vp9_encodeframe.c

  *rate = best_rate;
  *dist = best_dist;

  if (best_rate == INT_MAX)
    return;

  // update mode info array
  subsize = get_subsize(bsize, pc_tree->partitioning);
  fill_mode_info_sb(cm, x, mi_row, mi_col, bsize, subsize,
                    pc_tree);

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

    // Check the projected output rate for this SB against it's target
    // and and if necessary apply a Q delta using segmentation to get
    // closer to the target.
    if ((oxcf->aq_mode == COMPLEXITY_AQ) && cm->seg.update_map) {
      vp9_select_in_frame_q_segment(cpi, mi_row, mi_col, output_enabled,
                                    best_rate);
    }

    if (oxcf->aq_mode == CYCLIC_REFRESH_AQ)
      vp9_cyclic_refresh_set_rate_and_dist_sb(cpi->cyclic_refresh,
                                              best_rate, best_dist);

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

  if (bsize == BLOCK_64X64) {
    assert(tp_orig < *tp);
    assert(best_rate < INT_MAX);
    assert(best_dist < INT64_MAX);
  } else {
    assert(tp_orig == *tp);
  }
}

static void nonrd_use_partition(VP9_COMP *cpi,
                                const TileInfo *const tile,
                                MODE_INFO *mi,
                                TOKENEXTRA **tp,
                                int mi_row, int mi_col,
                                BLOCK_SIZE bsize, int output_enabled,
                                int *totrate, int64_t *totdist,
                                PC_TREE *pc_tree) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  const int bsl = b_width_log2_lookup[bsize], hbs = (1 << bsl) / 4;
  const int mis = cm->mi_stride;
  PARTITION_TYPE partition;
  BLOCK_SIZE subsize;
  int rate = INT_MAX;
  int64_t dist = INT64_MAX;

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

  subsize = (bsize >= BLOCK_8X8) ? mi[0].src_mi->mbmi.sb_type : BLOCK_4X4;
  partition = partition_lookup[bsl][subsize];

  switch (partition) {
    case PARTITION_NONE:
      nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col, totrate, totdist,
                          subsize, &pc_tree->none);
      pc_tree->none.mic.mbmi = xd->mi[0].src_mi->mbmi;
      pc_tree->none.skip_txfm[0] = x->skip_txfm[0];
      pc_tree->none.skip = x->skip;
      break;
    case PARTITION_VERT:
      nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col, totrate, totdist,
                          subsize, &pc_tree->vertical[0]);
      pc_tree->vertical[0].mic.mbmi = xd->mi[0].src_mi->mbmi;
      pc_tree->vertical[0].skip_txfm[0] = x->skip_txfm[0];
      pc_tree->vertical[0].skip = x->skip;
      if (mi_col + hbs < cm->mi_cols) {
        nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col + hbs,
                            &rate, &dist, subsize, &pc_tree->vertical[1]);
        pc_tree->vertical[1].mic.mbmi = xd->mi[0].src_mi->mbmi;
        pc_tree->vertical[1].skip_txfm[0] = x->skip_txfm[0];
        pc_tree->vertical[1].skip = x->skip;
        if (rate != INT_MAX && dist != INT64_MAX &&
            *totrate != INT_MAX && *totdist != INT64_MAX) {
          *totrate += rate;
          *totdist += dist;
        }
      }
      break;
    case PARTITION_HORZ:
      nonrd_pick_sb_modes(cpi, tile, mi_row, mi_col, totrate, totdist,
                          subsize, &pc_tree->horizontal[0]);
      pc_tree->horizontal[0].mic.mbmi = xd->mi[0].src_mi->mbmi;
      pc_tree->horizontal[0].skip_txfm[0] = x->skip_txfm[0];
      pc_tree->horizontal[0].skip = x->skip;
      if (mi_row + hbs < cm->mi_rows) {
        nonrd_pick_sb_modes(cpi, tile, mi_row + hbs, mi_col,
                            &rate, &dist, subsize, &pc_tree->horizontal[0]);
        pc_tree->horizontal[1].mic.mbmi = xd->mi[0].src_mi->mbmi;
        pc_tree->horizontal[1].skip_txfm[0] = x->skip_txfm[0];
        pc_tree->horizontal[1].skip = x->skip;
        if (rate != INT_MAX && dist != INT64_MAX &&
            *totrate != INT_MAX && *totdist != INT64_MAX) {
          *totrate += rate;
          *totdist += dist;
        }
      }
      break;
    case PARTITION_SPLIT:
      subsize = get_subsize(bsize, PARTITION_SPLIT);
      nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col,
                          subsize, output_enabled, totrate, totdist,
                          pc_tree->split[0]);
      nonrd_use_partition(cpi, tile, mi + hbs, tp,
                          mi_row, mi_col + hbs, subsize, output_enabled,
                          &rate, &dist, pc_tree->split[1]);
      if (rate != INT_MAX && dist != INT64_MAX &&
          *totrate != INT_MAX && *totdist != INT64_MAX) {
        *totrate += rate;
        *totdist += dist;
      }
      nonrd_use_partition(cpi, tile, mi + hbs * mis, tp,
                          mi_row + hbs, mi_col, subsize, output_enabled,
                          &rate, &dist, pc_tree->split[2]);
      if (rate != INT_MAX && dist != INT64_MAX &&
          *totrate != INT_MAX && *totdist != INT64_MAX) {
        *totrate += rate;
        *totdist += dist;
      }
      nonrd_use_partition(cpi, tile, mi + hbs * mis + hbs, tp,
                          mi_row + hbs, mi_col + hbs, subsize, output_enabled,
                          &rate, &dist, pc_tree->split[3]);
      if (rate != INT_MAX && dist != INT64_MAX &&
          *totrate != INT_MAX && *totdist != INT64_MAX) {
        *totrate += rate;
        *totdist += dist;
      }
      break;
    default:
      assert("Invalid partition type.");
      break;
  }

  if (bsize == BLOCK_64X64 && output_enabled) {
    if (cpi->oxcf.aq_mode == CYCLIC_REFRESH_AQ)
      vp9_cyclic_refresh_set_rate_and_dist_sb(cpi->cyclic_refresh,
                                              *totrate, *totdist);
    encode_sb_rt(cpi, tile, tp, mi_row, mi_col, 1, bsize, pc_tree);
  }
}

static void encode_nonrd_sb_row(VP9_COMP *cpi, const TileInfo *const tile,
                                int mi_row, TOKENEXTRA **tp) {
  SPEED_FEATURES *const sf = &cpi->sf;
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  int mi_col;

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

  // Code each SB in the row
  for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
       mi_col += MI_BLOCK_SIZE) {
    int dummy_rate = 0;
    int64_t dummy_dist = 0;
    const int idx_str = cm->mi_stride * mi_row + mi_col;
    MODE_INFO *mi = cm->mi + idx_str;
    MODE_INFO *prev_mi = (cm->prev_mip + cm->mi_stride + 1 + idx_str)->src_mi;
    BLOCK_SIZE bsize;
    x->in_static_area = 0;
    x->source_variance = UINT_MAX;
    vp9_zero(x->pred_mv);

    // Set the partition type of the 64X64 block
    switch (sf->partition_search_type) {
      case VAR_BASED_PARTITION:
        choose_partitioning(cpi, tile, mi_row, mi_col);
        nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
                            1, &dummy_rate, &dummy_dist, cpi->pc_root);
        break;
      case SOURCE_VAR_BASED_PARTITION:
        set_source_var_based_partition(cpi, tile, mi, mi_row, mi_col);
        nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
                            1, &dummy_rate, &dummy_dist, cpi->pc_root);
        break;
      case VAR_BASED_FIXED_PARTITION:
      case FIXED_PARTITION:
        bsize = sf->partition_search_type == FIXED_PARTITION ?
                sf->always_this_block_size :
                get_nonrd_var_based_fixed_partition(cpi, mi_row, mi_col);
        set_fixed_partitioning(cpi, tile, mi, mi_row, mi_col, bsize);
        nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col, BLOCK_64X64,
                            1, &dummy_rate, &dummy_dist, cpi->pc_root);
        break;
      case REFERENCE_PARTITION:
        if (sf->partition_check ||
            !(x->in_static_area = is_background(cpi, tile, mi_row, mi_col))) {
          set_modeinfo_offsets(cm, xd, mi_row, mi_col);
          auto_partition_range(cpi, tile, mi_row, mi_col,
                               &sf->min_partition_size,
                               &sf->max_partition_size);
          nonrd_pick_partition(cpi, tile, tp, mi_row, mi_col, BLOCK_64X64,
                               &dummy_rate, &dummy_dist, 1, INT64_MAX,
                               cpi->pc_root);
        } else {
          copy_partitioning(cm, mi, prev_mi);
          nonrd_use_partition(cpi, tile, mi, tp, mi_row, mi_col,
                              BLOCK_64X64, 1, &dummy_rate, &dummy_dist,
                              cpi->pc_root);
        }
        break;
      default:
        assert(0);
        break;
    }
  }
}
// end RTC play code

static int set_var_thresh_from_histogram(VP9_COMP *cpi) {
  const SPEED_FEATURES *const sf = &cpi->sf;
  const VP9_COMMON *const cm = &cpi->common;

  const uint8_t *src = cpi->Source->y_buffer;
  const uint8_t *last_src = cpi->Last_Source->y_buffer;
  const int src_stride = cpi->Source->y_stride;
  const int last_stride = cpi->Last_Source->y_stride;

  // Pick cutoff threshold
  const int cutoff = (MIN(cm->width, cm->height) >= 720) ?
      (cm->MBs * VAR_HIST_LARGE_CUT_OFF / 100) :
      (cm->MBs * VAR_HIST_SMALL_CUT_OFF / 100);
  DECLARE_ALIGNED_ARRAY(16, int, hist, VAR_HIST_BINS);
  diff *var16 = cpi->source_diff_var;

  int sum = 0;
  int i, j;

  vpx_memset(hist, 0, VAR_HIST_BINS * sizeof(hist[0]));

  for (i = 0; i < cm->mb_rows; i++) {
    for (j = 0; j < cm->mb_cols; j++) {
#if CONFIG_VP9_HIGHBITDEPTH
      if (cm->use_highbitdepth) {
        switch (cm->bit_depth) {
          case VPX_BITS_8:
            vp9_highbd_get16x16var(src, src_stride, last_src, last_stride,
                                   &var16->sse, &var16->sum);
            break;
          case VPX_BITS_10:
            vp9_highbd_10_get16x16var(src, src_stride, last_src, last_stride,
                                    &var16->sse, &var16->sum);
            break;
          case VPX_BITS_12:
            vp9_highbd_12_get16x16var(src, src_stride, last_src, last_stride,
                                      &var16->sse, &var16->sum);
            break;
          default:
            assert(0 && "cm->bit_depth should be VPX_BITS_8, VPX_BITS_10"
                   " or VPX_BITS_12");
            return -1;
        }
      } else {
        vp9_get16x16var(src, src_stride, last_src, last_stride,
                        &var16->sse, &var16->sum);
      }
#else
      vp9_get16x16var(src, src_stride, last_src, last_stride,
                      &var16->sse, &var16->sum);
#endif  // CONFIG_VP9_HIGHBITDEPTH
      var16->var = var16->sse -
          (((uint32_t)var16->sum * var16->sum) >> 8);

      if (var16->var >= VAR_HIST_MAX_BG_VAR)
        hist[VAR_HIST_BINS - 1]++;
      else
        hist[var16->var / VAR_HIST_FACTOR]++;

      src += 16;
      last_src += 16;
      var16++;
    }

    src = src - cm->mb_cols * 16 + 16 * src_stride;
    last_src = last_src - cm->mb_cols * 16 + 16 * last_stride;
  }

  cpi->source_var_thresh = 0;

  if (hist[VAR_HIST_BINS - 1] < cutoff) {
    for (i = 0; i < VAR_HIST_BINS - 1; i++) {
      sum += hist[i];

      if (sum > cutoff) {
        cpi->source_var_thresh = (i + 1) * VAR_HIST_FACTOR;
        return 0;
      }
    }
  }

  return sf->search_type_check_frequency;
}

static void source_var_based_partition_search_method(VP9_COMP *cpi) {
  VP9_COMMON *const cm = &cpi->common;
  SPEED_FEATURES *const sf = &cpi->sf;

  if (cm->frame_type == KEY_FRAME) {
    // For key frame, use SEARCH_PARTITION.
    sf->partition_search_type = SEARCH_PARTITION;
  } else if (cm->intra_only) {
    sf->partition_search_type = FIXED_PARTITION;
  } else {
    if (cm->last_width != cm->width || cm->last_height != cm->height) {
      if (cpi->source_diff_var)
        vpx_free(cpi->source_diff_var);

        CHECK_MEM_ERROR(cm, cpi->source_diff_var,
                        vpx_calloc(cm->MBs, sizeof(diff)));
      }

    if (!cpi->frames_till_next_var_check)
      cpi->frames_till_next_var_check = set_var_thresh_from_histogram(cpi);

    if (cpi->frames_till_next_var_check > 0) {
      sf->partition_search_type = FIXED_PARTITION;
      cpi->frames_till_next_var_check--;
    }
  }
}

static int get_skip_encode_frame(const VP9_COMMON *cm) {
  unsigned int intra_count = 0, inter_count = 0;
  int j;

  for (j = 0; j < INTRA_INTER_CONTEXTS; ++j) {
    intra_count += cm->counts.intra_inter[j][0];
    inter_count += cm->counts.intra_inter[j][1];
  }

  return (intra_count << 2) < inter_count &&
         cm->frame_type != KEY_FRAME &&
         cm->show_frame;
}

static void encode_tiles(VP9_COMP *cpi) {
  const VP9_COMMON *const cm = &cpi->common;
  const int tile_cols = 1 << cm->log2_tile_cols;
  const int tile_rows = 1 << cm->log2_tile_rows;
  int tile_col, tile_row;
  TOKENEXTRA *tok = cpi->tok;

  for (tile_row = 0; tile_row < tile_rows; ++tile_row) {
    for (tile_col = 0; tile_col < tile_cols; ++tile_col) {
      TileInfo tile;
      TOKENEXTRA *old_tok = tok;
      int mi_row;

      vp9_tile_init(&tile, cm, tile_row, tile_col);
      for (mi_row = tile.mi_row_start; mi_row < tile.mi_row_end;
           mi_row += MI_BLOCK_SIZE) {
        if (cpi->sf.use_nonrd_pick_mode && !frame_is_intra_only(cm))
          encode_nonrd_sb_row(cpi, &tile, mi_row, &tok);
        else
          encode_rd_sb_row(cpi, &tile, mi_row, &tok);
      }
      cpi->tok_count[tile_row][tile_col] = (unsigned int)(tok - old_tok);
      assert(tok - cpi->tok <= get_token_alloc(cm->mb_rows, cm->mb_cols));
    }
  }
}

#if CONFIG_FP_MB_STATS
static int input_fpmb_stats(FIRSTPASS_MB_STATS *firstpass_mb_stats,
                            VP9_COMMON *cm, uint8_t **this_frame_mb_stats) {
  uint8_t *mb_stats_in = firstpass_mb_stats->mb_stats_start +
      cm->current_video_frame * cm->MBs * sizeof(uint8_t);

  if (mb_stats_in > firstpass_mb_stats->mb_stats_end)
    return EOF;

  *this_frame_mb_stats = mb_stats_in;

  return 1;
}
#endif

static void encode_frame_internal(VP9_COMP *cpi) {
  SPEED_FEATURES *const sf = &cpi->sf;
  RD_OPT *const rd_opt = &cpi->rd;
  MACROBLOCK *const x = &cpi->mb;
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCKD *const xd = &x->e_mbd;

  xd->mi = cm->mi;
  xd->mi[0].src_mi = &xd->mi[0];

  vp9_zero(cm->counts);
  vp9_zero(cpi->coef_counts);
  vp9_zero(rd_opt->comp_pred_diff);
  vp9_zero(rd_opt->filter_diff);
  vp9_zero(rd_opt->tx_select_diff);
  vp9_zero(rd_opt->tx_select_threshes);

  xd->lossless = cm->base_qindex == 0 &&
                 cm->y_dc_delta_q == 0 &&
                 cm->uv_dc_delta_q == 0 &&
                 cm->uv_ac_delta_q == 0;

  cm->tx_mode = select_tx_mode(cpi);

#if CONFIG_VP9_HIGHBITDEPTH
  if (cm->use_highbitdepth)
    x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vp9_fdct4x4;
  else
    x->fwd_txm4x4 = xd->lossless ? vp9_highbd_fwht4x4 : vp9_highbd_fdct4x4;
  x->highbd_itxm_add = xd->lossless ? vp9_highbd_iwht4x4_add :
                                      vp9_highbd_idct4x4_add;
#else
  x->fwd_txm4x4 = xd->lossless ? vp9_fwht4x4 : vp9_fdct4x4;
#endif  // CONFIG_VP9_HIGHBITDEPTH
  x->itxm_add = xd->lossless ? vp9_iwht4x4_add : vp9_idct4x4_add;

  if (xd->lossless) {
    x->optimize = 0;
    cm->lf.filter_level = 0;
    cpi->zbin_mode_boost_enabled = 0;
  }

  vp9_frame_init_quantizer(cpi);

  vp9_initialize_rd_consts(cpi);
  vp9_initialize_me_consts(cpi, cm->base_qindex);
  init_encode_frame_mb_context(cpi);
  set_prev_mi(cm);

  x->quant_fp = cpi->sf.use_quant_fp;
  vp9_zero(x->skip_txfm);
  if (sf->use_nonrd_pick_mode) {
    // Initialize internal buffer pointers for rtc coding, where non-RD
    // mode decision is used and hence no buffer pointer swap needed.
    int i;
    struct macroblock_plane *const p = x->plane;
    struct macroblockd_plane *const pd = xd->plane;
    PICK_MODE_CONTEXT *ctx = &cpi->pc_root->none;

    for (i = 0; i < MAX_MB_PLANE; ++i) {
      p[i].coeff = ctx->coeff_pbuf[i][0];
      p[i].qcoeff = ctx->qcoeff_pbuf[i][0];
      pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][0];
      p[i].eobs = ctx->eobs_pbuf[i][0];
    }
    vp9_zero(x->zcoeff_blk);

    if (sf->partition_search_type == SOURCE_VAR_BASED_PARTITION)
      source_var_based_partition_search_method(cpi);
  }

  {
    struct vpx_usec_timer emr_timer;
    vpx_usec_timer_start(&emr_timer);

#if CONFIG_FP_MB_STATS
  if (cpi->use_fp_mb_stats) {
    input_fpmb_stats(&cpi->twopass.firstpass_mb_stats, cm,
                     &cpi->twopass.this_frame_mb_stats);
  }
#endif

    encode_tiles(cpi);

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

  sf->skip_encode_frame = sf->skip_encode_sb ? get_skip_encode_frame(cm) : 0;

#if 0
  // Keep record of the total distortion this time around for future use
  cpi->last_frame_distortion = cpi->frame_distortion;
#endif
}

static INTERP_FILTER get_interp_filter(
    const int64_t threshes[SWITCHABLE_FILTER_CONTEXTS], int is_alt_ref) {
  if (!is_alt_ref &&
      threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP] &&
      threshes[EIGHTTAP_SMOOTH] > threshes[EIGHTTAP_SHARP] &&
      threshes[EIGHTTAP_SMOOTH] > threshes[SWITCHABLE - 1]) {
    return EIGHTTAP_SMOOTH;
  } else if (threshes[EIGHTTAP_SHARP] > threshes[EIGHTTAP] &&
             threshes[EIGHTTAP_SHARP] > threshes[SWITCHABLE - 1]) {
    return EIGHTTAP_SHARP;
  } else if (threshes[EIGHTTAP] > threshes[SWITCHABLE - 1]) {
    return EIGHTTAP;
  } else {
    return SWITCHABLE;
  }
}

void vp9_encode_frame(VP9_COMP *cpi) {
  VP9_COMMON *const cm = &cpi->common;
  RD_OPT *const rd_opt = &cpi->rd;

  // In the longer term the encoder should be generalized to match the
  // decoder such that we allow compound where one of the 3 buffers has a
  // different sign bias and that buffer is then the fixed ref. However, this
  // requires further work in the rd loop. For now the only supported encoder
  // side behavior is where the ALT ref buffer has opposite sign bias to
  // the other two.
  if (!frame_is_intra_only(cm)) {
    if ((cm->ref_frame_sign_bias[ALTREF_FRAME] ==
             cm->ref_frame_sign_bias[GOLDEN_FRAME]) ||
        (cm->ref_frame_sign_bias[ALTREF_FRAME] ==
             cm->ref_frame_sign_bias[LAST_FRAME])) {
      cm->allow_comp_inter_inter = 0;
    } else {
      cm->allow_comp_inter_inter = 1;
      cm->comp_fixed_ref = ALTREF_FRAME;
      cm->comp_var_ref[0] = LAST_FRAME;
      cm->comp_var_ref[1] = GOLDEN_FRAME;
    }
  }

  if (cpi->sf.frame_parameter_update) {
    int i;

    // This code does a single RD pass over the whole frame assuming
    // either compound, single or hybrid prediction as per whatever has
    // worked best for that type of frame in the past.
    // It also predicts whether another coding mode would have worked
    // better that this coding mode. If that is the case, it remembers
    // that for subsequent frames.
    // It does the same analysis for transform size selection also.
    const MV_REFERENCE_FRAME frame_type = get_frame_type(cpi);
    int64_t *const mode_thrs = rd_opt->prediction_type_threshes[frame_type];
    int64_t *const filter_thrs = rd_opt->filter_threshes[frame_type];
    int *const tx_thrs = rd_opt->tx_select_threshes[frame_type];
    const int is_alt_ref = frame_type == ALTREF_FRAME;

    /* prediction (compound, single or hybrid) mode selection */
    if (is_alt_ref || !cm->allow_comp_inter_inter)
      cm->reference_mode = SINGLE_REFERENCE;
    else if (mode_thrs[COMPOUND_REFERENCE] > mode_thrs[SINGLE_REFERENCE] &&
             mode_thrs[COMPOUND_REFERENCE] >
                 mode_thrs[REFERENCE_MODE_SELECT] &&
             check_dual_ref_flags(cpi) &&
             cpi->static_mb_pct == 100)
      cm->reference_mode = COMPOUND_REFERENCE;
    else if (mode_thrs[SINGLE_REFERENCE] > mode_thrs[REFERENCE_MODE_SELECT])
      cm->reference_mode = SINGLE_REFERENCE;
    else
      cm->reference_mode = REFERENCE_MODE_SELECT;

    if (cm->interp_filter == SWITCHABLE)
      cm->interp_filter = get_interp_filter(filter_thrs, is_alt_ref);

    encode_frame_internal(cpi);

    for (i = 0; i < REFERENCE_MODES; ++i)
      mode_thrs[i] = (mode_thrs[i] + rd_opt->comp_pred_diff[i] / cm->MBs) / 2;

    for (i = 0; i < SWITCHABLE_FILTER_CONTEXTS; ++i)
      filter_thrs[i] = (filter_thrs[i] + rd_opt->filter_diff[i] / cm->MBs) / 2;

    for (i = 0; i < TX_MODES; ++i) {
      int64_t pd = rd_opt->tx_select_diff[i];
      if (i == TX_MODE_SELECT)
        pd -= RDCOST(cpi->mb.rdmult, cpi->mb.rddiv, 2048 * (TX_SIZES - 1), 0);
      tx_thrs[i] = (tx_thrs[i] + (int)(pd / cm->MBs)) / 2;
    }

    if (cm->reference_mode == REFERENCE_MODE_SELECT) {
      int single_count_zero = 0;
      int comp_count_zero = 0;

      for (i = 0; i < COMP_INTER_CONTEXTS; i++) {
        single_count_zero += cm->counts.comp_inter[i][0];
        comp_count_zero += cm->counts.comp_inter[i][1];
      }

      if (comp_count_zero == 0) {
        cm->reference_mode = SINGLE_REFERENCE;
        vp9_zero(cm->counts.comp_inter);
      } else if (single_count_zero == 0) {
        cm->reference_mode = COMPOUND_REFERENCE;
        vp9_zero(cm->counts.comp_inter);
      }
    }

    if (cm->tx_mode == TX_MODE_SELECT) {
      int count4x4 = 0;
      int count8x8_lp = 0, count8x8_8x8p = 0;
      int count16x16_16x16p = 0, count16x16_lp = 0;
      int count32x32 = 0;

      for (i = 0; i < TX_SIZE_CONTEXTS; ++i) {
        count4x4 += cm->counts.tx.p32x32[i][TX_4X4];
        count4x4 += cm->counts.tx.p16x16[i][TX_4X4];
        count4x4 += cm->counts.tx.p8x8[i][TX_4X4];

        count8x8_lp += cm->counts.tx.p32x32[i][TX_8X8];
        count8x8_lp += cm->counts.tx.p16x16[i][TX_8X8];
        count8x8_8x8p += cm->counts.tx.p8x8[i][TX_8X8];

        count16x16_16x16p += cm->counts.tx.p16x16[i][TX_16X16];
        count16x16_lp += cm->counts.tx.p32x32[i][TX_16X16];
        count32x32 += cm->counts.tx.p32x32[i][TX_32X32];
      }

      if (count4x4 == 0 && count16x16_lp == 0 && count16x16_16x16p == 0 &&
          count32x32 == 0) {
        cm->tx_mode = ALLOW_8X8;
        reset_skip_tx_size(cm, TX_8X8);
      } else if (count8x8_8x8p == 0 && count16x16_16x16p == 0 &&
                 count8x8_lp == 0 && count16x16_lp == 0 && count32x32 == 0) {
        cm->tx_mode = ONLY_4X4;
        reset_skip_tx_size(cm, TX_4X4);
      } else if (count8x8_lp == 0 && count16x16_lp == 0 && count4x4 == 0) {
        cm->tx_mode = ALLOW_32X32;
      } else if (count32x32 == 0 && count8x8_lp == 0 && count4x4 == 0) {
        cm->tx_mode = ALLOW_16X16;
        reset_skip_tx_size(cm, TX_16X16);
      }
    }
  } else {
    cm->reference_mode = SINGLE_REFERENCE;
    encode_frame_internal(cpi);
  }
}

static void sum_intra_stats(FRAME_COUNTS *counts, const MODE_INFO *mi) {
  const PREDICTION_MODE y_mode = mi->mbmi.mode;
  const PREDICTION_MODE uv_mode = mi->mbmi.uv_mode;
  const BLOCK_SIZE bsize = mi->mbmi.sb_type;

  if (bsize < BLOCK_8X8) {
    int idx, idy;
    const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
    const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
    for (idy = 0; idy < 2; idy += num_4x4_h)
      for (idx = 0; idx < 2; idx += num_4x4_w)
        ++counts->y_mode[0][mi->bmi[idy * 2 + idx].as_mode];
  } else {
    ++counts->y_mode[size_group_lookup[bsize]][y_mode];
  }

  ++counts->uv_mode[y_mode][uv_mode];
}

static int get_zbin_mode_boost(const MB_MODE_INFO *mbmi, int enabled) {
  if (enabled) {
    if (is_inter_block(mbmi)) {
      if (mbmi->mode == ZEROMV) {
        return mbmi->ref_frame[0] != LAST_FRAME ? GF_ZEROMV_ZBIN_BOOST
                                                : LF_ZEROMV_ZBIN_BOOST;
      } else {
        return mbmi->sb_type < BLOCK_8X8 ? SPLIT_MV_ZBIN_BOOST
                                         : MV_ZBIN_BOOST;
      }
    } else {
      return INTRA_ZBIN_BOOST;
    }
  } else {
    return 0;
  }
}

static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, int output_enabled,
                              int mi_row, int mi_col, BLOCK_SIZE bsize,
                              PICK_MODE_CONTEXT *ctx) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  MODE_INFO *mi_8x8 = xd->mi;
  MODE_INFO *mi = mi_8x8;
  MB_MODE_INFO *mbmi = &mi->mbmi;
  const int seg_skip = vp9_segfeature_active(&cm->seg, mbmi->segment_id,
                                             SEG_LVL_SKIP);
  const int mis = cm->mi_stride;
  const int mi_width = num_8x8_blocks_wide_lookup[bsize];
  const int mi_height = num_8x8_blocks_high_lookup[bsize];

  x->skip_recode = !x->select_tx_size && mbmi->sb_type >= BLOCK_8X8 &&
                   cpi->oxcf.aq_mode != COMPLEXITY_AQ &&
                   cpi->oxcf.aq_mode != CYCLIC_REFRESH_AQ &&
                   cpi->sf.allow_skip_recode;

  if (!x->skip_recode && !cpi->sf.use_nonrd_pick_mode)
    vpx_memset(x->skip_txfm, 0, sizeof(x->skip_txfm));

  x->skip_optimize = ctx->is_coded;
  ctx->is_coded = 1;
  x->use_lp32x32fdct = cpi->sf.use_lp32x32fdct;
  x->skip_encode = (!output_enabled && cpi->sf.skip_encode_frame &&
                    x->q_index < QIDX_SKIP_THRESH);

  if (x->skip_encode)
    return;

  set_ref_ptrs(cm, xd, mbmi->ref_frame[0], mbmi->ref_frame[1]);

  // Experimental code. Special case for gf and arf zeromv modes.
  // Increase zbin size to suppress noise
  cpi->zbin_mode_boost = get_zbin_mode_boost(mbmi,
                                             cpi->zbin_mode_boost_enabled);
  vp9_update_zbin_extra(cpi, x);

  if (!is_inter_block(mbmi)) {
    int plane;
    mbmi->skip = 1;
    for (plane = 0; plane < MAX_MB_PLANE; ++plane)
      vp9_encode_intra_block_plane(x, MAX(bsize, BLOCK_8X8), plane);
    if (output_enabled)
      sum_intra_stats(&cm->counts, mi);
    vp9_tokenize_sb(cpi, t, !output_enabled, MAX(bsize, BLOCK_8X8));
  } else {
    int ref;
    const int is_compound = has_second_ref(mbmi);
    for (ref = 0; ref < 1 + is_compound; ++ref) {
      YV12_BUFFER_CONFIG *cfg = get_ref_frame_buffer(cpi,
                                                     mbmi->ref_frame[ref]);
      vp9_setup_pre_planes(xd, ref, cfg, mi_row, mi_col,
                           &xd->block_refs[ref]->sf);
    }
    if (!cpi->sf.reuse_inter_pred_sby || seg_skip)
      vp9_build_inter_predictors_sby(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));

    vp9_build_inter_predictors_sbuv(xd, mi_row, mi_col, MAX(bsize, BLOCK_8X8));

    vp9_encode_sb(x, MAX(bsize, BLOCK_8X8));
    vp9_tokenize_sb(cpi, t, !output_enabled, MAX(bsize, BLOCK_8X8));
  }

  if (output_enabled) {
    if (cm->tx_mode == TX_MODE_SELECT &&
        mbmi->sb_type >= BLOCK_8X8  &&
        !(is_inter_block(mbmi) && (mbmi->skip || seg_skip))) {
      ++get_tx_counts(max_txsize_lookup[bsize], vp9_get_tx_size_context(xd),
                      &cm->counts.tx)[mbmi->tx_size];
    } else {
      int x, y;
      TX_SIZE tx_size;
      // The new intra coding scheme requires no change of transform size
      if (is_inter_block(&mi->mbmi)) {
        tx_size = MIN(tx_mode_to_biggest_tx_size[cm->tx_mode],
                      max_txsize_lookup[bsize]);
      } else {
        tx_size = (bsize >= BLOCK_8X8) ? mbmi->tx_size : TX_4X4;
      }

      for (y = 0; y < mi_height; y++)
        for (x = 0; x < mi_width; x++)
          if (mi_col + x < cm->mi_cols && mi_row + y < cm->mi_rows)
            mi_8x8[mis * y + x].src_mi->mbmi.tx_size = tx_size;
    }
  }
}