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_ports/mem.h"
#include "vpx_ports/vpx_timer.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_systemdependent.h"
#include "vp9/common/vp9_tile_common.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);

// This is used as a reference when computing the source variance for the
//  purposes 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 ROUND_POWER_OF_TWO(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;
  MB_MODE_INFO *mbmi;
  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;

  set_skip_context(xd, mi_row, mi_col);

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

  mbmi = &xd->mi[0]->mbmi;

  // 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.
  x->mv_row_min = -(((mi_row + mi_height) * MI_SIZE) + VP9_INTERP_EXTEND);
  x->mv_col_min = -(((mi_col + mi_width) * MI_SIZE) + VP9_INTERP_EXTEND);
  x->mv_row_max = (cm->mi_rows - mi_row) * MI_SIZE + VP9_INTERP_EXTEND;
  x->mv_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) {
      const uint8_t *const map = seg->update_map ? cpi->segmentation_map
                                                 : cm->last_frame_seg_map;
      mbmi->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[mbmi->segment_id];
  } else {
    mbmi->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 = num_8x8_blocks_wide_lookup[bsize];
  const int block_height = num_8x8_blocks_high_lookup[bsize];
  int i, j;
  for (j = 0; j < block_height; ++j)
    for (i = 0; i < block_width; ++i) {
      if (mi_row + j < cm->mi_rows && mi_col + i < cm->mi_cols)
        xd->mi[j * xd->mi_stride + i] = xd->mi[0];
    }
}

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]->mbmi.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];
  const int low_res = (cm->width <= 352 && cm->height <= 288);

  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 (low_res || 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 (low_res || 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;
}

// 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) {
  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;
  const 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 {
    thresholds[1] = threshold_base;
    if (cm->width <= 352 && cm->height <= 288) {
      thresholds[0] = threshold_base >> 2;
      thresholds[2] = threshold_base << 3;
    } else {
      thresholds[0] = threshold_base;
      thresholds[1] = (5 * threshold_base) >> 2;
      if (cm->width >= 1920 && cm->height >= 1080)
        thresholds[1] = (7 * threshold_base) >> 2;
      thresholds[2] = threshold_base << cpi->oxcf.speed;
    }
  }
}

void vp9_set_variance_partition_thresholds(VP9_COMP *cpi, int q) {
  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);
    // The thresholds below are not changed locally.
    if (is_key_frame) {
      cpi->vbp_threshold_sad = 0;
      cpi->vbp_bsize_min = BLOCK_8X8;
    } else {
      if (cm->width <= 352 && cm->height <= 288)
        cpi->vbp_threshold_sad = 100;
      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;
    }
    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) {
        vp9_highbd_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
                              d + y8_idx * dp + x8_idx, dp,
                              &min, &max);
      } else {
        vp9_minmax_8x8(s + y8_idx * sp + x8_idx, sp,
                       d + y8_idx * dp + x8_idx, dp,
                       &min, &max);
      }
#else
      vp9_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 = vp9_highbd_avg_4x4(s + y4_idx * sp + x4_idx, sp);
        if (!is_key_frame)
          d_avg = vp9_highbd_avg_4x4(d + y4_idx * dp + x4_idx, dp);
      } else {
        s_avg = vp9_avg_4x4(s + y4_idx * sp + x4_idx, sp);
        if (!is_key_frame)
          d_avg = vp9_avg_4x4(d + y4_idx * dp + x4_idx, dp);
      }
#else
      s_avg = vp9_avg_4x4(s + y4_idx * sp + x4_idx, sp);
      if (!is_key_frame)
        d_avg = vp9_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 = vp9_highbd_avg_8x8(s + y8_idx * sp + x8_idx, sp);
        if (!is_key_frame)
          d_avg = vp9_highbd_avg_8x8(d + y8_idx * dp + x8_idx, dp);
      } else {
        s_avg = vp9_avg_8x8(s + y8_idx * sp + x8_idx, sp);
        if (!is_key_frame)
          d_avg = vp9_avg_8x8(d + y8_idx * dp + x8_idx, dp);
      }
#else
      s_avg = vp9_avg_8x8(s + y8_idx * sp + x8_idx, sp);
      if (!is_key_frame)
        d_avg = vp9_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);
  }
}

// This function chooses partitioning based on the variance between source and
// reconstructed last, where variance is computed for down-sampled inputs.
static int choose_partitioning(VP9_COMP *cpi,
                                const TileInfo *const tile,
                                MACROBLOCK *x,
                                int mi_row, int mi_col) {
  VP9_COMMON * const cm = &cpi->common;
  MACROBLOCKD *xd = &x->e_mbd;
  int i, j, k, m;
  v64x64 vt;
  v16x16 vt2[16];
  int force_split[21];
  uint8_t *s;
  const uint8_t *d;
  int sp;
  int dp;
  int pixels_wide = 64, pixels_high = 64;
  int64_t thresholds[4] = {cpi->vbp_thresholds[0], cpi->vbp_thresholds[1],
      cpi->vbp_thresholds[2], cpi->vbp_thresholds[3]};

  // Always use 4x4 partition for key frame.
  const int is_key_frame = (cm->frame_type == KEY_FRAME);
  const int use_4x4_partition = is_key_frame;
  const int low_res = (cm->width <= 352 && cm->height <= 288);
  int variance4x4downsample[16];

  int segment_id = CR_SEGMENT_ID_BASE;
  if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ && cm->seg.enabled) {
    const uint8_t *const map = cm->seg.update_map ? cpi->segmentation_map :
                                                    cm->last_frame_seg_map;
    segment_id = get_segment_id(cm, map, BLOCK_64X64, mi_row, mi_col);

    if (cyclic_refresh_segment_id_boosted(segment_id)) {
      int q = vp9_get_qindex(&cm->seg, segment_id, cm->base_qindex);
      set_vbp_thresholds(cpi, thresholds, q);
    }
  }

  set_offsets(cpi, tile, x, mi_row, mi_col, BLOCK_64X64);

  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);

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

  if (!is_key_frame && !(is_one_pass_cbr_svc(cpi) &&
      cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame)) {
    // In the case of spatial/temporal scalable coding, the assumption here is
    // that the temporal reference frame will always be of type LAST_FRAME.
    // TODO(marpan): If that assumption is broken, we need to revisit this code.
    MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
    unsigned int uv_sad;
    const YV12_BUFFER_CONFIG *yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);

    const YV12_BUFFER_CONFIG *yv12_g = NULL;
    unsigned int y_sad, y_sad_g;
    const BLOCK_SIZE bsize = BLOCK_32X32
        + (mi_col + 4 < cm->mi_cols) * 2 + (mi_row + 4 < cm->mi_rows);

    assert(yv12 != NULL);

    if (!(is_one_pass_cbr_svc(cpi) && cpi->svc.spatial_layer_id)) {
      // For now, GOLDEN will not be used for non-zero spatial layers, since
      // it may not be a temporal reference.
      yv12_g = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
    }

    if (yv12_g && yv12_g != yv12) {
      vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
                           &cm->frame_refs[GOLDEN_FRAME - 1].sf);
      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);
    } else {
      y_sad_g = UINT_MAX;
    }

    vp9_setup_pre_planes(xd, 0, yv12, mi_row, mi_col,
                         &cm->frame_refs[LAST_FRAME - 1].sf);
    mbmi->ref_frame[0] = LAST_FRAME;
    mbmi->ref_frame[1] = NONE;
    mbmi->sb_type = BLOCK_64X64;
    mbmi->mv[0].as_int = 0;
    mbmi->interp_filter = BILINEAR;

    y_sad = vp9_int_pro_motion_estimation(cpi, x, bsize, mi_row, mi_col);
    if (y_sad_g < y_sad) {
      vp9_setup_pre_planes(xd, 0, yv12_g, mi_row, mi_col,
                           &cm->frame_refs[GOLDEN_FRAME - 1].sf);
      mbmi->ref_frame[0] = GOLDEN_FRAME;
      mbmi->mv[0].as_int = 0;
      y_sad = y_sad_g;
    } else {
      x->pred_mv[LAST_FRAME] = mbmi->mv[0].as_mv;
    }

    vp9_build_inter_predictors_sb(xd, mi_row, mi_col, BLOCK_64X64);

    for (i = 1; i <= 2; ++i) {
      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 = UINT_MAX;
      else
        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);
    }

    d = xd->plane[0].dst.buf;
    dp = xd->plane[0].dst.stride;

    // If the y_sad is very small, take 64x64 as partition and exit.
    // Don't check on boosted segment for now, as 64x64 is suppressed there.
    if (segment_id == CR_SEGMENT_ID_BASE &&
        y_sad < cpi->vbp_threshold_sad) {
      const int block_width = num_8x8_blocks_wide_lookup[BLOCK_64X64];
      const int block_height = num_8x8_blocks_high_lookup[BLOCK_64X64];
      if (mi_col + block_width / 2 < cm->mi_cols &&
          mi_row + block_height / 2 < cm->mi_rows) {
        set_block_size(cpi, x, xd, mi_row, mi_col, BLOCK_64X64);
        return 0;
      }
    }
  } else {
    d = VP9_VAR_OFFS;
    dp = 0;
#if CONFIG_VP9_HIGHBITDEPTH
    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
      switch (xd->bd) {
        case 10:
          d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_10);
          break;
        case 12:
          d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_12);
          break;
        case 8:
        default:
          d = CONVERT_TO_BYTEPTR(VP9_HIGH_VAR_OFFS_8);
          break;
      }
    }
#endif  // CONFIG_VP9_HIGHBITDEPTH
  }

  // Index for force_split: 0 for 64x64, 1-4 for 32x32 blocks,
  // 5-20 for the 16x16 blocks.
  force_split[0] = 0;
  // Fill in the entire tree of 8x8 (or 4x4 under some conditions) variances
  // for splits.
  for (i = 0; i < 4; i++) {
    const int x32_idx = ((i & 1) << 5);
    const int y32_idx = ((i >> 1) << 5);
    const int i2 = i << 2;
    force_split[i + 1] = 0;
    for (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 = 5 + i2 + j;
      v16x16 *vst = &vt.split[i].split[j];
      force_split[split_index] = 0;
      variance4x4downsample[i2 + j] = 0;
      if (!is_key_frame) {
        fill_variance_8x8avg(s, sp, d, dp, x16_idx, y16_idx, vst,
#if CONFIG_VP9_HIGHBITDEPTH
                            xd->cur_buf->flags,
#endif
                            pixels_wide,
                            pixels_high,
                            is_key_frame);
        fill_variance_tree(&vt.split[i].split[j], BLOCK_16X16);
        get_variance(&vt.split[i].split[j].part_variances.none);
        if (vt.split[i].split[j].part_variances.none.variance >
            thresholds[2]) {
          // 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[i + 1] = 1;
          force_split[0] = 1;
        } else if (vt.split[i].split[j].part_variances.none.variance >
                   thresholds[1] &&
                   !cyclic_refresh_segment_id_boosted(segment_id)) {
          // 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(s, sp, d, dp, x16_idx, y16_idx,
#if CONFIG_VP9_HIGHBITDEPTH
                                          xd->cur_buf->flags,
#endif
                                          pixels_wide, pixels_high);
          if (minmax > cpi->vbp_threshold_minmax) {
            force_split[split_index] = 1;
            force_split[i + 1] = 1;
            force_split[0] = 1;
          }
        }
      }
      // TODO(marpan): There is an issue with variance based on 4x4 average in
      // svc mode, don't allow it for now.
      if (is_key_frame || (low_res && !cpi->use_svc &&
          vt.split[i].split[j].part_variances.none.variance >
          (thresholds[1] << 1))) {
        force_split[split_index] = 0;
        // Go down to 4x4 down-sampling for variance.
        variance4x4downsample[i2 + j] = 1;
        for (k = 0; k < 4; k++) {
          int x8_idx = x16_idx + ((k & 1) << 3);
          int y8_idx = y16_idx + ((k >> 1) << 3);
          v8x8 *vst2 = is_key_frame ? &vst->split[k] :
              &vt2[i2 + j].split[k];
          fill_variance_4x4avg(s, sp, d, dp, x8_idx, y8_idx, vst2,
#if CONFIG_VP9_HIGHBITDEPTH
                               xd->cur_buf->flags,
#endif
                               pixels_wide,
                               pixels_high,
                               is_key_frame);
        }
      }
    }
  }

  // Fill the rest of the variance tree by summing split partition values.
  for (i = 0; i < 4; i++) {
    const int i2 = i << 2;
    for (j = 0; j < 4; j++) {
      if (variance4x4downsample[i2 + j] == 1) {
        v16x16 *vtemp = (!is_key_frame) ? &vt2[i2 + j] :
            &vt.split[i].split[j];
        for (m = 0; m < 4; m++)
          fill_variance_tree(&vtemp->split[m], BLOCK_8X8);
        fill_variance_tree(vtemp, BLOCK_16X16);
      }
    }
    fill_variance_tree(&vt.split[i], BLOCK_32X32);
    // If variance of this 32x32 block is above the threshold, force the block
    // to split. This also forces a split on the upper (64x64) level.
    if (!force_split[i + 1]) {
      get_variance(&vt.split[i].part_variances.none);
      if (vt.split[i].part_variances.none.variance > thresholds[1]) {
        force_split[i + 1] = 1;
        force_split[0] = 1;
      }
    }
  }
  if (!force_split[0]) {
    fill_variance_tree(&vt, BLOCK_64X64);
    get_variance(&vt.part_variances.none);
  }

  // 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 ( mi_col + 8 > cm->mi_cols || mi_row + 8 > cm->mi_rows ||
      !set_vt_partitioning(cpi, x, xd, &vt, BLOCK_64X64, mi_row, mi_col,
                           thresholds[0], BLOCK_16X16, force_split[0])) {
    for (i = 0; i < 4; ++i) {
      const int x32_idx = ((i & 1) << 2);
      const int y32_idx = ((i >> 1) << 2);
      const int i2 = i << 2;
      if (!set_vt_partitioning(cpi, x, xd, &vt.split[i], BLOCK_32X32,
                               (mi_row + y32_idx), (mi_col + x32_idx),
                               thresholds[1], BLOCK_16X16,
                               force_split[i + 1])) {
        for (j = 0; j < 4; ++j) {
          const int x16_idx = ((j & 1) << 1);
          const int y16_idx = ((j >> 1) << 1);
          // 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.
          v16x16 *vtemp = (!is_key_frame &&
                           variance4x4downsample[i2 + j] == 1) ?
                           &vt2[i2 + j] : &vt.split[i].split[j];
          if (!set_vt_partitioning(cpi, x, xd, vtemp, BLOCK_16X16,
                                   mi_row + y32_idx + y16_idx,
                                   mi_col + x32_idx + x16_idx,
                                   thresholds[2],
                                   cpi->vbp_bsize_min,
                                   force_split[5 + i2  + j])) {
            for (k = 0; k < 4; ++k) {
              const int x8_idx = (k & 1);
              const int y8_idx = (k >> 1);
              if (use_4x4_partition) {
                if (!set_vt_partitioning(cpi, x, xd, &vtemp->split[k],
                                         BLOCK_8X8,
                                         mi_row + y32_idx + y16_idx + y8_idx,
                                         mi_col + x32_idx + x16_idx + x8_idx,
                                         thresholds[3], BLOCK_8X8, 0)) {
                  set_block_size(cpi, x, xd,
                                 (mi_row + y32_idx + y16_idx + y8_idx),
                                 (mi_col + x32_idx + x16_idx + x8_idx),
                                 BLOCK_4X4);
                }
              } else {
                set_block_size(cpi, x, xd,
                               (mi_row + y32_idx + y16_idx + y8_idx),
                               (mi_col + x32_idx + x16_idx + x8_idx),
                               BLOCK_8X8);
              }
            }
          }
        }
      }
    }
  }
  return 0;
}

static void update_state(VP9_COMP *cpi, ThreadData *td,
                         PICK_MODE_CONTEXT *ctx,
                         int mi_row, int mi_col, BLOCK_SIZE bsize,
                         int output_enabled) {
  int i, x_idx, y;
  VP9_COMMON *const cm = &cpi->common;
  RD_COUNTS *const rdc = &td->rd_counts;
  MACROBLOCK *const x = &td->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  struct macroblock_plane *const p = x->plane;
  struct macroblockd_plane *const pd = xd->plane;
  MODE_INFO *mi = &ctx->mic;
  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
  MODE_INFO *mi_addr = xd->mi[0];
  const struct segmentation *const seg = &cm->seg;
  const int bw = num_8x8_blocks_wide_lookup[mi->mbmi.sb_type];
  const int bh = num_8x8_blocks_high_lookup[mi->mbmi.sb_type];
  const int x_mis = MIN(bw, cm->mi_cols - mi_col);
  const int y_mis = MIN(bh, cm->mi_rows - mi_row);
  MV_REF *const frame_mvs =
      cm->cur_frame->mvs + mi_row * cm->mi_cols + mi_col;
  int w, h;

  const int mis = cm->mi_stride;
  const int mi_width = num_8x8_blocks_wide_lookup[bsize];
  const int mi_height = num_8x8_blocks_high_lookup[bsize];
  int max_plane;

  assert(mi->mbmi.sb_type == bsize);

  *mi_addr = *mi;
  *x->mbmi_ext = ctx->mbmi_ext;

  // If segmentation in use
  if (seg->enabled) {
    // For in frame complexity AQ copy the segment id from the segment map.
    if (cpi->oxcf.aq_mode == COMPLEXITY_AQ) {