vp9_encodeframe.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 "./vpx_config.h"
#include "vp9/encoder/vp9_encodeframe.h"
#include "vp9/encoder/vp9_encodemb.h"
#include "vp9/encoder/vp9_encodemv.h"
#include "vp9/common/vp9_common.h"
#include "vp9/encoder/vp9_onyx_int.h"
#include "vp9/common/vp9_extend.h"
#include "vp9/common/vp9_entropy.h"
#include "vp9/common/vp9_entropymode.h"
#include "vp9/common/vp9_quant_common.h"
#include "vp9/encoder/vp9_segmentation.h"
#include "vp9/common/vp9_setupintrarecon.h"
#include "vp9/encoder/vp9_encodeintra.h"
#include "vp9/common/vp9_reconinter.h"
#include "vp9/common/vp9_invtrans.h"
#include "vp9/encoder/vp9_rdopt.h"
#include "vp9/common/vp9_findnearmv.h"
#include "vp9/common/vp9_reconintra.h"
#include "vp9/common/vp9_seg_common.h"
#include "vp9/common/vp9_tile_common.h"
#include "vp9/encoder/vp9_tokenize.h"
#include "./vp9_rtcd.h"
#include <stdio.h>
#include <math.h>
#include <limits.h>
#include "vpx_ports/vpx_timer.h"
#include "vp9/common/vp9_pred_common.h"
#include "vp9/common/vp9_mvref_common.h"

#define DBG_PRNT_SEGMAP 0

// #define ENC_DEBUG
#ifdef ENC_DEBUG
int enc_debug = 0;
#endif

void vp9_select_interp_filter_type(VP9_COMP *cpi);

static void encode_macroblock(VP9_COMP *cpi, TOKENEXTRA **t,
                              int output_enabled, int mb_row, int mb_col);

static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t,
                              int output_enabled, int mb_row, int mb_col,
                              BLOCK_SIZE_TYPE bsize);

static void adjust_act_zbin(VP9_COMP *cpi, MACROBLOCK *x);

#ifdef MODE_STATS
unsigned int inter_y_modes[MB_MODE_COUNT];
unsigned int inter_uv_modes[VP9_UV_MODES];
unsigned int inter_b_modes[B_MODE_COUNT];
unsigned int y_modes[VP9_YMODES];
unsigned int i8x8_modes[VP9_I8X8_MODES];
unsigned int uv_modes[VP9_UV_MODES];
unsigned int uv_modes_y[VP9_YMODES][VP9_UV_MODES];
unsigned int b_modes[B_MODE_COUNT];
#endif


/* activity_avg must be positive, or flat regions could get a zero weight
 *  (infinite lambda), which confounds analysis.
 * This also avoids the need for divide by zero checks in
 *  vp9_activity_masking().
 */
#define VP9_ACTIVITY_AVG_MIN (64)

/* This is used as a reference when computing the source variance for the
 *  purposes of activity masking.
 * Eventually this should be replaced by custom no-reference routines,
 *  which will be faster.
 */
static const uint8_t VP9_VAR_OFFS[16] = {
  128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128, 128
};


// Original activity measure from Tim T's code.
static unsigned int tt_activity_measure(VP9_COMP *cpi, MACROBLOCK *x) {
  unsigned int act;
  unsigned int sse;
  /* TODO: This could also be done over smaller areas (8x8), but that would
   *  require extensive changes elsewhere, as lambda is assumed to be fixed
   *  over an entire MB in most of the code.
   * Another option is to compute four 8x8 variances, and pick a single
   *  lambda using a non-linear combination (e.g., the smallest, or second
   *  smallest, etc.).
   */
  act = vp9_variance16x16(x->src.y_buffer, x->src.y_stride, VP9_VAR_OFFS, 0,
                          &sse);
  act <<= 4;

  /* If the region is flat, lower the activity some more. */
  if (act < 8 << 12)
    act = act < 5 << 12 ? act : 5 << 12;

  return act;
}

// Stub for alternative experimental activity measures.
static unsigned int alt_activity_measure(VP9_COMP *cpi,
                                         MACROBLOCK *x, int use_dc_pred) {
  return vp9_encode_intra(cpi, x, use_dc_pred);
}


// Measure the activity of the current macroblock
// What we measure here is TBD so abstracted to this function
#define ALT_ACT_MEASURE 1
static unsigned int mb_activity_measure(VP9_COMP *cpi, MACROBLOCK *x,
                                        int mb_row, int mb_col) {
  unsigned int mb_activity;

  if (ALT_ACT_MEASURE) {
    int use_dc_pred = (mb_col || mb_row) && (!mb_col || !mb_row);

    // Or use and alternative.
    mb_activity = alt_activity_measure(cpi, x, use_dc_pred);
  } else {
    // Original activity measure from Tim T's code.
    mb_activity = tt_activity_measure(cpi, x);
  }

  if (mb_activity < VP9_ACTIVITY_AVG_MIN)
    mb_activity = VP9_ACTIVITY_AVG_MIN;

  return mb_activity;
}

// Calculate an "average" mb activity value for the frame
#define ACT_MEDIAN 0
static void calc_av_activity(VP9_COMP *cpi, int64_t activity_sum) {
#if ACT_MEDIAN
  // Find median: Simple n^2 algorithm for experimentation
  {
    unsigned int median;
    unsigned int i, j;
    unsigned int *sortlist;
    unsigned int tmp;

    // Create a list to sort to
    CHECK_MEM_ERROR(sortlist,
    vpx_calloc(sizeof(unsigned int),
    cpi->common.MBs));

    // Copy map to sort list
    vpx_memcpy(sortlist, cpi->mb_activity_map,
    sizeof(unsigned int) * cpi->common.MBs);


    // Ripple each value down to its correct position
    for (i = 1; i < cpi->common.MBs; i ++) {
      for (j = i; j > 0; j --) {
        if (sortlist[j] < sortlist[j - 1]) {
          // Swap values
          tmp = sortlist[j - 1];
          sortlist[j - 1] = sortlist[j];
          sortlist[j] = tmp;
        } else
          break;
      }
    }

    // Even number MBs so estimate median as mean of two either side.
    median = (1 + sortlist[cpi->common.MBs >> 1] +
              sortlist[(cpi->common.MBs >> 1) + 1]) >> 1;

    cpi->activity_avg = median;

    vpx_free(sortlist);
  }
#else
  // Simple mean for now
  cpi->activity_avg = (unsigned int)(activity_sum / cpi->common.MBs);
#endif

  if (cpi->activity_avg < VP9_ACTIVITY_AVG_MIN)
    cpi->activity_avg = VP9_ACTIVITY_AVG_MIN;

  // Experimental code: return fixed value normalized for several clips
  if (ALT_ACT_MEASURE)
    cpi->activity_avg = 100000;
}

#define USE_ACT_INDEX   0
#define OUTPUT_NORM_ACT_STATS   0

#if USE_ACT_INDEX
// Calculate an activity index for each mb
static void calc_activity_index(VP9_COMP *cpi, MACROBLOCK *x) {
  VP9_COMMON *const cm = &cpi->common;
  int mb_row, mb_col;

  int64_t act;
  int64_t a;
  int64_t b;

#if OUTPUT_NORM_ACT_STATS
  FILE *f = fopen("norm_act.stt", "a");
  fprintf(f, "\n%12d\n", cpi->activity_avg);
#endif

  // Reset pointers to start of activity map
  x->mb_activity_ptr = cpi->mb_activity_map;

  // Calculate normalized mb activity number.
  for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
    // for each macroblock col in image
    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
      // Read activity from the map
      act = *(x->mb_activity_ptr);

      // Calculate a normalized activity number
      a = act + 4 * cpi->activity_avg;
      b = 4 * act + cpi->activity_avg;

      if (b >= a)
        *(x->activity_ptr) = (int)((b + (a >> 1)) / a) - 1;
      else
        *(x->activity_ptr) = 1 - (int)((a + (b >> 1)) / b);

#if OUTPUT_NORM_ACT_STATS
      fprintf(f, " %6d", *(x->mb_activity_ptr));
#endif
      // Increment activity map pointers
      x->mb_activity_ptr++;
    }

#if OUTPUT_NORM_ACT_STATS
    fprintf(f, "\n");
#endif

  }

#if OUTPUT_NORM_ACT_STATS
  fclose(f);
#endif

}
#endif

// Loop through all MBs. Note activity of each, average activity and
// calculate a normalized activity for each
static void build_activity_map(VP9_COMP *cpi) {
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *xd = &x->e_mbd;
  VP9_COMMON *const cm = &cpi->common;

#if ALT_ACT_MEASURE
  YV12_BUFFER_CONFIG *new_yv12 = &cm->yv12_fb[cm->new_fb_idx];
  int recon_yoffset;
  int recon_y_stride = new_yv12->y_stride;
#endif

  int mb_row, mb_col;
  unsigned int mb_activity;
  int64_t activity_sum = 0;

  x->mb_activity_ptr = cpi->mb_activity_map;

  // for each macroblock row in image
  for (mb_row = 0; mb_row < cm->mb_rows; mb_row++) {
#if ALT_ACT_MEASURE
    // reset above block coeffs
    xd->up_available = (mb_row != 0);
    recon_yoffset = (mb_row * recon_y_stride * 16);
#endif
    // for each macroblock col in image
    for (mb_col = 0; mb_col < cm->mb_cols; mb_col++) {
#if ALT_ACT_MEASURE
      xd->plane[0].dst.buf = new_yv12->y_buffer + recon_yoffset;
      xd->left_available = (mb_col != 0);
      recon_yoffset += 16;
#endif

      // measure activity
      mb_activity = mb_activity_measure(cpi, x, mb_row, mb_col);

      // Keep frame sum
      activity_sum += mb_activity;

      // Store MB level activity details.
      *x->mb_activity_ptr = mb_activity;

      // Increment activity map pointer
      x->mb_activity_ptr++;

      // adjust to the next column of source macroblocks
      x->src.y_buffer += 16;
    }


    // adjust to the next row of mbs
    x->src.y_buffer += 16 * x->src.y_stride - 16 * cm->mb_cols;

#if ALT_ACT_MEASURE
    // extend the recon for intra prediction
    vp9_extend_mb_row(new_yv12, xd->plane[0].dst.buf + 16,
                      xd->plane[1].dst.buf + 8, xd->plane[2].dst.buf + 8);
#endif

  }

  // Calculate an "average" MB activity
  calc_av_activity(cpi, activity_sum);

#if USE_ACT_INDEX
  // Calculate an activity index number of each mb
  calc_activity_index(cpi, x);
#endif

}

// Macroblock activity masking
void vp9_activity_masking(VP9_COMP *cpi, MACROBLOCK *x) {
#if USE_ACT_INDEX
  x->rdmult += *(x->mb_activity_ptr) * (x->rdmult >> 2);
  x->errorperbit = x->rdmult * 100 / (110 * x->rddiv);
  x->errorperbit += (x->errorperbit == 0);
#else
  int64_t a;
  int64_t b;
  int64_t act = *(x->mb_activity_ptr);

  // Apply the masking to the RD multiplier.
  a = act + (2 * cpi->activity_avg);
  b = (2 * act) + cpi->activity_avg;

  x->rdmult = (unsigned int)(((int64_t)x->rdmult * b + (a >> 1)) / a);
  x->errorperbit = x->rdmult * 100 / (110 * x->rddiv);
  x->errorperbit += (x->errorperbit == 0);
#endif

  // Activity based Zbin adjustment
  adjust_act_zbin(cpi, x);
}

static void update_state(VP9_COMP *cpi,
                         PICK_MODE_CONTEXT *ctx,
                         BLOCK_SIZE_TYPE bsize,
                         int output_enabled) {
  int i, x_idx, y;
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  MODE_INFO *mi = &ctx->mic;
  MB_MODE_INFO *const mbmi = &xd->mode_info_context->mbmi;
  int mb_mode = mi->mbmi.mode;
  int mb_mode_index = ctx->best_mode_index;
  const int mis = cpi->common.mode_info_stride;
  const int bh = 1 << mb_height_log2(bsize), bw = 1 << mb_width_log2(bsize);

#if CONFIG_DEBUG
  assert(mb_mode < MB_MODE_COUNT);
  assert(mb_mode_index < MAX_MODES);
  assert(mi->mbmi.ref_frame < MAX_REF_FRAMES);
#endif
  assert(mi->mbmi.sb_type == bsize);

  // Restore the coding context of the MB to that that was in place
  // when the mode was picked for it
  for (y = 0; y < bh; y++) {
    for (x_idx = 0; x_idx < bw; x_idx++) {
      if ((xd->mb_to_right_edge >> 7) + bw > x_idx &&
          (xd->mb_to_bottom_edge >> 7) + bh > y) {
        MODE_INFO *mi_addr = xd->mode_info_context + x_idx + y * mis;

        vpx_memcpy(mi_addr, mi, sizeof(MODE_INFO));
      }
    }
  }
  if (bsize < BLOCK_SIZE_SB32X32) {
    ctx->txfm_rd_diff[ALLOW_32X32] = ctx->txfm_rd_diff[ALLOW_16X16];
  }

  if (mb_mode == I4X4_PRED) {
    for (i = 0; i < 16; i++) {
      xd->block[i].bmi.as_mode = xd->mode_info_context->bmi[i].as_mode;
      assert(xd->block[i].bmi.as_mode.first < B_MODE_COUNT);
    }
  } else if (mb_mode == I8X8_PRED) {
    for (i = 0; i < 16; i++) {
      xd->block[i].bmi = xd->mode_info_context->bmi[i];
    }
  } else if (mb_mode == SPLITMV) {
    vpx_memcpy(x->partition_info, &ctx->partition_info,
               sizeof(PARTITION_INFO));

    mbmi->mv[0].as_int = x->partition_info->bmi[15].mv.as_int;
    mbmi->mv[1].as_int = x->partition_info->bmi[15].second_mv.as_int;
  }

  x->skip = ctx->skip;
  if (!output_enabled)
    return;

  {
    int segment_id = mbmi->segment_id, ref_pred_flag;
    if (!vp9_segfeature_active(xd, segment_id, SEG_LVL_SKIP)) {
      for (i = 0; i < NB_TXFM_MODES; i++) {
        cpi->rd_tx_select_diff[i] += ctx->txfm_rd_diff[i];
      }
    }

    // Did the chosen reference frame match its predicted value.
    ref_pred_flag = ((xd->mode_info_context->mbmi.ref_frame ==
                      vp9_get_pred_ref(cm, xd)));
    vp9_set_pred_flag(xd, PRED_REF, ref_pred_flag);
    if (!xd->segmentation_enabled ||
        !vp9_segfeature_active(xd, segment_id, SEG_LVL_REF_FRAME) ||
        vp9_check_segref(xd, segment_id, INTRA_FRAME)  +
        vp9_check_segref(xd, segment_id, LAST_FRAME)   +
        vp9_check_segref(xd, segment_id, GOLDEN_FRAME) +
        vp9_check_segref(xd, segment_id, ALTREF_FRAME) > 1) {
      // Get the prediction context and status
      int pred_context = vp9_get_pred_context(cm, xd, PRED_REF);

      // Count prediction success
      cpi->ref_pred_count[pred_context][ref_pred_flag]++;
    }
  }

  if (cpi->common.frame_type == KEY_FRAME) {
    // Restore the coding modes to that held in the coding context
    // if (mb_mode == I4X4_PRED)
    //    for (i = 0; i < 16; i++)
    //    {
    //        xd->block[i].bmi.as_mode =
    //                          xd->mode_info_context->bmi[i].as_mode;
    //        assert(xd->mode_info_context->bmi[i].as_mode < MB_MODE_COUNT);
    //    }
#if CONFIG_INTERNAL_STATS
    static const int kf_mode_index[] = {
      THR_DC /*DC_PRED*/,
      THR_V_PRED /*V_PRED*/,
      THR_H_PRED /*H_PRED*/,
      THR_D45_PRED /*D45_PRED*/,
      THR_D135_PRED /*D135_PRED*/,
      THR_D117_PRED /*D117_PRED*/,
      THR_D153_PRED /*D153_PRED*/,
      THR_D27_PRED /*D27_PRED*/,
      THR_D63_PRED /*D63_PRED*/,
      THR_TM /*TM_PRED*/,
      THR_I8X8_PRED /*I8X8_PRED*/,
      THR_B_PRED /*I4X4_PRED*/,
    };
    cpi->mode_chosen_counts[kf_mode_index[mb_mode]]++;
#endif
  } else {
    /*
            // Reduce the activation RD thresholds for the best choice mode
            if ((cpi->rd_baseline_thresh[mb_mode_index] > 0) &&
                (cpi->rd_baseline_thresh[mb_mode_index] < (INT_MAX >> 2)))
            {
                int best_adjustment = (cpi->rd_thresh_mult[mb_mode_index] >> 2);

                cpi->rd_thresh_mult[mb_mode_index] =
                        (cpi->rd_thresh_mult[mb_mode_index]
                         >= (MIN_THRESHMULT + best_adjustment)) ?
                                cpi->rd_thresh_mult[mb_mode_index] - best_adjustment :
                                MIN_THRESHMULT;
                cpi->rd_threshes[mb_mode_index] =
                        (cpi->rd_baseline_thresh[mb_mode_index] >> 7)
                        * cpi->rd_thresh_mult[mb_mode_index];

            }
    */
    // Note how often each mode chosen as best
    cpi->mode_chosen_counts[mb_mode_index]++;
    if (mbmi->mode == SPLITMV || mbmi->mode == NEWMV) {
      int_mv best_mv, best_second_mv;
      MV_REFERENCE_FRAME rf = mbmi->ref_frame;
      best_mv.as_int = ctx->best_ref_mv.as_int;
      best_second_mv.as_int = ctx->second_best_ref_mv.as_int;
      if (mbmi->mode == NEWMV) {
        best_mv.as_int = mbmi->ref_mvs[rf][0].as_int;
        best_second_mv.as_int = mbmi->ref_mvs[mbmi->second_ref_frame][0].as_int;
      }
      mbmi->best_mv.as_int = best_mv.as_int;
      mbmi->best_second_mv.as_int = best_second_mv.as_int;
      vp9_update_nmv_count(cpi, x, &best_mv, &best_second_mv);
    }
#if CONFIG_COMP_INTERINTRA_PRED
    if (mbmi->mode >= NEARESTMV && mbmi->mode < SPLITMV &&
        mbmi->second_ref_frame <= INTRA_FRAME) {
      if (mbmi->second_ref_frame == INTRA_FRAME) {
        ++cpi->interintra_count[1];
        ++cpi->ymode_count[mbmi->interintra_mode];
#if SEPARATE_INTERINTRA_UV
        ++cpi->y_uv_mode_count[mbmi->interintra_mode][mbmi->interintra_uv_mode];
#endif
      } else {
        ++cpi->interintra_count[0];
      }
    }
#endif
    if (cpi->common.mcomp_filter_type == SWITCHABLE &&
        mbmi->mode >= NEARESTMV &&
        mbmi->mode <= SPLITMV) {
      ++cpi->switchable_interp_count
          [vp9_get_pred_context(&cpi->common, xd, PRED_SWITCHABLE_INTERP)]
          [vp9_switchable_interp_map[mbmi->interp_filter]];
    }

    cpi->rd_comp_pred_diff[SINGLE_PREDICTION_ONLY] += ctx->single_pred_diff;
    cpi->rd_comp_pred_diff[COMP_PREDICTION_ONLY]   += ctx->comp_pred_diff;
    cpi->rd_comp_pred_diff[HYBRID_PREDICTION]      += ctx->hybrid_pred_diff;
  }
}

static unsigned find_seg_id(uint8_t *buf, BLOCK_SIZE_TYPE bsize,
                            int start_y, int height, int start_x, int width) {
  const int bw = 1 << mb_width_log2(bsize), bh = 1 << mb_height_log2(bsize);
  const int end_x = MIN(start_x + bw, width);
  const int end_y = MIN(start_y + bh, height);
  int x, y;
  unsigned seg_id = -1;

  buf += width * start_y;
  for (y = start_y; y < end_y; y++, buf += width) {
    for (x = start_x; x < end_x; x++) {
      seg_id = MIN(seg_id, buf[x]);
    }
  }

  return seg_id;
}

static void set_offsets(VP9_COMP *cpi,
                        int mb_row, int mb_col, BLOCK_SIZE_TYPE bsize) {
  MACROBLOCK *const x = &cpi->mb;
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *mbmi;
  const int dst_fb_idx = cm->new_fb_idx;
  const int idx_map = mb_row * cm->mb_cols + mb_col;
  const int idx_str = xd->mode_info_stride * mb_row + mb_col;
  const int bw = 1 << mb_width_log2(bsize), bh = 1 << mb_height_log2(bsize);

  // entropy context structures
  xd->above_context = cm->above_context + mb_col;
  xd->left_context  = cm->left_context + (mb_row & 3);

  // Activity map pointer
  x->mb_activity_ptr = &cpi->mb_activity_map[idx_map];
  x->active_ptr = cpi->active_map + idx_map;

  /* pointers to mode info contexts */
  x->partition_info          = x->pi + idx_str;
  xd->mode_info_context      = cm->mi + idx_str;
  mbmi = &xd->mode_info_context->mbmi;
  xd->prev_mode_info_context = cm->prev_mi + idx_str;

  // Set up destination pointers
  setup_dst_planes(xd, &cm->yv12_fb[dst_fb_idx], mb_row, mb_col);

  /* Set up limit values for MV components to prevent them from
   * extending beyond the UMV borders assuming 16x16 block size */
  x->mv_row_min = -((mb_row * 16) + VP9BORDERINPIXELS - VP9_INTERP_EXTEND);
  x->mv_col_min = -((mb_col * 16) + VP9BORDERINPIXELS - VP9_INTERP_EXTEND);
  x->mv_row_max = ((cm->mb_rows - mb_row) * 16 +
                   (VP9BORDERINPIXELS - 16 * bh - VP9_INTERP_EXTEND));
  x->mv_col_max = ((cm->mb_cols - mb_col) * 16 +
                   (VP9BORDERINPIXELS - 16 * bw - VP9_INTERP_EXTEND));

  // Set up distance of MB to edge of frame in 1/8th pel units
  assert(!(mb_col & (bw - 1)) && !(mb_row & (bh - 1)));
  set_mb_row(cm, xd, mb_row, bh);
  set_mb_col(cm, xd, mb_col, bw);

  /* set up source buffers */
  setup_pred_block(&x->src, cpi->Source, mb_row, mb_col, NULL, NULL);

  /* R/D setup */
  x->rddiv = cpi->RDDIV;
  x->rdmult = cpi->RDMULT;

  /* segment ID */
  if (xd->segmentation_enabled) {
    if (xd->update_mb_segmentation_map) {
      mbmi->segment_id = find_seg_id(cpi->segmentation_map, bsize,
                                     mb_row, cm->mb_rows, mb_col, cm->mb_cols);
    } else {
      mbmi->segment_id = find_seg_id(cm->last_frame_seg_map, bsize,
                                     mb_row, cm->mb_rows, mb_col, cm->mb_cols);
    }
    assert(mbmi->segment_id <= (MAX_MB_SEGMENTS-1));
    vp9_mb_init_quantizer(cpi, x);

    if (xd->segmentation_enabled && cpi->seg0_cnt > 0 &&
        !vp9_segfeature_active(xd, 0, SEG_LVL_REF_FRAME) &&
        vp9_segfeature_active(xd, 1, SEG_LVL_REF_FRAME) &&
        vp9_check_segref(xd, 1, INTRA_FRAME)  +
        vp9_check_segref(xd, 1, LAST_FRAME)   +
        vp9_check_segref(xd, 1, GOLDEN_FRAME) +
        vp9_check_segref(xd, 1, ALTREF_FRAME) == 1) {
      cpi->seg0_progress = (cpi->seg0_idx << 16) / cpi->seg0_cnt;
    } else {
      const int y = mb_row & ~3;
      const int x = mb_col & ~3;
      const int p16 = ((mb_row & 1) << 1) +  (mb_col & 1);
      const int p32 = ((mb_row & 2) << 2) + ((mb_col & 2) << 1);
      const int tile_progress = cm->cur_tile_mb_col_start * cm->mb_rows;
      const int mb_cols = cm->cur_tile_mb_col_end - cm->cur_tile_mb_col_start;

      cpi->seg0_progress =
          ((y * mb_cols + x * 4 + p32 + p16 + tile_progress) << 16) / cm->MBs;
    }
  } else {
    mbmi->segment_id = 0;
  }
}

static int pick_mb_mode(VP9_COMP *cpi,
                        int mb_row,
                        int mb_col,
                        TOKENEXTRA **tp,
                        int *totalrate,
                        int *totaldist) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  int splitmodes_used = 0;
  MB_MODE_INFO *mbmi;

  set_offsets(cpi, mb_row, mb_col, BLOCK_SIZE_MB16X16);

  if (cpi->oxcf.tuning == VP8_TUNE_SSIM)
    vp9_activity_masking(cpi, x);

  mbmi = &xd->mode_info_context->mbmi;
  mbmi->sb_type = BLOCK_SIZE_MB16X16;

  // Find best coding mode & reconstruct the MB so it is available
  // as a predictor for MBs that follow in the SB
  if (cm->frame_type == KEY_FRAME) {
    vp9_rd_pick_intra_mode(cpi, x, totalrate, totaldist);

    // Save the coding context
    vpx_memcpy(&x->mb_context[xd->sb_index][xd->mb_index].mic,
               xd->mode_info_context, sizeof(MODE_INFO));
  } else {
    vp9_pick_mode_inter_macroblock(cpi, x, mb_row, mb_col,
                                   totalrate, totaldist);
    splitmodes_used += (mbmi->mode == SPLITMV);

    if (cpi->mb.e_mbd.segmentation_enabled && mbmi->segment_id == 0) {
      cpi->seg0_idx++;
    }
  }

  return splitmodes_used;
}

static void pick_sb_modes(VP9_COMP *cpi, int mb_row, int mb_col,
                          TOKENEXTRA **tp, int *totalrate, int *totaldist,
                          BLOCK_SIZE_TYPE bsize, PICK_MODE_CONTEXT *ctx) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;

  set_offsets(cpi, mb_row, mb_col, bsize);
  xd->mode_info_context->mbmi.sb_type = bsize;
  if (cpi->oxcf.tuning == VP8_TUNE_SSIM)
    vp9_activity_masking(cpi, x);

  /* Find best coding mode & reconstruct the MB so it is available
   * as a predictor for MBs that follow in the SB */
  if (cm->frame_type == KEY_FRAME) {
    vp9_rd_pick_intra_mode_sb(cpi, x, totalrate, totaldist, bsize, ctx);
  } else {
    vp9_rd_pick_inter_mode_sb(cpi, x, mb_row, mb_col, totalrate, totaldist,
                              bsize, ctx);
  }
}

static void update_stats(VP9_COMP *cpi, int mb_row, int mb_col) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  MODE_INFO *mi = xd->mode_info_context;
  MB_MODE_INFO *const mbmi = &mi->mbmi;

  if (cm->frame_type == KEY_FRAME) {
#ifdef MODE_STATS
    y_modes[mbmi->mode]++;
#endif
  } else {
    int segment_id, seg_ref_active;

    if (mbmi->ref_frame) {
      int pred_context = vp9_get_pred_context(cm, xd, PRED_COMP);

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

#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 = vp9_segfeature_active(xd, segment_id,
                                           SEG_LVL_REF_FRAME);
    if (!seg_ref_active ||
        ((vp9_check_segref(xd, segment_id, INTRA_FRAME) +
          vp9_check_segref(xd, segment_id, LAST_FRAME) +
          vp9_check_segref(xd, segment_id, GOLDEN_FRAME) +
          vp9_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++;
  }
}

static void encode_sb(VP9_COMP *cpi,
                      int mb_row,
                      int mb_col,
                      int output_enabled,
                      TOKENEXTRA **tp, BLOCK_SIZE_TYPE is_sb) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  BLOCK_SIZE_TYPE bsize = BLOCK_SIZE_SB32X32;

  if (is_sb == BLOCK_SIZE_SB32X32) {
    set_offsets(cpi, mb_row, mb_col, bsize);
    update_state(cpi, &x->sb32_context[xd->sb_index],
                 bsize, output_enabled);

    encode_superblock(cpi, tp,
                      output_enabled, mb_row, mb_col, bsize);
    if (output_enabled) {
      update_stats(cpi, mb_row, mb_col);
      cpi->partition_count[partition_plane(bsize)][PARTITION_NONE]++;

      (*tp)->token = EOSB_TOKEN;
      (*tp)++;
    }
#if CONFIG_SBSEGMENT
  } else if (is_sb == BLOCK_SIZE_SB16X32) {
    int i;

    if (output_enabled)
      cpi->partition_count[partition_plane(bsize)][PARTITION_VERT]++;
    for (i = 0; i < 2 && mb_col + i != cm->mb_cols; i++) {
      set_offsets(cpi, mb_row, mb_col + i, BLOCK_SIZE_SB16X32);
      update_state(cpi, &x->sb16x32_context[xd->sb_index][i],
                   BLOCK_SIZE_SB16X32, output_enabled);
      encode_superblock(cpi, tp,
                        output_enabled, mb_row, mb_col + i, BLOCK_SIZE_SB16X32);
      if (output_enabled) {
        update_stats(cpi, mb_row, mb_col + i);

        (*tp)->token = EOSB_TOKEN;
        (*tp)++;
      }
    }
  } else if (is_sb == BLOCK_SIZE_SB32X16) {
    int i;

    if (output_enabled)
      cpi->partition_count[partition_plane(bsize)][PARTITION_HORZ]++;
    for (i = 0; i < 2 && mb_row + i != cm->mb_rows; i++) {
      set_offsets(cpi, mb_row + i, mb_col, BLOCK_SIZE_SB32X16);
      update_state(cpi, &x->sb32x16_context[xd->sb_index][i],
                   BLOCK_SIZE_SB32X16, output_enabled);
      encode_superblock(cpi, tp,
                        output_enabled, mb_row + i, mb_col, BLOCK_SIZE_SB32X16);
      if (output_enabled) {
        update_stats(cpi, mb_row + i, mb_col);

        (*tp)->token = EOSB_TOKEN;
        (*tp)++;
      }
    }
#endif
  } else {
    int i;
    if (output_enabled)
      cpi->partition_count[partition_plane(bsize)][PARTITION_SPLIT]++;

    for (i = 0; i < 4; i++) {
      const int x_idx = i & 1, y_idx = i >> 1;

      if ((mb_row + y_idx >= cm->mb_rows) || (mb_col + x_idx >= cm->mb_cols)) {
        // MB lies outside frame, move on
        continue;
      }

      set_offsets(cpi, mb_row + y_idx, mb_col + x_idx, BLOCK_SIZE_MB16X16);
      xd->mb_index = i;
      update_state(cpi, &x->mb_context[xd->sb_index][i],
                   BLOCK_SIZE_MB16X16, output_enabled);

      if (cpi->oxcf.tuning == VP8_TUNE_SSIM)
        vp9_activity_masking(cpi, x);

      encode_macroblock(cpi, tp,
                        output_enabled, mb_row + y_idx, mb_col + x_idx);
      if (output_enabled) {
        update_stats(cpi, mb_row + y_idx, mb_col + x_idx);

        (*tp)->token = EOSB_TOKEN;
        (*tp)++;
      }
    }
  }

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

static void encode_sb64(VP9_COMP *cpi,
                        int mb_row,
                        int mb_col,
                        TOKENEXTRA **tp, BLOCK_SIZE_TYPE is_sb[4]) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  BLOCK_SIZE_TYPE bsize = BLOCK_SIZE_SB64X64;

  if (is_sb[0] == BLOCK_SIZE_SB64X64) {
    set_offsets(cpi, mb_row, mb_col, bsize);
    update_state(cpi, &x->sb64_context, bsize, 1);
    encode_superblock(cpi, tp,
                      1, mb_row, mb_col, bsize);
    update_stats(cpi, mb_row, mb_col);

    (*tp)->token = EOSB_TOKEN;
    (*tp)++;
    cpi->partition_count[partition_plane(bsize)][PARTITION_NONE]++;
#if CONFIG_SBSEGMENT
  } else if (is_sb[0] == BLOCK_SIZE_SB32X64) {
    int i;

    cpi->partition_count[partition_plane(bsize)][PARTITION_VERT]++;
    for (i = 0; i < 2 && mb_col + i * 2 != cm->mb_cols; i++) {
      set_offsets(cpi, mb_row, mb_col + i * 2, BLOCK_SIZE_SB32X64);
      update_state(cpi, &x->sb32x64_context[i], BLOCK_SIZE_SB32X64, 1);
      encode_superblock(cpi, tp,
                        1, mb_row, mb_col + i * 2, BLOCK_SIZE_SB32X64);
      update_stats(cpi, mb_row, mb_col + i * 2);

      (*tp)->token = EOSB_TOKEN;
      (*tp)++;
    }
  } else if (is_sb[0] == BLOCK_SIZE_SB64X32) {
    int i;

    cpi->partition_count[partition_plane(bsize)][PARTITION_HORZ]++;
    for (i = 0; i < 2 && mb_row + i * 2 != cm->mb_rows; i++) {
      set_offsets(cpi, mb_row + i * 2, mb_col, BLOCK_SIZE_SB64X32);
      update_state(cpi, &x->sb64x32_context[i], BLOCK_SIZE_SB64X32, 1);
      encode_superblock(cpi, tp,
                        1, mb_row + i * 2, mb_col, BLOCK_SIZE_SB64X32);
      update_stats(cpi, mb_row + i * 2, mb_col);

      (*tp)->token = EOSB_TOKEN;
      (*tp)++;
    }
#endif
  } else {
    int i;
    cpi->partition_count[partition_plane(bsize)][PARTITION_SPLIT]++;
    for (i = 0; i < 4; i++) {
      const int x_idx = i & 1, y_idx = i >> 1;

      if (mb_row + y_idx * 2 >= cm->mb_rows ||
          mb_col + x_idx * 2 >= cm->mb_cols) {
        // MB lies outside frame, move on
        continue;
      }
      xd->sb_index = i;
      encode_sb(cpi, mb_row + 2 * y_idx, mb_col + 2 * x_idx, 1, tp,
                is_sb[i]);
    }
  }
}

static void encode_sb_row(VP9_COMP *cpi,
                          int mb_row,
                          TOKENEXTRA **tp,
                          int *totalrate) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  int mb_col;

  // 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 = cm->cur_tile_mb_col_start;
       mb_col < cm->cur_tile_mb_col_end; mb_col += 4) {
    int i;
    BLOCK_SIZE_TYPE sb_partitioning[4];
    int sb64_rate = 0, sb64_dist = 0;
    int sb64_skip = 0;
    ENTROPY_CONTEXT_PLANES l[4], a[4];
    TOKENEXTRA *tp_orig = *tp;

    memcpy(&a, cm->above_context + mb_col, sizeof(a));
    memcpy(&l, cm->left_context, sizeof(l));
    for (i = 0; i < 4; i++) {
      const int x_idx = (i & 1) << 1, y_idx = i & 2;
      int sb32_rate = 0, sb32_dist = 0;
      int splitmodes_used = 0;
      int sb32_skip = 0;
      int j;
      ENTROPY_CONTEXT_PLANES l2[2], a2[2];

      if (mb_row + y_idx >= cm->mb_rows || mb_col + x_idx >= cm->mb_cols)
        continue;

      xd->sb_index = i;

      /* Function should not modify L & A contexts; save and restore on exit */
      vpx_memcpy(l2, cm->left_context + y_idx, sizeof(l2));
      vpx_memcpy(a2, cm->above_context + mb_col + x_idx, sizeof(a2));

      /* Encode MBs in raster order within the SB */
      sb_partitioning[i] = BLOCK_SIZE_MB16X16;
      for (j = 0; j < 4; j++) {
        const int x_idx_m = x_idx + (j & 1), y_idx_m = y_idx + (j >> 1);
        int r, d;

        if (mb_row + y_idx_m >= cm->mb_rows ||
            mb_col + x_idx_m >= cm->mb_cols) {
          // MB lies outside frame, move on
          continue;
        }

        // Index of the MB in the SB 0..3
        xd->mb_index = j;

        splitmodes_used += pick_mb_mode(cpi, mb_row + y_idx_m,
                                        mb_col + x_idx_m, tp, &r, &d);
        sb32_rate += r;
        sb32_dist += d;

        // Dummy encode, do not do the tokenization
        encode_macroblock(cpi, tp, 0, mb_row + y_idx_m,
                          mb_col + x_idx_m);
      }

      /* Restore L & A coding context to those in place on entry */
      vpx_memcpy(cm->left_context + y_idx, l2, sizeof(l2));
      vpx_memcpy(cm->above_context + mb_col + x_idx, a2, sizeof(a2));

      sb32_rate += x->partition_cost[partition_plane(BLOCK_SIZE_SB32X32)]
                                    [PARTITION_SPLIT];

      if (cpi->sf.splitmode_breakout) {
        sb32_skip = splitmodes_used;
        sb64_skip += splitmodes_used;
      }

#if CONFIG_SBSEGMENT
      // check 32x16
      if (mb_col + x_idx + 1 < cm->mb_cols) {
        int r, d;

        xd->mb_index = 0;
        pick_sb_modes(cpi, mb_row + y_idx, mb_col + x_idx,
                      tp, &r, &d, BLOCK_SIZE_SB32X16,