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 <math.h>
#include <limits.h>
#include <stdio.h>
#include "vp9/encoder/vp9_block.h"
#include "vp9/encoder/vp9_onyx_int.h"
#include "vp9/encoder/vp9_variance.h"
#include "vp9/encoder/vp9_encodeintra.h"
#include "vp9/encoder/vp9_mcomp.h"
#include "vp9/encoder/vp9_firstpass.h"
#include "vpx_scale/vpx_scale.h"
#include "vp9/encoder/vp9_encodeframe.h"
#include "vp9/encoder/vp9_encodemb.h"
#include "vp9/common/vp9_extend.h"
#include "vp9/common/vp9_systemdependent.h"
#include "vpx_mem/vpx_mem.h"
#include "vpx_scale/yv12config.h"
#include "vp9/encoder/vp9_quantize.h"
#include "vp9/encoder/vp9_rdopt.h"
#include "vp9/encoder/vp9_ratectrl.h"
#include "vp9/common/vp9_quant_common.h"
#include "vp9/common/vp9_entropymv.h"
#include "vp9/encoder/vp9_encodemv.h"
#include "vp9/encoder/vp9_vaq.h"
#include "./vpx_scale_rtcd.h"
// TODO(jkoleszar): for setup_dst_planes
#include "vp9/common/vp9_reconinter.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 POW1 (double)cpi->oxcf.two_pass_vbrbias/100.0
#define POW2 (double)cpi->oxcf.two_pass_vbrbias/100.0

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

static void find_next_key_frame(VP9_COMP *cpi, FIRSTPASS_STATS *this_frame);

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;
}


// Resets the first pass file to the given position using a relative seek from
// the current position.
static void reset_fpf_position(VP9_COMP *cpi, FIRSTPASS_STATS *position) {
  cpi->twopass.stats_in = position;
}

static int lookup_next_frame_stats(VP9_COMP *cpi, FIRSTPASS_STATS *next_frame) {
  if (cpi->twopass.stats_in >= cpi->twopass.stats_in_end)
    return EOF;

  *next_frame = *cpi->twopass.stats_in;
  return 1;
}

// Read frame stats at an offset from the current position
static int read_frame_stats(VP9_COMP *cpi,
                            FIRSTPASS_STATS *frame_stats,
                            int offset) {
  FIRSTPASS_STATS *fps_ptr = cpi->twopass.stats_in;

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

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

static int input_stats(VP9_COMP *cpi, FIRSTPASS_STATS *fps) {
  if (cpi->twopass.stats_in >= cpi->twopass.stats_in_end)
    return EOF;

  *fps = *cpi->twopass.stats_in;
  cpi->twopass.stats_in =
    (void *)((char *)cpi->twopass.stats_in + sizeof(FIRSTPASS_STATS));
  return 1;
}

static void output_stats(const VP9_COMP            *cpi,
                         struct vpx_codec_pkt_list *pktlist,
                         FIRSTPASS_STATS            *stats) {
  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(stdout, "%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(VP9_COMP *cpi,
                                     FIRSTPASS_STATS *this_frame) {
  const FIRSTPASS_STATS *const stats = &cpi->twopass.total_stats;
  const double av_err = stats->ssim_weighted_pred_err / stats->count;
  const double this_err = this_frame->ssim_weighted_pred_err;
  return av_err * pow(this_err / DOUBLE_DIVIDE_CHECK(av_err),
                      this_err > av_err ? POW1 : POW2);
}

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(YV12_BUFFER_CONFIG *source) {
  int i, j;

  uint8_t *src = source->y_buffer;
  double sum_weights = 0.0;

  // Loop through the Y plane examining levels and creating a weight for
  // the image.
  i = source->y_height;
  do {
    j = source->y_width;
    do {
      sum_weights += weight_table[ *src];
      src++;
    } while (--j);
    src -= source->y_width;
    src += source->y_stride;
  } while (--i);

  sum_weights /= (source->y_height * source->y_width);

  return sum_weights;
}


// This function returns the current per frame maximum bitrate target.
static int frame_max_bits(VP9_COMP *cpi) {
  // Max allocation for a single frame based on the max section guidelines
  // passed in and how many bits are left.
  // For VBR base this on the bits and frames left plus the
  // two_pass_vbrmax_section rate passed in by the user.
  const double max_bits = (1.0 * cpi->twopass.bits_left /
      (cpi->twopass.total_stats.count - cpi->common.current_video_frame)) *
      (cpi->oxcf.two_pass_vbrmax_section / 100.0);

  // Trap case where we are out of bits.
  return MAX((int)max_bits, 0);
}

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, cpi->output_pkt_list, &cpi->twopass.total_stats);
}

static void zz_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
                             YV12_BUFFER_CONFIG *recon_buffer,
                             int *best_motion_err, int recon_yoffset) {
  MACROBLOCKD *const xd = &x->e_mbd;

  // Set up pointers for this macro block recon buffer
  xd->plane[0].pre[0].buf = recon_buffer->y_buffer + recon_yoffset;

  switch (xd->mi_8x8[0]->mbmi.sb_type) {
    case BLOCK_8X8:
      vp9_mse8x8(x->plane[0].src.buf, x->plane[0].src.stride,
                 xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride,
                 (unsigned int *)(best_motion_err));
      break;
    case BLOCK_16X8:
      vp9_mse16x8(x->plane[0].src.buf, x->plane[0].src.stride,
                  xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride,
                  (unsigned int *)(best_motion_err));
      break;
    case BLOCK_8X16:
      vp9_mse8x16(x->plane[0].src.buf, x->plane[0].src.stride,
                  xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride,
                  (unsigned int *)(best_motion_err));
      break;
    default:
      vp9_mse16x16(x->plane[0].src.buf, x->plane[0].src.stride,
                   xd->plane[0].pre[0].buf, xd->plane[0].pre[0].stride,
                   (unsigned int *)(best_motion_err));
      break;
  }
}

static void first_pass_motion_search(VP9_COMP *cpi, MACROBLOCK *x,
                                     int_mv *ref_mv, MV *best_mv,
                                     YV12_BUFFER_CONFIG *recon_buffer,
                                     int *best_motion_err, int recon_yoffset) {
  MACROBLOCKD *const xd = &x->e_mbd;
  int num00;

  int_mv tmp_mv;
  int_mv ref_mv_full;

  int tmp_err;
  int step_param = 3;
  int further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
  int n;
  vp9_variance_fn_ptr_t v_fn_ptr =
      cpi->fn_ptr[xd->mi_8x8[0]->mbmi.sb_type];
  int new_mv_mode_penalty = 256;

  int sr = 0;
  int quart_frm = MIN(cpi->common.width, cpi->common.height);

  // refine the motion search range accroding to the frame dimension
  // for first pass test
  while ((quart_frm << sr) < MAX_FULL_PEL_VAL)
    sr++;
  if (sr)
    sr--;

  step_param    += sr;
  further_steps -= sr;

  // override the default variance function to use MSE
  switch (xd->mi_8x8[0]->mbmi.sb_type) {
    case BLOCK_8X8:
      v_fn_ptr.vf = vp9_mse8x8;
      break;
    case BLOCK_16X8:
      v_fn_ptr.vf = vp9_mse16x8;
      break;
    case BLOCK_8X16:
      v_fn_ptr.vf = vp9_mse8x16;
      break;
    default:
      v_fn_ptr.vf = vp9_mse16x16;
      break;
  }

  // Set up pointers for this macro block recon buffer
  xd->plane[0].pre[0].buf = recon_buffer->y_buffer + recon_yoffset;

  // Initial step/diamond search centred on best mv
  tmp_mv.as_int = 0;
  ref_mv_full.as_mv.col = ref_mv->as_mv.col >> 3;
  ref_mv_full.as_mv.row = ref_mv->as_mv.row >> 3;
  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 - 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.as_mv.row;
    best_mv->col = tmp_mv.as_mv.col;
  }

  // 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 - 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.as_mv.row;
        best_mv->col = tmp_mv.as_mv.col;
      }
    }
  }
}

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;
  PICK_MODE_CONTEXT *ctx = &x->sb64_context;
  int i;

  int recon_yoffset, recon_uvoffset;
  const int lst_yv12_idx = cm->ref_frame_map[cpi->lst_fb_idx];
  const int gld_yv12_idx = cm->ref_frame_map[cpi->gld_fb_idx];
  YV12_BUFFER_CONFIG *const lst_yv12 = &cm->yv12_fb[lst_yv12_idx];
  YV12_BUFFER_CONFIG *const gld_yv12 = &cm->yv12_fb[gld_yv12_idx];
  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;
  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;
  int 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;

  int_mv zero_ref_mv;

  zero_ref_mv.as_int = 0;

  vp9_clear_system_state();  // __asm emms;

  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;
  // required for vp9_frame_init_quantizer
  xd->mi_8x8[0] = cm->mi;

  setup_block_dptrs(&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];
    pd[i].qcoeff = ctx->qcoeff_pbuf[i][1];
    pd[i].dqcoeff = ctx->dqcoeff_pbuf[i][1];
    pd[i].eobs = ctx->eobs_pbuf[i][1];
  }


  // Initialise the MV cost table to the defaults
  // if( cm->current_video_frame == 0)
  // if ( 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 each macroblock row in image
  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 * 8);

    // 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 each macroblock col in image
    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
      int this_error;
      int gf_motion_error = INT_MAX;
      int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);
      double error_weight;

      vp9_clear_system_state();  // __asm emms;
      error_weight = 1.0;  // avoid uninitialized warnings

      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);

      if (mb_col * 2 + 1 < cm->mi_cols) {
        if (mb_row * 2 + 1 < cm->mi_rows) {
          xd->mi_8x8[0]->mbmi.sb_type = BLOCK_16X16;
        } else {
          xd->mi_8x8[0]->mbmi.sb_type = BLOCK_16X8;
        }
      } else {
        if (mb_row * 2 + 1 < cm->mi_rows) {
          xd->mi_8x8[0]->mbmi.sb_type = BLOCK_8X16;
        } else {
          xd->mi_8x8[0]->mbmi.sb_type = BLOCK_8X8;
        }
      }
      xd->mi_8x8[0]->mbmi.ref_frame[0] = INTRA_FRAME;
      set_mi_row_col(xd, &tile,
                     mb_row << 1,
                     num_8x8_blocks_high_lookup[xd->mi_8x8[0]->mbmi.sb_type],
                     mb_col << 1,
                     num_8x8_blocks_wide_lookup[xd->mi_8x8[0]->mbmi.sb_type],
                     cm->mi_rows, cm->mi_cols);

      if (cpi->sf.variance_adaptive_quantization) {
        int energy = vp9_block_energy(cpi, x, xd->mi_8x8[0]->mbmi.sb_type);
        error_weight = vp9_vaq_inv_q_ratio(energy);
      }

      // do intra 16x16 prediction
      this_error = vp9_encode_intra(x, use_dc_pred);
      if (cpi->sf.variance_adaptive_quantization) {
        vp9_clear_system_state();  // __asm emms;
        this_error *= error_weight;
      }

      // intrapenalty below deals with situations where the intra and inter
      // error scores are very low (eg 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;

      // Cumulative intra error total
      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;
        int motion_error = INT_MAX;
        int_mv mv, tmp_mv;

        // Simple 0,0 motion with no mv overhead
        zz_motion_search(cpi, x, lst_yv12, &motion_error, recon_yoffset);
        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,
                                 &mv.as_mv, lst_yv12,
                                 &motion_error, recon_yoffset);
        if (cpi->sf.variance_adaptive_quantization) {
          vp9_clear_system_state();  // __asm emms;
          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_ref_mv, &tmp_mv.as_mv,
                                   lst_yv12, &tmp_err, recon_yoffset);
          if (cpi->sf.variance_adaptive_quantization) {
            vp9_clear_system_state();  // __asm emms;
            tmp_err *= error_weight;
          }

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

        // Experimental search in an older reference frame
        if (cm->current_video_frame > 1) {
          // Simple 0,0 motion with no mv overhead
          zz_motion_search(cpi, x, gld_yv12,
                           &gf_motion_error, recon_yoffset);

          first_pass_motion_search(cpi, x, &zero_ref_mv,
                                   &tmp_mv.as_mv, gld_yv12,
                                   &gf_motion_error, recon_yoffset);
          if (cpi->sf.variance_adaptive_quantization) {
            vp9_clear_system_state();  // __asm emms;
            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;
        }
        /* Intra assumed 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;
          vp9_set_mbmode_and_mvs(x, NEWMV, &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,
                                         xd->mi_8x8[0]->mbmi.sb_type);
          vp9_encode_sby(x, xd->mi_8x8[0]->mbmi.sb_type);
          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;

          // Was the vector non-zero
          if (mv.as_int) {
            mvcount++;

            // 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 Row 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 macroblocks
      x->plane[0].src.buf += 16;
      x->plane[1].src.buf += 8;
      x->plane[2].src.buf += 8;

      recon_yoffset += 16;
      recon_uvoffset += 8;
    }

    // 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 += 8 * x->plane[1].src.stride - 8 * cm->mb_cols;
    x->plane[2].src.buf += 8 * x->plane[1].src.stride - 8 * cm->mb_cols;

    vp9_clear_system_state();  // __asm emms;
  }

  vp9_clear_system_state();  // __asm emms;
  {
    double weight = 0.0;

    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);
    weight = simple_weight(cpi->Source);


    if (weight < 0.1)
      weight = 0.1;

    fps.ssim_weighted_pred_err = fps.coded_error * weight;

    fps.pcnt_inter  = 0.0;
    fps.pcnt_motion = 0.0;
    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.count      = 1.0;

    fps.pcnt_inter   = 1.0 * (double)intercount / cm->MBs;
    fps.pcnt_second_ref = 1.0 * (double)second_ref_count / cm->MBs;
    fps.pcnt_neutral = 1.0 * (double)neutral_count / cm->MBs;

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

      fps.pcnt_motion = 1.0 * (double)mvcount / cpi->common.MBs;
    }

    // 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!
    cpi->twopass.this_frame_stats = fps;
    output_stats(cpi, cpi->output_pkt_list, &cpi->twopass.this_frame_stats);
    accumulate_stats(&cpi->twopass.total_stats, &fps);
  }

  // Copy the previous Last Frame back into gf and and arf buffers if
  // the prediction is good enough... but also dont allow it to lag too far
  if ((cpi->twopass.sr_update_lag > 3) ||
      ((cm->current_video_frame > 0) &&
       (cpi->twopass.this_frame_stats.pcnt_inter > 0.20) &&
       ((cpi->twopass.this_frame_stats.intra_error /
         DOUBLE_DIVIDE_CHECK(cpi->twopass.this_frame_stats.coded_error)) >
        2.0))) {
    vp8_yv12_copy_frame(lst_yv12, gld_yv12);
    cpi->twopass.sr_update_lag = 1;
  } else {
    cpi->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, cm->subsampling_x, cm->subsampling_y);

  // 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++;
}

// Estimate a cost per mb attributable to overheads such as the coding of
// modes and motion vectors.
// Currently simplistic in its assumptions for testing.
//


static double bitcost(double prob) {
  return -(log(prob) / log(2.0));
}

static int64_t estimate_modemvcost(VP9_COMP *cpi,
                                     FIRSTPASS_STATS *fpstats) {
#if 0
  int mv_cost;
  int mode_cost;

  double av_pct_inter = fpstats->pcnt_inter / fpstats->count;
  double av_pct_motion = fpstats->pcnt_motion / fpstats->count;
  double av_intra = (1.0 - av_pct_inter);

  double zz_cost;
  double motion_cost;
  double intra_cost;

  zz_cost = bitcost(av_pct_inter - av_pct_motion);
  motion_cost = bitcost(av_pct_motion);
  intra_cost = bitcost(av_intra);

  // Estimate of extra bits per mv overhead for mbs
  // << 9 is the normalization to the (bits * 512) used in vp9_bits_per_mb
  mv_cost = ((int)(fpstats->new_mv_count / fpstats->count) * 8) << 9;

  // Crude estimate of overhead cost from modes
  // << 9 is the normalization to (bits * 512) used in vp9_bits_per_mb
  mode_cost =
    (int)((((av_pct_inter - av_pct_motion) * zz_cost) +
           (av_pct_motion * motion_cost) +
           (av_intra * intra_cost)) * cpi->common.MBs) << 9;

  // return mv_cost + mode_cost;
  // TODO(paulwilkins): Fix overhead costs for extended Q range.
#endif
  return 0;
}

static double calc_correction_factor(double err_per_mb,
                                     double err_divisor,
                                     double pt_low,
                                     double pt_high,
                                     int q) {
  const double error_term = err_per_mb / err_divisor;

  // Adjustment based on actual quantizer to power term.
  const double power_term = MIN(vp9_convert_qindex_to_q(q) * 0.01 + pt_low,
                                pt_high);

  // Calculate correction factor
  if (power_term < 1.0)
    assert(error_term >= 0.0);

  return fclamp(pow(error_term, power_term), 0.05, 5.0);
}

// Given a current maxQ value sets a range for future values.
// PGW TODO..
// This code removes direct dependency on QIndex to determine the range
// (now uses the actual quantizer) but has not been tuned.
static void adjust_maxq_qrange(VP9_COMP *cpi) {
  int i;
  // Set the max corresponding to cpi->avg_q * 2.0
  double q = cpi->avg_q * 2.0;
  cpi->twopass.maxq_max_limit = cpi->worst_quality;
  for (i = cpi->best_quality; i <= cpi->worst_quality; i++) {
    cpi->twopass.maxq_max_limit = i;
    if (vp9_convert_qindex_to_q(i) >= q)
      break;
  }

  // Set the min corresponding to cpi->avg_q * 0.5
  q = cpi->avg_q * 0.5;
  cpi->twopass.maxq_min_limit = cpi->best_quality;
  for (i = cpi->worst_quality; i >= cpi->best_quality; i--) {
    cpi->twopass.maxq_min_limit = i;
    if (vp9_convert_qindex_to_q(i) <= q)
      break;
  }
}

static int estimate_max_q(VP9_COMP *cpi,
                          FIRSTPASS_STATS *fpstats,
                          int section_target_bandwitdh) {
  int q;
  int num_mbs = cpi->common.MBs;
  int target_norm_bits_per_mb;

  double section_err = fpstats->coded_error / fpstats->count;
  double sr_correction;
  double err_per_mb = section_err / num_mbs;
  double err_correction_factor;
  double speed_correction = 1.0;

  if (section_target_bandwitdh <= 0)
    return cpi->twopass.maxq_max_limit;          // Highest value allowed

  target_norm_bits_per_mb = section_target_bandwitdh < (1 << 20)
                              ? (512 * section_target_bandwitdh) / num_mbs
                              : 512 * (section_target_bandwitdh / num_mbs);

  // Look at the drop in prediction quality between the last frame
  // and the GF buffer (which contained an older frame).
  if (fpstats->sr_coded_error > fpstats->coded_error) {
    double sr_err_diff = (fpstats->sr_coded_error - fpstats->coded_error) /
                             (fpstats->count * cpi->common.MBs);
    sr_correction = fclamp(pow(sr_err_diff / 32.0, 0.25), 0.75, 1.25);
  } else {
    sr_correction = 0.75;
  }

  // Calculate a corrective factor based on a rolling ratio of bits spent
  // vs target bits
  if (cpi->rolling_target_bits > 0 &&
      cpi->active_worst_quality < cpi->worst_quality) {
    double rolling_ratio = (double)cpi->rolling_actual_bits /
                               (double)cpi->rolling_target_bits;

    if (rolling_ratio < 0.95)
      cpi->twopass.est_max_qcorrection_factor -= 0.005;
    else if (rolling_ratio > 1.05)
      cpi->twopass.est_max_qcorrection_factor += 0.005;

    cpi->twopass.est_max_qcorrection_factor = fclamp(
        cpi->twopass.est_max_qcorrection_factor, 0.1, 10.0);
  }

  // Corrections for higher compression speed settings
  // (reduced compression expected)
  // FIXME(jimbankoski): Once we settle on vp9 speed features we need to
  // change this code.
  if (cpi->compressor_speed == 1)
    speed_correction = cpi->oxcf.cpu_used <= 5 ?
                          1.04 + (/*cpi->oxcf.cpu_used*/0 * 0.04) :
                          1.25;

  // Try and pick a max Q that will be high enough to encode the
  // content at the given rate.
  for (q = cpi->twopass.maxq_min_limit; q < cpi->twopass.maxq_max_limit; q++) {
    int bits_per_mb_at_this_q;

    err_correction_factor = calc_correction_factor(err_per_mb,
                                                   ERR_DIVISOR, 0.4, 0.90, q) *
                                sr_correction * speed_correction *
                                cpi->twopass.est_max_qcorrection_factor;

    bits_per_mb_at_this_q = vp9_bits_per_mb(INTER_FRAME, q,
                                            err_correction_factor);

    if (bits_per_mb_at_this_q <= target_norm_bits_per_mb)
      break;
  }

  // Restriction on active max q for constrained quality mode.
  if (cpi->oxcf.end_usage == USAGE_CONSTRAINED_QUALITY &&
      q < cpi->cq_target_quality)
    q = cpi->cq_target_quality;

  // Adjust maxq_min_limit and maxq_max_limit limits based on
  // average q observed in clip for non kf/gf/arf frames
  // Give average a chance to settle though.
  // PGW TODO.. This code is broken for the extended Q range
  if (cpi->ni_frames > ((int)cpi->twopass.total_stats.count >> 8) &&
      cpi->ni_frames > 25)
    adjust_maxq_qrange(cpi);

  return q;
}

// For cq mode estimate a cq level that matches the observed
// complexity and data rate.
static int estimate_cq(VP9_COMP *cpi,
                       FIRSTPASS_STATS *fpstats,
                       int section_target_bandwitdh) {
  int q;
  int num_mbs = cpi->common.MBs;
  int target_norm_bits_per_mb;

  double section_err = (fpstats->coded_error / fpstats->count);
  double err_per_mb = section_err / num_mbs;
  double err_correction_factor;
  double sr_err_diff;
  double sr_correction;
  double speed_correction = 1.0;
  double clip_iiratio;
  double clip_iifactor;

  target_norm_bits_per_mb = (section_target_bandwitdh < (1 << 20))
                            ? (512 * section_target_bandwitdh) / num_mbs
                            : 512 * (section_target_bandwitdh / num_mbs);


  // Corrections for higher compression speed settings
  // (reduced compression expected)
  if (cpi->compressor_speed == 1) {
    if (cpi->oxcf.cpu_used <= 5)