/*
 * Copyright (c) 2019, 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 <limits.h>
#include <math.h>
#include <stdbool.h>
#include <stdio.h>

#include "config/aom_config.h"
#include "config/aom_dsp_rtcd.h"
#include "config/av1_rtcd.h"

#include "aom_dsp/aom_dsp_common.h"
#include "aom_dsp/binary_codes_writer.h"
#include "aom_ports/mem.h"
#include "aom_ports/aom_timer.h"
#include "aom_ports/system_state.h"

#include "av1/common/reconinter.h"
#include "av1/common/blockd.h"

#include "av1/encoder/encodeframe.h"
#include "av1/encoder/var_based_part.h"
#include "av1/encoder/reconinter_enc.h"

extern const uint8_t AV1_VAR_OFFS[];

typedef struct {
  VPVariance *part_variances;
  VPartVar *split[4];
} variance_node;

static AOM_INLINE void tree_to_node(void *data, BLOCK_SIZE bsize,
                                    variance_node *node) {
  int i;
  node->part_variances = NULL;
  switch (bsize) {
    case BLOCK_128X128: {
      VP128x128 *vt = (VP128x128 *)data;
      node->part_variances = &vt->part_variances;
      for (i = 0; i < 4; i++)
        node->split[i] = &vt->split[i].part_variances.none;
      break;
    }
    case BLOCK_64X64: {
      VP64x64 *vt = (VP64x64 *)data;
      node->part_variances = &vt->part_variances;
      for (i = 0; i < 4; i++)
        node->split[i] = &vt->split[i].part_variances.none;
      break;
    }
    case BLOCK_32X32: {
      VP32x32 *vt = (VP32x32 *)data;
      node->part_variances = &vt->part_variances;
      for (i = 0; i < 4; i++)
        node->split[i] = &vt->split[i].part_variances.none;
      break;
    }
    case BLOCK_16X16: {
      VP16x16 *vt = (VP16x16 *)data;
      node->part_variances = &vt->part_variances;
      for (i = 0; i < 4; i++)
        node->split[i] = &vt->split[i].part_variances.none;
      break;
    }
    case BLOCK_8X8: {
      VP8x8 *vt = (VP8x8 *)data;
      node->part_variances = &vt->part_variances;
      for (i = 0; i < 4; i++)
        node->split[i] = &vt->split[i].part_variances.none;
      break;
    }
    default: {
      VP4x4 *vt = (VP4x4 *)data;
      assert(bsize == BLOCK_4X4);
      node->part_variances = &vt->part_variances;
      for (i = 0; i < 4; i++) node->split[i] = &vt->split[i];
      break;
    }
  }
}

// Set variance values given sum square error, sum error, count.
static AOM_INLINE void fill_variance(uint32_t s2, int32_t s, int c,
                                     VPartVar *v) {
  v->sum_square_error = s2;
  v->sum_error = s;
  v->log2_count = c;
}

static AOM_INLINE void get_variance(VPartVar *v) {
  v->variance =
      (int)(256 * (v->sum_square_error -
                   (uint32_t)(((int64_t)v->sum_error * v->sum_error) >>
                              v->log2_count)) >>
            v->log2_count);
}

static AOM_INLINE void sum_2_variances(const VPartVar *a, const VPartVar *b,
                                       VPartVar *r) {
  assert(a->log2_count == b->log2_count);
  fill_variance(a->sum_square_error + b->sum_square_error,
                a->sum_error + b->sum_error, a->log2_count + 1, r);
}

static AOM_INLINE void fill_variance_tree(void *data, BLOCK_SIZE bsize) {
  variance_node node;
  memset(&node, 0, sizeof(node));
  tree_to_node(data, bsize, &node);
  sum_2_variances(node.split[0], node.split[1], &node.part_variances->horz[0]);
  sum_2_variances(node.split[2], node.split[3], &node.part_variances->horz[1]);
  sum_2_variances(node.split[0], node.split[2], &node.part_variances->vert[0]);
  sum_2_variances(node.split[1], node.split[3], &node.part_variances->vert[1]);
  sum_2_variances(&node.part_variances->vert[0], &node.part_variances->vert[1],
                  &node.part_variances->none);
}

static AOM_INLINE void set_block_size(AV1_COMP *const cpi, MACROBLOCK *const x,
                                      MACROBLOCKD *const xd, int mi_row,
                                      int mi_col, BLOCK_SIZE bsize) {
  if (cpi->common.mi_params.mi_cols > mi_col &&
      cpi->common.mi_params.mi_rows > mi_row) {
    set_mode_info_offsets(&cpi->common.mi_params, &cpi->mbmi_ext_info, x, xd,
                          mi_row, mi_col);
    xd->mi[0]->sb_type = bsize;
  }
}

static int set_vt_partitioning(AV1_COMP *cpi, MACROBLOCK *const x,
                               MACROBLOCKD *const xd,
                               const TileInfo *const tile, void *data,
                               BLOCK_SIZE bsize, int mi_row, int mi_col,
                               int64_t threshold, BLOCK_SIZE bsize_min,
                               int force_split) {
  AV1_COMMON *const cm = &cpi->common;
  variance_node vt;
  const int block_width = mi_size_wide[bsize];
  const int block_height = mi_size_high[bsize];

  assert(block_height == block_width);
  tree_to_node(data, bsize, &vt);

  if (force_split == 1) return 0;

  // For bsize=bsize_min (16x16/8x8 for 8x8/4x4 downsampling), select if
  // variance is below threshold, otherwise split will be selected.
  // No check for vert/horiz split as too few samples for variance.
  if (bsize == bsize_min) {
    // Variance already computed to set the force_split.
    if (frame_is_intra_only(cm)) get_variance(&vt.part_variances->none);
    if (mi_col + block_width <= tile->mi_col_end &&
        mi_row + block_height <= tile->mi_row_end &&
        vt.part_variances->none.variance < threshold) {
      set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
      return 1;
    }
    return 0;
  } else if (bsize > bsize_min) {
    // Variance already computed to set the force_split.
    if (frame_is_intra_only(cm)) get_variance(&vt.part_variances->none);
    // For key frame: take split for bsize above 32X32 or very high variance.
    if (frame_is_intra_only(cm) &&
        (bsize > BLOCK_32X32 ||
         vt.part_variances->none.variance > (threshold << 4))) {
      return 0;
    }
    // If variance is low, take the bsize (no split).
    if (mi_col + block_width <= tile->mi_col_end &&
        mi_row + block_height <= tile->mi_row_end &&
        vt.part_variances->none.variance < threshold) {
      set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
      return 1;
    }
    // Check vertical split.
    if (mi_row + block_height <= tile->mi_row_end &&
        mi_col + block_width / 2 <= tile->mi_col_end) {
      BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_VERT);
      get_variance(&vt.part_variances->vert[0]);
      get_variance(&vt.part_variances->vert[1]);
      if (vt.part_variances->vert[0].variance < threshold &&
          vt.part_variances->vert[1].variance < threshold &&
          get_plane_block_size(subsize, xd->plane[1].subsampling_x,
                               xd->plane[1].subsampling_y) < BLOCK_INVALID) {
        set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
        set_block_size(cpi, x, xd, mi_row, mi_col + block_width / 2, subsize);
        return 1;
      }
    }
    // Check horizontal split.
    if (mi_col + block_width <= tile->mi_col_end &&
        mi_row + block_height / 2 <= tile->mi_row_end) {
      BLOCK_SIZE subsize = get_partition_subsize(bsize, PARTITION_HORZ);
      get_variance(&vt.part_variances->horz[0]);
      get_variance(&vt.part_variances->horz[1]);
      if (vt.part_variances->horz[0].variance < threshold &&
          vt.part_variances->horz[1].variance < threshold &&
          get_plane_block_size(subsize, xd->plane[1].subsampling_x,
                               xd->plane[1].subsampling_y) < BLOCK_INVALID) {
        set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
        set_block_size(cpi, x, xd, mi_row + block_height / 2, mi_col, subsize);
        return 1;
      }
    }
    return 0;
  }
  return 0;
}

static AOM_INLINE void fill_variance_8x8avg(const uint8_t *s, int sp,
                                            const uint8_t *d, int dp,
                                            int x16_idx, int y16_idx,
                                            VP16x16 *vst, int highbd_flag,
                                            int pixels_wide, int pixels_high,
                                            int is_key_frame) {
  int k;
  for (k = 0; k < 4; k++) {
    int x8_idx = x16_idx + ((k & 1) << 3);
    int y8_idx = y16_idx + ((k >> 1) << 3);
    unsigned int sse = 0;
    int sum = 0;
    if (x8_idx < pixels_wide && y8_idx < pixels_high) {
      int s_avg;
      int d_avg = 128;
      if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
        s_avg = aom_highbd_avg_8x8(s + y8_idx * sp + x8_idx, sp);
        if (!is_key_frame)
          d_avg = aom_highbd_avg_8x8(d + y8_idx * dp + x8_idx, dp);
      } else {
        s_avg = aom_avg_8x8(s + y8_idx * sp + x8_idx, sp);
        if (!is_key_frame) d_avg = aom_avg_8x8(d + y8_idx * dp + x8_idx, dp);
      }

      sum = s_avg - d_avg;
      sse = sum * sum;
    }
    fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
  }
}

static int compute_minmax_8x8(const uint8_t *s, int sp, const uint8_t *d,
                              int dp, int x16_idx, int y16_idx, int highbd_flag,
                              int pixels_wide, int pixels_high) {
  int k;
  int minmax_max = 0;
  int minmax_min = 255;
  // Loop over the 4 8x8 subblocks.
  for (k = 0; k < 4; k++) {
    int x8_idx = x16_idx + ((k & 1) << 3);
    int y8_idx = y16_idx + ((k >> 1) << 3);
    int min = 0;
    int max = 0;
    if (x8_idx < pixels_wide && y8_idx < pixels_high) {
      if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
        aom_highbd_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
                              d + y8_idx * dp + x8_idx, dp, &min, &max);
      } else {
        aom_minmax_8x8(s + y8_idx * sp + x8_idx, sp, d + y8_idx * dp + x8_idx,
                       dp, &min, &max);
      }
      if ((max - min) > minmax_max) minmax_max = (max - min);
      if ((max - min) < minmax_min) minmax_min = (max - min);
    }
  }
  return (minmax_max - minmax_min);
}

static AOM_INLINE void fill_variance_4x4avg(const uint8_t *s, int sp,
                                            const uint8_t *d, int dp,
                                            int x8_idx, int y8_idx, VP8x8 *vst,
                                            int highbd_flag, int pixels_wide,
                                            int pixels_high, int is_key_frame) {
  int k;
  for (k = 0; k < 4; k++) {
    int x4_idx = x8_idx + ((k & 1) << 2);
    int y4_idx = y8_idx + ((k >> 1) << 2);
    unsigned int sse = 0;
    int sum = 0;
    if (x4_idx < pixels_wide && y4_idx < pixels_high) {
      int s_avg;
      int d_avg = 128;
      if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
        s_avg = aom_highbd_avg_4x4(s + y4_idx * sp + x4_idx, sp);
        if (!is_key_frame)
          d_avg = aom_highbd_avg_4x4(d + y4_idx * dp + x4_idx, dp);
      } else {
        s_avg = aom_avg_4x4(s + y4_idx * sp + x4_idx, sp);
        if (!is_key_frame) d_avg = aom_avg_4x4(d + y4_idx * dp + x4_idx, dp);
      }

      sum = s_avg - d_avg;
      sse = sum * sum;
    }
    fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
  }
}

// TODO(kyslov) Bring back threshold adjustment based on content state
static int64_t scale_part_thresh_content(int64_t threshold_base, int speed,
                                         int width, int height,
                                         int content_state,
                                         int non_reference_frame) {
  (void)width;
  (void)height;
  (void)content_state;
  int64_t threshold = threshold_base;
  if (non_reference_frame) threshold = (3 * threshold) >> 1;
  if (speed >= 8) {
    return (5 * threshold) >> 2;
  }
  return threshold;
}

static AOM_INLINE void set_vbp_thresholds(AV1_COMP *cpi, int64_t thresholds[],
                                          int q, int content_state) {
  AV1_COMMON *const cm = &cpi->common;
  const int is_key_frame = frame_is_intra_only(cm);
  const int threshold_multiplier = is_key_frame ? 40 : 1;
  EncQuantDequantParams *quant_params = &cpi->enc_quant_dequant_params;
#if CONFIG_EXTQUANT
  int64_t threshold_base =
      (int64_t)(threshold_multiplier *
                ROUND_POWER_OF_TWO(quant_params->dequants.y_dequant_QTX[q][1],
                                   QUANT_TABLE_BITS));
#else
  int64_t threshold_base = (int64_t)(
      threshold_multiplier * quant_params->dequants.y_dequant_QTX[q][1]);
#endif
  if (is_key_frame) {
    thresholds[0] = threshold_base;
    thresholds[1] = threshold_base;
    thresholds[2] = threshold_base >> 2;
    thresholds[3] = threshold_base >> 2;
    thresholds[4] = threshold_base << 2;
  } else {
    if (cpi->noise_estimate.enabled && cm->width >= 640 && cm->height >= 480) {
      NOISE_LEVEL noise_level =
          av1_noise_estimate_extract_level(&cpi->noise_estimate);
      if (noise_level == kHigh)
        threshold_base = 3 * threshold_base;
      else if (noise_level == kMedium)
        threshold_base = threshold_base << 1;
      else if (noise_level < kLow)
        threshold_base = (7 * threshold_base) >> 3;
    }

    // Increase base variance threshold based on content_state/sum_diff level.
    threshold_base = scale_part_thresh_content(
        threshold_base, cpi->oxcf.speed, cm->width, cm->height, content_state,
        cpi->svc.non_reference_frame);

    thresholds[0] = threshold_base >> 1;
    thresholds[1] = threshold_base;
    thresholds[3] = threshold_base << cpi->oxcf.speed;
    if (cm->width >= 1280 && cm->height >= 720)
      thresholds[3] = thresholds[3] << 1;
    if (cm->width * cm->height <= 352 * 288) {
      int last_qindex = cpi->rc.last_q[INTER_FRAME];
      if (last_qindex >= QINDEX_HIGH_THR) {
        threshold_base = (5 * threshold_base) >> 1;
        thresholds[1] = threshold_base >> 3;
        thresholds[2] = threshold_base << 2;
        thresholds[3] = threshold_base << 5;
      } else if (last_qindex < QINDEX_LOW_THR) {
        thresholds[1] = threshold_base >> 3;
        thresholds[2] = threshold_base >> 1;
        thresholds[3] = threshold_base << 3;
      } else {
        int64_t qi_diff_low = last_qindex - QINDEX_LOW_THR;
        int64_t qi_diff_high = QINDEX_HIGH_THR - last_qindex;
        int64_t threshold_diff = QINDEX_HIGH_THR - QINDEX_LOW_THR;
        int64_t threshold_base_high = (5 * threshold_base) >> 1;

        threshold_diff = threshold_diff > 0 ? threshold_diff : 1;
        threshold_base = (qi_diff_low * threshold_base_high +
                          qi_diff_high * threshold_base) /
                         threshold_diff;
        thresholds[1] = threshold_base >> 3;
        thresholds[2] = ((qi_diff_low * threshold_base) +
                         qi_diff_high * (threshold_base >> 1)) /
                        threshold_diff;
        thresholds[3] = ((qi_diff_low * (threshold_base << 5)) +
                         qi_diff_high * (threshold_base << 3)) /
                        threshold_diff;
      }
    } else if (cm->width < 1280 && cm->height < 720) {
      thresholds[2] = (5 * threshold_base) >> 2;
    } else if (cm->width < 1920 && cm->height < 1080) {
      thresholds[2] = threshold_base << 1;
    } else {
      thresholds[2] = (5 * threshold_base) >> 1;
    }
    if (cpi->sf.rt_sf.force_large_partition_blocks) {
      thresholds[1] <<= 2;
      thresholds[2] <<= 5;
      thresholds[3] = INT32_MAX;
    }
  }
}

// Set temporal variance low flag for superblock 64x64.
// Only first 25 in the array are used in this case.
static AOM_INLINE void set_low_temp_var_flag_64x64(
    CommonModeInfoParams *mi_params, PartitionSearchInfo *part_info,
    MACROBLOCKD *xd, VP64x64 *vt, const int64_t thresholds[], int mi_col,
    int mi_row) {
  if (xd->mi[0]->sb_type == BLOCK_64X64) {
    if ((vt->part_variances).none.variance < (thresholds[0] >> 1))
      part_info->variance_low[0] = 1;
  } else if (xd->mi[0]->sb_type == BLOCK_64X32) {
    for (int i = 0; i < 2; i++) {
      if (vt->part_variances.horz[i].variance < (thresholds[0] >> 2))
        part_info->variance_low[i + 1] = 1;
    }
  } else if (xd->mi[0]->sb_type == BLOCK_32X64) {
    for (int i = 0; i < 2; i++) {
      if (vt->part_variances.vert[i].variance < (thresholds[0] >> 2))
        part_info->variance_low[i + 3] = 1;
    }
  } else {
    static const int idx[4][2] = { { 0, 0 }, { 0, 8 }, { 8, 0 }, { 8, 8 } };
    for (int i = 0; i < 4; i++) {
      const int idx_str =
          mi_params->mi_stride * (mi_row + idx[i][0]) + mi_col + idx[i][1];
      MB_MODE_INFO **this_mi = mi_params->mi_grid_base + idx_str;

      if (mi_params->mi_cols <= mi_col + idx[i][1] ||
          mi_params->mi_rows <= mi_row + idx[i][0])
        continue;

      if (*this_mi == NULL) continue;

      if ((*this_mi)->sb_type == BLOCK_32X32) {
        int64_t threshold_32x32 = (5 * thresholds[1]) >> 3;
        if (vt->split[i].part_variances.none.variance < threshold_32x32)
          part_info->variance_low[i + 5] = 1;
      } else {
        // For 32x16 and 16x32 blocks, the flag is set on each 16x16 block
        // inside.
        if ((*this_mi)->sb_type == BLOCK_16X16 ||
            (*this_mi)->sb_type == BLOCK_32X16 ||
            (*this_mi)->sb_type == BLOCK_16X32) {
          for (int j = 0; j < 4; j++) {
            if (vt->split[i].split[j].part_variances.none.variance <
                (thresholds[2] >> 8))
              part_info->variance_low[(i << 2) + j + 9] = 1;
          }
        }
      }
    }
  }
}

static AOM_INLINE void set_low_temp_var_flag_128x128(
    CommonModeInfoParams *mi_params, PartitionSearchInfo *part_info,
    MACROBLOCKD *xd, VP128x128 *vt, const int64_t thresholds[], int mi_col,
    int mi_row) {
  if (xd->mi[0]->sb_type == BLOCK_128X128) {
    if (vt->part_variances.none.variance < (thresholds[0] >> 1))
      part_info->variance_low[0] = 1;
  } else if (xd->mi[0]->sb_type == BLOCK_128X64) {
    for (int i = 0; i < 2; i++) {
      if (vt->part_variances.horz[i].variance < (thresholds[0] >> 2))
        part_info->variance_low[i + 1] = 1;
    }
  } else if (xd->mi[0]->sb_type == BLOCK_64X128) {
    for (int i = 0; i < 2; i++) {
      if (vt->part_variances.vert[i].variance < (thresholds[0] >> 2))
        part_info->variance_low[i + 3] = 1;
    }
  } else {
    static const int idx64[4][2] = {
      { 0, 0 }, { 0, 16 }, { 16, 0 }, { 16, 16 }
    };
    static const int idx32[4][2] = { { 0, 0 }, { 0, 8 }, { 8, 0 }, { 8, 8 } };
    for (int i = 0; i < 4; i++) {
      const int idx_str =
          mi_params->mi_stride * (mi_row + idx64[i][0]) + mi_col + idx64[i][1];
      MB_MODE_INFO **mi_64 = mi_params->mi_grid_base + idx_str;
      if (*mi_64 == NULL) continue;
      if (mi_params->mi_cols <= mi_col + idx64[i][1] ||
          mi_params->mi_rows <= mi_row + idx64[i][0])
        continue;
      const int64_t threshold_64x64 = (5 * thresholds[1]) >> 3;
      if ((*mi_64)->sb_type == BLOCK_64X64) {
        if (vt->split[i].part_variances.none.variance < threshold_64x64)
          part_info->variance_low[5 + i] = 1;
      } else if ((*mi_64)->sb_type == BLOCK_64X32) {
        for (int j = 0; j < 2; j++)
          if (vt->split[i].part_variances.horz[j].variance <
              (threshold_64x64 >> 1))
            part_info->variance_low[9 + (i << 1) + j] = 1;
      } else if ((*mi_64)->sb_type == BLOCK_32X64) {
        for (int j = 0; j < 2; j++)
          if (vt->split[i].part_variances.vert[j].variance <
              (threshold_64x64 >> 1))
            part_info->variance_low[17 + (i << 1) + j] = 1;
      } else {
        for (int k = 0; k < 4; k++) {
          const int idx_str1 = mi_params->mi_stride * idx32[k][0] + idx32[k][1];
          MB_MODE_INFO **mi_32 = mi_params->mi_grid_base + idx_str + idx_str1;
          if (*mi_32 == NULL) continue;

          if (mi_params->mi_cols <= mi_col + idx64[i][1] + idx32[k][1] ||
              mi_params->mi_rows <= mi_row + idx64[i][0] + idx32[k][0])
            continue;
          const int64_t threshold_32x32 = (5 * thresholds[2]) >> 3;
          if ((*mi_32)->sb_type == BLOCK_32X32) {
            if (vt->split[i].split[k].part_variances.none.variance <
                threshold_32x32)
              part_info->variance_low[25 + (i << 2) + k] = 1;
          } else {
            // For 32x16 and 16x32 blocks, the flag is set on each 16x16 block
            // inside.
            if ((*mi_32)->sb_type == BLOCK_16X16 ||
                (*mi_32)->sb_type == BLOCK_32X16 ||
                (*mi_32)->sb_type == BLOCK_16X32) {
              for (int j = 0; j < 4; j++) {
                if (vt->split[i]
                        .split[k]
                        .split[j]
                        .part_variances.none.variance < (thresholds[3] >> 8))
                  part_info->variance_low[41 + (i << 4) + (k << 2) + j] = 1;
              }
            }
          }
        }
      }
    }
  }
}

static AOM_INLINE void set_low_temp_var_flag(
    AV1_COMP *cpi, PartitionSearchInfo *part_info, MACROBLOCKD *xd,
    VP128x128 *vt, int64_t thresholds[], MV_REFERENCE_FRAME ref_frame_partition,
    int mi_col, int mi_row) {
  AV1_COMMON *const cm = &cpi->common;
  const int mv_thr = cm->width > 640 ? 8 : 4;
  // Check temporal variance for bsize >= 16x16, if LAST_FRAME was selected and
  // int_pro mv is small. If the temporal variance is small set the flag
  // variance_low for the block. The variance threshold can be adjusted, the
  // higher the more aggressive.
  if (ref_frame_partition == LAST_FRAME &&
      (cpi->sf.rt_sf.short_circuit_low_temp_var == 1 ||
       (cpi->sf.rt_sf.estimate_motion_for_var_based_partition &&
        xd->mi[0]->mv[0].as_mv.col < mv_thr &&
        xd->mi[0]->mv[0].as_mv.col > -mv_thr &&
        xd->mi[0]->mv[0].as_mv.row < mv_thr &&
        xd->mi[0]->mv[0].as_mv.row > -mv_thr))) {
    const int is_small_sb = (cm->seq_params.sb_size == BLOCK_64X64);
    if (is_small_sb)
      set_low_temp_var_flag_64x64(&cm->mi_params, part_info, xd,
                                  &(vt->split[0]), thresholds, mi_col, mi_row);
    else
      set_low_temp_var_flag_128x128(&cm->mi_params, part_info, xd, vt,
                                    thresholds, mi_col, mi_row);
  }
}

void av1_set_variance_partition_thresholds(AV1_COMP *cpi, int q,
                                           int content_state) {
  SPEED_FEATURES *const sf = &cpi->sf;
  if (sf->part_sf.partition_search_type != VAR_BASED_PARTITION) {
    return;
  } else {
    set_vbp_thresholds(cpi, cpi->vbp_info.thresholds, q, content_state);
    // The threshold below is not changed locally.
    cpi->vbp_info.threshold_minmax = 15 + (q >> 3);
  }
}

static AOM_INLINE void chroma_check(AV1_COMP *cpi, MACROBLOCK *x,
                                    BLOCK_SIZE bsize, unsigned int y_sad,
                                    int is_key_frame) {
  int i;
  MACROBLOCKD *xd = &x->e_mbd;

  if (is_key_frame || cpi->oxcf.tool_cfg.enable_monochrome) return;

  for (i = 1; i <= 2; ++i) {
    unsigned int uv_sad = UINT_MAX;
    struct macroblock_plane *p = &x->plane[i];
    struct macroblockd_plane *pd = &xd->plane[i];
    const BLOCK_SIZE bs =
        get_plane_block_size(bsize, pd->subsampling_x, pd->subsampling_y);

    if (bs != BLOCK_INVALID)
      uv_sad = cpi->fn_ptr[bs].sdf(p->src.buf, p->src.stride, pd->dst.buf,
                                   pd->dst.stride);

    x->color_sensitivity[i - 1] = uv_sad > (y_sad >> 2);
  }
}

static void fill_variance_tree_leaves(
    AV1_COMP *cpi, MACROBLOCK *x, VP128x128 *vt, VP16x16 *vt2,
    unsigned char *force_split, int avg_16x16[][4], int maxvar_16x16[][4],
    int minvar_16x16[][4], int *variance4x4downsample, int64_t *thresholds,
    uint8_t *src, int src_stride, const uint8_t *dst, int dst_stride) {
  AV1_COMMON *cm = &cpi->common;
  MACROBLOCKD *xd = &x->e_mbd;
  const int is_key_frame = frame_is_intra_only(cm);
  const int is_small_sb = (cm->seq_params.sb_size == BLOCK_64X64);
  const int num_64x64_blocks = is_small_sb ? 1 : 4;
  // TODO(kyslov) Bring back compute_minmax_variance with content type detection
  const int compute_minmax_variance = 0;
  const int segment_id = xd->mi[0]->segment_id;
  int pixels_wide = 128, pixels_high = 128;

  if (is_small_sb) {
    pixels_wide = 64;
    pixels_high = 64;
  }
  if (xd->mb_to_right_edge < 0) pixels_wide += (xd->mb_to_right_edge >> 3);
  if (xd->mb_to_bottom_edge < 0) pixels_high += (xd->mb_to_bottom_edge >> 3);
  for (int m = 0; m < num_64x64_blocks; m++) {
    const int x64_idx = ((m & 1) << 6);
    const int y64_idx = ((m >> 1) << 6);
    const int m2 = m << 2;
    force_split[m + 1] = 0;

    for (int i = 0; i < 4; i++) {
      const int x32_idx = x64_idx + ((i & 1) << 5);
      const int y32_idx = y64_idx + ((i >> 1) << 5);
      const int i2 = (m2 + i) << 2;
      force_split[5 + m2 + i] = 0;
      avg_16x16[m][i] = 0;
      maxvar_16x16[m][i] = 0;
      minvar_16x16[m][i] = INT_MAX;
      for (int j = 0; j < 4; j++) {
        const int x16_idx = x32_idx + ((j & 1) << 4);
        const int y16_idx = y32_idx + ((j >> 1) << 4);
        const int split_index = 21 + i2 + j;
        VP16x16 *vst = &vt->split[m].split[i].split[j];
        force_split[split_index] = 0;
        variance4x4downsample[i2 + j] = 0;
        if (!is_key_frame) {
          fill_variance_8x8avg(src, src_stride, dst, dst_stride, x16_idx,
                               y16_idx, vst, xd->cur_buf->flags, pixels_wide,
                               pixels_high, is_key_frame);
          fill_variance_tree(&vt->split[m].split[i].split[j], BLOCK_16X16);
          get_variance(&vt->split[m].split[i].split[j].part_variances.none);
          avg_16x16[m][i] +=
              vt->split[m].split[i].split[j].part_variances.none.variance;
          if (vt->split[m].split[i].split[j].part_variances.none.variance <
              minvar_16x16[m][i])
            minvar_16x16[m][i] =
                vt->split[m].split[i].split[j].part_variances.none.variance;
          if (vt->split[m].split[i].split[j].part_variances.none.variance >
              maxvar_16x16[m][i])
            maxvar_16x16[m][i] =
                vt->split[m].split[i].split[j].part_variances.none.variance;
          if (vt->split[m].split[i].split[j].part_variances.none.variance >
              thresholds[3]) {
            // 16X16 variance is above threshold for split, so force split to
            // 8x8 for this 16x16 block (this also forces splits for upper
            // levels).
            force_split[split_index] = 1;
            force_split[5 + m2 + i] = 1;
            force_split[m + 1] = 1;
            force_split[0] = 1;
          } else if (!cyclic_refresh_segment_id_boosted(segment_id) &&
                     compute_minmax_variance &&
                     vt->split[m]
                             .split[i]
                             .split[j]
                             .part_variances.none.variance > thresholds[2]) {
            // We have some nominal amount of 16x16 variance (based on average),
            // compute the minmax over the 8x8 sub-blocks, and if above
            // threshold, force split to 8x8 block for this 16x16 block.
            int minmax = compute_minmax_8x8(
                src, src_stride, dst, dst_stride, x16_idx, y16_idx,
                xd->cur_buf->flags, pixels_wide, pixels_high);
            int thresh_minmax = (int)cpi->vbp_info.threshold_minmax;
            if (minmax > thresh_minmax) {
              force_split[split_index] = 1;
              force_split[5 + m2 + i] = 1;
              force_split[m + 1] = 1;
              force_split[0] = 1;
            }
          }
        }
        if (is_key_frame) {
          force_split[split_index] = 0;
          // Go down to 4x4 down-sampling for variance.
          variance4x4downsample[i2 + j] = 1;
          for (int k = 0; k < 4; k++) {
            int x8_idx = x16_idx + ((k & 1) << 3);
            int y8_idx = y16_idx + ((k >> 1) << 3);
            VP8x8 *vst2 = is_key_frame ? &vst->split[k] : &vt2[i2 + j].split[k];
            fill_variance_4x4avg(src, src_stride, dst, dst_stride, x8_idx,
                                 y8_idx, vst2, xd->cur_buf->flags, pixels_wide,
                                 pixels_high, is_key_frame);
          }
        }
      }
    }
  }
}

static void setup_planes(AV1_COMP *cpi, MACROBLOCK *x, unsigned int *y_sad,
                         unsigned int *y_sad_g,
                         MV_REFERENCE_FRAME *ref_frame_partition, int mi_row,
                         int mi_col) {
  AV1_COMMON *const cm = &cpi->common;
  MACROBLOCKD *xd = &x->e_mbd;
  const int num_planes = av1_num_planes(cm);
  const int is_small_sb = (cm->seq_params.sb_size == BLOCK_64X64);
  BLOCK_SIZE bsize = is_small_sb ? BLOCK_64X64 : BLOCK_128X128;
  // TODO(kyslov): we are assuming that the ref is LAST_FRAME! Check if it
  // is!!
  MB_MODE_INFO *mi = xd->mi[0];
  const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_yv12_buf(cm, LAST_FRAME);
  assert(yv12 != NULL);
  const YV12_BUFFER_CONFIG *yv12_g = NULL;

  // For non-SVC GOLDEN is another temporal reference. Check if it should be
  // used as reference for partitioning.
  if (!cpi->use_svc && (cpi->ref_frame_flags & AOM_GOLD_FLAG) &&
      cpi->sf.rt_sf.use_nonrd_pick_mode) {
    yv12_g = get_ref_frame_yv12_buf(cm, GOLDEN_FRAME);
    if (yv12_g && yv12_g != yv12) {
      av1_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
                           get_ref_scale_factors(cm, GOLDEN_FRAME), num_planes);
      *y_sad_g = cpi->fn_ptr[bsize].sdf(
          x->plane[0].src.buf, x->plane[0].src.stride, xd->plane[0].pre[0].buf,
          xd->plane[0].pre[0].stride);
    }
  }

  av1_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
                       get_ref_scale_factors(cm, LAST_FRAME), num_planes);
  mi->ref_frame[0] = LAST_FRAME;
  mi->ref_frame[1] = NONE_FRAME;
  mi->sb_type = cm->seq_params.sb_size;
  mi->mv[0].as_int = 0;
#if CONFIG_REMOVE_DUAL_FILTER
  mi->interp_fltr = BILINEAR;
#else
  mi->interp_filters = av1_broadcast_interp_filter(BILINEAR);
#endif  // CONFIG_REMOVE_DUAL_FILTER
  if (cpi->sf.rt_sf.estimate_motion_for_var_based_partition) {
    if (xd->mb_to_right_edge >= 0 && xd->mb_to_bottom_edge >= 0) {
      const MV dummy_mv = { 0, 0 };
      *y_sad = av1_int_pro_motion_estimation(cpi, x, cm->seq_params.sb_size,
                                             mi_row, mi_col, &dummy_mv);
    }
  }
  if (*y_sad == UINT_MAX) {
    *y_sad = cpi->fn_ptr[bsize].sdf(x->plane[0].src.buf, x->plane[0].src.stride,
                                    xd->plane[0].pre[0].buf,
                                    xd->plane[0].pre[0].stride);
  }

  // Pick the ref frame for partitioning, use golden frame only if its
  // lower sad.
  if (*y_sad_g < 0.9 * *y_sad) {
    av1_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
                         get_ref_scale_factors(cm, GOLDEN_FRAME), num_planes);
    mi->ref_frame[0] = GOLDEN_FRAME;
    mi->mv[0].as_int = 0;
    *y_sad = *y_sad_g;
    *ref_frame_partition = GOLDEN_FRAME;
    x->nonrd_prune_ref_frame_search = 0;
  } else {
    *ref_frame_partition = LAST_FRAME;
    x->nonrd_prune_ref_frame_search =
        cpi->sf.rt_sf.nonrd_prune_ref_frame_search;
  }

  set_ref_ptrs(cm, xd, mi->ref_frame[0], mi->ref_frame[1]);
  av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL,
                                cm->seq_params.sb_size, AOM_PLANE_Y,
                                AOM_PLANE_Y);
}

int av1_choose_var_based_partitioning(AV1_COMP *cpi, const TileInfo *const tile,
                                      ThreadData *td, MACROBLOCK *x, int mi_row,
                                      int mi_col) {
  AV1_COMMON *const cm = &cpi->common;
  MACROBLOCKD *xd = &x->e_mbd;
  const int64_t *const vbp_thresholds = cpi->vbp_info.thresholds;

  int i, j, k, m;
  VP128x128 *vt;
  VP16x16 *vt2 = NULL;
  unsigned char force_split[85];
  int avg_32x32;
  int avg_64x64;
  int max_var_32x32[4];
  int min_var_32x32[4];
  int var_32x32;
  int var_64x64;
  int min_var_64x64 = INT_MAX;
  int max_var_64x64 = 0;
  int avg_16x16[4][4];
  int maxvar_16x16[4][4];
  int minvar_16x16[4][4];
  int64_t threshold_4x4avg;
  int content_state = 0;
  uint8_t *s;
  const uint8_t *d;
  int sp;
  int dp;

  int is_key_frame =
      (frame_is_intra_only(cm) ||
       (cpi->use_svc &&
        cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame));

  assert(cm->seq_params.sb_size == BLOCK_64X64 ||
         cm->seq_params.sb_size == BLOCK_128X128);
  const int is_small_sb = (cm->seq_params.sb_size == BLOCK_64X64);
  const int num_64x64_blocks = is_small_sb ? 1 : 4;

  unsigned int y_sad = UINT_MAX;
  unsigned int y_sad_g = UINT_MAX;
  BLOCK_SIZE bsize = is_small_sb ? BLOCK_64X64 : BLOCK_128X128;

  // Ref frame used in partitioning.
  MV_REFERENCE_FRAME ref_frame_partition = LAST_FRAME;
  NOISE_LEVEL noise_level = kLow;

  CHECK_MEM_ERROR(cm, vt, aom_malloc(sizeof(*vt)));

  vt->split = td->vt64x64;

  int64_t thresholds[5] = { vbp_thresholds[0], vbp_thresholds[1],
                            vbp_thresholds[2], vbp_thresholds[3],
                            vbp_thresholds[4] };

  const int low_res = (cm->width <= 352 && cm->height <= 288);
  int variance4x4downsample[64];
  const int segment_id = xd->mi[0]->segment_id;

  if (cpi->oxcf.q_cfg.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled &&
      cyclic_refresh_segment_id_boosted(segment_id) &&
      cpi->sf.rt_sf.use_nonrd_pick_mode) {
#if CONFIG_EXTQUANT
    int q = av1_get_qindex(&cm->seg, segment_id, cm->quant_params.base_qindex,
                           cm->seq_params.bit_depth);
#else
    int q = av1_get_qindex(&cm->seg, segment_id, cm->quant_params.base_qindex);
#endif
    set_vbp_thresholds(cpi, thresholds, q, content_state);
  } else {
    set_vbp_thresholds(cpi, thresholds, cm->quant_params.base_qindex,
                       content_state);
  }

  // For non keyframes, disable 4x4 average for low resolution when speed = 8
  threshold_4x4avg = INT64_MAX;

  s = x->plane[0].src.buf;
  sp = x->plane[0].src.stride;

  // Index for force_split: 0 for 64x64, 1-4 for 32x32 blocks,
  // 5-20 for the 16x16 blocks.
  force_split[0] = 0;
  memset(x->part_search_info.variance_low, 0,
         sizeof(x->part_search_info.variance_low));

  // Check if LAST frame is NULL or if the resolution of LAST is
  // different than the current frame resolution, and if so, treat this frame
  // as a key frame, for the purpose of the superblock partitioning.
  // LAST == NULL can happen in cases where enhancement spatial layers are
  // enabled dyanmically and the only reference is the spatial(GOLDEN).
  // TODO(marpan): Check se of scaled references for the different resoln.
  if (!frame_is_intra_only(cm)) {
    const YV12_BUFFER_CONFIG *const ref =
        get_ref_frame_yv12_buf(cm, LAST_FRAME);
    if (ref == NULL || ref->y_crop_height != cm->height ||
        ref->y_crop_width != cm->width) {
      is_key_frame = 1;
    }
  }

  if (!is_key_frame) {
    setup_planes(cpi, x, &y_sad, &y_sad_g, &ref_frame_partition, mi_row,
                 mi_col);
    d = xd->plane[0].dst.buf;
    dp = xd->plane[0].dst.stride;
  } else {
    d = AV1_VAR_OFFS;
    dp = 0;
  }

  if (low_res && threshold_4x4avg < INT64_MAX)
    CHECK_MEM_ERROR(cm, vt2, aom_malloc(sizeof(*vt2)));
  // Fill in the entire tree of 8x8 (or 4x4 under some conditions) variances
  // for splits.
  fill_variance_tree_leaves(cpi, x, vt, vt2, force_split, avg_16x16,
                            maxvar_16x16, minvar_16x16, variance4x4downsample,
                            thresholds, s, sp, d, dp);

  // Fill the rest of the variance tree by summing split partition values.
  if (cpi->noise_estimate.enabled)
    noise_level = av1_noise_estimate_extract_level(&cpi->noise_estimate);

  avg_64x64 = 0;
  for (m = 0; m < num_64x64_blocks; ++m) {
    avg_32x32 = 0;
    max_var_32x32[m] = 0;
    min_var_32x32[m] = INT_MAX;
    const int m2 = m << 2;
    for (i = 0; i < 4; i++) {
      const int i2 = (m2 + i) << 2;
      for (j = 0; j < 4; j++) {
        const int split_index = 21 + i2 + j;
        if (variance4x4downsample[i2 + j] == 1) {
          VP16x16 *vtemp =
              (!is_key_frame) ? &vt2[i2 + j] : &vt->split[m].split[i].split[j];
          for (k = 0; k < 4; k++)
            fill_variance_tree(&vtemp->split[k], BLOCK_8X8);
          fill_variance_tree(vtemp, BLOCK_16X16);
          // If variance of this 16x16 block is above the threshold, force block
          // to split. This also forces a split on the upper levels.
          get_variance(&vtemp->part_variances.none);
          if (vtemp->part_variances.none.variance > thresholds[3]) {
            force_split[split_index] = 1;
            force_split[5 + m2 + i] = 1;
            force_split[m + 1] = 1;
            force_split[0] = 1;
          }
        }
      }
      fill_variance_tree(&vt->split[m].split[i], BLOCK_32X32);
      // If variance of this 32x32 block is above the threshold, or if its above
      // (some threshold of) the average variance over the sub-16x16 blocks,
      // then force this block to split. This also forces a split on the upper
      // (64x64) level.
      if (!force_split[5 + m2 + i]) {
        get_variance(&vt->split[m].split[i].part_variances.none);
        var_32x32 = vt->split[m].split[i].part_variances.none.variance;
        max_var_32x32[m] = AOMMAX(var_32x32, max_var_32x32[m]);
        min_var_32x32[m] = AOMMIN(var_32x32, min_var_32x32[m]);
        if (vt->split[m].split[i].part_variances.none.variance >
                thresholds[2] ||
            (!is_key_frame &&
             vt->split[m].split[i].part_variances.none.variance >
                 (thresholds[2] >> 1) &&
             vt->split[m].split[i].part_variances.none.variance >
                 (avg_16x16[m][i] >> 1))) {
          force_split[5 + m2 + i] = 1;
          force_split[m + 1] = 1;
          force_split[0] = 1;
        } else if (!is_key_frame &&
                   (!cpi->noise_estimate.enabled ||
                    (cpi->noise_estimate.enabled && noise_level < kLow)) &&
                   cm->height <= 360 &&
                   (maxvar_16x16[m][i] - minvar_16x16[m][i]) >
                       (thresholds[2] >> 1) &&
                   maxvar_16x16[m][i] > thresholds[2]) {
          force_split[5 + m2 + i] = 1;
          force_split[m + 1] = 1;
          force_split[0] = 1;
        }
        avg_32x32 += var_32x32;
      }
    }
    if (!force_split[1 + m]) {
      fill_variance_tree(&vt->split[m], BLOCK_64X64);
      get_variance(&vt->split[m].part_variances.none);
      var_64x64 = vt->split[m].part_variances.none.variance;
      max_var_64x64 = AOMMAX(var_64x64, max_var_64x64);
      min_var_64x64 = AOMMIN(var_64x64, min_var_64x64);
      // If variance of this 64x64 block is above (some threshold of) the
      // average variance over the sub-32x32 blocks, then force this block to
      // split. Only checking this for noise level >= medium for now.

      if (!is_key_frame &&
          (max_var_32x32[m] - min_var_32x32[m]) > 3 * (thresholds[1] >> 3) &&
          max_var_32x32[m] > thresholds[1] >> 1)
        force_split[1 + m] = 1;
      avg_64x64 += var_64x64;
    }
    if (is_small_sb) force_split[0] = 1;
  }

  if (!force_split[0]) {
    fill_variance_tree(vt, BLOCK_128X128);
    get_variance(&vt->part_variances.none);
    if (!is_key_frame &&
        (!cpi->noise_estimate.enabled ||
         (cpi->noise_estimate.enabled && noise_level >= kMedium)) &&
        vt->part_variances.none.variance > (9 * avg_64x64) >> 5)
      force_split[0] = 1;

    if (!is_key_frame &&
        (max_var_64x64 - min_var_64x64) > 3 * (thresholds[0] >> 3) &&
        max_var_64x64 > thresholds[0] >> 1)
      force_split[0] = 1;
  }

  if (mi_col + 32 > tile->mi_col_end || mi_row + 32 > tile->mi_row_end ||
      !set_vt_partitioning(cpi, x, xd, tile, vt, BLOCK_128X128, mi_row, mi_col,
                           thresholds[0], BLOCK_16X16, force_split[0])) {
    for (m = 0; m < num_64x64_blocks; ++m) {
      const int x64_idx = ((m & 1) << 4);
      const int y64_idx = ((m >> 1) << 4);
      const int m2 = m << 2;

      // Now go through the entire structure, splitting every block size until
      // we get to one that's got a variance lower than our threshold.
      if (!set_vt_partitioning(cpi, x, xd, tile, &vt->split[m], BLOCK_64X64,
                               mi_row + y64_idx, mi_col + x64_idx,
                               thresholds[1], BLOCK_16X16,
                               force_split[1 + m])) {
        for (i = 0; i < 4; ++i) {
          const int x32_idx = ((i & 1) << 3);
          const int y32_idx = ((i >> 1) << 3);
          const int i2 = (m2 + i) << 2;
          if (!set_vt_partitioning(cpi, x, xd, tile, &vt->split[m].split[i],
                                   BLOCK_32X32, (mi_row + y64_idx + y32_idx),
                                   (mi_col + x64_idx + x32_idx), thresholds[2],
                                   BLOCK_16X16, force_split[5 + m2 + i])) {
            for (j = 0; j < 4; ++j) {
              const int x16_idx = ((j & 1) << 2);
              const int y16_idx = ((j >> 1) << 2);
              const int split_index = 21 + i2 + j;
              // For inter frames: if variance4x4downsample[] == 1 for this
              // 16x16 block, then the variance is based on 4x4 down-sampling,
              // so use vt2 in set_vt_partioning(), otherwise use vt.
              VP16x16 *vtemp =
                  (!is_key_frame && variance4x4downsample[i2 + j] == 1)
                      ? &vt2[i2 + j]
                      : &vt->split[m].split[i].split[j];
              if (!set_vt_partitioning(cpi, x, xd, tile, vtemp, BLOCK_16X16,
                                       mi_row + y64_idx + y32_idx + y16_idx,
                                       mi_col + x64_idx + x32_idx + x16_idx,
                                       thresholds[3], BLOCK_8X8,
                                       force_split[split_index])) {
                for (k = 0; k < 4; ++k) {
                  const int x8_idx = (k & 1) << 1;
                  const int y8_idx = (k >> 1) << 1;
                  set_block_size(
                      cpi, x, xd,
                      (mi_row + y64_idx + y32_idx + y16_idx + y8_idx),
                      (mi_col + x64_idx + x32_idx + x16_idx + x8_idx),
                      BLOCK_8X8);
                }
              }
            }
          }
        }
      }
    }
  }

  if (cpi->sf.rt_sf.short_circuit_low_temp_var) {
    set_low_temp_var_flag(cpi, &x->part_search_info, xd, vt, thresholds,
                          ref_frame_partition, mi_col, mi_row);
  }
  chroma_check(cpi, x, bsize, y_sad, is_key_frame);

  if (vt2) aom_free(vt2);
  if (vt) aom_free(vt);
  return 0;
}