vp9_encodeframe.c

/*
 *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include <limits.h>
#include <math.h>
#include <stdio.h>

#include "./vp9_rtcd.h"
#include "./vpx_dsp_rtcd.h"
#include "./vpx_config.h"

#include "vpx_dsp/vpx_dsp_common.h"
#include "vpx_ports/mem.h"
#include "vpx_ports/vpx_timer.h"
#include "vpx_ports/system_state.h"

#include "vp9/common/vp9_common.h"
#include "vp9/common/vp9_entropy.h"
#include "vp9/common/vp9_entropymode.h"
#include "vp9/common/vp9_idct.h"
#include "vp9/common/vp9_mvref_common.h"
#include "vp9/common/vp9_pred_common.h"
#include "vp9/common/vp9_quant_common.h"
#include "vp9/common/vp9_reconintra.h"
#include "vp9/common/vp9_reconinter.h"
#include "vp9/common/vp9_seg_common.h"
#include "vp9/common/vp9_tile_common.h"

#include "vp9/encoder/vp9_aq_360.h"
#include "vp9/encoder/vp9_aq_complexity.h"
#include "vp9/encoder/vp9_aq_cyclicrefresh.h"
#include "vp9/encoder/vp9_aq_variance.h"
#include "vp9/encoder/vp9_encodeframe.h"
#include "vp9/encoder/vp9_encodemb.h"
#include "vp9/encoder/vp9_encodemv.h"
#include "vp9/encoder/vp9_ethread.h"
#include "vp9/encoder/vp9_extend.h"
#include "vp9/encoder/vp9_pickmode.h"
#include "vp9/encoder/vp9_rd.h"
#include "vp9/encoder/vp9_rdopt.h"
#include "vp9/encoder/vp9_segmentation.h"
#include "vp9/encoder/vp9_tokenize.h"

static void encode_superblock(VP9_COMP *cpi, ThreadData *td, TOKENEXTRA **t,
                              int output_enabled, int mi_row, int mi_col,
                              BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx);

// Machine learning-based early termination parameters.
static const double train_mean[24] = {
  303501.697372, 3042630.372158, 24.694696, 1.392182,
  689.413511,    162.027012,     1.478213,  0.0,
  135382.260230, 912738.513263,  28.845217, 1.515230,
  544.158492,    131.807995,     1.436863,  0.0,
  43682.377587,  208131.711766,  28.084737, 1.356677,
  138.254122,    119.522553,     1.252322,  0.0
};

static const double train_stdm[24] = {
  673689.212982, 5996652.516628, 0.024449, 1.989792,
  985.880847,    0.014638,       2.001898, 0.0,
  208798.775332, 1812548.443284, 0.018693, 1.838009,
  396.986910,    0.015657,       1.332541, 0.0,
  55888.847031,  448587.962714,  0.017900, 1.904776,
  98.652832,     0.016598,       1.320992, 0.0
};

// Error tolerance: 0.01%-0.0.05%-0.1%
static const double classifiers[24] = {
  0.111736, 0.289977, 0.042219, 0.204765, 0.120410, -0.143863,
  0.282376, 0.847811, 0.637161, 0.131570, 0.018636, 0.202134,
  0.112797, 0.028162, 0.182450, 1.124367, 0.386133, 0.083700,
  0.050028, 0.150873, 0.061119, 0.109318, 0.127255, 0.625211
};

// This is used as a reference when computing the source variance for the
//  purpose of activity masking.
// Eventually this should be replaced by custom no-reference routines,
//  which will be faster.
static const uint8_t VP9_VAR_OFFS[64] = {
  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128
};

#if CONFIG_VP9_HIGHBITDEPTH
static const uint16_t VP9_HIGH_VAR_OFFS_8[64] = {
  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128
};

static const uint16_t VP9_HIGH_VAR_OFFS_10[64] = {
  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4,
  128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4, 128 * 4
};

static const uint16_t VP9_HIGH_VAR_OFFS_12[64] = {
  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
  128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16, 128 * 16,
  128 * 16
};
#endif  // CONFIG_VP9_HIGHBITDEPTH

unsigned int vp9_get_sby_perpixel_variance(VP9_COMP *cpi,
                                           const struct buf_2d *ref,
                                           BLOCK_SIZE bs) {
  unsigned int sse;
  const unsigned int var =
      cpi->fn_ptr[bs].vf(ref->buf, ref->stride, VP9_VAR_OFFS, 0, &sse);
  return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
}

#if CONFIG_VP9_HIGHBITDEPTH
unsigned int vp9_high_get_sby_perpixel_variance(VP9_COMP *cpi,
                                                const struct buf_2d *ref,
                                                BLOCK_SIZE bs, int bd) {
  unsigned int var, sse;
  switch (bd) {
    case 10:
      var =
          cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
                             CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10), 0, &sse);
      break;
    case 12:
      var =
          cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
                             CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12), 0, &sse);
      break;
    case 8:
    default:
      var =
          cpi->fn_ptr[bs].vf(ref->buf, ref->stride,
                             CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8), 0, &sse);
      break;
  }
  return (unsigned int)ROUND64_POWER_OF_TWO((int64_t)var,
                                            num_pels_log2_lookup[bs]);
}
#endif  // CONFIG_VP9_HIGHBITDEPTH

static unsigned int get_sby_perpixel_diff_variance(VP9_COMP *cpi,
                                                   const struct buf_2d *ref,
                                                   int mi_row, int mi_col,
                                                   BLOCK_SIZE bs) {
  unsigned int sse, var;
  uint8_t *last_y;
  const YV12_BUFFER_CONFIG *last = get_ref_frame_buffer(cpi, LAST_FRAME);

  assert(last != NULL);
  last_y =
      &last->y_buffer[mi_row * MI_SIZE * last->y_stride + mi_col * MI_SIZE];
  var = cpi->fn_ptr[bs].vf(ref->buf, ref->stride, last_y, last->y_stride, &sse);
  return ROUND_POWER_OF_TWO(var, num_pels_log2_lookup[bs]);
}

static BLOCK_SIZE get_rd_var_based_fixed_partition(VP9_COMP *cpi, MACROBLOCK *x,
                                                   int mi_row, int mi_col) {
  unsigned int var = get_sby_perpixel_diff_variance(
      cpi, &x->plane[0].src, mi_row, mi_col, BLOCK_64X64);
  if (var < 8)
    return BLOCK_64X64;
  else if (var < 128)
    return BLOCK_32X32;
  else if (var < 2048)
    return BLOCK_16X16;
  else
    return BLOCK_8X8;
}

// Lighter version of set_offsets that only sets the mode info
// pointers.
static INLINE void set_mode_info_offsets(VP9_COMMON *const cm,
                                         MACROBLOCK *const x,
                                         MACROBLOCKD *const xd, int mi_row,
                                         int mi_col) {
  const int idx_str = xd->mi_stride * mi_row + mi_col;
  xd->mi = cm->mi_grid_visible + idx_str;
  xd->mi[0] = cm->mi + idx_str;
  x->mbmi_ext = x->mbmi_ext_base + (mi_row * cm->mi_cols + mi_col);
}

static void set_offsets(VP9_COMP *cpi, const TileInfo *const tile,
                        MACROBLOCK *const x, int mi_row, int mi_col,
                        BLOCK_SIZE bsize) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCKD *const xd = &x->e_mbd;
  MODE_INFO *mi;
  const int mi_width = num_8x8_blocks_wide_lookup[bsize];
  const int mi_height = num_8x8_blocks_high_lookup[bsize];
  const struct segmentation *const seg = &cm->seg;
  MvLimits *const mv_limits = &x->mv_limits;

  set_skip_context(xd, mi_row, mi_col);

  set_mode_info_offsets(cm, x, xd, mi_row, mi_col);

  mi = xd->mi[0];

  // Set up destination pointers.
  vp9_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);

  // Set up limit values for MV components.
  // Mv beyond the range do not produce new/different prediction block.
  mv_limits->row_min = -(((mi_row + mi_height) * MI_SIZE) + VP9_INTERP_EXTEND);
  mv_limits->col_min = -(((mi_col + mi_width) * MI_SIZE) + VP9_INTERP_EXTEND);
  mv_limits->row_max = (cm->mi_rows - mi_row) * MI_SIZE + VP9_INTERP_EXTEND;
  mv_limits->col_max = (cm->mi_cols - mi_col) * MI_SIZE + VP9_INTERP_EXTEND;

  // Set up distance of MB to edge of frame in 1/8th pel units.
  assert(!(mi_col & (mi_width - 1)) && !(mi_row & (mi_height - 1)));
  set_mi_row_col(xd, tile, mi_row, mi_height, mi_col, mi_width, cm->mi_rows,
                 cm->mi_cols);

  // Set up source buffers.
  vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);

  // R/D setup.
  x->rddiv = cpi->rd.RDDIV;
  x->rdmult = cpi->rd.RDMULT;

  // Setup segment ID.
  if (seg->enabled) {
    if (cpi->oxcf.aq_mode != VARIANCE_AQ && cpi->oxcf.aq_mode != LOOKAHEAD_AQ &&
        cpi->oxcf.aq_mode != EQUATOR360_AQ) {
      const uint8_t *const map =
          seg->update_map ? cpi->segmentation_map : cm->last_frame_seg_map;
      mi->segment_id = get_segment_id(cm, map, bsize, mi_row, mi_col);
    }
    vp9_init_plane_quantizers(cpi, x);

    x->encode_breakout = cpi->segment_encode_breakout[mi->segment_id];
  } else {
    mi->segment_id = 0;
    x->encode_breakout = cpi->encode_breakout;
  }

  // required by vp9_append_sub8x8_mvs_for_idx() and vp9_find_best_ref_mvs()
  xd->tile = *tile;
}

static void duplicate_mode_info_in_sb(VP9_COMMON *cm, MACROBLOCKD *xd,
                                      int mi_row, int mi_col,
                                      BLOCK_SIZE bsize) {
  const int block_width =
      VPXMIN(num_8x8_blocks_wide_lookup[bsize], cm->mi_cols - mi_col);
  const int block_height =
      VPXMIN(num_8x8_blocks_high_lookup[bsize], cm->mi_rows - mi_row);
  const int mi_stride = xd->mi_stride;
  MODE_INFO *const src_mi = xd->mi[0];
  int i, j;

  for (j = 0; j < block_height; ++j)
    for (i = 0; i < block_width; ++i) xd->mi[j * mi_stride + i] = src_mi;
}

static void set_block_size(VP9_COMP *const cpi, MACROBLOCK *const x,
                           MACROBLOCKD *const xd, int mi_row, int mi_col,
                           BLOCK_SIZE bsize) {
  if (cpi->common.mi_cols > mi_col && cpi->common.mi_rows > mi_row) {
    set_mode_info_offsets(&cpi->common, x, xd, mi_row, mi_col);
    xd->mi[0]->sb_type = bsize;
  }
}

typedef struct {
  int64_t sum_square_error;
  int64_t sum_error;
  int log2_count;
  int variance;
} var;

typedef struct {
  var none;
  var horz[2];
  var vert[2];
} partition_variance;

typedef struct {
  partition_variance part_variances;
  var split[4];
} v4x4;

typedef struct {
  partition_variance part_variances;
  v4x4 split[4];
} v8x8;

typedef struct {
  partition_variance part_variances;
  v8x8 split[4];
} v16x16;

typedef struct {
  partition_variance part_variances;
  v16x16 split[4];
} v32x32;

typedef struct {
  partition_variance part_variances;
  v32x32 split[4];
} v64x64;

typedef struct {
  partition_variance *part_variances;
  var *split[4];
} variance_node;

typedef enum {
  V16X16,
  V32X32,
  V64X64,
} TREE_LEVEL;

static void tree_to_node(void *data, BLOCK_SIZE bsize, variance_node *node) {
  int i;
  node->part_variances = NULL;
  switch (bsize) {
    case BLOCK_64X64: {
      v64x64 *vt = (v64x64 *)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: {
      v32x32 *vt = (v32x32 *)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: {
      v16x16 *vt = (v16x16 *)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: {
      v8x8 *vt = (v8x8 *)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_4X4: {
      v4x4 *vt = (v4x4 *)data;
      node->part_variances = &vt->part_variances;
      for (i = 0; i < 4; i++) node->split[i] = &vt->split[i];
      break;
    }
    default: {
      assert(0);
      break;
    }
  }
}

// Set variance values given sum square error, sum error, count.
static void fill_variance(int64_t s2, int64_t s, int c, var *v) {
  v->sum_square_error = s2;
  v->sum_error = s;
  v->log2_count = c;
}

static void get_variance(var *v) {
  v->variance =
      (int)(256 * (v->sum_square_error -
                   ((v->sum_error * v->sum_error) >> v->log2_count)) >>
            v->log2_count);
}

static void sum_2_variances(const var *a, const var *b, var *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 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 int set_vt_partitioning(VP9_COMP *cpi, MACROBLOCK *const x,
                               MACROBLOCKD *const xd, void *data,
                               BLOCK_SIZE bsize, int mi_row, int mi_col,
                               int64_t threshold, BLOCK_SIZE bsize_min,
                               int force_split) {
  VP9_COMMON *const cm = &cpi->common;
  variance_node vt;
  const int block_width = num_8x8_blocks_wide_lookup[bsize];
  const int block_height = num_8x8_blocks_high_lookup[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 (cm->frame_type == KEY_FRAME) get_variance(&vt.part_variances->none);
    if (mi_col + block_width / 2 < cm->mi_cols &&
        mi_row + block_height / 2 < cm->mi_rows &&
        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 (cm->frame_type == KEY_FRAME) get_variance(&vt.part_variances->none);
    // For key frame: take split for bsize above 32X32 or very high variance.
    if (cm->frame_type == KEY_FRAME &&
        (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 / 2 < cm->mi_cols &&
        mi_row + block_height / 2 < cm->mi_rows &&
        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 / 2 < cm->mi_rows) {
      BLOCK_SIZE subsize = get_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]) < 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 / 2 < cm->mi_cols) {
      BLOCK_SIZE subsize = get_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]) < 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;
}

int64_t scale_part_thresh_sumdiff(int64_t threshold_base, int speed, int width,
                                  int height, int content_state) {
  if (speed >= 8) {
    if (width <= 640 && height <= 480)
      return (5 * threshold_base) >> 2;
    else if ((content_state == kLowSadLowSumdiff) ||
             (content_state == kHighSadLowSumdiff) ||
             (content_state == kLowVarHighSumdiff))
      return (5 * threshold_base) >> 2;
  } else if (speed == 7) {
    if ((content_state == kLowSadLowSumdiff) ||
        (content_state == kHighSadLowSumdiff) ||
        (content_state == kLowVarHighSumdiff)) {
      return (5 * threshold_base) >> 2;
    }
  }
  return threshold_base;
}

// Set the variance split thresholds for following the block sizes:
// 0 - threshold_64x64, 1 - threshold_32x32, 2 - threshold_16x16,
// 3 - vbp_threshold_8x8. vbp_threshold_8x8 (to split to 4x4 partition) is
// currently only used on key frame.
static void set_vbp_thresholds(VP9_COMP *cpi, int64_t thresholds[], int q,
                               int content_state) {
  VP9_COMMON *const cm = &cpi->common;
  const int is_key_frame = (cm->frame_type == KEY_FRAME);
  const int threshold_multiplier = is_key_frame ? 20 : 1;
  int64_t threshold_base =
      (int64_t)(threshold_multiplier * cpi->y_dequant[q][1]);

  if (is_key_frame) {
    thresholds[0] = threshold_base;
    thresholds[1] = threshold_base >> 2;
    thresholds[2] = threshold_base >> 2;
    thresholds[3] = threshold_base << 2;
  } else {
    // Increase base variance threshold based on estimated noise level.
    if (cpi->noise_estimate.enabled && cm->width >= 640 && cm->height >= 480) {
      NOISE_LEVEL noise_level =
          vp9_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;
    }
#if CONFIG_VP9_TEMPORAL_DENOISING
    if (cpi->oxcf.noise_sensitivity > 0 && denoise_svc(cpi) &&
        cpi->oxcf.speed > 5 && cpi->denoiser.denoising_level >= kDenLow)
      threshold_base =
          vp9_scale_part_thresh(threshold_base, cpi->denoiser.denoising_level,
                                content_state, cpi->svc.temporal_layer_id);
    else
      threshold_base =
          scale_part_thresh_sumdiff(threshold_base, cpi->oxcf.speed, cm->width,
                                    cm->height, content_state);
#else
    // Increase base variance threshold based on content_state/sum_diff level.
    threshold_base = scale_part_thresh_sumdiff(
        threshold_base, cpi->oxcf.speed, cm->width, cm->height, content_state);
#endif
    thresholds[0] = threshold_base;
    thresholds[2] = threshold_base << cpi->oxcf.speed;
    if (cm->width <= 352 && cm->height <= 288) {
      thresholds[0] = threshold_base >> 3;
      thresholds[1] = threshold_base >> 1;
      thresholds[2] = threshold_base << 3;
    } else if (cm->width < 1280 && cm->height < 720) {
      thresholds[1] = (5 * threshold_base) >> 2;
    } else if (cm->width < 1920 && cm->height < 1080) {
      thresholds[1] = threshold_base << 1;
    } else {
      thresholds[1] = (5 * threshold_base) >> 1;
    }
  }
}

void vp9_set_variance_partition_thresholds(VP9_COMP *cpi, int q,
                                           int content_state) {
  VP9_COMMON *const cm = &cpi->common;
  SPEED_FEATURES *const sf = &cpi->sf;
  const int is_key_frame = (cm->frame_type == KEY_FRAME);
  if (sf->partition_search_type != VAR_BASED_PARTITION &&
      sf->partition_search_type != REFERENCE_PARTITION) {
    return;
  } else {
    set_vbp_thresholds(cpi, cpi->vbp_thresholds, q, content_state);
    // The thresholds below are not changed locally.
    if (is_key_frame) {
      cpi->vbp_threshold_sad = 0;
      cpi->vbp_threshold_copy = 0;
      cpi->vbp_bsize_min = BLOCK_8X8;
    } else {
      if (cm->width <= 352 && cm->height <= 288)
        cpi->vbp_threshold_sad = 10;
      else
        cpi->vbp_threshold_sad = (cpi->y_dequant[q][1] << 1) > 1000
                                     ? (cpi->y_dequant[q][1] << 1)
                                     : 1000;
      cpi->vbp_bsize_min = BLOCK_16X16;
      if (cm->width <= 352 && cm->height <= 288)
        cpi->vbp_threshold_copy = 4000;
      else if (cm->width <= 640 && cm->height <= 360)
        cpi->vbp_threshold_copy = 8000;
      else
        cpi->vbp_threshold_copy = (cpi->y_dequant[q][1] << 3) > 8000
                                      ? (cpi->y_dequant[q][1] << 3)
                                      : 8000;
    }
    cpi->vbp_threshold_minmax = 15 + (q >> 3);
  }
}

// Compute the minmax over the 8x8 subblocks.
static int compute_minmax_8x8(const uint8_t *s, int sp, const uint8_t *d,
                              int dp, int x16_idx, int y16_idx,
#if CONFIG_VP9_HIGHBITDEPTH
                              int highbd_flag,
#endif
                              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 CONFIG_VP9_HIGHBITDEPTH
      if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
        vpx_highbd_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
                              d + y8_idx * dp + x8_idx, dp, &min, &max);
      } else {
        vpx_minmax_8x8(s + y8_idx * sp + x8_idx, sp, d + y8_idx * dp + x8_idx,
                       dp, &min, &max);
      }
#else
      vpx_minmax_8x8(s + y8_idx * sp + x8_idx, sp, d + y8_idx * dp + x8_idx, dp,
                     &min, &max);
#endif
      if ((max - min) > minmax_max) minmax_max = (max - min);
      if ((max - min) < minmax_min) minmax_min = (max - min);
    }
  }
  return (minmax_max - minmax_min);
}

static void fill_variance_4x4avg(const uint8_t *s, int sp, const uint8_t *d,
                                 int dp, int x8_idx, int y8_idx, v8x8 *vst,
#if CONFIG_VP9_HIGHBITDEPTH
                                 int highbd_flag,
#endif
                                 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 CONFIG_VP9_HIGHBITDEPTH
      if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
        s_avg = vpx_highbd_avg_4x4(s + y4_idx * sp + x4_idx, sp);
        if (!is_key_frame)
          d_avg = vpx_highbd_avg_4x4(d + y4_idx * dp + x4_idx, dp);
      } else {
        s_avg = vpx_avg_4x4(s + y4_idx * sp + x4_idx, sp);
        if (!is_key_frame) d_avg = vpx_avg_4x4(d + y4_idx * dp + x4_idx, dp);
      }
#else
      s_avg = vpx_avg_4x4(s + y4_idx * sp + x4_idx, sp);
      if (!is_key_frame) d_avg = vpx_avg_4x4(d + y4_idx * dp + x4_idx, dp);
#endif
      sum = s_avg - d_avg;
      sse = sum * sum;
    }
    fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
  }
}

static void fill_variance_8x8avg(const uint8_t *s, int sp, const uint8_t *d,
                                 int dp, int x16_idx, int y16_idx, v16x16 *vst,
#if CONFIG_VP9_HIGHBITDEPTH
                                 int highbd_flag,
#endif
                                 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 CONFIG_VP9_HIGHBITDEPTH
      if (highbd_flag & YV12_FLAG_HIGHBITDEPTH) {
        s_avg = vpx_highbd_avg_8x8(s + y8_idx * sp + x8_idx, sp);
        if (!is_key_frame)
          d_avg = vpx_highbd_avg_8x8(d + y8_idx * dp + x8_idx, dp);
      } else {
        s_avg = vpx_avg_8x8(s + y8_idx * sp + x8_idx, sp);
        if (!is_key_frame) d_avg = vpx_avg_8x8(d + y8_idx * dp + x8_idx, dp);
      }
#else
      s_avg = vpx_avg_8x8(s + y8_idx * sp + x8_idx, sp);
      if (!is_key_frame) d_avg = vpx_avg_8x8(d + y8_idx * dp + x8_idx, dp);
#endif
      sum = s_avg - d_avg;
      sse = sum * sum;
    }
    fill_variance(sse, sum, 0, &vst->split[k].part_variances.none);
  }
}

// Check if most of the superblock is skin content, and if so, force split to
// 32x32, and set x->sb_is_skin for use in mode selection.
static int skin_sb_split(VP9_COMP *cpi, MACROBLOCK *x, const int low_res,
                         int mi_row, int mi_col, int *force_split) {
  VP9_COMMON *const cm = &cpi->common;
#if CONFIG_VP9_HIGHBITDEPTH
  if (cm->use_highbitdepth) return 0;
#endif
  // Avoid checking superblocks on/near boundary and avoid low resolutions.
  // Note superblock may still pick 64X64 if y_sad is very small
  // (i.e., y_sad < cpi->vbp_threshold_sad) below. For now leave this as is.
  if (!low_res && (mi_col >= 8 && mi_col + 8 < cm->mi_cols && mi_row >= 8 &&
                   mi_row + 8 < cm->mi_rows)) {
    int num_16x16_skin = 0;
    int num_16x16_nonskin = 0;
    uint8_t *ysignal = x->plane[0].src.buf;
    uint8_t *usignal = x->plane[1].src.buf;
    uint8_t *vsignal = x->plane[2].src.buf;
    int sp = x->plane[0].src.stride;
    int spuv = x->plane[1].src.stride;
    const int block_index = mi_row * cm->mi_cols + mi_col;
    const int bw = num_8x8_blocks_wide_lookup[BLOCK_64X64];
    const int bh = num_8x8_blocks_high_lookup[BLOCK_64X64];
    const int xmis = VPXMIN(cm->mi_cols - mi_col, bw);
    const int ymis = VPXMIN(cm->mi_rows - mi_row, bh);
    // Loop through the 16x16 sub-blocks.
    int i, j;
    for (i = 0; i < ymis; i += 2) {
      for (j = 0; j < xmis; j += 2) {
        int bl_index = block_index + i * cm->mi_cols + j;
        int bl_index1 = bl_index + 1;
        int bl_index2 = bl_index + cm->mi_cols;
        int bl_index3 = bl_index2 + 1;
        int consec_zeromv =
            VPXMIN(cpi->consec_zero_mv[bl_index],
                   VPXMIN(cpi->consec_zero_mv[bl_index1],
                          VPXMIN(cpi->consec_zero_mv[bl_index2],
                                 cpi->consec_zero_mv[bl_index3])));
        int is_skin = vp9_compute_skin_block(
            ysignal, usignal, vsignal, sp, spuv, BLOCK_16X16, consec_zeromv, 0);
        num_16x16_skin += is_skin;
        num_16x16_nonskin += (1 - is_skin);
        if (num_16x16_nonskin > 3) {
          // Exit loop if at least 4 of the 16x16 blocks are not skin.
          i = ymis;
          break;
        }
        ysignal += 16;
        usignal += 8;
        vsignal += 8;
      }
      ysignal += (sp << 4) - 64;
      usignal += (spuv << 3) - 32;
      vsignal += (spuv << 3) - 32;
    }
    if (num_16x16_skin > 12) {
      *force_split = 1;
      return 1;
    }
  }
  return 0;
}

static void set_low_temp_var_flag(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
                                  v64x64 *vt, int64_t thresholds[],
                                  MV_REFERENCE_FRAME ref_frame_partition,
                                  int mi_col, int mi_row) {
  int i, j;
  VP9_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.short_circuit_low_temp_var == 1 ||
       (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))) {
    if (xd->mi[0]->sb_type == BLOCK_64X64) {
      if ((vt->part_variances).none.variance < (thresholds[0] >> 1))
        x->variance_low[0] = 1;
    } else if (xd->mi[0]->sb_type == BLOCK_64X32) {
      for (i = 0; i < 2; i++) {
        if (vt->part_variances.horz[i].variance < (thresholds[0] >> 2))
          x->variance_low[i + 1] = 1;
      }
    } else if (xd->mi[0]->sb_type == BLOCK_32X64) {
      for (i = 0; i < 2; i++) {
        if (vt->part_variances.vert[i].variance < (thresholds[0] >> 2))
          x->variance_low[i + 3] = 1;
      }
    } else {
      for (i = 0; i < 4; i++) {
        const int idx[4][2] = { { 0, 0 }, { 0, 4 }, { 4, 0 }, { 4, 4 } };
        const int idx_str =
            cm->mi_stride * (mi_row + idx[i][0]) + mi_col + idx[i][1];
        MODE_INFO **this_mi = cm->mi_grid_visible + idx_str;

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

        if ((*this_mi)->sb_type == BLOCK_32X32) {
          int64_t threshold_32x32 = (cpi->sf.short_circuit_low_temp_var == 1 ||
                                     cpi->sf.short_circuit_low_temp_var == 3)
                                        ? ((5 * thresholds[1]) >> 3)
                                        : (thresholds[1] >> 1);
          if (vt->split[i].part_variances.none.variance < threshold_32x32)
            x->variance_low[i + 5] = 1;
        } else if (cpi->sf.short_circuit_low_temp_var >= 2) {
          // 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 (j = 0; j < 4; j++) {
              if (vt->split[i].split[j].part_variances.none.variance <
                  (thresholds[2] >> 8))
                x->variance_low[(i << 2) + j + 9] = 1;
            }
          }
        }
      }
    }
  }
}

static void copy_partitioning_helper(VP9_COMP *cpi, MACROBLOCK *x,
                                     MACROBLOCKD *xd, BLOCK_SIZE bsize,
                                     int mi_row, int mi_col) {
  VP9_COMMON *const cm = &cpi->common;
  BLOCK_SIZE *prev_part = cpi->prev_partition;
  int start_pos = mi_row * cm->mi_stride + mi_col;

  const int bsl = b_width_log2_lookup[bsize];
  const int bs = (1 << bsl) / 4;
  BLOCK_SIZE subsize;
  PARTITION_TYPE partition;

  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;

  partition = partition_lookup[bsl][prev_part[start_pos]];
  subsize = get_subsize(bsize, partition);

  if (subsize < BLOCK_8X8) {
    set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
  } else {
    switch (partition) {
      case PARTITION_NONE:
        set_block_size(cpi, x, xd, mi_row, mi_col, bsize);
        break;
      case PARTITION_HORZ:
        set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
        set_block_size(cpi, x, xd, mi_row + bs, mi_col, subsize);
        break;
      case PARTITION_VERT:
        set_block_size(cpi, x, xd, mi_row, mi_col, subsize);
        set_block_size(cpi, x, xd, mi_row, mi_col + bs, subsize);
        break;
      case PARTITION_SPLIT:
        copy_partitioning_helper(cpi, x, xd, subsize, mi_row, mi_col);
        copy_partitioning_helper(cpi, x, xd, subsize, mi_row + bs, mi_col);
        copy_partitioning_helper(cpi, x, xd, subsize, mi_row, mi_col + bs);
        copy_partitioning_helper(cpi, x, xd, subsize, mi_row + bs, mi_col + bs);
        break;
      default: assert(0);
    }
  }
}

static int copy_partitioning(VP9_COMP *cpi, MACROBLOCK *x, MACROBLOCKD *xd,
                             int mi_row, int mi_col, int segment_id,
                             int sb_offset) {
  int svc_copy_allowed = 1;
  int frames_since_key_thresh = 1;
  if (cpi->use_svc) {
    // For SVC, don't allow copy if base spatial layer is key frame, or if
    // frame is not a temporal enhancement layer frame.
    int layer = LAYER_IDS_TO_IDX(0, cpi->svc.temporal_layer_id,
                                 cpi->svc.number_temporal_layers);
    const LAYER_CONTEXT *lc = &cpi->svc.layer_context[layer];
    if (lc->is_key_frame ||
        (cpi->svc.temporal_layer_id != cpi->svc.number_temporal_layers - 1 &&
         cpi->svc.number_temporal_layers > 1))
      svc_copy_allowed = 0;
    frames_since_key_thresh = cpi->svc.number_spatial_layers << 1;
  }
  if (cpi->rc.frames_since_key > frames_since_key_thresh && svc_copy_allowed &&
      !cpi->resize_pending && segment_id == CR_SEGMENT_ID_BASE &&
      cpi->prev_segment_id[sb_offset] == CR_SEGMENT_ID_BASE &&
      cpi->copied_frame_cnt[sb_offset] < cpi->max_copied_frame) {
    if (cpi->prev_partition != NULL) {
      copy_partitioning_helper(cpi, x, xd, BLOCK_64X64, mi_row, mi_col);
      cpi->copied_frame_cnt[sb_offset] += 1;
      memcpy(x->variance_low, &(cpi->prev_variance_low[sb_offset * 25]),
             sizeof(x->variance_low));
      return 1;
    }
  }

  return 0;
}

static void update_prev_partition(VP9_COMP *cpi, BLOCK_SIZE bsize, int mi_row,
                                  int mi_col) {
  VP9_COMMON *const cm = &cpi->common;
  BLOCK_SIZE *prev_part = cpi->prev_partition;
  int start_pos = mi_row * cm->mi_stride + mi_col;
  const int bsl = b_width_log2_lookup[bsize];
  const int bs = (1 << bsl) / 4;
  BLOCK_SIZE subsize;
  PARTITION_TYPE partition;
  const MODE_INFO *mi = NULL;

  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols) return;

  mi = cm->mi_grid_visible[start_pos];
  partition = partition_lookup[bsl][mi->sb_type];
  subsize = get_subsize(bsize, partition);
  if (subsize < BLOCK_8X8) {
    prev_part[start_pos] = bsize;
  } else {
    switch (partition) {
      case PARTITION_NONE: prev_part[start_pos] = bsize; break;
      case PARTITION_HORZ:
        prev_part[start_pos] = subsize;
        if (mi_row + bs < cm->mi_rows)
          prev_part[start_pos + bs * cm->mi_stride] = subsize;
        break;
      case PARTITION_VERT:
        prev_part[start_pos] = subsize;
        if (mi_col + bs < cm->mi_cols) prev_part[start_pos + bs] = subsize;
        break;
      case PARTITION_SPLIT:
        update_prev_partition(cpi, subsize, mi_row, mi_col);
        update_prev_partition(cpi, subsize, mi_row + bs, mi_col);
        update_prev_partition(cpi, subsize, mi_row, mi_col + bs);
        update_prev_partition(cpi, subsize, mi_row + bs, mi_col + bs);
        break;
      default: assert(0);
    }
  }
}

static void chroma_check(VP9_COMP *cpi, MACROBLOCK *x, int bsize,
                         unsigned int y_sad, int is_key_frame) {
  int i;
  MACROBLOCKD *xd = &x->e_mbd;

  if (is_key_frame) return;

  // For speed >= 8, avoid the chroma check if y_sad is above threshold.
  if (cpi->oxcf.speed >= 8) {
    if (y_sad > cpi->vbp_thresholds[1] &&
        (!cpi->noise_estimate.enabled ||
         vp9_noise_estimate_extract_level(&cpi->noise_estimate) < kMedium))
      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);

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

    // TODO(marpan): Investigate if we should lower this threshold if