av1/encoder/partition_search.c

/*
 * Copyright (c) 2020, Alliance for Open Media. All rights reserved
 *
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
 */

#include "av1/common/av1_common_int.h"
#include "av1/common/blockd.h"
#include "av1/common/enums.h"
#include "av1/common/reconintra.h"

#include "av1/encoder/aq_complexity.h"
#include "av1/encoder/aq_variance.h"
#include "av1/encoder/context_tree.h"
#include "av1/encoder/encoder.h"
#include "av1/encoder/encodeframe.h"
#include "av1/encoder/encodeframe_utils.h"
#include "av1/encoder/encodemv.h"
#include "av1/encoder/motion_search_facade.h"
#include "av1/encoder/partition_search.h"
#include "av1/encoder/partition_strategy.h"
#include "av1/encoder/reconinter_enc.h"
#include "av1/encoder/tokenize.h"
#include "av1/encoder/var_based_part.h"
#include "av1/encoder/av1_ml_partition_models.h"

#if CONFIG_TUNE_VMAF
#include "av1/encoder/tune_vmaf.h"
#endif

#define COLLECT_MOTION_SEARCH_FEATURE_SB 0
#define ML_PARTITION_WHOLE_TREE_DECISION 0

void av1_reset_part_sf(PARTITION_SPEED_FEATURES *part_sf) {
  part_sf->partition_search_type = SEARCH_PARTITION;
  part_sf->less_rectangular_check_level = 0;
  part_sf->use_square_partition_only_threshold = BLOCK_128X128;
  part_sf->auto_max_partition_based_on_simple_motion = NOT_IN_USE;
  part_sf->default_max_partition_size = BLOCK_LARGEST;
  part_sf->default_min_partition_size = BLOCK_4X4;
  part_sf->adjust_var_based_rd_partitioning = 0;
  part_sf->max_intra_bsize = BLOCK_LARGEST;
  // This setting only takes effect when partition_search_type is set
  // to FIXED_PARTITION.
  part_sf->fixed_partition_size = BLOCK_16X16;
  // Recode loop tolerance %.
  part_sf->partition_search_breakout_dist_thr = 0;
  part_sf->partition_search_breakout_rate_thr = 0;
  part_sf->prune_ext_partition_types_search_level = 0;
  part_sf->prune_part4_search = 0;
  part_sf->ml_prune_partition = 0;
  part_sf->ml_early_term_after_part_split_level = 0;
  for (int i = 0; i < PARTITION_BLOCK_SIZES; ++i) {
    part_sf->ml_partition_search_breakout_thresh[i] =
        -1;  // -1 means not enabled.
  }
  part_sf->simple_motion_search_prune_agg = 0;
  part_sf->simple_motion_search_split = 0;
  part_sf->simple_motion_search_prune_rect = 0;
  part_sf->simple_motion_search_early_term_none = 0;
  part_sf->simple_motion_search_reduce_search_steps = 0;
  part_sf->intra_cnn_based_part_prune_level = 0;
  part_sf->ext_partition_eval_thresh = BLOCK_8X8;
  part_sf->rect_partition_eval_thresh = BLOCK_128X128;
  part_sf->prune_ext_part_using_split_info = 0;
  part_sf->prune_rectangular_split_based_on_qidx = 0;
  part_sf->early_term_after_none_split = 0;
  part_sf->ml_predict_breakout_level = 0;
  part_sf->prune_sub_8x8_partition_level = 0;
  part_sf->simple_motion_search_rect_split = 0;
  part_sf->reuse_prev_rd_results_for_part_ab = 0;
  part_sf->reuse_best_prediction_for_part_ab = 0;
  part_sf->use_best_rd_for_pruning = 0;
  part_sf->skip_non_sq_part_based_on_none = 0;
}

#if !CONFIG_REALTIME_ONLY
// If input |features| is NULL, write tpl stats to file for each super block.
// Otherwise, store tpl stats to |features|.
// The tpl stats is computed in the unit of tpl_bsize_1d (16x16).
// When writing to text file:
// The first row contains super block position, super block size,
// tpl unit length, number of units in the super block.
// The second row contains the intra prediction cost for each unit.
// The third row contains the inter prediction cost for each unit.
// The forth row contains the motion compensated dependency cost for each unit.
static void collect_tpl_stats_sb(const AV1_COMP *const cpi,
                                 const BLOCK_SIZE bsize, const int mi_row,
                                 const int mi_col,
                                 aom_partition_features_t *features) {
  const AV1_COMMON *const cm = &cpi->common;
  GF_GROUP *gf_group = &cpi->ppi->gf_group;
  if (gf_group->update_type[cpi->gf_frame_index] == INTNL_OVERLAY_UPDATE ||
      gf_group->update_type[cpi->gf_frame_index] == OVERLAY_UPDATE) {
    return;
  }

  TplParams *const tpl_data = &cpi->ppi->tpl_data;
  TplDepFrame *tpl_frame = &tpl_data->tpl_frame[cpi->gf_frame_index];
  TplDepStats *tpl_stats = tpl_frame->tpl_stats_ptr;
  // If tpl stats is not established, early return
  if (!tpl_data->ready || gf_group->max_layer_depth_allowed == 0) {
    features->sb_features.tpl_features.available = 0;
    return;
  }

  const int tpl_stride = tpl_frame->stride;
  const int step = 1 << tpl_data->tpl_stats_block_mis_log2;
  const int mi_width =
      AOMMIN(mi_size_wide[bsize], cm->mi_params.mi_cols - mi_col);
  const int mi_height =
      AOMMIN(mi_size_high[bsize], cm->mi_params.mi_rows - mi_row);
  const int col_steps = (mi_width / step) + ((mi_width % step) > 0);
  const int row_steps = (mi_height / step) + ((mi_height % step) > 0);
  const int num_blocks = col_steps * row_steps;

  if (features == NULL) {
    char filename[256];
    snprintf(filename, sizeof(filename), "%s/tpl_feature_sb%d",
             cpi->oxcf.partition_info_path, cpi->sb_counter);
    FILE *pfile = fopen(filename, "w");
    fprintf(pfile, "%d,%d,%d,%d,%d\n", mi_row, mi_col, bsize,
            tpl_data->tpl_bsize_1d, num_blocks);
    int count = 0;
    for (int row = 0; row < mi_height; row += step) {
      for (int col = 0; col < mi_width; col += step) {
        TplDepStats *this_stats =
            &tpl_stats[av1_tpl_ptr_pos(mi_row + row, mi_col + col, tpl_stride,
                                       tpl_data->tpl_stats_block_mis_log2)];
        fprintf(pfile, "%.0f", (double)this_stats->intra_cost);
        if (count < num_blocks - 1) fprintf(pfile, ",");
        ++count;
      }
    }
    fprintf(pfile, "\n");
    count = 0;
    for (int row = 0; row < mi_height; row += step) {
      for (int col = 0; col < mi_width; col += step) {
        TplDepStats *this_stats =
            &tpl_stats[av1_tpl_ptr_pos(mi_row + row, mi_col + col, tpl_stride,
                                       tpl_data->tpl_stats_block_mis_log2)];
        fprintf(pfile, "%.0f", (double)this_stats->inter_cost);
        if (count < num_blocks - 1) fprintf(pfile, ",");
        ++count;
      }
    }
    fprintf(pfile, "\n");
    count = 0;
    for (int row = 0; row < mi_height; row += step) {
      for (int col = 0; col < mi_width; col += step) {
        TplDepStats *this_stats =
            &tpl_stats[av1_tpl_ptr_pos(mi_row + row, mi_col + col, tpl_stride,
                                       tpl_data->tpl_stats_block_mis_log2)];
        const int64_t mc_dep_delta =
            RDCOST(tpl_frame->base_rdmult, this_stats->mc_dep_rate,
                   this_stats->mc_dep_dist);
        fprintf(pfile, "%.0f", (double)mc_dep_delta);
        if (count < num_blocks - 1) fprintf(pfile, ",");
        ++count;
      }
    }
    fclose(pfile);
  } else {
    features->sb_features.tpl_features.available = 1;
    features->sb_features.tpl_features.tpl_unit_length = tpl_data->tpl_bsize_1d;
    features->sb_features.tpl_features.num_units = num_blocks;
    int count = 0;
    for (int row = 0; row < mi_height; row += step) {
      for (int col = 0; col < mi_width; col += step) {
        TplDepStats *this_stats =
            &tpl_stats[av1_tpl_ptr_pos(mi_row + row, mi_col + col, tpl_stride,
                                       tpl_data->tpl_stats_block_mis_log2)];
        const int64_t mc_dep_delta =
            RDCOST(tpl_frame->base_rdmult, this_stats->mc_dep_rate,
                   this_stats->mc_dep_dist);
        features->sb_features.tpl_features.intra_cost[count] =
            this_stats->intra_cost;
        features->sb_features.tpl_features.inter_cost[count] =
            this_stats->inter_cost;
        features->sb_features.tpl_features.mc_dep_cost[count] = mc_dep_delta;
        ++count;
      }
    }
  }
}
#endif  // !CONFIG_REALTIME_ONLY

static void update_txfm_count(MACROBLOCK *x, MACROBLOCKD *xd,
                              FRAME_COUNTS *counts, TX_SIZE tx_size, int depth,
                              int blk_row, int blk_col,
                              uint8_t allow_update_cdf) {
  MB_MODE_INFO *mbmi = xd->mi[0];
  const BLOCK_SIZE bsize = mbmi->bsize;
  const int max_blocks_high = max_block_high(xd, bsize, 0);
  const int max_blocks_wide = max_block_wide(xd, bsize, 0);
  int ctx = txfm_partition_context(xd->above_txfm_context + blk_col,
                                   xd->left_txfm_context + blk_row, mbmi->bsize,
                                   tx_size);
  const int txb_size_index = av1_get_txb_size_index(bsize, blk_row, blk_col);
  const TX_SIZE plane_tx_size = mbmi->inter_tx_size[txb_size_index];

  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;
  assert(tx_size > TX_4X4);

  if (depth == MAX_VARTX_DEPTH) {
    // Don't add to counts in this case
    mbmi->tx_size = tx_size;
    txfm_partition_update(xd->above_txfm_context + blk_col,
                          xd->left_txfm_context + blk_row, tx_size, tx_size);
    return;
  }

  if (tx_size == plane_tx_size) {
#if CONFIG_ENTROPY_STATS
    ++counts->txfm_partition[ctx][0];
#endif
    if (allow_update_cdf)
      update_cdf(xd->tile_ctx->txfm_partition_cdf[ctx], 0, 2);
    mbmi->tx_size = tx_size;
    txfm_partition_update(xd->above_txfm_context + blk_col,
                          xd->left_txfm_context + blk_row, tx_size, tx_size);
  } else {
    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
    const int bsw = tx_size_wide_unit[sub_txs];
    const int bsh = tx_size_high_unit[sub_txs];

#if CONFIG_ENTROPY_STATS
    ++counts->txfm_partition[ctx][1];
#endif
    if (allow_update_cdf)
      update_cdf(xd->tile_ctx->txfm_partition_cdf[ctx], 1, 2);
    ++x->txfm_search_info.txb_split_count;

    if (sub_txs == TX_4X4) {
      mbmi->inter_tx_size[txb_size_index] = TX_4X4;
      mbmi->tx_size = TX_4X4;
      txfm_partition_update(xd->above_txfm_context + blk_col,
                            xd->left_txfm_context + blk_row, TX_4X4, tx_size);
      return;
    }

    for (int row = 0; row < tx_size_high_unit[tx_size]; row += bsh) {
      for (int col = 0; col < tx_size_wide_unit[tx_size]; col += bsw) {
        int offsetr = row;
        int offsetc = col;

        update_txfm_count(x, xd, counts, sub_txs, depth + 1, blk_row + offsetr,
                          blk_col + offsetc, allow_update_cdf);
      }
    }
  }
}

static void tx_partition_count_update(const AV1_COMMON *const cm, MACROBLOCK *x,
                                      BLOCK_SIZE plane_bsize,
                                      FRAME_COUNTS *td_counts,
                                      uint8_t allow_update_cdf) {
  MACROBLOCKD *xd = &x->e_mbd;
  const int mi_width = mi_size_wide[plane_bsize];
  const int mi_height = mi_size_high[plane_bsize];
  const TX_SIZE max_tx_size = get_vartx_max_txsize(xd, plane_bsize, 0);
  const int bh = tx_size_high_unit[max_tx_size];
  const int bw = tx_size_wide_unit[max_tx_size];

  xd->above_txfm_context =
      cm->above_contexts.txfm[xd->tile.tile_row] + xd->mi_col;
  xd->left_txfm_context =
      xd->left_txfm_context_buffer + (xd->mi_row & MAX_MIB_MASK);

  for (int idy = 0; idy < mi_height; idy += bh) {
    for (int idx = 0; idx < mi_width; idx += bw) {
      update_txfm_count(x, xd, td_counts, max_tx_size, 0, idy, idx,
                        allow_update_cdf);
    }
  }
}

static void set_txfm_context(MACROBLOCKD *xd, TX_SIZE tx_size, int blk_row,
                             int blk_col) {
  MB_MODE_INFO *mbmi = xd->mi[0];
  const BLOCK_SIZE bsize = mbmi->bsize;
  const int max_blocks_high = max_block_high(xd, bsize, 0);
  const int max_blocks_wide = max_block_wide(xd, bsize, 0);
  const int txb_size_index = av1_get_txb_size_index(bsize, blk_row, blk_col);
  const TX_SIZE plane_tx_size = mbmi->inter_tx_size[txb_size_index];

  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide) return;

  if (tx_size == plane_tx_size) {
    mbmi->tx_size = tx_size;
    txfm_partition_update(xd->above_txfm_context + blk_col,
                          xd->left_txfm_context + blk_row, tx_size, tx_size);

  } else {
    if (tx_size == TX_8X8) {
      mbmi->inter_tx_size[txb_size_index] = TX_4X4;
      mbmi->tx_size = TX_4X4;
      txfm_partition_update(xd->above_txfm_context + blk_col,
                            xd->left_txfm_context + blk_row, TX_4X4, tx_size);
      return;
    }
    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
    const int bsw = tx_size_wide_unit[sub_txs];
    const int bsh = tx_size_high_unit[sub_txs];
    const int row_end =
        AOMMIN(tx_size_high_unit[tx_size], max_blocks_high - blk_row);
    const int col_end =
        AOMMIN(tx_size_wide_unit[tx_size], max_blocks_wide - blk_col);
    for (int row = 0; row < row_end; row += bsh) {
      const int offsetr = blk_row + row;
      for (int col = 0; col < col_end; col += bsw) {
        const int offsetc = blk_col + col;
        set_txfm_context(xd, sub_txs, offsetr, offsetc);
      }
    }
  }
}

static void tx_partition_set_contexts(const AV1_COMMON *const cm,
                                      MACROBLOCKD *xd, BLOCK_SIZE plane_bsize) {
  const int mi_width = mi_size_wide[plane_bsize];
  const int mi_height = mi_size_high[plane_bsize];
  const TX_SIZE max_tx_size = get_vartx_max_txsize(xd, plane_bsize, 0);
  const int bh = tx_size_high_unit[max_tx_size];
  const int bw = tx_size_wide_unit[max_tx_size];

  xd->above_txfm_context =
      cm->above_contexts.txfm[xd->tile.tile_row] + xd->mi_col;
  xd->left_txfm_context =
      xd->left_txfm_context_buffer + (xd->mi_row & MAX_MIB_MASK);

  for (int idy = 0; idy < mi_height; idy += bh) {
    for (int idx = 0; idx < mi_width; idx += bw) {
      set_txfm_context(xd, max_tx_size, idy, idx);
    }
  }
}

static void update_zeromv_cnt(const AV1_COMP *const cpi,
                              const MB_MODE_INFO *const mi, int mi_row,
                              int mi_col, BLOCK_SIZE bsize) {
  if (mi->ref_frame[0] != LAST_FRAME || !is_inter_block(mi) ||
      mi->segment_id > CR_SEGMENT_ID_BOOST2) {
    return;
  }
  const AV1_COMMON *const cm = &cpi->common;
  const MV mv = mi->mv[0].as_mv;
  const int bw = mi_size_wide[bsize] >> 1;
  const int bh = mi_size_high[bsize] >> 1;
  const int xmis = AOMMIN((cm->mi_params.mi_cols - mi_col) >> 1, bw);
  const int ymis = AOMMIN((cm->mi_params.mi_rows - mi_row) >> 1, bh);
  const int block_index =
      (mi_row >> 1) * (cm->mi_params.mi_cols >> 1) + (mi_col >> 1);
  for (int y = 0; y < ymis; y++) {
    for (int x = 0; x < xmis; x++) {
      // consec_zero_mv is in the scale of 8x8 blocks
      const int map_offset = block_index + y * (cm->mi_params.mi_cols >> 1) + x;
      if (abs(mv.row) < 10 && abs(mv.col) < 10) {
        if (cpi->consec_zero_mv[map_offset] < 255)
          cpi->consec_zero_mv[map_offset]++;
      } else {
        cpi->consec_zero_mv[map_offset] = 0;
      }
    }
  }
}

static void encode_superblock(const AV1_COMP *const cpi, TileDataEnc *tile_data,
                              ThreadData *td, TokenExtra **t, RUN_TYPE dry_run,
                              BLOCK_SIZE bsize, int *rate) {
  const AV1_COMMON *const cm = &cpi->common;
  const int num_planes = av1_num_planes(cm);
  MACROBLOCK *const x = &td->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO **mi_4x4 = xd->mi;
  MB_MODE_INFO *mbmi = mi_4x4[0];
  const int seg_skip =
      segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP);
  const int mis = cm->mi_params.mi_stride;
  const int mi_width = mi_size_wide[bsize];
  const int mi_height = mi_size_high[bsize];
  const int is_inter = is_inter_block(mbmi);

  // Initialize tx_mode and tx_size_search_method
  TxfmSearchParams *txfm_params = &x->txfm_search_params;
  set_tx_size_search_method(
      cm, &cpi->winner_mode_params, txfm_params,
      cpi->sf.winner_mode_sf.enable_winner_mode_for_tx_size_srch, 1);

  const int mi_row = xd->mi_row;
  const int mi_col = xd->mi_col;
  if (!is_inter) {
    xd->cfl.store_y = store_cfl_required(cm, xd);
    mbmi->skip_txfm = 1;
    for (int plane = 0; plane < num_planes; ++plane) {
      av1_encode_intra_block_plane(cpi, x, bsize, plane, dry_run,
                                   cpi->optimize_seg_arr[mbmi->segment_id]);
    }

    // If there is at least one lossless segment, force the skip for intra
    // block to be 0, in order to avoid the segment_id to be changed by in
    // write_segment_id().
    if (!cpi->common.seg.segid_preskip && cpi->common.seg.update_map &&
        cpi->enc_seg.has_lossless_segment)
      mbmi->skip_txfm = 0;

    xd->cfl.store_y = 0;
    if (av1_allow_palette(cm->features.allow_screen_content_tools, bsize)) {
      for (int plane = 0; plane < AOMMIN(2, num_planes); ++plane) {
        if (mbmi->palette_mode_info.palette_size[plane] > 0) {
          if (!dry_run) {
            av1_tokenize_color_map(x, plane, t, bsize, mbmi->tx_size,
                                   PALETTE_MAP, tile_data->allow_update_cdf,
                                   td->counts);
          } else if (dry_run == DRY_RUN_COSTCOEFFS) {
            rate +=
                av1_cost_color_map(x, plane, bsize, mbmi->tx_size, PALETTE_MAP);
          }
        }
      }
    }

    av1_update_intra_mb_txb_context(cpi, td, dry_run, bsize,
                                    tile_data->allow_update_cdf);
  } else {
    int ref;
    const int is_compound = has_second_ref(mbmi);

    set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
    for (ref = 0; ref < 1 + is_compound; ++ref) {
      const YV12_BUFFER_CONFIG *cfg =
          get_ref_frame_yv12_buf(cm, mbmi->ref_frame[ref]);
      assert(IMPLIES(!is_intrabc_block(mbmi), cfg));
      av1_setup_pre_planes(xd, ref, cfg, mi_row, mi_col,
                           xd->block_ref_scale_factors[ref], num_planes);
    }
    const int start_plane = (cpi->sf.rt_sf.reuse_inter_pred_nonrd &&
                             cm->seq_params->bit_depth == AOM_BITS_8)
                                ? 1
                                : 0;
    av1_enc_build_inter_predictor(cm, xd, mi_row, mi_col, NULL, bsize,
                                  start_plane, av1_num_planes(cm) - 1);
    if (mbmi->motion_mode == OBMC_CAUSAL) {
      assert(cpi->oxcf.motion_mode_cfg.enable_obmc);
      av1_build_obmc_inter_predictors_sb(cm, xd);
    }

#if CONFIG_MISMATCH_DEBUG
    if (dry_run == OUTPUT_ENABLED) {
      for (int plane = 0; plane < num_planes; ++plane) {
        const struct macroblockd_plane *pd = &xd->plane[plane];
        int pixel_c, pixel_r;
        mi_to_pixel_loc(&pixel_c, &pixel_r, mi_col, mi_row, 0, 0,
                        pd->subsampling_x, pd->subsampling_y);
        if (!is_chroma_reference(mi_row, mi_col, bsize, pd->subsampling_x,
                                 pd->subsampling_y))
          continue;
        mismatch_record_block_pre(pd->dst.buf, pd->dst.stride,
                                  cm->current_frame.order_hint, plane, pixel_c,
                                  pixel_r, pd->width, pd->height,
                                  xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH);
      }
    }
#else
    (void)num_planes;
#endif

    av1_encode_sb(cpi, x, bsize, dry_run);
    av1_tokenize_sb_vartx(cpi, td, dry_run, bsize, rate,
                          tile_data->allow_update_cdf);
  }

  if (!dry_run) {
    if (av1_allow_intrabc(cm) && is_intrabc_block(mbmi)) td->intrabc_used = 1;
    if (txfm_params->tx_mode_search_type == TX_MODE_SELECT &&
        !xd->lossless[mbmi->segment_id] && mbmi->bsize > BLOCK_4X4 &&
        !(is_inter && (mbmi->skip_txfm || seg_skip))) {
      if (is_inter) {
        tx_partition_count_update(cm, x, bsize, td->counts,
                                  tile_data->allow_update_cdf);
      } else {
        if (mbmi->tx_size != max_txsize_rect_lookup[bsize])
          ++x->txfm_search_info.txb_split_count;
        if (block_signals_txsize(bsize)) {
          const int tx_size_ctx = get_tx_size_context(xd);
          const int32_t tx_size_cat = bsize_to_tx_size_cat(bsize);
          const int depth = tx_size_to_depth(mbmi->tx_size, bsize);
          const int max_depths = bsize_to_max_depth(bsize);

          if (tile_data->allow_update_cdf)
            update_cdf(xd->tile_ctx->tx_size_cdf[tx_size_cat][tx_size_ctx],
                       depth, max_depths + 1);
#if CONFIG_ENTROPY_STATS
          ++td->counts->intra_tx_size[tx_size_cat][tx_size_ctx][depth];
#endif
        }
      }
      assert(IMPLIES(is_rect_tx(mbmi->tx_size), is_rect_tx_allowed(xd, mbmi)));
    } else {
      int i, j;
      TX_SIZE intra_tx_size;
      // The new intra coding scheme requires no change of transform size
      if (is_inter) {
        if (xd->lossless[mbmi->segment_id]) {
          intra_tx_size = TX_4X4;
        } else {
          intra_tx_size =
              tx_size_from_tx_mode(bsize, txfm_params->tx_mode_search_type);
        }
      } else {
        intra_tx_size = mbmi->tx_size;
      }

      for (j = 0; j < mi_height; j++)
        for (i = 0; i < mi_width; i++)
          if (mi_col + i < cm->mi_params.mi_cols &&
              mi_row + j < cm->mi_params.mi_rows)
            mi_4x4[mis * j + i]->tx_size = intra_tx_size;

      if (intra_tx_size != max_txsize_rect_lookup[bsize])
        ++x->txfm_search_info.txb_split_count;
    }
  }

  if (txfm_params->tx_mode_search_type == TX_MODE_SELECT &&
      block_signals_txsize(mbmi->bsize) && is_inter &&
      !(mbmi->skip_txfm || seg_skip) && !xd->lossless[mbmi->segment_id]) {
    if (dry_run) tx_partition_set_contexts(cm, xd, bsize);
  } else {
    TX_SIZE tx_size = mbmi->tx_size;
    // The new intra coding scheme requires no change of transform size
    if (is_inter) {
      if (xd->lossless[mbmi->segment_id]) {
        tx_size = TX_4X4;
      } else {
        tx_size = tx_size_from_tx_mode(bsize, txfm_params->tx_mode_search_type);
      }
    } else {
      tx_size = (bsize > BLOCK_4X4) ? tx_size : TX_4X4;
    }
    mbmi->tx_size = tx_size;
    set_txfm_ctxs(tx_size, xd->width, xd->height,
                  (mbmi->skip_txfm || seg_skip) && is_inter_block(mbmi), xd);
  }

  if (is_inter_block(mbmi) && !xd->is_chroma_ref && is_cfl_allowed(xd)) {
    cfl_store_block(xd, mbmi->bsize, mbmi->tx_size);
  }
  if (!dry_run) {
    if (cpi->oxcf.pass == AOM_RC_ONE_PASS && cpi->svc.temporal_layer_id == 0 &&
        cpi->sf.rt_sf.use_temporal_noise_estimate &&
        (!cpi->ppi->use_svc ||
         (cpi->ppi->use_svc &&
          !cpi->svc.layer_context[cpi->svc.temporal_layer_id].is_key_frame &&
          cpi->svc.spatial_layer_id == cpi->svc.number_spatial_layers - 1)))
      update_zeromv_cnt(cpi, mbmi, mi_row, mi_col, bsize);
  }
}

static void setup_block_rdmult(const AV1_COMP *const cpi, MACROBLOCK *const x,
                               int mi_row, int mi_col, BLOCK_SIZE bsize,
                               AQ_MODE aq_mode, MB_MODE_INFO *mbmi) {
  x->rdmult = cpi->rd.RDMULT;

  if (aq_mode != NO_AQ) {
    assert(mbmi != NULL);
    if (aq_mode == VARIANCE_AQ) {
      if (cpi->vaq_refresh) {
        const int energy = bsize <= BLOCK_16X16
                               ? x->mb_energy
                               : av1_log_block_var(cpi, x, bsize);
        mbmi->segment_id = energy;
      }
      x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id);
    } else if (aq_mode == COMPLEXITY_AQ) {
      x->rdmult = set_segment_rdmult(cpi, x, mbmi->segment_id);
    } else if (aq_mode == CYCLIC_REFRESH_AQ) {
      // If segment is boosted, use rdmult for that segment.
      if (cyclic_refresh_segment_id_boosted(mbmi->segment_id))
        x->rdmult = av1_cyclic_refresh_get_rdmult(cpi->cyclic_refresh);
    }
  }

#if !CONFIG_REALTIME_ONLY
  const AV1_COMMON *const cm = &cpi->common;
  if (cm->delta_q_info.delta_q_present_flag &&
      !cpi->sf.rt_sf.use_nonrd_pick_mode) {
    x->rdmult =
        av1_get_hier_tpl_rdmult(cpi, x, bsize, mi_row, mi_col, x->rdmult);
  }
#endif  // !CONFIG_REALTIME_ONLY

  if (cpi->oxcf.tune_cfg.tuning == AOM_TUNE_SSIM) {
    av1_set_ssim_rdmult(cpi, &x->errorperbit, bsize, mi_row, mi_col,
                        &x->rdmult);
  }
#if CONFIG_TUNE_VMAF
  if (cpi->oxcf.tune_cfg.tuning == AOM_TUNE_VMAF_WITHOUT_PREPROCESSING ||
      cpi->oxcf.tune_cfg.tuning == AOM_TUNE_VMAF_MAX_GAIN ||
      cpi->oxcf.tune_cfg.tuning == AOM_TUNE_VMAF_NEG_MAX_GAIN) {
    av1_set_vmaf_rdmult(cpi, x, bsize, mi_row, mi_col, &x->rdmult);
  }
#endif
#if CONFIG_TUNE_BUTTERAUGLI
  if (cpi->oxcf.tune_cfg.tuning == AOM_TUNE_BUTTERAUGLI) {
    av1_set_butteraugli_rdmult(cpi, x, bsize, mi_row, mi_col, &x->rdmult);
  }
#endif
}

void av1_set_offsets_without_segment_id(const AV1_COMP *const cpi,
                                        const TileInfo *const tile,
                                        MACROBLOCK *const x, int mi_row,
                                        int mi_col, BLOCK_SIZE bsize) {
  const AV1_COMMON *const cm = &cpi->common;
  const int num_planes = av1_num_planes(cm);
  MACROBLOCKD *const xd = &x->e_mbd;
  assert(bsize < BLOCK_SIZES_ALL);
  const int mi_width = mi_size_wide[bsize];
  const int mi_height = mi_size_high[bsize];

  set_mode_info_offsets(&cpi->common.mi_params, &cpi->mbmi_ext_info, x, xd,
                        mi_row, mi_col);

  set_entropy_context(xd, mi_row, mi_col, num_planes);
  xd->above_txfm_context = cm->above_contexts.txfm[tile->tile_row] + mi_col;
  xd->left_txfm_context =
      xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);

  // Set up destination pointers.
  av1_setup_dst_planes(xd->plane, bsize, &cm->cur_frame->buf, mi_row, mi_col, 0,
                       num_planes);

  // Set up limit values for MV components.
  // Mv beyond the range do not produce new/different prediction block.
  av1_set_mv_limits(&cm->mi_params, &x->mv_limits, mi_row, mi_col, mi_height,
                    mi_width, cpi->oxcf.border_in_pixels);

  set_plane_n4(xd, mi_width, mi_height, num_planes);

  // 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_params.mi_rows, cm->mi_params.mi_cols);

  // Set up source buffers.
  av1_setup_src_planes(x, cpi->source, mi_row, mi_col, num_planes, bsize);

  // required by av1_append_sub8x8_mvs_for_idx() and av1_find_best_ref_mvs()
  xd->tile = *tile;
}

void av1_set_offsets(const AV1_COMP *const cpi, const TileInfo *const tile,
                     MACROBLOCK *const x, int mi_row, int mi_col,
                     BLOCK_SIZE bsize) {
  const AV1_COMMON *const cm = &cpi->common;
  const struct segmentation *const seg = &cm->seg;
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *mbmi;

  av1_set_offsets_without_segment_id(cpi, tile, x, mi_row, mi_col, bsize);

  // Setup segment ID.
  mbmi = xd->mi[0];
  mbmi->segment_id = 0;
  if (seg->enabled) {
    if (seg->enabled && !cpi->vaq_refresh) {
      const uint8_t *const map =
          seg->update_map ? cpi->enc_seg.map : cm->last_frame_seg_map;
      mbmi->segment_id =
          map ? get_segment_id(&cm->mi_params, map, bsize, mi_row, mi_col) : 0;
    }
    av1_init_plane_quantizers(cpi, x, mbmi->segment_id);
  }
}

/*!\brief Hybrid intra mode search.
 *
 * \ingroup intra_mode_search
 * \callgraph
 * \callergraph
 * This is top level function for mode search for intra frames in non-RD
 * optimized case. Depending on speed feature and block size it calls
 * either non-RD or RD optimized intra mode search.
 *
 * \param[in]    cpi            Top-level encoder structure
 * \param[in]    x              Pointer to structure holding all the data for
                                the current macroblock
 * \param[in]    rd_cost        Struct to keep track of the RD information
 * \param[in]    bsize          Current block size
 * \param[in]    ctx            Structure to hold snapshot of coding context
                                during the mode picking process
 *
 * \return Nothing is returned. Instead, the MB_MODE_INFO struct inside x
 * is modified to store information about the best mode computed
 * in this function. The rd_cost struct is also updated with the RD stats
 * corresponding to the best mode found.
 */

static AOM_INLINE void hybrid_intra_mode_search(AV1_COMP *cpi,
                                                MACROBLOCK *const x,
                                                RD_STATS *rd_cost,
                                                BLOCK_SIZE bsize,
                                                PICK_MODE_CONTEXT *ctx) {
  if (cpi->sf.rt_sf.hybrid_intra_pickmode && bsize < BLOCK_16X16)
    av1_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, INT64_MAX);
  else
    av1_nonrd_pick_intra_mode(cpi, x, rd_cost, bsize, ctx);
}

// For real time row-mt enabled multi-threaded encoding with cost update
// frequency set to COST_UPD_TILE/COST_UPD_OFF, tile ctxt is not updated at
// superblock level. Thus, it is not required for the encoding of top-right
// superblock be complete for updating tile ctxt. However, when encoding a block
// whose right edge is also the superblock edge, intra and inter mode evaluation
// (ref mv list population) require the encoding of the top-right superblock to
// be complete. So, here, we delay the waiting of threads until the need for the
// data from the top-right superblock region.
static AOM_INLINE void wait_for_top_right_sb(
    AV1EncRowMultiThreadInfo *enc_row_mt, AV1EncRowMultiThreadSync *row_mt_sync,
    TileInfo *tile_info, BLOCK_SIZE sb_size, int sb_mi_size_log2,
    BLOCK_SIZE bsize, int mi_row, int mi_col) {
  const int sb_size_in_mi = mi_size_wide[sb_size];
  const int bw_in_mi = mi_size_wide[bsize];
  const int blk_row_in_sb = mi_row & (sb_size_in_mi - 1);
  const int blk_col_in_sb = mi_col & (sb_size_in_mi - 1);
  const int top_right_block_in_sb =
      (blk_row_in_sb == 0) && (blk_col_in_sb + bw_in_mi >= sb_size_in_mi);

  // Don't wait if the block is the not the top-right block in the superblock.
  if (!top_right_block_in_sb) return;

  // Wait for the top-right superblock to finish encoding.
  const int sb_row_in_tile =
      (mi_row - tile_info->mi_row_start) >> sb_mi_size_log2;
  const int sb_col_in_tile =
      (mi_col - tile_info->mi_col_start) >> sb_mi_size_log2;

  (*(enc_row_mt->sync_read_ptr))(row_mt_sync, sb_row_in_tile, sb_col_in_tile);
}

/*!\brief Interface for AV1 mode search for an individual coding block
 *
 * \ingroup partition_search
 * \callgraph
 * \callergraph
 * Searches prediction modes, transform, and coefficient coding modes for an
 * individual coding block. This function is the top-level interface that
 * directs the encoder to the proper mode search function, among these
 * implemented for inter/intra + rd/non-rd + non-skip segment/skip segment.
 *
 * \param[in]    cpi            Top-level encoder structure
 * \param[in]    tile_data      Pointer to struct holding adaptive
 *                              data/contexts/models for the tile during
 *                              encoding
 * \param[in]    x              Pointer to structure holding all the data for
 *                              the current macroblock
 * \param[in]    mi_row         Row coordinate of the block in a step size of
 *                              MI_SIZE
 * \param[in]    mi_col         Column coordinate of the block in a step size of
 *                              MI_SIZE
 * \param[in]    rd_cost        Pointer to structure holding rate and distortion
 *                              stats for the current block
 * \param[in]    partition      Partition mode of the parent block
 * \param[in]    bsize          Current block size
 * \param[in]    ctx            Pointer to structure holding coding contexts and
 *                              chosen modes for the current block
 * \param[in]    best_rd        Upper bound of rd cost of a valid partition
 *
 * \return Nothing is returned. Instead, the chosen modes and contexts necessary
 * for reconstruction are stored in ctx, the rate-distortion stats are stored in
 * rd_cost. If no valid mode leading to rd_cost <= best_rd, the status will be
 * signalled by an INT64_MAX rd_cost->rdcost.
 */
static void pick_sb_modes(AV1_COMP *const cpi, TileDataEnc *tile_data,
                          MACROBLOCK *const x, int mi_row, int mi_col,
                          RD_STATS *rd_cost, PARTITION_TYPE partition,
                          BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
                          RD_STATS best_rd) {
  if (cpi->sf.part_sf.use_best_rd_for_pruning && best_rd.rdcost < 0) {
    ctx->rd_stats.rdcost = INT64_MAX;
    ctx->rd_stats.skip_txfm = 0;
    av1_invalid_rd_stats(rd_cost);
    return;
  }

  av1_set_offsets(cpi, &tile_data->tile_info, x, mi_row, mi_col, bsize);

  if (cpi->sf.part_sf.reuse_prev_rd_results_for_part_ab &&
      ctx->rd_mode_is_ready) {
    assert(ctx->mic.bsize == bsize);
    assert(ctx->mic.partition == partition);
    rd_cost->rate = ctx->rd_stats.rate;
    rd_cost->dist = ctx->rd_stats.dist;
    rd_cost->rdcost = ctx->rd_stats.rdcost;
    return;
  }

  AV1_COMMON *const cm = &cpi->common;
  const int num_planes = av1_num_planes(cm);
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *mbmi;
  struct macroblock_plane *const p = x->plane;
  struct macroblockd_plane *const pd = xd->plane;
  const AQ_MODE aq_mode = cpi->oxcf.q_cfg.aq_mode;
  TxfmSearchInfo *txfm_info = &x->txfm_search_info;

  int i;

  // This is only needed for real time row-mt enabled multi-threaded encoding
  // with cost update frequency set to COST_UPD_TILE/COST_UPD_OFF.
  wait_for_top_right_sb(&cpi->mt_info.enc_row_mt, &tile_data->row_mt_sync,
                        &tile_data->tile_info, cm->seq_params->sb_size,
                        cm->seq_params->mib_size_log2, bsize, mi_row, mi_col);

#if CONFIG_COLLECT_COMPONENT_TIMING
  start_timing(cpi, rd_pick_sb_modes_time);
#endif

  mbmi = xd->mi[0];
  mbmi->bsize = bsize;
  mbmi->partition = partition;

#if CONFIG_RD_DEBUG
  mbmi->mi_row = mi_row;
  mbmi->mi_col = mi_col;
#endif

  // Sets up the tx_type_map buffer in MACROBLOCKD.
  xd->tx_type_map = txfm_info->tx_type_map_;
  xd->tx_type_map_stride = mi_size_wide[bsize];

  for (i = 0; i < num_planes; ++i) {
    p[i].coeff = ctx->coeff[i];
    p[i].qcoeff = ctx->qcoeff[i];
    p[i].dqcoeff = ctx->dqcoeff[i];
    p[i].eobs = ctx->eobs[i];
    p[i].txb_entropy_ctx = ctx->txb_entropy_ctx[i];
  }

  for (i = 0; i < 2; ++i) pd[i].color_index_map = ctx->color_index_map[i];

  ctx->skippable = 0;
  // Set to zero to make sure we do not use the previous encoded frame stats
  mbmi->skip_txfm = 0;
  // Reset skip mode flag.
  mbmi->skip_mode = 0;

  if (is_cur_buf_hbd(xd)) {
    x->source_variance = av1_high_get_sby_perpixel_variance(
        cpi, &x->plane[0].src, bsize, xd->bd);
  } else {
    x->source_variance =
        av1_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
  }

  // Initialize default mode evaluation params
  set_mode_eval_params(cpi, x, DEFAULT_EVAL);

  // Save rdmult before it might be changed, so it can be restored later.
  const int orig_rdmult = x->rdmult;
  setup_block_rdmult(cpi, x, mi_row, mi_col, bsize, aq_mode, mbmi);
  // Set error per bit for current rdmult
  av1_set_error_per_bit(&x->errorperbit, x->rdmult);
  av1_rd_cost_update(x->rdmult, &best_rd);

  // If set best_rd.rdcost to INT64_MAX, the encoder will not use any previous
  // rdcost information for the following mode search.
  // Disabling the feature could get some coding gain, with encoder slowdown.
  if (!cpi->sf.part_sf.use_best_rd_for_pruning) {
    av1_invalid_rd_stats(&best_rd);
  }

  // Find best coding mode & reconstruct the MB so it is available
  // as a predictor for MBs that follow in the SB
  if (frame_is_intra_only(cm)) {
#if CONFIG_COLLECT_COMPONENT_TIMING
    start_timing(cpi, av1_rd_pick_intra_mode_sb_time);
#endif
    av1_rd_pick_intra_mode_sb(cpi, x, rd_cost, bsize, ctx, best_rd.rdcost);
#if CONFIG_COLLECT_COMPONENT_TIMING
    end_timing(cpi, av1_rd_pick_intra_mode_sb_time);
#endif
  } else {
#if CONFIG_COLLECT_COMPONENT_TIMING
    start_timing(cpi, av1_rd_pick_inter_mode_sb_time);
#endif
    if (segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
      av1_rd_pick_inter_mode_sb_seg_skip(cpi, tile_data, x, mi_row, mi_col,
                                         rd_cost, bsize, ctx, best_rd.rdcost);
    } else {
      av1_rd_pick_inter_mode(cpi, tile_data, x, rd_cost, bsize, ctx,
                             best_rd.rdcost);
    }
#if CONFIG_COLLECT_COMPONENT_TIMING
    end_timing(cpi, av1_rd_pick_inter_mode_sb_time);
#endif
  }

  // Examine the resulting rate and for AQ mode 2 make a segment choice.
  if (rd_cost->rate != INT_MAX && aq_mode == COMPLEXITY_AQ &&
      bsize >= BLOCK_16X16) {
    av1_caq_select_segment(cpi, x, bsize, mi_row, mi_col, rd_cost->rate);
  }

  x->rdmult = orig_rdmult;

  // TODO(jingning) The rate-distortion optimization flow needs to be
  // refactored to provide proper exit/return handle.
  if (rd_cost->rate == INT_MAX) rd_cost->rdcost = INT64_MAX;

  ctx->rd_stats.rate = rd_cost->rate;
  ctx->rd_stats.dist = rd_cost->dist;
  ctx->rd_stats.rdcost = rd_cost->rdcost;

#if CONFIG_COLLECT_COMPONENT_TIMING
  end_timing(cpi, rd_pick_sb_modes_time);
#endif
}

static void update_stats(const AV1_COMMON *const cm, ThreadData *td) {
  MACROBLOCK *x = &td->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  const MB_MODE_INFO *const mbmi = xd->mi[0];
  const MB_MODE_INFO_EXT *const mbmi_ext = &x->mbmi_ext;
  const CurrentFrame *const current_frame = &cm->current_frame;
  const BLOCK_SIZE bsize = mbmi->bsize;
  FRAME_CONTEXT *fc = xd->tile_ctx;
  const int seg_ref_active =
      segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_REF_FRAME);

  if (current_frame->skip_mode_info.skip_mode_flag && !seg_ref_active &&
      is_comp_ref_allowed(bsize)) {
    const int skip_mode_ctx = av1_get_skip_mode_context(xd);
#if CONFIG_ENTROPY_STATS
    td->counts->skip_mode[skip_mode_ctx][mbmi->skip_mode]++;
#endif
    update_cdf(fc->skip_mode_cdfs[skip_mode_ctx], mbmi->skip_mode, 2);
  }

  if (!mbmi->skip_mode && !seg_ref_active) {
    const int skip_ctx = av1_get_skip_txfm_context(xd);
#if CONFIG_ENTROPY_STATS
    td->counts->skip_txfm[skip_ctx][mbmi->skip_txfm]++;
#endif
    update_cdf(fc->skip_txfm_cdfs[skip_ctx], mbmi->skip_txfm, 2);
  }

#if CONFIG_ENTROPY_STATS
  // delta quant applies to both intra and inter
  const int super_block_upper_left =
      ((xd->mi_row & (cm->seq_params->mib_size - 1)) == 0) &&
      ((xd->mi_col & (cm->seq_params->mib_size - 1)) == 0);
  const DeltaQInfo *const delta_q_info = &cm->delta_q_info;
  if (delta_q_info->delta_q_present_flag &&
      (bsize != cm->seq_params->sb_size || !mbmi->skip_txfm) &&
      super_block_upper_left) {
    const int dq = (mbmi->current_qindex - xd->current_base_qindex) /
                   delta_q_info->delta_q_res;
    const int absdq = abs(dq);
    for (int i = 0; i < AOMMIN(absdq, DELTA_Q_SMALL); ++i) {
      td->counts->delta_q[i][1]++;
    }
    if (absdq < DELTA_Q_SMALL) td->counts->delta_q[absdq][0]++;
    if (delta_q_info->delta_lf_present_flag) {
      if (delta_q_info->delta_lf_multi) {
        const int frame_lf_count =
            av1_num_planes(cm) > 1 ? FRAME_LF_COUNT : FRAME_LF_COUNT - 2;
        for (int lf_id = 0; lf_id < frame_lf_count; ++lf_id) {
          const int delta_lf = (mbmi->delta_lf[lf_id] - xd->delta_lf[lf_id]) /
                               delta_q_info->delta_lf_res;
          const int abs_delta_lf = abs(delta_lf);
          for (int i = 0; i < AOMMIN(abs_delta_lf, DELTA_LF_SMALL); ++i) {
            td->counts->delta_lf_multi[lf_id][i][1]++;
          }
          if (abs_delta_lf < DELTA_LF_SMALL)
            td->counts->delta_lf_multi[lf_id][abs_delta_lf][0]++;
        }
      } else {
        const int delta_lf =
            (mbmi->delta_lf_from_base - xd->delta_lf_from_base) /
            delta_q_info->delta_lf_res;
        const int abs_delta_lf = abs(delta_lf);
        for (int i = 0; i < AOMMIN(abs_delta_lf, DELTA_LF_SMALL); ++i) {
          td->counts->delta_lf[i][1]++;
        }
        if (abs_delta_lf < DELTA_LF_SMALL)
          td->counts->delta_lf[abs_delta_lf][0]++;
      }
    }
  }
#endif

  if (!is_inter_block(mbmi)) {
    av1_sum_intra_stats(cm, td->counts, xd, mbmi, xd->above_mbmi, xd->left_mbmi,
                        frame_is_intra_only(cm));
  }

  if (av1_allow_intrabc(cm)) {
    const int is_intrabc = is_intrabc_block(mbmi);
    update_cdf(fc->intrabc_cdf, is_intrabc, 2);
#if CONFIG_ENTROPY_STATS
    ++td->counts->intrabc[is_intrabc];
#endif  // CONFIG_ENTROPY_STATS
    if (is_intrabc) {
      const int_mv dv_ref = x->mbmi_ext_frame->ref_mv_stack[0].this_mv;
      av1_update_mv_stats(&mbmi->mv[0].as_mv, &dv_ref.as_mv, &fc->ndvc,
                          MV_SUBPEL_NONE);
    }
  }

  if (frame_is_intra_only(cm) || mbmi->skip_mode) return;

  FRAME_COUNTS *const counts = td->counts;
  const int inter_block = is_inter_block(mbmi);

  if (!seg_ref_active) {
#if CONFIG_ENTROPY_STATS
    counts->intra_inter[av1_get_intra_inter_context(xd)][inter_block]++;
#endif
    update_cdf(fc->intra_inter_cdf[av1_get_intra_inter_context(xd)],
               inter_block, 2);
    // If the segment reference feature is enabled we have only a single
    // reference frame allowed for the segment so exclude it from
    // the reference frame counts used to work out probabilities.
    if (inter_block) {
      const MV_REFERENCE_FRAME ref0 = mbmi->ref_frame[0];
      const MV_REFERENCE_FRAME ref1 = mbmi->ref_frame[1];
      if (current_frame->reference_mode == REFERENCE_MODE_SELECT) {
        if (is_comp_ref_allowed(bsize)) {
#if CONFIG_ENTROPY_STATS
          counts->comp_inter[av1_get_reference_mode_context(xd)]
                            [has_second_ref(mbmi)]++;
#endif  // CONFIG_ENTROPY_STATS
          update_cdf(av1_get_reference_mode_cdf(xd), has_second_ref(mbmi), 2);
        }
      }

      if (has_second_ref(mbmi)) {
        const COMP_REFERENCE_TYPE comp_ref_type = has_uni_comp_refs(mbmi)
                                                      ? UNIDIR_COMP_REFERENCE
                                                      : BIDIR_COMP_REFERENCE;
        update_cdf(av1_get_comp_reference_type_cdf(xd), comp_ref_type,
                   COMP_REFERENCE_TYPES);
#if CONFIG_ENTROPY_STATS
        counts->comp_ref_type[av1_get_comp_reference_type_context(xd)]
                             [comp_ref_type]++;
#endif  // CONFIG_ENTROPY_STATS

        if (comp_ref_type == UNIDIR_COMP_REFERENCE) {
          const int bit = (ref0 == BWDREF_FRAME);
          update_cdf(av1_get_pred_cdf_uni_comp_ref_p(xd), bit, 2);
#if CONFIG_ENTROPY_STATS
          counts
              ->uni_comp_ref[av1_get_pred_context_uni_comp_ref_p(xd)][0][bit]++;
#endif  // CONFIG_ENTROPY_STATS
          if (!bit) {
            const int bit1 = (ref1 == LAST3_FRAME || ref1 == GOLDEN_FRAME);
            update_cdf(av1_get_pred_cdf_uni_comp_ref_p1(xd), bit1, 2);
#if CONFIG_ENTROPY_STATS
            counts->uni_comp_ref[av1_get_pred_context_uni_comp_ref_p1(xd)][1]
                                [bit1]++;
#endif  // CONFIG_ENTROPY_STATS
            if (bit1) {
              update_cdf(av1_get_pred_cdf_uni_comp_ref_p2(xd),
                         ref1 == GOLDEN_FRAME, 2);
#if CONFIG_ENTROPY_STATS
              counts->uni_comp_ref[av1_get_pred_context_uni_comp_ref_p2(xd)][2]
                                  [ref1 == GOLDEN_FRAME]++;
#endif  // CONFIG_ENTROPY_STATS
            }
          }
        } else {
          const int bit = (ref0 == GOLDEN_FRAME || ref0 == LAST3_FRAME);
          update_cdf(av1_get_pred_cdf_comp_ref_p(xd), bit, 2);
#if CONFIG_ENTROPY_STATS
          counts->comp_ref[av1_get_pred_context_comp_ref_p(xd)][0][bit]++;
#endif  // CONFIG_ENTROPY_STATS
          if (!bit) {
            update_cdf(av1_get_pred_cdf_comp_ref_p1(xd), ref0 == LAST2_FRAME,
                       2);
#if CONFIG_ENTROPY_STATS
            counts->comp_ref[av1_get_pred_context_comp_ref_p1(xd)][1]
                            [ref0 == LAST2_FRAME]++;
#endif  // CONFIG_ENTROPY_STATS
          } else {
            update_cdf(av1_get_pred_cdf_comp_ref_p2(xd), ref0 == GOLDEN_FRAME,
                       2);
#if CONFIG_ENTROPY_STATS
            counts->comp_ref[av1_get_pred_context_comp_ref_p2(xd)][2]
                            [ref0 == GOLDEN_FRAME]++;
#endif  // CONFIG_ENTROPY_STATS
          }
          update_cdf(av1_get_pred_cdf_comp_bwdref_p(xd), ref1 == ALTREF_FRAME,
                     2);
#if CONFIG_ENTROPY_STATS
          counts->comp_bwdref[av1_get_pred_context_comp_bwdref_p(xd)][0]
                             [ref1 == ALTREF_FRAME]++;
#endif  // CONFIG_ENTROPY_STATS
          if (ref1 != ALTREF_FRAME) {
            update_cdf(av1_get_pred_cdf_comp_bwdref_p1(xd),
                       ref1 == ALTREF2_FRAME, 2);
#if CONFIG_ENTROPY_STATS
            counts->comp_bwdref[av1_get_pred_context_comp_bwdref_p1(xd)][1]
                               [ref1 == ALTREF2_FRAME]++;
#endif  // CONFIG_ENTROPY_STATS
          }
        }
      } else {
        const int bit = (ref0 >= BWDREF_FRAME);
        update_cdf(av1_get_pred_cdf_single_ref_p1(xd), bit, 2);
#if CONFIG_ENTROPY_STATS
        counts->single_ref[av1_get_pred_context_single_ref_p1(xd)][0][bit]++;
#endif  // CONFIG_ENTROPY_STATS
        if (bit) {
          assert(ref0 <= ALTREF_FRAME);
          update_cdf(av1_get_pred_cdf_single_ref_p2(xd), ref0 == ALTREF_FRAME,
                     2);
#if CONFIG_ENTROPY_STATS
          counts->single_ref[av1_get_pred_context_single_ref_p2(xd)][1]
                            [ref0 == ALTREF_FRAME]++;
#endif  // CONFIG_ENTROPY_STATS
          if (ref0 != ALTREF_FRAME) {
            update_cdf(av1_get_pred_cdf_single_ref_p6(xd),
                       ref0 == ALTREF2_FRAME, 2);
#if CONFIG_ENTROPY_STATS
            counts->single_ref[av1_get_pred_context_single_ref_p6(xd)][5]
                              [ref0 == ALTREF2_FRAME]++;
#endif  // CONFIG_ENTROPY_STATS
          }
        } else {
          const int bit1 = !(ref0 == LAST2_FRAME || ref0 == LAST_FRAME);
          update_cdf(av1_get_pred_cdf_single_ref_p3(xd), bit1, 2);
#if CONFIG_ENTROPY_STATS
          counts->single_ref[av1_get_pred_context_single_ref_p3(xd)][2][bit1]++;
#endif  // CONFIG_ENTROPY_STATS
          if (!bit1) {
            update_cdf(av1_get_pred_cdf_single_ref_p4(xd), ref0 != LAST_FRAME,
                       2);
#if CONFIG_ENTROPY_STATS
            counts->single_ref[av1_get_pred_context_single_ref_p4(xd)][3]
                              [ref0 != LAST_FRAME]++;
#endif  // CONFIG_ENTROPY_STATS
          } else {
            update_cdf(av1_get_pred_cdf_single_ref_p5(xd), ref0 != LAST3_FRAME,
                       2);
#if CONFIG_ENTROPY_STATS
            counts->single_ref[av1_get_pred_context_single_ref_p5(xd)][4]
                              [ref0 != LAST3_FRAME]++;
#endif  // CONFIG_ENTROPY_STATS
          }
        }
      }

      if (cm->seq_params->enable_interintra_compound &&
          is_interintra_allowed(mbmi)) {
        const int bsize_group = size_group_lookup[bsize];
        if (mbmi->ref_frame[1] == INTRA_FRAME) {
#if CONFIG_ENTROPY_STATS
          counts->interintra[bsize_group][1]++;
#endif
          update_cdf(fc->interintra_cdf[bsize_group], 1, 2);
#if CONFIG_ENTROPY_STATS
          counts->interintra_mode[bsize_group][mbmi->interintra_mode]++;
#endif
          update_cdf(fc->interintra_mode_cdf[bsize_group],
                     mbmi->interintra_mode, INTERINTRA_MODES);
          if (av1_is_wedge_used(bsize)) {
#if CONFIG_ENTROPY_STATS
            counts->wedge_interintra[bsize][mbmi->use_wedge_interintra]++;
#endif
            update_cdf(fc->wedge_interintra_cdf[bsize],
                       mbmi->use_wedge_interintra, 2);
            if (mbmi->use_wedge_interintra) {
#if CONFIG_ENTROPY_STATS
              counts->wedge_idx[bsize][mbmi->interintra_wedge_index]++;
#endif
              update_cdf(fc->wedge_idx_cdf[bsize], mbmi->interintra_wedge_index,
                         16);
            }
          }
        } else {
#if CONFIG_ENTROPY_STATS
          counts->interintra[bsize_group][0]++;
#endif
          update_cdf(fc->interintra_cdf[bsize_group], 0, 2);
        }
      }

      const MOTION_MODE motion_allowed =
          cm->features.switchable_motion_mode
              ? motion_mode_allowed(xd->global_motion, xd, mbmi,
                                    cm->features.allow_warped_motion)
              : SIMPLE_TRANSLATION;
      if (mbmi->ref_frame[1] != INTRA_FRAME) {
        if (motion_allowed == WARPED_CAUSAL) {
#if CONFIG_ENTROPY_STATS
          counts->motion_mode[bsize][mbmi->motion_mode]++;
#endif
          update_cdf(fc->motion_mode_cdf[bsize], mbmi->motion_mode,
                     MOTION_MODES);
        } else if (motion_allowed == OBMC_CAUSAL) {
#if CONFIG_ENTROPY_STATS
          counts->obmc[bsize][mbmi->motion_mode == OBMC_CAUSAL]++;
#endif
          update_cdf(fc->obmc_cdf[bsize], mbmi->motion_mode == OBMC_CAUSAL, 2);
        }
      }

      if (has_second_ref(mbmi)) {
        assert(current_frame->reference_mode != SINGLE_REFERENCE &&
               is_inter_compound_mode(mbmi->mode) &&
               mbmi->motion_mode == SIMPLE_TRANSLATION);

        const int masked_compound_used = is_any_masked_compound_used(bsize) &&
                                         cm->seq_params->enable_masked_compound;
        if (masked_compound_used) {
          const int comp_group_idx_ctx = get_comp_group_idx_context(xd);
#if CONFIG_ENTROPY_STATS
          ++counts->comp_group_idx[comp_group_idx_ctx][mbmi->comp_group_idx];
#endif
          update_cdf(fc->comp_group_idx_cdf[comp_group_idx_ctx],
                     mbmi->comp_group_idx, 2);
        }

        if (mbmi->comp_group_idx == 0) {
          const int comp_index_ctx = get_comp_index_context(cm, xd);
#if CONFIG_ENTROPY_STATS
          ++counts->compound_index[comp_index_ctx][mbmi->compound_idx];
#endif
          update_cdf(fc->compound_index_cdf[comp_index_ctx], mbmi->compound_idx,
                     2);
        } else {
          assert(masked_compound_used);
          if (is_interinter_compound_used(COMPOUND_WEDGE, bsize)) {
#if CONFIG_ENTROPY_STATS
            ++counts->compound_type[bsize][mbmi->interinter_comp.type -
                                           COMPOUND_WEDGE];
#endif
            update_cdf(fc->compound_type_cdf[bsize],
                       mbmi->interinter_comp.type - COMPOUND_WEDGE,
                       MASKED_COMPOUND_TYPES);
          }
        }
      }
      if (mbmi->interinter_comp.type == COMPOUND_WEDGE) {
        if (is_interinter_compound_used(COMPOUND_WEDGE, bsize)) {
#if CONFIG_ENTROPY_STATS
          counts->wedge_idx[bsize][mbmi->interinter_comp.wedge_index]++;
#endif
          update_cdf(fc->wedge_idx_cdf[bsize],
                     mbmi->interinter_comp.wedge_index, 16);
        }
      }
    }
  }

  if (inter_block && cm->features.interp_filter == SWITCHABLE &&
      mbmi->motion_mode != WARPED_CAUSAL &&
      !is_nontrans_global_motion(xd, mbmi)) {
    update_filter_type_cdf(xd, mbmi, cm->seq_params->enable_dual_filter);
  }
  if (inter_block &&
      !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
    const PREDICTION_MODE mode = mbmi->mode;
    const int16_t mode_ctx =
        av1_mode_context_analyzer(mbmi_ext->mode_context, mbmi->ref_frame);
    if (has_second_ref(mbmi)) {
#if CONFIG_ENTROPY_STATS
      ++counts->inter_compound_mode[mode_ctx][INTER_COMPOUND_OFFSET(mode)];
#endif
      update_cdf(fc->inter_compound_mode_cdf[mode_ctx],
                 INTER_COMPOUND_OFFSET(mode), INTER_COMPOUND_MODES);
    } else {
      av1_update_inter_mode_stats(fc, counts, mode, mode_ctx);
    }

    const int new_mv = mbmi->mode == NEWMV || mbmi->mode == NEW_NEWMV;
    if (new_mv) {
      const uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
      for (int idx = 0; idx < 2; ++idx) {
        if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) {
          const uint8_t drl_ctx =
              av1_drl_ctx(mbmi_ext->weight[ref_frame_type], idx);
          update_cdf(fc->drl_cdf[drl_ctx], mbmi->ref_mv_idx != idx, 2);
#if CONFIG_ENTROPY_STATS
          ++counts->drl_mode[drl_ctx][mbmi->ref_mv_idx != idx];
#endif
          if (mbmi->ref_mv_idx == idx) break;
        }
      }
    }

    if (have_nearmv_in_inter_mode(mbmi->mode)) {
      const uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);
      for (int idx = 1; idx < 3; ++idx) {
        if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) {
          const uint8_t drl_ctx =
              av1_drl_ctx(mbmi_ext->weight[ref_frame_type], idx);
          update_cdf(fc->drl_cdf[drl_ctx], mbmi->ref_mv_idx != idx - 1, 2);
#if CONFIG_ENTROPY_STATS
          ++counts->drl_mode[drl_ctx][mbmi->ref_mv_idx != idx - 1];
#endif
          if (mbmi->ref_mv_idx == idx - 1) break;
        }
      }
    }
    if (have_newmv_in_inter_mode(mbmi->mode)) {
      const int allow_hp = cm->features.cur_frame_force_integer_mv
                               ? MV_SUBPEL_NONE
                               : cm->features.allow_high_precision_mv;
      if (new_mv) {
        for (int ref = 0; ref < 1 + has_second_ref(mbmi); ++ref) {
          const int_mv ref_mv = av1_get_ref_mv(x, ref);
          av1_update_mv_stats(&mbmi->mv[ref].as_mv, &ref_mv.as_mv, &fc->nmvc,
                              allow_hp);
        }
      } else if (mbmi->mode == NEAREST_NEWMV || mbmi->mode == NEAR_NEWMV) {
        const int ref = 1;
        const int_mv ref_mv = av1_get_ref_mv(x, ref);
        av1_update_mv_stats(&mbmi->mv[ref].as_mv, &ref_mv.as_mv, &fc->nmvc,
                            allow_hp);
      } else if (mbmi->mode == NEW_NEARESTMV || mbmi->mode == NEW_NEARMV) {
        const int ref = 0;
        const int_mv ref_mv = av1_get_ref_mv(x, ref);
        av1_update_mv_stats(&mbmi->mv[ref].as_mv, &ref_mv.as_mv, &fc->nmvc,
                            allow_hp);
      }
    }
  }
}

/*!\brief Reconstructs an individual coding block
 *
 * \ingroup partition_search
 * Reconstructs an individual coding block by applying the chosen modes stored
 * in ctx, also updates mode counts and entropy models.
 *
 * \param[in]    cpi       Top-level encoder structure
 * \param[in]    tile_data Pointer to struct holding adaptive
 *                         data/contexts/models for the tile during encoding
 * \param[in]    td        Pointer to thread data
 * \param[in]    tp        Pointer to the starting token
 * \param[in]    mi_row    Row coordinate of the block in a step size of MI_SIZE
 * \param[in]    mi_col    Column coordinate of the block in a step size of
 *                         MI_SIZE
 * \param[in]    dry_run   A code indicating whether it is part of the final
 *                         pass for reconstructing the superblock
 * \param[in]    bsize     Current block size
 * \param[in]    partition Partition mode of the parent block
 * \param[in]    ctx       Pointer to structure holding coding contexts and the
 *                         chosen modes for the current block
 * \param[in]    rate      Pointer to the total rate for the current block
 *
 * \return Nothing is returned. Instead, reconstructions (w/o in-loop filters)
 * will be updated in the pixel buffers in td->mb.e_mbd. Also, the chosen modes
 * will be stored in the MB_MODE_INFO buffer td->mb.e_mbd.mi[0].
 */
static void encode_b(const AV1_COMP *const cpi, TileDataEnc *tile_data,
                     ThreadData *td, TokenExtra **tp, int mi_row, int mi_col,
                     RUN_TYPE dry_run, BLOCK_SIZE bsize,
                     PARTITION_TYPE partition, PICK_MODE_CONTEXT *const ctx,
                     int *rate) {
  const AV1_COMMON *const cm = &cpi->common;
  TileInfo *const tile = &tile_data->tile_info;
  MACROBLOCK *const x = &td->mb;
  MACROBLOCKD *xd = &x->e_mbd;
  const int subsampling_x = cm->seq_params->subsampling_x;
  const int subsampling_y = cm->seq_params->subsampling_y;

  av1_set_offsets_without_segment_id(cpi, tile, x, mi_row, mi_col, bsize);
  const int origin_mult = x->rdmult;
  setup_block_rdmult(cpi, x, mi_row, mi_col, bsize, NO_AQ, NULL);
  MB_MODE_INFO *mbmi = xd->mi[0];
  mbmi->partition = partition;
  av1_update_state(cpi, td, ctx, mi_row, mi_col, bsize, dry_run);

  if (!dry_run) {
    set_cb_offsets(x->mbmi_ext_frame->cb_offset, x->cb_offset[PLANE_TYPE_Y],
                   x->cb_offset[PLANE_TYPE_UV]);
    assert(x->cb_offset[PLANE_TYPE_Y] <
           (1 << num_pels_log2_lookup[cpi->common.seq_params->sb_size]));
    assert(x->cb_offset[PLANE_TYPE_UV] <
           ((1 << num_pels_log2_lookup[cpi->common.seq_params->sb_size]) >>
            (subsampling_x + subsampling_y)));
  }

  encode_superblock(cpi, tile_data, td, tp, dry_run, bsize, rate);

  if (!dry_run) {
    update_cb_offsets(x, bsize, subsampling_x, subsampling_y);
    if (bsize == cpi->common.seq_params->sb_size && mbmi->skip_txfm == 1 &&
        cm->delta_q_info.delta_lf_present_flag) {
      const int frame_lf_count =
          av1_num_planes(cm) > 1 ? FRAME_LF_COUNT : FRAME_LF_COUNT - 2;
      for (int lf_id = 0; lf_id < frame_lf_count; ++lf_id)
        mbmi->delta_lf[lf_id] = xd->delta_lf[lf_id];
      mbmi->delta_lf_from_base = xd->delta_lf_from_base;
    }
    if (has_second_ref(mbmi)) {
      if (mbmi->compound_idx == 0 ||
          mbmi->interinter_comp.type == COMPOUND_AVERAGE)
        mbmi->comp_group_idx = 0;
      else
        mbmi->comp_group_idx = 1;
    }

    // delta quant applies to both intra and inter
    const int super_block_upper_left =
        ((mi_row & (cm->seq_params->mib_size - 1)) == 0) &&
        ((mi_col & (cm->seq_params->mib_size - 1)) == 0);
    const DeltaQInfo *const delta_q_info = &cm->delta_q_info;
    if (delta_q_info->delta_q_present_flag &&
        (bsize != cm->seq_params->sb_size || !mbmi->skip_txfm) &&
        super_block_upper_left) {
      xd->current_base_qindex = mbmi->current_qindex;
      if (delta_q_info->delta_lf_present_flag) {
        if (delta_q_info->delta_lf_multi) {
          const int frame_lf_count =
              av1_num_planes(cm) > 1 ? FRAME_LF_COUNT : FRAME_LF_COUNT - 2;
          for (int lf_id = 0; lf_id < frame_lf_count; ++lf_id) {
            xd->delta_lf[lf_id] = mbmi->delta_lf[lf_id];
          }
        } else {
          xd->delta_lf_from_base = mbmi->delta_lf_from_base;
        }
      }
    }

    RD_COUNTS *rdc = &td->rd_counts;
    if (mbmi->skip_mode) {
      assert(!frame_is_intra_only(cm));
      rdc->skip_mode_used_flag = 1;
      if (cm->current_frame.reference_mode == REFERENCE_MODE_SELECT) {
        assert(has_second_ref(mbmi));
        rdc->compound_ref_used_flag = 1;
      }
      set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
    } else {
      const int seg_ref_active =
          segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_REF_FRAME);
      if (!seg_ref_active) {
        // If the segment reference feature is enabled we have only a single
        // reference frame allowed for the segment so exclude it from
        // the reference frame counts used to work out probabilities.
        if (is_inter_block(mbmi)) {
          av1_collect_neighbors_ref_counts(xd);
          if (cm->current_frame.reference_mode == REFERENCE_MODE_SELECT) {
            if (has_second_ref(mbmi)) {
              // This flag is also updated for 4x4 blocks
              rdc->compound_ref_used_flag = 1;
            }
          }
          set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
        }
      }
    }

    if (tile_data->allow_update_cdf) update_stats(&cpi->common, td);

    // Gather obmc and warped motion count to update the probability.
    if ((cpi->sf.inter_sf.prune_obmc_prob_thresh > 0 &&
         cpi->sf.inter_sf.prune_obmc_prob_thresh < INT_MAX) ||
        (cm->features.allow_warped_motion &&
         cpi->sf.inter_sf.prune_warped_prob_thresh > 0)) {
      const int inter_block = is_inter_block(mbmi);
      const int seg_ref_active =
          segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_REF_FRAME);
      if (!seg_ref_active && inter_block) {
        const MOTION_MODE motion_allowed =
            cm->features.switchable_motion_mode
                ? motion_mode_allowed(xd->global_motion, xd, mbmi,
                                      cm->features.allow_warped_motion)
                : SIMPLE_TRANSLATION;

        if (mbmi->ref_frame[1] != INTRA_FRAME) {
          if (motion_allowed >= OBMC_CAUSAL) {
            td->rd_counts.obmc_used[bsize][mbmi->motion_mode == OBMC_CAUSAL]++;
          }
          if (motion_allowed == WARPED_CAUSAL) {
            td->rd_counts.warped_used[mbmi->motion_mode == WARPED_CAUSAL]++;
          }
        }
      }
    }
  }
  // TODO(Ravi/Remya): Move this copy function to a better logical place
  // This function will copy the best mode information from block
  // level (x->mbmi_ext) to frame level (cpi->mbmi_ext_info.frame_base). This
  // frame level buffer (cpi->mbmi_ext_info.frame_base) will be used during
  // bitstream preparation.
  av1_copy_mbmi_ext_to_mbmi_ext_frame(x->mbmi_ext_frame, &x->mbmi_ext,
                                      av1_ref_frame_type(xd->mi[0]->ref_frame));
  x->rdmult = origin_mult;
}

/*!\brief Reconstructs a partition (may contain multiple coding blocks)
 *
 * \ingroup partition_search
 * Reconstructs a sub-partition of the superblock by applying the chosen modes
 * and partition trees stored in pc_tree.
 *
 * \param[in]    cpi       Top-level encoder structure
 * \param[in]    td        Pointer to thread data
 * \param[in]    tile_data Pointer to struct holding adaptive
 *                         data/contexts/models for the tile during encoding
 * \param[in]    tp        Pointer to the starting token
 * \param[in]    mi_row    Row coordinate of the block in a step size of MI_SIZE
 * \param[in]    mi_col    Column coordinate of the block in a step size of
 *                         MI_SIZE
 * \param[in]    dry_run   A code indicating whether it is part of the final
 *                         pass for reconstructing the superblock
 * \param[in]    bsize     Current block size
 * \param[in]    pc_tree   Pointer to the PC_TREE node storing the picked
 *                         partitions and mode info for the current block
 * \param[in]    rate      Pointer to the total rate for the current block
 *
 * \return Nothing is returned. Instead, reconstructions (w/o in-loop filters)
 * will be updated in the pixel buffers in td->mb.e_mbd.
 */
static void encode_sb(const AV1_COMP *const cpi, ThreadData *td,
                      TileDataEnc *tile_data, TokenExtra **tp, int mi_row,
                      int mi_col, RUN_TYPE dry_run, BLOCK_SIZE bsize,
                      PC_TREE *pc_tree, int *rate) {
  assert(bsize < BLOCK_SIZES_ALL);
  const AV1_COMMON *const cm = &cpi->common;
  const CommonModeInfoParams *const mi_params = &cm->mi_params;
  MACROBLOCK *const x = &td->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  assert(bsize < BLOCK_SIZES_ALL);
  const int hbs = mi_size_wide[bsize] / 2;
  const int is_partition_root = bsize >= BLOCK_8X8;
  const int ctx = is_partition_root
                      ? partition_plane_context(xd, mi_row, mi_col, bsize)
                      : -1;
  const PARTITION_TYPE partition = pc_tree->partitioning;
  const BLOCK_SIZE subsize = get_partition_subsize(bsize, partition);
  int quarter_step = mi_size_wide[bsize] / 4;
  int i;
  BLOCK_SIZE bsize2 = get_partition_subsize(bsize, PARTITION_SPLIT);

  if (mi_row >= mi_params->mi_rows || mi_col >= mi_params->mi_cols) return;
  if (subsize == BLOCK_INVALID) return;

  if (!dry_run && ctx >= 0) {
    const int has_rows = (mi_row + hbs) < mi_params->mi_rows;
    const int has_cols = (mi_col + hbs) < mi_params->mi_cols;

    if (has_rows && has_cols) {
#if CONFIG_ENTROPY_STATS
      td->counts->partition[ctx][partition]++;
#endif

      if (tile_data->allow_update_cdf) {
        FRAME_CONTEXT *fc = xd->tile_ctx;
        update_cdf(fc->partition_cdf[ctx], partition,
                   partition_cdf_length(bsize));
      }
    }
  }

  switch (partition) {
    case PARTITION_NONE:
      encode_b(cpi, tile_data, td, tp, mi_row, mi_col, dry_run, subsize,
               partition, pc_tree->none, rate);
      break;
    case PARTITION_VERT:
      encode_b(cpi, tile_data, td, tp, mi_row, mi_col, dry_run, subsize,
               partition, pc_tree->vertical[0], rate);
      if (mi_col + hbs < mi_params->mi_cols) {
        encode_b(cpi, tile_data, td, tp, mi_row, mi_col + hbs, dry_run, subsize,
                 partition, pc_tree->vertical[1], rate);
      }
      break;
    case PARTITION_HORZ:
      encode_b(cpi, tile_data, td, tp, mi_row, mi_col, dry_run, subsize,
               partition, pc_tree->horizontal[0], rate);
      if (mi_row + hbs < mi_params->mi_rows) {
        encode_b(cpi, tile_data, td, tp, mi_row + hbs, mi_col, dry_run, subsize,
                 partition, pc_tree->horizontal[1], rate);
      }
      break;
    case PARTITION_SPLIT:
      encode_sb(cpi, td, tile_data, tp, mi_row, mi_col, dry_run, subsize,
                pc_tree->split[0], rate);
      encode_sb(cpi, td, tile_data, tp, mi_row, mi_col + hbs, dry_run, subsize,
                pc_tree->split[1], rate);
      encode_sb(cpi, td, tile_data, tp, mi_row + hbs, mi_col, dry_run, subsize,
                pc_tree->split[2], rate);
      encode_sb(cpi, td, tile_data, tp, mi_row + hbs, mi_col + hbs, dry_run,
                subsize, pc_tree->split[3], rate);
      break;

    case PARTITION_HORZ_A:
      encode_b(cpi, tile_data, td, tp, mi_row, mi_col, dry_run, bsize2,
               partition, pc_tree->horizontala[0], rate);
      encode_b(cpi, tile_data, td, tp, mi_row, mi_col + hbs, dry_run, bsize2,
               partition, pc_tree->horizontala[1], rate);
      encode_b(cpi, tile_data, td, tp, mi_row + hbs, mi_col, dry_run, subsize,
               partition, pc_tree->horizontala[2], rate);
      break;
    case PARTITION_HORZ_B:
      encode_b(cpi, tile_data, td, tp, mi_row, mi_col, dry_run, subsize,
               partition, pc_tree->horizontalb[0], rate);
      encode_b(cpi, tile_data, td, tp, mi_row + hbs, mi_col, dry_run, bsize2,
               partition, pc_tree->horizontalb[1], rate);
      encode_b(cpi, tile_data, td, tp, mi_row + hbs, mi_col + hbs, dry_run,
               bsize2, partition, pc_tree->horizontalb[2], rate);
      break;
    case PARTITION_VERT_A:
      encode_b(cpi, tile_data, td, tp, mi_row, mi_col, dry_run, bsize2,
               partition, pc_tree->verticala[0], rate);
      encode_b(cpi, tile_data, td, tp, mi_row + hbs, mi_col, dry_run, bsize2,
               partition, pc_tree->verticala[1], rate);
      encode_b(cpi, tile_data, td, tp, mi_row, mi_col + hbs, dry_run, subsize,
               partition, pc_tree->verticala[2], rate);

      break;
    case PARTITION_VERT_B:
      encode_b(cpi, tile_data, td, tp, mi_row, mi_col, dry_run, subsize,
               partition, pc_tree->verticalb[0], rate);
      encode_b(cpi, tile_data, td, tp, mi_row, mi_col + hbs, dry_run, bsize2,
               partition, pc_tree->verticalb[1], rate);
      encode_b(cpi, tile_data, td, tp, mi_row + hbs, mi_col + hbs, dry_run,
               bsize2, partition, pc_tree->verticalb[2], rate);
      break;
    case PARTITION_HORZ_4:
      for (i = 0; i < SUB_PARTITIONS_PART4; ++i) {
        int this_mi_row = mi_row + i * quarter_step;
        if (i > 0 && this_mi_row >= mi_params->mi_rows) break;

        encode_b(cpi, tile_data, td, tp, this_mi_row, mi_col, dry_run, subsize,
                 partition, pc_tree->horizontal4[i], rate);
      }
      break;
    case PARTITION_VERT_4:
      for (i = 0; i < SUB_PARTITIONS_PART4; ++i) {
        int this_mi_col = mi_col + i * quarter_step;
        if (i > 0 && this_mi_col >= mi_params->mi_cols) break;
        encode_b(cpi, tile_data, td, tp, mi_row, this_mi_col, dry_run, subsize,
                 partition, pc_tree->vertical4[i], rate);
      }
      break;
    default: assert(0 && "Invalid partition type."); break;
  }

  update_ext_partition_context(xd, mi_row, mi_col, subsize, bsize, partition);
}

static AOM_INLINE int is_adjust_var_based_part_enabled(
    AV1_COMMON *const cm, const PARTITION_SPEED_FEATURES *const part_sf,
    BLOCK_SIZE bsize) {
  if (part_sf->partition_search_type != VAR_BASED_PARTITION) return 0;
  if (part_sf->adjust_var_based_rd_partitioning == 0 ||
      part_sf->adjust_var_based_rd_partitioning > 3)
    return 0;

  if (part_sf->adjust_var_based_rd_partitioning == 1) {
    return bsize <= BLOCK_32X32;
  } else {
    if (bsize <= BLOCK_32X32) return 1;
    const int is_larger_qindex = cm->quant_params.base_qindex > 190;
    if (part_sf->adjust_var_based_rd_partitioning == 2) {
      const int is_360p_or_larger = AOMMIN(cm->width, cm->height) >= 360;
      return is_360p_or_larger && is_larger_qindex && bsize == BLOCK_64X64;
    }
  }
  return 0;
}

/*!\brief AV1 block partition search (partition estimation and partial search).
*
* \ingroup partition_search
* Encode the block by applying pre-calculated partition patterns that are
* represented by coding block sizes stored in the mbmi array. Minor partition
* adjustments are tested and applied if they lead to lower rd costs. The
* partition types are limited to a basic set: none, horz, vert, and split.
*
* \param[in]    cpi       Top-level encoder structure
* \param[in]    td        Pointer to thread data
* \param[in]    tile_data Pointer to struct holding adaptive
data/contexts/models for the tile during encoding
* \param[in]    mib       Array representing MB_MODE_INFO pointers for mi
blocks starting from the first pixel of the current
block
* \param[in]    tp        Pointer to the starting token
* \param[in]    mi_row    Row coordinate of the block in a step size of MI_SIZE
* \param[in]    mi_col    Column coordinate of the block in a step size of
MI_SIZE
* \param[in]    bsize     Current block size
* \param[in]    rate      Pointer to the final rate for encoding the current
block
* \param[in]    dist      Pointer to the final distortion of the current block
* \param[in]    do_recon  Whether the reconstruction function needs to be run,
either for finalizing a superblock or providing
reference for future sub-partitions
* \param[in]    pc_tree   Pointer to the PC_TREE node holding the picked
partitions and mode info for the current block
*
* \return Nothing is returned. The pc_tree struct is modified to store the
* picked partition and modes. The rate and dist are also updated with those
* corresponding to the best partition found.
*/
void av1_rd_use_partition(AV1_COMP *cpi, ThreadData *td, TileDataEnc *tile_data,
                          MB_MODE_INFO **mib, TokenExtra **tp, int mi_row,
                          int mi_col, BLOCK_SIZE bsize, int *rate,
                          int64_t *dist, int do_recon, PC_TREE *pc_tree) {
  AV1_COMMON *const cm = &cpi->common;
  const CommonModeInfoParams *const mi_params = &cm->mi_params;
  const int num_planes = av1_num_planes(cm);
  TileInfo *const tile_info = &tile_data->tile_info;
  MACROBLOCK *const x = &td->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  const ModeCosts *mode_costs = &x->mode_costs;
  const int bs = mi_size_wide[bsize];
  const int hbs = bs / 2;
  const int pl = (bsize >= BLOCK_8X8)
                     ? partition_plane_context(xd, mi_row, mi_col, bsize)
                     : 0;
  const PARTITION_TYPE partition =
      (bsize >= BLOCK_8X8) ? get_partition(cm, mi_row, mi_col, bsize)
                           : PARTITION_NONE;
  const BLOCK_SIZE subsize = get_partition_subsize(bsize, partition);
  RD_SEARCH_MACROBLOCK_CONTEXT x_ctx;
  RD_STATS last_part_rdc, none_rdc, chosen_rdc, invalid_rdc;
  BLOCK_SIZE sub_subsize = BLOCK_4X4;
  int splits_below = 0;
  BLOCK_SIZE bs_type = mib[0]->bsize;
  x->try_merge_partition = 0;

  if (pc_tree->none == NULL) {
    pc_tree->none = av1_alloc_pmc(cpi, bsize, &td->shared_coeff_buf);
  }
  PICK_MODE_CONTEXT *ctx_none = pc_tree->none;

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

  assert(mi_size_wide[bsize] == mi_size_high[bsize]);

  av1_invalid_rd_stats(&last_part_rdc);
  av1_invalid_rd_stats(&none_rdc);
  av1_invalid_rd_stats(&chosen_rdc);
  av1_invalid_rd_stats(&invalid_rdc);

  pc_tree->partitioning = partition;

  xd->above_txfm_context =
      cm->above_contexts.txfm[tile_info->tile_row] + mi_col;
  xd->left_txfm_context =
      xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);
  av1_save_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes);

  if (bsize == BLOCK_16X16 && cpi->vaq_refresh) {
    av1_set_offsets(cpi, tile_info, x, mi_row, mi_col, bsize);
    x->mb_energy = av1_log_block_var(cpi, x, bsize);
  }

  // Save rdmult before it might be changed, so it can be restored later.
  const int orig_rdmult = x->rdmult;
  setup_block_rdmult(cpi, x, mi_row, mi_col, bsize, NO_AQ, NULL);

  if (is_adjust_var_based_part_enabled(cm, &cpi->sf.part_sf, bsize)) {
    // Check if any of the sub blocks are further split.
    if (partition == PARTITION_SPLIT && subsize > BLOCK_8X8) {
      sub_subsize = get_partition_subsize(subsize, PARTITION_SPLIT);
      splits_below = 1;
      for (int i = 0; i < SUB_PARTITIONS_SPLIT; i++) {
        int jj = i >> 1, ii = i & 0x01;
        MB_MODE_INFO *this_mi = mib[jj * hbs * mi_params->mi_stride + ii * hbs];
        if (this_mi && this_mi->bsize >= 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 + hbs < mi_params->mi_rows &&
        mi_col + hbs < mi_params->mi_cols) {
      pc_tree->partitioning = PARTITION_NONE;
      x->try_merge_partition = 1;
      pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &none_rdc,
                    PARTITION_NONE, bsize, ctx_none, invalid_rdc);

      if (none_rdc.rate < INT_MAX) {
        none_rdc.rate += mode_costs->partition_cost[pl][PARTITION_NONE];
        none_rdc.rdcost = RDCOST(x->rdmult, none_rdc.rate, none_rdc.dist);
      }

      av1_restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes);
      mib[0]->bsize = bs_type;
      pc_tree->partitioning = partition;
    }
  }

  for (int i = 0; i < SUB_PARTITIONS_SPLIT; ++i) {
    pc_tree->split[i] = av1_alloc_pc_tree_node(subsize);
    pc_tree->split[i]->index = i;
  }
  switch (partition) {
    case PARTITION_NONE:
      pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
                    PARTITION_NONE, bsize, ctx_none, invalid_rdc);
      break;
    case PARTITION_HORZ:
      for (int i = 0; i < SUB_PARTITIONS_RECT; ++i) {
        pc_tree->horizontal[i] =
            av1_alloc_pmc(cpi, subsize, &td->shared_coeff_buf);
      }
      pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
                    PARTITION_HORZ, subsize, pc_tree->horizontal[0],
                    invalid_rdc);
      if (last_part_rdc.rate != INT_MAX && bsize >= BLOCK_8X8 &&
          mi_row + hbs < mi_params->mi_rows) {
        RD_STATS tmp_rdc;
        const PICK_MODE_CONTEXT *const ctx_h = pc_tree->horizontal[0];
        av1_init_rd_stats(&tmp_rdc);
        av1_update_state(cpi, td, ctx_h, mi_row, mi_col, subsize, 1);
        encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, subsize,
                          NULL);
        pick_sb_modes(cpi, tile_data, x, mi_row + hbs, mi_col, &tmp_rdc,
                      PARTITION_HORZ, subsize, pc_tree->horizontal[1],
                      invalid_rdc);
        if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
          av1_invalid_rd_stats(&last_part_rdc);
          break;
        }
        last_part_rdc.rate += tmp_rdc.rate;
        last_part_rdc.dist += tmp_rdc.dist;
        last_part_rdc.rdcost += tmp_rdc.rdcost;
      }
      break;
    case PARTITION_VERT:
      for (int i = 0; i < SUB_PARTITIONS_RECT; ++i) {
        pc_tree->vertical[i] =
            av1_alloc_pmc(cpi, subsize, &td->shared_coeff_buf);
      }
      pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &last_part_rdc,
                    PARTITION_VERT, subsize, pc_tree->vertical[0], invalid_rdc);
      if (last_part_rdc.rate != INT_MAX && bsize >= BLOCK_8X8 &&
          mi_col + hbs < mi_params->mi_cols) {
        RD_STATS tmp_rdc;
        const PICK_MODE_CONTEXT *const ctx_v = pc_tree->vertical[0];
        av1_init_rd_stats(&tmp_rdc);
        av1_update_state(cpi, td, ctx_v, mi_row, mi_col, subsize, 1);
        encode_superblock(cpi, tile_data, td, tp, DRY_RUN_NORMAL, subsize,
                          NULL);
        pick_sb_modes(cpi, tile_data, x, mi_row, mi_col + hbs, &tmp_rdc,
                      PARTITION_VERT, subsize,
                      pc_tree->vertical[bsize > BLOCK_8X8], invalid_rdc);
        if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
          av1_invalid_rd_stats(&last_part_rdc);
          break;
        }
        last_part_rdc.rate += tmp_rdc.rate;
        last_part_rdc.dist += tmp_rdc.dist;
        last_part_rdc.rdcost += tmp_rdc.rdcost;
      }
      break;
    case PARTITION_SPLIT:
      if (none_rdc.rate < INT_MAX && none_rdc.skip_txfm == 1) {
        const MB_MODE_INFO *mbmi = xd->mi[0];
        // Try to skip split partition evaluation based on none partition
        // characteristics.
        if (cpi->sf.part_sf.adjust_var_based_rd_partitioning == 1 ||
            (cpi->sf.part_sf.adjust_var_based_rd_partitioning == 2 &&
             is_inter_block(mbmi) && mbmi->mode != NEWMV)) {
          av1_invalid_rd_stats(&last_part_rdc);
          break;
        }
      }

      last_part_rdc.rate = 0;
      last_part_rdc.dist = 0;
      last_part_rdc.rdcost = 0;
      for (int i = 0; i < SUB_PARTITIONS_SPLIT; i++) {
        int x_idx = (i & 1) * hbs;
        int y_idx = (i >> 1) * hbs;
        int jj = i >> 1, ii = i & 0x01;
        RD_STATS tmp_rdc;
        if ((mi_row + y_idx >= mi_params->mi_rows) ||
            (mi_col + x_idx >= mi_params->mi_cols))
          continue;

        av1_init_rd_stats(&tmp_rdc);
        av1_rd_use_partition(
            cpi, td, tile_data,
            mib + jj * hbs * mi_params->mi_stride + ii * hbs, tp,
            mi_row + y_idx, mi_col + x_idx, subsize, &tmp_rdc.rate,
            &tmp_rdc.dist, i != (SUB_PARTITIONS_SPLIT - 1), pc_tree->split[i]);
        if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
          av1_invalid_rd_stats(&last_part_rdc);
          break;
        }
        last_part_rdc.rate += tmp_rdc.rate;
        last_part_rdc.dist += tmp_rdc.dist;
      }
      break;
    case PARTITION_VERT_A:
    case PARTITION_VERT_B:
    case PARTITION_HORZ_A:
    case PARTITION_HORZ_B:
    case PARTITION_HORZ_4:
    case PARTITION_VERT_4:
      assert(0 && "Cannot handle extended partition types");
    default: assert(0); break;
  }

  if (last_part_rdc.rate < INT_MAX) {
    last_part_rdc.rate += mode_costs->partition_cost[pl][partition];
    last_part_rdc.rdcost =
        RDCOST(x->rdmult, last_part_rdc.rate, last_part_rdc.dist);
  }

  if ((cpi->sf.part_sf.partition_search_type == VAR_BASED_PARTITION &&
       cpi->sf.part_sf.adjust_var_based_rd_partitioning > 3) &&
      partition != PARTITION_SPLIT && bsize > BLOCK_8X8 &&
      (mi_row + bs < mi_params->mi_rows ||
       mi_row + hbs == mi_params->mi_rows) &&
      (mi_col + bs < mi_params->mi_cols ||
       mi_col + hbs == mi_params->mi_cols)) {
    BLOCK_SIZE split_subsize = get_partition_subsize(bsize, PARTITION_SPLIT);
    chosen_rdc.rate = 0;
    chosen_rdc.dist = 0;

    av1_restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes);
    pc_tree->partitioning = PARTITION_SPLIT;

    // Split partition.
    for (int i = 0; i < SUB_PARTITIONS_SPLIT; i++) {
      int x_idx = (i & 1) * hbs;
      int y_idx = (i >> 1) * hbs;
      RD_STATS tmp_rdc;

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

      av1_save_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes);
      pc_tree->split[i]->partitioning = PARTITION_NONE;
      if (pc_tree->split[i]->none == NULL)
        pc_tree->split[i]->none =
            av1_alloc_pmc(cpi, split_subsize, &td->shared_coeff_buf);
      pick_sb_modes(cpi, tile_data, x, mi_row + y_idx, mi_col + x_idx, &tmp_rdc,
                    PARTITION_SPLIT, split_subsize, pc_tree->split[i]->none,
                    invalid_rdc);

      av1_restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes);
      if (tmp_rdc.rate == INT_MAX || tmp_rdc.dist == INT64_MAX) {
        av1_invalid_rd_stats(&chosen_rdc);
        break;
      }

      chosen_rdc.rate += tmp_rdc.rate;
      chosen_rdc.dist += tmp_rdc.dist;

      if (i != SUB_PARTITIONS_SPLIT - 1)
        encode_sb(cpi, td, tile_data, tp, mi_row + y_idx, mi_col + x_idx,
                  OUTPUT_ENABLED, split_subsize, pc_tree->split[i], NULL);

      chosen_rdc.rate += mode_costs->partition_cost[pl][PARTITION_NONE];
    }
    if (chosen_rdc.rate < INT_MAX) {
      chosen_rdc.rate += mode_costs->partition_cost[pl][PARTITION_SPLIT];
      chosen_rdc.rdcost = RDCOST(x->rdmult, chosen_rdc.rate, chosen_rdc.dist);
    }
  }

  // If last_part is better set the partitioning to that.
  if (last_part_rdc.rdcost < chosen_rdc.rdcost) {
    mib[0]->bsize = bsize;
    if (bsize >= BLOCK_8X8) pc_tree->partitioning = partition;
    chosen_rdc = last_part_rdc;
  }
  // If none was better set the partitioning to that.
  if (none_rdc.rdcost < chosen_rdc.rdcost) {
    if (bsize >= BLOCK_8X8) pc_tree->partitioning = PARTITION_NONE;
    chosen_rdc = none_rdc;
  }

  av1_restore_context(x, &x_ctx, mi_row, mi_col, bsize, num_planes);

  // We must have chosen a partitioning and encoding or we'll fail later on.
  // No other opportunities for success.
  if (bsize == cm->seq_params->sb_size)
    assert(chosen_rdc.rate < INT_MAX && chosen_rdc.dist < INT64_MAX);

#if CONFIG_COLLECT_COMPONENT_TIMING
  start_timing(cpi, encode_sb_time);
#endif
  if (do_recon) {
    if (bsize == cm->seq_params->sb_size) {
      // NOTE: To get estimate for rate due to the tokens, use:
      // int rate_coeffs = 0;
      // encode_sb(cpi, td, tile_data, tp, mi_row, mi_col, DRY_RUN_COSTCOEFFS,
      //           bsize, pc_tree, &rate_coeffs);
      set_cb_offsets(x->cb_offset, 0, 0);
      encode_sb(cpi, td, tile_data, tp, mi_row, mi_col, OUTPUT_ENABLED, bsize,
                pc_tree, NULL);
    } else {
      encode_sb(cpi, td, tile_data, tp, mi_row, mi_col, DRY_RUN_NORMAL, bsize,
                pc_tree, NULL);
    }
  }
#if CONFIG_COLLECT_COMPONENT_TIMING
  end_timing(cpi, encode_sb_time);
#endif

  *rate = chosen_rdc.rate;
  *dist = chosen_rdc.dist;
  x->rdmult = orig_rdmult;
}

static void encode_b_nonrd(const AV1_COMP *const cpi, TileDataEnc *tile_data,
                           ThreadData *td, TokenExtra **tp, int mi_row,
                           int mi_col, RUN_TYPE dry_run, BLOCK_SIZE bsize,
                           PARTITION_TYPE partition,
                           PICK_MODE_CONTEXT *const ctx, int *rate) {
  const AV1_COMMON *const cm = &cpi->common;
  TileInfo *const tile = &tile_data->tile_info;
  MACROBLOCK *const x = &td->mb;
  MACROBLOCKD *xd = &x->e_mbd;
  av1_set_offsets_without_segment_id(cpi, tile, x, mi_row, mi_col, bsize);
  const int origin_mult = x->rdmult;
  setup_block_rdmult(cpi, x, mi_row, mi_col, bsize, NO_AQ, NULL);
  MB_MODE_INFO *mbmi = xd->mi[0];
  mbmi->partition = partition;
  av1_update_state(cpi, td, ctx, mi_row, mi_col, bsize, dry_run);
  const int subsampling_x = cpi->common.seq_params->subsampling_x;
  const int subsampling_y = cpi->common.seq_params->subsampling_y;
  if (!dry_run) {
    set_cb_offsets(x->mbmi_ext_frame->cb_offset, x->cb_offset[PLANE_TYPE_Y],
                   x->cb_offset[PLANE_TYPE_UV]);
    assert(x->cb_offset[PLANE_TYPE_Y] <
           (1 << num_pels_log2_lookup[cpi->common.seq_params->sb_size]));
    assert(x->cb_offset[PLANE_TYPE_UV] <
           ((1 << num_pels_log2_lookup[cpi->common.seq_params->sb_size]) >>
            (subsampling_x + subsampling_y)));
  }
  encode_superblock(cpi, tile_data, td, tp, dry_run, bsize, rate);
  if (!dry_run) {
    update_cb_offsets(x, bsize, subsampling_x, subsampling_y);
    if (has_second_ref(mbmi)) {
      if (mbmi->compound_idx == 0 ||
          mbmi->interinter_comp.type == COMPOUND_AVERAGE)
        mbmi->comp_group_idx = 0;
      else
        mbmi->comp_group_idx = 1;
      mbmi->compound_idx = 1;
    }
    RD_COUNTS *const rdc = &td->rd_counts;
    if (mbmi->skip_mode) {
      assert(!frame_is_intra_only(cm));
      rdc->skip_mode_used_flag = 1;
      if (cm->current_frame.reference_mode == REFERENCE_MODE_SELECT &&
          has_second_ref(mbmi)) {
        rdc->compound_ref_used_flag = 1;
      }
      set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
    } else {
      const int seg_ref_active =
          segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_REF_FRAME);
      if (!seg_ref_active) {
        // If the segment reference feature is enabled we have only a single
        // reference frame allowed for the segment so exclude it from
        // the reference frame counts used to work out probabilities.
        if (is_inter_block(mbmi)) {
          av1_collect_neighbors_ref_counts(xd);
          if (cm->current_frame.reference_mode == REFERENCE_MODE_SELECT &&
              has_second_ref(mbmi)) {
            // This flag is also updated for 4x4 blocks
            rdc->compound_ref_used_flag = 1;
          }
          set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);
        }
      }
    }
    if (tile_data->allow_update_cdf) update_stats(&cpi->common, td);
  }
  if (cpi->oxcf.q_cfg.aq_mode == CYCLIC_REFRESH_AQ && mbmi->skip_txfm &&
      !cpi->cyclic_refresh->skip_over4x4)
    av1_cyclic_reset_segment_skip(cpi, x, mi_row, mi_col, bsize);
  // TODO(Ravi/Remya): Move this copy function to a better logical place
  // This function will copy the best mode information from block
  // level (x->mbmi_ext) to frame level (cpi->mbmi_ext_info.frame_base). This
  // frame level buffer (cpi->mbmi_ext_info.frame_base) will be used during
  // bitstream preparation.
  av1_copy_mbmi_ext_to_mbmi_ext_frame(x->mbmi_ext_frame, &x->mbmi_ext,
                                      av1_ref_frame_type(xd->mi[0]->ref_frame));
  x->rdmult = origin_mult;
}

/*!\brief Top level function to pick block mode for non-RD optimized case
 *
 * \ingroup partition_search
 * \callgraph
 * \callergraph
 * Searches prediction modes, transform, and coefficient coding modes for an
 * individual coding block. This function is the top-level function that is
 * used for non-RD optimized mode search (controlled by
 * \c cpi->sf.rt_sf.use_nonrd_pick_mode). Depending on frame type it calls
 * inter/skip/hybrid-intra mode search functions
 *
 * \param[in]    cpi            Top-level encoder structure
 * \param[in]    tile_data      Pointer to struct holding adaptive
 *                              data/contexts/models for the tile during
 *                              encoding
 * \param[in]    x              Pointer to structure holding all the data for
 *                              the current macroblock
 * \param[in]    mi_row         Row coordinate of the block in a step size of
 *                              MI_SIZE
 * \param[in]    mi_col         Column coordinate of the block in a step size of
 *                              MI_SIZE
 * \param[in]    rd_cost        Pointer to structure holding rate and distortion
 *                              stats for the current block
 * \param[in]    bsize          Current block size
 * \param[in]    ctx            Pointer to structure holding coding contexts and
 *                              chosen modes for the current block
 *
 * \return Nothing is returned. Instead, the chosen modes and contexts necessary
 * for reconstruction are stored in ctx, the rate-distortion stats are stored in
 * rd_cost. If no valid mode leading to rd_cost <= best_rd, the status will be
 * signalled by an INT64_MAX rd_cost->rdcost.
 */
static void pick_sb_modes_nonrd(AV1_COMP *const cpi, TileDataEnc *tile_data,
                                MACROBLOCK *const x, int mi_row, int mi_col,
                                RD_STATS *rd_cost, BLOCK_SIZE bsize,
                                PICK_MODE_CONTEXT *ctx) {
  av1_set_offsets(cpi, &tile_data->tile_info, x, mi_row, mi_col, bsize);
  AV1_COMMON *const cm = &cpi->common;
  const int num_planes = av1_num_planes(cm);
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *mbmi = xd->mi[0];
  struct macroblock_plane *const p = x->plane;
  struct macroblockd_plane *const pd = xd->plane;
  const AQ_MODE aq_mode = cpi->oxcf.q_cfg.aq_mode;
  TxfmSearchInfo *txfm_info = &x->txfm_search_info;
  int i;

  // This is only needed for real time row-mt enabled multi-threaded encoding
  // with cost update frequency set to COST_UPD_TILE/COST_UPD_OFF.
  wait_for_top_right_sb(&cpi->mt_info.enc_row_mt, &tile_data->row_mt_sync,
                        &tile_data->tile_info, cm->seq_params->sb_size,
                        cm->seq_params->mib_size_log2, bsize, mi_row, mi_col);

#if CONFIG_COLLECT_COMPONENT_TIMING
  start_timing(cpi, rd_pick_sb_modes_time);
#endif
  // Sets up the tx_type_map buffer in MACROBLOCKD.
  xd->tx_type_map = txfm_info->tx_type_map_;
  xd->tx_type_map_stride = mi_size_wide[bsize];
  for (i = 0; i < num_planes; ++i) {
    p[i].coeff = ctx->coeff[i];
    p[i].qcoeff = ctx->qcoeff[i];
    p[i].dqcoeff = ctx->dqcoeff[i];
    p[i].eobs = ctx->eobs[i];
    p[i].txb_entropy_ctx = ctx->txb_entropy_ctx[i];
  }
  for (i = 0; i < 2; ++i) pd[i].color_index_map = ctx->color_index_map[i];
  if (is_cur_buf_hbd(xd)) {
    x->source_variance = av1_high_get_sby_perpixel_variance(
        cpi, &x->plane[0].src, bsize, xd->bd);
  } else {
    x->source_variance =
        av1_get_sby_perpixel_variance(cpi, &x->plane[0].src, bsize);
  }
  // Save rdmult before it might be changed, so it can be restored later.
  const int orig_rdmult = x->rdmult;
  setup_block_rdmult(cpi, x, mi_row, mi_col, bsize, aq_mode, mbmi);
  // Set error per bit for current rdmult
  av1_set_error_per_bit(&x->errorperbit, x->rdmult);
  // Find best coding mode & reconstruct the MB so it is available
  // as a predictor for MBs that follow in the SB
  if (frame_is_intra_only(cm)) {
#if CONFIG_COLLECT_COMPONENT_TIMING
    start_timing(cpi, av1_rd_pick_intra_mode_sb_time);
#endif
    hybrid_intra_mode_search(cpi, x, rd_cost, bsize, ctx);
#if CONFIG_COLLECT_COMPONENT_TIMING
    end_timing(cpi, av1_rd_pick_intra_mode_sb_time);
#endif
  } else {
#if CONFIG_COLLECT_COMPONENT_TIMING
    start_timing(cpi, av1_rd_pick_inter_mode_sb_time);
#endif
    if (segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
      RD_STATS invalid_rd;
      av1_invalid_rd_stats(&invalid_rd);
      // TODO(kyslov): add av1_nonrd_pick_inter_mode_sb_seg_skip
      av1_rd_pick_inter_mode_sb_seg_skip(cpi, tile_data, x, mi_row, mi_col,
                                         rd_cost, bsize, ctx,
                                         invalid_rd.rdcost);
    } else {
      av1_nonrd_pick_inter_mode_sb(cpi, tile_data, x, rd_cost, bsize, ctx);
    }
#if CONFIG_COLLECT_COMPONENT_TIMING
    end_timing(cpi, av1_rd_pick_inter_mode_sb_time);
#endif
  }
  if (cpi->sf.rt_sf.skip_cdef_sb) {
    // Find the corresponding 64x64 block. It'll be the 128x128 block if that's
    // the block size.
    const int mi_row_sb = mi_row - mi_row % MI_SIZE_64X64;
    const int mi_col_sb = mi_col - mi_col % MI_SIZE_64X64;
    MB_MODE_INFO **mi_sb =
        cm->mi_params.mi_grid_base +
        get_mi_grid_idx(&cm->mi_params, mi_row_sb, mi_col_sb);
    // Do not skip if intra or new mv is picked, or color sensitivity is set.
    mi_sb[0]->skip_cdef_curr_sb =
        mi_sb[0]->skip_cdef_curr_sb &&
        !(x->color_sensitivity[0] || x->color_sensitivity[1]) &&
        !(mbmi->mode < INTRA_MODES || mbmi->mode == NEWMV);
    // Store in the pickmode context.
    ctx->mic.skip_cdef_curr_sb = mi_sb[0]->skip_cdef_curr_sb;
  }
  x->rdmult = orig_rdmult;
  ctx->rd_stats.rate = rd_cost->rate;
  ctx->rd_stats.dist = rd_cost->dist;
  ctx->rd_stats.rdcost = rd_cost->rdcost;
#if CONFIG_COLLECT_COMPONENT_TIMING
  end_timing(cpi, rd_pick_sb_modes_time);
#endif
}

/*!\brief AV1 block partition application (minimal RD search).
*
* \ingroup partition_search
* \callgraph
* \callergraph
* Encode the block by applying pre-calculated partition patterns that are
* represented by coding block sizes stored in the mbmi array. The only
* partition adjustment allowed is merging leaf split nodes if it leads to a
* lower rd cost. The partition types are limited to a basic set: none, horz,
* vert, and split. This function is only used in the real-time mode.
*
* \param[in]    cpi       Top-level encoder structure
* \param[in]    td        Pointer to thread data
* \param[in]    tile_data Pointer to struct holding adaptive
data/contexts/models for the tile during encoding
* \param[in]    mib       Array representing MB_MODE_INFO pointers for mi
blocks starting from the first pixel of the current
block
* \param[in]    tp        Pointer to the starting token
* \param[in]    mi_row    Row coordinate of the block in a step size of MI_SIZE
* \param[in]    mi_col    Column coordinate of the block in a step size of
MI_SIZE
* \param[in]    bsize     Current block size
* \param[in]    pc_tree   Pointer to the PC_TREE node holding the picked
partitions and mode info for the current block
*
* \return Nothing is returned. The pc_tree struct is modified to store the
* picked partition and modes.
*/
void av1_nonrd_use_partition(AV1_COMP *cpi, ThreadData *td,
                             TileDataEnc *tile_data, MB_MODE_INFO **mib,
                             TokenExtra **tp, int mi_row, int mi_col,
                             BLOCK_SIZE bsize, PC_TREE *pc_tree) {
  AV1_COMMON *const cm = &cpi->common;
  const CommonModeInfoParams *const mi_params = &cm->mi_params;
  TileInfo *const tile_info = &tile_data->tile_info;
  MACROBLOCK *const x = &td->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  const ModeCosts *mode_costs = &x->mode_costs;
  // Only square blocks from 8x8 to 128x128 are supported
  assert(bsize >= BLOCK_8X8 && bsize <= BLOCK_128X128);
  const int bs = mi_size_wide[bsize];
  const int hbs = bs / 2;
  const PARTITION_TYPE partition =
      (bsize >= BLOCK_8X8) ? get_partition(cm, mi_row, mi_col, bsize)
                           : PARTITION_NONE;
  BLOCK_SIZE subsize = get_partition_subsize(bsize, partition);
  assert(subsize <= BLOCK_LARGEST);
  const int pl = (bsize >= BLOCK_8X8)
                     ? partition_plane_context(xd, mi_row, mi_col, bsize)
                     : 0;

  RD_STATS dummy_cost;
  av1_invalid_rd_stats(&dummy_cost);

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

  assert(mi_size_wide[bsize] == mi_size_high[bsize]);

  pc_tree->partitioning = partition;

  xd->above_txfm_context =
      cm->above_contexts.txfm[tile_info->tile_row] + mi_col;
  xd->left_txfm_context =
      xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);

  // Initialize default mode evaluation params
  set_mode_eval_params(cpi, x, DEFAULT_EVAL);

  switch (partition) {
    case PARTITION_NONE:
      pc_tree->none = av1_alloc_pmc(cpi, bsize, &td->shared_coeff_buf);
      if (cpi->sf.rt_sf.nonrd_check_partition_split && do_split_check(bsize) &&
          !frame_is_intra_only(cm)) {
        RD_STATS split_rdc, none_rdc, block_rdc;
        RD_SEARCH_MACROBLOCK_CONTEXT x_ctx;

        av1_init_rd_stats(&split_rdc);
        av1_invalid_rd_stats(&none_rdc);

        av1_save_context(x, &x_ctx, mi_row, mi_col, bsize, 3);
        subsize = get_partition_subsize(bsize, PARTITION_SPLIT);
        pick_sb_modes_nonrd(cpi, tile_data, x, mi_row, mi_col, &none_rdc, bsize,
                            pc_tree->none);
        none_rdc.rate += mode_costs->partition_cost[pl][PARTITION_NONE];
        none_rdc.rdcost = RDCOST(x->rdmult, none_rdc.rate, none_rdc.dist);
        av1_restore_context(x, &x_ctx, mi_row, mi_col, bsize, 3);

        for (int i = 0; i < SUB_PARTITIONS_SPLIT; i++) {
          av1_invalid_rd_stats(&block_rdc);
          const int x_idx = (i & 1) * hbs;
          const int y_idx = (i >> 1) * hbs;
          if (mi_row + y_idx >= mi_params->mi_rows ||
              mi_col + x_idx >= mi_params->mi_cols)
            continue;
          xd->above_txfm_context =
              cm->above_contexts.txfm[tile_info->tile_row] + mi_col + x_idx;
          xd->left_txfm_context =
              xd->left_txfm_context_buffer + ((mi_row + y_idx) & MAX_MIB_MASK);
          pc_tree->split[i]->partitioning = PARTITION_NONE;
          pick_sb_modes_nonrd(cpi, tile_data, x, mi_row + y_idx, mi_col + x_idx,
                              &block_rdc, subsize, pc_tree->split[i]->none);
          split_rdc.rate += block_rdc.rate;
          split_rdc.dist += block_rdc.dist;

          encode_b_nonrd(cpi, tile_data, td, tp, mi_row + y_idx, mi_col + x_idx,
                         1, subsize, PARTITION_NONE, pc_tree->split[i]->none,
                         NULL);
        }
        split_rdc.rate += mode_costs->partition_cost[pl][PARTITION_SPLIT];
        split_rdc.rdcost = RDCOST(x->rdmult, split_rdc.rate, split_rdc.dist);
        av1_restore_context(x, &x_ctx, mi_row, mi_col, bsize, 3);

        if (none_rdc.rdcost < split_rdc.rdcost) {
          mib[0]->bsize = bsize;
          pc_tree->partitioning = PARTITION_NONE;
          encode_b_nonrd(cpi, tile_data, td, tp, mi_row, mi_col, 0, bsize,
                         partition, pc_tree->none, NULL);
        } else {
          mib[0]->bsize = subsize;
          pc_tree->partitioning = PARTITION_SPLIT;
          for (int i = 0; i < SUB_PARTITIONS_SPLIT; i++) {
            const int x_idx = (i & 1) * hbs;
            const int y_idx = (i >> 1) * hbs;
            if (mi_row + y_idx >= mi_params->mi_rows ||
                mi_col + x_idx >= mi_params->mi_cols)
              continue;
            encode_b_nonrd(cpi, tile_data, td, tp, mi_row + y_idx,
                           mi_col + x_idx, 0, subsize, PARTITION_NONE,
                           pc_tree->split[i]->none, NULL);
          }
        }

      } else {
        pick_sb_modes_nonrd(cpi, tile_data, x, mi_row, mi_col, &dummy_cost,
                            bsize, pc_tree->none);
        encode_b_nonrd(cpi, tile_data, td, tp, mi_row, mi_col, 0, bsize,
                       partition, pc_tree->none, NULL);
      }
      break;
    case PARTITION_VERT:
      for (int i = 0; i < SUB_PARTITIONS_RECT; ++i) {
        pc_tree->vertical[i] =
            av1_alloc_pmc(cpi, subsize, &td->shared_coeff_buf);
      }
      pick_sb_modes_nonrd(cpi, tile_data, x, mi_row, mi_col, &dummy_cost,
                          subsize, pc_tree->vertical[0]);
      encode_b_nonrd(cpi, tile_data, td, tp, mi_row, mi_col, 0, subsize,
                     PARTITION_VERT, pc_tree->vertical[0], NULL);
      if (mi_col + hbs < mi_params->mi_cols && bsize > BLOCK_8X8) {
        pick_sb_modes_nonrd(cpi, tile_data, x, mi_row, mi_col + hbs,
                            &dummy_cost, subsize, pc_tree->vertical[1]);
        encode_b_nonrd(cpi, tile_data, td, tp, mi_row, mi_col + hbs, 0, subsize,
                       PARTITION_VERT, pc_tree->vertical[1], NULL);
      }
      break;
    case PARTITION_HORZ:
      for (int i = 0; i < SUB_PARTITIONS_RECT; ++i) {
        pc_tree->horizontal[i] =
            av1_alloc_pmc(cpi, subsize, &td->shared_coeff_buf);
      }
      pick_sb_modes_nonrd(cpi, tile_data, x, mi_row, mi_col, &dummy_cost,
                          subsize, pc_tree->horizontal[0]);
      encode_b_nonrd(cpi, tile_data, td, tp, mi_row, mi_col, 0, subsize,
                     PARTITION_HORZ, pc_tree->horizontal[0], NULL);

      if (mi_row + hbs < mi_params->mi_rows && bsize > BLOCK_8X8) {
        pick_sb_modes_nonrd(cpi, tile_data, x, mi_row + hbs, mi_col,
                            &dummy_cost, subsize, pc_tree->horizontal[1]);
        encode_b_nonrd(cpi, tile_data, td, tp, mi_row + hbs, mi_col, 0, subsize,
                       PARTITION_HORZ, pc_tree->horizontal[1], NULL);
      }
      break;
    case PARTITION_SPLIT:
      for (int i = 0; i < SUB_PARTITIONS_SPLIT; ++i) {
        pc_tree->split[i] = av1_alloc_pc_tree_node(subsize);
        pc_tree->split[i]->index = i;
      }
      if (cpi->sf.rt_sf.nonrd_check_partition_merge_mode &&
          av1_is_leaf_split_partition(cm, mi_row, mi_col, bsize) &&
          !frame_is_intra_only(cm) && bsize <= BLOCK_64X64) {
        RD_SEARCH_MACROBLOCK_CONTEXT x_ctx;
        RD_STATS split_rdc, none_rdc;
        av1_invalid_rd_stats(&split_rdc);
        av1_invalid_rd_stats(&none_rdc);
        av1_save_context(x, &x_ctx, mi_row, mi_col, bsize, 3);
        xd->above_txfm_context =
            cm->above_contexts.txfm[tile_info->tile_row] + mi_col;
        xd->left_txfm_context =
            xd->left_txfm_context_buffer + (mi_row & MAX_MIB_MASK);
        pc_tree->partitioning = PARTITION_NONE;
        pc_tree->none = av1_alloc_pmc(cpi, bsize, &td->shared_coeff_buf);
        pick_sb_modes_nonrd(cpi, tile_data, x, mi_row, mi_col, &none_rdc, bsize,
                            pc_tree->none);
        none_rdc.rate += mode_costs->partition_cost[pl][PARTITION_NONE];
        none_rdc.rdcost = RDCOST(x->rdmult, none_rdc.rate, none_rdc.dist);
        av1_restore_context(x, &x_ctx, mi_row, mi_col, bsize, 3);
        if (cpi->sf.rt_sf.nonrd_check_partition_merge_mode != 2 ||
            none_rdc.skip_txfm != 1 || pc_tree->none->mic.mode == NEWMV) {
          av1_init_rd_stats(&split_rdc);
          for (int i = 0; i < SUB_PARTITIONS_SPLIT; i++) {
            RD_STATS block_rdc;
            av1_invalid_rd_stats(&block_rdc);
            int x_idx = (i & 1) * hbs;
            int y_idx = (i >> 1) * hbs;
            if ((mi_row + y_idx >= mi_params->mi_rows) ||
                (mi_col + x_idx >= mi_params->mi_cols))
              continue;
            xd->above_txfm_context =
                cm->above_contexts.txfm[tile_info->tile_row] + mi_col + x_idx;
            xd->left_txfm_context = xd->left_txfm_context_buffer +
                                    ((mi_row + y_idx) & MAX_MIB_MASK);
            if (pc_tree->split[i]->none == NULL)
              pc_tree->split[i]->none =
                  av1_alloc_pmc(cpi, subsize, &td->shared_coeff_buf);
            pc_tree->split[i]->partitioning = PARTITION_NONE;
            pick_sb_modes_nonrd(cpi, tile_data, x, mi_row + y_idx,
                                mi_col + x_idx, &block_rdc, subsize,
                                pc_tree->split[i]->none);
            split_rdc.rate += block_rdc.rate;
            split_rdc.dist += block_rdc.dist;

            encode_b_nonrd(cpi, tile_data, td, tp, mi_row + y_idx,
                           mi_col + x_idx, 1, subsize, PARTITION_NONE,
                           pc_tree->split[i]->none, NULL);
          }
          av1_restore_context(x, &x_ctx, mi_row, mi_col, bsize, 3);
          split_rdc.rate += mode_costs->partition_cost[pl][PARTITION_SPLIT];
          split_rdc.rdcost = RDCOST(x->rdmult, split_rdc.rate, split_rdc.dist);
        }
        if (none_rdc.rdcost < split_rdc.rdcost) {
          mib[0]->bsize = bsize;
          pc_tree->partitioning = PARTITION_NONE;
          encode_b_nonrd(cpi, tile_data, td, tp, mi_row, mi_col, 0, bsize,
                         partition, pc_tree->none, NULL);
        } else {
          mib[0]->bsize = subsize;
          pc_tree->partitioning = PARTITION_SPLIT;
          for (int i = 0; i < SUB_PARTITIONS_SPLIT; i++) {
            int x_idx = (i & 1) * hbs;
            int y_idx = (i >> 1) * hbs;
            if ((mi_row + y_idx >= mi_params->mi_rows) ||
                (mi_col + x_idx >= mi_params->mi_cols))
              continue;

            if (pc_tree->split[i]->none == NULL)
              pc_tree->split[i]->none =
                  av1_alloc_pmc(cpi, subsize, &td->shared_coeff_buf);
            encode_b_nonrd(cpi, tile_data, td, tp, mi_row + y_idx,
                           mi_col + x_idx, 0, subsize, PARTITION_NONE,
                           pc_tree->split[i]->none, NULL);
          }
        }
      } else {
        for (int i = 0; i < SUB_PARTITIONS_SPLIT; i++) {
          int x_idx = (i & 1) * hbs;
          int y_idx = (i >> 1) * hbs;
          int jj = i >> 1, ii = i & 0x01;
          if ((mi_row + y_idx >= mi_params->mi_rows) ||
              (mi_col + x_idx >= mi_params->mi_cols))
            continue;
          av1_nonrd_use_partition(
              cpi, td, tile_data,
              mib + jj * hbs * mi_params->mi_stride + ii * hbs, tp,
              mi_row + y_idx, mi_col + x_idx, subsize, pc_tree->split[i]);
        }
      }
      break;
    case PARTITION_VERT_A:
    case PARTITION_VERT_B:
    case PARTITION_HORZ_A:
    case PARTITION_HORZ_B:
    case PARTITION_HORZ_4:
    case PARTITION_VERT_4:
      assert(0 && "Cannot handle extended partition types");
    default: assert(0); break;
  }
}

#if !CONFIG_REALTIME_ONLY
// Try searching for an encoding for the given subblock. Returns zero if the
// rdcost is already too high (to tell the caller not to bother searching for
// encodings of further subblocks).
static int rd_try_subblock(AV1_COMP *const cpi, ThreadData *td,
                           TileDataEnc *tile_data, TokenExtra **tp, int is_last,
                           int mi_row, int mi_col, BLOCK_SIZE subsize,
                           RD_STATS best_rdcost, RD_STATS *sum_rdc,
                           PARTITION_TYPE partition,
                           PICK_MODE_CONTEXT *this_ctx) {
  MACROBLOCK *const x = &td->mb;
  const int orig_mult = x->rdmult;
  setup_block_rdmult(cpi, x, mi_row, mi_col, subsize, NO_AQ, NULL);

  av1_rd_cost_update(x->rdmult, &best_rdcost);

  RD_STATS rdcost_remaining;
  av1_rd_stats_subtraction(x->rdmult, &best_rdcost, sum_rdc, &rdcost_remaining);
  RD_STATS this_rdc;
  pick_sb_modes(cpi, tile_data, x, mi_row, mi_col, &this_rdc, partition,
                subsize, this_ctx, rdcost_remaining);

  if (this_rdc.rate == INT_MAX) {
    sum_rdc->