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

#include "./vp9_rtcd.h"
#include "./vpx_config.h"

#include "vpx_ports/vpx_timer.h"

#include "vp9/common/vp9_common.h"
#include "vp9/common/vp9_entropy.h"
#include "vp9/common/vp9_entropymode.h"
#include "vp9/common/vp9_extend.h"
#include "vp9/common/vp9_findnearmv.h"
#include "vp9/common/vp9_idct.h"
#include "vp9/common/vp9_mvref_common.h"
#include "vp9/common/vp9_pred_common.h"
#include "vp9/common/vp9_quant_common.h"
#include "vp9/common/vp9_reconintra.h"
#include "vp9/common/vp9_reconinter.h"
#include "vp9/common/vp9_seg_common.h"
#include "vp9/common/vp9_tile_common.h"

#include "vp9/encoder/vp9_encodeframe.h"
#include "vp9/encoder/vp9_encodeintra.h"
#include "vp9/encoder/vp9_encodemb.h"
#include "vp9/encoder/vp9_encodemv.h"
#include "vp9/encoder/vp9_onyx_int.h"
#include "vp9/encoder/vp9_rdopt.h"
#include "vp9/encoder/vp9_segmentation.h"
#include "vp9/common/vp9_systemdependent.h"
#include "vp9/encoder/vp9_tokenize.h"
#include "vp9/encoder/vp9_vaq.h"


#define DBG_PRNT_SEGMAP 0


static const TX_SIZE tx_mode_to_biggest_tx_size[TX_MODES] = {
  TX_4X4,  // ONLY_4X4
  TX_8X8,  // ONLY_8X8
  TX_16X16,  // ONLY_16X16
  TX_32X32,  // ONLY_32X32
  TX_32X32,  // TX_MODE_SELECT
};

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

static void encode_superblock(VP9_COMP *cpi, TOKENEXTRA **t, int output_enabled,
                              int mi_row, int mi_col, BLOCK_SIZE 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 ACTIVITY_AVG_MIN (64)

/* Motion vector component magnitude threshold for defining fast motion. */
#define FAST_MOTION_MV_THRESH (24)

/* 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[64] = {
  128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128,
  128, 128, 128, 128, 128, 128, 128, 128
};

static unsigned int get_sby_perpixel_variance(VP9_COMP *cpi, MACROBLOCK *x,
                                              BLOCK_SIZE bs) {
  unsigned int var, sse;
  var = cpi->fn_ptr[bs].vf(x->plane[0].src.buf,
                           x->plane[0].src.stride,
                           VP9_VAR_OFFS, 0, &sse);
  return (var + (1 << (num_pels_log2_lookup[bs] - 1))) >>
      num_pels_log2_lookup[bs];
}

// Original activity measure from Tim T's code.
static unsigned int tt_activity_measure(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(MACROBLOCK *x, int use_dc_pred) {
  return vp9_encode_intra(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(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(x, use_dc_pred);
  } else {
    // Original activity measure from Tim T's code.
    mb_activity = tt_activity_measure(x);
  }

  if (mb_activity < ACTIVITY_AVG_MIN)
    mb_activity = 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(&cpi->common, 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  // ACT_MEDIAN

  if (cpi->activity_avg < ACTIVITY_AVG_MIN)
    cpi->activity_avg = 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  // USE_ACT_INDEX

// 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(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 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->mi_8x8[0]->mbmi;
  MODE_INFO *mi_addr = xd->mi_8x8[0];

  int mb_mode_index = ctx->best_mode_index;
  const int mis = cm->mode_info_stride;
  const int mi_width = num_8x8_blocks_wide_lookup[bsize];
  const int mi_height = num_8x8_blocks_high_lookup[bsize];

  assert(mi->mbmi.mode < MB_MODE_COUNT);
  assert(mi->mbmi.ref_frame[0] < MAX_REF_FRAMES);
  assert(mi->mbmi.ref_frame[1] < MAX_REF_FRAMES);
  assert(mi->mbmi.sb_type == bsize);

  *mi_addr = *mi;

  // Restore the coding context of the MB to that that was in place
  // when the mode was picked for it
  for (y = 0; y < mi_height; y++)
    for (x_idx = 0; x_idx < mi_width; x_idx++)
      if ((xd->mb_to_right_edge >> (3 + MI_SIZE_LOG2)) + mi_width > x_idx
          && (xd->mb_to_bottom_edge >> (3 + MI_SIZE_LOG2)) + mi_height > y)
        xd->mi_8x8[x_idx + y * mis] = mi_addr;

  if (cpi->sf.variance_adaptive_quantization) {
    vp9_mb_init_quantizer(cpi, x);
  }

  // FIXME(rbultje) I'm pretty sure this should go to the end of this block
  // (i.e. after the output_enabled)
  if (bsize < BLOCK_32X32) {
    if (bsize < BLOCK_16X16)
      ctx->tx_rd_diff[ALLOW_16X16] = ctx->tx_rd_diff[ALLOW_8X8];
    ctx->tx_rd_diff[ALLOW_32X32] = ctx->tx_rd_diff[ALLOW_16X16];
  }

  if (is_inter_block(mbmi) && mbmi->sb_type < BLOCK_8X8) {
    mbmi->mv[0].as_int = mi->bmi[3].as_mv[0].as_int;
    mbmi->mv[1].as_int = mi->bmi[3].as_mv[1].as_int;
  }

  x->skip = ctx->skip;
  vpx_memcpy(x->zcoeff_blk[mbmi->tx_size], ctx->zcoeff_blk,
             sizeof(ctx->zcoeff_blk));

  if (!output_enabled)
    return;

  if (!vp9_segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
    for (i = 0; i < TX_MODES; i++)
      cpi->rd_tx_select_diff[i] += ctx->tx_rd_diff[i];
  }

  if (frame_is_intra_only(cm)) {
#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_D207_PRED /*D207_PRED*/,
      THR_D63_PRED /*D63_PRED*/,
      THR_TM /*TM_PRED*/,
    };
    cpi->mode_chosen_counts[kf_mode_index[mi->mbmi.mode]]++;
#endif
  } else {
    // Note how often each mode chosen as best
    cpi->mode_chosen_counts[mb_mode_index]++;
    if (is_inter_block(mbmi)
        && (mbmi->sb_type < BLOCK_8X8 || mbmi->mode == NEWMV)) {
      int_mv best_mv[2];
      const MV_REFERENCE_FRAME rf1 = mbmi->ref_frame[0];
      const MV_REFERENCE_FRAME rf2 = mbmi->ref_frame[1];
      best_mv[0].as_int = ctx->best_ref_mv.as_int;
      best_mv[1].as_int = ctx->second_best_ref_mv.as_int;
      if (mbmi->mode == NEWMV) {
        best_mv[0].as_int = mbmi->ref_mvs[rf1][0].as_int;
        if (rf2 > 0)
          best_mv[1].as_int = mbmi->ref_mvs[rf2][0].as_int;
      }
      mbmi->best_mv[0].as_int = best_mv[0].as_int;
      mbmi->best_mv[1].as_int = best_mv[1].as_int;
      vp9_update_mv_count(cpi, x, best_mv);
    }

    if (cm->mcomp_filter_type == SWITCHABLE && is_inter_mode(mbmi->mode)) {
      const int ctx = vp9_get_pred_context_switchable_interp(xd);
      ++cm->counts.switchable_interp[ctx][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;

    for (i = 0; i <= SWITCHABLE_FILTERS; i++)
      cpi->rd_filter_diff[i] += ctx->best_filter_diff[i];
  }
}

void vp9_setup_src_planes(MACROBLOCK *x, const YV12_BUFFER_CONFIG *src,
                          int mi_row, int mi_col) {
  uint8_t *const buffers[4] = {src->y_buffer, src->u_buffer, src->v_buffer,
                               src->alpha_buffer};
  const 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], mi_row, mi_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 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 mi_width = num_8x8_blocks_wide_lookup[bsize];
  const int mi_height = num_8x8_blocks_high_lookup[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;
  const struct segmentation *const seg = &cm->seg;

  set_skip_context(cm, xd, mi_row, mi_col);
  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;

  xd->mi_8x8 = cm->mi_grid_visible + idx_str;
  xd->prev_mi_8x8 = cm->prev_mi_grid_visible + idx_str;

  // Special case: if prev_mi is NULL, the previous mode info context
  // cannot be used.
  xd->last_mi = cm->prev_mi ? xd->prev_mi_8x8[0] : NULL;

  xd->mi_8x8[0] = cm->mi + idx_str;

  mbmi = &xd->mi_8x8[0]->mbmi;

  // 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
  // mv beyond the range do not produce new/different prediction block
  x->mv_row_min = -(((mi_row + mi_height) * MI_SIZE) + VP9_INTERP_EXTEND);
  x->mv_col_min = -(((mi_col + mi_width) * MI_SIZE) + VP9_INTERP_EXTEND);
  x->mv_row_max = (cm->mi_rows - mi_row) * MI_SIZE + VP9_INTERP_EXTEND;
  x->mv_col_max = (cm->mi_cols - mi_col) * MI_SIZE + VP9_INTERP_EXTEND;

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

  /* 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 (seg->enabled) {
    if (!cpi->sf.variance_adaptive_quantization) {
      uint8_t *map = seg->update_map ? cpi->segmentation_map
          : cm->last_frame_seg_map;
      mbmi->segment_id = vp9_get_segment_id(cm, map, bsize, mi_row, mi_col);
    }
    vp9_mb_init_quantizer(cpi, x);

    if (seg->enabled && cpi->seg0_cnt > 0
        && !vp9_segfeature_active(seg, 0, SEG_LVL_REF_FRAME)
        && vp9_segfeature_active(seg, 1, SEG_LVL_REF_FRAME)) {
      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;
    }

    x->encode_breakout = cpi->segment_encode_breakout[mbmi->segment_id];
  } else {
    mbmi->segment_id = 0;
    x->encode_breakout = cpi->oxcf.encode_breakout;
  }
}

static void pick_sb_modes(VP9_COMP *cpi, int mi_row, int mi_col,
                          int *totalrate, int64_t *totaldist,
                          BLOCK_SIZE bsize, PICK_MODE_CONTEXT *ctx,
                          int64_t best_rd) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  int orig_rdmult = x->rdmult;
  double rdmult_ratio;

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

  // Use the lower precision, but faster, 32x32 fdct for mode selection.
  x->use_lp32x32fdct = 1;

  if (bsize < BLOCK_8X8) {
    // When ab_index = 0 all sub-blocks are handled, so for ab_index != 0
    // there is nothing to be done.
    if (xd->ab_index != 0) {
      *totalrate = 0;
      *totaldist = 0;
      return;
    }
  }

  set_offsets(cpi, mi_row, mi_col, bsize);
  xd->mi_8x8[0]->mbmi.sb_type = bsize;

  // Set to zero to make sure we do not use the previous encoded frame stats
  xd->mi_8x8[0]->mbmi.skip_coeff = 0;

  x->source_variance = get_sby_perpixel_variance(cpi, x, bsize);

  if (cpi->sf.variance_adaptive_quantization) {
    int energy;
    if (bsize <= BLOCK_16X16) {
      energy = x->mb_energy;
    } else {
      energy = vp9_block_energy(cpi, x, bsize);
    }

    xd->mi_8x8[0]->mbmi.segment_id = vp9_vaq_segment_id(energy);
    rdmult_ratio = vp9_vaq_rdmult_ratio(energy);
    vp9_mb_init_quantizer(cpi, x);
  }

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

  if (cpi->sf.variance_adaptive_quantization) {
    vp9_clear_system_state();  // __asm emms;
    x->rdmult = round(x->rdmult * rdmult_ratio);
  }

  // Find best coding mode & reconstruct the MB so it is available
  // as a predictor for MBs that follow in the SB
  if (frame_is_intra_only(cm)) {
    vp9_rd_pick_intra_mode_sb(cpi, x, totalrate, totaldist, bsize, ctx,
                              best_rd);
  } else {
    if (bsize >= BLOCK_8X8)
      vp9_rd_pick_inter_mode_sb(cpi, x, mi_row, mi_col, totalrate, totaldist,
                                bsize, ctx, best_rd);
    else
      vp9_rd_pick_inter_mode_sub8x8(cpi, x, mi_row, mi_col, totalrate,
                                    totaldist, bsize, ctx, best_rd);
  }

  if (cpi->sf.variance_adaptive_quantization) {
    x->rdmult = orig_rdmult;
    if (*totalrate != INT_MAX) {
      vp9_clear_system_state();  // __asm emms;
      *totalrate = round(*totalrate * rdmult_ratio);
    }
  }
}

static void update_stats(VP9_COMP *cpi) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  MODE_INFO *mi = xd->mi_8x8[0];
  MB_MODE_INFO *const mbmi = &mi->mbmi;

  if (!frame_is_intra_only(cm)) {
    const int seg_ref_active = vp9_segfeature_active(&cm->seg, mbmi->segment_id,
                                                     SEG_LVL_REF_FRAME);

    if (!seg_ref_active)
      cpi->intra_inter_count[vp9_get_pred_context_intra_inter(xd)]
                            [is_inter_block(mbmi)]++;

    // If the segment reference feature is enabled we have only a single
    // reference frame allowed for the segment so exclude it from
    // the reference frame counts used to work out probabilities.
    if (is_inter_block(mbmi) && !seg_ref_active) {
      if (cm->comp_pred_mode == HYBRID_PREDICTION)
        cpi->comp_inter_count[vp9_get_pred_context_comp_inter_inter(cm, xd)]
                             [has_second_ref(mbmi)]++;

      if (has_second_ref(mbmi)) {
        cpi->comp_ref_count[vp9_get_pred_context_comp_ref_p(cm, xd)]
                           [mbmi->ref_frame[0] == GOLDEN_FRAME]++;
      } else {
        cpi->single_ref_count[vp9_get_pred_context_single_ref_p1(xd)][0]
                             [mbmi->ref_frame[0] != LAST_FRAME]++;
        if (mbmi->ref_frame[0] != LAST_FRAME)
          cpi->single_ref_count[vp9_get_pred_context_single_ref_p2(xd)][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 temporary buffers, when recursive
// partition down to 4x4 block size is enabled.
static PICK_MODE_CONTEXT *get_block_context(MACROBLOCK *x, BLOCK_SIZE bsize) {
  MACROBLOCKD *const xd = &x->e_mbd;

  switch (bsize) {
    case BLOCK_64X64:
      return &x->sb64_context;
    case BLOCK_64X32:
      return &x->sb64x32_context[xd->sb_index];
    case BLOCK_32X64:
      return &x->sb32x64_context[xd->sb_index];
    case BLOCK_32X32:
      return &x->sb32_context[xd->sb_index];
    case BLOCK_32X16:
      return &x->sb32x16_context[xd->sb_index][xd->mb_index];
    case BLOCK_16X32:
      return &x->sb16x32_context[xd->sb_index][xd->mb_index];
    case BLOCK_16X16:
      return &x->mb_context[xd->sb_index][xd->mb_index];
    case BLOCK_16X8:
      return &x->sb16x8_context[xd->sb_index][xd->mb_index][xd->b_index];
    case BLOCK_8X16:
      return &x->sb8x16_context[xd->sb_index][xd->mb_index][xd->b_index];
    case BLOCK_8X8:
      return &x->sb8x8_context[xd->sb_index][xd->mb_index][xd->b_index];
    case BLOCK_8X4:
      return &x->sb8x4_context[xd->sb_index][xd->mb_index][xd->b_index];
    case BLOCK_4X8:
      return &x->sb4x8_context[xd->sb_index][xd->mb_index][xd->b_index];
    case BLOCK_4X4:
      return &x->ab4x4_context[xd->sb_index][xd->mb_index][xd->b_index];
    default:
      assert(0);
      return NULL;
  }
}

static BLOCK_SIZE *get_sb_partitioning(MACROBLOCK *x, BLOCK_SIZE bsize) {
  MACROBLOCKD *const xd = &x->e_mbd;
  switch (bsize) {
    case BLOCK_64X64:
      return &x->sb64_partitioning;
    case BLOCK_32X32:
      return &x->sb_partitioning[xd->sb_index];
    case BLOCK_16X16:
      return &x->mb_partitioning[xd->sb_index][xd->mb_index];
    case BLOCK_8X8:
      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 bsize) {
  VP9_COMMON *const cm = &cpi->common;
  MACROBLOCK *const x = &cpi->mb;
  MACROBLOCKD *const xd = &x->e_mbd;
  int p;
  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
  int mi_width = num_8x8_blocks_wide_lookup[bsize];
  int mi_height = num_8x8_blocks_high_lookup[bsize];
  for (p = 0; p < MAX_MB_PLANE; p++) {
    vpx_memcpy(
        cm->above_context[p] + ((mi_col * 2) >> xd->plane[p].subsampling_x),
        a + num_4x4_blocks_wide * p,
        (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_wide) >>
        xd->plane[p].subsampling_x);
    vpx_memcpy(
        cm->left_context[p]
            + ((mi_row & MI_MASK) * 2 >> xd->plane[p].subsampling_y),
        l + num_4x4_blocks_high * p,
        (sizeof(ENTROPY_CONTEXT) * num_4x4_blocks_high) >>
        xd->plane[p].subsampling_y);
  }
  vpx_memcpy(cm->above_seg_context + mi_col, sa,
             sizeof(PARTITION_CONTEXT) * mi_width);
  vpx_memcpy(cm->left_seg_context + (mi_row & MI_MASK), sl,
             sizeof(PARTITION_CONTEXT) * mi_height);
}
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 bsize) {
  const VP9_COMMON *const cm = &cpi->common;
  const MACROBLOCK *const x = &cpi->mb;
  const MACROBLOCKD *const xd = &x->e_mbd;
  int p;
  const int num_4x4_blocks_wide = num_4x4_blocks_wide_lookup[bsize];
  const int num_4x4_blocks_high = num_4x4_blocks_high_lookup[bsize];
  int mi_width = num_8x8_blocks_wide_lookup[bsize];
  int mi_height = num_8x8_blocks_high_lookup[bsize];

  // buffer the above/left context information of the block in search.
  for (p = 0; p < MAX_MB_PLANE; ++p) {
    vpx_memcpy(
        a + num_4x4_blocks_wide * p,
        cm->above_context[p] + (mi_col * 2 >> xd->plane[p].subsampling_x),