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_rtcd.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/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_superblock(VP9_COMP *cpi, TOKENEXTRA **t,
                              int output_enabled, int mi_row, int mi_col,
                              BLOCK_SIZE_TYPE bsize);

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

/* 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->plane[0].src.buf, x->plane[0].src.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);
}

DECLARE_ALIGNED(16, static const uint8_t, vp9_64x64_zeros[64*64]) = { 0 };


// 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->plane[0].src.buf += 16;
    }


    // adjust to the next row of mbs
    x->plane[0].src.buf += 16 * x->plane[0].src.stride - 16 * cm->mb_cols;
  }

  // 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;
  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;
#if CONFIG_DEBUG || CONFIG_INTERNAL_STATS
  MB_PREDICTION_MODE mb_mode = mi->mbmi.mode;
#endif
  int mb_mode_index = ctx->best_mode_index;
  const int mis = cpi->common.mode_info_stride;
  const int bh = 1 << mi_height_log2(bsize), bw = 1 << mi_width_log2(bsize);

#if CONFIG_DEBUG
  assert(mb_mode < MB_MODE_COUNT);
  assert(mb_mode_index < MAX_MODES);
  assert(mi->mbmi.ref_frame[0] < MAX_REF_FRAMES);
  assert(mi->mbmi.ref_frame[1] < 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 >> (3 + LOG2_MI_SIZE)) + bw > x_idx &&
          (xd->mb_to_bottom_edge >> (3 + LOG2_MI_SIZE)) + bh > y) {
        MODE_INFO *mi_addr = xd->mode_info_context + x_idx + y * mis;
        *mi_addr = *mi;
      }
    }
  }
  if (bsize < BLOCK_SIZE_SB32X32) {
    if (bsize < BLOCK_SIZE_MB16X16)
      ctx->txfm_rd_diff[ALLOW_16X16] = ctx->txfm_rd_diff[ALLOW_8X8];
    ctx->txfm_rd_diff[ALLOW_32X32] = ctx->txfm_rd_diff[ALLOW_16X16];
  }

  if (mbmi->ref_frame[0] != INTRA_FRAME && mbmi->sb_type < BLOCK_SIZE_SB8X8) {
    *x->partition_info = ctx->partition_info;
    mbmi->mv[0].as_int = x->partition_info->bmi[3].mv.as_int;
    mbmi->mv[1].as_int = x->partition_info->bmi[3].second_mv.as_int;
  }

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

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

  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_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->ref_frame[0] != INTRA_FRAME &&
        (mbmi->sb_type < BLOCK_SIZE_SB8X8 || mbmi->mode == NEWMV)) {
      int_mv best_mv, best_second_mv;
      const MV_REFERENCE_FRAME rf1 = mbmi->ref_frame[0];
      const MV_REFERENCE_FRAME rf2 = mbmi->ref_frame[1];
      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[rf1][0].as_int;
        best_second_mv.as_int = mbmi->ref_mvs[rf2][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 (bsize > BLOCK_SIZE_SB8X8 && mbmi->mode == NEWMV) {
      int i, j;
      for (j = 0; j < bh; ++j)
        for (i = 0; i < bw; ++i)
          if ((xd->mb_to_right_edge >> (3 + LOG2_MI_SIZE)) + bw > i &&
              (xd->mb_to_bottom_edge >> (3 + LOG2_MI_SIZE)) + bh > j)
            xd->mode_info_context[mis * j + i].mbmi = *mbmi;
    }

    if (cpi->common.mcomp_filter_type == SWITCHABLE &&
        is_inter_mode(mbmi->mode)) {
      ++cpi->common.fc.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 << mi_width_log2(bsize), bh = 1 << mi_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;
}

void vp9_setup_src_planes(MACROBLOCK *x,
                          const YV12_BUFFER_CONFIG *src,
                          int mb_row, int mb_col) {
  uint8_t *buffers[4] = {src->y_buffer, src->u_buffer, src->v_buffer,
                         src->alpha_buffer};
  int strides[4] = {src->y_stride, src->uv_stride, src->uv_stride,
                    src->alpha_stride};
  int i;

  for (i = 0; i < MAX_MB_PLANE; i++) {
    setup_pred_plane(&x->plane[i].src,
                     buffers[i], strides[i],
                     mb_row, mb_col, NULL,
                     x->e_mbd.plane[i].subsampling_x,
                     x->e_mbd.plane[i].subsampling_y);
  }
}

static void set_offsets(VP9_COMP *cpi,
                        int mi_row, int mi_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_str = xd->mode_info_stride * mi_row + mi_col;
  const int bw = 1 << mi_width_log2(bsize), bh = 1 << mi_height_log2(bsize);
  const int mb_row = mi_row >> 1;
  const int mb_col = mi_col >> 1;
  const int idx_map = mb_row * cm->mb_cols + mb_col;
  int i;

  // entropy context structures
  for (i = 0; i < MAX_MB_PLANE; i++) {
    xd->plane[i].above_context = cm->above_context[i] +
        (mi_col * 2 >>  xd->plane[i].subsampling_x);
    xd->plane[i].left_context = cm->left_context[i] +
        (((mi_row * 2) & 15) >> xd->plane[i].subsampling_y);
  }

  // partition contexts
  set_partition_seg_context(cm, xd, mi_row, mi_col);

  // 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;
  // Special case: if prev_mi is NULL, the previous mode info context
  // cannot be used.
  xd->prev_mode_info_context = cm->prev_mi ?
                                 cm->prev_mi + idx_str : NULL;

  // Set up destination pointers
  setup_dst_planes(xd, &cm->yv12_fb[dst_fb_idx], mi_row, mi_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 = -((mi_row * MI_SIZE) + VP9BORDERINPIXELS - VP9_INTERP_EXTEND);
  x->mv_col_min = -((mi_col * MI_SIZE) + VP9BORDERINPIXELS - VP9_INTERP_EXTEND);
  x->mv_row_max = ((cm->mi_rows - mi_row) * MI_SIZE +
                   (VP9BORDERINPIXELS - MI_SIZE * bh - VP9_INTERP_EXTEND));
  x->mv_col_max = ((cm->mi_cols - mi_col) * MI_SIZE +
                   (VP9BORDERINPIXELS - MI_SIZE * bw - VP9_INTERP_EXTEND));

  // Set up distance of MB to edge of frame in 1/8th pel units
  assert(!(mi_col & (bw - 1)) && !(mi_row & (bh - 1)));
  set_mi_row_col(cm, xd, mi_row, bh, mi_col, bw);

  /* set up source buffers */
  vp9_setup_src_planes(x, cpi->Source, mi_row, mi_col);

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

  /* segment ID */
  if (xd->segmentation_enabled) {
    uint8_t *map = xd->update_mb_segmentation_map ? cpi->segmentation_map
                                                  : cm->last_frame_seg_map;
    mbmi->segment_id = find_seg_id(map, bsize, mi_row,
                                   cm->mi_rows, mi_col, cm->mi_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_mi_col_start * cm->mb_rows >> 1;
      const int mb_cols =
          (cm->cur_tile_mi_col_end - cm->cur_tile_mi_col_start) >> 1;

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

static void pick_sb_modes(VP9_COMP *cpi, int mi_row, int mi_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;

  if (bsize < BLOCK_SIZE_SB8X8)
    if (xd->ab_index != 0)
      return;

  set_offsets(cpi, mi_row, mi_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, mi_row, mi_col, totalrate, totaldist,
                              bsize, ctx);
  }
}

static void update_stats(VP9_COMP *cpi, int mi_row, int mi_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) {
    int segment_id, seg_ref_active;

    cpi->intra_inter_count[vp9_get_pred_context(cm, xd, PRED_INTRA_INTER)]
                          [mbmi->ref_frame[0] > INTRA_FRAME]++;

    // 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 (mbmi->ref_frame[0] > INTRA_FRAME &&
        (!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))) {
      if (cm->comp_pred_mode == HYBRID_PREDICTION)
        cpi->comp_inter_count[vp9_get_pred_context(cm, xd,
                                                   PRED_COMP_INTER_INTER)]
                             [mbmi->ref_frame[1] > INTRA_FRAME]++;

      if (mbmi->ref_frame[1] > INTRA_FRAME) {
        cpi->comp_ref_count[vp9_get_pred_context(cm, xd, PRED_COMP_REF_P)]
                           [mbmi->ref_frame[0] == GOLDEN_FRAME]++;
      } else {
        cpi->single_ref_count[vp9_get_pred_context(cm, xd, PRED_SINGLE_REF_P1)]
                             [0][mbmi->ref_frame[0] != LAST_FRAME]++;
        if (mbmi->ref_frame[0] != LAST_FRAME)
          cpi->single_ref_count[vp9_get_pred_context(cm, xd,
                                                     PRED_SINGLE_REF_P2)]
                               [1][mbmi->ref_frame[0] != GOLDEN_FRAME]++;
      }
    }
    // Count of last ref frame 0,0 usage
    if ((mbmi->mode == ZEROMV) && (mbmi->ref_frame[0] == LAST_FRAME))
      cpi->inter_zz_count++;
  }
}

// TODO(jingning): the variables used here are little complicated. need further
// refactoring on organizing the the temporary buffers, when recursive
// partition down to 4x4 block size is enabled.
static PICK_MODE_CONTEXT *get_block_context(MACROBLOCK *x,
                                            BLOCK_SIZE_TYPE bsize) {
  MACROBLOCKD *const xd = &x->e_mbd;

  switch (bsize) {
    case BLOCK_SIZE_SB64X64:
      return &x->sb64_context;
    case BLOCK_SIZE_SB64X32:
      return &x->sb64x32_context[xd->sb_index];
    case BLOCK_SIZE_SB32X64:
      return &x->sb32x64_context[xd->sb_index];
    case BLOCK_SIZE_SB32X32:
      return &x->sb32_context[xd->sb_index];
    case BLOCK_SIZE_SB32X16:
      return &x->sb32x16_context[xd->sb_index][xd->mb_index];
    case BLOCK_SIZE_SB16X32:
      return &x->sb16x32_context[xd->sb_index][xd->mb_index];
    case BLOCK_SIZE_MB16X16:
      return &x->mb_context[xd->sb_index][xd->mb_index];
    case BLOCK_SIZE_SB16X8:
      return &x->sb16x8_context[xd->sb_index][xd->mb_index][xd->b_index];
    case BLOCK_SIZE_SB8X16:
      return &x->sb8x16_context[xd->sb_index][xd->mb_index][xd->b_index];
    case BLOCK_SIZE_SB8X8:
      return &x->sb8x8_context[xd->sb_index][xd->mb_index][xd->b_index];
    case BLOCK_SIZE_SB8X4:
      return &x->sb8x4_context[xd->sb_index][xd->mb_index][xd->b_index];
    case BLOCK_SIZE_SB4X8:
      return &x->sb4x8_context[xd->sb_index][xd->mb_index][xd->b_index];
    case BLOCK_SIZE_AB4X4:
      return &x->ab4x4_context[xd->sb_index][xd->mb_index][xd->b_index];
    default:
      assert(0);
      return NULL;
  }
}

static BLOCK_SIZE_TYPE *get_sb_partitioning(MACROBLOCK *x,
                                            BLOCK_SIZE_TYPE bsize) {
  MACROBLOCKD *xd = &x->e_mbd;
  switch (bsize) {
    case BLOCK_SIZE_SB64X64:
      return &x->sb64_partitioning;
    case BLOCK_SIZE_SB32X32:
      return &x->sb_partitioning[xd->sb_index];
    case BLOCK_SIZE_MB16X16:
      return &x->mb_partitioning[xd->sb_index][xd->mb_index];
    case BLOCK_SIZE_SB8X8:
      return &x->b_partitioning[xd->sb_index][xd->mb_index][xd->b_index];
    default:
      assert(0);
      return NULL;
  }
}

static void restore_context(VP9_COMP *cpi, int mi_row, int mi_col,
                            ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
                            ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
                            PARTITION_CONTEXT sa[8],
                            PARTITION_CONTEXT sl[8],
                            BLOCK_SIZE_TYPE bsize) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  int p;
  int bwl = b_width_log2(bsize), bw = 1 << bwl;
  int bhl = b_height_log2(bsize), bh = 1 << bhl;
  int mwl = mi_width_log2(bsize), mw = 1 << mwl;
  int mhl = mi_height_log2(bsize), mh = 1 << mhl;
  for (p = 0; p < MAX_MB_PLANE; p++) {
    vpx_memcpy(cm->above_context[p] +
               ((mi_col * 2) >> xd->plane[p].subsampling_x),
               a + bw * p,
               sizeof(ENTROPY_CONTEXT) * bw >> xd->plane[p].subsampling_x);
    vpx_memcpy(cm->left_context[p] +
               ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
               l + bh * p,
               sizeof(ENTROPY_CONTEXT) * bh >> xd->plane[p].subsampling_y);
  }
  vpx_memcpy(cm->above_seg_context + mi_col, sa,
             sizeof(PARTITION_CONTEXT) * mw);
  vpx_memcpy(cm->left_seg_context + (mi_row & MI_MASK), sl,
             sizeof(PARTITION_CONTEXT) * mh);
}
static void save_context(VP9_COMP *cpi, int mi_row, int mi_col,
                          ENTROPY_CONTEXT a[16 * MAX_MB_PLANE],
                          ENTROPY_CONTEXT l[16 * MAX_MB_PLANE],
                          PARTITION_CONTEXT sa[8],
                          PARTITION_CONTEXT sl[8],
                          BLOCK_SIZE_TYPE bsize) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  int p;
  int bwl = b_width_log2(bsize), bw = 1 << bwl;
  int bhl = b_height_log2(bsize), bh = 1 << bhl;
  int mwl = mi_width_log2(bsize), mw = 1 << mwl;
  int mhl = mi_height_log2(bsize), mh = 1 << mhl;

  // buffer the above/left context information of the block in search.
  for (p = 0; p < MAX_MB_PLANE; ++p) {
    vpx_memcpy(a + bw * p, cm->above_context[p] +
               (mi_col * 2 >> xd->plane[p].subsampling_x),
               sizeof(ENTROPY_CONTEXT) * bw >> xd->plane[p].subsampling_x);
    vpx_memcpy(l + bh * p, cm->left_context[p] +
               ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
               sizeof(ENTROPY_CONTEXT) * bh >> xd->plane[p].subsampling_y);
  }
  vpx_memcpy(sa, cm->above_seg_context + mi_col,
             sizeof(PARTITION_CONTEXT) * mw);
  vpx_memcpy(sl, cm->left_seg_context + (mi_row & MI_MASK),
             sizeof(PARTITION_CONTEXT) * mh);
}

static void encode_b(VP9_COMP *cpi, TOKENEXTRA **tp,
                     int mi_row, int mi_col, int output_enabled,
                     BLOCK_SIZE_TYPE bsize, int sub_index) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;

  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
    return;

  if (sub_index != -1)
    *(get_sb_index(xd, bsize)) = sub_index;

  if (bsize < BLOCK_SIZE_SB8X8)
    if (xd->ab_index > 0)
      return;
  set_offsets(cpi, mi_row, mi_col, bsize);
  update_state(cpi, get_block_context(x, bsize), bsize, output_enabled);
  encode_superblock(cpi, tp, output_enabled, mi_row, mi_col, bsize);

  if (output_enabled) {
    update_stats(cpi, mi_row, mi_col);

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

static void encode_sb(VP9_COMP *cpi, TOKENEXTRA **tp,
                      int mi_row, int mi_col, int output_enabled,
                      BLOCK_SIZE_TYPE bsize) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  BLOCK_SIZE_TYPE c1 = BLOCK_SIZE_SB8X8;
  const int bsl = b_width_log2(bsize), bs = (1 << bsl) / 4;
  int bwl, bhl;
  int UNINITIALIZED_IS_SAFE(pl);

  if (mi_row >= cm->mi_rows || mi_col >= cm->mi_cols)
    return;

  c1 = BLOCK_SIZE_AB4X4;
  if (bsize >= BLOCK_SIZE_SB8X8) {
    set_partition_seg_context(cm, xd, mi_row, mi_col);
    pl = partition_plane_context(xd, bsize);
    c1 = *(get_sb_partitioning(x, bsize));
  }

  bwl = b_width_log2(c1), bhl = b_height_log2(c1);

  if (bsl == bwl && bsl == bhl) {
    if (output_enabled && bsize >= BLOCK_SIZE_SB8X8)
        cpi->partition_count[pl][PARTITION_NONE]++;
    encode_b(cpi, tp, mi_row, mi_col, output_enabled, c1, -1);
  } else if (bsl == bhl && bsl > bwl) {
    if (output_enabled)
      cpi->partition_count[pl][PARTITION_VERT]++;
    encode_b(cpi, tp, mi_row, mi_col,      output_enabled, c1, 0);
    encode_b(cpi, tp, mi_row, mi_col + bs, output_enabled, c1, 1);
  } else if (bsl == bwl && bsl > bhl) {
    if (output_enabled)
      cpi->partition_count[pl][PARTITION_HORZ]++;
    encode_b(cpi, tp, mi_row,      mi_col, output_enabled, c1, 0);
    encode_b(cpi, tp, mi_row + bs, mi_col, output_enabled, c1, 1);
  } else {
    BLOCK_SIZE_TYPE subsize;
    int i;

    assert(bwl < bsl && bhl < bsl);
    subsize = get_subsize(bsize, PARTITION_SPLIT);

    if (output_enabled)
      cpi->partition_count[pl][PARTITION_SPLIT]++;

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

      *(get_sb_index(xd, subsize)) = i;
      encode_sb(cpi, tp, mi_row + y_idx * bs, mi_col + x_idx * bs,
                output_enabled, subsize);
    }
  }

  if (bsize >= BLOCK_SIZE_SB8X8 &&
      (bsize == BLOCK_SIZE_SB8X8 || bsl == bwl || bsl == bhl)) {
    set_partition_seg_context(cm, xd, mi_row, mi_col);
    update_partition_context(xd, c1, bsize);
  }
}

static void set_partitioning(VP9_COMP *cpi, MODE_INFO *m,
                             BLOCK_SIZE_TYPE bsize) {
  VP9_COMMON *const cm = &cpi->common;
  const int mis = cm->mode_info_stride;
  int bsl = b_width_log2(bsize);
  int bs = (1 << bsl) / 2;  //
  int block_row, block_col;
  int row, col;

  // this test function sets the entire macroblock to the same bsize
  for (block_row = 0; block_row < 8; block_row += bs) {
    for (block_col = 0; block_col < 8; block_col += bs) {
      for (row = 0; row < bs; row++) {
        for (col = 0; col < bs; col++) {
          m[(block_row+row)*mis + block_col+col].mbmi.sb_type = bsize;
        }
      }
    }
  }
}

static void set_block_size(VP9_COMMON *const cm,
                           MODE_INFO *m, BLOCK_SIZE_TYPE bsize, int mis,
                           int mi_row, int mi_col) {
  int row, col;
  int bwl = b_width_log2(bsize);
  int bhl = b_height_log2(bsize);
  int bsl = (bwl > bhl ? bwl : bhl);

  int bs = (1 << bsl) / 2;  //
  MODE_INFO *m2 = m + mi_row * mis + mi_col;
  for (row = 0; row < bs; row++) {
    for (col = 0; col < bs; col++) {
      if (mi_row + row >= cm->mi_rows || mi_col + col >= cm->mi_cols)
        continue;
      m2[row*mis+col].mbmi.sb_type = bsize;
    }
  }
}
typedef struct {
  int64_t sum_square_error;
  int64_t sum_error;
  int count;
  int variance;
} var;

#define VT(TYPE, BLOCKSIZE) \
  typedef struct { \
    var none; \
    var horz[2]; \
    var vert[2]; \
    BLOCKSIZE split[4]; } TYPE;

VT(v8x8, var)
VT(v16x16, v8x8)
VT(v32x32, v16x16)
VT(v64x64, v32x32)

typedef enum {
  V16X16,
  V32X32,
  V64X64,
} TREE_LEVEL;

// Set variance values given sum square error, sum error, count.
static void fill_variance(var *v, int64_t s2, int64_t s, int c) {
  v->sum_square_error = s2;
  v->sum_error = s;
  v->count = c;
  v->variance = 256
      * (v->sum_square_error - v->sum_error * v->sum_error / v->count)
      / v->count;
}

// Combine 2 variance structures by summing the sum_error, sum_square_error,
// and counts and then calculating the new variance.
void sum_2_variances(var *r, var *a, var*b) {
  fill_variance(r, a->sum_square_error + b->sum_square_error,
                a->sum_error + b->sum_error, a->count + b->count);
}
// Fill one level of our variance tree,  by summing the split sums into each of
// the horizontal, vertical and none from split and recalculating variance.
#define fill_variance_tree(VT) \
  sum_2_variances(VT.horz[0], VT.split[0].none, VT.split[1].none); \
  sum_2_variances(VT.horz[1], VT.split[2].none, VT.split[3].none); \
  sum_2_variances(VT.vert[0], VT.split[0].none, VT.split[2].none); \
  sum_2_variances(VT.vert[1], VT.split[1].none, VT.split[3].none); \
  sum_2_variances(VT.none, VT.vert[0], VT.vert[1]);

// Set the blocksize in the macroblock info structure if the variance is less
// than our threshold to one of none, horz, vert.
#define set_vt_size(VT, BLOCKSIZE, R, C, ACTION) \
  if (VT.none.variance < threshold) { \
    set_block_size(cm, m, BLOCKSIZE, mis, R, C); \
    ACTION; \
  } \
  if (VT.horz[0].variance < threshold && VT.horz[1].variance < threshold ) { \
    set_block_size(cm, m, get_subsize(BLOCKSIZE, PARTITION_HORZ), mis, R, C); \
    ACTION; \
  } \
  if (VT.vert[0].variance < threshold && VT.vert[1].variance < threshold ) { \
    set_block_size(cm, m, get_subsize(BLOCKSIZE, PARTITION_VERT), mis, R, C); \
    ACTION; \
  }

static void choose_partitioning(VP9_COMP *cpi, MODE_INFO *m, int mi_row,
                                int mi_col) {
  VP9_COMMON * const cm = &cpi->common;
  MACROBLOCK *x = &cpi->mb;
  MACROBLOCKD *xd = &cpi->mb.e_mbd;
  const int mis = cm->mode_info_stride;
  // TODO(JBB): More experimentation or testing of this threshold;
  int64_t threshold = 4;
  int i, j, k;
  v64x64 vt;
  unsigned char * s;
  int sp;
  const unsigned char * d = xd->plane[0].pre->buf;
  int dp = xd->plane[0].pre->stride;
  int pixels_wide = 64, pixels_high = 64;

  vpx_memset(&vt, 0, sizeof(vt));

  set_offsets(cpi, mi_row, mi_col, BLOCK_SIZE_SB64X64);