av1/encoder/tpl_model.c

/*
 * Copyright (c) 2019, 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 <assert.h>
#include <float.h>
#include <stdint.h>

#include "config/aom_config.h"

#if CONFIG_THREE_PASS
#include "av1/encoder/thirdpass.h"
#endif
#include "config/aom_dsp_rtcd.h"
#include "config/aom_scale_rtcd.h"

#include "aom/aom_codec.h"
#include "aom_util/aom_pthread.h"

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

#include "av1/encoder/encoder.h"
#include "av1/encoder/ethread.h"
#include "av1/encoder/encodeframe_utils.h"
#include "av1/encoder/encode_strategy.h"
#include "av1/encoder/hybrid_fwd_txfm.h"
#include "av1/encoder/motion_search_facade.h"
#include "av1/encoder/rd.h"
#include "av1/encoder/rdopt.h"
#include "av1/encoder/reconinter_enc.h"
#include "av1/encoder/tpl_model.h"

static inline double exp_bounded(double v) {
  // When v > 700 or <-700, the exp function will be close to overflow
  // For details, see the "Notes" in the following link.
  // https://en.cppreference.com/w/c/numeric/math/exp
  if (v > 700) {
    return DBL_MAX;
  } else if (v < -700) {
    return 0;
  }
  return exp(v);
}

void av1_init_tpl_txfm_stats(TplTxfmStats *tpl_txfm_stats) {
  tpl_txfm_stats->ready = 0;
  tpl_txfm_stats->coeff_num = 256;
  tpl_txfm_stats->txfm_block_count = 0;
  memset(tpl_txfm_stats->abs_coeff_sum, 0,
         sizeof(tpl_txfm_stats->abs_coeff_sum[0]) * tpl_txfm_stats->coeff_num);
  memset(tpl_txfm_stats->abs_coeff_mean, 0,
         sizeof(tpl_txfm_stats->abs_coeff_mean[0]) * tpl_txfm_stats->coeff_num);
}

#if CONFIG_BITRATE_ACCURACY
void av1_accumulate_tpl_txfm_stats(const TplTxfmStats *sub_stats,
                                   TplTxfmStats *accumulated_stats) {
  accumulated_stats->txfm_block_count += sub_stats->txfm_block_count;
  for (int i = 0; i < accumulated_stats->coeff_num; ++i) {
    accumulated_stats->abs_coeff_sum[i] += sub_stats->abs_coeff_sum[i];
  }
}

void av1_record_tpl_txfm_block(TplTxfmStats *tpl_txfm_stats,
                               const tran_low_t *coeff) {
  // For transform larger than 16x16, the scale of coeff need to be adjusted.
  // It's not LOSSLESS_Q_STEP.
  assert(tpl_txfm_stats->coeff_num <= 256);
  for (int i = 0; i < tpl_txfm_stats->coeff_num; ++i) {
    tpl_txfm_stats->abs_coeff_sum[i] += abs(coeff[i]) / (double)LOSSLESS_Q_STEP;
  }
  ++tpl_txfm_stats->txfm_block_count;
}

void av1_tpl_txfm_stats_update_abs_coeff_mean(TplTxfmStats *txfm_stats) {
  if (txfm_stats->txfm_block_count > 0) {
    for (int j = 0; j < txfm_stats->coeff_num; j++) {
      txfm_stats->abs_coeff_mean[j] =
          txfm_stats->abs_coeff_sum[j] / txfm_stats->txfm_block_count;
    }
    txfm_stats->ready = 1;
  } else {
    txfm_stats->ready = 0;
  }
}

static inline void av1_tpl_store_txfm_stats(TplParams *tpl_data,
                                            const TplTxfmStats *tpl_txfm_stats,
                                            const int frame_index) {
  tpl_data->txfm_stats_list[frame_index] = *tpl_txfm_stats;
}
#endif  // CONFIG_BITRATE_ACCURACY

static inline void get_quantize_error(const MACROBLOCK *x, int plane,
                                      const tran_low_t *coeff,
                                      tran_low_t *qcoeff, tran_low_t *dqcoeff,
                                      TX_SIZE tx_size, uint16_t *eob,
                                      int64_t *recon_error, int64_t *sse) {
  const struct macroblock_plane *const p = &x->plane[plane];
  const MACROBLOCKD *xd = &x->e_mbd;
  const SCAN_ORDER *const scan_order = &av1_scan_orders[tx_size][DCT_DCT];
  int pix_num = 1 << num_pels_log2_lookup[txsize_to_bsize[tx_size]];
  const int shift = tx_size == TX_32X32 ? 0 : 2;

  QUANT_PARAM quant_param;
  av1_setup_quant(tx_size, 0, AV1_XFORM_QUANT_FP, 0, &quant_param);

#if CONFIG_AV1_HIGHBITDEPTH
  if (is_cur_buf_hbd(xd)) {
    av1_highbd_quantize_fp_facade(coeff, pix_num, p, qcoeff, dqcoeff, eob,
                                  scan_order, &quant_param);
    *recon_error =
        av1_highbd_block_error(coeff, dqcoeff, pix_num, sse, xd->bd) >> shift;
  } else {
    av1_quantize_fp_facade(coeff, pix_num, p, qcoeff, dqcoeff, eob, scan_order,
                           &quant_param);
    *recon_error = av1_block_error(coeff, dqcoeff, pix_num, sse) >> shift;
  }
#else
  (void)xd;
  av1_quantize_fp_facade(coeff, pix_num, p, qcoeff, dqcoeff, eob, scan_order,
                         &quant_param);
  *recon_error = av1_block_error(coeff, dqcoeff, pix_num, sse) >> shift;
#endif  // CONFIG_AV1_HIGHBITDEPTH

  *recon_error = AOMMAX(*recon_error, 1);

  *sse = (*sse) >> shift;
  *sse = AOMMAX(*sse, 1);
}

static inline void set_tpl_stats_block_size(uint8_t *block_mis_log2,
                                            uint8_t *tpl_bsize_1d) {
  // tpl stats bsize: 2 means 16x16
  *block_mis_log2 = 2;
  // Block size used in tpl motion estimation
  *tpl_bsize_1d = 16;
  // MIN_TPL_BSIZE_1D = 16;
  assert(*tpl_bsize_1d >= 16);
}

void av1_setup_tpl_buffers(AV1_PRIMARY *const ppi,
                           CommonModeInfoParams *const mi_params, int width,
                           int height, int byte_alignment, int lag_in_frames) {
  SequenceHeader *const seq_params = &ppi->seq_params;
  TplParams *const tpl_data = &ppi->tpl_data;
  set_tpl_stats_block_size(&tpl_data->tpl_stats_block_mis_log2,
                           &tpl_data->tpl_bsize_1d);
  const uint8_t block_mis_log2 = tpl_data->tpl_stats_block_mis_log2;
  tpl_data->border_in_pixels =
      ALIGN_POWER_OF_TWO(tpl_data->tpl_bsize_1d + 2 * AOM_INTERP_EXTEND, 5);

  const int alloc_y_plane_only =
      ppi->cpi->sf.tpl_sf.use_y_only_rate_distortion ? 1 : 0;
  for (int frame = 0; frame < MAX_LENGTH_TPL_FRAME_STATS; ++frame) {
    const int mi_cols =
        ALIGN_POWER_OF_TWO(mi_params->mi_cols, MAX_MIB_SIZE_LOG2);
    const int mi_rows =
        ALIGN_POWER_OF_TWO(mi_params->mi_rows, MAX_MIB_SIZE_LOG2);
    TplDepFrame *tpl_frame = &tpl_data->tpl_stats_buffer[frame];
    tpl_frame->is_valid = 0;
    tpl_frame->width = mi_cols >> block_mis_log2;
    tpl_frame->height = mi_rows >> block_mis_log2;
    tpl_frame->stride = tpl_data->tpl_stats_buffer[frame].width;
    tpl_frame->mi_rows = mi_params->mi_rows;
    tpl_frame->mi_cols = mi_params->mi_cols;
  }
  tpl_data->tpl_frame = &tpl_data->tpl_stats_buffer[REF_FRAMES + 1];

  // If lag_in_frames <= 1, TPL module is not invoked. Hence dynamic memory
  // allocations are avoided for buffers in tpl_data.
  if (lag_in_frames <= 1) return;

  AOM_CHECK_MEM_ERROR(&ppi->error, tpl_data->txfm_stats_list,
                      aom_calloc(MAX_LENGTH_TPL_FRAME_STATS,
                                 sizeof(*tpl_data->txfm_stats_list)));

  for (int frame = 0; frame < lag_in_frames; ++frame) {
    AOM_CHECK_MEM_ERROR(
        &ppi->error, tpl_data->tpl_stats_pool[frame],
        aom_calloc(tpl_data->tpl_stats_buffer[frame].width *
                       tpl_data->tpl_stats_buffer[frame].height,
                   sizeof(*tpl_data->tpl_stats_buffer[frame].tpl_stats_ptr)));

    if (aom_alloc_frame_buffer(
            &tpl_data->tpl_rec_pool[frame], width, height,
            seq_params->subsampling_x, seq_params->subsampling_y,
            seq_params->use_highbitdepth, tpl_data->border_in_pixels,
            byte_alignment, false, alloc_y_plane_only))
      aom_internal_error(&ppi->error, AOM_CODEC_MEM_ERROR,
                         "Failed to allocate frame buffer");
  }
}

static inline int32_t tpl_get_satd_cost(BitDepthInfo bd_info, int16_t *src_diff,
                                        int diff_stride, const uint8_t *src,
                                        int src_stride, const uint8_t *dst,
                                        int dst_stride, tran_low_t *coeff,
                                        int bw, int bh, TX_SIZE tx_size) {
  const int pix_num = bw * bh;

  av1_subtract_block(bd_info, bh, bw, src_diff, diff_stride, src, src_stride,
                     dst, dst_stride);
  av1_quick_txfm(/*use_hadamard=*/0, tx_size, bd_info, src_diff, bw, coeff);
  return aom_satd(coeff, pix_num);
}

static int rate_estimator(const tran_low_t *qcoeff, int eob, TX_SIZE tx_size) {
  const SCAN_ORDER *const scan_order = &av1_scan_orders[tx_size][DCT_DCT];

  assert((1 << num_pels_log2_lookup[txsize_to_bsize[tx_size]]) >= eob);
  int rate_cost = 1;

  for (int idx = 0; idx < eob; ++idx) {
    unsigned int abs_level = abs(qcoeff[scan_order->scan[idx]]);
    rate_cost += get_msb(abs_level + 1) + 1 + (abs_level > 0);
  }

  return (rate_cost << AV1_PROB_COST_SHIFT);
}

static inline void txfm_quant_rdcost(
    const MACROBLOCK *x, int16_t *src_diff, int diff_stride, uint8_t *src,
    int src_stride, uint8_t *dst, int dst_stride, tran_low_t *coeff,
    tran_low_t *qcoeff, tran_low_t *dqcoeff, int bw, int bh, TX_SIZE tx_size,
    int do_recon, int *rate_cost, int64_t *recon_error, int64_t *sse) {
  const MACROBLOCKD *xd = &x->e_mbd;
  const BitDepthInfo bd_info = get_bit_depth_info(xd);
  uint16_t eob;
  av1_subtract_block(bd_info, bh, bw, src_diff, diff_stride, src, src_stride,
                     dst, dst_stride);
  av1_quick_txfm(/*use_hadamard=*/0, tx_size, bd_info, src_diff, bw, coeff);

  get_quantize_error(x, 0, coeff, qcoeff, dqcoeff, tx_size, &eob, recon_error,
                     sse);

  *rate_cost = rate_estimator(qcoeff, eob, tx_size);

  if (do_recon)
    av1_inverse_transform_block(xd, dqcoeff, 0, DCT_DCT, tx_size, dst,
                                dst_stride, eob, 0);
}

static uint32_t motion_estimation(AV1_COMP *cpi, MACROBLOCK *x,
                                  uint8_t *cur_frame_buf,
                                  uint8_t *ref_frame_buf, int stride,
                                  int ref_stride, int width, int ref_width,
                                  BLOCK_SIZE bsize, MV center_mv,
                                  int_mv *best_mv) {
  AV1_COMMON *cm = &cpi->common;
  MACROBLOCKD *const xd = &x->e_mbd;
  TPL_SPEED_FEATURES *tpl_sf = &cpi->sf.tpl_sf;
  int step_param;
  uint32_t bestsme = UINT_MAX;
  FULLPEL_MV_STATS best_mv_stats;
  int distortion;
  uint32_t sse;
  int cost_list[5];
  FULLPEL_MV start_mv = get_fullmv_from_mv(&center_mv);

  // Setup frame pointers
  x->plane[0].src.buf = cur_frame_buf;
  x->plane[0].src.stride = stride;
  x->plane[0].src.width = width;
  xd->plane[0].pre[0].buf = ref_frame_buf;
  xd->plane[0].pre[0].stride = ref_stride;
  xd->plane[0].pre[0].width = ref_width;

  step_param = tpl_sf->reduce_first_step_size;
  step_param = AOMMIN(step_param, MAX_MVSEARCH_STEPS - 2);

  const search_site_config *search_site_cfg =
      cpi->mv_search_params.search_site_cfg[SS_CFG_SRC];
  if (search_site_cfg->stride != ref_stride)
    search_site_cfg = cpi->mv_search_params.search_site_cfg[SS_CFG_LOOKAHEAD];
  assert(search_site_cfg->stride == ref_stride);

  FULLPEL_MOTION_SEARCH_PARAMS full_ms_params;
  av1_make_default_fullpel_ms_params(&full_ms_params, cpi, x, bsize, &center_mv,
                                     start_mv, search_site_cfg,
                                     tpl_sf->search_method,
                                     /*fine_search_interval=*/0);

  bestsme = av1_full_pixel_search(start_mv, &full_ms_params, step_param,
                                  cond_cost_list(cpi, cost_list),
                                  &best_mv->as_fullmv, &best_mv_stats, NULL);

  // When sub-pel motion search is skipped, populate sub-pel precision MV and
  // return.
  if (tpl_sf->subpel_force_stop == FULL_PEL) {
    best_mv->as_mv = get_mv_from_fullmv(&best_mv->as_fullmv);
    return bestsme;
  }

  SUBPEL_MOTION_SEARCH_PARAMS ms_params;
  av1_make_default_subpel_ms_params(&ms_params, cpi, x, bsize, &center_mv,
                                    cost_list);
  ms_params.forced_stop = tpl_sf->subpel_force_stop;
  ms_params.var_params.subpel_search_type = USE_2_TAPS;
  ms_params.mv_cost_params.mv_cost_type = MV_COST_NONE;
  best_mv_stats.err_cost = 0;
  MV subpel_start_mv = get_mv_from_fullmv(&best_mv->as_fullmv);
  assert(av1_is_subpelmv_in_range(&ms_params.mv_limits, subpel_start_mv));
  bestsme = cpi->mv_search_params.find_fractional_mv_step(
      xd, cm, &ms_params, subpel_start_mv, &best_mv_stats, &best_mv->as_mv,
      &distortion, &sse, NULL);

  return bestsme;
}

typedef struct {
  int_mv mv;
  int sad;
} center_mv_t;

static int compare_sad(const void *a, const void *b) {
  const int diff = ((center_mv_t *)a)->sad - ((center_mv_t *)b)->sad;
  if (diff < 0)
    return -1;
  else if (diff > 0)
    return 1;
  return 0;
}

static int is_alike_mv(int_mv candidate_mv, center_mv_t *center_mvs,
                       int center_mvs_count, int skip_alike_starting_mv) {
  // MV difference threshold is in 1/8 precision.
  const int mv_diff_thr[3] = { 1, (8 << 3), (16 << 3) };
  int thr = mv_diff_thr[skip_alike_starting_mv];
  int i;

  for (i = 0; i < center_mvs_count; i++) {
    if (abs(center_mvs[i].mv.as_mv.col - candidate_mv.as_mv.col) < thr &&
        abs(center_mvs[i].mv.as_mv.row - candidate_mv.as_mv.row) < thr)
      return 1;
  }

  return 0;
}

static void get_rate_distortion(
    int *rate_cost, int64_t *recon_error, int64_t *pred_error,
    int16_t *src_diff, tran_low_t *coeff, tran_low_t *qcoeff,
    tran_low_t *dqcoeff, AV1_COMMON *cm, MACROBLOCK *x,
    const YV12_BUFFER_CONFIG *ref_frame_ptr[2], uint8_t *rec_buffer_pool[3],
    const int rec_stride_pool[3], TX_SIZE tx_size, PREDICTION_MODE best_mode,
    int mi_row, int mi_col, int use_y_only_rate_distortion, int do_recon,
    TplTxfmStats *tpl_txfm_stats) {
  const SequenceHeader *seq_params = cm->seq_params;
  *rate_cost = 0;
  *recon_error = 1;
  *pred_error = 1;

  (void)tpl_txfm_stats;

  MACROBLOCKD *xd = &x->e_mbd;
  int is_compound = (best_mode == NEW_NEWMV);
  int num_planes = use_y_only_rate_distortion ? 1 : MAX_MB_PLANE;

  uint8_t *src_buffer_pool[MAX_MB_PLANE] = {
    xd->cur_buf->y_buffer,
    xd->cur_buf->u_buffer,
    xd->cur_buf->v_buffer,
  };
  const int src_stride_pool[MAX_MB_PLANE] = {
    xd->cur_buf->y_stride,
    xd->cur_buf->uv_stride,
    xd->cur_buf->uv_stride,
  };

  const int_interpfilters kernel =
      av1_broadcast_interp_filter(EIGHTTAP_REGULAR);

  for (int plane = 0; plane < num_planes; ++plane) {
    struct macroblockd_plane *pd = &xd->plane[plane];
    BLOCK_SIZE bsize_plane =
        av1_ss_size_lookup[txsize_to_bsize[tx_size]][pd->subsampling_x]
                          [pd->subsampling_y];

    int dst_buffer_stride = rec_stride_pool[plane];
    int dst_mb_offset =
        ((mi_row * MI_SIZE * dst_buffer_stride) >> pd->subsampling_y) +
        ((mi_col * MI_SIZE) >> pd->subsampling_x);
    uint8_t *dst_buffer = rec_buffer_pool[plane] + dst_mb_offset;
    for (int ref = 0; ref < 1 + is_compound; ++ref) {
      if (!is_inter_mode(best_mode)) {
        av1_predict_intra_block(
            xd, seq_params->sb_size, seq_params->enable_intra_edge_filter,
            block_size_wide[bsize_plane], block_size_high[bsize_plane],
            max_txsize_rect_lookup[bsize_plane], best_mode, 0, 0,
            FILTER_INTRA_MODES, dst_buffer, dst_buffer_stride, dst_buffer,
            dst_buffer_stride, 0, 0, plane);
      } else {
        int_mv best_mv = xd->mi[0]->mv[ref];
        uint8_t *ref_buffer_pool[MAX_MB_PLANE] = {
          ref_frame_ptr[ref]->y_buffer,
          ref_frame_ptr[ref]->u_buffer,
          ref_frame_ptr[ref]->v_buffer,
        };
        InterPredParams inter_pred_params;
        struct buf_2d ref_buf = {
          NULL, ref_buffer_pool[plane],
          plane ? ref_frame_ptr[ref]->uv_width : ref_frame_ptr[ref]->y_width,
          plane ? ref_frame_ptr[ref]->uv_height : ref_frame_ptr[ref]->y_height,
          plane ? ref_frame_ptr[ref]->uv_stride : ref_frame_ptr[ref]->y_stride
        };
        av1_init_inter_params(&inter_pred_params, block_size_wide[bsize_plane],
                              block_size_high[bsize_plane],
                              (mi_row * MI_SIZE) >> pd->subsampling_y,
                              (mi_col * MI_SIZE) >> pd->subsampling_x,
                              pd->subsampling_x, pd->subsampling_y, xd->bd,
                              is_cur_buf_hbd(xd), 0,
                              xd->block_ref_scale_factors[0], &ref_buf, kernel);
        if (is_compound) av1_init_comp_mode(&inter_pred_params);
        inter_pred_params.conv_params = get_conv_params_no_round(
            ref, plane, xd->tmp_conv_dst, MAX_SB_SIZE, is_compound, xd->bd);

        av1_enc_build_one_inter_predictor(dst_buffer, dst_buffer_stride,
                                          &best_mv.as_mv, &inter_pred_params);
      }
    }

    int src_stride = src_stride_pool[plane];
    int src_mb_offset = ((mi_row * MI_SIZE * src_stride) >> pd->subsampling_y) +
                        ((mi_col * MI_SIZE) >> pd->subsampling_x);

    int this_rate = 1;
    int64_t this_recon_error = 1;
    int64_t sse;
    txfm_quant_rdcost(
        x, src_diff, block_size_wide[bsize_plane],
        src_buffer_pool[plane] + src_mb_offset, src_stride, dst_buffer,
        dst_buffer_stride, coeff, qcoeff, dqcoeff, block_size_wide[bsize_plane],
        block_size_high[bsize_plane], max_txsize_rect_lookup[bsize_plane],
        do_recon, &this_rate, &this_recon_error, &sse);

#if CONFIG_BITRATE_ACCURACY
    if (plane == 0 && tpl_txfm_stats) {
      // We only collect Y plane's transform coefficient
      av1_record_tpl_txfm_block(tpl_txfm_stats, coeff);
    }
#endif  // CONFIG_BITRATE_ACCURACY

    *recon_error += this_recon_error;
    *pred_error += sse;
    *rate_cost += this_rate;
  }
}

static inline int32_t get_inter_cost(const AV1_COMP *cpi, MACROBLOCKD *xd,
                                     const uint8_t *src_mb_buffer,
                                     int src_stride,
                                     TplBuffers *tpl_tmp_buffers,
                                     BLOCK_SIZE bsize, TX_SIZE tx_size,
                                     int mi_row, int mi_col, int rf_idx,
                                     MV *rfidx_mv, int use_pred_sad) {
  const BitDepthInfo bd_info = get_bit_depth_info(xd);
  TplParams *tpl_data = &cpi->ppi->tpl_data;
  const YV12_BUFFER_CONFIG *const ref_frame_ptr =
      tpl_data->src_ref_frame[rf_idx];
  int16_t *src_diff = tpl_tmp_buffers->src_diff;
  tran_low_t *coeff = tpl_tmp_buffers->coeff;
  const int bw = 4 << mi_size_wide_log2[bsize];
  const int bh = 4 << mi_size_high_log2[bsize];
  int32_t inter_cost;

  if (cpi->sf.tpl_sf.subpel_force_stop != FULL_PEL) {
    const int_interpfilters kernel =
        av1_broadcast_interp_filter(EIGHTTAP_REGULAR);
    uint8_t *predictor8 = tpl_tmp_buffers->predictor8;
    uint8_t *predictor =
        is_cur_buf_hbd(xd) ? CONVERT_TO_BYTEPTR(predictor8) : predictor8;
    struct buf_2d ref_buf = { NULL, ref_frame_ptr->y_buffer,
                              ref_frame_ptr->y_width, ref_frame_ptr->y_height,
                              ref_frame_ptr->y_stride };
    InterPredParams inter_pred_params;
    av1_init_inter_params(&inter_pred_params, bw, bh, mi_row * MI_SIZE,
                          mi_col * MI_SIZE, 0, 0, xd->bd, is_cur_buf_hbd(xd), 0,
                          &tpl_data->sf, &ref_buf, kernel);
    inter_pred_params.conv_params = get_conv_params(0, 0, xd->bd);

    av1_enc_build_one_inter_predictor(predictor, bw, rfidx_mv,
                                      &inter_pred_params);

    if (use_pred_sad) {
      inter_cost = (int)cpi->ppi->fn_ptr[bsize].sdf(src_mb_buffer, src_stride,
                                                    predictor, bw);
    } else {
      inter_cost =
          tpl_get_satd_cost(bd_info, src_diff, bw, src_mb_buffer, src_stride,
                            predictor, bw, coeff, bw, bh, tx_size);
    }
  } else {
    int ref_mb_offset =
        mi_row * MI_SIZE * ref_frame_ptr->y_stride + mi_col * MI_SIZE;
    uint8_t *ref_mb = ref_frame_ptr->y_buffer + ref_mb_offset;
    int ref_stride = ref_frame_ptr->y_stride;
    const FULLPEL_MV fullmv = get_fullmv_from_mv(rfidx_mv);
    // Since sub-pel motion search is not performed, use the prediction pixels
    // directly from the reference block ref_mb
    if (use_pred_sad) {
      inter_cost = (int)cpi->ppi->fn_ptr[bsize].sdf(
          src_mb_buffer, src_stride,
          &ref_mb[fullmv.row * ref_stride + fullmv.col], ref_stride);
    } else {
      inter_cost =
          tpl_get_satd_cost(bd_info, src_diff, bw, src_mb_buffer, src_stride,
                            &ref_mb[fullmv.row * ref_stride + fullmv.col],
                            ref_stride, coeff, bw, bh, tx_size);
    }
  }
  return inter_cost;
}

static inline void mode_estimation(AV1_COMP *cpi, TplTxfmStats *tpl_txfm_stats,
                                   TplBuffers *tpl_tmp_buffers, MACROBLOCK *x,
                                   int mi_row, int mi_col, BLOCK_SIZE bsize,
                                   TX_SIZE tx_size, TplDepStats *tpl_stats) {
  AV1_COMMON *cm = &cpi->common;
  const GF_GROUP *gf_group = &cpi->ppi->gf_group;
  TPL_SPEED_FEATURES *tpl_sf = &cpi->sf.tpl_sf;

  (void)gf_group;

  MACROBLOCKD *xd = &x->e_mbd;
  const BitDepthInfo bd_info = get_bit_depth_info(xd);
  TplParams *tpl_data = &cpi->ppi->tpl_data;
  TplDepFrame *tpl_frame = &tpl_data->tpl_frame[tpl_data->frame_idx];
  const uint8_t block_mis_log2 = tpl_data->tpl_stats_block_mis_log2;

  const int bw = 4 << mi_size_wide_log2[bsize];
  const int bh = 4 << mi_size_high_log2[bsize];

  int32_t best_intra_cost = INT32_MAX;
  int32_t intra_cost;
  PREDICTION_MODE best_mode = DC_PRED;

  const int mb_y_offset =
      mi_row * MI_SIZE * xd->cur_buf->y_stride + mi_col * MI_SIZE;
  uint8_t *src_mb_buffer = xd->cur_buf->y_buffer + mb_y_offset;
  const int src_stride = xd->cur_buf->y_stride;
  const int src_width = xd->cur_buf->y_width;

  int dst_mb_offset =
      mi_row * MI_SIZE * tpl_frame->rec_picture->y_stride + mi_col * MI_SIZE;
  uint8_t *dst_buffer = tpl_frame->rec_picture->y_buffer + dst_mb_offset;
  int dst_buffer_stride = tpl_frame->rec_picture->y_stride;
  int use_y_only_rate_distortion = tpl_sf->use_y_only_rate_distortion;

  uint8_t *rec_buffer_pool[3] = {
    tpl_frame->rec_picture->y_buffer,
    tpl_frame->rec_picture->u_buffer,
    tpl_frame->rec_picture->v_buffer,
  };

  const int rec_stride_pool[3] = {
    tpl_frame->rec_picture->y_stride,
    tpl_frame->rec_picture->uv_stride,
    tpl_frame->rec_picture->uv_stride,
  };

  for (int plane = 1; plane < MAX_MB_PLANE; ++plane) {
    struct macroblockd_plane *pd = &xd->plane[plane];
    pd->subsampling_x = xd->cur_buf->subsampling_x;
    pd->subsampling_y = xd->cur_buf->subsampling_y;
  }

  uint8_t *predictor8 = tpl_tmp_buffers->predictor8;
  int16_t *src_diff = tpl_tmp_buffers->src_diff;
  tran_low_t *coeff = tpl_tmp_buffers->coeff;
  tran_low_t *qcoeff = tpl_tmp_buffers->qcoeff;
  tran_low_t *dqcoeff = tpl_tmp_buffers->dqcoeff;
  uint8_t *predictor =
      is_cur_buf_hbd(xd) ? CONVERT_TO_BYTEPTR(predictor8) : predictor8;
  int64_t recon_error = 1;
  int64_t pred_error = 1;

  memset(tpl_stats, 0, sizeof(*tpl_stats));
  tpl_stats->ref_frame_index[0] = -1;
  tpl_stats->ref_frame_index[1] = -1;

  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_mi_row_col(xd, &xd->tile, mi_row, mi_height, mi_col, mi_width,
                 cm->mi_params.mi_rows, cm->mi_params.mi_cols);
  set_plane_n4(xd, mi_size_wide[bsize], mi_size_high[bsize],
               av1_num_planes(cm));
  xd->mi[0]->bsize = bsize;
  xd->mi[0]->motion_mode = SIMPLE_TRANSLATION;

  // Intra prediction search
  xd->mi[0]->ref_frame[0] = INTRA_FRAME;

  // Pre-load the bottom left line.
  if (xd->left_available &&
      mi_row + tx_size_high_unit[tx_size] < xd->tile.mi_row_end) {
    if (is_cur_buf_hbd(xd)) {
      uint16_t *dst = CONVERT_TO_SHORTPTR(dst_buffer);
      for (int i = 0; i < bw; ++i)
        dst[(bw + i) * dst_buffer_stride - 1] =
            dst[(bw - 1) * dst_buffer_stride - 1];
    } else {
      for (int i = 0; i < bw; ++i)
        dst_buffer[(bw + i) * dst_buffer_stride - 1] =
            dst_buffer[(bw - 1) * dst_buffer_stride - 1];
    }
  }

  // if cpi->sf.tpl_sf.prune_intra_modes is on, then search only DC_PRED,
  // H_PRED, and V_PRED
  const PREDICTION_MODE last_intra_mode =
      tpl_sf->prune_intra_modes ? D45_PRED : INTRA_MODE_END;
  const SequenceHeader *seq_params = cm->seq_params;
  for (PREDICTION_MODE mode = INTRA_MODE_START; mode < last_intra_mode;
       ++mode) {
    av1_predict_intra_block(xd, seq_params->sb_size,
                            seq_params->enable_intra_edge_filter,
                            block_size_wide[bsize], block_size_high[bsize],
                            tx_size, mode, 0, 0, FILTER_INTRA_MODES, dst_buffer,
                            dst_buffer_stride, predictor, bw, 0, 0, 0);

    if (tpl_frame->use_pred_sad) {
      intra_cost = (int32_t)cpi->ppi->fn_ptr[bsize].sdf(
          src_mb_buffer, src_stride, predictor, bw);
    } else {
      intra_cost =
          tpl_get_satd_cost(bd_info, src_diff, bw, src_mb_buffer, src_stride,
                            predictor, bw, coeff, bw, bh, tx_size);
    }

    if (intra_cost < best_intra_cost) {
      best_intra_cost = intra_cost;
      best_mode = mode;
    }
  }
  // Calculate SATD of the best intra mode if SAD was used for mode decision
  // as best_intra_cost is used in ML model to skip intra mode evaluation.
  if (tpl_frame->use_pred_sad) {
    av1_predict_intra_block(
        xd, seq_params->sb_size, seq_params->enable_intra_edge_filter,
        block_size_wide[bsize], block_size_high[bsize], tx_size, best_mode, 0,
        0, FILTER_INTRA_MODES, dst_buffer, dst_buffer_stride, predictor, bw, 0,
        0, 0);
    best_intra_cost =
        tpl_get_satd_cost(bd_info, src_diff, bw, src_mb_buffer, src_stride,
                          predictor, bw, coeff, bw, bh, tx_size);
  }

  int rate_cost = 1;

  if (cpi->use_ducky_encode) {
    get_rate_distortion(&rate_cost, &recon_error, &pred_error, src_diff, coeff,
                        qcoeff, dqcoeff, cm, x, NULL, rec_buffer_pool,
                        rec_stride_pool, tx_size, best_mode, mi_row, mi_col,
                        use_y_only_rate_distortion, 1 /*do_recon*/, NULL);

    tpl_stats->intra_dist = recon_error << TPL_DEP_COST_SCALE_LOG2;
    tpl_stats->intra_sse = pred_error << TPL_DEP_COST_SCALE_LOG2;
    tpl_stats->intra_rate = rate_cost;
  }

#if CONFIG_THREE_PASS
  const int frame_offset = tpl_data->frame_idx - cpi->gf_frame_index;

  if (cpi->third_pass_ctx &&
      frame_offset < cpi->third_pass_ctx->frame_info_count &&
      tpl_data->frame_idx < gf_group->size) {
    double ratio_h, ratio_w;
    av1_get_third_pass_ratio(cpi->third_pass_ctx, frame_offset, cm->height,
                             cm->width, &ratio_h, &ratio_w);
    THIRD_PASS_MI_INFO *this_mi = av1_get_third_pass_mi(
        cpi->third_pass_ctx, frame_offset, mi_row, mi_col, ratio_h, ratio_w);

    PREDICTION_MODE third_pass_mode = this_mi->pred_mode;

    if (third_pass_mode >= last_intra_mode &&
        third_pass_mode < INTRA_MODE_END) {
      av1_predict_intra_block(
          xd, seq_params->sb_size, seq_params->enable_intra_edge_filter,
          block_size_wide[bsize], block_size_high[bsize], tx_size,
          third_pass_mode, 0, 0, FILTER_INTRA_MODES, dst_buffer,
          dst_buffer_stride, predictor, bw, 0, 0, 0);

      intra_cost =
          tpl_get_satd_cost(bd_info, src_diff, bw, src_mb_buffer, src_stride,
                            predictor, bw, coeff, bw, bh, tx_size);

      if (intra_cost < best_intra_cost) {
        best_intra_cost = intra_cost;
        best_mode = third_pass_mode;
      }
    }
  }
#endif  // CONFIG_THREE_PASS

  // Motion compensated prediction
  xd->mi[0]->ref_frame[0] = INTRA_FRAME;
  xd->mi[0]->ref_frame[1] = NONE_FRAME;
  xd->mi[0]->compound_idx = 1;

  int best_rf_idx = -1;
  int_mv best_mv[2];
  int32_t inter_cost;
  int32_t best_inter_cost = INT32_MAX;
  int rf_idx;
  int_mv single_mv[INTER_REFS_PER_FRAME];

  best_mv[0].as_int = INVALID_MV;
  best_mv[1].as_int = INVALID_MV;

  for (rf_idx = 0; rf_idx < INTER_REFS_PER_FRAME; ++rf_idx) {
    single_mv[rf_idx].as_int = INVALID_MV;
    if (tpl_data->ref_frame[rf_idx] == NULL ||
        tpl_data->src_ref_frame[rf_idx] == NULL) {
      tpl_stats->mv[rf_idx].as_int = INVALID_MV;
      continue;
    }

    const YV12_BUFFER_CONFIG *ref_frame_ptr = tpl_data->src_ref_frame[rf_idx];
    const int ref_mb_offset =
        mi_row * MI_SIZE * ref_frame_ptr->y_stride + mi_col * MI_SIZE;
    uint8_t *ref_mb = ref_frame_ptr->y_buffer + ref_mb_offset;
    const int ref_stride = ref_frame_ptr->y_stride;
    const int ref_width = ref_frame_ptr->y_width;

    int_mv best_rfidx_mv = { 0 };
    uint32_t bestsme = UINT32_MAX;

    center_mv_t center_mvs[4] = { { { 0 }, INT_MAX },
                                  { { 0 }, INT_MAX },
                                  { { 0 }, INT_MAX },
                                  { { 0 }, INT_MAX } };
    int refmv_count = 1;
    int idx;

    if (xd->up_available) {
      TplDepStats *ref_tpl_stats = &tpl_frame->tpl_stats_ptr[av1_tpl_ptr_pos(
          mi_row - mi_height, mi_col, tpl_frame->stride, block_mis_log2)];
      if (!is_alike_mv(ref_tpl_stats->mv[rf_idx], center_mvs, refmv_count,
                       tpl_sf->skip_alike_starting_mv)) {
        center_mvs[refmv_count].mv.as_int = ref_tpl_stats->mv[rf_idx].as_int;
        ++refmv_count;
      }
    }

    if (xd->left_available) {
      TplDepStats *ref_tpl_stats = &tpl_frame->tpl_stats_ptr[av1_tpl_ptr_pos(
          mi_row, mi_col - mi_width, tpl_frame->stride, block_mis_log2)];
      if (!is_alike_mv(ref_tpl_stats->mv[rf_idx], center_mvs, refmv_count,
                       tpl_sf->skip_alike_starting_mv)) {
        center_mvs[refmv_count].mv.as_int = ref_tpl_stats->mv[rf_idx].as_int;
        ++refmv_count;
      }
    }

    if (xd->up_available && mi_col + mi_width < xd->tile.mi_col_end) {
      TplDepStats *ref_tpl_stats = &tpl_frame->tpl_stats_ptr[av1_tpl_ptr_pos(
          mi_row - mi_height, mi_col + mi_width, tpl_frame->stride,
          block_mis_log2)];
      if (!is_alike_mv(ref_tpl_stats->mv[rf_idx], center_mvs, refmv_count,
                       tpl_sf->skip_alike_starting_mv)) {
        center_mvs[refmv_count].mv.as_int = ref_tpl_stats->mv[rf_idx].as_int;
        ++refmv_count;
      }
    }

#if CONFIG_THREE_PASS
    if (cpi->third_pass_ctx &&
        frame_offset < cpi->third_pass_ctx->frame_info_count &&
        tpl_data->frame_idx < gf_group->size) {
      double ratio_h, ratio_w;
      av1_get_third_pass_ratio(cpi->third_pass_ctx, frame_offset, cm->height,
                               cm->width, &ratio_h, &ratio_w);
      THIRD_PASS_MI_INFO *this_mi = av1_get_third_pass_mi(
          cpi->third_pass_ctx, frame_offset, mi_row, mi_col, ratio_h, ratio_w);

      int_mv tp_mv = av1_get_third_pass_adjusted_mv(this_mi, ratio_h, ratio_w,
                                                    rf_idx + LAST_FRAME);
      if (tp_mv.as_int != INVALID_MV &&
          !is_alike_mv(tp_mv, center_mvs + 1, refmv_count - 1,
                       tpl_sf->skip_alike_starting_mv)) {
        center_mvs[0].mv = tp_mv;
      }
    }
#endif  // CONFIG_THREE_PASS

    // Prune starting mvs
    if (tpl_sf->prune_starting_mv && refmv_count > 1) {
      // Get each center mv's sad.
      for (idx = 0; idx < refmv_count; ++idx) {
        FULLPEL_MV mv = get_fullmv_from_mv(&center_mvs[idx].mv.as_mv);
        clamp_fullmv(&mv, &x->mv_limits);
        center_mvs[idx].sad = (int)cpi->ppi->fn_ptr[bsize].sdf(
            src_mb_buffer, src_stride, &ref_mb[mv.row * ref_stride + mv.col],
            ref_stride);
      }

      // Rank center_mv using sad.
      qsort(center_mvs, refmv_count, sizeof(center_mvs[0]), compare_sad);

      refmv_count = AOMMIN(4 - tpl_sf->prune_starting_mv, refmv_count);
      // Further reduce number of refmv based on sad difference.
      if (refmv_count > 1) {
        int last_sad = center_mvs[refmv_count - 1].sad;
        int second_to_last_sad = center_mvs[refmv_count - 2].sad;
        if ((last_sad - second_to_last_sad) * 5 > second_to_last_sad)
          refmv_count--;
      }
    }

    for (idx = 0; idx < refmv_count; ++idx) {
      int_mv this_mv;
      uint32_t thissme = motion_estimation(
          cpi, x, src_mb_buffer, ref_mb, src_stride, ref_stride, src_width,
          ref_width, bsize, center_mvs[idx].mv.as_mv, &this_mv);

      if (thissme < bestsme) {
        bestsme = thissme;
        best_rfidx_mv = this_mv;
      }
    }

    tpl_stats->mv[rf_idx].as_int = best_rfidx_mv.as_int;
    single_mv[rf_idx] = best_rfidx_mv;

    inter_cost = get_inter_cost(
        cpi, xd, src_mb_buffer, src_stride, tpl_tmp_buffers, bsize, tx_size,
        mi_row, mi_col, rf_idx, &best_rfidx_mv.as_mv, tpl_frame->use_pred_sad);
    // Store inter cost for each ref frame. This is used to prune inter modes.
    tpl_stats->pred_error[rf_idx] = AOMMAX(1, inter_cost);

    if (inter_cost < best_inter_cost) {
      best_rf_idx = rf_idx;

      best_inter_cost = inter_cost;
      best_mv[0].as_int = best_rfidx_mv.as_int;
    }
  }
  // Calculate SATD of the best inter mode if SAD was used for mode decision
  // as best_inter_cost is used in ML model to skip intra mode evaluation.
  if (best_inter_cost < INT32_MAX && tpl_frame->use_pred_sad) {
    assert(best_rf_idx != -1);
    best_inter_cost = get_inter_cost(
        cpi, xd, src_mb_buffer, src_stride, tpl_tmp_buffers, bsize, tx_size,
        mi_row, mi_col, best_rf_idx, &best_mv[0].as_mv, 0 /* use_pred_sad */);
  }

  if (best_rf_idx != -1 && best_inter_cost < best_intra_cost) {
    best_mode = NEWMV;
    xd->mi[0]->ref_frame[0] = best_rf_idx + LAST_FRAME;
    xd->mi[0]->mv[0].as_int = best_mv[0].as_int;
  }

  // Start compound predition search.
  int comp_ref_frames[3][2] = {
    { 0, 4 },
    { 0, 6 },
    { 3, 6 },
  };

  int start_rf = 0;
  int end_rf = 3;
  if (!tpl_sf->allow_compound_pred) end_rf = 0;
#if CONFIG_THREE_PASS
  if (cpi->third_pass_ctx &&
      frame_offset < cpi->third_pass_ctx->frame_info_count &&
      tpl_data->frame_idx < gf_group->size) {
    double ratio_h, ratio_w;
    av1_get_third_pass_ratio(cpi->third_pass_ctx, frame_offset, cm->height,
                             cm->width, &ratio_h, &ratio_w);
    THIRD_PASS_MI_INFO *this_mi = av1_get_third_pass_mi(
        cpi->third_pass_ctx, frame_offset, mi_row, mi_col, ratio_h, ratio_w);

    if (this_mi->ref_frame[0] >= LAST_FRAME &&
        this_mi->ref_frame[1] >= LAST_FRAME) {
      int found = 0;
      for (int i = 0; i < 3; i++) {
        if (comp_ref_frames[i][0] + LAST_FRAME == this_mi->ref_frame[0] &&
            comp_ref_frames[i][1] + LAST_FRAME == this_mi->ref_frame[1]) {
          found = 1;
          break;
        }
      }
      if (!found || !tpl_sf->allow_compound_pred) {
        comp_ref_frames[2][0] = this_mi->ref_frame[0] - LAST_FRAME;
        comp_ref_frames[2][1] = this_mi->ref_frame[1] - LAST_FRAME;
        if (!tpl_sf->allow_compound_pred) {
          start_rf = 2;
          end_rf = 3;
        }
      }
    }
  }
#endif  // CONFIG_THREE_PASS

  xd->mi_row = mi_row;
  xd->mi_col = mi_col;
  int best_cmp_rf_idx = -1;
  const int_interpfilters kernel =
      av1_broadcast_interp_filter(EIGHTTAP_REGULAR);
  for (int cmp_rf_idx = start_rf; cmp_rf_idx < end_rf; ++cmp_rf_idx) {
    int rf_idx0 = comp_ref_frames[cmp_rf_idx][0];
    int rf_idx1 = comp_ref_frames[cmp_rf_idx][1];

    if (tpl_data->ref_frame[rf_idx0] == NULL ||
        tpl_data->src_ref_frame[rf_idx0] == NULL ||
        tpl_data->ref_frame[rf_idx1] == NULL ||
        tpl_data->src_ref_frame[rf_idx1] == NULL) {
      continue;
    }

    const YV12_BUFFER_CONFIG *ref_frame_ptr[2] = {
      tpl_data->src_ref_frame[rf_idx0],
      tpl_data->src_ref_frame[rf_idx1],
    };

    xd->mi[0]->ref_frame[0] = rf_idx0 + LAST_FRAME;
    xd->mi[0]->ref_frame[1] = rf_idx1 + LAST_FRAME;
    xd->mi[0]->mode = NEW_NEWMV;
    const int8_t ref_frame_type = av1_ref_frame_type(xd->mi[0]->ref_frame);
    // Set up ref_mv for av1_joint_motion_search().
    CANDIDATE_MV *this_ref_mv_stack = x->mbmi_ext.ref_mv_stack[ref_frame_type];
    this_ref_mv_stack[xd->mi[0]->ref_mv_idx].this_mv = single_mv[rf_idx0];
    this_ref_mv_stack[xd->mi[0]->ref_mv_idx].comp_mv = single_mv[rf_idx1];

    struct buf_2d yv12_mb[2][MAX_MB_PLANE];
    for (int i = 0; i < 2; ++i) {
      av1_setup_pred_block(xd, yv12_mb[i], ref_frame_ptr[i],
                           xd->block_ref_scale_factors[i],
                           xd->block_ref_scale_factors[i], MAX_MB_PLANE);
      for (int plane = 0; plane < MAX_MB_PLANE; ++plane) {
        xd->plane[plane].pre[i] = yv12_mb[i][plane];
      }
    }

    int_mv tmp_mv[2] = { single_mv[rf_idx0], single_mv[rf_idx1] };
    int rate_mv;
    av1_joint_motion_search(cpi, x, bsize, tmp_mv, NULL, 0, &rate_mv,
                            !cpi->sf.mv_sf.disable_second_mv,
                            NUM_JOINT_ME_REFINE_ITER);

    for (int ref = 0; ref < 2; ++ref) {
      struct buf_2d ref_buf = { NULL, ref_frame_ptr[ref]->y_buffer,
                                ref_frame_ptr[ref]->y_width,
                                ref_frame_ptr[ref]->y_height,
                                ref_frame_ptr[ref]->y_stride };
      InterPredParams inter_pred_params;
      av1_init_inter_params(&inter_pred_params, bw, bh, mi_row * MI_SIZE,
                            mi_col * MI_SIZE, 0, 0, xd->bd, is_cur_buf_hbd(xd),
                            0, &tpl_data->sf, &ref_buf, kernel);
      av1_init_comp_mode(&inter_pred_params);

      inter_pred_params.conv_params = get_conv_params_no_round(
          ref, 0, xd->tmp_conv_dst, MAX_SB_SIZE, 1, xd->bd);

      av1_enc_build_one_inter_predictor(predictor, bw, &tmp_mv[ref].as_mv,
                                        &inter_pred_params);
    }
    inter_cost =
        tpl_get_satd_cost(bd_info, src_diff, bw, src_mb_buffer, src_stride,
                          predictor, bw, coeff, bw, bh, tx_size);
    if (inter_cost < best_inter_cost) {
      best_cmp_rf_idx = cmp_rf_idx;
      best_inter_cost = inter_cost;
      best_mv[0] = tmp_mv[0];
      best_mv[1] = tmp_mv[1];
    }
  }

  if (best_cmp_rf_idx != -1 && best_inter_cost < best_intra_cost) {
    best_mode = NEW_NEWMV;
    const int best_rf_idx0 = comp_ref_frames[best_cmp_rf_idx][0];
    const int best_rf_idx1 = comp_ref_frames[best_cmp_rf_idx][1];
    xd->mi[0]->ref_frame[0] = best_rf_idx0 + LAST_FRAME;
    xd->mi[0]->ref_frame[1] = best_rf_idx1 + LAST_FRAME;
  }

  if (best_inter_cost < INT32_MAX && is_inter_mode(best_mode)) {
    xd->mi[0]->mv[0].as_int = best_mv[0].as_int;
    xd->mi[0]->mv[1].as_int = best_mv[1].as_int;
    const YV12_BUFFER_CONFIG *ref_frame_ptr[2] = {
      best_cmp_rf_idx >= 0
          ? tpl_data->src_ref_frame[comp_ref_frames[best_cmp_rf_idx][0]]
          : tpl_data->src_ref_frame[best_rf_idx],
      best_cmp_rf_idx >= 0
          ? tpl_data->src_ref_frame[comp_ref_frames[best_cmp_rf_idx][1]]
          : NULL,
    };
    rate_cost = 1;
    get_rate_distortion(&rate_cost, &recon_error, &pred_error, src_diff, coeff,
                        qcoeff, dqcoeff, cm, x, ref_frame_ptr, rec_buffer_pool,
                        rec_stride_pool, tx_size, best_mode, mi_row, mi_col,
                        use_y_only_rate_distortion, 0 /*do_recon*/, NULL);
    tpl_stats->srcrf_rate = rate_cost;
  }

  best_intra_cost = AOMMAX(best_intra_cost, 1);
  best_inter_cost = AOMMIN(best_intra_cost, best_inter_cost);
  tpl_stats->inter_cost = best_inter_cost;
  tpl_stats->intra_cost = best_intra_cost;

  tpl_stats->srcrf_dist = recon_error << TPL_DEP_COST_SCALE_LOG2;
  tpl_stats->srcrf_sse = pred_error << TPL_DEP_COST_SCALE_LOG2;

  // Final encode
  rate_cost = 0;
  const YV12_BUFFER_CONFIG *ref_frame_ptr[2];

  ref_frame_ptr[0] =
      best_mode == NEW_NEWMV
          ? tpl_data->ref_frame[comp_ref_frames[best_cmp_rf_idx][0]]
      : best_rf_idx >= 0 ? tpl_data->ref_frame[best_rf_idx]
                         : NULL;
  ref_frame_ptr[1] =
      best_mode == NEW_NEWMV
          ? tpl_data->ref_frame[comp_ref_frames[best_cmp_rf_idx][1]]
          : NULL;
  get_rate_distortion(&rate_cost, &recon_error, &pred_error, src_diff, coeff,
                      qcoeff, dqcoeff, cm, x, ref_frame_ptr, rec_buffer_pool,
                      rec_stride_pool, tx_size, best_mode, mi_row, mi_col,
                      use_y_only_rate_distortion, 1 /*do_recon*/,
                      tpl_txfm_stats);

  tpl_stats->recrf_dist = recon_error << TPL_DEP_COST_SCALE_LOG2;
  tpl_stats->recrf_sse = pred_error << TPL_DEP_COST_SCALE_LOG2;
  tpl_stats->recrf_rate = rate_cost;

  if (!is_inter_mode(best_mode)) {
    tpl_stats->srcrf_dist = recon_error << TPL_DEP_COST_SCALE_LOG2;
    tpl_stats->srcrf_rate = rate_cost;
    tpl_stats->srcrf_sse = pred_error << TPL_DEP_COST_SCALE_LOG2;
  }

  tpl_stats->recrf_dist = AOMMAX(tpl_stats->srcrf_dist, tpl_stats->recrf_dist);
  tpl_stats->recrf_rate = AOMMAX(tpl_stats->srcrf_rate, tpl_stats->recrf_rate);

  if (best_mode == NEW_NEWMV) {
    ref_frame_ptr[0] = tpl_data->ref_frame[comp_ref_frames[best_cmp_rf_idx][0]];
    ref_frame_ptr[1] =
        tpl_data->src_ref_frame[comp_ref_frames[best_cmp_rf_idx][1]];
    get_rate_distortion(&rate_cost, &recon_error, &pred_error, src_diff, coeff,
                        qcoeff, dqcoeff, cm, x, ref_frame_ptr, rec_buffer_pool,
                        rec_stride_pool, tx_size, best_mode, mi_row, mi_col,
                        use_y_only_rate_distortion, 1 /*do_recon*/, NULL);
    tpl_stats->cmp_recrf_dist[0] = recon_error << TPL_DEP_COST_SCALE_LOG2;
    tpl_stats->cmp_recrf_rate[0] = rate_cost;

    tpl_stats->cmp_recrf_dist[0] =
        AOMMAX(tpl_stats->srcrf_dist, tpl_stats->cmp_recrf_dist[0]);
    tpl_stats->cmp_recrf_rate[0] =
        AOMMAX(tpl_stats->srcrf_rate, tpl_stats->cmp_recrf_rate[0]);

    tpl_stats->cmp_recrf_dist[0] =
        AOMMIN(tpl_stats->recrf_dist, tpl_stats->cmp_recrf_dist[0]);
    tpl_stats->cmp_recrf_rate[0] =
        AOMMIN(tpl_stats->recrf_rate, tpl_stats->cmp_recrf_rate[0]);

    rate_cost = 0;
    ref_frame_ptr[0] =
        tpl_data->src_ref_frame[comp_ref_frames[best_cmp_rf_idx][0]];
    ref_frame_ptr[1] = tpl_data->ref_frame[comp_ref_frames[best_cmp_rf_idx][1]];
    get_rate_distortion(&rate_cost, &recon_error, &pred_error, src_diff, coeff,
                        qcoeff, dqcoeff, cm, x, ref_frame_ptr, rec_buffer_pool,
                        rec_stride_pool, tx_size, best_mode, mi_row, mi_col,
                        use_y_only_rate_distortion, 1 /*do_recon*/, NULL);
    tpl_stats->cmp_recrf_dist[1] = recon_error << TPL_DEP_COST_SCALE_LOG2;
    tpl_stats->cmp_recrf_rate[1] = rate_cost;

    tpl_stats->cmp_recrf_dist[1] =
        AOMMAX(tpl_stats->srcrf_dist, tpl_stats->cmp_recrf_dist[1]);
    tpl_stats->cmp_recrf_rate[1] =
        AOMMAX(tpl_stats->srcrf_rate, tpl_stats->cmp_recrf_rate[1]);

    tpl_stats->cmp_recrf_dist[1] =
        AOMMIN(tpl_stats->recrf_dist, tpl_stats->cmp_recrf_dist[1]);
    tpl_stats->cmp_recrf_rate[1] =
        AOMMIN(tpl_stats->recrf_rate, tpl_stats->cmp_recrf_rate[1]);
  }

  if (best_mode == NEWMV) {
    tpl_stats->mv[best_rf_idx] = best_mv[0];
    tpl_stats->ref_frame_index[0] = best_rf_idx;
    tpl_stats->ref_frame_index[1] = NONE_FRAME;
  } else if (best_mode == NEW_NEWMV) {
    tpl_stats->ref_frame_index[0] = comp_ref_frames[best_cmp_rf_idx][0];
    tpl_stats->ref_frame_index[1] = comp_ref_frames[best_cmp_rf_idx][1];
    tpl_stats->mv[tpl_stats->ref_frame_index[0]] = best_mv[0];
    tpl_stats->mv[tpl_stats->ref_frame_index[1]] = best_mv[1];
  }

  for (int idy = 0; idy < mi_height; ++idy) {
    for (int idx = 0; idx < mi_width; ++idx) {
      if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > idx &&
          (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > idy) {
        xd->mi[idx + idy * cm->mi_params.mi_stride] = xd->mi[0];
      }
    }
  }
}

static int round_floor(int ref_pos, int bsize_pix) {
  int round;
  if (ref_pos < 0)
    round = -(1 + (-ref_pos - 1) / bsize_pix);
  else
    round = ref_pos / bsize_pix;

  return round;
}

int av1_get_overlap_area(int row_a, int col_a, int row_b, int col_b, int width,
                         int height) {
  int min_row = AOMMAX(row_a, row_b);
  int max_row = AOMMIN(row_a + height, row_b + height);
  int min_col = AOMMAX(col_a, col_b);
  int max_col = AOMMIN(col_a + width, col_b + width);
  if (min_row < max_row && min_col < max_col) {
    return (max_row - min_row) * (max_col - min_col);
  }
  return 0;
}

int av1_tpl_ptr_pos(int mi_row, int mi_col, int stride, uint8_t right_shift) {
  return (mi_row >> right_shift) * stride + (mi_col >> right_shift);
}

int64_t av1_delta_rate_cost(int64_t delta_rate, int64_t recrf_dist,
                            int64_t srcrf_dist, int pix_num) {
  double beta = (double)srcrf_dist / recrf_dist;
  int64_t rate_cost = delta_rate;

  if (srcrf_dist <= 128) return rate_cost;

  double dr =
      (double)(delta_rate >> (TPL_DEP_COST_SCALE_LOG2 + AV1_PROB_COST_SHIFT)) /
      pix_num;

  double log_den = log(beta) / log(2.0) + 2.0 * dr;

  if (log_den > log(10.0) / log(2.0)) {
    rate_cost = (int64_t)((log(1.0 / beta) * pix_num) / log(2.0) / 2.0);
    rate_cost <<= (TPL_DEP_COST_SCALE_LOG2 + AV1_PROB_COST_SHIFT);
    return rate_cost;
  }

  double num = pow(2.0, log_den);
  double den = num * beta + (1 - beta) * beta;

  rate_cost = (int64_t)((pix_num * log(num / den)) / log(2.0) / 2.0);

  rate_cost <<= (TPL_DEP_COST_SCALE_LOG2 + AV1_PROB_COST_SHIFT);

  return rate_cost;
}

static inline void tpl_model_update_b(TplParams *const tpl_data, int mi_row,
                                      int mi_col, const BLOCK_SIZE bsize,
                                      int frame_idx, int ref) {
  TplDepFrame *tpl_frame_ptr = &tpl_data->tpl_frame[frame_idx];
  TplDepStats *tpl_ptr = tpl_frame_ptr->tpl_stats_ptr;
  TplDepFrame *tpl_frame = tpl_data->tpl_frame;
  const uint8_t block_mis_log2 = tpl_data->tpl_stats_block_mis_log2;
  TplDepStats *tpl_stats_ptr = &tpl_ptr[av1_tpl_ptr_pos(
      mi_row, mi_col, tpl_frame->stride, block_mis_log2)];

  int is_compound = tpl_stats_ptr->ref_frame_index[1] >= 0;

  if (tpl_stats_ptr->ref_frame_index[ref] < 0) return;
  const int ref_frame_index = tpl_stats_ptr->ref_frame_index[ref];
  TplDepFrame *ref_tpl_frame =
      &tpl_frame[tpl_frame[frame_idx].ref_map_index[ref_frame_index]];
  TplDepStats *ref_stats_ptr = ref_tpl_frame->tpl_stats_ptr;

  if (tpl_frame[frame_idx].ref_map_index[ref_frame_index] < 0) return;

  const FULLPEL_MV full_mv =
      get_fullmv_from_mv(&tpl_stats_ptr->mv[ref_frame_index].as_mv);
  const int ref_pos_row = mi_row * MI_SIZE + full_mv.row;
  const int ref_pos_col = mi_col * MI_SIZE + full_mv.col;

  const int bw = 4 << mi_size_wide_log2[bsize];
  const int bh = 4 << mi_size_high_log2[bsize];
  const int mi_height = mi_size_high[bsize];
  const int mi_width = mi_size_wide[bsize];
  const int pix_num = bw * bh;

  // top-left on grid block location in pixel
  int grid_pos_row_base = round_floor(ref_pos_row, bh) * bh;
  int grid_pos_col_base = round_floor(ref_pos_col, bw) * bw;
  int block;

  int64_t srcrf_dist = is_compound ? tpl_stats_ptr->cmp_recrf_dist[!ref]
                                   : tpl_stats_ptr->srcrf_dist;
  int64_t srcrf_rate =
      is_compound
          ? (tpl_stats_ptr->cmp_recrf_rate[!ref] << TPL_DEP_COST_SCALE_LOG2)
          : (tpl_stats_ptr->srcrf_rate << TPL_DEP_COST_SCALE_LOG2);

  int64_t cur_dep_dist = tpl_stats_ptr->recrf_dist - srcrf_dist;
  int64_t mc_dep_dist =
      (int64_t)(tpl_stats_ptr->mc_dep_dist *
                ((double)(tpl_stats_ptr->recrf_dist - srcrf_dist) /
                 tpl_stats_ptr->recrf_dist));
  int64_t delta_rate =
      (tpl_stats_ptr->recrf_rate << TPL_DEP_COST_SCALE_LOG2) - srcrf_rate;
  int64_t mc_dep_rate =
      av1_delta_rate_cost(tpl_stats_ptr->mc_dep_rate, tpl_stats_ptr->recrf_dist,
                          srcrf_dist, pix_num);

  for (block = 0; block < 4; ++block) {
    int grid_pos_row = grid_pos_row_base + bh * (block >> 1);
    int grid_pos_col = grid_pos_col_base + bw * (block & 0x01);

    if (grid_pos_row >= 0 && grid_pos_row < ref_tpl_frame->mi_rows * MI_SIZE &&
        grid_pos_col >= 0 && grid_pos_col < ref_tpl_frame->mi_cols * MI_SIZE) {
      int overlap_area = av1_get_overlap_area(grid_pos_row, grid_pos_col,
                                              ref_pos_row, ref_pos_col, bw, bh);
      int ref_mi_row = round_floor(grid_pos_row, bh) * mi_height;
      int ref_mi_col = round_floor(grid_pos_col, bw) * mi_width;
      assert((1 << block_mis_log2) == mi_height);
      assert((1 << block_mis_log2) == mi_width);
      TplDepStats *des_stats = &ref_stats_ptr[av1_tpl_ptr_pos(
          ref_mi_row, ref_mi_col, ref_tpl_frame->stride, block_mis_log2)];
      des_stats->mc_dep_dist +=
          ((cur_dep_dist + mc_dep_dist) * overlap_area) / pix_num;
      des_stats->mc_dep_rate +=
          ((delta_rate + mc_dep_rate) * overlap_area) / pix_num;
    }
  }
}

static inline void tpl_model_update(TplParams *const tpl_data, int mi_row,
                                    int mi_col, int frame_idx) {
  const BLOCK_SIZE tpl_stats_block_size =
      convert_length_to_bsize(MI_SIZE << tpl_data->tpl_stats_block_mis_log2);
  tpl_model_update_b(tpl_data, mi_row, mi_col, tpl_stats_block_size, frame_idx,
                     0);
  tpl_model_update_b(tpl_data, mi_row, mi_col, tpl_stats_block_size, frame_idx,
                     1);
}

static inline void tpl_model_store(TplDepStats *tpl_stats_ptr, int mi_row,
                                   int mi_col, int stride,
                                   const TplDepStats *src_stats,
                                   uint8_t block_mis_log2) {
  int index = av1_tpl_ptr_pos(mi_row, mi_col, stride, block_mis_log2);
  TplDepStats *tpl_ptr = &tpl_stats_ptr[index];
  *tpl_ptr = *src_stats;
  tpl_ptr->intra_cost = AOMMAX(1, tpl_ptr->intra_cost);
  tpl_ptr->inter_cost = AOMMAX(1, tpl_ptr->inter_cost);
  tpl_ptr->srcrf_dist = AOMMAX(1, tpl_ptr->srcrf_dist);
  tpl_ptr->srcrf_sse = AOMMAX(1, tpl_ptr->srcrf_sse);
  tpl_ptr->recrf_dist = AOMMAX(1, tpl_ptr->recrf_dist);
  tpl_ptr->srcrf_rate = AOMMAX(1, tpl_ptr->srcrf_rate);
  tpl_ptr->recrf_rate = AOMMAX(1, tpl_ptr->recrf_rate);
  tpl_ptr->cmp_recrf_dist[0] = AOMMAX(1, tpl_ptr->cmp_recrf_dist[0]);
  tpl_ptr->cmp_recrf_dist[1] = AOMMAX(1, tpl_ptr->cmp_recrf_dist[1]);
  tpl_ptr->cmp_recrf_rate[0] = AOMMAX(1, tpl_ptr->cmp_recrf_rate[0]);
  tpl_ptr->cmp_recrf_rate[1] = AOMMAX(1, tpl_ptr->cmp_recrf_rate[1]);
}

// Reset the ref and source frame pointers of tpl_data.
static inline void tpl_reset_src_ref_frames(TplParams *tpl_data) {
  for (int i = 0; i < INTER_REFS_PER_FRAME; ++i) {
    tpl_data->ref_frame[i] = NULL;
    tpl_data->src_ref_frame[i] = NULL;
  }
}

static inline int get_gop_length(const GF_GROUP *gf_group) {
  int gop_length = AOMMIN(gf_group->size, MAX_TPL_FRAME_IDX - 1);
  return gop_length;
}

// Initialize the mc_flow parameters used in computing tpl data.
static inline void init_mc_flow_dispenser(AV1_COMP *cpi, int frame_idx,
                                          int pframe_qindex) {
  TplParams *const tpl_data = &cpi->ppi->tpl_data;
  TplDepFrame *tpl_frame = &tpl_data->tpl_frame[frame_idx];
  const YV12_BUFFER_CONFIG *this_frame = tpl_frame->gf_picture;
  const YV12_BUFFER_CONFIG *ref_frames_ordered[INTER_REFS_PER_FRAME];
  uint32_t ref_frame_display_indices[INTER_REFS_PER_FRAME];
  const GF_GROUP *gf_group = &cpi->ppi->gf_group;
  TPL_SPEED_FEATURES *tpl_sf = &cpi->sf.tpl_sf;
  int ref_pruning_enabled = is_frame_eligible_for_ref_pruning(
      gf_group, cpi->sf.inter_sf.selective_ref_frame,
      tpl_sf->prune_ref_frames_in_tpl, frame_idx);
  int gop_length = get_gop_length(gf_group);
  int ref_frame_flags;
  AV1_COMMON *cm = &cpi->common;
  int rdmult, idx;
  ThreadData *td = &cpi->td;
  MACROBLOCK *x = &td->mb;
  MACROBLOCKD *xd = &x->e_mbd;
  TplTxfmStats *tpl_txfm_stats = &td->tpl_txfm_stats;
  tpl_data->frame_idx = frame_idx;
  tpl_reset_src_ref_frames(tpl_data);
  av1_tile_init(&xd->tile, cm, 0, 0);

  const int boost_index = AOMMIN(15, (cpi->ppi->p_rc.gfu_boost / 100));
  const int layer_depth = AOMMIN(gf_group->layer_depth[cpi->gf_frame_index], 6);
  const FRAME_TYPE frame_type = cm->current_frame.frame_type;

  // Setup scaling factor
  av1_setup_scale_factors_for_frame(
      &tpl_data->sf, this_frame->y_crop_width, this_frame->y_crop_height,
      this_frame->y_crop_width, this_frame->y_crop_height);

  xd->cur_buf = this_frame;

  for (idx = 0; idx < INTER_REFS_PER_FRAME; ++idx) {
    TplDepFrame *tpl_ref_frame =
        &tpl_data->tpl_frame[tpl_frame->ref_map_index[idx]];
    tpl_data->ref_frame[idx] = tpl_ref_frame->rec_picture;
    tpl_data->src_ref_frame[idx] = tpl_ref_frame->gf_picture;
    ref_frame_display_indices[idx] = tpl_ref_frame->frame_display_index;
  }

  // Store the reference frames based on priority order
  for (int i = 0; i < INTER_REFS_PER_FRAME; ++i) {
    ref_frames_ordered[i] =
        tpl_data->ref_frame[ref_frame_priority_order[i] - 1];
  }

  // Work out which reference frame slots may be used.
  ref_frame_flags =
      get_ref_frame_flags(&cpi->sf, is_one_pass_rt_params(cpi),
                          ref_frames_ordered, cpi->ext_flags.ref_frame_flags);

  enforce_max_ref_frames(cpi, &ref_frame_flags, ref_frame_display_indices,
                         tpl_frame->frame_display_index);

  // Prune reference frames
  for (idx = 0; idx < INTER_REFS_PER_FRAME; ++idx) {
    if ((ref_frame_flags & (1 << idx)) == 0) {
      tpl_data->ref_frame[idx] = NULL;
    }
  }

  // Skip motion estimation w.r.t. reference frames which are not
  // considered in RD search, using "selective_ref_frame" speed feature.
  // The reference frame pruning is not enabled for frames beyond the gop
  // length, as there are fewer reference frames and the reference frames
  // differ from the frames considered during RD search.
  if (ref_pruning_enabled && (frame_idx < gop_length)) {
    for (idx = 0; idx < INTER_REFS_PER_FRAME; ++idx) {
      const MV_REFERENCE_FRAME refs[2] = { idx + 1, NONE_FRAME };
      if (prune_ref_by_selective_ref_frame(cpi, NULL, refs,
                                           ref_frame_display_indices)) {
        tpl_data->ref_frame[idx] = NULL;
      }
    }
  }

  // Make a temporary mbmi for tpl model
  MB_MODE_INFO mbmi;
  memset(&mbmi, 0, sizeof(mbmi));
  MB_MODE_INFO *mbmi_ptr = &mbmi;
  xd->mi = &mbmi_ptr;

  xd->block_ref_scale_factors[0] = &tpl_data->sf;
  xd->block_ref_scale_factors[1] = &tpl_data->sf;

  const int base_qindex =
      cpi->use_ducky_encode ? gf_group->q_val[frame_idx] : pframe_qindex;
  // Get rd multiplier set up.
  rdmult = (int)av1_compute_rd_mult(
      base_qindex, cm->seq_params->bit_depth,
      cpi->ppi->gf_group.update_type[cpi->gf_frame_index], layer_depth,
      boost_index, frame_type, cpi->oxcf.q_cfg.use_fixed_qp_offsets,
      is_stat_consumption_stage(cpi));

  if (rdmult < 1) rdmult = 1;
  av1_set_error_per_bit(&x->errorperbit, rdmult);
  av1_set_sad_per_bit(cpi, &x->sadperbit, base_qindex);

  tpl_frame->is_valid = 1;

  cm->quant_params.base_qindex = base_qindex;
  av1_frame_init_quantizer(cpi);

  const BitDepthInfo bd_info = get_bit_depth_info(xd);
  const FRAME_UPDATE_TYPE update_type =
      gf_group->update_type[cpi->gf_frame_index];
  tpl_frame->base_rdmult = av1_compute_rd_mult_based_on_qindex(
                               bd_info.bit_depth, update_type, base_qindex) /
                           6;

  if (cpi->use_ducky_encode)
    tpl_frame->base_rdmult = gf_group->rdmult_val[frame_idx];

  av1_init_tpl_txfm_stats(tpl_txfm_stats);

  // Initialize x->mbmi_ext when compound predictions are enabled.
  if (tpl_sf->allow_compound_pred) av1_zero(x->mbmi_ext);

  // Set the pointer to null since mbmi is only allocated inside this function.
  assert(xd->mi == &mbmi_ptr);
  xd->mi = NULL;

  // Tpl module is called before the setting of speed features at frame level.
  // Thus, turning off this speed feature for key frame is done here and not
  // integrated into the speed feature setting itself.
  const int layer_depth_th = (tpl_sf->use_sad_for_mode_decision == 1) ? 5 : 0;
  tpl_frame->use_pred_sad =
      tpl_sf->use_sad_for_mode_decision &&
      gf_group->update_type[cpi->gf_frame_index] != KF_UPDATE &&
      gf_group->layer_depth[frame_idx] >= layer_depth_th;
}

// This function stores the motion estimation dependencies of all the blocks in
// a row
void av1_mc_flow_dispenser_row(AV1_COMP *cpi, TplTxfmStats *tpl_txfm_stats,
                               TplBuffers *tpl_tmp_buffers, MACROBLOCK *x,
                               int mi_row, BLOCK_SIZE bsize, TX_SIZE tx_size) {
  AV1_COMMON *const cm = &cpi->common;
  MultiThreadInfo *const mt_info = &cpi->mt_info;
  AV1TplRowMultiThreadInfo *const tpl_row_mt = &mt_info->tpl_row_mt;
  const CommonModeInfoParams *const mi_params = &cm->mi_params;
  const int mi_width = mi_size_wide[bsize];
  TplParams *const tpl_data = &cpi->ppi->tpl_data;
  TplDepFrame *tpl_frame = &tpl_data->tpl_frame[tpl_data->frame_idx];
  MACROBLOCKD *xd = &x->e_mbd;

  const int tplb_cols_in_tile =
      ROUND_POWER_OF_TWO(mi_params->mi_cols, mi_size_wide_log2[bsize]);
  const int tplb_row = ROUND_POWER_OF_TWO(mi_row, mi_size_high_log2[bsize]);
  assert(mi_size_high[bsize] == (1 << tpl_data->tpl_stats_block_mis_log2));
  assert(mi_size_wide[bsize] == (1 << tpl_data->tpl_stats_block_mis_log2));

  for (int mi_col = 0, tplb_col_in_tile = 0; mi_col < mi_params->mi_cols;
       mi_col += mi_width, tplb_col_in_tile++) {
    (*tpl_row_mt->sync_read_ptr)(&tpl_data->tpl_mt_sync, tplb_row,
                                 tplb_col_in_tile);

#if CONFIG_MULTITHREAD
    if (mt_info->num_workers > 1) {
      pthread_mutex_lock(tpl_row_mt->mutex_);
      const bool tpl_mt_exit = tpl_row_mt->tpl_mt_exit;
      pthread_mutex_unlock(tpl_row_mt->mutex_);
      // Exit in case any worker has encountered an error.
      if (tpl_mt_exit) return;
    }
#endif

    TplDepStats tpl_stats;

    // Motion estimation column boundary
    av1_set_mv_col_limits(mi_params, &x->mv_limits, mi_col, mi_width,
                          tpl_data->border_in_pixels);
    xd->mb_to_left_edge = -GET_MV_SUBPEL(mi_col * MI_SIZE);
    xd->mb_to_right_edge =
        GET_MV_SUBPEL(mi_params->mi_cols - mi_width - mi_col);
    mode_estimation(cpi, tpl_txfm_stats, tpl_tmp_buffers, x, mi_row, mi_col,
                    bsize, tx_size, &tpl_stats);

    // Motion flow dependency dispenser.
    tpl_model_store(tpl_frame->tpl_stats_ptr, mi_row, mi_col, tpl_frame->stride,
                    &tpl_stats, tpl_data->tpl_stats_block_mis_log2);
    (*tpl_row_mt->sync_write_ptr)(&tpl_data->tpl_mt_sync, tplb_row,
                                  tplb_col_in_tile, tplb_cols_in_tile);
  }
}

static inline void mc_flow_dispenser(AV1_COMP *cpi) {
  AV1_COMMON *cm = &cpi->common;
  const CommonModeInfoParams *const mi_params = &cm->mi_params;
  ThreadData *td = &cpi->td;
  MACROBLOCK *x = &td->mb;
  MACROBLOCKD *xd = &x->e_mbd;
  const BLOCK_SIZE bsize =
      convert_length_to_bsize(cpi->ppi->tpl_data.tpl_bsize_1d);
  const TX_SIZE tx_size = max_txsize_lookup[bsize];
  const int mi_height = mi_size_high[bsize];
  for (int mi_row = 0; mi_row < mi_params->mi_rows; mi_row += mi_height) {
    // Motion estimation row boundary
    av1_set_mv_row_limits(mi_params, &x->mv_limits, mi_row, mi_height,
                          cpi->ppi->tpl_data.border_in_pixels);
    xd->mb_to_top_edge = -GET_MV_SUBPEL(mi_row * MI_SIZE);
    xd->mb_to_bottom_edge =
        GET_MV_SUBPEL((mi_params->mi_rows - mi_height - mi_row) * MI_SIZE);
    av1_mc_flow_dispenser_row(cpi, &td->tpl_txfm_stats, &td->tpl_tmp_buffers, x,
                              mi_row, bsize, tx_size);
  }
}

static void mc_flow_synthesizer(TplParams *tpl_data, int frame_idx, int mi_rows,
                                int mi_cols) {
  if (!frame_idx) {
    return;
  }
  const BLOCK_SIZE bsize = convert_length_to_bsize(tpl_data->tpl_bsize_1d);
  const int mi_height = mi_size_high[bsize];
  const int mi_width = mi_size_wide[bsize];
  assert(mi_height == (1 << tpl_data->tpl_stats_block_mis_log2));
  assert(mi_width == (1 << tpl_data->tpl_stats_block_mis_log2));

  for (int mi_row = 0; mi_row < mi_rows; mi_row += mi_height) {
    for (int mi_col = 0; mi_col < mi_cols; mi_col += mi_width) {
      tpl_model_update(tpl_data, mi_row, mi_col, frame_idx);
    }
  }
}

static inline void init_gop_frames_for_tpl(
    AV1_COMP *cpi, const EncodeFrameParams *const init_frame_params,
    GF_GROUP *gf_group, int *tpl_group_frames, int *pframe_qindex) {
  AV1_COMMON *cm = &cpi->common;
  assert(cpi->gf_frame_index == 0);
  *pframe_qindex = 0;

  RefFrameMapPair ref_frame_map_pairs[REF_FRAMES];
  init_ref_map_pair(cpi, ref_frame_map_pairs);

  int remapped_ref_idx[REF_FRAMES];

  EncodeFrameParams frame_params = *init_frame_params;
  TplParams *const tpl_data = &cpi->ppi->tpl_data;

  int ref_picture_map[REF_FRAMES];

  for (int i = 0; i < REF_FRAMES; ++i) {
    if (frame_params.frame_type == KEY_FRAME) {
      tpl_data->tpl_frame[-i - 1].gf_picture = NULL;
      tpl_data->tpl_frame[-i - 1].rec_picture = NULL;
      tpl_data->tpl_frame[-i - 1].frame_display_index = 0;
    } else {
      tpl_data->tpl_frame[-i - 1].gf_picture = &cm->ref_frame_map[i]->buf;
      tpl_data->tpl_frame[-i - 1].rec_picture = &cm->ref_frame_map[i]->buf;
      tpl_data->tpl_frame[-i - 1].frame_display_index =
          cm->ref_frame_map[i]->display_order_hint;
    }

    ref_picture_map[i] = -i - 1;
  }

  *tpl_group_frames = 0;

  int gf_index;
  int process_frame_count = 0;
  const int gop_length = get_gop_length(gf_group);

  for (gf_index = 0; gf_index < gop_length; ++gf_index) {
    TplDepFrame *tpl_frame = &tpl_data->tpl_frame[gf_index];
    FRAME_UPDATE_TYPE frame_update_type = gf_group->update_type[gf_index];
    int lookahead_index =
        gf_group->cur_frame_idx[gf_index] + gf_group->arf_src_offset[gf_index];
    frame_params.show_frame = frame_update_type != ARF_UPDATE &&
                              frame_update_type != INTNL_ARF_UPDATE;
    frame_params.show_existing_frame =
        frame_update_type == INTNL_OVERLAY_UPDATE ||
        frame_update_type == OVERLAY_UPDATE;
    frame_params.frame_type = gf_group->frame_type[gf_index];

    if (frame_update_type == LF_UPDATE)
      *pframe_qindex = gf_group->q_val[gf_index];

    const struct lookahead_entry *buf = av1_lookahead_peek(
        cpi->ppi->lookahead, lookahead_index, cpi->compressor_stage);
    if (buf == NULL) break;
    tpl_frame->gf_picture = &buf->img;

    // Use filtered frame buffer if available. This will make tpl stats more
    // precise.
    FRAME_DIFF frame_diff;
    const YV12_BUFFER_CONFIG *tf_buf =
        av1_tf_info_get_filtered_buf(&cpi->ppi->tf_info, gf_index, &frame_diff);
    if (tf_buf != NULL) {
      tpl_frame->gf_picture = tf_buf;
    }

    // 'cm->current_frame.frame_number' is the display number
    // of the current frame.
    // 'lookahead_index' is frame offset within the gf group.
    // 'lookahead_index + cm->current_frame.frame_number'
    // is the display index of the frame.
    tpl_frame->frame_display_index =
        lookahead_index + cm->current_frame.frame_number;
    assert(buf->display_idx ==
           cpi->frame_index_set.show_frame_count + lookahead_index);

    if (frame_update_type != OVERLAY_UPDATE &&
        frame_update_type != INTNL_OVERLAY_UPDATE) {
      tpl_frame->rec_picture = &tpl_data->tpl_rec_pool[process_frame_count];
      tpl_frame->tpl_stats_ptr = tpl_data->tpl_stats_pool[process_frame_count];
      ++process_frame_count;
    }
    const int true_disp = (int)(tpl_frame->frame_display_index);

    av1_get_ref_frames(ref_frame_map_pairs, true_disp, cpi, gf_index, 0,
                       remapped_ref_idx);

    int refresh_mask =
        av1_get_refresh_frame_flags(cpi, &frame_params, frame_update_type,
                                    gf_index, true_disp, ref_frame_map_pairs);

    // Make the frames marked as is_frame_non_ref to non-reference frames.
    if (cpi->ppi->gf_group.is_frame_non_ref[gf_index]) refresh_mask = 0;

    int refresh_frame_map_index = av1_get_refresh_ref_frame_map(refresh_mask);

    if (refresh_frame_map_index < REF_FRAMES &&
        refresh_frame_map_index != INVALID_IDX) {
      ref_frame_map_pairs[refresh_frame_map_index].disp_order =
          AOMMAX(0, true_disp);
      ref_frame_map_pairs[refresh_frame_map_index].pyr_level =
          get_true_pyr_level(gf_group->layer_depth[gf_index], true_disp,
                             cpi->ppi->gf_group.max_layer_depth);
    }

    for (int i = LAST_FRAME; i <= ALTREF_FRAME; ++i)
      tpl_frame->ref_map_index[i - LAST_FRAME] =
          ref_picture_map[remapped_ref_idx[i - LAST_FRAME]];

    if (refresh_mask) ref_picture_map[refresh_frame_map_index] = gf_index;

    ++*tpl_group_frames;
  }

  const int tpl_extend = cpi->oxcf.gf_cfg.lag_in_frames - MAX_GF_INTERVAL;
  int extend_frame_count = 0;
  int extend_frame_length = AOMMIN(
      tpl_extend, cpi->rc.frames_to_key - cpi->ppi->p_rc.baseline_gf_interval);

  int frame_display_index = gf_group->cur_frame_idx[gop_length - 1] +
                            gf_group->arf_src_offset[gop_length - 1] + 1;

  for (;
       gf_index < MAX_TPL_FRAME_IDX && extend_frame_count < extend_frame_length;
       ++gf_index) {
    TplDepFrame *tpl_frame = &tpl_data->tpl_frame[gf_index];
    FRAME_UPDATE_TYPE frame_update_type = LF_UPDATE;
    frame_params.show_frame = frame_update_type != ARF_UPDATE &&
                              frame_update_type != INTNL_ARF_UPDATE;
    frame_params.show_existing_frame =
        frame_update_type == INTNL_OVERLAY_UPDATE;
    frame_params.frame_type = INTER_FRAME;

    int lookahead_index = frame_display_index;
    struct lookahead_entry *buf = av1_lookahead_peek(
        cpi->ppi->lookahead, lookahead_index, cpi->compressor_stage);

    if (buf == NULL) break;

    tpl_frame->gf_picture = &buf->img;
    tpl_frame->rec_picture = &tpl_data->tpl_rec_pool[process_frame_count];
    tpl_frame->tpl_stats_ptr = tpl_data->tpl_stats_pool[process_frame_count];
    // 'cm->current_frame.frame_number' is the display number
    // of the current frame.
    // 'frame_display_index' is frame offset within the gf group.
    // 'frame_display_index + cm->current_frame.frame_number'
    // is the display index of the frame.
    tpl_frame->frame_display_index =
        frame_display_index + cm->current_frame.frame_number;

    ++process_frame_count;

    gf_group->update_type[gf_index] = LF_UPDATE;

#if CONFIG_BITRATE_ACCURACY && CONFIG_THREE_PASS
    if (cpi->oxcf.pass == AOM_RC_SECOND_PASS) {
      if (cpi->oxcf.rc_cfg.mode == AOM_Q) {
        *pframe_qindex = cpi->oxcf.rc_cfg.cq_level;
      } else if (cpi->oxcf.rc_cfg.mode == AOM_VBR) {
        // TODO(angiebird): Find a more adaptive method to decide pframe_qindex
        // override the pframe_qindex in the second pass when bitrate accuracy
        // is on. We found that setting this pframe_qindex make the tpl stats
        // more stable.
        *pframe_qindex = 128;
      }
    }
#endif  // CONFIG_BITRATE_ACCURACY && CONFIG_THREE_PASS
    gf_group->q_val[gf_index] = *pframe_qindex;
    const int true_disp = (int)(tpl_frame->frame_display_index);
    av1_get_ref_frames(ref_frame_map_pairs, true_disp, cpi, gf_index, 0,
                       remapped_ref_idx);
    int refresh_mask =
        av1_get_refresh_frame_flags(cpi, &frame_params, frame_update_type,
                                    gf_index, true_disp, ref_frame_map_pairs);
    int refresh_frame_map_index = av1_get_refresh_ref_frame_map(refresh_mask);

    if (refresh_frame_map_index < REF_FRAMES &&
        refresh_frame_map_index != INVALID_IDX) {
      ref_frame_map_pairs[refresh_frame_map_index].disp_order =
          AOMMAX(0, true_disp);
      ref_frame_map_pairs[refresh_frame_map_index].pyr_level =
          get_true_pyr_level(gf_group->layer_depth[gf_index], true_disp,
                             cpi->ppi->gf_group.max_layer_depth);
    }

    for (int i = LAST_FRAME; i <= ALTREF_FRAME; ++i)
      tpl_frame->ref_map_index[i - LAST_FRAME] =
          ref_picture_map[remapped_ref_idx[i - LAST_FRAME]];

    tpl_frame->ref_map_index[ALTREF_FRAME - LAST_FRAME] = -1;
    tpl_frame->ref_map_index[LAST3_FRAME - LAST_FRAME] = -1;
    tpl_frame->ref_map_index[BWDREF_FRAME - LAST_FRAME] = -1;
    tpl_frame->ref_map_index[ALTREF2_FRAME - LAST_FRAME] = -1;

    if (refresh_mask) ref_picture_map[refresh_frame_map_index] = gf_index;

    ++*tpl_group_frames;
    ++extend_frame_count;
    ++frame_display_index;
  }
}

void av1_init_tpl_stats(TplParams *const tpl_data) {
  tpl_data->ready = 0;
  set_tpl_stats_block_size(&tpl_data->tpl_stats_block_mis_log2,
                           &tpl_data->tpl_bsize_1d);
  for (int frame_idx = 0; frame_idx < MAX_LENGTH_TPL_FRAME_STATS; ++frame_idx) {
    TplDepFrame *tpl_frame = &tpl_data->tpl_stats_buffer[frame_idx];
    tpl_frame->is_valid = 0;
  }
  for (int frame_idx = 0; frame_idx < MAX_LAG_BUFFERS; ++frame_idx) {
    TplDepFrame *tpl_frame = &tpl_data->tpl_stats_buffer[frame_idx];
    if (tpl_data->tpl_stats_pool[frame_idx] == NULL) continue;
    memset(tpl_data->tpl_stats_pool[frame_idx], 0,
           tpl_frame->height * tpl_frame->width *
               sizeof(*tpl_frame->tpl_stats_ptr));
  }
}

int av1_tpl_stats_ready(const TplParams *tpl_data, int gf_frame_index) {
  if (tpl_data->ready == 0) {
    return 0;
  }
  if (gf_frame_index >= MAX_TPL_FRAME_IDX) {
    // The sub-GOP length exceeds the TPL buffer capacity.
    // Hence the TPL related functions are disabled hereafter.
    return 0;
  }
  return tpl_data->tpl_frame[gf_frame_index].is_valid;
}

static inline int eval_gop_length(double *beta, int gop_eval) {
  switch (gop_eval) {
    case 1:
      // Allow larger GOP size if the base layer ARF has higher dependency
      // factor than the intermediate ARF and both ARFs have reasonably high
      // dependency factors.
      return (beta[0] >= beta[1] + 0.7) && beta[0] > 3.0;
    case 2:
      if ((beta[0] >= beta[1] + 0.4) && beta[0] > 1.6)
        return 1;  // Don't shorten the gf interval
      else if ((beta[0] < beta[1] + 0.1) || beta[0] <= 1.4)
        return 0;  // Shorten the gf interval
      else
        return 2;  // Cannot decide the gf interval, so redo the
                   // tpl stats calculation.
    case 3: return beta[0] > 1.1;
    default: return 2;
  }
}

// TODO(jingning): Restructure av1_rc_pick_q_and_bounds() to narrow down
// the scope of input arguments.
void av1_tpl_preload_rc_estimate(AV1_COMP *cpi,
                                 const EncodeFrameParams *const frame_params) {
  AV1_COMMON *cm = &cpi->common;
  GF_GROUP *gf_group = &cpi->ppi->gf_group;
  int bottom_index, top_index;
  if (cpi->use_ducky_encode) return;

  cm->current_frame.frame_type = frame_params->frame_type;
  for (int gf_index = cpi->gf_frame_index; gf_index < gf_group->size;
       ++gf_index) {
    cm->current_frame.frame_type = gf_group->frame_type[gf_index];
    cm->show_frame = gf_group->update_type[gf_index] != ARF_UPDATE &&
                     gf_group->update_type[gf_index] != INTNL_ARF_UPDATE;
    gf_group->q_val[gf_index] = av1_rc_pick_q_and_bounds(
        cpi, cm->width, cm->height, gf_index, &bottom_index, &top_index);
  }
}

static inline int skip_tpl_for_frame(const GF_GROUP *gf_group, int frame_idx,
                                     int gop_eval, int approx_gop_eval,
                                     int reduce_num_frames) {
  // When gop_eval is set to 2, tpl stats calculation is done for ARFs from base
  // layer, (base+1) layer and (base+2) layer. When gop_eval is set to 3,
  // tpl stats calculation is limited to ARFs from base layer and (base+1)
  // layer.
  const int num_arf_layers = (gop_eval == 2) ? 3 : 2;
  const int gop_length = get_gop_length(gf_group);

  if (gf_group->update_type[frame_idx] == INTNL_OVERLAY_UPDATE ||
      gf_group->update_type[frame_idx] == OVERLAY_UPDATE)
    return 1;

  // When approx_gop_eval = 1, skip tpl stats calculation for higher layer
  // frames and for frames beyond gop length.
  if (approx_gop_eval && (gf_group->layer_depth[frame_idx] > num_arf_layers ||
                          frame_idx >= gop_length))
    return 1;

  if (reduce_num_frames && gf_group->update_type[frame_idx] == LF_UPDATE &&
      frame_idx < gop_length)
    return 1;

  return 0;
}

int av1_tpl_setup_stats(AV1_COMP *cpi, int gop_eval,
                        const EncodeFrameParams *const frame_params) {
#if CONFIG_COLLECT_COMPONENT_TIMING
  start_timing(cpi, av1_tpl_setup_stats_time);
#endif
  assert(cpi->gf_frame_index == 0);
  AV1_COMMON *cm = &cpi->common;
  MultiThreadInfo *const mt_info = &cpi->mt_info;
  AV1TplRowMultiThreadInfo *const tpl_row_mt = &mt_info->tpl_row_mt;
  GF_GROUP *gf_group = &cpi->ppi->gf_group;
  EncodeFrameParams this_frame_params = *frame_params;
  TplParams *const tpl_data = &cpi->ppi->tpl_data;
  int approx_gop_eval = (gop_eval > 1);

  if (cpi->superres_mode != AOM_SUPERRES_NONE) {
    assert(cpi->superres_mode != AOM_SUPERRES_AUTO);
    av1_init_tpl_stats(tpl_data);
    return 0;
  }

  cm->current_frame.frame_type = frame_params->frame_type;
  for (int gf_index = cpi->gf_frame_index; gf_index < gf_group->size;
       ++gf_index) {
    cm->current_frame.frame_type = gf_group->frame_type[gf_index];
    av1_configure_buffer_updates(cpi, &this_frame_params.refresh_frame,
                                 gf_group->update_type[gf_index],
                                 gf_group->refbuf_state[gf_index], 0);

    memcpy(&cpi->refresh_frame, &this_frame_params.refresh_frame,
           sizeof(cpi->refresh_frame));
  }

  int pframe_qindex;
  int tpl_gf_group_frames;
  init_gop_frames_for_tpl(cpi, frame_params, gf_group, &tpl_gf_group_frames,
                          &pframe_qindex);

  cpi->ppi->p_rc.base_layer_qp = pframe_qindex;

  av1_init_tpl_stats(tpl_data);

  TplBuffers *tpl_tmp_buffers = &cpi->td.tpl_tmp_buffers;
  if (!tpl_alloc_temp_buffers(tpl_tmp_buffers, tpl_data->tpl_bsize_1d)) {
    aom_internal_error(cpi->common.error, AOM_CODEC_MEM_ERROR,
                       "Error allocating tpl data");
  }

  tpl_row_mt->sync_read_ptr = av1_tpl_row_mt_sync_read_dummy;
  tpl_row_mt->sync_write_ptr = av1_tpl_row_mt_sync_write_dummy;

  av1_setup_scale_factors_for_frame(&cm->sf_identity, cm->width, cm->height,
                                    cm->width, cm->height);

  if (frame_params->frame_type == KEY_FRAME) {
    av1_init_mv_probs(cm);
  }
  av1_fill_mv_costs(&cm->fc->nmvc, cm->features.cur_frame_force_integer_mv,
                    cm->features.allow_high_precision_mv, cpi->td.mb.mv_costs);

  const int num_planes =
      cpi->sf.tpl_sf.use_y_only_rate_distortion ? 1 : av1_num_planes(cm);
  // As tpl module is called before the setting of speed features at frame
  // level, turning off this speed feature for the first GF group of the
  // key-frame interval is done here.
  int reduce_num_frames =
      cpi->sf.tpl_sf.reduce_num_frames &&
      gf_group->update_type[cpi->gf_frame_index] != KF_UPDATE &&
      gf_group->max_layer_depth > 2;
  // TPL processing is skipped for frames of type LF_UPDATE when
  // 'reduce_num_frames' is 1, which affects the r0 calcuation. Thus, a factor
  // to adjust r0 is used. The value of 1.6 corresponds to using ~60% of the
  // frames in the gf group on an average.
  tpl_data->r0_adjust_factor = reduce_num_frames ? 1.6 : 1.0;

  // Backward propagation from tpl_group_frames to 1.
  for (int frame_idx = cpi->gf_frame_index; frame_idx < tpl_gf_group_frames;
       ++frame_idx) {
    if (skip_tpl_for_frame(gf_group, frame_idx, gop_eval, approx_gop_eval,
                           reduce_num_frames))
      continue;

    init_mc_flow_dispenser(cpi, frame_idx, pframe_qindex);
    if (mt_info->num_workers > 1) {
      tpl_row_mt->sync_read_ptr = av1_tpl_row_mt_sync_read;
      tpl_row_mt->sync_write_ptr = av1_tpl_row_mt_sync_write;
      av1_mc_flow_dispenser_mt(cpi);
    } else {
      mc_flow_dispenser(cpi);
    }
#if CONFIG_BITRATE_ACCURACY
    av1_tpl_txfm_stats_update_abs_coeff_mean(&cpi->td.tpl_txfm_stats);
    av1_tpl_store_txfm_stats(tpl_data, &cpi->td.tpl_txfm_stats, frame_idx);
#endif  // CONFIG_BITRATE_ACCURACY
#if CONFIG_RATECTRL_LOG && CONFIG_THREE_PASS && CONFIG_BITRATE_ACCURACY
    if (cpi->oxcf.pass == AOM_RC_THIRD_PASS) {
      int frame_coding_idx =
          av1_vbr_rc_frame_coding_idx(&cpi->vbr_rc_info, frame_idx);
      rc_log_frame_stats(&cpi->rc_log, frame_coding_idx,
                         &cpi->td.tpl_txfm_stats);
    }
#endif  // CONFIG_RATECTRL_LOG

    aom_extend_frame_borders(tpl_data->tpl_frame[frame_idx].rec_picture,
                             num_planes);
  }

  for (int frame_idx = tpl_gf_group_frames - 1;
       frame_idx >= cpi->gf_frame_index; --frame_idx) {
    if (skip_tpl_for_frame(gf_group, frame_idx, gop_eval, approx_gop_eval,
                           reduce_num_frames))
      continue;

    mc_flow_synthesizer(tpl_data, frame_idx, cm->mi_params.mi_rows,
                        cm->mi_params.mi_cols);
  }

  av1_configure_buffer_updates(cpi, &this_frame_params.refresh_frame,
                               gf_group->update_type[cpi->gf_frame_index],
                               gf_group->update_type[cpi->gf_frame_index], 0);
  cm->current_frame.frame_type = frame_params->frame_type;
  cm->show_frame = frame_params->show_frame;

#if CONFIG_COLLECT_COMPONENT_TIMING
  // Record the time if the function returns.
  if (cpi->common.tiles.large_scale || gf_group->max_layer_depth_allowed == 0 ||
      !gop_eval)
    end_timing(cpi, av1_tpl_setup_stats_time);
#endif

  tpl_dealloc_temp_buffers(tpl_tmp_buffers);

  if (!approx_gop_eval) {
    tpl_data->ready = 1;
  }
  if (cpi->common.tiles.large_scale) return 0;
  if (gf_group->max_layer_depth_allowed == 0) return 1;
  if (!gop_eval) return 0;
  assert(gf_group->arf_index >= 0);

  double beta[2] = { 0.0 };
  const int frame_idx_0 = gf_group->arf_index;
  const int frame_idx_1 =
      AOMMIN(tpl_gf_group_frames - 1, gf_group->arf_index + 1);
  beta[0] = av1_tpl_get_frame_importance(tpl_data, frame_idx_0);
  beta[1] = av1_tpl_get_frame_importance(tpl_data, frame_idx_1);
#if CONFIG_COLLECT_COMPONENT_TIMING
  end_timing(cpi, av1_tpl_setup_stats_time);
#endif
  return eval_gop_length(beta, gop_eval);
}

void av1_tpl_rdmult_setup(AV1_COMP *cpi) {
  const AV1_COMMON *const cm = &cpi->common;
  const int tpl_idx = cpi->gf_frame_index;

  assert(
      IMPLIES(cpi->ppi->gf_group.size > 0, tpl_idx < cpi->ppi->gf_group.size));

  TplParams *const tpl_data = &cpi->ppi->tpl_data;
  const TplDepFrame *const tpl_frame = &tpl_data->tpl_frame[tpl_idx];

  if (!tpl_frame->is_valid) return;

  const TplDepStats *const tpl_stats = tpl_frame->tpl_stats_ptr;
  const int tpl_stride = tpl_frame->stride;
  const int mi_cols_sr = av1_pixels_to_mi(cm->superres_upscaled_width);

  const int block_size = BLOCK_16X16;
  const int num_mi_w = mi_size_wide[block_size];
  const int num_mi_h = mi_size_high[block_size];
  const int num_cols = (mi_cols_sr + num_mi_w - 1) / num_mi_w;
  const int num_rows = (cm->mi_params.mi_rows + num_mi_h - 1) / num_mi_h;
  const double c = 1.2;
  const int step = 1 << tpl_data->tpl_stats_block_mis_log2;

  // Loop through each 'block_size' X 'block_size' block.
  for (int row = 0; row < num_rows; row++) {
    for (int col = 0; col < num_cols; col++) {
      double intra_cost = 0.0, mc_dep_cost = 0.0;
      // Loop through each mi block.
      for (int mi_row = row * num_mi_h; mi_row < (row + 1) * num_mi_h;
           mi_row += step) {
        for (int mi_col = col * num_mi_w; mi_col < (col + 1) * num_mi_w;
             mi_col += step) {
          if (mi_row >= cm->mi_params.mi_rows || mi_col >= mi_cols_sr) continue;
          const TplDepStats *this_stats = &tpl_stats[av1_tpl_ptr_pos(
              mi_row, mi_col, tpl_stride, tpl_data->tpl_stats_block_mis_log2)];
          int64_t mc_dep_delta =
              RDCOST(tpl_frame->base_rdmult, this_stats->mc_dep_rate,
                     this_stats->mc_dep_dist);
          intra_cost += (double)(this_stats->recrf_dist << RDDIV_BITS);
          mc_dep_cost +=
              (double)(this_stats->recrf_dist << RDDIV_BITS) + mc_dep_delta;
        }
      }
      const double rk = intra_cost / mc_dep_cost;
      const int index = row * num_cols + col;
      cpi->tpl_rdmult_scaling_factors[index] = rk / cpi->rd.r0 + c;
    }
  }
}

void av1_tpl_rdmult_setup_sb(AV1_COMP *cpi, MACROBLOCK *const x,
                             BLOCK_SIZE sb_size, int mi_row, int mi_col) {
  AV1_COMMON *const cm = &cpi->common;
  GF_GROUP *gf_group = &cpi->ppi->gf_group;
  assert(IMPLIES(cpi->ppi->gf_group.size > 0,
                 cpi->gf_frame_index < cpi->ppi->gf_group.size));
  const int tpl_idx = cpi->gf_frame_index;

  const int boost_index = AOMMIN(15, (cpi->ppi->p_rc.gfu_boost / 100));
  const int layer_depth = AOMMIN(gf_group->layer_depth[cpi->gf_frame_index], 6);
  const FRAME_TYPE frame_type = cm->current_frame.frame_type;

  if (tpl_idx >= MAX_TPL_FRAME_IDX) return;
  TplDepFrame *tpl_frame = &cpi->ppi->tpl_data.tpl_frame[tpl_idx];
  if (!tpl_frame->is_valid) return;
  if (!is_frame_tpl_eligible(gf_group, cpi->gf_frame_index)) return;
  if (cpi->oxcf.q_cfg.aq_mode != NO_AQ) return;

  const int mi_col_sr =
      coded_to_superres_mi(mi_col, cm->superres_scale_denominator);
  const int mi_cols_sr = av1_pixels_to_mi(cm->superres_upscaled_width);
  const int sb_mi_width_sr = coded_to_superres_mi(
      mi_size_wide[sb_size], cm->superres_scale_denominator);

  const int bsize_base = BLOCK_16X16;
  const int num_mi_w = mi_size_wide[bsize_base];
  const int num_mi_h = mi_size_high[bsize_base];
  const int num_cols = (mi_cols_sr + num_mi_w - 1) / num_mi_w;
  const int num_rows = (cm->mi_params.mi_rows + num_mi_h - 1) / num_mi_h;
  const int num_bcols = (sb_mi_width_sr + num_mi_w - 1) / num_mi_w;
  const int num_brows = (mi_size_high[sb_size] + num_mi_h - 1) / num_mi_h;
  int row, col;

  double base_block_count = 0.0;
  double log_sum = 0.0;

  for (row = mi_row / num_mi_w;
       row < num_rows && row < mi_row / num_mi_w + num_brows; ++row) {
    for (col = mi_col_sr / num_mi_h;
         col < num_cols && col < mi_col_sr / num_mi_h + num_bcols; ++col) {
      const int index = row * num_cols + col;
      log_sum += log(cpi->tpl_rdmult_scaling_factors[index]);
      base_block_count += 1.0;
    }
  }

  const CommonQuantParams *quant_params = &cm->quant_params;

  const int orig_qindex_rdmult =
      quant_params->base_qindex + quant_params->y_dc_delta_q;
  const int orig_rdmult = av1_compute_rd_mult(
      orig_qindex_rdmult, cm->seq_params->bit_depth,
      cpi->ppi->gf_group.update_type[cpi->gf_frame_index], layer_depth,
      boost_index, frame_type, cpi->oxcf.q_cfg.use_fixed_qp_offsets,
      is_stat_consumption_stage(cpi));

  const int new_qindex_rdmult = quant_params->base_qindex +
                                x->rdmult_delta_qindex +
                                quant_params->y_dc_delta_q;
  const int new_rdmult = av1_compute_rd_mult(
      new_qindex_rdmult, cm->seq_params->bit_depth,
      cpi->ppi->gf_group.update_type[cpi->gf_frame_index], layer_depth,
      boost_index, frame_type, cpi->oxcf.q_cfg.use_fixed_qp_offsets,
      is_stat_consumption_stage(cpi));

  const double scaling_factor = (double)new_rdmult / (double)orig_rdmult;

  double scale_adj = log(scaling_factor) - log_sum / base_block_count;
  scale_adj = exp_bounded(scale_adj);

  for (row = mi_row / num_mi_w;
       row < num_rows && row < mi_row / num_mi_w + num_brows; ++row) {
    for (col = mi_col_sr / num_mi_h;
         col < num_cols && col < mi_col_sr / num_mi_h + num_bcols; ++col) {
      const int index = row * num_cols + col;
      cpi->ppi->tpl_sb_rdmult_scaling_factors[index] =
          scale_adj * cpi->tpl_rdmult_scaling_factors[index];
    }
  }
}

double av1_exponential_entropy(double q_step, double b) {
  b = AOMMAX(b, TPL_EPSILON);
  double z = fmax(exp_bounded(-q_step / b), TPL_EPSILON);
  return -log2(1 - z) - z * log2(z) / (1 - z);
}

double av1_laplace_entropy(double q_step, double b, double zero_bin_ratio) {
  // zero bin's size is zero_bin_ratio * q_step
  // non-zero bin's size is q_step
  b = AOMMAX(b, TPL_EPSILON);
  double z = fmax(exp_bounded(-zero_bin_ratio / 2 * q_step / b), TPL_EPSILON);
  double h = av1_exponential_entropy(q_step, b);
  double r = -(1 - z) * log2(1 - z) - z * log2(z) + z * (h + 1);
  return r;
}

double av1_laplace_estimate_frame_rate(int q_index, int block_count,
                                       const double *abs_coeff_mean,
                                       int coeff_num) {
  double zero_bin_ratio = 2;
  double dc_q_step = av1_dc_quant_QTX(q_index, 0, AOM_BITS_8) / 4.;
  double ac_q_step = av1_ac_quant_QTX(q_index, 0, AOM_BITS_8) / 4.;
  double est_rate = 0;
  // dc coeff
  est_rate += av1_laplace_entropy(dc_q_step, abs_coeff_mean[0], zero_bin_ratio);
  // ac coeff
  for (int i = 1; i < coeff_num; ++i) {
    est_rate +=
        av1_laplace_entropy(ac_q_step, abs_coeff_mean[i], zero_bin_ratio);
  }
  est_rate *= block_count;
  return est_rate;
}

double av1_estimate_coeff_entropy(double q_step, double b,
                                  double zero_bin_ratio, int qcoeff) {
  b = AOMMAX(b, TPL_EPSILON);
  int abs_qcoeff = abs(qcoeff);
  double z0 = fmax(exp_bounded(-zero_bin_ratio / 2 * q_step / b), TPL_EPSILON);
  if (abs_qcoeff == 0) {
    double r = -log2(1 - z0);
    return r;
  } else {
    double z = fmax(exp_bounded(-q_step / b), TPL_EPSILON);
    double r = 1 - log2(z0) - log2(1 - z) - (abs_qcoeff - 1) * log2(z);
    return r;
  }
}

#if CONFIG_RD_COMMAND
void av1_read_rd_command(const char *filepath, RD_COMMAND *rd_command) {
  FILE *fptr = fopen(filepath, "r");
  fscanf(fptr, "%d", &rd_command->frame_count);
  rd_command->frame_index = 0;
  for (int i = 0; i < rd_command->frame_count; ++i) {
    int option;
    fscanf(fptr, "%d", &option);
    rd_command->option_ls[i] = (RD_OPTION)option;
    if (option == RD_OPTION_SET_Q) {
      fscanf(fptr, "%d", &rd_command->q_index_ls[i]);
    } else if (option == RD_OPTION_SET_Q_RDMULT) {
      fscanf(fptr, "%d", &rd_command->q_index_ls[i]);
      fscanf(fptr, "%d", &rd_command->rdmult_ls[i]);
    }
  }
  fclose(fptr);
}
#endif  // CONFIG_RD_COMMAND

double av1_tpl_get_frame_importance(const TplParams *tpl_data,
                                    int gf_frame_index) {
  const TplDepFrame *tpl_frame = &tpl_data->tpl_frame[gf_frame_index];
  const TplDepStats *tpl_stats = tpl_frame->tpl_stats_ptr;

  const int tpl_stride = tpl_frame->stride;
  double intra_cost_base = 0;
  double mc_dep_cost_base = 0;
  double cbcmp_base = 1;
  const int step = 1 << tpl_data->tpl_stats_block_mis_log2;

  for (int row = 0; row < tpl_frame->mi_rows; row += step) {
    for (int col = 0; col < tpl_frame->mi_cols; col += step) {
      const TplDepStats *this_stats = &tpl_stats[av1_tpl_ptr_pos(
          row, col, tpl_stride, tpl_data->tpl_stats_block_mis_log2)];
      double cbcmp = (double)this_stats->srcrf_dist;
      const int64_t mc_dep_delta =
          RDCOST(tpl_frame->base_rdmult, this_stats->mc_dep_rate,
                 this_stats->mc_dep_dist);
      double dist_scaled = (double)(this_stats->recrf_dist << RDDIV_BITS);
      dist_scaled = AOMMAX(dist_scaled, 1);
      intra_cost_base += log(dist_scaled) * cbcmp;
      mc_dep_cost_base += log(dist_scaled + mc_dep_delta) * cbcmp;
      cbcmp_base += cbcmp;
    }
  }
  return exp((mc_dep_cost_base - intra_cost_base) / cbcmp_base);
}

double av1_tpl_get_qstep_ratio(const TplParams *tpl_data, int gf_frame_index) {
  if (!av1_tpl_stats_ready(tpl_data, gf_frame_index)) {
    return 1;
  }
  const double frame_importance =
      av1_tpl_get_frame_importance(tpl_data, gf_frame_index);
  return sqrt(1 / frame_importance);
}

int av1_get_q_index_from_qstep_ratio(int leaf_qindex, double qstep_ratio,
                                     aom_bit_depth_t bit_depth) {
  const double leaf_qstep = av1_dc_quant_QTX(leaf_qindex, 0, bit_depth);
  const double target_qstep = leaf_qstep * qstep_ratio;
  int qindex = leaf_qindex;
  if (qstep_ratio < 1.0) {
    for (qindex = leaf_qindex; qindex > 0; --qindex) {
      const double qstep = av1_dc_quant_QTX(qindex, 0, bit_depth);
      if (qstep <= target_qstep) break;
    }
  } else {
    for (qindex = leaf_qindex; qindex <= MAXQ; ++qindex) {
      const double qstep = av1_dc_quant_QTX(qindex, 0, bit_depth);
      if (qstep >= target_qstep) break;
    }
  }
  return qindex;
}

int av1_tpl_get_q_index(const TplParams *tpl_data, int gf_frame_index,
                        int leaf_qindex, aom_bit_depth_t bit_depth) {
  const double qstep_ratio = av1_tpl_get_qstep_ratio(tpl_data, gf_frame_index);
  return av1_get_q_index_from_qstep_ratio(leaf_qindex, qstep_ratio, bit_depth);
}

#if CONFIG_BITRATE_ACCURACY
void av1_vbr_rc_init(VBR_RATECTRL_INFO *vbr_rc_info, double total_bit_budget,
                     int show_frame_count) {
  av1_zero(*vbr_rc_info);
  vbr_rc_info->ready = 0;
  vbr_rc_info->total_bit_budget = total_bit_budget;
  vbr_rc_info->show_frame_count = show_frame_count;
  const double scale_factors[FRAME_UPDATE_TYPES] = { 0.94559, 0.94559, 1,
                                                     0.94559, 1,       1,
                                                     0.94559 };

  // TODO(angiebird): Based on the previous code, only the scale factor 0.94559
  // will be used in most of the cases with --limi=17. Figure out if the
  // following scale factors works better.
  // const double scale_factors[FRAME_UPDATE_TYPES] = { 0.94559, 0.12040, 1,
  //                                                    1.10199, 1,       1,
  //                                                    0.16393 };

  const double mv_scale_factors[FRAME_UPDATE_TYPES] = { 3, 3, 3, 3, 3, 3, 3 };
  memcpy(vbr_rc_info->scale_factors, scale_factors,
         sizeof(scale_factors[0]) * FRAME_UPDATE_TYPES);
  memcpy(vbr_rc_info->mv_scale_factors, mv_scale_factors,
         sizeof(mv_scale_factors[0]) * FRAME_UPDATE_TYPES);

  vbr_rc_reset_gop_data(vbr_rc_info);
#if CONFIG_THREE_PASS
  // TODO(angiebird): Explain why we use -1 here
  vbr_rc_info->cur_gop_idx = -1;
  vbr_rc_info->gop_count = 0;
  vbr_rc_info->total_frame_count = 0;
#endif  // CONFIG_THREE_PASS
}

#if CONFIG_THREE_PASS
int av1_vbr_rc_frame_coding_idx(const VBR_RATECTRL_INFO *vbr_rc_info,
                                int gf_frame_index) {
  int gop_idx = vbr_rc_info->cur_gop_idx;
  int gop_start_idx = vbr_rc_info->gop_start_idx_list[gop_idx];
  return gop_start_idx + gf_frame_index;
}

void av1_vbr_rc_append_tpl_info(VBR_RATECTRL_INFO *vbr_rc_info,
                                const TPL_INFO *tpl_info) {
  int gop_start_idx = vbr_rc_info->total_frame_count;
  vbr_rc_info->gop_start_idx_list[vbr_rc_info->gop_count] = gop_start_idx;
  vbr_rc_info->gop_length_list[vbr_rc_info->gop_count] = tpl_info->gf_length;
  assert(gop_start_idx + tpl_info->gf_length <= VBR_RC_INFO_MAX_FRAMES);
  for (int i = 0; i < tpl_info->gf_length; ++i) {
    vbr_rc_info->txfm_stats_list[gop_start_idx + i] =
        tpl_info->txfm_stats_list[i];
    vbr_rc_info->qstep_ratio_list[gop_start_idx + i] =
        tpl_info->qstep_ratio_ls[i];
    vbr_rc_info->update_type_list[gop_start_idx + i] =
        tpl_info->update_type_list[i];
  }
  vbr_rc_info->total_frame_count += tpl_info->gf_length;
  vbr_rc_info->gop_count++;
}
#endif  // CONFIG_THREE_PASS

void av1_vbr_rc_set_gop_bit_budget(VBR_RATECTRL_INFO *vbr_rc_info,
                                   int gop_showframe_count) {
  vbr_rc_info->gop_showframe_count = gop_showframe_count;
  vbr_rc_info->gop_bit_budget = vbr_rc_info->total_bit_budget *
                                gop_showframe_count /
                                vbr_rc_info->show_frame_count;
}

void av1_vbr_rc_compute_q_indices(int base_q_index, int frame_count,
                                  const double *qstep_ratio_list,
                                  aom_bit_depth_t bit_depth,
                                  int *q_index_list) {
  for (int i = 0; i < frame_count; ++i) {
    q_index_list[i] = av1_get_q_index_from_qstep_ratio(
        base_q_index, qstep_ratio_list[i], bit_depth);
  }
}

double av1_vbr_rc_info_estimate_gop_bitrate(
    int base_q_index, aom_bit_depth_t bit_depth,
    const double *update_type_scale_factors, int frame_count,
    const FRAME_UPDATE_TYPE *update_type_list, const double *qstep_ratio_list,
    const TplTxfmStats *stats_list, int *q_index_list,
    double *estimated_bitrate_byframe) {
  av1_vbr_rc_compute_q_indices(base_q_index, frame_count, qstep_ratio_list,
                               bit_depth, q_index_list);
  double estimated_gop_bitrate = 0;
  for (int frame_index = 0; frame_index < frame_count; frame_index++) {
    const TplTxfmStats *frame_stats = &stats_list[frame_index];
    double frame_bitrate = 0;
    if (frame_stats->ready) {
      int q_index = q_index_list[frame_index];

      frame_bitrate = av1_laplace_estimate_frame_rate(
          q_index, frame_stats->txfm_block_count, frame_stats->abs_coeff_mean,
          frame_stats->coeff_num);
    }
    FRAME_UPDATE_TYPE update_type = update_type_list[frame_index];
    estimated_gop_bitrate +=
        frame_bitrate * update_type_scale_factors[update_type];
    if (estimated_bitrate_byframe != NULL) {
      estimated_bitrate_byframe[frame_index] = frame_bitrate;
    }
  }
  return estimated_gop_bitrate;
}

int av1_vbr_rc_info_estimate_base_q(
    double bit_budget, aom_bit_depth_t bit_depth,
    const double *update_type_scale_factors, int frame_count,
    const FRAME_UPDATE_TYPE *update_type_list, const double *qstep_ratio_list,
    const TplTxfmStats *stats_list, int *q_index_list,
    double *estimated_bitrate_byframe) {
  int q_max = 255;  // Maximum q value.
  int q_min = 0;    // Minimum q value.
  int q = (q_max + q_min) / 2;

  double q_max_estimate = av1_vbr_rc_info_estimate_gop_bitrate(
      q_max, bit_depth, update_type_scale_factors, frame_count,
      update_type_list, qstep_ratio_list, stats_list, q_index_list,
      estimated_bitrate_byframe);

  double q_min_estimate = av1_vbr_rc_info_estimate_gop_bitrate(
      q_min, bit_depth, update_type_scale_factors, frame_count,
      update_type_list, qstep_ratio_list, stats_list, q_index_list,
      estimated_bitrate_byframe);
  while (q_min + 1 < q_max) {
    double estimate = av1_vbr_rc_info_estimate_gop_bitrate(
        q, bit_depth, update_type_scale_factors, frame_count, update_type_list,
        qstep_ratio_list, stats_list, q_index_list, estimated_bitrate_byframe);
    if (estimate > bit_budget) {
      q_min = q;
      q_min_estimate = estimate;
    } else {
      q_max = q;
      q_max_estimate = estimate;
    }
    q = (q_max + q_min) / 2;
  }
  // Pick the estimate that lands closest to the budget.
  if (fabs(q_max_estimate - bit_budget) < fabs(q_min_estimate - bit_budget)) {
    q = q_max;
  } else {
    q = q_min;
  }
  // Update q_index_list and vbr_rc_info.
  av1_vbr_rc_info_estimate_gop_bitrate(
      q, bit_depth, update_type_scale_factors, frame_count, update_type_list,
      qstep_ratio_list, stats_list, q_index_list, estimated_bitrate_byframe);
  return q;
}
void av1_vbr_rc_update_q_index_list(VBR_RATECTRL_INFO *vbr_rc_info,
                                    const TplParams *tpl_data,
                                    const GF_GROUP *gf_group,
                                    aom_bit_depth_t bit_depth) {
  vbr_rc_info->q_index_list_ready = 1;
  double gop_bit_budget = vbr_rc_info->gop_bit_budget;

  for (int i = 0; i < gf_group->size; i++) {
    vbr_rc_info->qstep_ratio_list[i] = av1_tpl_get_qstep_ratio(tpl_data, i);
  }

  double mv_bits = 0;
  for (int i = 0; i < gf_group->size; i++) {
    double frame_mv_bits = 0;
    if (av1_tpl_stats_ready(tpl_data, i)) {
      TplDepFrame *tpl_frame = &tpl_data->tpl_frame[i];
      frame_mv_bits = av1_tpl_compute_frame_mv_entropy(
          tpl_frame, tpl_data->tpl_stats_block_mis_log2);
      FRAME_UPDATE_TYPE updae_type = gf_group->update_type[i];
      mv_bits += frame_mv_bits * vbr_rc_info->mv_scale_factors[updae_type];
    }
  }

  mv_bits = AOMMIN(mv_bits, 0.6 * gop_bit_budget);
  gop_bit_budget -= mv_bits;

  vbr_rc_info->base_q_index = av1_vbr_rc_info_estimate_base_q(
      gop_bit_budget, bit_depth, vbr_rc_info->scale_factors, gf_group->size,
      gf_group->update_type, vbr_rc_info->qstep_ratio_list,
      tpl_data->txfm_stats_list, vbr_rc_info->q_index_list, NULL);
}

#endif  // CONFIG_BITRATE_ACCURACY

// Use upper and left neighbor block as the reference MVs.
// Compute the minimum difference between current MV and reference MV.
int_mv av1_compute_mv_difference(const TplDepFrame *tpl_frame, int row, int col,
                                 int step, int tpl_stride, int right_shift) {
  const TplDepStats *tpl_stats =
      &tpl_frame
           ->tpl_stats_ptr[av1_tpl_ptr_pos(row, col, tpl_stride, right_shift)];
  int_mv current_mv = tpl_stats->mv[tpl_stats->ref_frame_index[0]];
  int current_mv_magnitude =
      abs(current_mv.as_mv.row) + abs(current_mv.as_mv.col);

  // Retrieve the up and left neighbors.
  int up_error = INT_MAX;
  int_mv up_mv_diff;
  if (row - step >= 0) {
    tpl_stats = &tpl_frame->tpl_stats_ptr[av1_tpl_ptr_pos(
        row - step, col, tpl_stride, right_shift)];
    up_mv_diff = tpl_stats->mv[tpl_stats->ref_frame_index[0]];
    up_mv_diff.as_mv.row = current_mv.as_mv.row - up_mv_diff.as_mv.row;
    up_mv_diff.as_mv.col = current_mv.as_mv.col - up_mv_diff.as_mv.col;
    up_error = abs(up_mv_diff.as_mv.row) + abs(up_mv_diff.as_mv.col);
  }

  int left_error = INT_MAX;
  int_mv left_mv_diff;
  if (col - step >= 0) {
    tpl_stats = &tpl_frame->tpl_stats_ptr[av1_tpl_ptr_pos(
        row, col - step, tpl_stride, right_shift)];
    left_mv_diff = tpl_stats->mv[tpl_stats->ref_frame_index[0]];
    left_mv_diff.as_mv.row = current_mv.as_mv.row - left_mv_diff.as_mv.row;
    left_mv_diff.as_mv.col = current_mv.as_mv.col - left_mv_diff.as_mv.col;
    left_error = abs(left_mv_diff.as_mv.row) + abs(left_mv_diff.as_mv.col);
  }

  // Return the MV with the minimum distance from current.
  if (up_error < left_error && up_error < current_mv_magnitude) {
    return up_mv_diff;
  } else if (left_error < up_error && left_error < current_mv_magnitude) {
    return left_mv_diff;
  }
  return current_mv;
}

/* Compute the entropy of motion vectors for a single frame. */
double av1_tpl_compute_frame_mv_entropy(const TplDepFrame *tpl_frame,
                                        uint8_t right_shift) {
  if (!tpl_frame->is_valid) {
    return 0;
  }

  int count_row[500] = { 0 };
  int count_col[500] = { 0 };
  int n = 0;  // number of MVs to process

  const int tpl_stride = tpl_frame->stride;
  const int step = 1 << right_shift;

  for (int row = 0; row < tpl_frame->mi_rows; row += step) {
    for (int col = 0; col < tpl_frame->mi_cols; col += step) {
      int_mv mv = av1_compute_mv_difference(tpl_frame, row, col, step,
                                            tpl_stride, right_shift);
      count_row[clamp(mv.as_mv.row, 0, 499)] += 1;
      count_col[clamp(mv.as_mv.row, 0, 499)] += 1;
      n += 1;
    }
  }

  // Estimate the bits used using the entropy formula.
  double rate_row = 0;
  double rate_col = 0;
  for (int i = 0; i < 500; i++) {
    if (count_row[i] != 0) {
      double p = count_row[i] / (double)n;
      rate_row += count_row[i] * -log2(p);
    }
    if (count_col[i] != 0) {
      double p = count_col[i] / (double)n;
      rate_col += count_col[i] * -log2(p);
    }
  }

  return rate_row + rate_col;
}