vp9_firstpass.c

/*
 *  Copyright (c) 2010 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include <limits.h>
#include <math.h>
#include <stdio.h>

#include "./vpx_scale_rtcd.h"

#include "vpx_mem/vpx_mem.h"
#include "vpx_scale/vpx_scale.h"
#include "vpx_scale/yv12config.h"

#include "vp9/common/vp9_entropymv.h"
#include "vp9/common/vp9_quant_common.h"
#include "vp9/common/vp9_reconinter.h"  // setup_dst_planes()
#include "vp9/common/vp9_systemdependent.h"

#include "vp9/encoder/vp9_block.h"
#include "vp9/encoder/vp9_encodeframe.h"
#include "vp9/encoder/vp9_encodemb.h"
#include "vp9/encoder/vp9_encodemv.h"
#include "vp9/encoder/vp9_extend.h"
#include "vp9/encoder/vp9_firstpass.h"
#include "vp9/encoder/vp9_mcomp.h"
#include "vp9/encoder/vp9_onyx_int.h"
#include "vp9/encoder/vp9_quantize.h"
#include "vp9/encoder/vp9_ratectrl.h"
#include "vp9/encoder/vp9_rdopt.h"
#include "vp9/encoder/vp9_vaq.h"
#include "vp9/encoder/vp9_variance.h"

#define OUTPUT_FPF 0

#define IIFACTOR   12.5
#define IIKFACTOR1 12.5
#define IIKFACTOR2 15.0
#define RMAX       512.0
#define GF_RMAX    96.0
#define ERR_DIVISOR   150.0
#define MIN_DECAY_FACTOR 0.1

#define KF_MB_INTRA_MIN 150
#define GF_MB_INTRA_MIN 100

#define DOUBLE_DIVIDE_CHECK(x) ((x) < 0 ? (x) - 0.000001 : (x) + 0.000001)

#define MIN_KF_BOOST        300

#define DISABLE_RC_LONG_TERM_MEM 0

static void swap_yv12(YV12_BUFFER_CONFIG *a, YV12_BUFFER_CONFIG *b) {
  YV12_BUFFER_CONFIG temp = *a;
  *a = *b;
  *b = temp;
}

static int select_cq_level(int qindex) {
  int ret_val = QINDEX_RANGE - 1;
  int i;

  double target_q = (vp9_convert_qindex_to_q(qindex) * 0.5847) + 1.0;

  for (i = 0; i < QINDEX_RANGE; ++i) {
    if (target_q <= vp9_convert_qindex_to_q(i)) {
      ret_val = i;
      break;
    }
  }

  return ret_val;
}

static int gfboost_qadjust(int qindex) {
  const double q = vp9_convert_qindex_to_q(qindex);
  return (int)((0.00000828 * q * q * q) +
               (-0.0055 * q * q) +
               (1.32 * q) + 79.3);
}

static int kfboost_qadjust(int qindex) {
  const double q = vp9_convert_qindex_to_q(qindex);
  return (int)((0.00000973 * q * q * q) +
               (-0.00613 * q * q) +
               (1.316 * q) + 121.2);
}

// Resets the first pass file to the given position using a relative seek from
// the current position.
static void reset_fpf_position(struct twopass_rc *p,
                               FIRSTPASS_STATS *position) {
  p->stats_in = position;
}

static int lookup_next_frame_stats(const struct twopass_rc *p,
                                   FIRSTPASS_STATS *next_frame) {
  if (p->stats_in >= p->stats_in_end)
    return EOF;

  *next_frame = *p->stats_in;
  return 1;
}


// Read frame stats at an offset from the current position.
static int read_frame_stats(const struct twopass_rc *p,
                            FIRSTPASS_STATS *frame_stats, int offset) {
  const FIRSTPASS_STATS *fps_ptr = p->stats_in;

  // Check legality of offset.
  if (offset >= 0) {
    if (&fps_ptr[offset] >= p->stats_in_end)
      return EOF;
  } else if (offset < 0) {
    if (&fps_ptr[offset] < p->stats_in_start)
      return EOF;
  }

  *frame_stats = fps_ptr[offset];
  return 1;
}

static int input_stats(struct twopass_rc *p, FIRSTPASS_STATS *fps) {
  if (p->stats_in >= p->stats_in_end)
    return EOF;

  *fps = *p->stats_in;
  ++p->stats_in;
  return 1;
}

static void output_stats(FIRSTPASS_STATS *stats,
                         struct vpx_codec_pkt_list *pktlist) {
  struct vpx_codec_cx_pkt pkt;
  pkt.kind = VPX_CODEC_STATS_PKT;
  pkt.data.twopass_stats.buf = stats;
  pkt.data.twopass_stats.sz = sizeof(FIRSTPASS_STATS);
  vpx_codec_pkt_list_add(pktlist, &pkt);

// TEMP debug code
#if OUTPUT_FPF
  {
    FILE *fpfile;
    fpfile = fopen("firstpass.stt", "a");

    fprintf(fpfile, "%12.0f %12.0f %12.0f %12.0f %12.0f %12.4f %12.4f"
            "%12.4f %12.4f %12.4f %12.4f %12.4f %12.4f %12.4f"
            "%12.0f %12.0f %12.4f %12.0f %12.0f %12.4f\n",
            stats->frame,
            stats->intra_error,
            stats->coded_error,
            stats->sr_coded_error,
            stats->ssim_weighted_pred_err,
            stats->pcnt_inter,
            stats->pcnt_motion,
            stats->pcnt_second_ref,
            stats->pcnt_neutral,
            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
}

static void zero_stats(FIRSTPASS_STATS *section) {
  section->frame      = 0.0;
  section->intra_error = 0.0;
  section->coded_error = 0.0;
  section->sr_coded_error = 0.0;
  section->ssim_weighted_pred_err = 0.0;
  section->pcnt_inter  = 0.0;
  section->pcnt_motion  = 0.0;
  section->pcnt_second_ref = 0.0;
  section->pcnt_neutral = 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;
}

static void accumulate_stats(FIRSTPASS_STATS *section, FIRSTPASS_STATS *frame) {
  section->frame += frame->frame;
  section->intra_error += frame->intra_error;
  section->coded_error += frame->coded_error;
  section->sr_coded_error += frame->sr_coded_error;
  section->ssim_weighted_pred_err += frame->ssim_weighted_pred_err;
  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->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 void subtract_stats(FIRSTPASS_STATS *section, FIRSTPASS_STATS *frame) {
  section->frame -= frame->frame;
  section->intra_error -= frame->intra_error;
  section->coded_error -= frame->coded_error;
  section->sr_coded_error -= frame->sr_coded_error;
  section->ssim_weighted_pred_err -= frame->ssim_weighted_pred_err;
  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->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 void avg_stats(FIRSTPASS_STATS *section) {
  if (section->count < 1.0)
    return;

  section->intra_error /= section->count;
  section->coded_error /= section->count;
  section->sr_coded_error /= section->count;
  section->ssim_weighted_pred_err /= section->count;
  section->pcnt_inter  /= section->count;
  section->pcnt_second_ref /= section->count;
  section->pcnt_neutral /= section->count;
  section->pcnt_motion /= section->count;
  section->MVr        /= section->count;
  section->mvr_abs     /= section->count;
  section->MVc        /= section->count;
  section->mvc_abs     /= section->count;
  section->MVrv       /= section->count;
  section->MVcv       /= section->count;
  section->mv_in_out_count   /= section->count;
  section->duration   /= section->count;
}

// Calculate a modified Error used in distributing bits between easier and
// harder frames.
static double calculate_modified_err(const VP9_COMP *cpi,
                                     const FIRSTPASS_STATS *this_frame) {
  const struct twopass_rc *const twopass = &cpi->twopass;
  const FIRSTPASS_STATS *const stats = &twopass->total_stats;
  const double av_err = stats->ssim_weighted_pred_err / stats->count;
  double modified_error = av_err * pow(this_frame->ssim_weighted_pred_err /
                                           DOUBLE_DIVIDE_CHECK(av_err),
                                       cpi->oxcf.two_pass_vbrbias / 100.0);

  return fclamp(modified_error,
                twopass->modified_error_min, twopass->modified_error_max);
}

static const double weight_table[256] = {
  0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000,
  0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000,
  0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000,
  0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.020000,
  0.020000, 0.020000, 0.020000, 0.020000, 0.020000, 0.031250, 0.062500,
  0.093750, 0.125000, 0.156250, 0.187500, 0.218750, 0.250000, 0.281250,
  0.312500, 0.343750, 0.375000, 0.406250, 0.437500, 0.468750, 0.500000,
  0.531250, 0.562500, 0.593750, 0.625000, 0.656250, 0.687500, 0.718750,
  0.750000, 0.781250, 0.812500, 0.843750, 0.875000, 0.906250, 0.937500,
  0.968750, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000, 1.000000,
  1.000000, 1.000000, 1.000000, 1.000000
};

static double simple_weight(const YV12_BUFFER_CONFIG *buf) {
  int i, j;
  double sum = 0.0;
  const int w = buf->y_crop_width;
  const int h = buf->y_crop_height;
  const uint8_t *row = buf->y_buffer;

  for (i = 0; i < h; ++i) {
    const uint8_t *pixel = row;
    for (j = 0; j < w; ++j)
      sum += weight_table[*pixel++];
    row += buf->y_stride;
  }

  return MAX(0.1, sum / (w * h));
}

// This function returns the maximum target rate per frame.
static int frame_max_bits(const VP9_COMP *cpi) {
  int64_t max_bits =
    ((int64_t)cpi->rc.av_per_frame_bandwidth *
     (int64_t)cpi->oxcf.two_pass_vbrmax_section) / 100;

  if (max_bits < 0)
    max_bits = 0;
  else if (max_bits > cpi->rc.max_frame_bandwidth)
    max_bits = cpi->rc.max_frame_bandwidth;

  return (int)max_bits;
}

void vp9_init_first_pass(VP9_COMP *cpi) {
  zero_stats(&cpi->twopass.total_stats);
}

void vp9_end_first_pass(VP9_COMP *cpi) {
  output_stats(&cpi->twopass.total_stats, cpi->output_pkt_list);
}

static vp9_variance_fn_t get_block_variance_fn(BLOCK_SIZE bsize) {
  switch (bsize) {
    case BLOCK_8X8:
      return vp9_mse8x8;
    case BLOCK_16X8:
      return vp9_mse16x8;
    case BLOCK_8X16:
      return vp9_mse8x16;
    default:
      return vp9_mse16x16;
  }
}

static unsigned int zz_motion_search(const MACROBLOCK *x) {
  const MACROBLOCKD *const xd = &x->e_mbd;
  const uint8_t *const src = x->plane[0].src.buf;
  const int src_stride = x->plane[0].src.stride;
  const uint8_t *const ref = xd->plane[0].pre[0].buf;
  const int ref_stride = xd->plane[0].pre[0].stride;
  unsigned int sse;
  vp9_variance_fn_t fn = get_block_variance_fn(xd->mi_8x8[0]->mbmi.sb_type);
  fn(src, src_stride, ref, ref_stride, &sse);
  return sse;
}

static void first_pass_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
                                     const MV *ref_mv, MV *best_mv,
                                     int *best_motion_err) {
  MACROBLOCKD *const xd = &x->e_mbd;
  MV tmp_mv = {0, 0};
  MV ref_mv_full = {ref_mv->row >> 3, ref_mv->col >> 3};
  int num00, tmp_err, n, sr = 0;
  int step_param = 3;
  int further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
  const BLOCK_SIZE bsize = xd->mi_8x8[0]->mbmi.sb_type;
  vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[bsize];
  int new_mv_mode_penalty = 256;
  const int quart_frm = MIN(cpi->common.width, cpi->common.height);

  // Refine the motion search range according to the frame dimension
  // for first pass test.
  while ((quart_frm << sr) < MAX_FULL_PEL_VAL)
    ++sr;

  step_param += sr;
  further_steps -= sr;

  // Override the default variance function to use MSE.
  v_fn_ptr.vf = get_block_variance_fn(bsize);

  // Center the initial step/diamond search on best mv.
  tmp_err = cpi->diamond_search_sad(x, &ref_mv_full, &tmp_mv,
                                    step_param,
                                    x->sadperbit16, &num00, &v_fn_ptr,
                                    x->nmvjointcost,
                                    x->mvcost, ref_mv);
  if (tmp_err < INT_MAX)
    tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
  if (tmp_err < INT_MAX - new_mv_mode_penalty)
    tmp_err += new_mv_mode_penalty;

  if (tmp_err < *best_motion_err) {
    *best_motion_err = tmp_err;
    best_mv->row = tmp_mv.row;
    best_mv->col = tmp_mv.col;
  }

  // Carry out further step/diamond searches as necessary.
  n = num00;
  num00 = 0;

  while (n < further_steps) {
    ++n;

    if (num00) {
      --num00;
    } else {
      tmp_err = cpi->diamond_search_sad(x, &ref_mv_full, &tmp_mv,
                                        step_param + n, x->sadperbit16,
                                        &num00, &v_fn_ptr,
                                        x->nmvjointcost,
                                        x->mvcost, ref_mv);
      if (tmp_err < INT_MAX)
        tmp_err = vp9_get_mvpred_var(x, &tmp_mv, ref_mv, &v_fn_ptr, 1);
      if (tmp_err < INT_MAX - new_mv_mode_penalty)
        tmp_err += new_mv_mode_penalty;

      if (tmp_err < *best_motion_err) {
        *best_motion_err = tmp_err;
        best_mv->row = tmp_mv.row;
        best_mv->col = tmp_mv.col;
      }
    }
  }
}

static BLOCK_SIZE get_bsize(const VP9_COMMON *cm, int mb_row, int mb_col) {
  if (2 * mb_col + 1 < cm->mi_cols) {
    return 2 * mb_row + 1 < cm->mi_rows ? BLOCK_16X16
                                        : BLOCK_16X8;
  } else {
    return 2 * mb_row + 1 < cm->mi_rows ? BLOCK_8X16
                                        : BLOCK_8X8;
  }
}

void vp9_first_pass(VP9_COMP *cpi) {
  int mb_row, mb_col;
  MACROBLOCK *const x = &cpi->mb;
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCKD *const xd = &x->e_mbd;
  TileInfo tile;
  struct macroblock_plane *const p = x->plane;
  struct macroblockd_plane *const pd = xd->plane;
  const PICK_MODE_CONTEXT *ctx = &x->sb64_context;
  int i;

  int recon_yoffset, recon_uvoffset;
  YV12_BUFFER_CONFIG *const lst_yv12 = get_ref_frame_buffer(cpi, LAST_FRAME);
  YV12_BUFFER_CONFIG *const gld_yv12 = get_ref_frame_buffer(cpi, GOLDEN_FRAME);
  YV12_BUFFER_CONFIG *const new_yv12 = get_frame_new_buffer(cm);
  const int recon_y_stride = lst_yv12->y_stride;
  const int recon_uv_stride = lst_yv12->uv_stride;
  const int uv_mb_height = 16 >> (lst_yv12->y_height > lst_yv12->uv_height);
  int64_t intra_error = 0;
  int64_t coded_error = 0;
  int64_t sr_coded_error = 0;

  int sum_mvr = 0, sum_mvc = 0;
  int sum_mvr_abs = 0, sum_mvc_abs = 0;
  int64_t sum_mvrs = 0, sum_mvcs = 0;
  int mvcount = 0;
  int intercount = 0;
  int second_ref_count = 0;
  int intrapenalty = 256;
  int neutral_count = 0;
  int new_mv_count = 0;
  int sum_in_vectors = 0;
  uint32_t lastmv_as_int = 0;
  struct twopass_rc *const twopass = &cpi->twopass;
  const MV zero_mv = {0, 0};

  vp9_clear_system_state();

  vp9_setup_src_planes(x, cpi->Source, 0, 0);
  setup_pre_planes(xd, 0, lst_yv12, 0, 0, NULL);
  setup_dst_planes(xd, new_yv12, 0, 0);

  xd->mi_8x8 = cm->mi_grid_visible;
  xd->mi_8x8[0] = cm->mi;

  vp9_setup_block_planes(&x->e_mbd, cm->subsampling_x, cm->subsampling_y);

  vp9_frame_init_quantizer(cpi);

  for (i = 0; i < MAX_MB_PLANE; ++i) {
    p[i].coeff = ctx->coeff_pbuf[i][1];
    p[i].qcoeff = ctx->qcoeff_pbuf[i][1];
    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
    p[i].eobs = ctx->eobs_pbuf[i][1];
  }
  x->skip_recode = 0;

  vp9_init_mv_probs(cm);
  vp9_initialize_rd_consts(cpi);

  // Tiling is ignored in the first pass.
  vp9_tile_init(&tile, cm, 0, 0);

  for (mb_row = 0; mb_row < cm->mb_rows; ++mb_row) {
    int_mv best_ref_mv;

    best_ref_mv.as_int = 0;

    // Reset above block coeffs.
    xd->up_available = (mb_row != 0);
    recon_yoffset = (mb_row * recon_y_stride * 16);
    recon_uvoffset = (mb_row * recon_uv_stride * uv_mb_height);

    // Set up limit values for motion vectors to prevent them extending
    // outside the UMV borders.
    x->mv_row_min = -((mb_row * 16) + BORDER_MV_PIXELS_B16);
    x->mv_row_max = ((cm->mb_rows - 1 - mb_row) * 16)
                    + BORDER_MV_PIXELS_B16;

    for (mb_col = 0; mb_col < cm->mb_cols; ++mb_col) {
      int this_error;
      const int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
      double error_weight = 1.0;
      const BLOCK_SIZE bsize = get_bsize(cm, mb_row, mb_col);

      vp9_clear_system_state();

      xd->plane[0].dst.buf = new_yv12->y_buffer + recon_yoffset;
      xd->plane[1].dst.buf = new_yv12->u_buffer + recon_uvoffset;
      xd->plane[2].dst.buf = new_yv12->v_buffer + recon_uvoffset;
      xd->left_available = (mb_col != 0);
      xd->mi_8x8[0]->mbmi.sb_type = bsize;
      xd->mi_8x8[0]->mbmi.ref_frame[0] = INTRA_FRAME;
      set_mi_row_col(xd, &tile,
                     mb_row << 1, num_8x8_blocks_high_lookup[bsize],
                     mb_col << 1, num_8x8_blocks_wide_lookup[bsize],
                     cm->mi_rows, cm->mi_cols);

      if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
        const int energy = vp9_block_energy(cpi, x, bsize);
        error_weight = vp9_vaq_inv_q_ratio(energy);
      }

      // Do intra 16x16 prediction.
      this_error = vp9_encode_intra(x, use_dc_pred);
      if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
        vp9_clear_system_state();
        this_error = (int)(this_error * error_weight);
      }

      // 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_error += intrapenalty;

      // Accumulate the intra error.
      intra_error += (int64_t)this_error;

      // Set up limit values for motion vectors to prevent them extending
      // outside the UMV borders.
      x->mv_col_min = -((mb_col * 16) + BORDER_MV_PIXELS_B16);
      x->mv_col_max = ((cm->mb_cols - 1 - mb_col) * 16) + BORDER_MV_PIXELS_B16;

      // Other than for the first frame do a motion search.
      if (cm->current_video_frame > 0) {
        int tmp_err, motion_error;
        int_mv mv, tmp_mv;

        xd->plane[0].pre[0].buf = lst_yv12->y_buffer + recon_yoffset;
        motion_error = zz_motion_search(x);
        // Assume 0,0 motion with no mv overhead.
        mv.as_int = tmp_mv.as_int = 0;

        // Test last reference frame using the previous best mv as the
        // starting point (best reference) for the search.
        first_pass_motion_search(cpi, x, &best_ref_mv.as_mv, &mv.as_mv,
                                 &motion_error);
        if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
          vp9_clear_system_state();
          motion_error = (int)(motion_error * error_weight);
        }

        // If the current best reference mv is not centered on 0,0 then do a 0,0
        // based search as well.
        if (best_ref_mv.as_int) {
          tmp_err = INT_MAX;
          first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv.as_mv,
                                   &tmp_err);
          if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
            vp9_clear_system_state();
            tmp_err = (int)(tmp_err * error_weight);
          }

          if (tmp_err < motion_error) {
            motion_error = tmp_err;
            mv.as_int = tmp_mv.as_int;
          }
        }

        // Search in an older reference frame.
        if (cm->current_video_frame > 1) {
          // Assume 0,0 motion with no mv overhead.
          int gf_motion_error;

          xd->plane[0].pre[0].buf = gld_yv12->y_buffer + recon_yoffset;
          gf_motion_error = zz_motion_search(x);

          first_pass_motion_search(cpi, x, &zero_mv, &tmp_mv.as_mv,
                                   &gf_motion_error);
          if (cpi->oxcf.aq_mode == VARIANCE_AQ) {
            vp9_clear_system_state();
            gf_motion_error = (int)(gf_motion_error * error_weight);
          }

          if (gf_motion_error < motion_error && gf_motion_error < this_error)
            ++second_ref_count;

          // Reset to last frame as reference buffer.
          xd->plane[0].pre[0].buf = lst_yv12->y_buffer + recon_yoffset;
          xd->plane[1].pre[0].buf = lst_yv12->u_buffer + recon_uvoffset;
          xd->plane[2].pre[0].buf = lst_yv12->v_buffer + recon_uvoffset;

          // In accumulating a score for the older 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 (gf_motion_error < this_error)
            sr_coded_error += gf_motion_error;
          else
            sr_coded_error += this_error;
        } else {
          sr_coded_error += motion_error;
        }
        // Start by assuming that intra mode is best.
        best_ref_mv.as_int = 0;

        if (motion_error <= this_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_error - intrapenalty) * 9 <= motion_error * 10) &&
              this_error < 2 * intrapenalty)
            ++neutral_count;

          mv.as_mv.row *= 8;
          mv.as_mv.col *= 8;
          this_error = motion_error;
          xd->mi_8x8[0]->mbmi.mode = NEWMV;
          xd->mi_8x8[0]->mbmi.mv[0] = mv;
          xd->mi_8x8[0]->mbmi.tx_size = TX_4X4;
          xd->mi_8x8[0]->mbmi.ref_frame[0] = LAST_FRAME;
          xd->mi_8x8[0]->mbmi.ref_frame[1] = NONE;
          vp9_build_inter_predictors_sby(xd, mb_row << 1, mb_col << 1, bsize);
          vp9_encode_sby_pass1(x, bsize);
          sum_mvr += mv.as_mv.row;
          sum_mvr_abs += abs(mv.as_mv.row);
          sum_mvc += mv.as_mv.col;
          sum_mvc_abs += abs(mv.as_mv.col);
          sum_mvrs += mv.as_mv.row * mv.as_mv.row;
          sum_mvcs += mv.as_mv.col * mv.as_mv.col;
          ++intercount;

          best_ref_mv.as_int = mv.as_int;

          if (mv.as_int) {
            ++mvcount;

            // Non-zero vector, was it different from the last non zero vector?
            if (mv.as_int != lastmv_as_int)
              ++new_mv_count;
            lastmv_as_int = mv.as_int;

            // Does the row vector point inwards or outwards?
            if (mb_row < cm->mb_rows / 2) {
              if (mv.as_mv.row > 0)
                --sum_in_vectors;
              else if (mv.as_mv.row < 0)
                ++sum_in_vectors;
            } else if (mb_row > cm->mb_rows / 2) {
              if (mv.as_mv.row > 0)
                ++sum_in_vectors;
              else if (mv.as_mv.row < 0)
                --sum_in_vectors;
            }

            // Does the col vector point inwards or outwards?
            if (mb_col < cm->mb_cols / 2) {
              if (mv.as_mv.col > 0)
                --sum_in_vectors;
              else if (mv.as_mv.col < 0)
                ++sum_in_vectors;
            } else if (mb_col > cm->mb_cols / 2) {
              if (mv.as_mv.col > 0)
                ++sum_in_vectors;
              else if (mv.as_mv.col < 0)
                --sum_in_vectors;
            }
          }
        }
      } else {
        sr_coded_error += (int64_t)this_error;
      }
      coded_error += (int64_t)this_error;

      // Adjust to the next column of MBs.
      x->plane[0].src.buf += 16;
      x->plane[1].src.buf += uv_mb_height;
      x->plane[2].src.buf += uv_mb_height;

      recon_yoffset += 16;
      recon_uvoffset += uv_mb_height;
    }

    // Adjust to the next row of MBs.
    x->plane[0].src.buf += 16 * x->plane[0].src.stride - 16 * cm->mb_cols;
    x->plane[1].src.buf += uv_mb_height * x->plane[1].src.stride -
                           uv_mb_height * cm->mb_cols;
    x->plane[2].src.buf += uv_mb_height * x->plane[1].src.stride -
                           uv_mb_height * cm->mb_cols;

    vp9_clear_system_state();
  }

  vp9_clear_system_state();
  {
    FIRSTPASS_STATS fps;

    fps.frame = cm->current_video_frame;
    fps.intra_error = (double)(intra_error >> 8);
    fps.coded_error = (double)(coded_error >> 8);
    fps.sr_coded_error = (double)(sr_coded_error >> 8);
    fps.ssim_weighted_pred_err = fps.coded_error * simple_weight(cpi->Source);
    fps.count = 1.0;
    fps.pcnt_inter = (double)intercount / cm->MBs;
    fps.pcnt_second_ref = (double)second_ref_count / cm->MBs;
    fps.pcnt_neutral = (double)neutral_count / cm->MBs;

    if (mvcount > 0) {
      fps.MVr = (double)sum_mvr / mvcount;
      fps.mvr_abs = (double)sum_mvr_abs / mvcount;
      fps.MVc = (double)sum_mvc / mvcount;
      fps.mvc_abs = (double)sum_mvc_abs / mvcount;
      fps.MVrv = ((double)sum_mvrs - (fps.MVr * fps.MVr / mvcount)) / mvcount;
      fps.MVcv = ((double)sum_mvcs - (fps.MVc * fps.MVc / mvcount)) / mvcount;
      fps.mv_in_out_count = (double)sum_in_vectors / (mvcount * 2);
      fps.new_mv_count = new_mv_count;
      fps.pcnt_motion = (double)mvcount / cm->MBs;
    } else {
      fps.MVr = 0.0;
      fps.mvr_abs = 0.0;
      fps.MVc = 0.0;
      fps.mvc_abs = 0.0;
      fps.MVrv = 0.0;
      fps.MVcv = 0.0;
      fps.mv_in_out_count = 0.0;
      fps.new_mv_count = 0.0;
      fps.pcnt_motion = 0.0;
    }

    // TODO(paulwilkins):  Handle the case when duration is set to 0, or
    // something less than the full time between subsequent values of
    // cpi->source_time_stamp.
    fps.duration = (double)(cpi->source->ts_end - cpi->source->ts_start);

    // Don't want to do output stats with a stack variable!
    twopass->this_frame_stats = fps;
    output_stats(&twopass->this_frame_stats, cpi->output_pkt_list);
    accumulate_stats(&twopass->total_stats, &fps);
  }

  // Copy the previous Last Frame back into gf and and arf buffers if
  // the prediction is good enough... but also don't allow it to lag too far.
  if ((twopass->sr_update_lag > 3) ||
      ((cm->current_video_frame > 0) &&
       (twopass->this_frame_stats.pcnt_inter > 0.20) &&
       ((twopass->this_frame_stats.intra_error /
         DOUBLE_DIVIDE_CHECK(twopass->this_frame_stats.coded_error)) > 2.0))) {
    vp8_yv12_copy_frame(lst_yv12, gld_yv12);
    twopass->sr_update_lag = 1;
  } else {
    ++twopass->sr_update_lag;
  }
  // Swap frame pointers so last frame refers to the frame we just compressed.
  swap_yv12(lst_yv12, new_yv12);

  vp9_extend_frame_borders(lst_yv12);

  // Special case for the first frame. Copy into the GF buffer as a second
  // reference.
  if (cm->current_video_frame == 0)
    vp8_yv12_copy_frame(lst_yv12, gld_yv12);

  // Use this to see what the first pass reconstruction looks like.
  if (0) {
    char filename[512];
    FILE *recon_file;
    snprintf(filename, sizeof(filename), "enc%04d.yuv",
             (int)cm->current_video_frame);

    if (cm->current_video_frame == 0)
      recon_file = fopen(filename, "wb");
    else
      recon_file = fopen(filename, "ab");

    (void)fwrite(lst_yv12->buffer_alloc, lst_yv12->frame_size, 1, recon_file);
    fclose(recon_file);
  }

  ++cm->current_video_frame;
}