encodeframe.c

      int seg_ref_active;

      if (xd->mode_info_context->mbmi.ref_frame) {
        unsigned char pred_context;

        pred_context = vp9_get_pred_context(cm, xd, PRED_COMP);

        if (xd->mode_info_context->mbmi.second_ref_frame == INTRA_FRAME)
          cpi->single_pred_count[pred_context]++;
        else
          cpi->comp_pred_count[pred_context]++;
      }

#if CONFIG_SUPERBLOCKS
      if (xd->mode_info_context->mbmi.encoded_as_sb)
        vp8cx_encode_inter_superblock(cpi, x, tp, recon_yoffset, recon_uvoffset, mb_col, mb_row);
      else
#endif
        vp8cx_encode_inter_macroblock(cpi, x, tp,
                                      recon_yoffset, recon_uvoffset, 1);
        // Note the encoder may have changed the segment_id

#ifdef MODE_STATS
      inter_y_modes[mbmi->mode]++;

      if (mbmi->mode == SPLITMV) {
        int b;

        for (b = 0; b < x->partition_info->count; b++) {
          inter_b_modes[x->partition_info->bmi[b].mode]++;
        }
      }

#endif

      // If we have just a single reference frame coded for a segment then
      // exclude from the reference frame counts used to work out
      // probabilities. NOTE: At the moment we dont support custom trees
      // for the reference frame coding for each segment but this is a
      // possible future action.
      segment_id = &mbmi->segment_id;
      seg_ref_active = segfeature_active(xd, *segment_id, SEG_LVL_REF_FRAME);
      if (!seg_ref_active ||
          ((check_segref(xd, *segment_id, INTRA_FRAME) +
            check_segref(xd, *segment_id, LAST_FRAME) +
            check_segref(xd, *segment_id, GOLDEN_FRAME) +
            check_segref(xd, *segment_id, ALTREF_FRAME)) > 1)) {
        {
          cpi->count_mb_ref_frame_usage[mbmi->ref_frame]++;
        }
      }

      // Count of last ref frame 0,0 usage
      if ((mbmi->mode == ZEROMV) && (mbmi->ref_frame == LAST_FRAME))
        cpi->inter_zz_count++;
    }

#if CONFIG_SUPERBLOCKS
    if (xd->mode_info_context->mbmi.encoded_as_sb) {
      x->src.y_buffer += 32;
      x->src.u_buffer += 16;
      x->src.v_buffer += 16;

      x->gf_active_ptr      += 2;
      x->partition_info     += 2;
      xd->mode_info_context += 2;
      xd->prev_mode_info_context += 2;

      (*tp)->Token = EOSB_TOKEN;
      (*tp)++;
      if (mb_row < cm->mb_rows) cpi->tplist[mb_row].stop = *tp;
      break;
    }
#endif

    // Next MB
    mb_row += dy;
    mb_col += dx;

    x->src.y_buffer += 16 * (dx + dy * x->src.y_stride);
    x->src.u_buffer += 8  * (dx + dy * x->src.uv_stride);
    x->src.v_buffer += 8  * (dx + dy * x->src.uv_stride);

    x->gf_active_ptr      += offset_unextended;
    x->partition_info     += offset_extended;
    xd->mode_info_context += offset_extended;
    xd->prev_mode_info_context += offset_extended;

#if CONFIG_DEBUG
    assert((xd->prev_mode_info_context - cpi->common.prev_mip) ==
           (xd->mode_info_context - cpi->common.mip));
#endif
    (*tp)->Token = EOSB_TOKEN;
    (*tp)++;
    if (mb_row < cm->mb_rows) cpi->tplist[mb_row].stop = *tp;
  }

  // debug output
#if DBG_PRNT_SEGMAP
  {
    FILE *statsfile;
    statsfile = fopen("segmap2.stt", "a");
    fprintf(statsfile, "\n");
    fclose(statsfile);
  }
#endif
}

static
void encode_sb_row(VP8_COMP *cpi,
                   VP8_COMMON *cm,
                   int mb_row,
                   MACROBLOCK  *x,
                   MACROBLOCKD *xd,
                   TOKENEXTRA **tp,
                   int *totalrate) {
  int mb_col;
  int mb_cols = cm->mb_cols;

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

  // Code each SB in the row
  for (mb_col = 0; mb_col < mb_cols; mb_col += 2) {
    int mb_rate = 0, mb_dist = 0;
#if CONFIG_SUPERBLOCKS
    int sb_rate = INT_MAX, sb_dist;
#endif

#if CONFIG_DEBUG
    MODE_INFO *mic = xd->mode_info_context;
    PARTITION_INFO *pi = x->partition_info;
    signed char  *gfa = x->gf_active_ptr;
    unsigned char *yb = x->src.y_buffer;
    unsigned char *ub = x->src.u_buffer;
    unsigned char *vb = x->src.v_buffer;
#endif

#if CONFIG_SUPERBLOCKS
    // Pick modes assuming the SB is coded as 4 independent MBs
    xd->mode_info_context->mbmi.encoded_as_sb = 0;
#endif
    pick_mb_modes(cpi, cm, mb_row, mb_col, x, xd, tp, &mb_rate, &mb_dist);
#if CONFIG_SUPERBLOCKS
    mb_rate += vp8_cost_bit(cm->sb_coded, 0);
#endif

    x->src.y_buffer -= 32;
    x->src.u_buffer -= 16;
    x->src.v_buffer -= 16;

    x->gf_active_ptr -= 2;
    x->partition_info -= 2;
    xd->mode_info_context -= 2;
    xd->prev_mode_info_context -= 2;

#if CONFIG_DEBUG
    assert(x->gf_active_ptr == gfa);
    assert(x->partition_info == pi);
    assert(xd->mode_info_context == mic);
    assert(x->src.y_buffer == yb);
    assert(x->src.u_buffer == ub);
    assert(x->src.v_buffer == vb);
#endif

#if CONFIG_SUPERBLOCKS
    if (!(((    mb_cols & 1) && mb_col ==     mb_cols - 1) ||
          ((cm->mb_rows & 1) && mb_row == cm->mb_rows - 1))) {
      /* Pick a mode assuming that it applies to all 4 of the MBs in the SB */
      xd->mode_info_context->mbmi.encoded_as_sb = 1;
      pick_sb_modes(cpi, cm, mb_row, mb_col, x, xd, tp, &sb_rate, &sb_dist);
      sb_rate += vp8_cost_bit(cm->sb_coded, 1);
    }

    /* Decide whether to encode as a SB or 4xMBs */
    if (sb_rate < INT_MAX &&
        RDCOST(x->rdmult, x->rddiv, sb_rate, sb_dist) <
          RDCOST(x->rdmult, x->rddiv, mb_rate, mb_dist)) {
      xd->mode_info_context->mbmi.encoded_as_sb = 1;
      xd->mode_info_context[1].mbmi.encoded_as_sb = 1;
      xd->mode_info_context[cm->mode_info_stride].mbmi.encoded_as_sb = 1;
      xd->mode_info_context[1 + cm->mode_info_stride].mbmi.encoded_as_sb = 1;
      *totalrate += sb_rate;
    } else
#endif
    {
#if CONFIG_SUPERBLOCKS
      xd->mode_info_context->mbmi.encoded_as_sb = 0;
      if (cm->mb_cols - 1 > mb_col)
        xd->mode_info_context[1].mbmi.encoded_as_sb = 0;
      if (cm->mb_rows - 1 > mb_row) {
        xd->mode_info_context[cm->mode_info_stride].mbmi.encoded_as_sb = 0;
        if (cm->mb_cols - 1 > mb_col)
          xd->mode_info_context[1 + cm->mode_info_stride].mbmi.encoded_as_sb = 0;
      }
#endif
      *totalrate += mb_rate;
    }

    /* Encode SB using best computed mode(s) */
    encode_sb(cpi, cm, mb_row, mb_col, x, xd, tp);

#if CONFIG_DEBUG
    assert(x->gf_active_ptr == gfa + 2);
    assert(x->partition_info == pi + 2);
    assert(xd->mode_info_context == mic + 2);
    assert(x->src.y_buffer == yb + 32);
    assert(x->src.u_buffer == ub + 16);
    assert(x->src.v_buffer == vb + 16);
#endif
  }

  // this is to account for the border
  x->gf_active_ptr += mb_cols - (mb_cols & 0x1);
  x->partition_info += xd->mode_info_stride + 1 - (mb_cols & 0x1);
  xd->mode_info_context += xd->mode_info_stride + 1 - (mb_cols & 0x1);
  xd->prev_mode_info_context += xd->mode_info_stride + 1 - (mb_cols & 0x1);

#if CONFIG_DEBUG
  assert((xd->prev_mode_info_context - cpi->common.prev_mip) ==
         (xd->mode_info_context - cpi->common.mip));
#endif
}

void init_encode_frame_mb_context(VP8_COMP *cpi) {
  MACROBLOCK *const x = &cpi->mb;
  VP8_COMMON *const cm = &cpi->common;
  MACROBLOCKD *const xd = &x->e_mbd;

  // GF active flags data structure
  x->gf_active_ptr = (signed char *)cpi->gf_active_flags;

  // Activity map pointer
  x->mb_activity_ptr = cpi->mb_activity_map;

  x->act_zbin_adj = 0;
  cpi->seg0_idx = 0;
  vpx_memset(cpi->ref_pred_count, 0, sizeof(cpi->ref_pred_count));

  x->partition_info = x->pi;

  xd->mode_info_context = cm->mi;
  xd->mode_info_stride = cm->mode_info_stride;
  xd->prev_mode_info_context = cm->prev_mi;

  xd->frame_type = cm->frame_type;

  xd->frames_since_golden = cm->frames_since_golden;
  xd->frames_till_alt_ref_frame = cm->frames_till_alt_ref_frame;

  // reset intra mode contexts
  if (cm->frame_type == KEY_FRAME)
    vp8_init_mbmode_probs(cm);

  // Copy data over into macro block data structures.
  x->src = * cpi->Source;
  xd->pre = cm->yv12_fb[cm->lst_fb_idx];
  xd->dst = cm->yv12_fb[cm->new_fb_idx];

  // set up frame for intra coded blocks
  vp8_setup_intra_recon(&cm->yv12_fb[cm->new_fb_idx]);

  vp8_build_block_offsets(x);

  vp8_setup_block_dptrs(&x->e_mbd);

  vp8_setup_block_ptrs(x);

  xd->mode_info_context->mbmi.mode = DC_PRED;
  xd->mode_info_context->mbmi.uv_mode = DC_PRED;

  vp8_zero(cpi->count_mb_ref_frame_usage)
  vp8_zero(cpi->bmode_count)
  vp8_zero(cpi->ymode_count)
  vp8_zero(cpi->i8x8_mode_count)
  vp8_zero(cpi->y_uv_mode_count)
  vp8_zero(cpi->sub_mv_ref_count)
  vp8_zero(cpi->mbsplit_count)
  vp8_zero(cpi->common.fc.mv_ref_ct)
  vp8_zero(cpi->common.fc.mv_ref_ct_a)
#if CONFIG_SUPERBLOCKS
  vp8_zero(cpi->sb_ymode_count)
  cpi->sb_count = 0;
#endif
  // vp8_zero(cpi->uv_mode_count)

  vpx_memset(cm->above_context, 0,
             sizeof(ENTROPY_CONTEXT_PLANES) * cm->mb_cols);

  xd->fullpixel_mask = 0xffffffff;
  if (cm->full_pixel)
    xd->fullpixel_mask = 0xfffffff8;
}

static void encode_frame_internal(VP8_COMP *cpi) {
  int mb_row;
  MACROBLOCK *const x = &cpi->mb;
  VP8_COMMON *const cm = &cpi->common;
  MACROBLOCKD *const xd = &x->e_mbd;

  TOKENEXTRA *tp = cpi->tok;
  int totalrate;

  //printf("encode_frame_internal\n");

  // Compute a modified set of reference frame probabilities to use when
  // prediction fails. These are based on the current general estimates for
  // this frame which may be updated with each iteration of the recode loop.
  vp9_compute_mod_refprobs(cm);

#if CONFIG_NEW_MVREF
  // temp stats reset
  vp8_zero( cpi->best_ref_index_counts );
#endif

// debug output
#if DBG_PRNT_SEGMAP
  {
    FILE *statsfile;
    statsfile = fopen("segmap2.stt", "a");
    fprintf(statsfile, "\n");
    fclose(statsfile);
  }
#endif

  totalrate = 0;

  // Functions setup for all frame types so we can use MC in AltRef
  vp8_setup_interp_filters(xd, cm->mcomp_filter_type, cm);

  // Reset frame count of inter 0,0 motion vector usage.
  cpi->inter_zz_count = 0;

  cpi->prediction_error = 0;
  cpi->intra_error = 0;
  cpi->skip_true_count[0] = cpi->skip_true_count[1] = cpi->skip_true_count[2] = 0;
  cpi->skip_false_count[0] = cpi->skip_false_count[1] = cpi->skip_false_count[2] = 0;

#if CONFIG_PRED_FILTER
  if (cm->current_video_frame == 0) {
    // Initially assume that we'll signal the prediction filter
    // state at the frame level and that it is off.
    cpi->common.pred_filter_mode = 0;
    cpi->common.prob_pred_filter_off = 128;
  }
  cpi->pred_filter_on_count = 0;
  cpi->pred_filter_off_count = 0;
#endif
  vp8_zero(cpi->switchable_interp_count);

  xd->mode_info_context = cm->mi;
  xd->prev_mode_info_context = cm->prev_mi;

  vp8_zero(cpi->NMVcount);
  vp8_zero(cpi->coef_counts);
  vp8_zero(cpi->hybrid_coef_counts);
  vp8_zero(cpi->coef_counts_8x8);
  vp8_zero(cpi->hybrid_coef_counts_8x8);
  vp8_zero(cpi->coef_counts_16x16);
  vp8_zero(cpi->hybrid_coef_counts_16x16);

  vp8cx_frame_init_quantizer(cpi);

  vp8_initialize_rd_consts(cpi, cm->base_qindex + cm->y1dc_delta_q);
  vp8cx_initialize_me_consts(cpi, cm->base_qindex);

  if (cpi->oxcf.tuning == VP8_TUNE_SSIM) {
    // Initialize encode frame context.
    init_encode_frame_mb_context(cpi);

    // Build a frame level activity map
    build_activity_map(cpi);
  }

  // re-initencode frame context.
  init_encode_frame_mb_context(cpi);

  vpx_memset(cpi->rd_comp_pred_diff, 0, sizeof(cpi->rd_comp_pred_diff));
  vpx_memset(cpi->single_pred_count, 0, sizeof(cpi->single_pred_count));
  vpx_memset(cpi->comp_pred_count, 0, sizeof(cpi->comp_pred_count));
  vpx_memset(cpi->txfm_count, 0, sizeof(cpi->txfm_count));
  vpx_memset(cpi->txfm_count_8x8p, 0, sizeof(cpi->txfm_count_8x8p));
  vpx_memset(cpi->rd_tx_select_diff, 0, sizeof(cpi->rd_tx_select_diff));
  {
    struct vpx_usec_timer  emr_timer;
    vpx_usec_timer_start(&emr_timer);

    {
      // For each row of SBs in the frame
      for (mb_row = 0; mb_row < cm->mb_rows; mb_row += 2) {
        int offset = (cm->mb_cols + 1) & ~0x1;

        encode_sb_row(cpi, cm, mb_row, x, xd, &tp, &totalrate);

        // adjust to the next row of SBs
        x->src.y_buffer += 32 * x->src.y_stride - 16 * offset;
        x->src.u_buffer += 16 * x->src.uv_stride - 8 * offset;
        x->src.v_buffer += 16 * x->src.uv_stride - 8 * offset;
      }

      cpi->tok_count = tp - cpi->tok;
    }

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

  }

  // 256 rate units to the bit,
  // projected_frame_size in units of BYTES
  cpi->projected_frame_size = totalrate >> 8;


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

}

static int check_dual_ref_flags(VP8_COMP *cpi) {
  MACROBLOCKD *xd = &cpi->mb.e_mbd;
  int ref_flags = cpi->ref_frame_flags;

  if (segfeature_active(xd, 1, SEG_LVL_REF_FRAME)) {
    if ((ref_flags & (VP8_LAST_FLAG | VP8_GOLD_FLAG)) == (VP8_LAST_FLAG | VP8_GOLD_FLAG) &&
        check_segref(xd, 1, LAST_FRAME))
      return 1;
    if ((ref_flags & (VP8_GOLD_FLAG | VP8_ALT_FLAG)) == (VP8_GOLD_FLAG | VP8_ALT_FLAG) &&
        check_segref(xd, 1, GOLDEN_FRAME))
      return 1;
    if ((ref_flags & (VP8_ALT_FLAG  | VP8_LAST_FLAG)) == (VP8_ALT_FLAG  | VP8_LAST_FLAG) &&
        check_segref(xd, 1, ALTREF_FRAME))
      return 1;
    return 0;
  } else {
    return (!!(ref_flags & VP8_GOLD_FLAG) +
            !!(ref_flags & VP8_LAST_FLAG) +
            !!(ref_flags & VP8_ALT_FLAG)) >= 2;
  }
}

static void reset_skip_txfm_size(VP8_COMP *cpi, TX_SIZE txfm_max) {
  VP8_COMMON *cm = &cpi->common;
  int mb_row, mb_col, mis = cm->mode_info_stride;
  MODE_INFO *mi, *mi_ptr = cm->mi;
  MB_MODE_INFO *mbmi;
  MACROBLOCK *x = &cpi->mb;
  MACROBLOCKD *xd = &x->e_mbd;

  for (mb_row = 0; mb_row < cm->mb_rows; mb_row++, mi_ptr += mis) {
    mi = mi_ptr;
    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++, mi++) {
      mbmi = &mi->mbmi;
      if (mbmi->txfm_size > txfm_max) {
        int segment_id = mbmi->segment_id;
        xd->mode_info_context = mi;
        assert((segfeature_active(xd, segment_id, SEG_LVL_EOB) &&
                get_segdata(xd, segment_id, SEG_LVL_EOB) == 0) ||
               (cm->mb_no_coeff_skip && mbmi->mb_skip_coeff));
        mbmi->txfm_size = txfm_max;
      }
    }
  }
}

void vp8_encode_frame(VP8_COMP *cpi) {
  if (cpi->sf.RD) {
    int i, frame_type, pred_type;
    TXFM_MODE txfm_type;

    /*
     * This code does a single RD pass over the whole frame assuming
     * either compound, single or hybrid prediction as per whatever has
     * worked best for that type of frame in the past.
     * It also predicts whether another coding mode would have worked
     * better that this coding mode. If that is the case, it remembers
     * that for subsequent frames.
     * It does the same analysis for transform size selection also.
     */
    if (cpi->common.frame_type == KEY_FRAME)
      frame_type = 0;
    else if (cpi->is_src_frame_alt_ref && cpi->common.refresh_golden_frame)
      frame_type = 3;
    else if (cpi->common.refresh_golden_frame || cpi->common.refresh_alt_ref_frame)
      frame_type = 1;
    else
      frame_type = 2;

    /* prediction (compound, single or hybrid) mode selection */
    if (frame_type == 3)
      pred_type = SINGLE_PREDICTION_ONLY;
    else if (cpi->rd_prediction_type_threshes[frame_type][1] >
                 cpi->rd_prediction_type_threshes[frame_type][0] &&
             cpi->rd_prediction_type_threshes[frame_type][1] >
                 cpi->rd_prediction_type_threshes[frame_type][2] &&
             check_dual_ref_flags(cpi) && cpi->static_mb_pct == 100)
      pred_type = COMP_PREDICTION_ONLY;
    else if (cpi->rd_prediction_type_threshes[frame_type][0] >
                 cpi->rd_prediction_type_threshes[frame_type][2])
      pred_type = SINGLE_PREDICTION_ONLY;
    else
      pred_type = HYBRID_PREDICTION;

    /* transform size (4x4, 8x8, 16x16 or select-per-mb) selection */
#if CONFIG_LOSSLESS
    if (cpi->oxcf.lossless) {
      txfm_type = ONLY_4X4;
    } else
#endif
    /* FIXME (rbultje)
     * this is a hack (no really), basically to work around the complete
     * nonsense coefficient cost prediction for keyframes. The probabilities
     * are reset to defaults, and thus we basically have no idea how expensive
     * a 4x4 vs. 8x8 will really be. The result is that any estimate at which
     * of the two is better is utterly bogus.
     * I'd like to eventually remove this hack, but in order to do that, we
     * need to move the frame reset code from the frame encode init to the
     * bitstream write code, or alternatively keep a backup of the previous
     * keyframe's probabilities as an estimate of what the current keyframe's
     * coefficient cost distributions may look like. */
    if (frame_type == 0) {
      txfm_type = ALLOW_16X16;
    } else
#if 0
    /* FIXME (rbultje)
     * this code is disabled for a similar reason as the code above; the
     * problem is that each time we "revert" to 4x4 only (or even 8x8 only),
     * the coefficient probabilities for 16x16 (and 8x8) start lagging behind,
     * thus leading to them lagging further behind and not being chosen for
     * subsequent frames either. This is essentially a local minimum problem
     * that we can probably fix by estimating real costs more closely within
     * a frame, perhaps by re-calculating costs on-the-fly as frame encoding
     * progresses. */
    if (cpi->rd_tx_select_threshes[frame_type][TX_MODE_SELECT] >
            cpi->rd_tx_select_threshes[frame_type][ONLY_4X4] &&
        cpi->rd_tx_select_threshes[frame_type][TX_MODE_SELECT] >
            cpi->rd_tx_select_threshes[frame_type][ALLOW_16X16] &&
        cpi->rd_tx_select_threshes[frame_type][TX_MODE_SELECT] >
            cpi->rd_tx_select_threshes[frame_type][ALLOW_8X8]) {
      txfm_type = TX_MODE_SELECT;
    } else if (cpi->rd_tx_select_threshes[frame_type][ONLY_4X4] >
                  cpi->rd_tx_select_threshes[frame_type][ALLOW_8X8]
            && cpi->rd_tx_select_threshes[frame_type][ONLY_4X4] >
                  cpi->rd_tx_select_threshes[frame_type][ALLOW_16X16]
               ) {
      txfm_type = ONLY_4X4;
    } else if (cpi->rd_tx_select_threshes[frame_type][ALLOW_16X16] >=
                  cpi->rd_tx_select_threshes[frame_type][ALLOW_8X8]) {
      txfm_type = ALLOW_16X16;
    } else
      txfm_type = ALLOW_8X8;
#else
    txfm_type = cpi->rd_tx_select_threshes[frame_type][ALLOW_16X16] >=
                 cpi->rd_tx_select_threshes[frame_type][TX_MODE_SELECT] ?
    ALLOW_16X16 : TX_MODE_SELECT;
#endif
    cpi->common.txfm_mode = txfm_type;
    if (txfm_type != TX_MODE_SELECT) {
      cpi->common.prob_tx[0] = 128;
      cpi->common.prob_tx[1] = 128;
    }
    cpi->common.comp_pred_mode = pred_type;
    encode_frame_internal(cpi);

    for (i = 0; i < NB_PREDICTION_TYPES; ++i) {
      const int diff = cpi->rd_comp_pred_diff[i] / cpi->common.MBs;
      cpi->rd_prediction_type_threshes[frame_type][i] += diff;
      cpi->rd_prediction_type_threshes[frame_type][i] >>= 1;
    }

    for (i = 0; i < NB_TXFM_MODES; ++i) {
      int64_t pd = cpi->rd_tx_select_diff[i];
      int diff;
      if (i == TX_MODE_SELECT)
        pd -= RDCOST(cpi->mb.rdmult, cpi->mb.rddiv, 2048 * (TX_SIZE_MAX - 1), 0);
      diff = pd / cpi->common.MBs;
      cpi->rd_tx_select_threshes[frame_type][i] += diff;
      cpi->rd_tx_select_threshes[frame_type][i] /= 2;
    }

    if (cpi->common.comp_pred_mode == HYBRID_PREDICTION) {
      int single_count_zero = 0;
      int comp_count_zero = 0;

      for (i = 0; i < COMP_PRED_CONTEXTS; i++) {
        single_count_zero += cpi->single_pred_count[i];
        comp_count_zero += cpi->comp_pred_count[i];
      }

      if (comp_count_zero == 0) {
        cpi->common.comp_pred_mode = SINGLE_PREDICTION_ONLY;
      } else if (single_count_zero == 0) {
        cpi->common.comp_pred_mode = COMP_PREDICTION_ONLY;
      }
    }

    if (cpi->common.txfm_mode == TX_MODE_SELECT) {
      const int count4x4 = cpi->txfm_count[TX_4X4] + cpi->txfm_count_8x8p[TX_4X4];
      const int count8x8 = cpi->txfm_count[TX_8X8];
      const int count8x8_8x8p = cpi->txfm_count_8x8p[TX_8X8];
      const int count16x16 = cpi->txfm_count[TX_16X16];

      if (count4x4 == 0 && count16x16 == 0) {
        cpi->common.txfm_mode = ALLOW_8X8;
        reset_skip_txfm_size(cpi, TX_8X8);
      } else if (count8x8 == 0 && count16x16 == 0 && count8x8_8x8p == 0) {
        cpi->common.txfm_mode = ONLY_4X4;
        reset_skip_txfm_size(cpi, TX_4X4);
      } else if (count8x8 == 0 && count4x4 == 0) {
        cpi->common.txfm_mode = ALLOW_16X16;
      }
    }
  } else {
    encode_frame_internal(cpi);
  }

}

void vp8_setup_block_ptrs(MACROBLOCK *x) {
  int r, c;
  int i;

  for (r = 0; r < 4; r++) {
    for (c = 0; c < 4; c++) {
      x->block[r * 4 + c].src_diff = x->src_diff + r * 4 * 16 + c * 4;
    }
  }

  for (r = 0; r < 2; r++) {
    for (c = 0; c < 2; c++) {
      x->block[16 + r * 2 + c].src_diff = x->src_diff + 256 + r * 4 * 8 + c * 4;
    }
  }


  for (r = 0; r < 2; r++) {
    for (c = 0; c < 2; c++) {
      x->block[20 + r * 2 + c].src_diff = x->src_diff + 320 + r * 4 * 8 + c * 4;
    }
  }

  x->block[24].src_diff = x->src_diff + 384;


  for (i = 0; i < 25; i++) {
    x->block[i].coeff = x->coeff + i * 16;
  }
}

void vp8_build_block_offsets(MACROBLOCK *x) {
  int block = 0;
  int br, bc;

  vp8_build_block_doffsets(&x->e_mbd);

  // y blocks
  x->thismb_ptr = &x->thismb[0];
  for (br = 0; br < 4; br++) {
    for (bc = 0; bc < 4; bc++) {
      BLOCK *this_block = &x->block[block];
      // this_block->base_src = &x->src.y_buffer;
      // this_block->src_stride = x->src.y_stride;
      // this_block->src = 4 * br * this_block->src_stride + 4 * bc;
      this_block->base_src = &x->thismb_ptr;
      this_block->src_stride = 16;
      this_block->src = 4 * br * 16 + 4 * bc;
      ++block;
    }
  }

  // u blocks
  for (br = 0; br < 2; br++) {
    for (bc = 0; bc < 2; bc++) {
      BLOCK *this_block = &x->block[block];
      this_block->base_src = &x->src.u_buffer;
      this_block->src_stride = x->src.uv_stride;
      this_block->src = 4 * br * this_block->src_stride + 4 * bc;
      ++block;
    }
  }

  // v blocks
  for (br = 0; br < 2; br++) {
    for (bc = 0; bc < 2; bc++) {
      BLOCK *this_block = &x->block[block];
      this_block->base_src = &x->src.v_buffer;
      this_block->src_stride = x->src.uv_stride;
      this_block->src = 4 * br * this_block->src_stride + 4 * bc;
      ++block;
    }
  }
}

static void sum_intra_stats(VP8_COMP *cpi, MACROBLOCK *x) {
  const MACROBLOCKD *xd = &x->e_mbd;
  const MB_PREDICTION_MODE m = xd->mode_info_context->mbmi.mode;
  const MB_PREDICTION_MODE uvm = xd->mode_info_context->mbmi.uv_mode;

#ifdef MODE_STATS
  const int is_key = cpi->common.frame_type == KEY_FRAME;

  ++ (is_key ? uv_modes : inter_uv_modes)[uvm];
  ++ uv_modes_y[m][uvm];

  if (m == B_PRED) {
    unsigned int *const bct = is_key ? b_modes : inter_b_modes;

    int b = 0;

    do {
      ++ bct[xd->block[b].bmi.as_mode.first];
    } while (++b < 16);
  }

  if (m == I8X8_PRED) {
    i8x8_modes[xd->block[0].bmi.as_mode.first]++;
    i8x8_modes[xd->block[2].bmi.as_mode.first]++;
    i8x8_modes[xd->block[8].bmi.as_mode.first]++;
    i8x8_modes[xd->block[10].bmi.as_mode.first]++;
  }
#endif

#if CONFIG_SUPERBLOCKS
  if (xd->mode_info_context->mbmi.encoded_as_sb) {
    ++cpi->sb_ymode_count[m];
  } else
#endif
    ++cpi->ymode_count[m];
  if (m != I8X8_PRED)
    ++cpi->y_uv_mode_count[m][uvm];
  else {
    cpi->i8x8_mode_count[xd->block[0].bmi.as_mode.first]++;
    cpi->i8x8_mode_count[xd->block[2].bmi.as_mode.first]++;
    cpi->i8x8_mode_count[xd->block[8].bmi.as_mode.first]++;
    cpi->i8x8_mode_count[xd->block[10].bmi.as_mode.first]++;
  }
  if (m == B_PRED) {
    int b = 0;
    do {
      ++ cpi->bmode_count[xd->block[b].bmi.as_mode.first];
    } while (++b < 16);
  }
}

// Experimental stub function to create a per MB zbin adjustment based on
// some previously calculated measure of MB activity.
static void adjust_act_zbin(VP8_COMP *cpi, MACROBLOCK *x) {
#if USE_ACT_INDEX
  x->act_zbin_adj = *(x->mb_activity_ptr);
#else
  int64_t a;
  int64_t b;
  int64_t act = *(x->mb_activity_ptr);

  // Apply the masking to the RD multiplier.
  a = act + 4 * cpi->activity_avg;
  b = 4 * act + cpi->activity_avg;

  if (act > cpi->activity_avg)
    x->act_zbin_adj = (int)(((int64_t)b + (a >> 1)) / a) - 1;
  else
    x->act_zbin_adj = 1 - (int)(((int64_t)a + (b >> 1)) / b);
#endif
}

#if CONFIG_SUPERBLOCKS
static void update_sb_skip_coeff_state(VP8_COMP *cpi,
                                       MACROBLOCK *x,
                                       ENTROPY_CONTEXT_PLANES ta[4],
                                       ENTROPY_CONTEXT_PLANES tl[4],
                                       TOKENEXTRA *t[4],
                                       TOKENEXTRA **tp,
                                       int skip[4])
{
  TOKENEXTRA tokens[4][16 * 24];
  int n_tokens[4], n;

  // if there were no skips, we don't need to do anything
  if (!skip[0] && !skip[1] && !skip[2] && !skip[3])
    return;

  // if we don't do coeff skipping for this frame, we don't
  // need to do anything here
  if (!cpi->common.mb_no_coeff_skip)
    return;

  // if all 4 MBs skipped coeff coding, nothing to be done
  if (skip[0] && skip[1] && skip[2] && skip[3])
    return;

  // so the situation now is that we want to skip coeffs
  // for some MBs, but not all, and we didn't code EOB
  // coefficients for them. However, the skip flag for this
  // SB will be 0 overall, so we need to insert EOBs in the
  // middle of the token tree. Do so here.
  n_tokens[0] = t[1] - t[0];
  n_tokens[1] = t[2] - t[1];
  n_tokens[2] = t[3] - t[2];
  n_tokens[3] = *tp  - t[3];
  if (n_tokens[0])
    memcpy(tokens[0], t[0], n_tokens[0] * sizeof(*t[0]));
  if (n_tokens[1])
    memcpy(tokens[1], t[1], n_tokens[1] * sizeof(*t[0]));
  if (n_tokens[2])
    memcpy(tokens[2], t[2], n_tokens[2] * sizeof(*t[0]));
  if (n_tokens[3])
    memcpy(tokens[3], t[3], n_tokens[3] * sizeof(*t[0]));

  // reset pointer, stuff EOBs where necessary
  *tp = t[0];
  for (n = 0; n < 4; n++) {
    if (skip[n]) {
      x->e_mbd.above_context = &ta[n];
      x->e_mbd.left_context  = &tl[n];
      vp8_stuff_mb(cpi, &x->e_mbd, tp, 0);
    } else {
      if (n_tokens[n]) {
        memcpy(*tp, tokens[n], sizeof(*t[0]) * n_tokens[n]);
      }
      (*tp) += n_tokens[n];
    }
  }
}

void vp8cx_encode_intra_super_block(VP8_COMP *cpi,
                                    MACROBLOCK *x,
                                    TOKENEXTRA **t,
                                    int mb_col) {
  const int output_enabled = 1;
  int n;
  MACROBLOCKD *xd = &x->e_mbd;
  VP8_COMMON *cm = &cpi->common;
  const uint8_t *src = x->src.y_buffer;
  uint8_t *dst = xd->dst.y_buffer;
  const uint8_t *usrc = x->src.u_buffer;
  uint8_t *udst = xd->dst.u_buffer;
  const uint8_t *vsrc = x->src.v_buffer;
  uint8_t *vdst = xd->dst.v_buffer;
  int src_y_stride = x->src.y_stride, dst_y_stride = xd->dst.y_stride;
  int src_uv_stride = x->src.uv_stride, dst_uv_stride = xd->dst.uv_stride;
  const VP8_ENCODER_RTCD *rtcd = IF_RTCD(&cpi->rtcd);
  TOKENEXTRA *tp[4];
  int skip[4];
  MODE_INFO *mi = x->e_mbd.mode_info_context;
  ENTROPY_CONTEXT_PLANES ta[4], tl[4];

  if ((cpi->oxcf.tuning == VP8_TUNE_SSIM) && output_enabled) {
    adjust_act_zbin(cpi, x);
    vp8_update_zbin_extra(cpi, x);
  }

  vp8_build_intra_predictors_sby_s(&x->e_mbd);
  vp8_build_intra_predictors_sbuv_s(&x->e_mbd);

  assert(x->e_mbd.mode_info_context->mbmi.txfm_size == TX_8X8);
  for (n = 0; n < 4; n++)
  {
    int x_idx = n & 1, y_idx = n >> 1;

    xd->above_context = cm->above_context + mb_col + (n & 1);
    xd->left_context = cm->left_context + (n >> 1);

    vp8_subtract_mby_s_c(x->src_diff,
                         src + x_idx * 16 + y_idx * 16 * src_y_stride,
                         src_y_stride,
                         dst + x_idx * 16 + y_idx * 16 * dst_y_stride,
                         dst_y_stride);
    vp8_subtract_mbuv_s_c(x->src_diff,
                          usrc + x_idx * 8 + y_idx * 8 * src_uv_stride,
                          vsrc + x_idx * 8 + y_idx * 8 * src_uv_stride,
                          src_uv_stride,
                          udst + x_idx * 8 + y_idx * 8 * dst_uv_stride,
                          vdst + x_idx * 8 + y_idx * 8 * dst_uv_stride,
                          dst_uv_stride);
    vp8_transform_mb_8x8(x);
    vp8_quantize_mb_8x8(x);
    if (x->optimize) {
      vp8_optimize_mby_8x8(x, rtcd);
      vp8_optimize_mbuv_8x8(x, rtcd);
    }
    vp8_inverse_transform_mb_8x8(IF_RTCD(&rtcd->common->idct), &x->e_mbd);
    vp8_recon_mby_s_c(&x->e_mbd, dst + x_idx * 16 + y_idx * 16 * dst_y_stride);
    vp8_recon_mbuv_s_c(&x->e_mbd,
                       udst + x_idx * 8 + y_idx * 8 * dst_uv_stride,
                       vdst + x_idx * 8 + y_idx * 8 * dst_uv_stride);

    if (output_enabled) {
      memcpy(&ta[n], xd->above_context, sizeof(ta[n]));
      memcpy(&tl[n], xd->left_context, sizeof(tl[n]));
      tp[n] = *t;
      xd->mode_info_context = mi + x_idx + y_idx * cm->mode_info_stride;
      vp8_tokenize_mb(cpi, &x->e_mbd, t, 0);
      skip[n] = xd->mode_info_context->mbmi.mb_skip_coeff;
    }
  }

  if (output_enabled) {
    // Tokenize
    xd->mode_info_context = mi;
    sum_intra_stats(cpi, x);
    update_sb_skip_coeff_state(cpi, x, ta, tl, tp, t, skip);
  }
}
#endif /* CONFIG_SUPERBLOCKS */

void vp8cx_encode_intra_macro_block(VP8_COMP *cpi,
                                    MACROBLOCK *x,
                                    TOKENEXTRA **t,
                                    int output_enabled) {
  MB_MODE_INFO * mbmi = &x->e_mbd.mode_info_context->mbmi;
  if ((cpi->oxcf.tuning == VP8_TUNE_SSIM) && output_enabled) {
    adjust_act_zbin(cpi, x);
    vp8_update_zbin_extra(cpi, x);
  }
  if (mbmi->mode == I8X8_PRED) {
    vp8_encode_intra8x8mby(IF_RTCD(&cpi->rtcd), x);
    vp8_encode_intra8x8mbuv(IF_RTCD(&cpi->rtcd), x);
  } else if (mbmi->mode == B_PRED) {
    vp8_encode_intra4x4mby(IF_RTCD(&cpi->rtcd), x);
  } else {
    vp8_encode_intra16x16mby(IF_RTCD(&cpi->rtcd), x);
  }

  if (mbmi->mode != I8X8_PRED) {
    vp8_encode_intra16x16mbuv(IF_RTCD(&cpi->rtcd), x);
  }

  if (output_enabled) {
    int segment_id = mbmi->segment_id;

    // Tokenize
    sum_intra_stats(cpi, x);
    vp8_tokenize_mb(cpi, &x->e_mbd, t, 0);

    if (cpi->common.txfm_mode == TX_MODE_SELECT &&
        !((cpi->common.mb_no_coeff_skip && mbmi->mb_skip_coeff) ||
          (segfeature_active(&x->e_mbd, segment_id, SEG_LVL_EOB) &&
           get_segdata(&x->e_mbd, segment_id, SEG_LVL_EOB) == 0))) {
      if (mbmi->mode != B_PRED && mbmi->mode != I8X8_PRED) {
        cpi->txfm_count[mbmi->txfm_size]++;
      } else if (mbmi->mode == I8X8_PRED) {
        cpi->txfm_count_8x8p[mbmi->txfm_size]++;
      }
    } else if (cpi->common.txfm_mode >= ALLOW_16X16 && mbmi->mode <= TM_PRED) {
      mbmi->txfm_size = TX_16X16;
    } else
    if (cpi->common.txfm_mode >= ALLOW_8X8 && mbmi->mode != B_PRED) {
      mbmi->txfm_size = TX_8X8;
    } else {
      mbmi->txfm_size = TX_4X4;
    }
  }
#if CONFIG_NEWBESTREFMV
  else
    vp8_tokenize_mb(cpi, &x->e_mbd, t, 1);
#endif
}
#ifdef SPEEDSTATS
extern int cnt_pm;
#endif

extern void vp8_fix_contexts(MACROBLOCKD *xd);

void vp8cx_encode_inter_macroblock (VP8_COMP *cpi, MACROBLOCK *x,
                                    TOKENEXTRA **t, int recon_yoffset,
                                    int recon_uvoffset, int output_enabled) {
  VP8_COMMON *cm = &cpi->common;
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO * mbmi = &xd->mode_info_context->mbmi;
  unsigned char *segment_id = &mbmi->segment_id;
  int seg_ref_active;
  unsigned char ref_pred_flag;

  x->skip = 0;
#if CONFIG_SUPERBLOCKS
  assert(!xd->mode_info_context->mbmi.encoded_as_sb);
#endif

  vp8_setup_interp_filters(xd, mbmi->interp_filter, cm);
  if (cpi->oxcf.tuning == VP8_TUNE_SSIM) {
    // Adjust the zbin based on this MB rate.
    adjust_act_zbin(cpi, x);
  }

  {
    // Experimental code. Special case for gf and arf zeromv modes.
    // Increase zbin size to suppress noise
    cpi->zbin_mode_boost = 0;
    if (cpi->zbin_mode_boost_enabled) {
      if (mbmi->ref_frame != INTRA_FRAME) {
        if (mbmi->mode == ZEROMV) {
          if (mbmi->ref_frame != LAST_FRAME)
            cpi->zbin_mode_boost = GF_ZEROMV_ZBIN_BOOST;
          else
            cpi->zbin_mode_boost = LF_ZEROMV_ZBIN_BOOST;
        } else if (mbmi->mode == SPLITMV)
          cpi->zbin_mode_boost = 0;
        else
          cpi->zbin_mode_boost = MV_ZBIN_BOOST;