bitstream.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 <assert.h>
#include <stdio.h>
#include <limits.h>

#include "vpx/vpx_encoder.h"
#include "vpx_dsp/bitwriter_buffer.h"
#include "vpx_dsp/vpx_dsp_common.h"
#include "vpx_mem/vpx_mem.h"
#include "vpx_ports/mem_ops.h"
#include "vpx_ports/system_state.h"

#include "vp10/common/entropy.h"
#include "vp10/common/entropymode.h"
#include "vp10/common/entropymv.h"
#include "vp10/common/mvref_common.h"
#include "vp10/common/pred_common.h"
#include "vp10/common/seg_common.h"
#include "vp10/common/tile_common.h"

#include "vp10/encoder/cost.h"
#include "vp10/encoder/bitstream.h"
#include "vp10/encoder/encodemv.h"
#include "vp10/encoder/mcomp.h"
#include "vp10/encoder/segmentation.h"
#include "vp10/encoder/subexp.h"
#include "vp10/encoder/tokenize.h"

static const struct vp10_token intra_mode_encodings[INTRA_MODES] = {
  {0, 1}, {6, 3}, {28, 5}, {30, 5}, {58, 6}, {59, 6}, {126, 7}, {127, 7},
  {62, 6}, {2, 2}};
#if CONFIG_EXT_INTERP && SWITCHABLE_FILTERS == 4
static const struct vp10_token switchable_interp_encodings[SWITCHABLE_FILTERS] =
  {{0, 1}, {4, 3}, {3, 2}, {5, 3}};
#else
static const struct vp10_token switchable_interp_encodings[SWITCHABLE_FILTERS] =
  {{0, 1}, {2, 2}, {3, 2}};
#endif  // CONFIG_EXT_INTERP && SWITCHABLE_FILTERS == 4
static const struct vp10_token partition_encodings[PARTITION_TYPES] =
  {{0, 1}, {2, 2}, {6, 3}, {7, 3}};
static const struct vp10_token inter_mode_encodings[INTER_MODES] =
  {{2, 2}, {6, 3}, {0, 1}, {7, 3}};

static INLINE void write_uniform(vpx_writer *w, int n, int v) {
  int l = get_unsigned_bits(n);
  int m = (1 << l) - n;
  if (l == 0)
    return;
  if (v < m) {
    vpx_write_literal(w, v, l - 1);
  } else {
    vpx_write_literal(w, m + ((v - m) >> 1), l - 1);
    vpx_write_literal(w, (v - m) & 1, 1);
  }
}

#if CONFIG_EXT_TX
static struct vp10_token ext_tx_inter_encodings[EXT_TX_SETS_INTER][TX_TYPES];
static struct vp10_token ext_tx_intra_encodings[EXT_TX_SETS_INTRA][TX_TYPES];
#endif  // CONFIG_EXT_TX

void vp10_encode_token_init() {
#if CONFIG_EXT_TX
  int s;
  for (s = 1; s < EXT_TX_SETS_INTER; ++s) {
    vp10_tokens_from_tree(ext_tx_inter_encodings[s], vp10_ext_tx_inter_tree[s]);
  }
  for (s = 1; s < EXT_TX_SETS_INTRA; ++s) {
    vp10_tokens_from_tree(ext_tx_intra_encodings[s], vp10_ext_tx_intra_tree[s]);
  }
#endif  // CONFIG_EXT_TX
}

static void write_intra_mode(vpx_writer *w, PREDICTION_MODE mode,
                             const vpx_prob *probs) {
  vp10_write_token(w, vp10_intra_mode_tree, probs, &intra_mode_encodings[mode]);
}

static void write_inter_mode(vpx_writer *w, PREDICTION_MODE mode,
                             const vpx_prob *probs) {
  assert(is_inter_mode(mode));
  vp10_write_token(w, vp10_inter_mode_tree, probs,
                  &inter_mode_encodings[INTER_OFFSET(mode)]);
}

static void encode_unsigned_max(struct vpx_write_bit_buffer *wb,
                                int data, int max) {
  vpx_wb_write_literal(wb, data, get_unsigned_bits(max));
}

static void prob_diff_update(const vpx_tree_index *tree,
                             vpx_prob probs[/*n - 1*/],
                             const unsigned int counts[/*n - 1*/],
                             int n, vpx_writer *w) {
  int i;
  unsigned int branch_ct[32][2];

  // Assuming max number of probabilities <= 32
  assert(n <= 32);

  vp10_tree_probs_from_distribution(tree, branch_ct, counts);
  for (i = 0; i < n - 1; ++i)
    vp10_cond_prob_diff_update(w, &probs[i], branch_ct[i]);
}

static int prob_diff_update_savings(const vpx_tree_index *tree,
                                    vpx_prob probs[/*n - 1*/],
                                    const unsigned int counts[/*n - 1*/],
                                    int n) {
  int i;
  unsigned int branch_ct[32][2];
  int savings = 0;

  // Assuming max number of probabilities <= 32
  assert(n <= 32);
  vp10_tree_probs_from_distribution(tree, branch_ct, counts);
  for (i = 0; i < n - 1; ++i) {
    savings += vp10_cond_prob_diff_update_savings(&probs[i],
                                                  branch_ct[i]);
  }
  return savings;
}

#if CONFIG_VAR_TX
static void write_tx_size_inter(const VP10_COMMON *cm,
                                const MACROBLOCKD *xd,
                                const MB_MODE_INFO *mbmi,
                                TX_SIZE tx_size, int blk_row, int blk_col,
                                vpx_writer *w) {
  const int tx_idx = (blk_row >> 1) * 8 + (blk_col >> 1);
  int max_blocks_high = num_4x4_blocks_high_lookup[mbmi->sb_type];
  int max_blocks_wide = num_4x4_blocks_wide_lookup[mbmi->sb_type];
  int ctx = txfm_partition_context(xd->above_txfm_context + (blk_col >> 1),
                                   xd->left_txfm_context + (blk_row >> 1),
                                   tx_size);

  if (xd->mb_to_bottom_edge < 0)
    max_blocks_high += xd->mb_to_bottom_edge >> 5;
  if (xd->mb_to_right_edge < 0)
     max_blocks_wide += xd->mb_to_right_edge >> 5;

  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide)
     return;

  if (tx_size == mbmi->inter_tx_size[tx_idx]) {
    vpx_write(w, 0, cm->fc->txfm_partition_prob[ctx]);
    txfm_partition_update(xd->above_txfm_context + (blk_col >> 1),
                          xd->left_txfm_context + (blk_row >> 1), tx_size);
  } else {
    const BLOCK_SIZE bsize = txsize_to_bsize[tx_size];
    int bsl = b_width_log2_lookup[bsize];
    int i;
    vpx_write(w, 1, cm->fc->txfm_partition_prob[ctx]);

    if (tx_size == TX_8X8) {
      txfm_partition_update(xd->above_txfm_context + (blk_col >> 1),
                            xd->left_txfm_context + (blk_row >> 1), TX_4X4);
      return;
    }

    assert(bsl > 0);
    --bsl;
    for (i = 0; i < 4; ++i) {
      int offsetr = blk_row + ((i >> 1) << bsl);
      int offsetc = blk_col + ((i & 0x01) << bsl);
      write_tx_size_inter(cm, xd, mbmi, tx_size - 1, offsetr, offsetc, w);
    }
  }
}

static void update_txfm_partition_probs(VP10_COMMON *cm, vpx_writer *w,
                                        FRAME_COUNTS *counts) {
  int k;
  for (k = 0; k < TXFM_PARTITION_CONTEXTS; ++k)
    vp10_cond_prob_diff_update(w, &cm->fc->txfm_partition_prob[k],
                               counts->txfm_partition[k]);
}
#endif

static void write_selected_tx_size(const VP10_COMMON *cm,
                                   const MACROBLOCKD *xd, vpx_writer *w) {
  TX_SIZE tx_size = xd->mi[0]->mbmi.tx_size;
  BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
  const TX_SIZE max_tx_size = max_txsize_lookup[bsize];
  const vpx_prob *const tx_probs = get_tx_probs2(max_tx_size, xd,
                                                 &cm->fc->tx_probs);
  vpx_write(w, tx_size != TX_4X4, tx_probs[0]);
  if (tx_size != TX_4X4 && max_tx_size >= TX_16X16) {
    vpx_write(w, tx_size != TX_8X8, tx_probs[1]);
    if (tx_size != TX_8X8 && max_tx_size >= TX_32X32)
      vpx_write(w, tx_size != TX_16X16, tx_probs[2]);
  }
}

static int write_skip(const VP10_COMMON *cm, const MACROBLOCKD *xd,
                      int segment_id, const MODE_INFO *mi, vpx_writer *w) {
  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)) {
    return 1;
  } else {
    const int skip = mi->mbmi.skip;
    vpx_write(w, skip, vp10_get_skip_prob(cm, xd));
    return skip;
  }
}

static void update_skip_probs(VP10_COMMON *cm, vpx_writer *w,
                              FRAME_COUNTS *counts) {
  int k;

  for (k = 0; k < SKIP_CONTEXTS; ++k)
    vp10_cond_prob_diff_update(w, &cm->fc->skip_probs[k], counts->skip[k]);
}

static void update_switchable_interp_probs(VP10_COMMON *cm, vpx_writer *w,
                                           FRAME_COUNTS *counts) {
  int j;
  for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j)
    prob_diff_update(vp10_switchable_interp_tree,
                     cm->fc->switchable_interp_prob[j],
                     counts->switchable_interp[j], SWITCHABLE_FILTERS, w);
}

#if CONFIG_EXT_TX
static void update_ext_tx_probs(VP10_COMMON *cm, vpx_writer *w) {
  const int savings_thresh = vp10_cost_one(GROUP_DIFF_UPDATE_PROB) -
                             vp10_cost_zero(GROUP_DIFF_UPDATE_PROB);
  int i, j;
  int s;
  for (s = 1; s < EXT_TX_SETS_INTER; ++s) {
    int savings = 0;
    int do_update = 0;
    for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
      if (!use_inter_ext_tx_for_txsize[s][i]) continue;
      savings += prob_diff_update_savings(
          vp10_ext_tx_inter_tree[s], cm->fc->inter_ext_tx_prob[s][i],
          cm->counts.inter_ext_tx[s][i], num_ext_tx_set_inter[s]);
    }
    do_update = savings > savings_thresh;
    vpx_write(w, do_update, GROUP_DIFF_UPDATE_PROB);
    if (do_update) {
      for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
        if (!use_inter_ext_tx_for_txsize[s][i]) continue;
        prob_diff_update(vp10_ext_tx_inter_tree[s],
                         cm->fc->inter_ext_tx_prob[s][i],
                         cm->counts.inter_ext_tx[s][i],
                         num_ext_tx_set_inter[s], w);
      }
    }
  }

  for (s = 1; s < EXT_TX_SETS_INTRA; ++s) {
    int savings = 0;
    int do_update = 0;
    for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
      if (!use_intra_ext_tx_for_txsize[s][i]) continue;
      for (j = 0; j < INTRA_MODES; ++j)
        savings += prob_diff_update_savings(
            vp10_ext_tx_intra_tree[s], cm->fc->intra_ext_tx_prob[s][i][j],
            cm->counts.intra_ext_tx[s][i][j], num_ext_tx_set_intra[s]);
    }
    do_update = savings > savings_thresh;
    vpx_write(w, do_update, GROUP_DIFF_UPDATE_PROB);
    if (do_update) {
      for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
        if (!use_intra_ext_tx_for_txsize[s][i]) continue;
        for (j = 0; j < INTRA_MODES; ++j)
          prob_diff_update(vp10_ext_tx_intra_tree[s],
                           cm->fc->intra_ext_tx_prob[s][i][j],
                           cm->counts.intra_ext_tx[s][i][j],
                           num_ext_tx_set_intra[s], w);
      }
    }
  }
}
#endif  // CONFIG_EXT_TX

static void pack_mb_tokens(vpx_writer *w,
                           TOKENEXTRA **tp, const TOKENEXTRA *const stop,
                           vpx_bit_depth_t bit_depth, const TX_SIZE tx) {
  TOKENEXTRA *p = *tp;
#if CONFIG_VAR_TX
  int count = 0;
  const int seg_eob = 16 << (tx << 1);
#endif

  while (p < stop && p->token != EOSB_TOKEN) {
    const int t = p->token;
    const struct vp10_token *const a = &vp10_coef_encodings[t];
    int v = a->value;
    int n = a->len;
#if CONFIG_VP9_HIGHBITDEPTH
    const vp10_extra_bit *b;
    if (bit_depth == VPX_BITS_12)
      b = &vp10_extra_bits_high12[t];
    else if (bit_depth == VPX_BITS_10)
      b = &vp10_extra_bits_high10[t];
    else
      b = &vp10_extra_bits[t];
#else
    const vp10_extra_bit *const b = &vp10_extra_bits[t];
    (void) bit_depth;
#endif  // CONFIG_VP9_HIGHBITDEPTH

    /* skip one or two nodes */
    if (p->skip_eob_node)
      n -= p->skip_eob_node;
    else
      vpx_write(w, t != EOB_TOKEN, p->context_tree[0]);

    if (t != EOB_TOKEN) {
      vpx_write(w, t != ZERO_TOKEN, p->context_tree[1]);

      if (t != ZERO_TOKEN) {
        vpx_write(w, t != ONE_TOKEN, p->context_tree[2]);

        if (t != ONE_TOKEN) {
          int len = UNCONSTRAINED_NODES - p->skip_eob_node;
          vp10_write_tree(w, vp10_coef_con_tree,
                          vp10_pareto8_full[p->context_tree[PIVOT_NODE] - 1],
                          v, n - len, 0);
        }
      }
    }

    if (b->base_val) {
      const int e = p->extra, l = b->len;
      int skip_bits =
          (b->base_val == CAT6_MIN_VAL) ? TX_SIZES - 1 - tx : 0;

      if (l) {
        const unsigned char *pb = b->prob;
        int v = e >> 1;
        int n = l;              /* number of bits in v, assumed nonzero */
        int i = 0;

        do {
          const int bb = (v >> --n) & 1;
          if (skip_bits) {
            skip_bits--;
            assert(!bb);
          } else {
            vpx_write(w, bb, pb[i >> 1]);
          }
          i = b->tree[i + bb];
        } while (n);
      }

      vpx_write_bit(w, e & 1);
    }
    ++p;

#if CONFIG_VAR_TX
    ++count;
    if (t == EOB_TOKEN || count == seg_eob)
      break;
#endif
  }

  *tp = p;
}

#if CONFIG_VAR_TX
static void pack_txb_tokens(vpx_writer *w,
                           TOKENEXTRA **tp, const TOKENEXTRA *const tok_end,
                           MACROBLOCKD *xd, MB_MODE_INFO *mbmi, int plane,
                           BLOCK_SIZE plane_bsize,
                           vpx_bit_depth_t bit_depth,
                           int block,
                           int blk_row, int blk_col, TX_SIZE tx_size) {
  const struct macroblockd_plane *const pd = &xd->plane[plane];
  const BLOCK_SIZE bsize = txsize_to_bsize[tx_size];
  int tx_idx = (blk_row >> (1 - pd->subsampling_y)) * 8 +
               (blk_col >> (1 - pd->subsampling_x));
  TX_SIZE plane_tx_size = plane ?
      get_uv_tx_size_impl(mbmi->inter_tx_size[tx_idx], bsize, 0, 0) :
      mbmi->inter_tx_size[tx_idx];
  int max_blocks_high = num_4x4_blocks_high_lookup[plane_bsize];
  int max_blocks_wide = num_4x4_blocks_wide_lookup[plane_bsize];

  if (xd->mb_to_bottom_edge < 0)
    max_blocks_high += xd->mb_to_bottom_edge >> (5 + pd->subsampling_y);
  if (xd->mb_to_right_edge < 0)
    max_blocks_wide += xd->mb_to_right_edge >> (5 + pd->subsampling_x);

  if (blk_row >= max_blocks_high || blk_col >= max_blocks_wide)
    return;

  if (tx_size == plane_tx_size) {
    pack_mb_tokens(w, tp, tok_end, bit_depth, tx_size);
  } else {
    int bsl = b_width_log2_lookup[bsize];
    int i;

    assert(bsl > 0);
    --bsl;

    for (i = 0; i < 4; ++i) {
      const int offsetr = blk_row + ((i >> 1) << bsl);
      const int offsetc = blk_col + ((i & 0x01) << bsl);
      int step = 1 << (2 * (tx_size - 1));

      if (offsetr >= max_blocks_high || offsetc >= max_blocks_wide)
        continue;

      pack_txb_tokens(w, tp, tok_end, xd, mbmi, plane,
                      plane_bsize, bit_depth, block + i * step,
                      offsetr, offsetc, tx_size - 1);
    }
  }
}
#endif

static void write_segment_id(vpx_writer *w, const struct segmentation *seg,
                             const struct segmentation_probs *segp,
                             int segment_id) {
  if (seg->enabled && seg->update_map)
    vp10_write_tree(w, vp10_segment_tree, segp->tree_probs, segment_id, 3, 0);
}

// This function encodes the reference frame
static void write_ref_frames(const VP10_COMMON *cm, const MACROBLOCKD *xd,
                             vpx_writer *w) {
  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
  const int is_compound = has_second_ref(mbmi);
  const int segment_id = mbmi->segment_id;

  // If segment level coding of this signal is disabled...
  // or the segment allows multiple reference frame options
  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME)) {
    assert(!is_compound);
    assert(mbmi->ref_frame[0] ==
               get_segdata(&cm->seg, segment_id, SEG_LVL_REF_FRAME));
  } else {
    // does the feature use compound prediction or not
    // (if not specified at the frame/segment level)
    if (cm->reference_mode == REFERENCE_MODE_SELECT) {
      vpx_write(w, is_compound, vp10_get_reference_mode_prob(cm, xd));
    } else {
      assert(!is_compound == (cm->reference_mode == SINGLE_REFERENCE));
    }

    if (is_compound) {
#if CONFIG_EXT_REFS
      const int bit = (mbmi->ref_frame[0] == GOLDEN_FRAME ||
                       mbmi->ref_frame[0] == LAST3_FRAME ||
                       mbmi->ref_frame[0] == LAST4_FRAME);
#else
      const int bit = mbmi->ref_frame[0] == GOLDEN_FRAME;
#endif  // CONFIG_EXT_REFS
      vpx_write(w, bit, vp10_get_pred_prob_comp_ref_p(cm, xd));

#if CONFIG_EXT_REFS
      if (!bit) {
        const int bit1 = mbmi->ref_frame[0] == LAST_FRAME;
        vpx_write(w, bit1, vp10_get_pred_prob_comp_ref_p1(cm, xd));
      } else {
        const int bit2 = mbmi->ref_frame[0] == GOLDEN_FRAME;
        vpx_write(w, bit2, vp10_get_pred_prob_comp_ref_p2(cm, xd));
        if (!bit2) {
          const int bit3 = mbmi->ref_frame[0] == LAST3_FRAME;
          vpx_write(w, bit3, vp10_get_pred_prob_comp_ref_p3(cm, xd));
        }
      }
#endif  // CONFIG_EXT_REFS
    } else {
#if CONFIG_EXT_REFS
      const int bit0 = (mbmi->ref_frame[0] == GOLDEN_FRAME ||
                        mbmi->ref_frame[0] == ALTREF_FRAME);
      vpx_write(w, bit0, vp10_get_pred_prob_single_ref_p1(cm, xd));

      if (bit0) {
        const int bit1 = mbmi->ref_frame[0] != GOLDEN_FRAME;
        vpx_write(w, bit1, vp10_get_pred_prob_single_ref_p2(cm, xd));
      } else {
        const int bit2 = (mbmi->ref_frame[0] == LAST3_FRAME ||
                          mbmi->ref_frame[0] == LAST4_FRAME);
        vpx_write(w, bit2, vp10_get_pred_prob_single_ref_p3(cm, xd));

        if (!bit2) {
          const int bit3 = mbmi->ref_frame[0] != LAST_FRAME;
          vpx_write(w, bit3, vp10_get_pred_prob_single_ref_p4(cm, xd));
        } else {
          const int bit4 = mbmi->ref_frame[0] != LAST3_FRAME;
          vpx_write(w, bit4, vp10_get_pred_prob_single_ref_p5(cm, xd));
        }
      }
#else
      const int bit0 = mbmi->ref_frame[0] != LAST_FRAME;
      vpx_write(w, bit0, vp10_get_pred_prob_single_ref_p1(cm, xd));
      if (bit0) {
        const int bit1 = mbmi->ref_frame[0] != GOLDEN_FRAME;
        vpx_write(w, bit1, vp10_get_pred_prob_single_ref_p2(cm, xd));
      }
#endif  // CONFIG_EXT_REFS
    }
  }
}

#if CONFIG_EXT_INTRA
static void write_ext_intra_mode_info(const VP10_COMMON *const cm,
                                      const MB_MODE_INFO *const mbmi,
                                      vpx_writer *w) {
#if !ALLOW_FILTER_INTRA_MODES
  return;
#endif
  if (mbmi->mode == DC_PRED) {
    vpx_write(w, mbmi->ext_intra_mode_info.use_ext_intra_mode[0],
              cm->fc->ext_intra_probs[0]);
    if (mbmi->ext_intra_mode_info.use_ext_intra_mode[0]) {
      EXT_INTRA_MODE mode = mbmi->ext_intra_mode_info.ext_intra_mode[0];
      write_uniform(w, FILTER_INTRA_MODES, mode);
    }
  }
  if (mbmi->uv_mode == DC_PRED) {
    vpx_write(w, mbmi->ext_intra_mode_info.use_ext_intra_mode[1],
              cm->fc->ext_intra_probs[1]);
    if (mbmi->ext_intra_mode_info.use_ext_intra_mode[1]) {
      EXT_INTRA_MODE mode = mbmi->ext_intra_mode_info.ext_intra_mode[1];
      write_uniform(w, FILTER_INTRA_MODES, mode);
    }
  }
}
#endif  // CONFIG_EXT_INTRA

static void write_switchable_interp_filter(VP10_COMP *cpi,
                                           const MACROBLOCKD *xd,
                                           vpx_writer *w) {
  VP10_COMMON *const cm = &cpi->common;
  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
  if (cm->interp_filter == SWITCHABLE) {
    const int ctx = vp10_get_pred_context_switchable_interp(xd);
#if CONFIG_EXT_INTERP
    if (!vp10_is_interp_needed(xd)) {
      // if (mbmi->interp_filter != EIGHTTAP)
      //   printf("Error [%d]\n", mbmi->sb_type);
      assert(mbmi->interp_filter == EIGHTTAP);
      return;
    }
#endif
    vp10_write_token(w, vp10_switchable_interp_tree,
                     cm->fc->switchable_interp_prob[ctx],
                     &switchable_interp_encodings[mbmi->interp_filter]);
    ++cpi->interp_filter_selected[0][mbmi->interp_filter];
  }
}

static void pack_inter_mode_mvs(VP10_COMP *cpi, const MODE_INFO *mi,
                                vpx_writer *w) {
  VP10_COMMON *const cm = &cpi->common;
  const nmv_context *nmvc = &cm->fc->nmvc;
  const MACROBLOCK *x = &cpi->td.mb;
  const MACROBLOCKD *xd = &x->e_mbd;
  const struct segmentation *const seg = &cm->seg;
  const struct segmentation_probs *const segp = &cm->fc->seg;
  const MB_MODE_INFO *const mbmi = &mi->mbmi;
  const MB_MODE_INFO_EXT *const mbmi_ext = x->mbmi_ext;
  const PREDICTION_MODE mode = mbmi->mode;
  const int segment_id = mbmi->segment_id;
  const BLOCK_SIZE bsize = mbmi->sb_type;
  const int allow_hp = cm->allow_high_precision_mv;
  const int is_inter = is_inter_block(mbmi);
  const int is_compound = has_second_ref(mbmi);
  int skip, ref;

  if (seg->update_map) {
    if (seg->temporal_update) {
      const int pred_flag = mbmi->seg_id_predicted;
      vpx_prob pred_prob = vp10_get_pred_prob_seg_id(segp, xd);
      vpx_write(w, pred_flag, pred_prob);
      if (!pred_flag)
        write_segment_id(w, seg, segp, segment_id);
    } else {
      write_segment_id(w, seg, segp, segment_id);
    }
  }

  skip = write_skip(cm, xd, segment_id, mi, w);

  if (!segfeature_active(seg, segment_id, SEG_LVL_REF_FRAME))
    vpx_write(w, is_inter, vp10_get_intra_inter_prob(cm, xd));

  if (bsize >= BLOCK_8X8 && cm->tx_mode == TX_MODE_SELECT &&
      !(is_inter && skip)) {
#if CONFIG_VAR_TX
    if (is_inter) {  // This implies skip flag is 0.
      const TX_SIZE max_tx_size = max_txsize_lookup[bsize];
      const int txb_size = txsize_to_bsize[max_tx_size];
      const int bs = num_4x4_blocks_wide_lookup[txb_size];
      const int width  = num_4x4_blocks_wide_lookup[bsize];
      const int height = num_4x4_blocks_high_lookup[bsize];
      int idx, idy;
      for (idy = 0; idy < height; idy += bs)
        for (idx = 0; idx < width; idx += bs)
          write_tx_size_inter(cm, xd, mbmi, max_tx_size, idy, idx, w);
    } else {
      set_txfm_ctx(xd->left_txfm_context, mbmi->tx_size, xd->n8_h);
      set_txfm_ctx(xd->above_txfm_context, mbmi->tx_size, xd->n8_w);

      write_selected_tx_size(cm, xd, w);
    }
  } else {
    set_txfm_ctx(xd->left_txfm_context, mbmi->tx_size, xd->n8_h);
    set_txfm_ctx(xd->above_txfm_context, mbmi->tx_size, xd->n8_w);
#else
  write_selected_tx_size(cm, xd, w);
#endif
  }

  if (!is_inter) {
    if (bsize >= BLOCK_8X8) {
      write_intra_mode(w, mode, cm->fc->y_mode_prob[size_group_lookup[bsize]]);
#if CONFIG_EXT_INTRA
      if (mode != DC_PRED && mode != TM_PRED) {
        write_uniform(w, 2 * MAX_ANGLE_DELTAS + 1,
                      MAX_ANGLE_DELTAS + mbmi->angle_delta[0]);
      }
#endif  // CONFIG_EXT_INTRA
    } else {
      int idx, idy;
      const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
      const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
      for (idy = 0; idy < 2; idy += num_4x4_h) {
        for (idx = 0; idx < 2; idx += num_4x4_w) {
          const PREDICTION_MODE b_mode = mi->bmi[idy * 2 + idx].as_mode;
          write_intra_mode(w, b_mode, cm->fc->y_mode_prob[0]);
        }
      }
    }
    write_intra_mode(w, mbmi->uv_mode, cm->fc->uv_mode_prob[mode]);
#if CONFIG_EXT_INTRA
    if (mbmi->uv_mode != DC_PRED && mbmi->uv_mode != TM_PRED &&
        bsize >= BLOCK_8X8)
      write_uniform(w, 2 * MAX_ANGLE_DELTAS + 1,
                    MAX_ANGLE_DELTAS + mbmi->angle_delta[1]);

    if (bsize >= BLOCK_8X8)
      write_ext_intra_mode_info(cm, mbmi, w);
#endif  // CONFIG_EXT_INTRA
  } else {
    const int mode_ctx = mbmi_ext->mode_context[mbmi->ref_frame[0]];
    const vpx_prob *const inter_probs = cm->fc->inter_mode_probs[mode_ctx];
    write_ref_frames(cm, xd, w);

    // If segment skip is not enabled code the mode.
    if (!segfeature_active(seg, segment_id, SEG_LVL_SKIP)) {
      if (bsize >= BLOCK_8X8) {
        write_inter_mode(w, mode, inter_probs);
      }
    }

#if !CONFIG_EXT_INTERP
    write_switchable_interp_filter(cpi, xd, w);
#endif  // !CONFIG_EXT_INTERP

    if (bsize < BLOCK_8X8) {
      const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
      const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
      int idx, idy;
      for (idy = 0; idy < 2; idy += num_4x4_h) {
        for (idx = 0; idx < 2; idx += num_4x4_w) {
          const int j = idy * 2 + idx;
          const PREDICTION_MODE b_mode = mi->bmi[j].as_mode;
          write_inter_mode(w, b_mode, inter_probs);
          if (b_mode == NEWMV) {
            for (ref = 0; ref < 1 + is_compound; ++ref)
              vp10_encode_mv(cpi, w, &mi->bmi[j].as_mv[ref].as_mv,
                            &mbmi_ext->ref_mvs[mbmi->ref_frame[ref]][0].as_mv,
                            nmvc, allow_hp);
          }
        }
      }
    } else {
      if (mode == NEWMV) {
        for (ref = 0; ref < 1 + is_compound; ++ref)
          vp10_encode_mv(cpi, w, &mbmi->mv[ref].as_mv,
                        &mbmi_ext->ref_mvs[mbmi->ref_frame[ref]][0].as_mv, nmvc,
                        allow_hp);
      }
    }
#if CONFIG_EXT_INTERP
    write_switchable_interp_filter(cpi, xd, w);
#endif  // CONFIG_EXT_INTERP
  }

#if CONFIG_EXT_TX
  if (get_ext_tx_types(mbmi->tx_size, bsize, is_inter) > 1 &&
      cm->base_qindex > 0 && !mbmi->skip &&
      !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP)) {
    int eset = get_ext_tx_set(mbmi->tx_size, bsize, is_inter);
    if (is_inter) {
      if (eset > 0)
        vp10_write_token(w, vp10_ext_tx_inter_tree[eset],
                         cm->fc->inter_ext_tx_prob[eset][mbmi->tx_size],
                         &ext_tx_inter_encodings[eset][mbmi->tx_type]);
    } else if (ALLOW_INTRA_EXT_TX) {
      if (eset > 0)
        vp10_write_token(
            w, vp10_ext_tx_intra_tree[eset],
            cm->fc->intra_ext_tx_prob[eset][mbmi->tx_size][mbmi->mode],
            &ext_tx_intra_encodings[eset][mbmi->tx_type]);
    }
  }
#endif  // CONFIG_EXT_TX
}

static void write_mb_modes_kf(const VP10_COMMON *cm, const MACROBLOCKD *xd,
                              MODE_INFO **mi_8x8, vpx_writer *w) {
  const struct segmentation *const seg = &cm->seg;
  const struct segmentation_probs *const segp = &cm->fc->seg;
  const MODE_INFO *const mi = mi_8x8[0];
  const MODE_INFO *const above_mi = xd->above_mi;
  const MODE_INFO *const left_mi = xd->left_mi;
  const MB_MODE_INFO *const mbmi = &mi->mbmi;
  const BLOCK_SIZE bsize = mbmi->sb_type;

  if (seg->update_map)
    write_segment_id(w, seg, segp, mbmi->segment_id);

  write_skip(cm, xd, mbmi->segment_id, mi, w);

  if (bsize >= BLOCK_8X8 && cm->tx_mode == TX_MODE_SELECT)
    write_selected_tx_size(cm, xd, w);

  if (bsize >= BLOCK_8X8) {
    write_intra_mode(w, mbmi->mode,
                     get_y_mode_probs(cm, mi, above_mi, left_mi, 0));
#if CONFIG_EXT_INTRA
    if (mbmi->mode != DC_PRED && mbmi->mode != TM_PRED)
      write_uniform(w, 2 * MAX_ANGLE_DELTAS + 1,
                    MAX_ANGLE_DELTAS + mbmi->angle_delta[0]);
#endif  // CONFIG_EXT_INTRA
  } else {
    const int num_4x4_w = num_4x4_blocks_wide_lookup[bsize];
    const int num_4x4_h = num_4x4_blocks_high_lookup[bsize];
    int idx, idy;

    for (idy = 0; idy < 2; idy += num_4x4_h) {
      for (idx = 0; idx < 2; idx += num_4x4_w) {
        const int block = idy * 2 + idx;
        write_intra_mode(w, mi->bmi[block].as_mode,
                         get_y_mode_probs(cm, mi, above_mi, left_mi, block));
      }
    }
  }

  write_intra_mode(w, mbmi->uv_mode, cm->fc->uv_mode_prob[mbmi->mode]);
#if CONFIG_EXT_INTRA
  if (mbmi->uv_mode != DC_PRED && mbmi->uv_mode != TM_PRED &&
      bsize >= BLOCK_8X8)
    write_uniform(w, 2 * MAX_ANGLE_DELTAS + 1,
                  MAX_ANGLE_DELTAS + mbmi->angle_delta[1]);
#endif  // CONFIG_EXT_INTRA


#if CONFIG_EXT_TX
  if (get_ext_tx_types(mbmi->tx_size, bsize, 0) > 1 &&
      cm->base_qindex > 0 && !mbmi->skip &&
      !segfeature_active(&cm->seg, mbmi->segment_id, SEG_LVL_SKIP) &&
      ALLOW_INTRA_EXT_TX) {
    int eset = get_ext_tx_set(mbmi->tx_size, bsize, 0);
    if (eset > 0)
      vp10_write_token(
          w, vp10_ext_tx_intra_tree[eset],
          cm->fc->intra_ext_tx_prob[eset][mbmi->tx_size][mbmi->mode],
          &ext_tx_intra_encodings[eset][mbmi->tx_type]);
  }
#endif  // CONFIG_EXT_TX

#if CONFIG_EXT_INTRA
  if (bsize >= BLOCK_8X8)
      write_ext_intra_mode_info(cm, mbmi, w);
#endif  // CONFIG_EXT_INTRA
}

static void write_modes_b(VP10_COMP *cpi, const TileInfo *const tile,
                          vpx_writer *w, TOKENEXTRA **tok,
                          const TOKENEXTRA *const tok_end,
                          int mi_row, int mi_col) {
  const VP10_COMMON *const cm = &cpi->common;
  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
  MODE_INFO *m;
  int plane;

  xd->mi = cm->mi_grid_visible + (mi_row * cm->mi_stride + mi_col);
  m = xd->mi[0];

  cpi->td.mb.mbmi_ext = cpi->mbmi_ext_base + (mi_row * cm->mi_cols + mi_col);

  set_mi_row_col(xd, tile,
                 mi_row, num_8x8_blocks_high_lookup[m->mbmi.sb_type],
                 mi_col, num_8x8_blocks_wide_lookup[m->mbmi.sb_type],
                 cm->mi_rows, cm->mi_cols);
  if (frame_is_intra_only(cm)) {
    write_mb_modes_kf(cm, xd, xd->mi, w);
  } else {
#if CONFIG_VAR_TX
    xd->above_txfm_context = cm->above_txfm_context + mi_col;
    xd->left_txfm_context = xd->left_txfm_context_buffer + (mi_row & 0x07);
#endif
    pack_inter_mode_mvs(cpi, m, w);
  }

  if (!m->mbmi.skip) {
    assert(*tok < tok_end);
    for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
#if CONFIG_VAR_TX
      const struct macroblockd_plane *const pd = &xd->plane[plane];
      MB_MODE_INFO *mbmi = &m->mbmi;
      BLOCK_SIZE bsize = mbmi->sb_type;
      const BLOCK_SIZE plane_bsize =
          get_plane_block_size(VPXMAX(bsize, BLOCK_8X8), pd);

      const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
      const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
      int row, col;

      if (is_inter_block(mbmi)) {
        const TX_SIZE max_tx_size = max_txsize_lookup[plane_bsize];
        const BLOCK_SIZE txb_size = txsize_to_bsize[max_tx_size];
        int bw = num_4x4_blocks_wide_lookup[txb_size];
        int block = 0;
        const int step = 1 << (max_tx_size << 1);
        for (row = 0; row < num_4x4_h; row += bw) {
          for (col = 0; col < num_4x4_w; col += bw) {
            pack_txb_tokens(w, tok, tok_end, xd, mbmi, plane, plane_bsize,
                            cm->bit_depth, block, row, col, max_tx_size);
            block += step;
          }
        }
      } else {
        TX_SIZE tx = plane ? get_uv_tx_size(&m->mbmi, &xd->plane[plane])
                           : m->mbmi.tx_size;
        BLOCK_SIZE txb_size = txsize_to_bsize[tx];
        int bw = num_4x4_blocks_wide_lookup[txb_size];

        for (row = 0; row < num_4x4_h; row += bw)
          for (col = 0; col < num_4x4_w; col += bw)
            pack_mb_tokens(w, tok, tok_end, cm->bit_depth, tx);
      }
#else
      TX_SIZE tx = plane ? get_uv_tx_size(&m->mbmi, &xd->plane[plane])
                         : m->mbmi.tx_size;
      pack_mb_tokens(w, tok, tok_end, cm->bit_depth, tx);
#endif
      assert(*tok < tok_end && (*tok)->token == EOSB_TOKEN);
      (*tok)++;
    }
  }
}

static void write_partition(const VP10_COMMON *const cm,
                            const MACROBLOCKD *const xd,
                            int hbs, int mi_row, int mi_col,
                            PARTITION_TYPE p, BLOCK_SIZE bsize, vpx_writer *w) {
  const int ctx = partition_plane_context(xd, mi_row, mi_col, bsize);
  const vpx_prob *const probs = cm->fc->partition_prob[ctx];
  const int has_rows = (mi_row + hbs) < cm->mi_rows;
  const int has_cols = (mi_col + hbs) < cm->mi_cols;

  if (has_rows && has_cols) {
    vp10_write_token(w, vp10_partition_tree, probs, &partition_encodings[p]);
  } else if (!has_rows && has_cols) {
    assert(p == PARTITION_SPLIT || p == PARTITION_HORZ);
    vpx_write(w, p == PARTITION_SPLIT, probs[1]);
  } else if (has_rows && !has_cols) {
    assert(p == PARTITION_SPLIT || p == PARTITION_VERT);
    vpx_write(w, p == PARTITION_SPLIT, probs[2]);
  } else {
    assert(p == PARTITION_SPLIT);
  }
}

static void write_modes_sb(VP10_COMP *cpi,
                           const TileInfo *const tile, vpx_writer *w,
                           TOKENEXTRA **tok, const TOKENEXTRA *const tok_end,
                           int mi_row, int mi_col, BLOCK_SIZE bsize) {
  const VP10_COMMON *const cm = &cpi->common;
  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;

  const int bsl = b_width_log2_lookup[bsize];
  const int bs = (1 << bsl) / 4;
  PARTITION_TYPE partition;
  BLOCK_SIZE subsize;
  const MODE_INFO *m = NULL;

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

  m = cm->mi_grid_visible[mi_row * cm->mi_stride + mi_col];

  partition = partition_lookup[bsl][m->mbmi.sb_type];
  write_partition(cm, xd, bs, mi_row, mi_col, partition, bsize, w);
  subsize = get_subsize(bsize, partition);
  if (subsize < BLOCK_8X8) {
    write_modes_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
  } else {
    switch (partition) {
      case PARTITION_NONE:
        write_modes_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
        break;
      case PARTITION_HORZ:
        write_modes_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
        if (mi_row + bs < cm->mi_rows)
          write_modes_b(cpi, tile, w, tok, tok_end, mi_row + bs, mi_col);
        break;
      case PARTITION_VERT:
        write_modes_b(cpi, tile, w, tok, tok_end, mi_row, mi_col);
        if (mi_col + bs < cm->mi_cols)
          write_modes_b(cpi, tile, w, tok, tok_end, mi_row, mi_col + bs);
        break;
      case PARTITION_SPLIT:
        write_modes_sb(cpi, tile, w, tok, tok_end, mi_row, mi_col, subsize);
        write_modes_sb(cpi, tile, w, tok, tok_end, mi_row, mi_col + bs,
                       subsize);
        write_modes_sb(cpi, tile, w, tok, tok_end, mi_row + bs, mi_col,
                       subsize);
        write_modes_sb(cpi, tile, w, tok, tok_end, mi_row + bs, mi_col + bs,
                       subsize);
        break;
      default:
        assert(0);
    }
  }

  // update partition context
  if (bsize >= BLOCK_8X8 &&
      (bsize == BLOCK_8X8 || partition != PARTITION_SPLIT))
    update_partition_context(xd, mi_row, mi_col, subsize, bsize);
}

static void write_modes(VP10_COMP *cpi,
                        const TileInfo *const tile, vpx_writer *w,
                        TOKENEXTRA **tok, const TOKENEXTRA *const tok_end) {
  MACROBLOCKD *const xd = &cpi->td.mb.e_mbd;
  int mi_row, mi_col;

  for (mi_row = tile->mi_row_start; mi_row < tile->mi_row_end;
       mi_row += MI_BLOCK_SIZE) {
    vp10_zero(xd->left_seg_context);
#if CONFIG_VAR_TX
    vp10_zero(xd->left_txfm_context_buffer);
#endif
    for (mi_col = tile->mi_col_start; mi_col < tile->mi_col_end;
         mi_col += MI_BLOCK_SIZE)
      write_modes_sb(cpi, tile, w, tok, tok_end, mi_row, mi_col,
                     BLOCK_64X64);
  }
}

static void build_tree_distribution(VP10_COMP *cpi, TX_SIZE tx_size,
                                    vp10_coeff_stats *coef_branch_ct,
                                    vp10_coeff_probs_model *coef_probs) {
  vp10_coeff_count *coef_counts = cpi->td.rd_counts.coef_counts[tx_size];
  unsigned int (*eob_branch_ct)[REF_TYPES][COEF_BANDS][COEFF_CONTEXTS] =
      cpi->common.counts.eob_branch[tx_size];
  int i, j, k, l, m;

  for (i = 0; i < PLANE_TYPES; ++i) {
    for (j = 0; j < REF_TYPES; ++j) {
      for (k = 0; k < COEF_BANDS; ++k) {
        for (l = 0; l < BAND_COEFF_CONTEXTS(k); ++l) {
          vp10_tree_probs_from_distribution(vp10_coef_tree,
                                           coef_branch_ct[i][j][k][l],
                                           coef_counts[i][j][k][l]);
          coef_branch_ct[i][j][k][l][0][1] = eob_branch_ct[i][j][k][l] -
                                             coef_branch_ct[i][j][k][l][0][0];
          for (m = 0; m < UNCONSTRAINED_NODES; ++m)
            coef_probs[i][j][k][l][m] = get_binary_prob(
                                            coef_branch_ct[i][j][k][l][m][0],
                                            coef_branch_ct[i][j][k][l][m][1]);
        }
      }
    }
  }
}

static void update_coef_probs_common(vpx_writer* const bc, VP10_COMP *cpi,
                                     TX_SIZE tx_size,
                                     vp10_coeff_stats *frame_branch_ct,
                                     vp10_coeff_probs_model *new_coef_probs) {
  vp10_coeff_probs_model *old_coef_probs = cpi->common.fc->coef_probs[tx_size];
  const vpx_prob upd = DIFF_UPDATE_PROB;
  const int entropy_nodes_update = UNCONSTRAINED_NODES;
  int i, j, k, l, t;
  int stepsize = cpi->sf.coeff_prob_appx_step;

  switch (cpi->sf.use_fast_coef_updates) {
    case TWO_LOOP: {
      /* dry run to see if there is any update at all needed */
      int savings = 0;
      int update[2] = {0, 0};