/*

 * Copyright (c) 2016, 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 <limits.h>

#include <math.h>

#include <stdio.h>

#include "config/aom_dsp_rtcd.h"

#include "config/aom_scale_rtcd.h"

#include "aom_dsp/aom_dsp_common.h"

#include "aom_dsp/variance.h"

#include "aom_mem/aom_mem.h"

#include "aom_ports/mem.h"

#include "aom_scale/yv12config.h"

#include "aom_util/aom_pthread.h"

#include "av1/common/entropymv.h"

#include "av1/common/quant_common.h"

#include "av1/common/reconinter.h"  // av1_setup_dst_planes()

#include "av1/common/reconintra.h"

#include "av1/common/txb_common.h"

#include "av1/encoder/aq_variance.h"

#include "av1/encoder/av1_quantize.h"

#include "av1/encoder/block.h"

#include "av1/encoder/dwt.h"

#include "av1/encoder/encodeframe.h"

#include "av1/encoder/encodeframe_utils.h"

#include "av1/encoder/encodemb.h"

#include "av1/encoder/encodemv.h"

#include "av1/encoder/encoder.h"

#include "av1/encoder/encoder_utils.h"

#include "av1/encoder/encode_strategy.h"

#include "av1/encoder/ethread.h"

#include "av1/encoder/extend.h"

#include "av1/encoder/firstpass.h"

#include "av1/encoder/mcomp.h"

#include "av1/encoder/rd.h"

#include "av1/encoder/reconinter_enc.h"

#define OUTPUT_FPF 0

#define FIRST_PASS_Q 10.0

#define INTRA_MODE_PENALTY 1024

#define NEW_MV_MODE_PENALTY 32

#define DARK_THRESH 64

#define NCOUNT_INTRA_THRESH 8192

#define NCOUNT_INTRA_FACTOR 3

#define INVALID_FP_STATS_TO_PREDICT_FLAT_GOP -1

static inline void output_stats(FIRSTPASS_STATS *stats,

                                struct aom_codec_pkt_list *pktlist) {

  struct aom_codec_cx_pkt pkt;

  pkt.kind = AOM_CODEC_STATS_PKT;

  pkt.data.twopass_stats.buf = stats;

  pkt.data.twopass_stats.sz = sizeof(FIRSTPASS_STATS);

  if (pktlist != NULL) aom_codec_pkt_list_add(pktlist, &pkt);

// TEMP debug code

#if OUTPUT_FPF

  {

    FILE *fpfile;

    fpfile = fopen("firstpass.stt", "a");

    fprintf(fpfile,

            "%12.0lf %12.4lf %12.0lf %12.0lf %12.0lf %12.4lf %12.4lf"

            "%12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.4lf %12.4lf"

            "%12.4lf %12.4lf %12.0lf %12.0lf %12.0lf %12.4lf %12.4lf\n",

            stats->frame, stats->weight, stats->intra_error, stats->coded_error,

            stats->sr_coded_error, stats->pcnt_inter, stats->pcnt_motion,

            stats->pcnt_second_ref, stats->pcnt_neutral, stats->intra_skip_pct,

            stats->inactive_zone_rows, stats->inactive_zone_cols, stats->MVr,

            stats->mvr_abs, stats->MVc, stats->mvc_abs, stats->MVrv,

            stats->MVcv, stats->mv_in_out_count, stats->new_mv_count,

            stats->count, stats->duration);

    fclose(fpfile);

  }

#endif

}

void av1_twopass_zero_stats(FIRSTPASS_STATS *section) {

  section->frame = 0.0;

  section->weight = 0.0;

  section->intra_error = 0.0;

  section->frame_avg_wavelet_energy = 0.0;

  section->coded_error = 0.0;

  section->log_intra_error = 0.0;

  section->log_coded_error = 0.0;

  section->sr_coded_error = 0.0;

  section->lt_coded_error = 0.0;

  section->pcnt_inter = 0.0;

  section->pcnt_motion = 0.0;

  section->pcnt_second_ref = 0.0;

  section->pcnt_neutral = 0.0;

  section->intra_skip_pct = 0.0;

  section->inactive_zone_rows = 0.0;

  section->inactive_zone_cols = 0.0;

  section->MVr = 0.0;

  section->mvr_abs = 0.0;

  section->MVc = 0.0;

  section->mvc_abs = 0.0;

  section->MVrv = 0.0;

  section->MVcv = 0.0;

  section->mv_in_out_count = 0.0;

  section->new_mv_count = 0.0;

  section->count = 0.0;

  section->duration = 1.0;

  section->is_flash = 0;

  section->noise_var = 0;

  section->cor_coeff = 1.0;

}

void av1_accumulate_stats(FIRSTPASS_STATS *section,

                          const FIRSTPASS_STATS *frame) {

  section->frame += frame->frame;

  section->weight += frame->weight;

  section->intra_error += frame->intra_error;

  section->log_intra_error += log1p(frame->intra_error);

  section->log_coded_error += log1p(frame->coded_error);

  section->frame_avg_wavelet_energy += frame->frame_avg_wavelet_energy;

  section->coded_error += frame->coded_error;

  section->sr_coded_error += frame->sr_coded_error;

  section->lt_coded_error += frame->lt_coded_error;

  section->pcnt_inter += frame->pcnt_inter;

  section->pcnt_motion += frame->pcnt_motion;

  section->pcnt_second_ref += frame->pcnt_second_ref;

  section->pcnt_neutral += frame->pcnt_neutral;

  section->intra_skip_pct += frame->intra_skip_pct;

  section->inactive_zone_rows += frame->inactive_zone_rows;

  section->inactive_zone_cols += frame->inactive_zone_cols;

  section->MVr += frame->MVr;

  section->mvr_abs += frame->mvr_abs;

  section->MVc += frame->MVc;

  section->mvc_abs += frame->mvc_abs;

  section->MVrv += frame->MVrv;

  section->MVcv += frame->MVcv;

  section->mv_in_out_count += frame->mv_in_out_count;

  section->new_mv_count += frame->new_mv_count;

  section->count += frame->count;

  section->duration += frame->duration;

}

static int get_unit_rows(const BLOCK_SIZE fp_block_size, const int mb_rows) {

  const int height_mi_log2 = mi_size_high_log2[fp_block_size];

  const int mb_height_mi_log2 = mi_size_high_log2[BLOCK_16X16];

  if (height_mi_log2 > mb_height_mi_log2) {

    return mb_rows >> (height_mi_log2 - mb_height_mi_log2);

  }

  return mb_rows << (mb_height_mi_log2 - height_mi_log2);

}

static int get_unit_cols(const BLOCK_SIZE fp_block_size, const int mb_cols) {

  const int width_mi_log2 = mi_size_wide_log2[fp_block_size];

  const int mb_width_mi_log2 = mi_size_wide_log2[BLOCK_16X16];

  if (width_mi_log2 > mb_width_mi_log2) {

    return mb_cols >> (width_mi_log2 - mb_width_mi_log2);

  }

  return mb_cols << (mb_width_mi_log2 - width_mi_log2);

}

// TODO(chengchen): can we simplify it even if resize has to be considered?

static int get_num_mbs(const BLOCK_SIZE fp_block_size,

                       const int num_mbs_16X16) {

  const int width_mi_log2 = mi_size_wide_log2[fp_block_size];

  const int height_mi_log2 = mi_size_high_log2[fp_block_size];

  const int mb_width_mi_log2 = mi_size_wide_log2[BLOCK_16X16];

  const int mb_height_mi_log2 = mi_size_high_log2[BLOCK_16X16];

  // TODO(chengchen): Now this function assumes a square block is used.

  // It does not support rectangular block sizes.

  assert(width_mi_log2 == height_mi_log2);

  if (width_mi_log2 > mb_width_mi_log2) {

    return num_mbs_16X16 >> ((width_mi_log2 - mb_width_mi_log2) +

                             (height_mi_log2 - mb_height_mi_log2));

  }

  return num_mbs_16X16 << ((mb_width_mi_log2 - width_mi_log2) +

                           (mb_height_mi_log2 - height_mi_log2));

}

void av1_end_first_pass(AV1_COMP *cpi) {

  if (cpi->ppi->twopass.stats_buf_ctx->total_stats && !cpi->ppi->lap_enabled)

    output_stats(cpi->ppi->twopass.stats_buf_ctx->total_stats,

                 cpi->ppi->output_pkt_list);

}

static aom_variance_fn_t get_block_variance_fn(BLOCK_SIZE bsize) {

  switch (bsize) {

    case BLOCK_8X8: return aom_mse8x8;

    case BLOCK_16X8: return aom_mse16x8;

    case BLOCK_8X16: return aom_mse8x16;

    default: return aom_mse16x16;

  }

}

static unsigned int get_prediction_error(BLOCK_SIZE bsize,

                                         const struct buf_2d *src,

                                         const struct buf_2d *ref) {

  unsigned int sse;

  const aom_variance_fn_t fn = get_block_variance_fn(bsize);

  fn(src->buf, src->stride, ref->buf, ref->stride, &sse);

  return sse;

}

#if CONFIG_AV1_HIGHBITDEPTH

static aom_variance_fn_t highbd_get_block_variance_fn(BLOCK_SIZE bsize,

                                                      int bd) {

  switch (bd) {

    default:

      switch (bsize) {

        case BLOCK_8X8: return aom_highbd_8_mse8x8;

        case BLOCK_16X8: return aom_highbd_8_mse16x8;

        case BLOCK_8X16: return aom_highbd_8_mse8x16;

        default: return aom_highbd_8_mse16x16;

      }

    case 10:

      switch (bsize) {

        case BLOCK_8X8: return aom_highbd_10_mse8x8;

        case BLOCK_16X8: return aom_highbd_10_mse16x8;

        case BLOCK_8X16: return aom_highbd_10_mse8x16;

        default: return aom_highbd_10_mse16x16;

      }

    case 12:

      switch (bsize) {

        case BLOCK_8X8: return aom_highbd_12_mse8x8;

        case BLOCK_16X8: return aom_highbd_12_mse16x8;

        case BLOCK_8X16: return aom_highbd_12_mse8x16;

        default: return aom_highbd_12_mse16x16;

      }

  }

}

static unsigned int highbd_get_prediction_error(BLOCK_SIZE bsize,

                                                const struct buf_2d *src,

                                                const struct buf_2d *ref,

                                                int bd) {

  unsigned int sse;

  const aom_variance_fn_t fn = highbd_get_block_variance_fn(bsize, bd);

  fn(src->buf, src->stride, ref->buf, ref->stride, &sse);

  return sse;

}

#endif  // CONFIG_AV1_HIGHBITDEPTH

// Refine the motion search range according to the frame dimension

// for first pass test.

static int get_search_range(int width, int height) {

  int sr = 0;

  int dim = AOMMIN(width, height);

  dim = AOMMAX(dim, MI_SIZE);

  while ((dim << sr) < MAX_FULL_PEL_VAL) ++sr;

  return sr;

}

static inline const search_site_config *av1_get_first_pass_search_site_config(

    const AV1_COMP *cpi, MACROBLOCK *x, SEARCH_METHODS search_method) {

  const int ref_stride = x->e_mbd.plane[0].pre[0].stride;

  // For AVIF applications, even the source frames can have changing resolution,

  // so we need to manually check for the strides :(

  // AV1_COMP::mv_search_params.search_site_config is a compressor level cache

  // that's shared by multiple threads. In most cases where all frames have the

  // same resolution, the cache contains the search site config that we need.

  const MotionVectorSearchParams *mv_search_params = &cpi->mv_search_params;

  if (ref_stride == mv_search_params->search_site_cfg[SS_CFG_FPF]->stride) {

    return mv_search_params->search_site_cfg[SS_CFG_FPF];

  }

  // If the cache does not contain the correct stride, then we will need to rely

  // on the thread level config MACROBLOCK::search_site_cfg_buf. If even the

  // thread level config doesn't match, then we need to update it.

  search_method = search_method_lookup[search_method];

  assert(search_method_lookup[search_method] == search_method &&

         "The search_method_lookup table should be idempotent.");

  if (ref_stride != x->search_site_cfg_buf[search_method].stride) {

    av1_refresh_search_site_config(x->search_site_cfg_buf, search_method,

                                   ref_stride);

  }

  return x->search_site_cfg_buf;

}

static inline void first_pass_motion_search(AV1_COMP *cpi, MACROBLOCK *x,

                                            const MV *ref_mv,

                                            FULLPEL_MV *best_mv,

                                            int *best_motion_err) {

  AV1_COMMON *const cm = &cpi->common;

  MACROBLOCKD *const xd = &x->e_mbd;

  FULLPEL_MV start_mv = get_fullmv_from_mv(ref_mv);

  int tmp_err;

  const BLOCK_SIZE bsize = xd->mi[0]->bsize;

  const int new_mv_mode_penalty = NEW_MV_MODE_PENALTY;

  const int sr = get_search_range(cm->width, cm->height);

  const int step_param = cpi->sf.fp_sf.reduce_mv_step_param + sr;

  const search_site_config *first_pass_search_sites =

      av1_get_first_pass_search_site_config(cpi, x, NSTEP);

  const int fine_search_interval =

      cpi->is_screen_content_type && cm->features.allow_intrabc;

  FULLPEL_MOTION_SEARCH_PARAMS ms_params;

  av1_make_default_fullpel_ms_params(&ms_params, cpi, x, bsize, ref_mv,

                                     start_mv, first_pass_search_sites, NSTEP,

                                     fine_search_interval);

  FULLPEL_MV this_best_mv;

  FULLPEL_MV_STATS best_mv_stats;

  tmp_err = av1_full_pixel_search(start_mv, &ms_params, step_param, NULL,

                                  &this_best_mv, &best_mv_stats, NULL);

  if (tmp_err < INT_MAX) {

    aom_variance_fn_ptr_t v_fn_ptr = cpi->ppi->fn_ptr[bsize];

    const MSBuffers *ms_buffers = &ms_params.ms_buffers;

    tmp_err = av1_get_mvpred_sse(&ms_params.mv_cost_params, this_best_mv,

                                 &v_fn_ptr, ms_buffers->src, ms_buffers->ref) +

              new_mv_mode_penalty;

  }

  if (tmp_err < *best_motion_err) {

    *best_motion_err = tmp_err;

    *best_mv = this_best_mv;

  }

}

static BLOCK_SIZE get_bsize(const CommonModeInfoParams *const mi_params,

                            const BLOCK_SIZE fp_block_size, const int unit_row,

                            const int unit_col) {

  const int unit_width = mi_size_wide[fp_block_size];

  const int unit_height = mi_size_high[fp_block_size];

  const int is_half_width =

      unit_width * unit_col + unit_width / 2 >= mi_params->mi_cols;

  const int is_half_height =

      unit_height * unit_row + unit_height / 2 >= mi_params->mi_rows;

  const int max_dimension =

      AOMMAX(block_size_wide[fp_block_size], block_size_high[fp_block_size]);

  int square_block_size = 0;

  // 4X4, 8X8, 16X16, 32X32, 64X64, 128X128

  switch (max_dimension) {

    case 4: square_block_size = 0; break;

    case 8: square_block_size = 1; break;

    case 16: square_block_size = 2; break;

    case 32: square_block_size = 3; break;

    case 64: square_block_size = 4; break;

    case 128: square_block_size = 5; break;

    default: assert(0 && "First pass block size is not supported!"); break;

  }

  if (is_half_width && is_half_height) {

    return subsize_lookup[PARTITION_SPLIT][square_block_size];

  } else if (is_half_width) {

    return subsize_lookup[PARTITION_VERT][square_block_size];

  } else if (is_half_height) {

    return subsize_lookup[PARTITION_HORZ][square_block_size];

  } else {

    return fp_block_size;

  }

}

static int find_fp_qindex(aom_bit_depth_t bit_depth) {

  return av1_find_qindex(FIRST_PASS_Q, bit_depth, 0, QINDEX_RANGE - 1);

}

static double raw_motion_error_stdev(int *raw_motion_err_list,

                                     int raw_motion_err_counts) {

  int64_t sum_raw_err = 0;

  double raw_err_avg = 0;

  double raw_err_stdev = 0;

  if (raw_motion_err_counts == 0) return 0;

  int i;

  for (i = 0; i < raw_motion_err_counts; i++) {

    sum_raw_err += raw_motion_err_list[i];

  }

  raw_err_avg = (double)sum_raw_err / raw_motion_err_counts;

  for (i = 0; i < raw_motion_err_counts; i++) {

    raw_err_stdev += (raw_motion_err_list[i] - raw_err_avg) *

                     (raw_motion_err_list[i] - raw_err_avg);

  }

  // Calculate the standard deviation for the motion error of all the inter

  // blocks of the 0,0 motion using the last source

  // frame as the reference.

  raw_err_stdev = sqrt(raw_err_stdev / raw_motion_err_counts);

  return raw_err_stdev;

}

static inline int calc_wavelet_energy(const AV1EncoderConfig *oxcf) {

  return oxcf->q_cfg.deltaq_mode == DELTA_Q_PERCEPTUAL;

}

typedef struct intra_pred_block_pass1_args {

  const SequenceHeader *seq_params;

  MACROBLOCK *x;

  bool do_border_pad;

} intra_pred_block_pass1_args;

static inline void copy_rect(uint8_t *dst, int dstride, const uint8_t *src,

                             int sstride, int width, int height, int use_hbd) {

#if CONFIG_AV1_HIGHBITDEPTH

  if (use_hbd) {

    aom_highbd_convolve_copy(CONVERT_TO_SHORTPTR(src), sstride,

                             CONVERT_TO_SHORTPTR(dst), dstride, width, height);

  } else {

    aom_convolve_copy(src, sstride, dst, dstride, width, height);

  }

#else

  (void)use_hbd;

  aom_convolve_copy(src, sstride, dst, dstride, width, height);

#endif

}

static void first_pass_intra_pred_and_calc_diff(int plane, int block,

                                                int blk_row, int blk_col,

                                                BLOCK_SIZE plane_bsize,

                                                TX_SIZE tx_size, void *arg) {

  (void)block;

  struct intra_pred_block_pass1_args *const args = arg;

  MACROBLOCK *const x = args->x;

  MACROBLOCKD *const xd = &x->e_mbd;

  MACROBLOCKD_PLANE *const pd = &xd->plane[plane];

  MACROBLOCK_PLANE *const p = &x->plane[plane];

  const int dst_stride = pd->dst.stride;

  uint8_t *dst = &pd->dst.buf[(blk_row * dst_stride + blk_col) << MI_SIZE_LOG2];

  const MB_MODE_INFO *const mbmi = xd->mi[0];

  const SequenceHeader *seq_params = args->seq_params;

  const int src_stride = p->src.stride;

  uint8_t *src = &p->src.buf[(blk_row * src_stride + blk_col) << MI_SIZE_LOG2];

  av1_predict_intra_block(

      xd, seq_params->sb_size, seq_params->enable_intra_edge_filter, pd->width,

      pd->height, tx_size, mbmi->mode, 0, 0, FILTER_INTRA_MODES, src,

      src_stride, dst, dst_stride, blk_col, blk_row, plane);

  av1_subtract_txb(x, plane, plane_bsize, blk_col, blk_row, tx_size, DCT_DCT,

                   args->do_border_pad);

}

static void first_pass_predict_intra_block_for_luma_plane(

    const SequenceHeader *seq_params, MACROBLOCK *x, BLOCK_SIZE bsize,

    bool do_border_pad) {

  assert(bsize < BLOCK_SIZES_ALL);

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

  const int plane = AOM_PLANE_Y;

  const MACROBLOCKD_PLANE *const pd = &xd->plane[plane];

  const int ss_x = pd->subsampling_x;

  const int ss_y = pd->subsampling_y;

  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, ss_x, ss_y);

  const int dst_stride = pd->dst.stride;

  uint8_t *dst = pd->dst.buf;

  const MACROBLOCK_PLANE *const p = &x->plane[plane];

  const int src_stride = p->src.stride;

  const uint8_t *src = p->src.buf;

  intra_pred_block_pass1_args args = { seq_params, x, do_border_pad };

  av1_foreach_transformed_block_in_plane(

      xd, plane_bsize, plane, first_pass_intra_pred_and_calc_diff, &args);

  // copy source data to recon buffer, as the recon buffer will be used as a

  // reference frame subsequently.

  copy_rect(dst, dst_stride, src, src_stride, block_size_wide[bsize],

            block_size_high[bsize], seq_params->use_highbitdepth);

}

#define UL_INTRA_THRESH 50

#define INVALID_ROW -1

// Computes and returns the intra pred error of a block.

// intra pred error: sum of squared error of the intra predicted residual.

// Inputs:

//   cpi: the encoder setting. Only a few params in it will be used.

//   this_frame: the current frame buffer.

//   tile: tile information (not used in first pass, already init to zero)

//   unit_row: row index in the unit of first pass block size.

//   unit_col: column index in the unit of first pass block size.

//   y_offset: the offset of y frame buffer, indicating the starting point of

//             the current block.

//   uv_offset: the offset of u and v frame buffer, indicating the starting

//              point of the current block.

//   fp_block_size: first pass block size.

//   qindex: quantization step size to encode the frame.

//   stats: frame encoding stats.

// Modifies:

//   stats->intra_skip_count

//   stats->image_data_start_row

//   stats->intra_factor

//   stats->brightness_factor

//   stats->intra_error

//   stats->frame_avg_wavelet_energy

// Returns:

//   this_intra_error.

static int firstpass_intra_prediction(

    AV1_COMP *cpi, ThreadData *td, YV12_BUFFER_CONFIG *const this_frame,

    const TileInfo *const tile, const int unit_row, const int unit_col,

    const int y_offset, const int uv_offset, const BLOCK_SIZE fp_block_size,

    const int qindex, FRAME_STATS *const stats) {

  const AV1_COMMON *const cm = &cpi->common;

  const CommonModeInfoParams *const mi_params = &cm->mi_params;

  const SequenceHeader *const seq_params = cm->seq_params;

  MACROBLOCK *const x = &td->mb;

  MACROBLOCKD *const xd = &x->e_mbd;

  const int unit_scale = mi_size_wide[fp_block_size];

  const int num_planes = av1_num_planes(cm);

  const BLOCK_SIZE bsize =

      get_bsize(mi_params, fp_block_size, unit_row, unit_col);

  set_mi_offsets(mi_params, xd, unit_row * unit_scale, unit_col * unit_scale);

  xd->plane[0].dst.buf = this_frame->y_buffer + y_offset;

  if (num_planes > 1) {

    xd->plane[1].dst.buf = this_frame->u_buffer + uv_offset;

    xd->plane[2].dst.buf = this_frame->v_buffer + uv_offset;

  }

  xd->left_available = (unit_col != 0);

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

  xd->mi[0]->ref_frame[0] = INTRA_FRAME;

  set_mi_row_col(xd, tile, unit_row * unit_scale, mi_size_high[bsize],

                 unit_col * unit_scale, mi_size_wide[bsize], mi_params->mi_rows,

                 mi_params->mi_cols);

  set_pixels_to_frame_edge(x, mi_size_wide[bsize], mi_size_high[bsize],

                           unit_col * unit_scale, unit_row * unit_scale,

                           mi_params->mi_cols, mi_params->mi_rows, cm->width,

                           cm->height, cpi->do_border_pad);

  set_plane_n4(xd, mi_size_wide[bsize], mi_size_high[bsize], num_planes);

  xd->mi[0]->segment_id = 0;

  xd->lossless[xd->mi[0]->segment_id] = (qindex == 0);

  xd->mi[0]->mode = DC_PRED;

  xd->mi[0]->tx_size = TX_4X4;

  xd->mi[0]->skip_txfm = 0;

  if (cpi->sf.fp_sf.disable_recon)

    first_pass_predict_intra_block_for_luma_plane(seq_params, x, bsize,

                                                  cpi->do_border_pad);

  else

    av1_encode_intra_block_plane(cpi, x, bsize, 0, DRY_RUN_NORMAL, 0);

  int this_intra_error = aom_get_mb_ss(x->plane[0].src_diff);

  if (seq_params->use_highbitdepth) {

    switch (seq_params->bit_depth) {

      case AOM_BITS_8: break;

      case AOM_BITS_10: this_intra_error >>= 4; break;

      case AOM_BITS_12: this_intra_error >>= 8; break;

      default:

        assert(0 &&

               "seq_params->bit_depth should be AOM_BITS_8, "

               "AOM_BITS_10 or AOM_BITS_12");

        return -1;

    }

  }

  if (this_intra_error < UL_INTRA_THRESH) {

    ++stats->intra_skip_count;

  } else if ((unit_col > 0) && (stats->image_data_start_row == INVALID_ROW)) {

    stats->image_data_start_row = unit_row;

  }

  double log_intra = log1p(this_intra_error);

  if (log_intra < 10.0) {

    stats->intra_factor += 1.0 + ((10.0 - log_intra) * 0.05);

  } else {

    stats->intra_factor += 1.0;

  }

  int level_sample;

  if (seq_params->use_highbitdepth) {

    level_sample = CONVERT_TO_SHORTPTR(x->plane[0].src.buf)[0];

  } else {

    level_sample = x->plane[0].src.buf[0];

  }

  if (seq_params->use_highbitdepth) {

    switch (seq_params->bit_depth) {

      case AOM_BITS_8: break;

      case AOM_BITS_10: level_sample >>= 2; break;

      case AOM_BITS_12: level_sample >>= 4; break;

      default:

        assert(0 &&

               "seq_params->bit_depth should be AOM_BITS_8, "

               "AOM_BITS_10 or AOM_BITS_12");

        return -1;

    }

  }

  if ((level_sample < DARK_THRESH) && (log_intra < 9.0)) {

    stats->brightness_factor += 1.0 + (0.01 * (DARK_THRESH - level_sample));

  } else {

    stats->brightness_factor += 1.0;

  }

  // Intrapenalty below deals with situations where the intra and inter

  // error scores are very low (e.g. a plain black frame).

  // We do not have special cases in first pass for 0,0 and nearest etc so

  // all inter modes carry an overhead cost estimate for the mv.

  // When the error score is very low this causes us to pick all or lots of

  // INTRA modes and throw lots of key frames.

  // This penalty adds a cost matching that of a 0,0 mv to the intra case.

  this_intra_error += INTRA_MODE_PENALTY;

  // Accumulate the intra error.

  stats->intra_error += (int64_t)this_intra_error;

  // Stats based on wavelet energy is used in the following cases :

  // 1. ML model which predicts if a flat structure (golden-frame only structure

  // without ALT-REF and Internal-ARFs) is better. This ML model is enabled in

  // constant quality mode under certain conditions.

  // 2. Delta qindex mode is set as DELTA_Q_PERCEPTUAL.

  // Thus, wavelet energy calculation is enabled for the above cases.

  if (calc_wavelet_energy(&cpi->oxcf)) {

    const int hbd = is_cur_buf_hbd(xd);

    const int stride = x->plane[0].src.stride;

    const int num_8x8_rows = block_size_high[fp_block_size] / 8;

    const int num_8x8_cols = block_size_wide[fp_block_size] / 8;

    const uint8_t *buf = x->plane[0].src.buf;

    stats->frame_avg_wavelet_energy += av1_haar_ac_sad_mxn_uint8_input(

        buf, stride, hbd, num_8x8_rows, num_8x8_cols);

  } else {

    stats->frame_avg_wavelet_energy = INVALID_FP_STATS_TO_PREDICT_FLAT_GOP;

  }

  return this_intra_error;

}

// Returns the sum of square error between source and reference blocks.

static int get_prediction_error_bitdepth(const int is_high_bitdepth,

                                         const int bitdepth,

                                         const BLOCK_SIZE block_size,

                                         const struct buf_2d *src,

                                         const struct buf_2d *ref) {

  (void)is_high_bitdepth;

  (void)bitdepth;

#if CONFIG_AV1_HIGHBITDEPTH

  if (is_high_bitdepth) {

    return highbd_get_prediction_error(block_size, src, ref, bitdepth);

  }

#endif  // CONFIG_AV1_HIGHBITDEPTH

  return get_prediction_error(block_size, src, ref);

}

// Accumulates motion vector stats.

// Modifies member variables of "stats".

static void accumulate_mv_stats(const MV best_mv, const FULLPEL_MV mv,

                                const int mb_row, const int mb_col,

                                const int mb_rows, const int mb_cols,

                                MV *last_non_zero_mv, FRAME_STATS *stats) {

  if (is_zero_mv(&best_mv)) return;

  ++stats->mv_count;

  // Non-zero vector, was it different from the last non zero vector?

  if (!is_equal_mv(&best_mv, last_non_zero_mv)) ++stats->new_mv_count;

  *last_non_zero_mv = best_mv;

  // Does the row vector point inwards or outwards?

  if (mb_row < mb_rows / 2) {

    if (mv.row > 0) {

      --stats->sum_in_vectors;

    } else if (mv.row < 0) {

      ++stats->sum_in_vectors;

    }

  } else if (mb_row > mb_rows / 2) {

    if (mv.row > 0) {

      ++stats->sum_in_vectors;

    } else if (mv.row < 0) {

      --stats->sum_in_vectors;

    }

  }

  // Does the col vector point inwards or outwards?

  if (mb_col < mb_cols / 2) {

    if (mv.col > 0) {

      --stats->sum_in_vectors;

    } else if (mv.col < 0) {

      ++stats->sum_in_vectors;

    }

  } else if (mb_col > mb_cols / 2) {

    if (mv.col > 0) {

      ++stats->sum_in_vectors;

    } else if (mv.col < 0) {

      --stats->sum_in_vectors;

    }

  }

}

// Computes and returns the inter prediction error from the last frame.

// Computes inter prediction errors from the golden and alt ref frams and

// Updates stats accordingly.

// Inputs:

//   cpi: the encoder setting. Only a few params in it will be used.

//   last_frame: the frame buffer of the last frame.

//   golden_frame: the frame buffer of the golden frame.

//   unit_row: row index in the unit of first pass block size.

//   unit_col: column index in the unit of first pass block size.

//   recon_yoffset: the y offset of the reconstructed frame buffer,

//                  indicating the starting point of the current block.

//   recon_uvoffset: the u/v offset of the reconstructed frame buffer,

//                   indicating the starting point of the current block.

//   src_yoffset: the y offset of the source frame buffer.

//   fp_block_size: first pass block size.

//   this_intra_error: the intra prediction error of this block.

//   raw_motion_err_counts: the count of raw motion vectors.

//   raw_motion_err_list: the array that records the raw motion error.

//   ref_mv: the reference used to start the motion search

//   best_mv: the best mv found

//   last_non_zero_mv: the last non zero mv found in this tile row.

//   stats: frame encoding stats.

//  Modifies:

//    raw_motion_err_list

//    best_ref_mv

//    last_mv

//    stats: many member params in it.

//  Returns:

//    this_inter_error

static int firstpass_inter_prediction(

    AV1_COMP *cpi, ThreadData *td, const YV12_BUFFER_CONFIG *const last_frame,

    const YV12_BUFFER_CONFIG *const golden_frame,

    const YV12_BUFFER_CONFIG *const last2_frame, const int unit_row,

    const int unit_col, const int recon_yoffset, const int recon_uvoffset,

    const int src_yoffset, const BLOCK_SIZE fp_block_size,

    const int this_intra_error, const int raw_motion_err_counts,

    int *raw_motion_err_list, const MV ref_mv, MV *best_mv,

    MV *last_non_zero_mv, FRAME_STATS *stats) {

  int this_inter_error = this_intra_error;

  AV1_COMMON *const cm = &cpi->common;

  const CommonModeInfoParams *const mi_params = &cm->mi_params;

  CurrentFrame *const current_frame = &cm->current_frame;

  MACROBLOCK *const x = &td->mb;

  MACROBLOCKD *const xd = &x->e_mbd;

  const int is_high_bitdepth = is_cur_buf_hbd(xd);

  const int bitdepth = xd->bd;

  const int unit_scale = mi_size_wide[fp_block_size];

  const BLOCK_SIZE bsize =

      get_bsize(mi_params, fp_block_size, unit_row, unit_col);

  const int fp_block_size_height = block_size_wide[fp_block_size];

  const int unit_width = mi_size_wide[fp_block_size];

  const int unit_rows = get_unit_rows(fp_block_size, mi_params->mb_rows);

  const int unit_cols = get_unit_cols(fp_block_size, mi_params->mb_cols);

  // Assume 0,0 motion with no mv overhead.

  FULLPEL_MV mv = kZeroFullMv;

  xd->plane[0].pre[0].buf = last_frame->y_buffer + recon_yoffset;

  // Set up limit values for motion vectors to prevent them extending

  // outside the UMV borders.

  av1_set_mv_col_limits(mi_params, &x->mv_limits, unit_col * unit_width,

                        fp_block_size_height >> MI_SIZE_LOG2,

                        cpi->oxcf.border_in_pixels);

  int motion_error =

      get_prediction_error_bitdepth(is_high_bitdepth, bitdepth, bsize,

                                    &x->plane[0].src, &xd->plane[0].pre[0]);

  // Compute the motion error of the 0,0 motion using the last source

  // frame as the reference. Skip the further motion search on

  // reconstructed frame if this error is small.

  // TODO(chiyotsai): The unscaled last source might be different dimension

  // as the current source. See BUG=aomedia:3413

  struct buf_2d unscaled_last_source_buf_2d;

  unscaled_last_source_buf_2d.buf =

      cpi->unscaled_last_source->y_buffer + src_yoffset;

  unscaled_last_source_buf_2d.stride = cpi->unscaled_last_source->y_stride;

  const int raw_motion_error = get_prediction_error_bitdepth(

      is_high_bitdepth, bitdepth, bsize, &x->plane[0].src,

      &unscaled_last_source_buf_2d);

  raw_motion_err_list[raw_motion_err_counts] = raw_motion_error;

  const FIRST_PASS_SPEED_FEATURES *const fp_sf = &cpi->sf.fp_sf;

  if (raw_motion_error > fp_sf->skip_motion_search_threshold) {

    // Test last reference frame using the previous best mv as the

    // starting point (best reference) for the search.

    first_pass_motion_search(cpi, x, &ref_mv, &mv, &motion_error);

    // If the current best reference mv is not centered on 0,0 then do a

    // 0,0 based search as well.

    if ((fp_sf->skip_zeromv_motion_search == 0) && !is_zero_mv(&ref_mv)) {

      FULLPEL_MV tmp_mv = kZeroFullMv;

      int tmp_err = INT_MAX;

      first_pass_motion_search(cpi, x, &kZeroMv, &tmp_mv, &tmp_err);

      if (tmp_err < motion_error) {

        motion_error = tmp_err;

        mv = tmp_mv;

      }

    }

  }

  int last2_motion_error = motion_error;

  if ((current_frame->frame_number > 2) && last2_frame != NULL) {

    FULLPEL_MV tmp_mv = kZeroFullMv;

    // Assume 0,0 motion with no mv overhead.

    av1_setup_pre_planes(xd, 0, last2_frame, 0, 0, NULL, 1);

    xd->plane[0].pre[0].buf += recon_yoffset;

    last2_motion_error =

        get_prediction_error_bitdepth(is_high_bitdepth, bitdepth, bsize,

                                      &x->plane[0].src, &xd->plane[0].pre[0]);

    first_pass_motion_search(cpi, x, &kZeroMv, &tmp_mv, &last2_motion_error);

    stats->lt_coded_error += AOMMIN(last2_motion_error, this_intra_error);

  }

  // Motion search in 2nd reference frame.

  int gf_motion_error = motion_error;

  if ((current_frame->frame_number > 1) && golden_frame != NULL) {

    FULLPEL_MV tmp_mv = kZeroFullMv;

    // Assume 0,0 motion with no mv overhead.

    av1_setup_pre_planes(xd, 0, golden_frame, 0, 0, NULL, 1);

    xd->plane[0].pre[0].buf += recon_yoffset;

    gf_motion_error =

        get_prediction_error_bitdepth(is_high_bitdepth, bitdepth, bsize,

                                      &x->plane[0].src, &xd->plane[0].pre[0]);

    first_pass_motion_search(cpi, x, &kZeroMv, &tmp_mv, &gf_motion_error);

  }

  if (gf_motion_error < motion_error && gf_motion_error < this_intra_error) {

    ++stats->second_ref_count;

  }

  // In accumulating a score for the 2nd reference frame take the

  // best of the motion predicted score and the intra coded error

  // (just as will be done for) accumulation of "coded_error" for

  // the last frame.

  if ((current_frame->frame_number > 1) && golden_frame != NULL) {

    stats->sr_coded_error += AOMMIN(gf_motion_error, this_intra_error);

  } else {

    // TODO(chengchen): I believe logically this should also be changed to

    // stats->sr_coded_error += AOMMIN(gf_motion_error, this_intra_error).

    stats->sr_coded_error += motion_error;

  }

  // Reset to last frame as reference buffer.

  xd->plane[0].pre[0].buf = last_frame->y_buffer + recon_yoffset;

  if (av1_num_planes(&cpi->common) > 1) {

    xd->plane[1].pre[0].buf = last_frame->u_buffer + recon_uvoffset;

    xd->plane[2].pre[0].buf = last_frame->v_buffer + recon_uvoffset;

  }

  // Start by assuming that intra mode is best.

  *best_mv = kZeroMv;

  if (motion_error <= this_intra_error) {

    // Keep a count of cases where the inter and intra were very close

    // and very low. This helps with scene cut detection for example in

    // cropped clips with black bars at the sides or top and bottom.

    if (((this_intra_error - INTRA_MODE_PENALTY) * 9 <= motion_error * 10) &&

        (this_intra_error < (2 * INTRA_MODE_PENALTY))) {

      stats->neutral_count += 1.0;

      // Also track cases where the intra is not much worse than the inter

      // and use this in limiting the GF/arf group length.

    } else if ((this_intra_error > NCOUNT_INTRA_THRESH) &&

               (this_intra_error < (NCOUNT_INTRA_FACTOR * motion_error))) {

      stats->neutral_count +=

          (double)motion_error / DOUBLE_DIVIDE_CHECK((double)this_intra_error);

    }

    *best_mv = get_mv_from_fullmv(&mv);

    this_inter_error = motion_error;

    xd->mi[0]->mode = NEWMV;

    xd->mi[0]->mv[0].as_mv = *best_mv;

    xd->mi[0]->tx_size = TX_4X4;

    xd->mi[0]->ref_frame[0] = LAST_FRAME;

    xd->mi[0]->ref_frame[1] = NONE_FRAME;

    if (fp_sf->disable_recon == 0) {

      av1_enc_build_inter_predictor(cm, xd, unit_row * unit_scale,

                                    unit_col * unit_scale, NULL, bsize,

                                    AOM_PLANE_Y, AOM_PLANE_Y);

      av1_encode_sby_pass1(cpi, x, bsize);

    }

    stats->sum_mvr += best_mv->row;

    stats->sum_mvr_abs += abs(best_mv->row);

    stats->sum_mvc += best_mv->col;

    stats->sum_mvc_abs += abs(best_mv->col);

    stats->sum_mvrs += best_mv->row * best_mv->row;

    stats->sum_mvcs += best_mv->col * best_mv->col;

    ++stats->inter_count;

    accumulate_mv_stats(*best_mv, mv, unit_row, unit_col, unit_rows, unit_cols,

                        last_non_zero_mv, stats);

  }

  return this_inter_error;

}

// Normalize the first pass stats.

// Error / counters are normalized to each MB.

// MVs are normalized to the width/height of the frame.

static void normalize_firstpass_stats(FIRSTPASS_STATS *fps,

                                      double num_mbs_16x16, double f_w,

                                      double f_h) {

  fps->coded_error /= num_mbs_16x16;

  fps->sr_coded_error /= num_mbs_16x16;

  fps->lt_coded_error /= num_mbs_16x16;

  fps->intra_error /= num_mbs_16x16;

  fps->frame_avg_wavelet_energy /= num_mbs_16x16;

  fps->log_coded_error = log1p(fps->coded_error);

  fps->log_intra_error = log1p(fps->intra_error);

  fps->MVr /= f_h;

  fps->mvr_abs /= f_h;

  fps->MVc /= f_w;

  fps->mvc_abs /= f_w;

  fps->MVrv /= (f_h * f_h);

  fps->MVcv /= (f_w * f_w);

  fps->new_mv_count /= num_mbs_16x16;

}

// Updates the first pass stats of this frame.

// Input:

//   cpi: the encoder setting. Only a few params in it will be used.

//   stats: stats accumulated for this frame.

//   raw_err_stdev: the statndard deviation for the motion error of all the

//                  inter blocks of the (0,0) motion using the last source

//                  frame as the reference.

//   frame_number: current frame number.