_col, p_row,
        plane, ref, mi, build_for_obmc, xd, can_use_previous);
  }
  conv_params->dst = orig_conv_dst;
}

void av1_make_inter_predictor(
    const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride,
    const SubpelParams *subpel_params, const struct scale_factors *sf, int w,
    int h, ConvolveParams *conv_params, int_interpfilters interp_filters,
    const WarpTypesAllowed *warp_types, int p_col, int p_row, int plane,
    int ref, const MB_MODE_INFO *mi, int build_for_obmc, const MACROBLOCKD *xd,
    int can_use_previous, const int border) {
  CONV_BUF_TYPE *orig_conv_dst = conv_params->dst;

  src -= (src_stride * border + border);
  dst -= (dst_stride * border + border);
  p_row -= border;
  p_col -= border;
  // Build the top row of the extension in 8x8 blocks.
  make_inter_pred_8x8(src, src_stride, dst, dst_stride, subpel_params, sf,
                      border + w, border, w, h, conv_params, interp_filters,
                      warp_types, p_col, p_row, plane, ref, mi, build_for_obmc,
                      xd, can_use_previous);
  src += src_stride * border;
  dst += dst_stride * border;
  p_row += border;
  conv_params->dst += conv_params->dst_stride * border;

  // Build the left edge (not including corners).
  make_inter_pred_8x8(src, src_stride, dst, dst_stride, subpel_params, sf,
                      border, h, w, h, conv_params, interp_filters, warp_types,
                      p_col, p_row, plane, ref, mi, build_for_obmc, xd,
                      can_use_previous);
  src += border;
  dst += border;
  p_col += border;
  conv_params->dst += border;

  // Build the original inter-predicator.
  av1_make_inter_predictor_aux(src, src_stride, dst, dst_stride, subpel_params,
                               sf, w, h, w, h, conv_params, interp_filters,
                               warp_types, p_col, p_row, plane, ref, mi,
                               build_for_obmc, xd, can_use_previous);
  conv_params->dst = orig_conv_dst;
}

static void av1_make_inter_predictor_aux(
    const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride,
    const SubpelParams *subpel_params, const struct scale_factors *sf, int w,
    // orig_w and orig_h refer to the width and height of the predictor
    // without the extended region.
    int h, int orig_w, int orig_h, ConvolveParams *conv_params,
    int_interpfilters interp_filters, const WarpTypesAllowed *warp_types,
    int p_col, int p_row, int plane, int ref, const MB_MODE_INFO *mi,
    int build_for_obmc, const MACROBLOCKD *xd, int can_use_previous) {
  // Make sure the selected motion mode is valid for this configuration
  assert_motion_mode_valid(mi->motion_mode, xd->global_motion, xd, mi,
                           can_use_previous);
  assert(IMPLIES(conv_params->is_compound, conv_params->dst != NULL));

  WarpedMotionParams final_warp_params;
  const int do_warp =
#if CONFIG_EXT_WARP && CONFIG_SUB8X8_WARP
      orig_w >= 4 && orig_h >= 4
#else
      orig_w >= 8 && orig_h >= 8
#endif  // CONFIG_EXT_WARP && CONFIG_SUB8X8_WARP
      && av1_allow_warp(mi, warp_types, &xd->global_motion[mi->ref_frame[ref]],
                        build_for_obmc, sf, &final_warp_params);
  const int is_intrabc = mi->use_intrabc;
  assert(IMPLIES(is_intrabc, !do_warp));

  if (do_warp && xd->cur_frame_force_integer_mv == 0) {
    const struct macroblockd_plane *const pd = &xd->plane[plane];
    const struct buf_2d *const pre_buf = &pd->pre[ref];
    av1_warp_plane(&final_warp_params, is_cur_buf_hbd(xd), xd->bd,
                   pre_buf->buf0, pre_buf->width, pre_buf->height,
                   pre_buf->stride, dst, p_col, p_row, w, h, dst_stride,
                   pd->subsampling_x, pd->subsampling_y,
#if CONFIG_EXT_WARP && CONFIG_SUB8X8_WARP
                   (orig_w < 8 || orig_h < 8),
#endif  // CONFIG_EXT_WARP && CONFIG_SUB8X8_WARP
                   conv_params);

  } else if (is_cur_buf_hbd(xd)) {
    highbd_inter_predictor(src, src_stride, dst, dst_stride, subpel_params, w,
                           h, orig_w, orig_h, conv_params, interp_filters,
                           is_intrabc, xd->bd);
  } else {
    inter_predictor(src, src_stride, dst, dst_stride, subpel_params, w, h,
                    orig_w, orig_h, conv_params, interp_filters, is_intrabc);
  }
}

static void build_inter_predictors_sub8x8(
    const AV1_COMMON *cm, MACROBLOCKD *xd, int plane, const MB_MODE_INFO *mi,
    int bw, int bh, int mi_x, int mi_y,
    CalcSubpelParamsFunc calc_subpel_params_func,
    const void *const calc_subpel_params_func_args, uint8_t *orig_dst,
    int orig_dst_stride) {
  const BLOCK_SIZE bsize = mi->sb_type;
  struct macroblockd_plane *const pd = &xd->plane[plane];
  const bool ss_x = pd->subsampling_x;
  const bool ss_y = pd->subsampling_y;

  // block size
  const int b4_w = block_size_wide[bsize] >> ss_x;
  const int b4_h = block_size_high[bsize] >> ss_y;
  const BLOCK_SIZE plane_bsize = plane ? mi->chroma_ref_info.bsize_base : bsize;
  const int b8_w = block_size_wide[plane_bsize] >> ss_x;
  const int b8_h = block_size_high[plane_bsize] >> ss_y;
  assert(!is_intrabc_block(mi));

  // For sub8x8 chroma blocks, we may be covering more than one luma block's
  // worth of pixels. Thus (mi_x, mi_y) may not be the correct coordinates
  // for the top-left corner of the prediction source - the correct top-left
  // corner is at (pre_x, pre_y).
  const int mi_row = -xd->mb_to_top_edge >> (3 + MI_SIZE_LOG2);
  const int mi_col = -xd->mb_to_left_edge >> (3 + MI_SIZE_LOG2);
  const int row_start =
      plane ? mi->chroma_ref_info.mi_row_chroma_base - mi_row : 0;
  const int col_start =
      plane ? mi->chroma_ref_info.mi_col_chroma_base - mi_col : 0;

  const int pre_x = (mi_x + MI_SIZE * col_start) >> ss_x;
  const int pre_y = (mi_y + MI_SIZE * row_start) >> ss_y;

  const struct buf_2d orig_pred_buf[2] = { pd->pre[0], pd->pre[1] };

  const WarpedMotionParams *const wm = &xd->global_motion[mi->ref_frame[0]];
  bool is_global = is_global_mv_block(mi, wm->wmtype);

  int row = row_start;
  for (int y = 0; y < b8_h; y += b4_h) {
    int col = col_start;
    for (int x = 0; x < b8_w; x += b4_w) {
      MB_MODE_INFO *this_mbmi = xd->mi[row * xd->mi_stride + col];
#if CONFIG_EXT_RECUR_PARTITIONS
      // TODO(yuec): enabling compound prediction in none sub8x8 mbs in the
      // group
      bool is_compound = 0;
#else
      bool is_compound = has_second_ref(this_mbmi);
#endif  // CONFIG_EXT_RECUR_PARTITIONS
      const int tmp_dst_stride = 8;
#if !CONFIG_EXT_RECUR_PARTITIONS
      assert(bw < 8 || bh < 8);
#endif  // !CONFIG_EXT_RECUR_PARTITIONS
      ConvolveParams conv_params = get_conv_params_no_round(
          0, plane, xd->tmp_conv_dst, tmp_dst_stride, is_compound, xd->bd);
      conv_params.use_dist_wtd_comp_avg = 0;
      uint8_t *dst = orig_dst + orig_dst_stride * y + x;

      const RefCntBuffer *ref_buf =
          get_ref_frame_buf(cm, this_mbmi->ref_frame[0]);
      const struct scale_factors *ref_scale_factors =
          get_ref_scale_factors_const(cm, this_mbmi->ref_frame[0]);

      pd->pre[0].buf0 =
          (plane == 1) ? ref_buf->buf.u_buffer : ref_buf->buf.v_buffer;
      pd->pre[0].buf =
          pd->pre[0].buf0 + scaled_buffer_offset(pre_x, pre_y,
                                                 ref_buf->buf.uv_stride,
                                                 ref_scale_factors);
      pd->pre[0].width = ref_buf->buf.uv_crop_width;
      pd->pre[0].height = ref_buf->buf.uv_crop_height;
      pd->pre[0].stride = ref_buf->buf.uv_stride;

      const struct scale_factors *const sf = ref_scale_factors;
      struct buf_2d *const pre_buf = &pd->pre[0];

#if CONFIG_DERIVED_MV
      const MV mv = (this_mbmi->derived_mv_allowed && this_mbmi->use_derived_mv)
                        ? this_mbmi->derived_mv[0]
                        : this_mbmi->mv[0].as_mv;
#else
      const MV mv = this_mbmi->mv[0].as_mv;
#endif  // CONFIG_DERIVED_MV

      const WarpTypesAllowed warp_types = { is_global, this_mbmi->motion_mode ==
                                                           WARPED_CAUSAL };

      uint8_t *pre;
      SubpelParams subpel_params;
      int src_stride;
      calc_subpel_params_func(xd, sf, &mv, plane, pre_x, pre_y, x, y, pre_buf,
                              bw, bh, &warp_types, 0 /* ref */,
#if CONFIG_OPTFLOW_REFINEMENT
                              0,
#endif  // CONFIG_OPTFLOW_REFINEMENT
                              calc_subpel_params_func_args, &pre,
                              &subpel_params, &src_stride);

      conv_params.do_average = 0;

      // Border computation does not currnetly work in sub-8x8.
      const int border = 0;
      av1_make_inter_predictor(
          pre, src_stride, dst, orig_dst_stride, &subpel_params, sf, b4_w, b4_h,
          &conv_params, this_mbmi->interp_filters, &warp_types,
          (mi_x >> pd->subsampling_x) + x, (mi_y >> pd->subsampling_y) + y,
          plane, 0 /* ref */, mi, false /* build_for_obmc */, xd,
          cm->allow_warped_motion, border);
      ++col;
    }
    ++row;
  }

  for (int ref = 0; ref < 2; ++ref) {
    pd->pre[ref] = orig_pred_buf[ref];
  }
}

static void build_masked_compound_no_round(
    uint8_t *dst, int dst_stride, const CONV_BUF_TYPE *src0, int src0_stride,
    const CONV_BUF_TYPE *src1, int src1_stride,
    const INTERINTER_COMPOUND_DATA *const comp_data, BLOCK_SIZE sb_type, int h,
    int w, ConvolveParams *conv_params, MACROBLOCKD *xd) {
  // Derive subsampling from h and w passed in. May be refactored to
  // pass in subsampling factors directly.
  const int subh = (2 << mi_size_high_log2[sb_type]) == h;
  const int subw = (2 << mi_size_wide_log2[sb_type]) == w;
  const uint8_t *mask = av1_get_compound_type_mask(comp_data, sb_type);
  if (is_cur_buf_hbd(xd)) {
    aom_highbd_blend_a64_d16_mask(dst, dst_stride, src0, src0_stride, src1,
                                  src1_stride, mask, block_size_wide[sb_type],
                                  w, h, subw, subh, conv_params, xd->bd);
  } else {
    aom_lowbd_blend_a64_d16_mask(dst, dst_stride, src0, src0_stride, src1,
                                 src1_stride, mask, block_size_wide[sb_type], w,
                                 h, subw, subh, conv_params);
  }
}

static void av1_make_masked_inter_predictor(
    const uint8_t *pre, int pre_stride, uint8_t *dst, int dst_stride,
    const SubpelParams *subpel_params, const struct scale_factors *sf, int w,
    int h, ConvolveParams *conv_params, int_interpfilters interp_filters,
    int plane, const WarpTypesAllowed *warp_types, int p_col, int p_row,
    int ref, MACROBLOCKD *xd, int can_use_previous) {
  // Inter-predictor extended border not supported yet.
  assert(av1_calc_border(xd, plane, false) == 0);
  MB_MODE_INFO *mi = xd->mi[0];
  mi->interinter_comp.seg_mask = xd->seg_mask;
  const INTERINTER_COMPOUND_DATA *comp_data = &mi->interinter_comp;

// We're going to call av1_make_inter_predictor to generate a prediction into
// a temporary buffer, then will blend that temporary buffer with that from
// the other reference.
//
#define INTER_PRED_BYTES_PER_PIXEL 2

  DECLARE_ALIGNED(32, uint8_t,
                  tmp_buf[INTER_PRED_BYTES_PER_PIXEL * MAX_SB_SQUARE]);
#undef INTER_PRED_BYTES_PER_PIXEL

  const int tmp_buf_stride = MAX_SB_SIZE;
  CONV_BUF_TYPE *org_dst = conv_params->dst;
  int org_dst_stride = conv_params->dst_stride;
  CONV_BUF_TYPE *tmp_buf16 = (CONV_BUF_TYPE *)tmp_buf;
  conv_params->dst = tmp_buf16;
  conv_params->dst_stride = tmp_buf_stride;
  assert(conv_params->do_average == 0);

  // This will generate a prediction in tmp_buf for the second reference
  const int border = 0;
  av1_make_inter_predictor(pre, pre_stride, dst, dst_stride, subpel_params, sf,
                           w, h, conv_params, interp_filters, warp_types, p_col,
                           p_row, plane, ref, mi, 0, xd, can_use_previous,
                           border);

  if (!plane && comp_data->type == COMPOUND_DIFFWTD) {
#if CONFIG_CTX_ADAPT_LOG_WEIGHT || CONFIG_DIFFWTD_42
    av1_build_compound_diffwtd_mask_d16_c(
        comp_data->seg_mask, comp_data->mask_type, org_dst, org_dst_stride,
        tmp_buf16, tmp_buf_stride, h, w, conv_params, xd->bd);
#else
    av1_build_compound_diffwtd_mask_d16(
        comp_data->seg_mask, comp_data->mask_type, org_dst, org_dst_stride,
        tmp_buf16, tmp_buf_stride, h, w, conv_params, xd->bd);
#endif  // CONFIG_CTX_ADAPT_LOG_WEIGHT || CONFIG_DIFFWTD_42
  }
  build_masked_compound_no_round(dst, dst_stride, org_dst, org_dst_stride,
                                 tmp_buf16, tmp_buf_stride, comp_data,
                                 mi->sb_type, h, w, conv_params, xd);
}

static void build_inter_predictors(
    const AV1_COMMON *cm, MACROBLOCKD *xd, int plane, MB_MODE_INFO *mi,
    int build_for_obmc, int bw, int bh, int mi_x, int mi_y,
    CalcSubpelParamsFunc calc_subpel_params_func,
    const void *const calc_subpel_params_func_args, uint8_t *dst,
    int dst_stride, const int border) {
  int is_compound = has_second_ref(mi);
  ConvolveParams conv_params = get_conv_params_no_round(
      0, plane, xd->tmp_conv_dst, MAX_SB_SIZE, is_compound, xd->bd);
  av1_dist_wtd_comp_weight_assign(
      cm, mi, 0, &conv_params.fwd_offset, &conv_params.bck_offset,
      &conv_params.use_dist_wtd_comp_avg, is_compound);

  struct macroblockd_plane *const pd = &xd->plane[plane];
  struct buf_2d *const dst_buf = &pd->dst;
  const int is_intrabc = is_intrabc_block(mi);

  int is_global[2] = { 0, 0 };
  for (int ref = 0; ref < 1 + is_compound; ++ref) {
    const WarpedMotionParams *const wm = &xd->global_motion[mi->ref_frame[ref]];
    is_global[ref] = is_global_mv_block(mi, wm->wmtype);
  }

  const int ss_x = pd->subsampling_x;
  const int ss_y = pd->subsampling_y;
  int row_start = 0;
  int col_start = 0;
  if (!build_for_obmc) {
    const int mi_row = -xd->mb_to_top_edge >> (3 + MI_SIZE_LOG2);
    const int mi_col = -xd->mb_to_left_edge >> (3 + MI_SIZE_LOG2);
    int mi_row_offset, mi_col_offset;
    mi_row_offset =
        plane ? (mi_row - mi->chroma_ref_info.mi_row_chroma_base) : 0;
    mi_col_offset =
        plane ? (mi_col - mi->chroma_ref_info.mi_col_chroma_base) : 0;
    row_start = -mi_row_offset;
    col_start = -mi_col_offset;
  }
  const int pre_x = (mi_x + MI_SIZE * col_start) >> ss_x;
  const int pre_y = (mi_y + MI_SIZE * row_start) >> ss_y;

#if CONFIG_DERIVED_MV
  const int use_derived_mv = mi->derived_mv_allowed && mi->use_derived_mv;
#endif  // CONFIG_DERIVED_MV
#if CONFIG_OPTFLOW_REFINEMENT
  const int use_optflow_prec = (mi->mode > NEW_NEWMV) && is_compound;
  assert(IMPLIES(use_optflow_prec, !build_for_obmc));

  // Obtain offset MVs in Y plane
  if (use_optflow_prec && plane == 0) {
    // Initialize refined mv
#if CONFIG_DERIVED_MV
    const MV mv0 = use_derived_mv ? mi->derived_mv[0] : mi->mv[0].as_mv;
    const MV mv1 = use_derived_mv ? mi->derived_mv[1] : mi->mv[1].as_mv;
#else
    const MV mv0 = mi->mv[0].as_mv;
    const MV mv1 = mi->mv[1].as_mv;
#endif  // CONFIG_DERIVED_MV
    for (int mvi = 0; mvi < N_OF_OFFSETS; mvi++) {
      mi->mv_refined[mvi * 2].as_mv = mv0;
      mi->mv_refined[mvi * 2 + 1].as_mv = mv1;
    }
    av1_get_optflow_based_mv(cm, xd, plane, mi, mi->mv_refined, bw, bh, mi_x,
                             mi_y, calc_subpel_params_func,
                             calc_subpel_params_func_args);
  }

#endif  // CONFIG_OPTFLOW_REFINEMENT
  assert(IMPLIES(is_intrabc || is_compound, dst == dst_buf->buf));
  for (int ref = 0; ref < 1 + is_compound; ++ref) {
    const struct scale_factors *const sf =
        is_intrabc ? &cm->sf_identity : xd->block_ref_scale_factors[ref];
    struct buf_2d *const pre_buf = is_intrabc ? dst_buf : &pd->pre[ref];
#if CONFIG_DERIVED_MV
    const MV mv = use_derived_mv ? mi->derived_mv[ref] : mi->mv[ref].as_mv;
#else
    const MV mv = mi->mv[ref].as_mv;
#endif  // CONFIG_DERIVED_MV
    const WarpTypesAllowed warp_types = { is_global[ref],
                                          mi->motion_mode == WARPED_CAUSAL };
    uint8_t *pre;
    SubpelParams subpel_params;
    int src_stride;
#if CONFIG_OPTFLOW_REFINEMENT
    // For luma, always apply offset MVs. For chroma, use the MVs derived for
    // luma if luma subblock size is 8x8 (i.e., chroma block size > 8x8),
    // and otherwise apply normal compound average.
    if (use_optflow_prec && (plane == 0 || bh > 8 || bw > 8)) {
      conv_params.do_average = ref;
#if USE_OF_NXN
      int n = plane ? OF_BSIZE / 2
                    : ((bh <= 16 && bw <= 16) ? OF_MIN_BSIZE : OF_BSIZE);
      make_inter_pred_of_nxn(
          dst, dst_buf->stride, &subpel_params, sf, bw, bh, &conv_params,
          mi->interp_filters, &warp_types, mi_x >> pd->subsampling_x,
          mi_y >> pd->subsampling_y, plane, ref, mi, xd,
          cm->allow_warped_motion, n, mi->mv_refined, pre_x, pre_y, pre_buf,
          calc_subpel_params_func, calc_subpel_params_func_args);
#else
      // Compute subpel params with refined mv
      calc_subpel_params_func(xd, sf, &(mi->mv_refined[ref].as_mv), plane,
                              pre_x, pre_y, 0, 0, pre_buf, bw, bh, &warp_types,
                              ref, 1, calc_subpel_params_func_args, &pre,
                              &subpel_params, &src_stride);
      av1_make_inter_predictor(
          pre, src_stride, dst, dst_buf->stride, &subpel_params, sf, bw, bh,
          &conv_params, mi->interp_filters, &warp_types,
          mi_x >> pd->subsampling_x, mi_y >> pd->subsampling_y, plane, ref, mi,
          0, xd, cm->allow_warped_motion, 0 /* border */);
#endif  // USE_OF_NXN
        // Predictor already built
      continue;
    } else {
      calc_subpel_params_func(xd, sf, &mv, plane, pre_x, pre_y, 0, 0, pre_buf,
                              bw, bh, &warp_types, ref, 0,
                              calc_subpel_params_func_args, &pre,
                              &subpel_params, &src_stride);
    }
#else
    calc_subpel_params_func(xd, sf, &mv, plane, pre_x, pre_y, 0, 0, pre_buf, bw,
                            bh, &warp_types, ref, calc_subpel_params_func_args,
                            &pre, &subpel_params, &src_stride);
#endif  // CONFIG_OPTFLOW_REFINEMENT

    if (ref && is_masked_compound_type(mi->interinter_comp.type)) {
      // masked compound type has its own average mechanism
      conv_params.do_average = 0;
      assert(!build_for_obmc);
      av1_make_masked_inter_predictor(
          pre, src_stride, dst, dst_stride, &subpel_params, sf, bw, bh,
          &conv_params, mi->interp_filters, plane, &warp_types,
          mi_x >> pd->subsampling_x, mi_y >> pd->subsampling_y, ref, xd,
          cm->allow_warped_motion);
    } else {
      conv_params.do_average = ref;

#if CONFIG_EXT_IBC_MODES
      // IBC+ Winners Only : Extract Predicted block & translate accordingly
      if (is_intrabc && mi->ibc_mode) {
        uint8_t ibcWinMode = mi->ibc_mode;

        // Allocate & Extract/Fetch predicted block
        uint16_t *pred_block = NULL;

        uint8_t *dst_block = CONVERT_TO_BYTEPTR(xd->ibc_pred);
        uint8_t dst_block_stride = 128;

        if (bw != bh &&
            (ibcWinMode == ROTATION_90 || ibcWinMode == ROTATION_270 ||
             ibcWinMode == MIRROR_45 || ibcWinMode == MIRROR_135)) {
          av1_make_inter_predictor(
              pre, src_stride, dst_block, dst_block_stride, &subpel_params, sf,
              bh, bw, &conv_params, mi->interp_filters, &warp_types,
              mi_x >> pd->subsampling_x, mi_y >> pd->subsampling_y, plane, ref,
              mi, build_for_obmc, xd, cm->allow_warped_motion, border);

          av1_intrabc_allocate_sb(&pred_block, bh, bw);
          av1_fetch_prediction_sb(dst_block, dst_block_stride, pred_block, bh,
                                  bw);
        } else {
          av1_make_inter_predictor(
              pre, src_stride, dst_block, dst_block_stride, &subpel_params, sf,
              bw, bh, &conv_params, mi->interp_filters, &warp_types,
              mi_x >> pd->subsampling_x, mi_y >> pd->subsampling_y, plane, ref,
              mi, build_for_obmc, xd, cm->allow_warped_motion, border);

          av1_intrabc_allocate_sb(&pred_block, bw, bh);
          av1_fetch_prediction_sb(dst_block, dst_block_stride, pred_block, bw,
                                  bh);
        }

        switch (ibcWinMode) {
          case MIRROR_90:
            av1_intrabc_mirror90_sb(xd->ibc_pred, pred_block, bw, bh);
            break;

          case MIRROR_0:
            av1_intrabc_mirror0_sb(xd->ibc_pred, pred_block, bw, bh);
            break;

          case ROTATION_180:
            av1_intrabc_rotate180_sb(xd->ibc_pred, pred_block, bw, bh);
            break;

          case ROTATION_90:
            av1_intrabc_rotate270_sb(xd->ibc_pred, pred_block, bh, bw);
            break;

          case MIRROR_135:
            av1_intrabc_mirror135_sb(xd->ibc_pred, pred_block, bh, bw);
            break;

          case MIRROR_45:
            av1_intrabc_mirror45_sb(xd->ibc_pred, pred_block, bh, bw);
            break;

          case ROTATION_270:
            av1_intrabc_rotate90_sb(xd->ibc_pred, pred_block, bh, bw);
            break;

          default: break;  // assert(0);
        }

        av1_write_prediction_sb(dst, dst_stride, xd->ibc_pred, bw, bh);

        // Deallocate 1D arrays
        aom_free(pred_block);
      } else {  // Regular IBC
        av1_make_inter_predictor(
            pre, src_stride, dst, dst_stride, &subpel_params, sf, bw, bh,
            &conv_params, mi->interp_filters, &warp_types,
            mi_x >> pd->subsampling_x, mi_y >> pd->subsampling_y, plane, ref,
            mi, build_for_obmc, xd, cm->allow_warped_motion, border);
      }
#else
      av1_make_inter_predictor(
          pre, src_stride, dst, dst_stride, &subpel_params, sf, bw, bh,
          &conv_params, mi->interp_filters, &warp_types,
          mi_x >> pd->subsampling_x, mi_y >> pd->subsampling_y, plane, ref, mi,
          build_for_obmc, xd, cm->allow_warped_motion, border);
#endif  // CONFIG_EXT_IBC_MODES
    }
  }
}

#if CONFIG_DERIVED_MV
#define DERIVED_MV_REF_LINES 4
#define REFINE_SUBPEL_RANGE 8
#define REFINE_FULLPEL_RANGE 4
#define REFINE_FULLPEL_STEP 1
#define DERIVED_MV_IDX_RANGE 1
#define DERIVED_MV_MAX_BSIZE 64
#define DERIVED_MV_MIN_BSIZE 4

int av1_derived_mv_allowed(MACROBLOCKD *const xd, MB_MODE_INFO *const mbmi) {
  const BLOCK_SIZE bsize = mbmi->sb_type;
  const int bw = block_size_wide[bsize];
  const int bh = block_size_high[bsize];
  return !is_cur_buf_hbd(xd) &&
         (mbmi->mode == NEARMV ||
#if CONFIG_OPTFLOW_REFINEMENT
          mbmi->mode == NEAR_NEARMV_OPTFLOW ||
#endif  // CONFIG_OPTFLOW_REFINEMENT
          mbmi->mode == NEAR_NEARMV) &&
         bw <= DERIVED_MV_MAX_BSIZE && bh <= DERIVED_MV_MAX_BSIZE &&
         bw >= DERIVED_MV_MIN_BSIZE && bh >= DERIVED_MV_MIN_BSIZE &&
         xd->mi_row + mi_size_high[bsize] <= xd->tile.mi_row_end &&
         xd->mi_col + mi_size_wide[bsize] <= xd->tile.mi_col_end &&
         xd->mi_row > xd->tile.mi_row_start &&
         xd->mi_col > xd->tile.mi_col_start;
}

// A mesh full pel search around the reference MV.
static MV full_pel_refine(const AV1_COMMON *const cm, MACROBLOCKD *xd, int ref,
                          BLOCK_SIZE bsize, MV ref_mv, const uint8_t *top,
                          const uint8_t *left, const uint8_t *top_left,
                          int stride, aom_subpixvariance_fn_t svp_fn_top,
                          aom_subpixvariance_fn_t svp_fn_left,
                          uint32_t *best_error) {
  struct macroblockd_plane *const pd = &xd->plane[0];
  const uint8_t *ref_buf = pd->pre[ref].buf;
  const int ref_stride = pd->pre[ref].stride;
  *best_error = UINT32_MAX;
  uint32_t sse;
  MV best_mv = ref_mv;
  ref_mv.row = ref_mv.row >> 3;
  ref_mv.col = ref_mv.col >> 3;
  const int x = xd->mi_col * 4 * 8;
  const int y = xd->mi_row * 4 * 8;
  for (int r = ref_mv.row - REFINE_FULLPEL_RANGE;
       r <= ref_mv.row + REFINE_FULLPEL_RANGE; r += REFINE_FULLPEL_STEP) {
    for (int c = ref_mv.col - REFINE_FULLPEL_RANGE;
         c <= ref_mv.col + REFINE_FULLPEL_RANGE; c += REFINE_FULLPEL_STEP) {
      // Do not consider it when falling out of frame boundary. It may cause
      // mismatches because boarder extenstion is handled differently on the
      // encoder and decoder side.
      if (x + c * 8 - DERIVED_MV_REF_LINES * 8 < 0 ||
          y + r * 8 - DERIVED_MV_REF_LINES * 8 < 0 ||
          xd->mi_col * 4 + block_size_wide[bsize] + c >= cm->width ||
          xd->mi_row * 4 + block_size_high[bsize] + r >= cm->height) {
        continue;
      }
      const uint8_t *pre_top =
          ref_buf + (r - DERIVED_MV_REF_LINES) * ref_stride + c;
      const uint8_t *pre_left =
          ref_buf + r * ref_stride + c - DERIVED_MV_REF_LINES;
      const uint8_t *pre_top_left = ref_buf +
                                    (r - DERIVED_MV_REF_LINES) * ref_stride +
                                    c - DERIVED_MV_REF_LINES;
      const uint32_t top_error =
          svp_fn_top(pre_top, ref_stride, 0, 0, top, stride, &sse);
      const uint32_t left_error =
          svp_fn_left(pre_left, ref_stride, 0, 0, left, stride, &sse);
      const uint32_t top_left_error = aom_sub_pixel_variance4x4(
          pre_top_left, ref_stride, 0, 0, top_left, stride, &sse);
      const uint32_t this_error = top_error + left_error + top_left_error;
      if (this_error < *best_error) {
        *best_error = this_error;
        best_mv.row = r * 8;
        best_mv.col = c * 8;
      }
    }
  }
  return best_mv;
}

MV av1_derive_mv(const AV1_COMMON *const cm, MACROBLOCKD *xd, int ref,
                 MB_MODE_INFO *mbmi, uint8_t *recon_buf, int recon_stride) {
  struct macroblockd_plane *const pd = &xd->plane[0];
  const uint8_t *ref_buf = pd->pre[ref].buf;
  const int ref_stride = pd->pre[ref].stride;
  const uint8_t *recon_top = recon_buf - DERIVED_MV_REF_LINES * recon_stride;
  const uint8_t *recon_left = recon_buf - DERIVED_MV_REF_LINES;
  const uint8_t *recon_top_left =
      recon_buf - DERIVED_MV_REF_LINES * recon_stride - DERIVED_MV_REF_LINES;
  const BLOCK_SIZE bsize = mbmi->sb_type;
  const int bwl = mi_size_wide_log2[bsize];
  const int bhl = mi_size_high_log2[bsize];
  aom_subpixvariance_fn_t svp_fn_top, svp_fn_left;
  switch (bwl) {
    case 0: svp_fn_top = aom_sub_pixel_variance4x4; break;
    case 1: svp_fn_top = aom_sub_pixel_variance8x4; break;
    case 2: svp_fn_top = aom_sub_pixel_variance16x4; break;
    case 3: svp_fn_top = aom_sub_pixel_variance32x4; break;
    default: svp_fn_top = aom_sub_pixel_variance64x4; break;
  }
  switch (bhl) {
    case 0: svp_fn_left = aom_sub_pixel_variance4x4; break;
    case 1: svp_fn_left = aom_sub_pixel_variance4x8; break;
    case 2: svp_fn_left = aom_sub_pixel_variance4x16; break;
    case 3: svp_fn_left = aom_sub_pixel_variance4x32; break;
    default: svp_fn_left = aom_sub_pixel_variance4x64; break;
  }
  uint32_t best_error = UINT32_MAX;
  uint32_t sse;
  int16_t inter_mode_ctx[MODE_CTX_REF_FRAMES];
  int_mv ref_mvs[MODE_CTX_REF_FRAMES][MAX_MV_REF_CANDIDATES] = { { { 0 } } };
  MV_REFERENCE_FRAME ref_frame = av1_ref_frame_type(mbmi->ref_frame);
  av1_find_mv_refs(cm, xd, mbmi, ref_frame, &xd->ref_mv_info, ref_mvs, NULL,
                   inter_mode_ctx);
  MV best_mv = ref ? xd->ref_mv_info.ref_mv_stack[ref_frame][0].comp_mv.as_mv
                   : xd->ref_mv_info.ref_mv_stack[ref_frame][0].this_mv.as_mv;
  int step = 1;
  if (cm->fr_mv_precision == MV_SUBPEL_NONE) {
    step = 8;
  } else if (cm->fr_mv_precision == MV_SUBPEL_HALF_PRECISION) {
    step = 4;
  } else if (cm->fr_mv_precision == MV_SUBPEL_QTR_PRECISION) {
    step = 2;
  }
  const int x = xd->mi_col * 4 * 8;
  const int y = xd->mi_row * 4 * 8;
  const int bw = block_size_wide[bsize];
  const int bh = block_size_high[bsize];
  // Full pixel motion search around each reference MV.
  for (int i = 0; i < AOMMIN(DERIVED_MV_IDX_RANGE,
                             xd->ref_mv_info.ref_mv_count[ref_frame]);
       ++i) {
    MV ref_mv = ref ? xd->ref_mv_info.ref_mv_stack[ref_frame][i].comp_mv.as_mv
                    : xd->ref_mv_info.ref_mv_stack[ref_frame][i].this_mv.as_mv;
    uint32_t full_pel_error;
    // Search the full pel locations around the ref_mv.
    ref_mv = full_pel_refine(cm, xd, ref, bsize, ref_mv, recon_top, recon_left,
                             recon_top_left, recon_stride, svp_fn_top,
                             svp_fn_left, &full_pel_error);
    if (full_pel_error < best_error) {
      best_error = full_pel_error;
      best_mv = ref_mv;
    }
  }

  const MV best_full_pel_mv = best_mv;
  // Subpel search around the best full pel MV.
  for (int r = best_full_pel_mv.row - REFINE_SUBPEL_RANGE;
       r <= best_full_pel_mv.row + REFINE_SUBPEL_RANGE; r += step) {
    for (int c = best_full_pel_mv.col - REFINE_SUBPEL_RANGE;
         c <= best_full_pel_mv.col + REFINE_SUBPEL_RANGE; c += step) {
      if (x + c - DERIVED_MV_REF_LINES * 8 < 0 ||
          y + r - DERIVED_MV_REF_LINES * 8 < 0 ||
          (xd->mi_col * 4 + bw) * 8 + c >= cm->width * 8 ||
          (xd->mi_row * 4 + bh) * 8 + r >= cm->height * 8) {
        continue;
      }
      const int r_int = r >> 3;
      const int c_int = c >> 3;
      const int r_sub = r & 7;
      const int c_sub = c & 7;
      const uint8_t *pre_top =
          ref_buf + (r_int - DERIVED_MV_REF_LINES) * ref_stride + c_int;
      const uint8_t *pre_left =
          ref_buf + r_int * ref_stride + c_int - DERIVED_MV_REF_LINES;
      const uint8_t *pre_top_left =
          ref_buf + (r_int - DERIVED_MV_REF_LINES) * ref_stride + c_int -
          DERIVED_MV_REF_LINES;
      const uint32_t top_error = svp_fn_top(pre_top, ref_stride, c_sub, r_sub,
                                            recon_top, recon_stride, &sse);
      const uint32_t left_error = svp_fn_left(
          pre_left, ref_stride, c_sub, r_sub, recon_left, recon_stride, &sse);
      const uint32_t top_left_error =
          aom_sub_pixel_variance4x4(pre_top_left, ref_stride, c_sub, r_sub,
                                    recon_top_left, recon_stride, &sse);
      const uint32_t this_error = top_error + left_error + top_left_error;
      if (this_error < best_error) {
        best_error = this_error;
        best_mv.row = r;
        best_mv.col = c;
      }
    }
  }

  return best_mv;
}
#endif  // CONFIG_DERIVED_MV

// True if the following hold:
//  1. Not intrabc and not build_for_obmc
//  2. A U or V plane
//  3. If the block size differs from the base block size
//  4. If sub-sampled, none of the previous blocks around the sub-sample
//     are intrabc or inter-blocks
static bool is_sub8x8_inter(const MACROBLOCKD *xd, int plane,
                            const MB_MODE_INFO *mi, int build_for_obmc) {
  const int is_intrabc = is_intrabc_block(mi);
  if (is_intrabc || build_for_obmc) {
    return false;
  }

  const BLOCK_SIZE bsize = mi->sb_type;
  int sub8x8_inter = plane && (bsize != mi->chroma_ref_info.bsize_base);

  if (!sub8x8_inter) {
    return false;
  }
  // For sub8x8 chroma blocks, we may be covering more than one luma block's
  // worth of pixels.
  const int mi_row = -xd->mb_to_top_edge >> (3 + MI_SIZE_LOG2);
  const int mi_col = -xd->mb_to_left_edge >> (3 + MI_SIZE_LOG2);
  const int row_start =
      plane ? mi->chroma_ref_info.mi_row_chroma_base - mi_row : 0;
  const int col_start =
      plane ? mi->chroma_ref_info.mi_col_chroma_base - mi_col : 0;

  for (int row = row_start; row <= 0; ++row) {
    for (int col = col_start; col <= 0; ++col) {
      const MB_MODE_INFO *this_mbmi = xd->mi[row * xd->mi_stride + col];
      if (!is_inter_block(this_mbmi) || is_intrabc_block(this_mbmi)) {
        return false;
      }
    }
  }
  return true;
}

int av1_calc_border(const MACROBLOCKD *xd, int plane, int build_for_obmc) {
#if CONFIG_INTERINTRA_BORDER
  // Intra-block copy will set the source pointer to a different location in
  // the destination buffer. It's possible that the border pixels around
  // that region have not been initialized. Compound mode does not currently
  // work as the masked inter-predictor needs to increase its region used
  // for the mask.
  const MB_MODE_INFO *mi = xd->mi[0];
  const bool is_compound = has_second_ref(mi);
  const bool intra_bc = mi->use_intrabc;
  if (is_compound || intra_bc) {
    return 0;
  }
  // Not implemented for sub-8x8 blocks.
  if (is_sub8x8_inter(xd, plane, mi, build_for_obmc)) {
    return 0;
  }
  return INTERINTRA_PRED_BORDER;
#endif  // CONFIG_INTERINTRA_BORDER
  (void)xd;
  (void)plane;
  (void)build_for_obmc;
  return 0;
}

void av1_build_inter_predictors(const AV1_COMMON *cm, MACROBLOCKD *xd,
                                int plane, MB_MODE_INFO *mi, int build_for_obmc,
                                int bw, int bh, int mi_x, int mi_y,
                                CalcSubpelParamsFunc calc_subpel_params_func,
                                const void *const calc_subpel_params_func_args,
                                uint8_t *dst, int dst_stride, int border) {
  if (is_sub8x8_inter(xd, plane, mi, build_for_obmc)) {
    assert(border == 0);  // Not yet supported for sub8x8.
    build_inter_predictors_sub8x8(
        cm, xd, plane, mi, bw, bh, mi_x, mi_y, calc_subpel_params_func,
        calc_subpel_params_func_args, dst, dst_stride);
  } else {
    build_inter_predictors(cm, xd, plane, mi, build_for_obmc, bw, bh, mi_x,
                           mi_y, calc_subpel_params_func,
                           calc_subpel_params_func_args, dst, dst_stride,
                           border);
  }
}

#if USE_PRECOMPUTED_WEDGE_MASK
static const uint8_t wedge_master_oblique_odd[MASK_MASTER_SIZE] = {
  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,  0,  0,  1,  2,  6,  18,
  37, 53, 60, 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, 64, 64,
};
static const uint8_t wedge_master_oblique_even[MASK_MASTER_SIZE] = {
  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,  0,  0,  1,  4,  11, 27,
  46, 58, 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, 64, 64,
};
static const uint8_t wedge_master_vertical[MASK_MASTER_SIZE] = {
  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,  0,  0,  0,  2,  7,  21,
  43, 57, 62, 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, 64, 64, 64,
};

static void shift_copy(const uint8_t *src, uint8_t *dst, int shift, int width) {
  if (shift >= 0) {
    memcpy(dst + shift, src, width - shift);
    memset(dst, src[0], shift);
  } else {
    shift = -shift;
    memcpy(dst, src + shift, width - shift);
    memset(dst + width - shift, src[width - 1], shift);
  }
}
#endif  // USE_PRECOMPUTED_WEDGE_MASK

#if USE_PRECOMPUTED_WEDGE_SIGN
/* clang-format off */
DECLARE_ALIGNED(16, static uint8_t,
                wedge_signflip_lookup[BLOCK_SIZES_ALL][MAX_WEDGE_TYPES]) = {
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
  { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 0, 1, },
  { 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 0, 1, },
  { 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 0, 1, },
  { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 0, 1, },
  { 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 0, 1, },
  { 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 1, 1, 0, 1, },
  { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 0, 1, },
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
  { 1, 1, 1, 1, 0, 1, 1, 1, 0, 1, 0, 1, 1, 1, 0, 1, },
  { 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 0, 1, 0, 1, 0, 1, },
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
#if CONFIG_FLEX_PARTITION
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
  { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, },  // not used
#endif  // CONFIG_FLEX_PARTITION
};
/* clang-format on */
#else
DECLARE_ALIGNED(16, static uint8_t,
                wedge_signflip_lookup[BLOCK_SIZES_ALL][MAX_WEDGE_TYPES]);
#endif  // USE_PRECOMPUTED_WEDGE_SIGN

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

// 4 * 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 * 4 * MAX_WEDGE_SQUARE]);

static wedge_masks_type wedge_masks[BLOCK_SIZES_ALL][2];

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 av1_wedge_params_lookup[BLOCK_SIZES_ALL] = {
  { 0, NULL, NULL, NULL },
  { 0, NULL, NULL, NULL },
  { 0, NULL, NULL, NULL },
  { 4, wedge_codebook_16_heqw, wedge_signflip_lookup[BLOCK_8X8],
    wedge_masks[BLOCK_8X8] },
  { 4, wedge_codebook_16_hgtw, wedge_signflip_lookup[BLOCK_8X16],
    wedge_masks[BLOCK_8X16] },
  { 4, wedge_codebook_16_hltw, wedge_signflip_lookup[BLOCK_16X8],
    wedge_masks[BLOCK_16X8] },
  { 4, wedge_codebook_16_heqw, wedge_signflip_lookup[BLOCK_16X16],
    wedge_masks[BLOCK_16X16] },
  { 4, wedge_codebook_16_hgtw, wedge_signflip_lookup[BLOCK_16X32],
    wedge_masks[BLOCK_16X32] },
  { 4, wedge_codebook_16_hltw, wedge_signflip_lookup[BLOCK_32X16],
    wedge_masks[BLOCK_32X16] },
  { 4, wedge_codebook_16_heqw, wedge_signflip_lookup[BLOCK_32X32],
    wedge_masks[BLOCK_32X32] },
  { 0, NULL, NULL, NULL },
  { 0, NULL, NULL, NULL },
  { CONFIG_SEGMENT_BASED_PARTITIONING, NULL, NULL, NULL },
  { 0, NULL, NULL, NULL },
  { 0, NULL, NULL, NULL },
  { CONFIG_SEGMENT_BASED_PARTITIONING, NULL, NULL, NULL },
  { 0, NULL, NULL, NULL },
  { 0, NULL, NULL, NULL },
  { 4, wedge_codebook_16_hgtw, wedge_signflip_lookup[BLOCK_8X32],
    wedge_masks[BLOCK_8X32] },
  { 4, wedge_codebook_16_hltw, wedge_signflip_lookup[BLOCK_32X8],
    wedge_masks[BLOCK_32X8] },
  { 0, NULL, NULL, NULL },
  { 0, NULL, NULL, NULL },
#if CONFIG_FLEX_PARTITION
  { 0, NULL, NULL, NULL },
  { 0, NULL, NULL, NULL },
  { 0, NULL, NULL, NULL },
  { 0, NULL, NULL, NULL },
  { 0, NULL, NULL, NULL },
  { 0, NULL, NULL, NULL },
#endif  // CONFIG_FLEX_PARTITION
};

static const uint8_t *get_wedge_mask_inplace(int wedge_index, int neg,
                                             BLOCK_SIZE sb_type) {
  const uint8_t *master;
  const int bh = block_size_high[sb_type];
  const int bw = block_size_wide[sb_type];
  const wedge_code_type *a =
      av1_wedge_params_lookup[sb_type].codebook + wedge_index;
  int woff, hoff;
  const uint8_t wsignflip =
      av1_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[neg ^ wsignflip][a->direction] +
           MASK_MASTER_STRIDE * (MASK_MASTER_SIZE / 2 - hoff) +
           MASK_MASTER_SIZE / 2 - woff;
  return master;
}

const uint8_t *av1_get_compound_type_mask(
    const INTERINTER_COMPOUND_DATA *const comp_data, BLOCK_SIZE sb_type) {
  assert(is_masked_compound_type(comp_data->type));
  (void)sb_type;
  switch (comp_data->type) {
    case COMPOUND_WEDGE:
#if CONFIG_SEGMENT_BASED_PARTITIONING
      if (av1_wedge_params_lookup[sb_type].codebook == NULL) {
        // We are using an arbitrary mask, stored earlier.
        return comp_data->seg_mask;
      }
#endif  // CONFIG_SEGMENT_BASED_PARTITIONING
      return av1_get_contiguous_soft_mask(comp_data->wedge_index,
                                          comp_data->wedge_sign, sb_type);
    case COMPOUND_DIFFWTD: return comp_data->seg_mask;
    default: assert(0); return NULL;
  }
}

static void diffwtd_mask_d16(uint8_t *mask, int which_inverse, int mask_base,
                             const CONV_BUF_TYPE *src0, int src0_stride,
                             const CONV_BUF_TYPE *src1, int src1_stride, int h,
                             int w, ConvolveParams *conv_params, int bd) {
#if CONFIG_CTX_ADAPT_LOG_WEIGHT
  (void)bd;
  (void)mask_base;
  double *R0 = NULL;
  double *R1 = NULL;
  int m;
  const CONV_BUF_TYPE *pred0 = which_inverse ? src1 : src0;
  const CONV_BUF_TYPE *pred1 = which_inverse ? src0 : src1;
  int stride0 = which_inverse ? src1_stride : src0_stride;
  int stride1 = which_inverse ? src0_stride : src1_stride;

  R0 = gen_correlation(pred0, stride0, h, w, log_k, LOG_K_SIZE, conv_params);
  R1 = gen_correlation(pred1, stride1, h, w, log_k, LOG_K_SIZE, conv_params);
  int i, j;
  for (i = 0; i < h; ++i) {
    for (j = 0; j < w; ++j) {
      double_t edge_diff = fabs(R0[i * w + j]) - fabs(R1[i * w + j]);
      if (edge_diff < DIFFLOG_THR) {
        m = LOG_WEIGHT_1;
      } else {
        m = LOG_WEIGHT_0;
      }
      mask[i * w + j] = which_inverse ? AOM_BLEND_A64_MAX_ALPHA - m : m;
    }
  }
  free(R0);
  free(R1);
#else
  int round =
      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1 + (bd - 8);
  int i, j, m, diff;
  for (i = 0; i < h; ++i) {
    for (j = 0; j < w; ++j) {
      diff = abs(src0[i * src0_stride + j] - src1[i * src1_stride + j]);
      diff = ROUND_POWER_OF_TWO(diff, round);
      m = clamp(mask_base + (diff / DIFF_FACTOR), 0, AOM_BLEND_A64_MAX_ALPHA);
      mask[i * w + j] = which_inverse ? AOM_BLEND_A64_MAX_ALPHA - m : m;
    }
  }
#endif  // CONFIG_CTX_ADAPT_LOG_WEIGHT
}

void av1_build_compound_diffwtd_mask_d16_c(
    uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const CONV_BUF_TYPE *src0,
    int src0_stride, const CONV_BUF_TYPE *src1, int src1_stride, int h, int w,
    ConvolveParams *conv_params, int bd) {
  switch (mask_type) {
    case NORMAL_MASK:
      diffwtd_mask_d16(mask, 0, DIFFWTD_MASK_VAL, src0, src0_stride, src1,
                       src1_stride, h, w, conv_params, bd);
      break;
    case INVERSE_MASK:
      diffwtd_mask_d16(mask, 1, DIFFWTD_MASK_VAL, src0, src0_stride, src1,
                       src1_stride, h, w, conv_params, bd);
      break;
    default: assert(0);
  }
}

static void diffwtd_mask(uint8_t *mask, int which_inverse, int mask_base,
                         const uint8_t *src0, int src0_stride,
                         const uint8_t *src1, int src1_stride, int h, int w) {
#if CONFIG_CTX_ADAPT_LOG_WEIGHT
  (void)mask_base;
  double *R0 = NULL;
  double *R1 = NULL;
  int m;
  const uint8_t *pred0 = which_inverse ? src1 : src0;
  const uint8_t *pred1 = which_inverse ? src0 : src1;
  int stride0 = which_inverse ? src1_stride : src0_stride;
  int stride1 = which_inverse ? src0_stride : src1_stride;
  R0 = gen_correlation_uint8(pred0, stride0, h, w, log_k, LOG_K_SIZE);
  R1 = gen_correlation_uint8(pred1, stride1, h, w, log_k, LOG_K_SIZE);
  for (int i = 0; i < h; ++i) {
    for (int j = 0; j < w; ++j) {
      // this one works better
      double_t edge_diff = fabs(R0[i * w + j]) - fabs(R1[i * w + j]);
      if (edge_diff < DIFFLOG_THR) {
        m = LOG_WEIGHT_1;
      } else {
        m = LOG_WEIGHT_0;
      }
      mask[i * w + j] = which_inverse ? AOM_BLEND_A64_MAX_ALPHA - m : m;
    }
  }
  free(R0);
  free(R1);
#else
  int i, j, m, diff;
  for (i = 0; i < h; ++i) {
    for (j = 0; j < w; ++j) {
      diff =
          abs((int)src0[i * src0_stride + j] - (int)src1[i * src1_stride + j]);
      m = clamp(mask_base + (diff / DIFF_FACTOR), 0, AOM_BLEND_A64_MAX_ALPHA);
      mask[i * w + j] = which_inverse ? AOM_BLEND_A64_MAX_ALPHA - m : m;
    }
  }
#endif  // CONFIG_CTX_ADAPT_LOG_WEIGHT
}

void av1_build_compound_diffwtd_mask_c(uint8_t *mask,
                                       DIFFWTD_MASK_TYPE mask_type,
                                       const uint8_t *src0, int src0_stride,
                                       const uint8_t *src1, int src1_stride,
                                       int h, int w) {
  switch (mask_type) {
    case NORMAL_MASK:
      diffwtd_mask(mask, 0, DIFFWTD_MASK_VAL, src0, src0_stride, src1,
                   src1_stride, h, w);
      break;
    case INVERSE_MASK:
      diffwtd_mask(mask, 1, DIFFWTD_MASK_VAL, src0, src0_stride, src1,
                   src1_stride, h, w);
      break;
    default: assert(0);
  }
}

static AOM_FORCE_INLINE void diffwtd_mask_highbd(
    uint8_t *mask, int which_inverse, int mask_base, const uint16_t *src0,
    int src0_stride, const uint16_t *src1, int src1_stride, int h, int w,
    const unsigned int bd) {
  assert(bd >= 8);
  if (bd == 8) {
    if (which_inverse) {
      for (int i = 0; i < h; ++i) {
        for (int j = 0; j < w; ++j) {
          int diff = abs((int)src0[j] - (int)src1[j]) / DIFF_FACTOR;
          unsigned int m = negative_to_zero(mask_base + diff);
          m = AOMMIN(m, AOM_BLEND_A64_MAX_ALPHA);
          mask[j] = AOM_BLEND_A64_MAX_ALPHA - m;
        }
        src0 += src0_stride;
        src1 += src1_stride;
        mask += w;
      }
    } else {
      for (int i = 0; i < h; ++i) {
        for (int j = 0; j < w; ++j) {
          int diff = abs((int)src0[j] - (int)src1[j]) / DIFF_FACTOR;
          unsigned int m = negative_to_zero(mask_base + diff);
          m = AOMMIN(m, AOM_BLEND_A64_MAX_ALPHA);
          mask[j] = m;
        }
        src0 += src0_stride;
        src1 += src1_stride;
        mask += w;
      }
    }
  } else {
    const unsigned int bd_shift = bd - 8;
    if (which_inverse) {
      for (int i = 0; i < h; ++i) {
        for (int j = 0; j < w; ++j) {
          int diff =
              (abs((int)src0[j] - (int)src1[j]) >> bd_shift) / DIFF_FACTOR;
          unsigned int m = negative_to_zero(mask_base + diff);
          m = AOMMIN(m, AOM_BLEND_A64_MAX_ALPHA);
          mask[j] = AOM_BLEND_A64_MAX_ALPHA - m;
        }
        src0 += src0_stride;
        src1 += src1_stride;
        mask += w;
      }
    } else {
      for (int i = 0; i < h; ++i) {
        for (int j = 0; j < w; ++j) {
          int diff =
              (abs((int)src0[j] - (int)src1[j]) >> bd_shift) / DIFF_FACTOR;
          unsigned int m = negative_to_zero(mask_base + diff);
          m = AOMMIN(m, AOM_BLEND_A64_MAX_ALPHA);
          mask[j] = m;
        }
        src0 += src0_stride;
        src1 += src1_stride;
        mask += w;
      }
    }
  }
}

void av1_build_compound_diffwtd_mask_highbd_c(
    uint8_t *mask, DIFFWTD_MASK_TYPE mask_type, const uint8_t *src0,
    int src0_stride, const uint8_t *src1, int src1_stride, int h, int w,
    int bd) {
  switch (mask_type) {
    case NORMAL_MASK:
      diffwtd_mask_highbd(mask, 0, DIFFWTD_MASK_VAL, CONVERT_TO_SHORTPTR(src0),
                          src0_stride, CONVERT_TO_SHORTPTR(src1), src1_stride,
                          h, w, bd);
      break;
    case INVERSE_MASK:
      diffwtd_mask_highbd(mask, 1, DIFFWTD_MASK_VAL, CONVERT_TO_SHORTPTR(src0),
                          src0_stride, CONVERT_TO_SHORTPTR(src1), src1_stride,
                          h, w, bd);
      break;
    default: assert(0);
  }
}

static void init_wedge_master_masks() {
  int i, j;
  const int w = MASK_MASTER_SIZE;
  const int h = MASK_MASTER_SIZE;
  const int stride = MASK_MASTER_STRIDE;
// Note: index [0] stores the masters, and [1] its complement.
#if USE_PRECOMPUTED_WEDGE_MASK
  // Generate prototype by shifting the masters
  int shift = h / 4;
  for (i = 0; i < h; i += 2) {
    shift_copy(wedge_master_oblique_even,
               &wedge_mask_obl[0][WEDGE_OBLIQUE63][i * stride], shift,
               MASK_MASTER_SIZE);
    shift--;
    shift_copy(wedge_master_oblique_odd,
               &wedge_mask_obl[0][WEDGE_OBLIQUE63][(i + 1) * stride], shift,
               MASK_MASTER_SIZE);
    memcpy(&wedge_mask_obl[0][WEDGE_VERTICAL][i * stride],
           wedge_master_vertical,
           MASK_MASTER_SIZE * sizeof(wedge_master_vertical[0]));
    memcpy(&wedge_mask_obl[0][WEDGE_VERTICAL][(i + 1) * stride],
           wedge_master_vertical,
           MASK_MASTER_SIZE * sizeof(wedge_master_vertical[0]));
  }
#else
  static const double smoother_param = 2.85;
  const int a[2] = { 2, 1 };
  const double asqrt = sqrt(a[0] * a[0] + a[1] * a[1]);
  for (i = 0; i < h; i++) {
    for (j = 0; j < w; ++j) {
      int x = (2 * j + 1 - w);
      int y = (2 * i + 1 - h);
      double d = (a[0] * x + a[1] * y) / asqrt;
      const int msk = (int)rint((1.0 + tanh(d / smoother_param)) * 32);
      wedge_mask_obl[0][WEDGE_OBLIQUE63][i * stride + j] = msk;
      const int mskx = (int)rint((1.0 + tanh(x / smoother_param)) * 32);
      wedge_mask_obl[0][WEDGE_VERTICAL][i * stride + j] = mskx;
    }
  }
#endif  // USE_PRECOMPUTED_WEDGE_MASK
  for (i = 0; i < h; ++i) {
    for (j = 0; j < w; ++j) {
      const int msk = wedge_mask_obl[0][WEDGE_OBLIQUE63][i * stride + j];
      wedge_mask_obl[0][WEDGE_OBLIQUE27][j * stride + i] = msk;
      wedge_mask_obl[0][WEDGE_OBLIQUE117][i * stride + w - 1 - j] =
          wedge_mask_obl[0][WEDGE_OBLIQUE153][(w - 1 - j) * stride + i] =
              (1 << WEDGE_WEIGHT_BITS) - msk;
      wedge_mask_obl[1][WEDGE_OBLIQUE63][i * stride + j] =
          wedge_mask_obl[1][WEDGE_OBLIQUE27][j * stride + i] =
              (1 << WEDGE_WEIGHT_BITS) - msk;
      wedge_mask_obl[1][WEDGE_OBLIQUE117][i * stride + w - 1 - j] =
          wedge_mask_obl[1][WEDGE_OBLIQUE153][(w - 1 - j) * stride + i] = msk;
      const int mskx = wedge_mask_obl[0][WEDGE_VERTICAL][i * stride + j];
      wedge_mask_obl[0][WEDGE_HORIZONTAL][j * stride + i] = mskx;
      wedge_mask_obl[1][WEDGE_VERTICAL][i * stride + j] =
          wedge_mask_obl[1][WEDGE_HORIZONTAL][j * stride + i] =
              (1 << WEDGE_WEIGHT_BITS) - mskx;
    }
  }
}

#if !USE_PRECOMPUTED_WEDGE_SIGN
// 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_ALL; ++sb_type) {
    const int bw = block_size_wide[sb_type];
    const int bh = block_size_high[sb_type];
    const wedge_params_type wedge_params = av1_wedge_params_lookup[sb_type];
    const int wbits = wedge_params.bits;
    const int wtypes = 1 << wbits;
    int i, w;
    if (wbits) {
      for (w = 0; w < wtypes; ++w) {
        // Get the mask master, i.e. index [0]
        const uint8_t *mask = get_wedge_mask_inplace(w, 0, sb_type);
        int avg = 0;
        for (i = 0; i < bw; ++i) avg += mask[i];
        for (i = 1; i < bh; ++i) avg += mask[i * MASK_MASTER_STRIDE];
        avg = (avg + (bw + bh - 1) / 2) / (bw + bh - 1);
        // Default sign of this wedge is 1 if the average < 32, 0 otherwise.
        // If default sign is 1:
        //   If sign requested is 0, we need to flip the sign and return
        //   the complement i.e. index [1] instead. If sign requested is 1
        //   we need to flip the sign and return index [0] instead.
        // If default sign is 0:
        //   If sign requested is 0, we need to return index [0] the master
        //   if sign requested is 1, we need to return the complement index
        //   [1] instead.
        wedge_params.signflip[w] = (avg < 32);
      }
    }
  }
}
#endif  // !USE_PRECOMPUTED_WEDGE_SIGN

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_ALL; ++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 = &av1_wedge_params_lookup[bsize];
    const int wbits = wedge_params->bits;
    if (wbits == 0 || wedge_params->codebook == NULL) continue;
    const int wtypes = 1 << wbits;
    int w;
    for (w = 0; w < wtypes; ++w) {
      mask = get_wedge_mask_inplace(w, 0, bsize);
      aom_convolve_copy(mask, MASK_MASTER_STRIDE, dst, bw, 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, 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();
#if !USE_PRECOMPUTED_WEDGE_SIGN
  init_wedge_signs();
#endif  // !USE_PRECOMPUTED_WEDGE_SIGN
  init_wedge_masks();
}

void av1_dist_wtd_comp_weight_assign(const AV1_COMMON *cm,
                                     const MB_MODE_INFO *mbmi, int order_idx,
                                     int *fwd_offset, int *bck_offset,
                                     int *use_dist_wtd_comp_avg,
                                     int is_compound) {
  assert(fwd_offset != NULL && bck_offset != NULL);
#if CONFIG_OPTFLOW_REFINEMENT && OPFL_DISTWTD_AVG
  if (!is_compound || (mbmi->compound_idx && mbmi->mode <= NEW_NEWMV)) {
#else
  if (!is_compound || mbmi->compound_idx) {
#endif
    *use_dist_wtd_comp_avg = 0;
    return;
  }

  *use_dist_wtd_comp_avg = 1;
  const RefCntBuffer *const bck_buf = get_ref_frame_buf(cm, mbmi->ref_frame[0]);
  const RefCntBuffer *const fwd_buf = get_ref_frame_buf(cm, mbmi->ref_frame[1]);
  const int cur_frame_index = cm->cur_frame->order_hint;
  int bck_frame_index = 0, fwd_frame_index = 0;

  if (bck_buf != NULL) bck_frame_index = bck_buf->order_hint;
  if (fwd_buf != NULL) fwd_frame_index = fwd_buf->order_hint;

  int d0 = clamp(abs(get_relative_dist(&cm->seq_params.order_hint_info,
                                       fwd_frame_index, cur_frame_index)),
                 0, MAX_FRAME_DISTANCE);
  int d1 = clamp(abs(get_relative_dist(&cm->seq_params.order_hint_info,
                                       cur_frame_index, bck_frame_index)),
                 0, MAX_FRAME_DISTANCE);

  const int order = d0 <= d1;

  if (d0 == 0 || d1 == 0) {
    *fwd_offset = quant_dist_lookup_table[order_idx][3][order];
    *bck_offset = quant_dist_lookup_table[order_idx][3][1 - order];
    return;
  }

#if CONFIG_OPTFLOW_REFINEMENT && OPFL_DISTWTD_AVG
  // Typically in COMPOUND_DISTWTD, when d0 and d1 are equal, weights are still
  // chosen to be different (7/16 for ref1 and 9/16 for ref0) because the case
  // where equal weights are applied can be signaled in COMPOUND_AVERAGE
  // anyway. However, in CONFIG_OPTFLOW_REFINEMENT we always apply
  // COMPOUND_DISTWTD without signaling other compound types, so d0 == d1
  // is handled as a special case
  if (mbmi->mode > NEW_NEWMV && d0 == d1) {
    *fwd_offset = 8;
    *bck_offset = 8;
    return;
  }
#endif  // CONFIG_OPTFLOW_REFINEMENT

  int i;
  for (i = 0; i < 3; ++i) {
    int c0 = quant_dist_weight[i][order];
    int c1 = quant_dist_weight[i][!order];
    int d0_c0 = d0 * c0;
    int d1_c1 = d1 * c1;
    if ((d0 > d1 && d0_c0 < d1_c1) || (d0 <= d1 && d0_c0 > d1_c1)) break;
  }

  *fwd_offset = quant_dist_lookup_table[order_idx][i][order];
  *bck_offset = quant_dist_lookup_table[order_idx][i][1 - order];
}

void av1_setup_dst_planes(struct macroblockd_plane *planes,
                          const YV12_BUFFER_CONFIG *src, int mi_row, int mi_col,
                          const int plane_start, const int plane_end,
                          const CHROMA_REF_INFO *chr_ref_info) {
  // We use AOMMIN(num_planes, MAX_MB_PLANE) instead of num_planes to quiet
  // the static analysis warnings.
  for (int i = plane_start; i < AOMMIN(plane_end, MAX_MB_PLANE); ++i) {
    struct macroblockd_plane *const pd = &planes[i];
    const int is_uv = i > 0;
    setup_pred_plane(&pd->dst, src->buffers[i], src->crop_widths[is_uv],
                     src->crop_heights[is_uv], src->strides[is_uv], mi_row,
                     mi_col, NULL, pd->subsampling_x, pd->subsampling_y, is_uv,
                     chr_ref_info);
  }
}

void av1_setup_pre_planes(MACROBLOCKD *xd, int idx,
                          const YV12_BUFFER_CONFIG *src, int mi_row, int mi_col,
                          const struct scale_factors *sf, const int num_planes,
                          const CHROMA_REF_INFO *chr_ref_info) {
  if (src != NULL) {
    // We use AOMMIN(num_planes, MAX_MB_PLANE) instead of num_planes to
    // quiet the static analysis warnings.
    for (int i = 0; i < AOMMIN(num_planes, MAX_MB_PLANE); ++i) {
      struct macroblockd_plane *const pd = &xd->plane[i];
      const int is_uv = i > 0;
      setup_pred_plane(&pd->pre[idx], src->buffers[i], src->crop_widths[is_uv],
                       src->crop_heights[is_uv], src->strides[is_uv], mi_row,
                       mi_col, sf, pd->subsampling_x, pd->subsampling_y, is_uv,
                       chr_ref_info);
    }
  }
}

// obmc_mask_N[overlap_position]
static const uint8_t obmc_mask_1[1] = { 64 };
DECLARE_ALIGNED(2, static const uint8_t, obmc_mask_2[2]) = { 45, 64 };

DECLARE_ALIGNED(4, static const uint8_t, obmc_mask_4[4]) = { 39, 50, 59, 64 };

static const uint8_t obmc_mask_8[8] = { 36, 42, 48, 53, 57, 61, 64, 64 };

static const uint8_t obmc_mask_16[16] = { 34, 37, 40, 43, 46, 49, 52, 54,
                                          56, 58, 60, 61, 64, 64, 64, 64 };

static const uint8_t obmc_mask_32[32] = { 33, 35, 36, 38, 40, 41, 43, 44,
                                          45, 47, 48, 50, 51, 52, 53, 55,
                                          56, 57, 58, 59, 60, 60, 61, 62,
                                          64, 64, 64, 64, 64, 64, 64, 64 };

static const uint8_t obmc_mask_64[64] = {
  33, 34, 35, 35, 36, 37, 38, 39, 40, 40, 41, 42, 43, 44, 44, 44,
  45, 46, 47, 47, 48, 49, 50, 51, 51, 51, 52, 52, 53, 54, 55, 56,
  56, 56, 57, 57, 58, 58, 59, 60, 60, 60, 60, 60, 61, 62, 62, 62,
  62, 62, 63, 63, 63, 63, 64, 64, 64, 64, 64, 64, 64, 64, 64, 64,
};

const uint8_t *av1_get_obmc_mask(int length) {
  switch (length) {
    case 1: return obmc_mask_1;
    case 2: return obmc_mask_2;
    case 4: return obmc_mask_4;
    case 8: return obmc_mask_8;
    case 16: return obmc_mask_16;
    case 32: return obmc_mask_32;
    case 64: return obmc_mask_64;
    default: assert(0); return NULL;
  }
}

static INLINE void increment_int_ptr(MACROBLOCKD *xd, int rel_mi_rc,
                                     uint8_t mi_hw, MB_MODE_INFO *mi,
                                     void *fun_ctxt, const int num_planes) {
  (void)xd;
  (void)rel_mi_rc;
  (void)mi_hw;
  (void)mi;
  ++*(int *)fun_ctxt;
  (void)num_planes;
}

void av1_count_overlappable_neighbors(const AV1_COMMON *cm, MACROBLOCKD *xd) {
  MB_MODE_INFO *mbmi = xd->mi[0];

  mbmi->overlappable_neighbors[0] = 0;
  mbmi->overlappable_neighbors[1] = 0;

#if CONFIG_EXT_WARP && CONFIG_SUB8X8_WARP
  if (block_size_wide[mbmi->sb_type] > 4)
    foreach_overlappable_nb_above(cm, xd, INT_MAX, increment_int_ptr,
                                  &mbmi->overlappable_neighbors[0]);
  else
    foreach_inter_nb_above(xd, &mbmi->overlappable_neighbors[0]);

  if (block_size_high[mbmi->sb_type] > 4)
    foreach_overlappable_nb_left(cm, xd, INT_MAX, increment_int_ptr,
                                 &mbmi->overlappable_neighbors[1]);
  else
    foreach_inter_nb_left(xd, &mbmi->overlappable_neighbors[1]);
#else
  if (!is_motion_variation_allowed_bsize(mbmi->sb_type, xd->mi_row, xd->mi_col))
    return;

  foreach_overlappable_nb_above(cm, xd, INT_MAX, increment_int_ptr,
                                &mbmi->overlappable_neighbors[0]);
  foreach_overlappable_nb_left(cm, xd, INT_MAX, increment_int_ptr,
                               &mbmi->overlappable_neighbors[1]);
#endif  // CONFIG_EXT_WARP && CONFIG_SUB8X8_WARP
}

// HW does not support < 4x4 prediction. To limit the bandwidth requirement, if
// block-size of current plane is smaller than 8x8, always only blend with the
// left neighbor(s) (skip blending with the above side).
#define DISABLE_CHROMA_U8X8_OBMC 0  // 0: one-sided obmc; 1: disable

int av1_skip_u4x4_pred_in_obmc(int mi_row, int mi_col, BLOCK_SIZE bsize,
                               const struct macroblockd_plane *pd, int dir) {
  (void)mi_row;
  (void)mi_col;
  assert(is_motion_variation_allowed_bsize(bsize, mi_row, mi_col));

  const BLOCK_SIZE bsize_plane =
      get_plane_block_size(bsize, pd->subsampling_x, pd->subsampling_y);
  switch (bsize_plane) {
#if DISABLE_CHROMA_U8X8_OBMC
    case BLOCK_4X4:
    case BLOCK_8X4:
    case BLOCK_4X8: return 1; break;
#else
    case BLOCK_4X4:
    case BLOCK_8X4:
    case BLOCK_4X8: return dir == 0; break;
#endif
    default: return 0;
  }
}

void av1_modify_neighbor_predictor_for_obmc(MB_MODE_INFO *mbmi) {
  mbmi->ref_frame[1] = NONE_FRAME;
  mbmi->interinter_comp.type = COMPOUND_AVERAGE;

  return;
}

struct obmc_inter_pred_ctxt {
  uint8_t **adjacent;
  int *adjacent_stride;
};

static INLINE void build_obmc_inter_pred_above(MACROBLOCKD *xd, int rel_mi_col,
                                               uint8_t above_mi_width,
                                               MB_MODE_INFO *above_mi,
                                               void *fun_ctxt,
                                               const int num_planes) {
  (void)above_mi;
  struct obmc_inter_pred_ctxt *ctxt = (struct obmc_inter_pred_ctxt *)fun_ctxt;
  const BLOCK_SIZE bsize = xd->mi[0]->sb_type;
  const int is_hbd = is_cur_buf_hbd(xd);
  const int mi_row = -xd->mb_to_top_edge >> (3 + MI_SIZE_LOG2);
  const int mi_col = -xd->mb_to_left_edge >> (3 + MI_SIZE_LOG2);
  const int overlap =
      AOMMIN(block_size_high[bsize], block_size_high[BLOCK_64X64]) >> 1;

  for (int plane = 0; plane < num_planes; ++plane) {
    const struct macroblockd_plane *pd = &xd->plane[plane];
    const int bw = (above_mi_width * MI_SIZE) >> pd->subsampling_x;
    const int bh = overlap >> pd->subsampling_y;
    const int plane_col = (rel_mi_col * MI_SIZE) >> pd->subsampling_x;

    if (av1_skip_u4x4_pred_in_obmc(mi_row, mi_col, bsize, pd, 0)) continue;

    const int dst_stride = pd->dst.stride;
    uint8_t *const dst = &pd->dst.buf[plane_col];
    const int tmp_stride = ctxt->adjacent_stride[plane];
    const uint8_t *const tmp = &ctxt->adjacent[plane][plane_col];
    const uint8_t *const mask = av1_get_obmc_mask(bh);

    if (is_hbd)
      aom_highbd_blend_a64_vmask(dst, dst_stride, dst, dst_stride, tmp,
                                 tmp_stride, mask, bw, bh, xd->bd);
    else
      aom_blend_a64_vmask(dst, dst_stride, dst, dst_stride, tmp, tmp_stride,
                          mask, bw, bh);
  }
}

static INLINE void build_obmc_inter_pred_left(MACROBLOCKD *xd, int rel_mi_row,
                                              uint8_t left_mi_height,
                                              MB_MODE_INFO *left_mi,
                                              void *fun_ctxt,
                                              const int num_planes) {
  (void)left_mi;
  struct obmc_inter_pred_ctxt *ctxt = (struct obmc_inter_pred_ctxt *)fun_ctxt;
  const BLOCK_SIZE bsize = xd->mi[0]->sb_type;
  const int overlap =
      AOMMIN(block_size_wide[bsize], block_size_wide[BLOCK_64X64]) >> 1;
  const int is_hbd = is_cur_buf_hbd(xd);
  const int mi_row = -xd->mb_to_top_edge >> (3 + MI_SIZE_LOG2);
  const int mi_col = -xd->mb_to_left_edge >> (3 + MI_SIZE_LOG2);

  for (int plane = 0; plane < num_planes; ++plane) {
    const struct macroblockd_plane *pd = &xd->plane[plane];
    const int bw = overlap >> pd->subsampling_x;
    const int bh = (left_mi_height * MI_SIZE) >> pd->subsampling_y;
    const int plane_row = (rel_mi_row * MI_SIZE) >> pd->subsampling_y;

    if (av1_skip_u4x4_pred_in_obmc(mi_row, mi_col, bsize, pd, 1)) continue;

    const int dst_stride = pd->dst.stride;
    uint8_t *const dst = &pd->dst.buf[plane_row * dst_stride];
    const int tmp_stride = ctxt->adjacent_stride[plane];
    const uint8_t *const tmp = &ctxt->adjacent[plane][plane_row * tmp_stride];
    const uint8_t *const mask = av1_get_obmc_mask(bw);

    if (is_hbd)
      aom_highbd_blend_a64_hmask(dst, dst_stride, dst, dst_stride, tmp,
                                 tmp_stride, mask, bw, bh, xd->bd);
    else
      aom_blend_a64_hmask(dst, dst_stride, dst, dst_stride, tmp, tmp_stride,
                          mask, bw, bh);
  }
}

// This function combines motion compensated predictions that are generated
// by top/left neighboring blocks' inter predictors with the regular inter
// prediction. We assume the original prediction (bmc) is stored in
// xd->plane[].dst.buf
void av1_build_obmc_inter_prediction(const AV1_COMMON *cm, MACROBLOCKD *xd,
                                     uint8_t *above[MAX_MB_PLANE],
                                     int above_stride[MAX_MB_PLANE],
                                     uint8_t *left[MAX_MB_PLANE],
                                     int left_stride[MAX_MB_PLANE]) {
  const BLOCK_SIZE bsize = xd->mi[0]->sb_type;

  // handle above row
  struct obmc_inter_pred_ctxt ctxt_above = { above, above_stride };
  foreach_overlappable_nb_above(cm, xd,
                                max_neighbor_obmc[mi_size_wide_log2[bsize]],
                                build_obmc_inter_pred_above, &ctxt_above);

  // handle left column
  struct obmc_inter_pred_ctxt ctxt_left = { left, left_stride };
  foreach_overlappable_nb_left(cm, xd,
                               max_neighbor_obmc[mi_size_high_log2[bsize]],
                               build_obmc_inter_pred_left, &ctxt_left);
}

void av1_setup_build_prediction_by_above_pred(
    MACROBLOCKD *xd, int rel_mi_col, uint8_t above_mi_width,
    MB_MODE_INFO *above_mbmi, struct build_prediction_ctxt *ctxt,
    const int num_planes) {
  const int above_mi_col = xd->mi_col + rel_mi_col;

  av1_modify_neighbor_predictor_for_obmc(above_mbmi);

  for (int j = 0; j < num_planes; ++j) {
    struct macroblockd_plane *const pd = &xd->plane[j];
    setup_pred_plane(&pd->dst, ctxt->tmp_buf[j], ctxt->tmp_width[j],
                     ctxt->tmp_height[j], ctxt->tmp_stride[j], 0, rel_mi_col,
                     NULL, pd->subsampling_x, pd->subsampling_y, j > 0, NULL);
  }

  const int num_refs = 1 + has_second_ref(above_mbmi);

  for (int ref = 0; ref < num_refs; ++ref) {
    const MV_REFERENCE_FRAME frame = above_mbmi->ref_frame[ref];

    const RefCntBuffer *const ref_buf = get_ref_frame_buf(ctxt->cm, frame);
    const struct scale_factors *const sf =
        get_ref_scale_factors_const(ctxt->cm, frame);
    xd->block_ref_scale_factors[ref] = sf;
    if ((!av1_is_valid_scale(sf)))
      aom_internal_error(xd->error_info, AOM_CODEC_UNSUP_BITSTREAM,
                         "Reference frame has invalid dimensions");
    av1_setup_pre_planes(xd, ref, &ref_buf->buf, xd->mi_row, above_mi_col, sf,
                         num_planes, NULL);
  }

  xd->mb_to_left_edge = 8 * MI_SIZE * (-above_mi_col);
  xd->mb_to_right_edge = ctxt->mb_to_far_edge +
                         (xd->n4_w - rel_mi_col - above_mi_width) * MI_SIZE * 8;
}

void av1_setup_build_prediction_by_left_pred(MACROBLOCKD *xd, int rel_mi_row,
                                             uint8_t left_mi_height,
                                             MB_MODE_INFO *left_mbmi,
                                             struct build_prediction_ctxt *ctxt,
                                             const int num_planes) {
  const int left_mi_row = xd->mi_row + rel_mi_row;

  av1_modify_neighbor_predictor_for_obmc(left_mbmi);

  for (int j = 0; j < num_planes; ++j) {
    struct macroblockd_plane *const pd = &xd->plane[j];
    setup_pred_plane(&pd->dst, ctxt->tmp_buf[j], ctxt->tmp_width[j],
                     ctxt->tmp_height[j], ctxt->tmp_stride[j], rel_mi_row, 0,
                     NULL, pd->subsampling_x, pd->subsampling_y, j > 0, NULL);
  }

  const int num_refs = 1 + has_second_ref(left_mbmi);

  for (int ref = 0; ref < num_refs; ++ref) {
    const MV_REFERENCE_FRAME frame = left_mbmi->ref_frame[ref];

    const RefCntBuffer *const ref_buf = get_ref_frame_buf(ctxt->cm, frame);
    const struct scale_factors *const ref_scale_factors =
        get_ref_scale_factors_const(ctxt->cm, frame);

    xd->block_ref_scale_factors[ref] = ref_scale_factors;
    if ((!av1_is_valid_scale(ref_scale_factors)))
      aom_internal_error(xd->error_info, AOM_CODEC_UNSUP_BITSTREAM,
                         "Reference frame has invalid dimensions");
    av1_setup_pre_planes(xd, ref, &ref_buf->buf, left_mi_row, xd->mi_col,
                         ref_scale_factors, num_planes, NULL);
  }

  xd->mb_to_top_edge = 8 * MI_SIZE * (-left_mi_row);
  xd->mb_to_bottom_edge =
      ctxt->mb_to_far_edge +
      (xd->n4_h - rel_mi_row - left_mi_height) * MI_SIZE * 8;
}

/* clang-format off */
static const uint8_t ii_weights1d[MAX_SB_SIZE] = {
  60, 58, 56, 54, 52, 50, 48, 47, 45, 44, 42, 41, 39, 38, 37, 35, 34, 33, 32,
  31, 30, 29, 28, 27, 26, 25, 24, 23, 22, 22, 21, 20, 19, 19, 18, 18, 17, 16,
  16, 15, 15, 14, 14, 13, 13, 12, 12, 12, 11, 11, 10, 10, 10,  9,  9,  9,  8,
  8,  8,  8,  7,  7,  7,  7,  6,  6,  6,  6,  6,  5,  5,  5,  5,  5,  4,  4,
  4,  4,  4,  4,  4,  4,  3,  3,  3,  3,  3,  3,  3,  3,  3,  2,  2,  2,  2,
  2,  2,  2,  2,  2,  2,  2,  2,  2,  2,  2,  1,  1,  1,  1,  1,  1,  1,  1,
  1,  1,  1,  1,  1,  1,  1,  1,  1,  1,  1,  1,  1,  1
};
static uint8_t ii_size_scales[BLOCK_SIZES_ALL] = {
    32, 16, 16, 16, 8, 8, 8, 4,
    4,  4,  2,  2,  2, 1, 1, 1,
    8,  8,  4,  4,  2, 2,
#if CONFIG_FLEX_PARTITION
    8,  8,  4,  4,  8, 8,
#endif  // CONFIG_FLEX_PARTITION
};
/* clang-format on */

static void build_smooth_interintra_mask(uint8_t *mask, int stride,
                                         BLOCK_SIZE plane_bsize,
                                         INTERINTRA_MODE mode) {
  int i, j;
  const int bw = block_size_wide[plane_bsize];
  const int bh = block_size_high[plane_bsize];
  const int size_scale = ii_size_scales[plane_bsize];

  switch (mode) {
    case II_V_PRED:
      for (i = 0; i < bh; ++i) {
        memset(mask, ii_weights1d[i * size_scale], bw * sizeof(mask[0]));
        mask += stride;
      }
      break;

    case II_H_PRED:
      for (i = 0; i < bh; ++i) {
        for (j = 0; j < bw; ++j) mask[j] = ii_weights1d[j * size_scale];
        mask += stride;
      }
      break;

    case II_SMOOTH_PRED:
      for (i = 0; i < bh; ++i) {
        for (j = 0; j < bw; ++j)
          mask[j] = ii_weights1d[(i < j ? i : j) * size_scale];
        mask += stride;
      }
      break;

    case II_DC_PRED:
#if CONFIG_ILLUM_MCOMP
    case II_ILLUM_MCOMP_PRED:
#endif  // CONFIG_ILLUM_MCOMP
    default:
      for (i = 0; i < bh; ++i) {
        memset(mask, 32, bw * sizeof(mask[0]));
        mask += stride;
      }
      break;
  }
}

#if CONFIG_ILLUM_MCOMP

// Only analyze the 4 pixel border around the inter/intra predictors.
#define ILLUM_MCOMP_BORDER 4

// Toggle between 'old' method (using difference of averages) and
// 'new' method (linear regression).
#define ILLUM_MCOMP_OLD 0
#if ILLUM_MCOMP_OLD
// Defines a function that can be used to obtain the average of the
// extended region.
#define ILLUM_MCOMP_COMPUTE_DC(INT_TYPE, suffix)                               \
  static int illum_mcomp_compute_dc_##suffix(const INT_TYPE *pred, int stride, \
                                             int bw, int bh) {                 \
    const int border = ILLUM_MCOMP_BORDER;                                     \
    int sum = 0;                                                               \
    for (int i = -border; i < 0; ++i) {                                        \
      for (int j = -border; j < bw; ++j) {                                     \
        sum += pred[i * stride + j];                                           \
      }                                                                        \
    }                                                                          \
    for (int i = 0; i < bh; ++i) {                                             \
      for (int j = -border; j < 0; ++j) {                                      \
        sum += pred[i * stride + j];                                           \
      }                                                                        \
    }                                                                          \
    /* Add "0.5" so we round "half-up" instead of "down". */                   \
    const int count = border * (border + bw) + bh * border;                    \
    int expected_dc = (sum + (count >> 1)) / count;                            \
    return expected_dc;                                                        \
  }

ILLUM_MCOMP_COMPUTE_DC(uint8_t, lowbd);
ILLUM_MCOMP_COMPUTE_DC(uint16_t, highbd);

#endif  // ILLUM_MCOMP_OLD

#define ILLUM_MCOMP_PREC_BITS 8
#define ILLUM_MCOMP_PREC (1 << ILLUM_MCOMP_PREC_BITS)
static void illum_mcomp_linear_model_lowbd(const uint8_t *inter_pred,
                                           int inter_stride,
                                           const uint8_t *intra_pred,
                                           int intra_stride, int bw, int bh,
                                           int bd, int *alpha, int *beta) {
  assert(bd == 8);
#if ILLUM_MCOMP_OLD
  *alpha = 1 << ILLUM_MCOMP_PREC_BITS;
  int intra_dc = illum_mcomp_compute_dc_lowbd(intra_pred, intra_stride, bw, bh);
  int inter_dc = illum_mcomp_compute_dc_lowbd(inter_pred, inter_stride, bw, bh);
  *beta = (intra_dc - inter_dc) << ILLUM_MCOMP_PREC_BITS;
  return;
#endif  // ILLUM_MCOMP_OLD

  int64_t n = 0;
  int64_t sx2 = 0;
  int64_t sx = 0;
  int64_t sxy = 0;
  int64_t sy = 0;
  const int border = ILLUM_MCOMP_BORDER;
  for (int i = -border; i < 0; ++i) {
    for (int j = -border; j < bw; ++j) {
      const int x = inter_pred[i * inter_stride + j];
      const int y = intra_pred[i * intra_stride + j];
      sx += x;
      sy += y;
      sx2 += x * x;
      sxy += x * y;
      n++;
    }
  }
  for (int i = 0; i < bh; ++i) {
    for (int j = -border; j < 0; ++j) {
      const int x = inter_pred[i * inter_stride + j];
      const int y = intra_pred[i * intra_stride + j];
      sx += x;
      sy += y;
      sx2 += x * x;
      sxy += x * y;
      n++;
    }
  }
  const int64_t Pa = (n * sxy - sx * sy) * ILLUM_MCOMP_PREC;
  const int64_t Pb = (-sx * sxy + sx2 * sy) * ILLUM_MCOMP_PREC;
  const int64_t D = sx2 * n - sx * sx;

  int64_t a;
  int64_t b;
  if (D != 0) {
    a = DIVIDE_AND_ROUND_SIGNED(Pa, D);
    b = DIVIDE_AND_ROUND_SIGNED(Pb, D);
  } else {
    // Set to extreme values.
    a = ILLUM_MCOMP_PREC * 4;
    const int sign = (Pb < 0) ? -1 : 1;
    b = sign * (1 << (bd - 2)) * ILLUM_MCOMP_PREC;
  }
  // Clamp to reasonable range
  *alpha = (int)clamp64(a, ILLUM_MCOMP_PREC / 4, ILLUM_MCOMP_PREC * 4);
  *beta = (int)clamp64(b, -(1 << (bd - 2)) * ILLUM_MCOMP_PREC,
                       (1 << (bd - 2)) * ILLUM_MCOMP_PREC);
}

static void illum_mcomp_linear_model_highbd(const uint16_t *inter_pred,
                                            int inter_stride,
                                            const uint16_t *intra_pred,
                                            int intra_stride, int bw, int bh,
                                            int bd, int *alpha, int *beta) {
  assert(bd > 8);
#if ILLUM_MCOMP_OLD
  *alpha = 1 << ILLUM_MCOMP_PREC_BITS;
  int intra_dc =
      illum_mcomp_compute_dc_highbd(intra_pred, intra_stride, bw, bh);
  int inter_dc =
      illum_mcomp_compute_dc_highbd(inter_pred, inter_stride, bw, bh);
  *beta = (intra_dc - inter_dc) << ILLUM_MCOMP_PREC_BITS;
  return;
#endif  // ILLUM_MCOMP_OLD

  int64_t n = 0;
  int64_t sx2 = 0;
  int64_t sx = 0;
  int64_t sxy = 0;
  int64_t sy = 0;
  const int border = ILLUM_MCOMP_BORDER;
  for (int i = -border; i < 0; ++i) {
    for (int j = -border; j < bw; ++j) {
      const int x = inter_pred[i * inter_stride + j];
      const int y = intra_pred[i * intra_stride + j];
      sx += x;
      sy += y;
      sx2 += x * x;
      sxy += x * y;
      n++;
    }
  }
  for (int i = 0; i < bh; ++i) {
    for (int j = -border; j < 0; ++j) {
      const int x = inter_pred[i * inter_stride + j];
      const int y = intra_pred[i * intra_stride + j];
      sx += x;
      sy += y;
      sx2 += x * x;
      sxy += x * y;
      n++;
    }
  }
  const int64_t Pa = (n * sxy - sx * sy) * ILLUM_MCOMP_PREC;
  const int64_t Pb = (-sx * sxy + sx2 * sy) * ILLUM_MCOMP_PREC;
  const int64_t D = sx2 * n - sx * sx;

  int64_t a;
  int64_t b;
  if (D != 0) {
    a = DIVIDE_AND_ROUND_SIGNED(Pa, D);
    b = DIVIDE_AND_ROUND_SIGNED(Pb, D);
  } else {
    // Set to extreme values.
    a = ILLUM_MCOMP_PREC * 4;
    const int sign = (Pb < 0) ? -1 : 1;
    b = sign * (1 << (bd - 2)) * ILLUM_MCOMP_PREC;
  }
  // Clamp to reasonable range
  *alpha = (int)clamp64(a, ILLUM_MCOMP_PREC / 4, ILLUM_MCOMP_PREC * 4);
  *beta = (int)clamp64(b, -(1 << (bd - 2)) * ILLUM_MCOMP_PREC,
                       (1 << (bd - 2)) * ILLUM_MCOMP_PREC);
}

static void illum_combine_interintra(
    int8_t use_wedge_interintra, int8_t wedge_index, int8_t wedge_sign,
    BLOCK_SIZE bsize, BLOCK_SIZE plane_bsize, uint8_t *comp_pred,
    int comp_stride, const uint8_t *inter_pred, int inter_stride,
    const uint8_t *intra_pred, int intra_stride, int border) {
  assert(border >= ILLUM_MCOMP_BORDER);
  (void)border;
  const int bw = block_size_wide[plane_bsize];
  const int bh = block_size_high[plane_bsize];

  // Compute the linearly projected predictor.
  uint8_t *projected = comp_pred;
  int projected_stride = comp_stride;
  const bool wedge_case =
      use_wedge_interintra && is_interintra_wedge_used(bsize);

  if (wedge_case) {
    projected = aom_memalign(64, bw * bh * sizeof(*projected));
    projected_stride = bw;
  }

  int alpha, beta;
  illum_mcomp_linear_model_lowbd(inter_pred, inter_stride, intra_pred,
                                 intra_stride, bw, bh, 8, &alpha, &beta);
  for (int i = 0; i < bh; ++i) {
    for (int j = 0; j < bw; ++j) {
      int32_t r = inter_pred[i * inter_stride + j];
      r *= alpha;
      r += beta;
      r >>= ILLUM_MCOMP_PREC_BITS;
      projected[i * projected_stride + j] = clip_pixel_highbd(r, 8);
    }
  }

  // If this is a wedge case, blend the intra-predictor.
  if (wedge_case) {
    const uint8_t *mask =
        av1_get_contiguous_soft_mask(wedge_index, wedge_sign, bsize);
    const int subw = 2 * mi_size_wide[bsize] == bw;
    const int subh = 2 * mi_size_high[bsize] == bh;
    aom_blend_a64_mask(comp_pred, comp_stride, intra_pred, intra_stride,
                       projected, projected_stride, mask,
                       block_size_wide[bsize], bw, bh, subw, subh);
    aom_free(projected);
  }
}

static void illum_combine_interintra_highbd(
    int8_t use_wedge_interintra, int8_t wedge_index, int8_t wedge_sign,
    BLOCK_SIZE bsize, BLOCK_SIZE plane_bsize, uint8_t *comp_pred8,
    int comp_stride, const uint8_t *inter_pred8, int inter_stride,
    const uint8_t *intra_pred8, int intra_stride, int bd, int border) {
  assert(border >= ILLUM_MCOMP_BORDER);
  (void)border;
  const int bw = block_size_wide[plane_bsize];
  const int bh = block_size_high[plane_bsize];

  uint16_t *projected = CONVERT_TO_SHORTPTR(comp_pred8);
  int projected_stride = comp_stride;
  const bool wedge_case =
      use_wedge_interintra && is_interintra_wedge_used(bsize);
  uint16_t *inter_pred = CONVERT_TO_SHORTPTR(inter_pred8);
  uint16_t *intra_pred = CONVERT_TO_SHORTPTR(intra_pred8);

  if (wedge_case) {
    projected = aom_memalign(64, bw * bh * sizeof(*projected));
    projected_stride = bw;
  }
  int alpha, beta;
  illum_mcomp_linear_model_highbd(inter_pred, inter_stride, intra_pred,
                                  intra_stride, bw, bh, bd, &alpha, &beta);
  for (int i = 0; i < bh; ++i) {
    for (int j = 0; j < bw; ++j) {
      int32_t r = inter_pred[i * inter_stride + j];
      r *= alpha;
      r += beta;
      r >>= ILLUM_MCOMP_PREC_BITS;
      projected[i * projected_stride + j] = clip_pixel_highbd(r, bd);
    }
  }

  if (wedge_case) {
    const uint8_t *mask =
        av1_get_contiguous_soft_mask(wedge_index, wedge_sign, bsize);
    const int subh = 2 * mi_size_high[bsize] == bh;
    const int subw = 2 * mi_size_wide[bsize] == bw;
    aom_highbd_blend_a64_mask(comp_pred8, comp_stride, intra_pred8,
                              intra_stride, CONVERT_TO_BYTEPTR(projected),
                              projected_stride, mask, block_size_wide[bsize],
                              bw, bh, subw, subh, bd);
    aom_free(projected);
  }
}
#endif  // CONFIG_ILLUM_MCOMP

static void combine_interintra(INTERINTRA_MODE mode,
                               int8_t use_wedge_interintra, int8_t wedge_index,
                               int8_t wedge_sign, BLOCK_SIZE bsize,
                               BLOCK_SIZE plane_bsize, int plane,
                               uint8_t *comppred, int compstride,
                               const uint8_t *interpred, int interstride,
                               const uint8_t *intrapred, int intrastride,
                               int border, bool is_ii_ml_mode) {
  assert(plane_bsize < BLOCK_SIZES_ALL);
  (void)plane;
  (void)border;
  (void)is_ii_ml_mode;
#if CONFIG_ILLUM_MCOMP
  if (mode == II_ILLUM_MCOMP_PRED) {
    illum_combine_interintra(use_wedge_interintra, wedge_index, wedge_sign,
                             bsize, plane_bsize, comppred, compstride,
                             interpred, interstride, intrapred, intrastride,
                             border);
    return;
  }
#endif  // CONFIG_ILLUM_MCOMP
#if CONFIG_INTERINTRA_ML
  if (is_ii_ml_mode) {
    assert(!use_wedge_interintra);
    av1_combine_interintra_ml(mode, bsize, plane_bsize, plane, comppred,
                              compstride, interpred, interstride, intrapred,
                              intrastride, border);
    return;
  }
#endif  // CONFIG_INTERINTRA_ML

  const int bw = block_size_wide[plane_bsize];
  const int bh = block_size_high[plane_bsize];

  if (use_wedge_interintra) {
    if (is_interintra_wedge_used(bsize)) {
      const uint8_t *mask =
          av1_get_contiguous_soft_mask(wedge_index, wedge_sign, bsize);
      const int subw = 2 * mi_size_wide[bsize] == bw;
      const int subh = 2 * mi_size_high[bsize] == bh;
      aom_blend_a64_mask(comppred, compstride, intrapred, intrastride,
                         interpred, interstride, mask, block_size_wide[bsize],
                         bw, bh, subw, subh);
    }
    return;
  }

  uint8_t mask[MAX_SB_SQUARE];
  build_smooth_interintra_mask(mask, bw, plane_bsize, mode);
  aom_blend_a64_mask(comppred, compstride, intrapred, intrastride, interpred,
                     interstride, mask, bw, bw, bh, 0, 0);
}

static void combine_interintra_highbd(
    INTERINTRA_MODE mode, int8_t use_wedge_interintra, int8_t wedge_index,
    int8_t wedge_sign, BLOCK_SIZE bsize, BLOCK_SIZE plane_bsize, int plane,
    uint8_t *comppred8, int compstride, const uint8_t *interpred8,
    int interstride, const uint8_t *intrapred8, int intrastride, int bd,
    int border, bool is_ii_ml_mode) {
  assert(plane_bsize < BLOCK_SIZES_ALL);
  (void)plane;
  (void)border;
  (void)is_ii_ml_mode;
#if CONFIG_ILLUM_MCOMP
  if (mode == II_ILLUM_MCOMP_PRED) {
    illum_combine_interintra_highbd(use_wedge_interintra, wedge_index,
                                    wedge_sign, bsize, plane_bsize, comppred8,
                                    compstride, interpred8, interstride,
                                    intrapred8, intrastride, bd, border);
    return;
  }
#endif  // CONFIG_ILLUM_MCOMP
#if CONFIG_INTERINTRA_ML
  if (is_ii_ml_mode) {
    assert(!use_wedge_interintra);
    av1_combine_interintra_ml_highbd(mode, bsize, plane_bsize, plane, comppred8,
                                     compstride, interpred8, interstride,
                                     intrapred8, intrastride, bd, border);
    return;
  }
#endif  // CONFIG_INTERINTRA_ML
  const int bw = block_size_wide[plane_bsize];
  const int bh = block_size_high[plane_bsize];

  if (use_wedge_interintra) {
    if (is_interintra_wedge_used(bsize)) {
      const uint8_t *mask =
          av1_get_contiguous_soft_mask(wedge_index, wedge_sign, bsize);
      const int subh = 2 * mi_size_high[bsize] == bh;
      const int subw = 2 * mi_size_wide[bsize] == bw;
      aom_highbd_blend_a64_mask(comppred8, compstride, intrapred8, intrastride,
                                interpred8, interstride, mask,
                                block_size_wide[bsize], bw, bh, subw, subh, bd);
    }
    return;
  }

  uint8_t mask[MAX_SB_SQUARE];
  build_smooth_interintra_mask(mask, bw, plane_bsize, mode);
  aom_highbd_blend_a64_mask(comppred8, compstride, intrapred8, intrastride,
                            interpred8, interstride, mask, bw, bw, bh, 0, 0,
                            bd);
}

void av1_build_intra_predictors_for_interintra(
    const AV1_COMMON *cm, MACROBLOCKD *xd, BLOCK_SIZE bsize, int plane,
    const BUFFER_SET *ctx, uint8_t *dst, int dst_stride, int border) {
  struct macroblockd_plane *const pd = &xd->plane[plane];
  const int ssx = xd->plane[plane].subsampling_x;
  const int ssy = xd->plane[plane].subsampling_y;
  MB_MODE_INFO *mbmi = xd->mi[0];
  const BLOCK_SIZE bsize_base =
      plane ? mbmi->chroma_ref_info.bsize_base : bsize;
  BLOCK_SIZE plane_bsize = get_plane_block_size(bsize_base, ssx, ssy);
#if CONFIG_DERIVED_INTRA_MODE
  const int use_derived_mode = mbmi->use_derived_intra_mode[0];
  const PREDICTION_MODE mode =
      use_derived_mode ? av1_get_derived_intra_mode(xd, bsize, mbmi)
                       : interintra_to_intra_mode[mbmi->interintra_mode];
#else
  const PREDICTION_MODE mode = interintra_to_intra_mode[mbmi->interintra_mode];
#endif
  assert(mbmi->angle_delta[PLANE_TYPE_Y] == 0);
  assert(mbmi->angle_delta[PLANE_TYPE_UV] == 0);
  assert(mbmi->filter_intra_mode_info.use_filter_intra == 0);
  assert(mbmi->use_intrabc == 0);
  assert(plane_bsize < BLOCK_SIZES_ALL);

  const TX_SIZE tx_size = max_txsize_rect_lookup[plane_bsize];
  av1_predict_intra_block(
      cm, xd, pd->width, pd->height, tx_size, mode, 0, 0, FILTER_INTRA_MODES,
#if CONFIG_ADAPT_FILTER_INTRA
      ADAPT_FILTER_INTRA_MODES,
#endif
#if CONFIG_DERIVED_INTRA_MODE
      use_derived_mode ? mbmi->derived_angle : 0,
#endif  // CONFIG_DERIVED_INTRA_MODE
      ctx->plane[plane], ctx->stride[plane], dst, dst_stride, 0, 0, plane);

  if (border > 0) {
    av1_extend_intra_border(ctx->plane[plane], ctx->stride[plane], dst,
                            dst_stride, av1_intra_top_available(xd, plane),
                            av1_intra_right_unavailable(xd, plane, tx_size),
                            av1_intra_left_available(xd, plane),
                            av1_intra_bottom_unavailable(xd, plane, tx_size),
                            xd->plane[plane].width, xd->plane[plane].height,
                            border, xd->bd, is_cur_buf_hbd(xd));
  }
}

void av1_combine_interintra(MACROBLOCKD *xd, BLOCK_SIZE bsize, int plane,
                            const uint8_t *inter_pred, int inter_stride,
                            const uint8_t *intra_pred, int intra_stride,
                            int border) {
  assert(border >= 0);
  const INTERINTRA_MODE mode = xd->mi[0]->interintra_mode;
  const int ssx = xd->plane[plane].subsampling_x;
  const int ssy = xd->plane[plane].subsampling_y;
  const BLOCK_SIZE bsize_base =
      plane ? xd->mi[0]->chroma_ref_info.bsize_base : bsize;
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize_base, ssx, ssy);
#if CONFIG_INTERINTRA_ML
  const bool is_ii_ml_mode =
      is_interintra_ml_supported(xd, xd->mi[0]->use_wedge_interintra);
#else
  const bool is_ii_ml_mode = false;
#endif
  if (is_cur_buf_hbd(xd)) {
    combine_interintra_highbd(
        mode, xd->mi[0]->use_wedge_interintra,
        xd->mi[0]->interintra_wedge_index, INTERINTRA_WEDGE_SIGN, bsize,
        plane_bsize, plane, xd->plane[plane].dst.buf,
        xd->plane[plane].dst.stride, inter_pred, inter_stride, intra_pred,
        intra_stride, xd->bd, border, is_ii_ml_mode);
    return;
  }
  combine_interintra(mode, xd->mi[0]->use_wedge_interintra,
                     xd->mi[0]->interintra_wedge_index, INTERINTRA_WEDGE_SIGN,
                     bsize, plane_bsize, plane, xd->plane[plane].dst.buf,
                     xd->plane[plane].dst.stride, inter_pred, inter_stride,
                     intra_pred, intra_stride, border, is_ii_ml_mode);
}

void av1_alloc_buf_with_border(uint8_t **buf, int *buf_stride, int border,
                               bool is_hbd) {
  const int border16 = border % 16 == 0 ? border : 16 * (1 + border / 16);
  *buf_stride = MAX_SB_SIZE + border16;
  assert(*buf_stride % 16 == 0);
  const int intrapred_buf_size = *buf_stride * *buf_stride;
  const int bytes_per_val = is_hbd ? sizeof(uint16_t) : sizeof(uint8_t);
  *buf = aom_memalign(16, intrapred_buf_size * bytes_per_val);
  // Offset past the border.
  *buf += (border16 + border16 * *buf_stride) * bytes_per_val;
}

void av1_free_buf_with_border(uint8_t *buf, int buf_stride, int border,
                              bool is_hbd) {
  const int border16 = border % 16 == 0 ? border : 16 * (1 + border / 16);
  const int bytes_per_val = is_hbd ? sizeof(uint16_t) : sizeof(uint8_t);
  aom_free(buf - (border16 + border16 * buf_stride) * bytes_per_val);
}

// build interintra_predictors for one plane
void av1_build_interintra_predictors_sbp(const AV1_COMMON *cm, MACROBLOCKD *xd,
                                         uint8_t *pred, int stride,
                                         const BUFFER_SET *ctx, int plane,
                                         BLOCK_SIZE bsize, int border) {
  assert(bsize < BLOCK_SIZES_ALL);
  uint8_t *intrapred_buf;
  int intrapred_stride;
  av1_alloc_buf_with_border(&intrapred_buf, &intrapred_stride, border,
                            is_cur_buf_hbd(xd));
  if (is_cur_buf_hbd(xd)) {
    uint16_t *intrapredictor = (uint16_t *)intrapred_buf;
    av1_build_intra_predictors_for_interintra(
        cm, xd, bsize, plane, ctx, CONVERT_TO_BYTEPTR(intrapredictor),
        intrapred_stride, border);
    av1_combine_interintra(xd, bsize, plane, pred, stride,
                           CONVERT_TO_BYTEPTR(intrapredictor), intrapred_stride,
                           border);
  } else {
    av1_build_intra_predictors_for_interintra(
        cm, xd, bsize, plane, ctx, intrapred_buf, intrapred_stride, border);
    av1_combine_interintra(xd, bsize, plane, pred, stride, intrapred_buf,
                           intrapred_stride, border);
  }
  av1_free_buf_with_border(intrapred_buf, intrapred_stride, border,
                           is_cur_buf_hbd(xd));
}