vp9_encodeframe.c

      *bh = num_8x8_blocks_high_lookup[bsize];
      *bw = num_8x8_blocks_wide_lookup[bsize];
      if ((*bh <= rows_left) && (*bw <= cols_left)) {
        break;
      }
    }
  }
  return bsize;
}

// This function attempts to set all mode info entries in a given SB64
// to the same block partition size.
// However, at the bottom and right borders of the image the requested size
// may not be allowed in which case this code attempts to choose the largest
// allowable partition.
static void set_partitioning(VP9_COMP *cpi, const TileInfo *const tile,
                             MODE_INFO **mi_8x8, int mi_row, int mi_col,
                             BLOCK_SIZE bsize) {
  VP9_COMMON *const cm = &cpi->common;
  const int mis = cm->mode_info_stride;
  int row8x8_remaining = tile->mi_row_end - mi_row;
  int col8x8_remaining = tile->mi_col_end - mi_col;
  int block_row, block_col;
  MODE_INFO *mi_upper_left = cm->mi + mi_row * mis + mi_col;
  int bh = num_8x8_blocks_high_lookup[bsize];
  int bw = num_8x8_blocks_wide_lookup[bsize];

  assert((row8x8_remaining > 0) && (col8x8_remaining > 0));

  // Apply the requested partition size to the SB64 if it is all "in image"
  if ((col8x8_remaining >= MI_BLOCK_SIZE) &&
      (row8x8_remaining >= MI_BLOCK_SIZE)) {
    for (block_row = 0; block_row < MI_BLOCK_SIZE; block_row += bh) {
      for (block_col = 0; block_col < MI_BLOCK_SIZE; block_col += bw) {
        int index = block_row * mis + block_col;
        mi_8x8[index] = mi_upper_left + index;
        mi_8x8[index]->mbmi.sb_type = bsize;
      }
    }
  } else {
    // Else this is a partial SB64.
    for (block_row = 0; block_row < MI_BLOCK_SIZE; block_row += bh) {
      for (block_col = 0; block_col < MI_BLOCK_SIZE; block_col += bw) {
        int index = block_row * mis + block_col;
        // Find a partition size that fits
        bsize = find_partition_size(bsize,
                                    (row8x8_remaining - block_row),
                                    (col8x8_remaining - block_col), &bh, &bw);
        mi_8x8[index] = mi_upper_left + index;
        mi_8x8[index]->mbmi.sb_type = bsize;
      }
    }
  }
}

static void copy_partitioning(VP9_COMMON *cm, MODE_INFO **mi_8x8,
                              MODE_INFO **prev_mi_8x8) {
  const int mis = cm->mode_info_stride;
  int block_row, block_col;

  for (block_row = 0; block_row < 8; ++block_row) {
    for (block_col = 0; block_col < 8; ++block_col) {
      MODE_INFO *const prev_mi = prev_mi_8x8[block_row * mis + block_col];
      const BLOCK_SIZE sb_type = prev_mi ? prev_mi->mbmi.sb_type : 0;
      if (prev_mi) {
        const ptrdiff_t offset = prev_mi - cm->prev_mi;
        mi_8x8[block_row * mis + block_col] = cm->mi + offset;
        mi_8x8[block_row * mis + block_col]->mbmi.sb_type = sb_type;
      }
    }
  }
}

static int sb_has_motion(const VP9_COMMON *cm, MODE_INFO **prev_mi_8x8) {
  const int mis = cm->mode_info_stride;
  int block_row, block_col;

  if (cm->prev_mi) {
    for (block_row = 0; block_row < 8; ++block_row) {
      for (block_col = 0; block_col < 8; ++block_col) {
        const MODE_INFO *prev_mi = prev_mi_8x8[block_row * mis + block_col];
        if (prev_mi) {
          if (abs(prev_mi->mbmi.mv[0].as_mv.row) >= 8 ||
              abs(prev_mi->mbmi.mv[0].as_mv.col) >= 8)
            return 1;
        }
      }
    }
  }
  return 0;
}

static void update_state_rt(VP9_COMP *cpi, PICK_MODE_CONTEXT *ctx,
                         BLOCK_SIZE bsize, int output_enabled) {
  int i;
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *const mbmi = &xd->mi_8x8[0]->mbmi;

  x->skip = ctx->skip;

#if CONFIG_INTERNAL_STATS
  if (frame_is_intra_only(cm)) {
    static const int kf_mode_index[] = {
      THR_DC /*DC_PRED*/,
      THR_V_PRED /*V_PRED*/,
      THR_H_PRED /*H_PRED*/,
      THR_D45_PRED /*D45_PRED*/,
      THR_D135_PRED /*D135_PRED*/,
      THR_D117_PRED /*D117_PRED*/,
      THR_D153_PRED /*D153_PRED*/,
      THR_D207_PRED /*D207_PRED*/,
      THR_D63_PRED /*D63_PRED*/,
      THR_TM /*TM_PRED*/,
    };
    ++cpi->mode_chosen_counts[kf_mode_index[mbmi->mode]];
  } else {
    // Note how often each mode chosen as best
    ++cpi->mode_chosen_counts[ctx->best_mode_index];
  }
#endif
  if (!frame_is_intra_only(cm)) {
    if (is_inter_block(mbmi)) {
      if (mbmi->sb_type < BLOCK_8X8 || mbmi->mode == NEWMV) {
        int_mv best_mv[2];
        for (i = 0; i < 1 + has_second_ref(mbmi); ++i)
          best_mv[i].as_int = mbmi->ref_mvs[mbmi->ref_frame[i]][0].as_int;
        vp9_update_mv_count(cpi, x, best_mv);
      }

      if (cm->interp_filter == SWITCHABLE) {
        const int ctx = vp9_get_pred_context_switchable_interp(xd);
        ++cm->counts.switchable_interp[ctx][mbmi->interp_filter];
      }
    }
  }
}

static void encode_b_rt(VP9_COMP *cpi, const TileInfo *const tile,
                     TOKENEXTRA **tp, int mi_row, int mi_col,
                     int output_enabled, BLOCK_SIZE bsize) {
  MACROBLOCK *const x = &cpi->mb;

  if (bsize < BLOCK_8X8) {
    // When ab_index = 0 all sub-blocks are handled, so for ab_index != 0
    // there is nothing to be done.
    if (x->ab_index > 0)
      return;
  }
  set_offsets(cpi, tile, mi_row, mi_col, bsize);
  update_state_rt(cpi, get_block_context(x, bsize), bsize, output_enabled);

  encode_superblock(cpi, tp, output_enabled, mi_row, mi_col, bsize);
  update_stats(cpi);

  (*tp)->token = EOSB_TOKEN;
  (*tp)++;
}

static void encode_sb_rt(VP9_COMP *cpi, const TileInfo *const tile,
                         TOKENEXTRA **tp, int mi_row, int mi_col,
                         int output_enabled, BLOCK_SIZE bsize) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  const int bsl = b_width_log2(bsize), hbs = (1 << bsl) / 4;
  int ctx;
  PARTITION_TYPE partition;
  BLOCK_SIZE subsize;

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

  if (bsize >= BLOCK_8X8) {
    MACROBLOCKD *const xd = &cpi->mb.e_mbd;
    const int idx_str = xd->mode_info_stride * mi_row + mi_col;
    MODE_INFO ** mi_8x8 = cm->mi_grid_visible + idx_str;
    ctx = partition_plane_context(cpi->above_seg_context, cpi->left_seg_context,
                                 mi_row, mi_col, bsize);
    subsize = mi_8x8[0]->mbmi.sb_type;
  } else {
    ctx = 0;
    subsize = BLOCK_4X4;
  }

  partition = partition_lookup[bsl][subsize];

  switch (partition) {
    case PARTITION_NONE:
      if (output_enabled && bsize >= BLOCK_8X8)
        cm->counts.partition[ctx][PARTITION_NONE]++;
      encode_b_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize);
      break;
    case PARTITION_VERT:
      if (output_enabled)
        cm->counts.partition[ctx][PARTITION_VERT]++;
      *get_sb_index(x, subsize) = 0;
      encode_b_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize);
      if (mi_col + hbs < cm->mi_cols) {
        *get_sb_index(x, subsize) = 1;
        encode_b_rt(cpi, tile, tp, mi_row, mi_col + hbs, output_enabled,
                    subsize);
      }
      break;
    case PARTITION_HORZ:
      if (output_enabled)
        cm->counts.partition[ctx][PARTITION_HORZ]++;
      *get_sb_index(x, subsize) = 0;
      encode_b_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize);
      if (mi_row + hbs < cm->mi_rows) {
        *get_sb_index(x, subsize) = 1;
        encode_b_rt(cpi, tile, tp, mi_row + hbs, mi_col, output_enabled,
                    subsize);
      }
      break;
    case PARTITION_SPLIT:
      subsize = get_subsize(bsize, PARTITION_SPLIT);
      if (output_enabled)
        cm->counts.partition[ctx][PARTITION_SPLIT]++;

      *get_sb_index(x, subsize) = 0;
      encode_sb_rt(cpi, tile, tp, mi_row, mi_col, output_enabled, subsize);
      *get_sb_index(x, subsize) = 1;
      encode_sb_rt(cpi, tile, tp, mi_row, mi_col + hbs, output_enabled,
                   subsize);
      *get_sb_index(x, subsize) = 2;
      encode_sb_rt(cpi, tile, tp, mi_row + hbs, mi_col, output_enabled,
                   subsize);
      *get_sb_index(x, subsize) = 3;
      encode_sb_rt(cpi, tile, tp, mi_row + hbs, mi_col + hbs, output_enabled,
                subsize);
      break;
    default:
      assert("Invalid partition type.");
  }

  if (partition != PARTITION_SPLIT || bsize == BLOCK_8X8)
    update_partition_context(cpi->above_seg_context, cpi->left_seg_context,
                             mi_row, mi_col, subsize, bsize);
}

static void rd_use_partition(VP9_COMP *cpi,
                             const TileInfo *const tile,
                             MODE_INFO **mi_8x8,
                             TOKENEXTRA **tp, int mi_row, int mi_col,
                             BLOCK_SIZE bsize, int *rate, int64_t *dist,
                             int do_recon) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  const int mis = cm->mode_info_stride;
  const int bsl = b_width_log2(bsize);
  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
  const int ms = num_4x4_blocks_wide / 2;
  const int mh = num_4x4_blocks_high / 2;
  const int bss = (1 << bsl) / 4;
  int i, pl;
  PARTITION_TYPE partition = PARTITION_NONE;
  BLOCK_SIZE subsize;
  ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
  PARTITION_CONTEXT sl[8], sa[8];
  int last_part_rate = INT_MAX;
  int64_t last_part_dist = INT64_MAX;
  int64_t last_part_rd = INT64_MAX;
  int none_rate = INT_MAX;
  int64_t none_dist = INT64_MAX;
  int64_t none_rd = INT64_MAX;
  int chosen_rate = INT_MAX;
  int64_t chosen_dist = INT64_MAX;
  int64_t chosen_rd = INT64_MAX;
  BLOCK_SIZE sub_subsize = BLOCK_4X4;
  int splits_below = 0;
  BLOCK_SIZE bs_type = mi_8x8[0]->mbmi.sb_type;

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

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

  if (bsize < BLOCK_8X8) {
    // When ab_index = 0 all sub-blocks are handled, so for ab_index != 0
    // there is nothing to be done.
    if (x->ab_index != 0) {
      *rate = 0;
      *dist = 0;
      return;
    }
  } else {
    *(get_sb_partitioning(x, bsize)) = subsize;
  }
  save_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);

  if (bsize == BLOCK_16X16) {
    set_offsets(cpi, tile, mi_row, mi_col, bsize);
    x->mb_energy = vp9_block_energy(cpi, x, bsize);
  }

  if (cpi->sf.partition_search_type == SEARCH_PARTITION &&
      cpi->sf.adjust_partitioning_from_last_frame) {
    // Check if any of the sub blocks are further split.
    if (partition == PARTITION_SPLIT && subsize > BLOCK_8X8) {
      sub_subsize = get_subsize(subsize, PARTITION_SPLIT);
      splits_below = 1;
      for (i = 0; i < 4; i++) {
        int jj = i >> 1, ii = i & 0x01;
        MODE_INFO * this_mi = mi_8x8[jj * bss * mis + ii * bss];
        if (this_mi && this_mi->mbmi.sb_type >= sub_subsize) {
          splits_below = 0;
        }
      }
    }

    // If partition is not none try none unless each of the 4 splits are split
    // even further..
    if (partition != PARTITION_NONE && !splits_below &&
        mi_row + (ms >> 1) < cm->mi_rows &&
        mi_col + (ms >> 1) < cm->mi_cols) {
      *(get_sb_partitioning(x, bsize)) = bsize;
      rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &none_rate, &none_dist, bsize,
                       get_block_context(x, bsize), INT64_MAX);

      pl = partition_plane_context(cpi->above_seg_context,
                                   cpi->left_seg_context,
                                   mi_row, mi_col, bsize);

      if (none_rate < INT_MAX) {
        none_rate += x->partition_cost[pl][PARTITION_NONE];
        none_rd = RDCOST(x->rdmult, x->rddiv, none_rate, none_dist);
      }

      restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
      mi_8x8[0]->mbmi.sb_type = bs_type;
      *(get_sb_partitioning(x, bsize)) = subsize;
    }
  }

  switch (partition) {
    case PARTITION_NONE:
      rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &last_part_rate,
                       &last_part_dist, bsize,
                       get_block_context(x, bsize), INT64_MAX);
      break;
    case PARTITION_HORZ:
      *get_sb_index(x, subsize) = 0;
      rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &last_part_rate,
                       &last_part_dist, subsize,
                       get_block_context(x, subsize), INT64_MAX);
      if (last_part_rate != INT_MAX &&
          bsize >= BLOCK_8X8 && mi_row + (mh >> 1) < cm->mi_rows) {
        int rt = 0;
        int64_t dt = 0;
        update_state(cpi, get_block_context(x, subsize), subsize, 0);
        encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize);
        *get_sb_index(x, subsize) = 1;
        rd_pick_sb_modes(cpi, tile, mi_row + (ms >> 1), mi_col, &rt, &dt,
                         subsize, get_block_context(x, subsize), INT64_MAX);
        if (rt == INT_MAX || dt == INT64_MAX) {
          last_part_rate = INT_MAX;
          last_part_dist = INT64_MAX;
          break;
        }

        last_part_rate += rt;
        last_part_dist += dt;
      }
      break;
    case PARTITION_VERT:
      *get_sb_index(x, subsize) = 0;
      rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &last_part_rate,
                       &last_part_dist, subsize,
                       get_block_context(x, subsize), INT64_MAX);
      if (last_part_rate != INT_MAX &&
          bsize >= BLOCK_8X8 && mi_col + (ms >> 1) < cm->mi_cols) {
        int rt = 0;
        int64_t dt = 0;
        update_state(cpi, get_block_context(x, subsize), subsize, 0);
        encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize);
        *get_sb_index(x, subsize) = 1;
        rd_pick_sb_modes(cpi, tile, mi_row, mi_col + (ms >> 1), &rt, &dt,
                         subsize, get_block_context(x, subsize), INT64_MAX);
        if (rt == INT_MAX || dt == INT64_MAX) {
          last_part_rate = INT_MAX;
          last_part_dist = INT64_MAX;
          break;
        }
        last_part_rate += rt;
        last_part_dist += dt;
      }
      break;
    case PARTITION_SPLIT:
      // Split partition.
      last_part_rate = 0;
      last_part_dist = 0;
      for (i = 0; i < 4; i++) {
        int x_idx = (i & 1) * (ms >> 1);
        int y_idx = (i >> 1) * (ms >> 1);
        int jj = i >> 1, ii = i & 0x01;
        int rt;
        int64_t dt;

        if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
          continue;

        *get_sb_index(x, subsize) = i;

        rd_use_partition(cpi, tile, mi_8x8 + jj * bss * mis + ii * bss, tp,
                         mi_row + y_idx, mi_col + x_idx, subsize, &rt, &dt,
                         i != 3);
        if (rt == INT_MAX || dt == INT64_MAX) {
          last_part_rate = INT_MAX;
          last_part_dist = INT64_MAX;
          break;
        }
        last_part_rate += rt;
        last_part_dist += dt;
      }
      break;
    default:
      assert(0);
  }

  pl = partition_plane_context(cpi->above_seg_context, cpi->left_seg_context,
                               mi_row, mi_col, bsize);
  if (last_part_rate < INT_MAX) {
    last_part_rate += x->partition_cost[pl][partition];
    last_part_rd = RDCOST(x->rdmult, x->rddiv, last_part_rate, last_part_dist);
  }

  if (cpi->sf.adjust_partitioning_from_last_frame
      && cpi->sf.partition_search_type == SEARCH_PARTITION
      && partition != PARTITION_SPLIT && bsize > BLOCK_8X8
      && (mi_row + ms < cm->mi_rows || mi_row + (ms >> 1) == cm->mi_rows)
      && (mi_col + ms < cm->mi_cols || mi_col + (ms >> 1) == cm->mi_cols)) {
    BLOCK_SIZE split_subsize = get_subsize(bsize, PARTITION_SPLIT);
    chosen_rate = 0;
    chosen_dist = 0;
    restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);

    // Split partition.
    for (i = 0; i < 4; i++) {
      int x_idx = (i & 1) * (num_4x4_blocks_wide >> 2);
      int y_idx = (i >> 1) * (num_4x4_blocks_wide >> 2);
      int rt = 0;
      int64_t dt = 0;
      ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
      PARTITION_CONTEXT sl[8], sa[8];

      if ((mi_row + y_idx >= cm->mi_rows) || (mi_col + x_idx >= cm->mi_cols))
        continue;

      *get_sb_index(x, split_subsize) = i;
      *get_sb_partitioning(x, bsize) = split_subsize;
      *get_sb_partitioning(x, split_subsize) = split_subsize;

      save_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);

      rd_pick_sb_modes(cpi, tile, mi_row + y_idx, mi_col + x_idx, &rt, &dt,
                       split_subsize, get_block_context(x, split_subsize),
                       INT64_MAX);

      restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);

      if (rt == INT_MAX || dt == INT64_MAX) {
        chosen_rate = INT_MAX;
        chosen_dist = INT64_MAX;
        break;
      }

      chosen_rate += rt;
      chosen_dist += dt;

      if (i != 3)
        encode_sb(cpi, tile, tp,  mi_row + y_idx, mi_col + x_idx, 0,
                  split_subsize);

      pl = partition_plane_context(cpi->above_seg_context,
                                   cpi->left_seg_context,
                                   mi_row + y_idx, mi_col + x_idx,
                                   split_subsize);
      chosen_rate += x->partition_cost[pl][PARTITION_NONE];
    }
    pl = partition_plane_context(cpi->above_seg_context, cpi->left_seg_context,
                                 mi_row, mi_col, bsize);
    if (chosen_rate < INT_MAX) {
      chosen_rate += x->partition_cost[pl][PARTITION_SPLIT];
      chosen_rd = RDCOST(x->rdmult, x->rddiv, chosen_rate, chosen_dist);
    }
  }

  // If last_part is better set the partitioning to that...
  if (last_part_rd < chosen_rd) {
    mi_8x8[0]->mbmi.sb_type = bsize;
    if (bsize >= BLOCK_8X8)
      *(get_sb_partitioning(x, bsize)) = subsize;
    chosen_rate = last_part_rate;
    chosen_dist = last_part_dist;
    chosen_rd = last_part_rd;
  }
  // If none was better set the partitioning to that...
  if (none_rd < chosen_rd) {
    if (bsize >= BLOCK_8X8)
      *(get_sb_partitioning(x, bsize)) = bsize;
    chosen_rate = none_rate;
    chosen_dist = none_dist;
  }

  restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);

  // We must have chosen a partitioning and encoding or we'll fail later on.
  // No other opportunities for success.
  if ( bsize == BLOCK_64X64)
    assert(chosen_rate < INT_MAX && chosen_dist < INT64_MAX);

  if (do_recon) {
    int output_enabled = (bsize == BLOCK_64X64);

    // Check the projected output rate for this SB against it's target
    // and and if necessary apply a Q delta using segmentation to get
    // closer to the target.
    if ((cpi->oxcf.aq_mode == COMPLEXITY_AQ) && cm->seg.update_map) {
      select_in_frame_q_segment(cpi, mi_row, mi_col,
                                output_enabled, chosen_rate);
    }

    encode_sb(cpi, tile, tp, mi_row, mi_col, output_enabled, bsize);
  }

  *rate = chosen_rate;
  *dist = chosen_dist;
}

static const BLOCK_SIZE min_partition_size[BLOCK_SIZES] = {
  BLOCK_4X4,   BLOCK_4X4,   BLOCK_4X4,
  BLOCK_4X4,   BLOCK_4X4,   BLOCK_4X4,
  BLOCK_8X8,   BLOCK_8X8,   BLOCK_8X8,
  BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,
  BLOCK_16X16
};

static const BLOCK_SIZE max_partition_size[BLOCK_SIZES] = {
  BLOCK_8X8,   BLOCK_16X16, BLOCK_16X16,
  BLOCK_16X16, BLOCK_32X32, BLOCK_32X32,
  BLOCK_32X32, BLOCK_64X64, BLOCK_64X64,
  BLOCK_64X64, BLOCK_64X64, BLOCK_64X64,
  BLOCK_64X64
};

// Look at all the mode_info entries for blocks that are part of this
// partition and find the min and max values for sb_type.
// At the moment this is designed to work on a 64x64 SB but could be
// adjusted to use a size parameter.
//
// The min and max are assumed to have been initialized prior to calling this
// function so repeat calls can accumulate a min and max of more than one sb64.
static void get_sb_partition_size_range(VP9_COMP *cpi, MODE_INFO ** mi_8x8,
                                        BLOCK_SIZE * min_block_size,
                                        BLOCK_SIZE * max_block_size ) {
  MACROBLOCKD *const xd = &cpi->mb.e_mbd;
  int sb_width_in_blocks = MI_BLOCK_SIZE;
  int sb_height_in_blocks  = MI_BLOCK_SIZE;
  int i, j;
  int index = 0;

  // Check the sb_type for each block that belongs to this region.
  for (i = 0; i < sb_height_in_blocks; ++i) {
    for (j = 0; j < sb_width_in_blocks; ++j) {
      MODE_INFO * mi = mi_8x8[index+j];
      BLOCK_SIZE sb_type = mi ? mi->mbmi.sb_type : 0;
      *min_block_size = MIN(*min_block_size, sb_type);
      *max_block_size = MAX(*max_block_size, sb_type);
    }
    index += xd->mode_info_stride;
  }
}

// Next square block size less or equal than current block size.
static const BLOCK_SIZE next_square_size[BLOCK_SIZES] = {
  BLOCK_4X4, BLOCK_4X4, BLOCK_4X4,
  BLOCK_8X8, BLOCK_8X8, BLOCK_8X8,
  BLOCK_16X16, BLOCK_16X16, BLOCK_16X16,
  BLOCK_32X32, BLOCK_32X32, BLOCK_32X32,
  BLOCK_64X64
};

// Look at neighboring blocks and set a min and max partition size based on
// what they chose.
static void rd_auto_partition_range(VP9_COMP *cpi, const TileInfo *const tile,
                                    int row, int col,
                                    BLOCK_SIZE *min_block_size,
                                    BLOCK_SIZE *max_block_size) {
  VP9_COMMON * const cm = &cpi->common;
  MACROBLOCKD *const xd = &cpi->mb.e_mbd;
  MODE_INFO ** mi_8x8 = xd->mi_8x8;
  MODE_INFO ** prev_mi_8x8 = xd->prev_mi_8x8;

  const int left_in_image = xd->left_available && mi_8x8[-1];
  const int above_in_image = xd->up_available &&
                             mi_8x8[-xd->mode_info_stride];
  MODE_INFO ** above_sb64_mi_8x8;
  MODE_INFO ** left_sb64_mi_8x8;

  int row8x8_remaining = tile->mi_row_end - row;
  int col8x8_remaining = tile->mi_col_end - col;
  int bh, bw;

  // Trap case where we do not have a prediction.
  if (!left_in_image && !above_in_image &&
      ((cm->frame_type == KEY_FRAME) || !cm->prev_mi)) {
    *min_block_size = BLOCK_4X4;
    *max_block_size = BLOCK_64X64;
  } else {
    // Default "min to max" and "max to min"
    *min_block_size = BLOCK_64X64;
    *max_block_size = BLOCK_4X4;

    // NOTE: each call to get_sb_partition_size_range() uses the previous
    // passed in values for min and max as a starting point.
    //
    // Find the min and max partition used in previous frame at this location
    if (cm->prev_mi && (cm->frame_type != KEY_FRAME)) {
      get_sb_partition_size_range(cpi, prev_mi_8x8,
                                  min_block_size, max_block_size);
    }

    // Find the min and max partition sizes used in the left SB64
    if (left_in_image) {
      left_sb64_mi_8x8 = &mi_8x8[-MI_BLOCK_SIZE];
      get_sb_partition_size_range(cpi, left_sb64_mi_8x8,
                                  min_block_size, max_block_size);
    }

    // Find the min and max partition sizes used in the above SB64.
    if (above_in_image) {
      above_sb64_mi_8x8 = &mi_8x8[-xd->mode_info_stride * MI_BLOCK_SIZE];
      get_sb_partition_size_range(cpi, above_sb64_mi_8x8,
                                  min_block_size, max_block_size);
    }
  }

  // adjust observed min and max
  if (cpi->sf.auto_min_max_partition_size == RELAXED_NEIGHBORING_MIN_MAX) {
    *min_block_size = min_partition_size[*min_block_size];
    *max_block_size = max_partition_size[*max_block_size];
  }

  // Check border cases where max and min from neighbours may not be legal.
  *max_block_size = find_partition_size(*max_block_size,
                                        row8x8_remaining, col8x8_remaining,
                                        &bh, &bw);
  *min_block_size = MIN(*min_block_size, *max_block_size);

  // When use_square_partition_only is true, make sure at least one square
  // partition is allowed by selecting the next smaller square size as
  // *min_block_size.
  if (cpi->sf.use_square_partition_only &&
      (*max_block_size - *min_block_size) < 2) {
    *min_block_size = next_square_size[*min_block_size];
  }
}

static INLINE void store_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
  vpx_memcpy(ctx->pred_mv, x->pred_mv, sizeof(x->pred_mv));
}

static INLINE void load_pred_mv(MACROBLOCK *x, PICK_MODE_CONTEXT *ctx) {
  vpx_memcpy(x->pred_mv, ctx->pred_mv, sizeof(x->pred_mv));
}

// TODO(jingning,jimbankoski,rbultje): properly skip partition types that are
// unlikely to be selected depending on previous rate-distortion optimization
// results, for encoding speed-up.
static void rd_pick_partition(VP9_COMP *cpi, const TileInfo *const tile,
                              TOKENEXTRA **tp, int mi_row,
                              int mi_col, BLOCK_SIZE bsize, int *rate,
                              int64_t *dist, int do_recon, int64_t best_rd) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  const int ms = num_8x8_blocks_wide_lookup[bsize] / 2;
  ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
  PARTITION_CONTEXT sl[8], sa[8];
  TOKENEXTRA *tp_orig = *tp;
  int i, pl;
  BLOCK_SIZE subsize;
  int this_rate, sum_rate = 0, best_rate = INT_MAX;
  int64_t this_dist, sum_dist = 0, best_dist = INT64_MAX;
  int64_t sum_rd = 0;
  int do_split = bsize >= BLOCK_8X8;
  int do_rect = 1;
  // Override skipping rectangular partition operations for edge blocks
  const int force_horz_split = (mi_row + ms >= cm->mi_rows);
  const int force_vert_split = (mi_col + ms >= cm->mi_cols);
  const int xss = x->e_mbd.plane[1].subsampling_x;
  const int yss = x->e_mbd.plane[1].subsampling_y;

  int partition_none_allowed = !force_horz_split && !force_vert_split;
  int partition_horz_allowed = !force_vert_split && yss <= xss &&
                               bsize >= BLOCK_8X8;
  int partition_vert_allowed = !force_horz_split && xss <= yss &&
                               bsize >= BLOCK_8X8;
  (void) *tp_orig;

  if (bsize < BLOCK_8X8) {
    // When ab_index = 0 all sub-blocks are handled, so for ab_index != 0
    // there is nothing to be done.
    if (x->ab_index != 0) {
      *rate = 0;
      *dist = 0;
      return;
    }
  }
  assert(num_8x8_blocks_wide_lookup[bsize] ==
             num_8x8_blocks_high_lookup[bsize]);

  if (bsize == BLOCK_16X16) {
    set_offsets(cpi, tile, mi_row, mi_col, bsize);
    x->mb_energy = vp9_block_energy(cpi, x, bsize);
  }

  // Determine partition types in search according to the speed features.
  // The threshold set here has to be of square block size.
  if (cpi->sf.auto_min_max_partition_size) {
    partition_none_allowed &= (bsize <= cpi->sf.max_partition_size &&
                               bsize >= cpi->sf.min_partition_size);
    partition_horz_allowed &= ((bsize <= cpi->sf.max_partition_size &&
                                bsize >  cpi->sf.min_partition_size) ||
                                force_horz_split);
    partition_vert_allowed &= ((bsize <= cpi->sf.max_partition_size &&
                                bsize >  cpi->sf.min_partition_size) ||
                                force_vert_split);
    do_split &= bsize > cpi->sf.min_partition_size;
  }
  if (cpi->sf.use_square_partition_only) {
    partition_horz_allowed &= force_horz_split;
    partition_vert_allowed &= force_vert_split;
  }

  save_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);

  if (cpi->sf.disable_split_var_thresh && partition_none_allowed) {
    unsigned int source_variancey;
    vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);
    source_variancey = get_sby_perpixel_variance(cpi, x, bsize);
    if (source_variancey < cpi->sf.disable_split_var_thresh) {
      do_split = 0;
      if (source_variancey < cpi->sf.disable_split_var_thresh / 2)
        do_rect = 0;
    }
  }

  // PARTITION_NONE
  if (partition_none_allowed) {
    rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &this_rate, &this_dist, bsize,
                     get_block_context(x, bsize), best_rd);
    if (this_rate != INT_MAX) {
      if (bsize >= BLOCK_8X8) {
        pl = partition_plane_context(cpi->above_seg_context,
                                     cpi->left_seg_context,
                                     mi_row, mi_col, bsize);
        this_rate += x->partition_cost[pl][PARTITION_NONE];
      }
      sum_rd = RDCOST(x->rdmult, x->rddiv, this_rate, this_dist);
      if (sum_rd < best_rd) {
        int64_t stop_thresh = 4096;
        int64_t stop_thresh_rd;

        best_rate = this_rate;
        best_dist = this_dist;
        best_rd = sum_rd;
        if (bsize >= BLOCK_8X8)
          *(get_sb_partitioning(x, bsize)) = bsize;

        // Adjust threshold according to partition size.
        stop_thresh >>= 8 - (b_width_log2_lookup[bsize] +
            b_height_log2_lookup[bsize]);

        stop_thresh_rd = RDCOST(x->rdmult, x->rddiv, 0, stop_thresh);
        // If obtained distortion is very small, choose current partition
        // and stop splitting.
        if (!x->e_mbd.lossless && best_rd < stop_thresh_rd) {
          do_split = 0;
          do_rect = 0;
        }
      }
    }
    restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
  }

  // store estimated motion vector
  if (cpi->sf.adaptive_motion_search)
    store_pred_mv(x, get_block_context(x, bsize));

  // PARTITION_SPLIT
  sum_rd = 0;
  // TODO(jingning): use the motion vectors given by the above search as
  // the starting point of motion search in the following partition type check.
  if (do_split) {
    subsize = get_subsize(bsize, PARTITION_SPLIT);
    for (i = 0; i < 4 && sum_rd < best_rd; ++i) {
      const int x_idx = (i & 1) * ms;
      const int y_idx = (i >> 1) * ms;

      if (mi_row + y_idx >= cm->mi_rows || mi_col + x_idx >= cm->mi_cols)
        continue;

      *get_sb_index(x, subsize) = i;
      if (cpi->sf.adaptive_motion_search)
        load_pred_mv(x, get_block_context(x, bsize));
      if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
          partition_none_allowed)
        get_block_context(x, subsize)->pred_interp_filter =
            get_block_context(x, bsize)->mic.mbmi.interp_filter;
      rd_pick_partition(cpi, tile, tp, mi_row + y_idx, mi_col + x_idx, subsize,
                        &this_rate, &this_dist, i != 3, best_rd - sum_rd);

      if (this_rate == INT_MAX) {
        sum_rd = INT64_MAX;
      } else {
        sum_rate += this_rate;
        sum_dist += this_dist;
        sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
      }
    }
    if (sum_rd < best_rd && i == 4) {
      pl = partition_plane_context(cpi->above_seg_context,
                                   cpi->left_seg_context,
                                   mi_row, mi_col, bsize);
      sum_rate += x->partition_cost[pl][PARTITION_SPLIT];
      sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
      if (sum_rd < best_rd) {
        best_rate = sum_rate;
        best_dist = sum_dist;
        best_rd = sum_rd;
        *(get_sb_partitioning(x, bsize)) = subsize;
      }
    } else {
      // skip rectangular partition test when larger block size
      // gives better rd cost
      if (cpi->sf.less_rectangular_check)
        do_rect &= !partition_none_allowed;
    }
    restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
  }

  // PARTITION_HORZ
  if (partition_horz_allowed && do_rect) {
    subsize = get_subsize(bsize, PARTITION_HORZ);
    *get_sb_index(x, subsize) = 0;
    if (cpi->sf.adaptive_motion_search)
      load_pred_mv(x, get_block_context(x, bsize));
    if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
        partition_none_allowed)
      get_block_context(x, subsize)->pred_interp_filter =
          get_block_context(x, bsize)->mic.mbmi.interp_filter;
    rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &sum_rate, &sum_dist, subsize,
                     get_block_context(x, subsize), best_rd);
    sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);

    if (sum_rd < best_rd && mi_row + ms < cm->mi_rows) {
      update_state(cpi, get_block_context(x, subsize), subsize, 0);
      encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize);

      *get_sb_index(x, subsize) = 1;
      if (cpi->sf.adaptive_motion_search)
        load_pred_mv(x, get_block_context(x, bsize));
      if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
          partition_none_allowed)
        get_block_context(x, subsize)->pred_interp_filter =
            get_block_context(x, bsize)->mic.mbmi.interp_filter;
      rd_pick_sb_modes(cpi, tile, mi_row + ms, mi_col, &this_rate,
                       &this_dist, subsize, get_block_context(x, subsize),
                       best_rd - sum_rd);
      if (this_rate == INT_MAX) {
        sum_rd = INT64_MAX;
      } else {
        sum_rate += this_rate;
        sum_dist += this_dist;
        sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
      }
    }
    if (sum_rd < best_rd) {
      pl = partition_plane_context(cpi->above_seg_context,
                                   cpi->left_seg_context,
                                   mi_row, mi_col, bsize);
      sum_rate += x->partition_cost[pl][PARTITION_HORZ];
      sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
      if (sum_rd < best_rd) {
        best_rd = sum_rd;
        best_rate = sum_rate;
        best_dist = sum_dist;
        *(get_sb_partitioning(x, bsize)) = subsize;
      }
    }
    restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
  }

  // PARTITION_VERT
  if (partition_vert_allowed && do_rect) {
    subsize = get_subsize(bsize, PARTITION_VERT);

    *get_sb_index(x, subsize) = 0;
    if (cpi->sf.adaptive_motion_search)
      load_pred_mv(x, get_block_context(x, bsize));
    if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
        partition_none_allowed)
      get_block_context(x, subsize)->pred_interp_filter =
          get_block_context(x, bsize)->mic.mbmi.interp_filter;
    rd_pick_sb_modes(cpi, tile, mi_row, mi_col, &sum_rate, &sum_dist, subsize,
                     get_block_context(x, subsize), best_rd);
    sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
    if (sum_rd < best_rd && mi_col + ms < cm->mi_cols) {
      update_state(cpi, get_block_context(x, subsize), subsize, 0);
      encode_superblock(cpi, tp, 0, mi_row, mi_col, subsize);

      *get_sb_index(x, subsize) = 1;
      if (cpi->sf.adaptive_motion_search)
        load_pred_mv(x, get_block_context(x, bsize));
      if (cpi->sf.adaptive_pred_interp_filter && bsize == BLOCK_8X8 &&
          partition_none_allowed)
        get_block_context(x, subsize)->pred_interp_filter =
            get_block_context(x, bsize)->mic.mbmi.interp_filter;
      rd_pick_sb_modes(cpi, tile, mi_row, mi_col + ms, &this_rate,
                       &this_dist, subsize, get_block_context(x, subsize),
                       best_rd - sum_rd);
      if (this_rate == INT_MAX) {
        sum_rd = INT64_MAX;
      } else {
        sum_rate += this_rate;
        sum_dist += this_dist;
        sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
      }
    }
    if (sum_rd < best_rd) {
      pl = partition_plane_context(cpi->above_seg_context,
                                   cpi->left_seg_context,
                                   mi_row, mi_col, bsize);
      sum_rate += x->partition_cost[pl][PARTITION_VERT];
      sum_rd = RDCOST(x->rdmult, x->rddiv, sum_rate, sum_dist);
      if (sum_rd < best_rd) {
        best_rate = sum_rate;
        best_dist = sum_dist;
        best_rd = sum_rd;
        *(get_sb_partitioning(x, bsize)) = subsize;
      }
    }
    restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
  }

  // TODO(jbb): This code added so that we avoid static analysis
  // warning related to the fact that best_rd isn't used after this
  // point.  This code should be refactored so that the duplicate
  // checks occur in some sub function and thus are used...
  (void) best_rd;
  *rate = best_rate;
  *dist = best_dist;

  if (best_rate < INT_MAX && best_dist < INT64_MAX && do_recon) {
    int output_enabled = (bsize == BLOCK_64X64);

    // Check the projected output rate for this SB against it's target
    // and and if necessary apply a Q delta using segmentation to get
    // closer to the target.
    if ((cpi->oxcf.aq_mode == COMPLEXITY_AQ) && cm->seg.update_map) {
      select_in_frame_q_segment(cpi, mi_row, mi_col, output_enabled, best_rate);
    }
    encode_sb(cpi, tile, tp, mi_row, mi_col, output_enabled, bsize);
  }
  if (bsize == BLOCK_64X64) {
    assert(tp_orig < *tp);
    assert(best_rate < INT_MAX);
    assert(best_dist < INT64_MAX);
  } else {
    assert(tp_orig == *tp);
  }
}

static void encode_rd_sb_row(VP9_COMP *cpi, const TileInfo *const tile,
                             int mi_row, TOKENEXTRA **tp) {
  VP9_COMMON *const cm = &cpi->common;
  int mi_col;

  // Initialize the left context for the new SB row
  vpx_memset(&cpi->left_context, 0, sizeof(cpi->left_context));
  vpx_memset(cpi->left_seg_context, 0, sizeof(cpi->left_seg_context));

  // Code each SB in the row
  for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
       mi_col += MI_BLOCK_SIZE) {
    int dummy_rate;
    int64_t dummy_dist;

    BLOCK_SIZE i;
    MACROBLOCK *x = &cpi->mb;

    if (cpi->sf.adaptive_pred_interp_filter) {
      for (i = BLOCK_4X4; i < BLOCK_8X8; ++i) {
        const int num_4x4_w = num_4x4_blocks_wide_lookup[i];
        const int num_4x4_h = num_4x4_blocks_high_lookup[i];
        const int num_4x4_blk = MAX(4, num_4x4_w * num_4x4_h);
        for (x->sb_index = 0; x->sb_index < 4; ++x->sb_index)