vp9_encodeframe.c

  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 (xd->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);
  }

  x->fast_ms = 0;
  x->subblock_ref = 0;

  if (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;
      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);
      none_rate += x->partition_cost[pl][PARTITION_NONE];

      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:
      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(xd, subsize) = 0;
      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(xd, subsize) = 1;
        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 == INT_MAX) {
          last_part_rate = INT_MAX;
          last_part_dist = INT_MAX;
          break;
        }

        last_part_rate += rt;
        last_part_dist += dt;
      }
      break;
    case PARTITION_VERT:
      *get_sb_index(xd, subsize) = 0;
      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(xd, subsize) = 1;
        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 == INT_MAX) {
          last_part_rate = INT_MAX;
          last_part_dist = INT_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(xd, 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 == INT_MAX) {
          last_part_rate = INT_MAX;
          last_part_dist = INT_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];

  if (cpi->sf.adjust_partitioning_from_last_frame
      && 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);
    split_rate = 0;
    split_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(xd, 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);

      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 == INT_MAX) {
        split_rate = INT_MAX;
        split_dist = INT_MAX;
        break;
      }

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

      split_rate += rt;
      split_dist += dt;
      pl = partition_plane_context(cpi->above_seg_context,
                                   cpi->left_seg_context,
                                   mi_row + y_idx, mi_col + x_idx, bsize);
      split_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 (split_rate < INT_MAX) {
      split_rate += x->partition_cost[pl][PARTITION_SPLIT];

      chosen_rate = split_rate;
      chosen_dist = split_dist;
    }
  }

  // If last_part is better set the partitioning to that...
  if (RDCOST(x->rdmult, x->rddiv, last_part_rate, last_part_dist)
      < RDCOST(x->rdmult, x->rddiv, chosen_rate, chosen_dist)) {
    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;
  }
  // If none was better set the partitioning to that...
  if (RDCOST(x->rdmult, x->rddiv, chosen_rate, chosen_dist)
      > RDCOST(x->rdmult, x->rddiv, none_rate, none_dist)) {
    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 < INT_MAX);

  if (do_recon)
    encode_sb(cpi, tile, tp, mi_row, mi_col, bsize == BLOCK_64X64, 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;
  }
}

// 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);
    }
  }

  // Give a bit of leaway either side of the observed min and 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);
}

static void compute_fast_motion_search_level(VP9_COMP *cpi, BLOCK_SIZE bsize) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;

  // Only use 8x8 result for non HD videos.
  // int use_8x8 = (MIN(cpi->common.width, cpi->common.height) < 720) ? 1 : 0;
  int use_8x8 = 1;

  if (cm->frame_type && !cpi->is_src_frame_alt_ref &&
      ((use_8x8 && bsize == BLOCK_16X16) ||
      bsize == BLOCK_32X32 || bsize == BLOCK_64X64)) {
    int ref0 = 0, ref1 = 0, ref2 = 0, ref3 = 0;
    PICK_MODE_CONTEXT *block_context = NULL;

    if (bsize == BLOCK_16X16) {
      block_context = x->sb8x8_context[xd->sb_index][xd->mb_index];
    } else if (bsize == BLOCK_32X32) {
      block_context = x->mb_context[xd->sb_index];
    } else if (bsize == BLOCK_64X64) {
      block_context = x->sb32_context;
    }

    if (block_context) {
      ref0 = block_context[0].mic.mbmi.ref_frame[0];
      ref1 = block_context[1].mic.mbmi.ref_frame[0];
      ref2 = block_context[2].mic.mbmi.ref_frame[0];
      ref3 = block_context[3].mic.mbmi.ref_frame[0];
    }

    // Currently, only consider 4 inter reference frames.
    if (ref0 && ref1 && ref2 && ref3) {
      int d01, d23, d02, d13;

      // Motion vectors for the four subblocks.
      int16_t mvr0 = block_context[0].mic.mbmi.mv[0].as_mv.row;
      int16_t mvc0 = block_context[0].mic.mbmi.mv[0].as_mv.col;
      int16_t mvr1 = block_context[1].mic.mbmi.mv[0].as_mv.row;
      int16_t mvc1 = block_context[1].mic.mbmi.mv[0].as_mv.col;
      int16_t mvr2 = block_context[2].mic.mbmi.mv[0].as_mv.row;
      int16_t mvc2 = block_context[2].mic.mbmi.mv[0].as_mv.col;
      int16_t mvr3 = block_context[3].mic.mbmi.mv[0].as_mv.row;
      int16_t mvc3 = block_context[3].mic.mbmi.mv[0].as_mv.col;

      // Adjust sign if ref is alt_ref.
      if (cm->ref_frame_sign_bias[ref0]) {
        mvr0 *= -1;
        mvc0 *= -1;
      }

      if (cm->ref_frame_sign_bias[ref1]) {
        mvr1 *= -1;
        mvc1 *= -1;
      }

      if (cm->ref_frame_sign_bias[ref2]) {
        mvr2 *= -1;
        mvc2 *= -1;
      }

      if (cm->ref_frame_sign_bias[ref3]) {
        mvr3 *= -1;
        mvc3 *= -1;
      }

      // Calculate mv distances.
      d01 = MAX(abs(mvr0 - mvr1), abs(mvc0 - mvc1));
      d23 = MAX(abs(mvr2 - mvr3), abs(mvc2 - mvc3));
      d02 = MAX(abs(mvr0 - mvr2), abs(mvc0 - mvc2));
      d13 = MAX(abs(mvr1 - mvr3), abs(mvc1 - mvc3));

      if (d01 < FAST_MOTION_MV_THRESH && d23 < FAST_MOTION_MV_THRESH &&
          d02 < FAST_MOTION_MV_THRESH && d13 < FAST_MOTION_MV_THRESH) {
        // Set fast motion search level.
        x->fast_ms = 1;

        if (ref0 == ref1 && ref1 == ref2 && ref2 == ref3 &&
            d01 < 2 && d23 < 2 && d02 < 2 && d13 < 2) {
          // Set fast motion search level.
          x->fast_ms = 2;

          if (!d01 && !d23 && !d02 && !d13) {
            x->fast_ms = 3;
            x->subblock_ref = ref0;
          }
        }
      }
    }
  }
}

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;
  MACROBLOCKD * const xd = &x->e_mbd;
  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);

  int partition_none_allowed = !force_horz_split && !force_vert_split;
  int partition_horz_allowed = !force_vert_split && bsize >= BLOCK_8X8;
  int partition_vert_allowed = !force_horz_split && bsize >= BLOCK_8X8;

  int partition_split_done = 0;
  (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 (xd->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) {
    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 = 2048;

        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]);

        // If obtained distortion is very small, choose current partition
        // and stop splitting.
        if (this_dist < stop_thresh) {
          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(xd, subsize) = i;
      if (cpi->sf.adaptive_motion_search)
        load_pred_mv(x, get_block_context(x, bsize));
      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;
    }
    partition_split_done = 1;
    restore_context(cpi, mi_row, mi_col, a, l, sa, sl, bsize);
  }

  x->fast_ms = 0;
  x->subblock_ref = 0;

  if (partition_split_done &&
      cpi->sf.using_small_partition_info) {
    compute_fast_motion_search_level(cpi, bsize);
  }

  // PARTITION_HORZ
  if (partition_horz_allowed && do_rect) {
    subsize = get_subsize(bsize, PARTITION_HORZ);
    *get_sb_index(xd, subsize) = 0;
    if (cpi->sf.adaptive_motion_search)
      load_pred_mv(x, get_block_context(x, bsize));
    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(xd, subsize) = 1;
      if (cpi->sf.adaptive_motion_search)
        load_pred_mv(x, get_block_context(x, bsize));
      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(xd, subsize) = 0;
    if (cpi->sf.adaptive_motion_search)
      load_pred_mv(x, get_block_context(x, bsize));
    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(xd, subsize) = 1;
      if (cpi->sf.adaptive_motion_search)
        load_pred_mv(x, get_block_context(x, bsize));
      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);
  }


  *rate = best_rate;
  *dist = best_dist;

  if (best_rate < INT_MAX && best_dist < INT64_MAX && do_recon)
    encode_sb(cpi, tile, tp, mi_row, mi_col, bsize == BLOCK_64X64, bsize);
  if (bsize == BLOCK_64X64) {
    assert(tp_orig < *tp);
    assert(best_rate < INT_MAX);
    assert(best_dist < INT_MAX);
  } else {
    assert(tp_orig == *tp);
  }
}

// Examines 64x64 block and chooses a best reference frame
static void rd_pick_reference_frame(VP9_COMP *cpi, const TileInfo *const tile,
                                    int mi_row, int mi_col) {
  VP9_COMMON * const cm = &cpi->common;
  MACROBLOCK * const x = &cpi->mb;
  int bsl = b_width_log2(BLOCK_64X64), bs = 1 << bsl;
  int ms = bs / 2;
  ENTROPY_CONTEXT l[16 * MAX_MB_PLANE], a[16 * MAX_MB_PLANE];
  PARTITION_CONTEXT sl[8], sa[8];
  int pl;
  int r;
  int64_t d;

  save_context(cpi, mi_row, mi_col, a, l, sa, sl, BLOCK_64X64);

  // Default is non mask (all reference frames allowed.
  cpi->ref_frame_mask = 0;

  // Do RD search for 64x64.
  if ((mi_row + (ms >> 1) < cm->mi_rows) &&
      (mi_col + (ms >> 1) < cm->mi_cols)) {
    cpi->set_ref_frame_mask = 1;
    pick_sb_modes(cpi, tile, mi_row, mi_col, &r, &d, BLOCK_64X64,
                  get_block_context(x, BLOCK_64X64), INT64_MAX);
    pl = partition_plane_context(cpi->above_seg_context, cpi->left_seg_context,
                                 mi_row, mi_col, BLOCK_64X64);
    r += x->partition_cost[pl][PARTITION_NONE];

    *(get_sb_partitioning(x, BLOCK_64X64)) = BLOCK_64X64;
    cpi->set_ref_frame_mask = 0;
  }

  restore_context(cpi, mi_row, mi_col, a, l, sa, sl, BLOCK_64X64);
}

static void encode_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;

    vp9_zero(cpi->mb.pred_mv);

    if (cpi->sf.reference_masking)
      rd_pick_reference_frame(cpi, tile, mi_row, mi_col);

    if (cpi->sf.use_lastframe_partitioning ||
        cpi->sf.use_one_partition_size_always ) {
      const int idx_str = cm->mode_info_stride * mi_row + mi_col;
      MODE_INFO **mi_8x8 = cm->mi_grid_visible + idx_str;
      MODE_INFO **prev_mi_8x8 = cm->prev_mi_grid_visible + idx_str;

      cpi->mb.source_variance = UINT_MAX;
      if (cpi->sf.use_one_partition_size_always) {
        set_offsets(cpi, tile, mi_row, mi_col, BLOCK_64X64);
        set_partitioning(cpi, tile, mi_8x8, mi_row, mi_col);
        rd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64,
                         &dummy_rate, &dummy_dist, 1);
      } else {
        if ((cpi->common.current_video_frame
            % cpi->sf.last_partitioning_redo_frequency) == 0
            || cm->prev_mi == 0
            || cpi->common.show_frame == 0
            || cpi->common.frame_type == KEY_FRAME
            || cpi->is_src_frame_alt_ref
            || ((cpi->sf.use_lastframe_partitioning ==
                 LAST_FRAME_PARTITION_LOW_MOTION) &&
                 sb_has_motion(cpi, prev_mi_8x8))) {
          // If required set upper and lower partition size limits
          if (cpi->sf.auto_min_max_partition_size) {
            set_offsets(cpi, tile, mi_row, mi_col, BLOCK_64X64);
            rd_auto_partition_range(cpi, tile, mi_row, mi_col,
                                    &cpi->sf.min_partition_size,
                                    &cpi->sf.max_partition_size);
          }
          rd_pick_partition(cpi, tile, tp, mi_row, mi_col, BLOCK_64X64,
                            &dummy_rate, &dummy_dist, 1, INT64_MAX);
        } else {
          copy_partitioning(cpi, mi_8x8, prev_mi_8x8);
          rd_use_partition(cpi, tile, mi_8x8, tp, mi_row, mi_col, BLOCK_64X64,
                           &dummy_rate, &dummy_dist, 1);
        }
      }
    } else {
      // If required set upper and lower partition size limits
      if (cpi->sf.auto_min_max_partition_size) {
        set_offsets(cpi, tile, mi_row, mi_col, BLOCK_64X64);
        rd_auto_partition_range(cpi, tile, mi_row, mi_col,
                                &cpi->sf.min_partition_size,
                                &cpi->sf.max_partition_size);
      }
      rd_pick_partition(cpi, tile, tp, mi_row, mi_col, BLOCK_64X64,
                        &dummy_rate, &dummy_dist, 1, INT64_MAX);
    }
  }
}

static void init_encode_frame_mb_context(VP9_COMP *cpi) {
  MACROBLOCK *const x = &cpi->mb;
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCKD *const xd = &x->e_mbd;
  const int aligned_mi_cols = mi_cols_aligned_to_sb(cm->mi_cols);

  x->act_zbin_adj = 0;
  cpi->seg0_idx = 0;

  xd->mode_info_stride = cm->mode_info_stride;

  // reset intra mode contexts
  if (frame_is_intra_only(cm))
    vp9_init_mbmode_probs(cm);

  // Copy data over into macro block data structures.
  vp9_setup_src_planes(x, cpi->Source, 0, 0);

  // TODO(jkoleszar): are these initializations required?
  setup_pre_planes(xd, 0, &cm->yv12_fb[cm->ref_frame_map[cpi->lst_fb_idx]],
                   0, 0, NULL);
  setup_dst_planes(xd, get_frame_new_buffer(cm), 0, 0);

  setup_block_dptrs(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);

  xd->mi_8x8[0]->mbmi.mode = DC_PRED;
  xd->mi_8x8[0]->mbmi.uv_mode = DC_PRED;

  vp9_zero(cpi->y_mode_count);
  vp9_zero(cpi->y_uv_mode_count);
  vp9_zero(cm->counts.inter_mode);
  vp9_zero(cpi->partition_count);
  vp9_zero(cpi->intra_inter_count);
  vp9_zero(cpi->comp_inter_count);
  vp9_zero(cpi->single_ref_count);
  vp9_zero(cpi->comp_ref_count);
  vp9_zero(cm->counts.tx);
  vp9_zero(cm->counts.mbskip);

  // Note: this memset assumes above_context[0], [1] and [2]
  // are allocated as part of the same buffer.
  vpx_memset(cpi->above_context[0], 0,
             sizeof(*cpi->above_context[0]) *
             2 * aligned_mi_cols * MAX_MB_PLANE);
  vpx_memset(cpi->above_seg_context, 0,
             sizeof(*cpi->above_seg_context) * aligned_mi_cols);
}

static void switch_lossless_mode(VP9_COMP *cpi, int lossless) {
  if (lossless) {
    // printf("Switching to lossless\n");
    cpi->mb.fwd_txm4x4 = vp9_fwht4x4;
    cpi->mb.e_mbd.itxm_add = vp9_iwht4x4_add;
    cpi->mb.optimize = 0;
    cpi->common.lf.filter_level = 0;
    cpi->zbin_mode_boost_enabled = 0;
    cpi->common.tx_mode = ONLY_4X4;
  } else {
    // printf("Not lossless\n");
    cpi->mb.fwd_txm4x4 = vp9_fdct4x4;
    cpi->mb.e_mbd.itxm_add = vp9_idct4x4_add;
  }
}

static void switch_tx_mode(VP9_COMP *cpi) {
  if (cpi->sf.tx_size_search_method == USE_LARGESTALL &&
      cpi->common.tx_mode >= ALLOW_32X32)
    cpi->common.tx_mode = ALLOW_32X32;
}

static void encode_frame_internal(VP9_COMP *cpi) {
  int mi_row;
  MACROBLOCK * const x = &cpi->mb;
  VP9_COMMON * const cm = &cpi->common;
  MACROBLOCKD * const xd = &x->e_mbd;

//  fprintf(stderr, "encode_frame_internal frame %d (%d) type %d\n",
//           cpi->common.current_video_frame, cpi->common.show_frame,
//           cm->frame_type);

// debug output
#if DBG_PRNT_SEGMAP
  {
    FILE *statsfile;
    statsfile = fopen("segmap2.stt", "a");
    fprintf(statsfile, "\n");
    fclose(statsfile);
  }
#endif

  vp9_zero(cm->counts.switchable_interp);
  vp9_zero(cpi->tx_stepdown_count);

  xd->mi_8x8 = cm->mi_grid_visible;
  // required for vp9_frame_init_quantizer
  xd->mi_8x8[0] = cm->mi;

  xd->last_mi = cm->prev_mi;

  vp9_zero(cpi->NMVcount);
  vp9_zero(cpi->coef_counts);
  vp9_zero(cm->counts.eob_branch);

  cpi->mb.e_mbd.lossless = cm->base_qindex == 0 && cm->y_dc_delta_q == 0
      && cm->uv_dc_delta_q == 0 && cm->uv_ac_delta_q == 0;
  switch_lossless_mode(cpi, cpi->mb.e_mbd.lossless);

  vp9_frame_init_quantizer(cpi);

  vp9_initialize_rd_consts(cpi);
  vp9_initialize_me_consts(cpi, cm->base_qindex);
  switch_tx_mode(cpi);

  if (cpi->oxcf.tuning == VP8_TUNE_SSIM) {
    // Initialize encode frame context.
    init_encode_frame_mb_context(cpi);

    // Build a frame level activity map
    build_activity_map(cpi);
  }

  // Re-initialize encode frame context.
  init_encode_frame_mb_context(cpi);

  vp9_zero(cpi->rd_comp_pred_diff);
  vp9_zero(cpi->rd_filter_diff);
  vp9_zero(cpi->rd_tx_select_diff);
  vp9_zero(cpi->rd_tx_select_threshes);

  set_prev_mi(cm);

  {
    struct vpx_usec_timer emr_timer;
    vpx_usec_timer_start(&emr_timer);

    {
      // Take tiles into account and give start/end MB
      int tile_col, tile_row;
      TOKENEXTRA *tp = cpi->tok;
      const int tile_cols = 1 << cm->log2_tile_cols;
      const int tile_rows = 1 << cm->log2_tile_rows;

      for (tile_row = 0; tile_row < tile_rows; tile_row++) {
        for (tile_col = 0; tile_col < tile_cols; tile_col++) {
          TileInfo tile;
          TOKENEXTRA *tp_old = tp;

          // For each row of SBs in the frame
          vp9_tile_init(&tile, cm, tile_row, tile_col);
          for (mi_row = tile.mi_row_start;
               mi_row < tile.mi_row_end; mi_row += 8)
            encode_sb_row(cpi, &tile, mi_row, &tp);

          cpi->tok_count[tile_row][tile_col] = (unsigned int)(tp - tp_old);
          assert(tp - cpi->tok <= get_token_alloc(cm->mb_rows, cm->mb_cols));
        }
      }
    }

    vpx_usec_timer_mark(&emr_timer);
    cpi->time_encode_sb_row += vpx_usec_timer_elapsed(&emr_timer);
  }