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

#include "aom/aom_encoder.h"
#include "aom_dsp/bitwriter_buffer.h"
#include "aom_dsp/aom_dsp_common.h"
#include "aom_mem/aom_mem.h"
#include "aom_ports/mem_ops.h"
#include "aom_ports/system_state.h"
#include "aom_util/debug_util.h"

#if CONFIG_CLPF
#include "av1/common/clpf.h"
#endif
#if CONFIG_DERING
#include "av1/common/dering.h"
#endif  // CONFIG_DERING
#include "av1/common/entropy.h"
#include "av1/common/entropymode.h"
#include "av1/common/entropymv.h"
#include "av1/common/mvref_common.h"
#include "av1/common/pred_common.h"
#include "av1/common/reconinter.h"
#include "av1/common/seg_common.h"
#include "av1/common/tile_common.h"

#if CONFIG_ANS
#include "av1/encoder/buf_ans.h"
#endif  // CONFIG_ANS
#include "av1/encoder/bitstream.h"
#include "av1/encoder/cost.h"
#include "av1/encoder/encodemv.h"
#include "av1/encoder/mcomp.h"
#include "av1/encoder/segmentation.h"
#include "av1/encoder/subexp.h"
#include "av1/encoder/tokenize.h"

static const struct av1_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
static const struct av1_token switchable_interp_encodings[SWITCHABLE_FILTERS] =
    { { 0, 1 }, { 4, 3 }, { 6, 3 }, { 5, 3 }, { 7, 3 } };
#else
static const struct av1_token switchable_interp_encodings[SWITCHABLE_FILTERS] =
    { { 0, 1 }, { 2, 2 }, { 3, 2 } };
#endif  // CONFIG_EXT_INTERP
#if CONFIG_EXT_PARTITION_TYPES
static const struct av1_token ext_partition_encodings[EXT_PARTITION_TYPES] = {
  { 0, 1 },  { 4, 3 },  { 12, 4 }, { 7, 3 },
  { 10, 4 }, { 11, 4 }, { 26, 5 }, { 27, 5 }
};
#endif
static const struct av1_token partition_encodings[PARTITION_TYPES] = {
  { 0, 1 }, { 2, 2 }, { 6, 3 }, { 7, 3 }
};
#if !CONFIG_REF_MV
static const struct av1_token inter_mode_encodings[INTER_MODES] =
#if CONFIG_EXT_INTER
    { { 2, 2 }, { 6, 3 }, { 0, 1 }, { 14, 4 }, { 15, 4 } };
#else
    { { 2, 2 }, { 6, 3 }, { 0, 1 }, { 7, 3 } };
#endif  // CONFIG_EXT_INTER
#endif
#if CONFIG_EXT_INTER
static const struct av1_token
    inter_compound_mode_encodings[INTER_COMPOUND_MODES] = {
      { 2, 2 },  { 50, 6 }, { 51, 6 }, { 24, 5 }, { 52, 6 },
      { 53, 6 }, { 54, 6 }, { 55, 6 }, { 0, 1 },  { 7, 3 }
    };
#endif  // CONFIG_EXT_INTER
static const struct av1_token palette_size_encodings[] = {
  { 0, 1 }, { 2, 2 }, { 6, 3 }, { 14, 4 }, { 30, 5 }, { 62, 6 }, { 63, 6 },
};
static const struct av1_token
    palette_color_encodings[PALETTE_MAX_SIZE - 1][PALETTE_MAX_SIZE] = {
      { { 0, 1 }, { 1, 1 } },                                  // 2 colors
      { { 0, 1 }, { 2, 2 }, { 3, 2 } },                        // 3 colors
      { { 0, 1 }, { 2, 2 }, { 6, 3 }, { 7, 3 } },              // 4 colors
      { { 0, 1 }, { 2, 2 }, { 6, 3 }, { 14, 4 }, { 15, 4 } },  // 5 colors
      { { 0, 1 },
        { 2, 2 },
        { 6, 3 },
        { 14, 4 },
        { 30, 5 },
        { 31, 5 } },  // 6 colors
      { { 0, 1 },
        { 2, 2 },
        { 6, 3 },
        { 14, 4 },
        { 30, 5 },
        { 62, 6 },
        { 63, 6 } },  // 7 colors
      { { 0, 1 },
        { 2, 2 },
        { 6, 3 },
        { 14, 4 },
        { 30, 5 },
        { 62, 6 },
        { 126, 7 },
        { 127, 7 } },  // 8 colors
    };

static const struct av1_token tx_size_encodings[TX_SIZES - 1][TX_SIZES] = {
  { { 0, 1 }, { 1, 1 } },                      // Max tx_size is 8X8
  { { 0, 1 }, { 2, 2 }, { 3, 2 } },            // Max tx_size is 16X16
  { { 0, 1 }, { 2, 2 }, { 6, 3 }, { 7, 3 } },  // Max tx_size is 32X32
};

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

#if CONFIG_EXT_TX
static struct av1_token ext_tx_inter_encodings[EXT_TX_SETS_INTER][TX_TYPES];
static struct av1_token ext_tx_intra_encodings[EXT_TX_SETS_INTRA][TX_TYPES];
#else
static struct av1_token ext_tx_encodings[TX_TYPES];
#endif  // CONFIG_EXT_TX
#if CONFIG_GLOBAL_MOTION
static struct av1_token global_motion_types_encodings[GLOBAL_MOTION_TYPES];
#endif  // CONFIG_GLOBAL_MOTION
#if CONFIG_EXT_INTRA
static struct av1_token intra_filter_encodings[INTRA_FILTERS];
#endif  // CONFIG_EXT_INTRA
#if CONFIG_EXT_INTER
static struct av1_token interintra_mode_encodings[INTERINTRA_MODES];
#endif  // CONFIG_EXT_INTER
#if CONFIG_OBMC || CONFIG_WARPED_MOTION
static struct av1_token motvar_encodings[MOTION_VARIATIONS];
#endif  // CONFIG_OBMC || CONFIG_WARPED_MOTION

void av1_encode_token_init(void) {
#if CONFIG_EXT_TX
  int s;
  for (s = 1; s < EXT_TX_SETS_INTER; ++s) {
    av1_tokens_from_tree(ext_tx_inter_encodings[s], av1_ext_tx_inter_tree[s]);
  }
  for (s = 1; s < EXT_TX_SETS_INTRA; ++s) {
    av1_tokens_from_tree(ext_tx_intra_encodings[s], av1_ext_tx_intra_tree[s]);
  }
#else
  av1_tokens_from_tree(ext_tx_encodings, av1_ext_tx_tree);
#endif  // CONFIG_EXT_TX
#if CONFIG_EXT_INTRA
  av1_tokens_from_tree(intra_filter_encodings, av1_intra_filter_tree);
#endif  // CONFIG_EXT_INTRA
#if CONFIG_EXT_INTER
  av1_tokens_from_tree(interintra_mode_encodings, av1_interintra_mode_tree);
#endif  // CONFIG_EXT_INTER
#if CONFIG_OBMC || CONFIG_WARPED_MOTION
  av1_tokens_from_tree(motvar_encodings, av1_motvar_tree);
#endif  // CONFIG_OBMC || CONFIG_WARPED_MOTION
#if CONFIG_GLOBAL_MOTION
  av1_tokens_from_tree(global_motion_types_encodings,
                       av1_global_motion_types_tree);
#endif  // CONFIG_GLOBAL_MOTION
}

static void write_intra_mode(aom_writer *w, PREDICTION_MODE mode,
                             const aom_prob *probs) {
  av1_write_token(w, av1_intra_mode_tree, probs, &intra_mode_encodings[mode]);
}

#if CONFIG_EXT_INTER
static void write_interintra_mode(aom_writer *w, INTERINTRA_MODE mode,
                                  const aom_prob *probs) {
  av1_write_token(w, av1_interintra_mode_tree, probs,
                  &interintra_mode_encodings[mode]);
}
#endif  // CONFIG_EXT_INTER

static void write_inter_mode(AV1_COMMON *cm, aom_writer *w,
                             PREDICTION_MODE mode,
#if CONFIG_REF_MV && CONFIG_EXT_INTER
                             int is_compound,
#endif  // CONFIG_REF_MV && CONFIG_EXT_INTER
                             const int16_t mode_ctx) {
#if CONFIG_REF_MV
  const int16_t newmv_ctx = mode_ctx & NEWMV_CTX_MASK;
  const aom_prob newmv_prob = cm->fc->newmv_prob[newmv_ctx];
#if CONFIG_EXT_INTER
  aom_write(w, mode != NEWMV && mode != NEWFROMNEARMV, newmv_prob);

  if (!is_compound && (mode == NEWMV || mode == NEWFROMNEARMV))
    aom_write(w, mode == NEWFROMNEARMV, cm->fc->new2mv_prob);

  if (mode != NEWMV && mode != NEWFROMNEARMV) {
#else
  aom_write(w, mode != NEWMV, newmv_prob);

  if (mode != NEWMV) {
#endif  // CONFIG_EXT_INTER
    const int16_t zeromv_ctx = (mode_ctx >> ZEROMV_OFFSET) & ZEROMV_CTX_MASK;
    const aom_prob zeromv_prob = cm->fc->zeromv_prob[zeromv_ctx];

    if (mode_ctx & (1 << ALL_ZERO_FLAG_OFFSET)) {
      assert(mode == ZEROMV);
      return;
    }

    aom_write(w, mode != ZEROMV, zeromv_prob);

    if (mode != ZEROMV) {
      int16_t refmv_ctx = (mode_ctx >> REFMV_OFFSET) & REFMV_CTX_MASK;
      aom_prob refmv_prob;

      if (mode_ctx & (1 << SKIP_NEARESTMV_OFFSET)) refmv_ctx = 6;
      if (mode_ctx & (1 << SKIP_NEARMV_OFFSET)) refmv_ctx = 7;
      if (mode_ctx & (1 << SKIP_NEARESTMV_SUB8X8_OFFSET)) refmv_ctx = 8;

      refmv_prob = cm->fc->refmv_prob[refmv_ctx];
      aom_write(w, mode != NEARESTMV, refmv_prob);
    }
  }
#else
  const aom_prob *const inter_probs = cm->fc->inter_mode_probs[mode_ctx];
  assert(is_inter_mode(mode));
  av1_write_token(w, av1_inter_mode_tree, inter_probs,
                  &inter_mode_encodings[INTER_OFFSET(mode)]);
#endif
}

#if CONFIG_REF_MV
static void write_drl_idx(const AV1_COMMON *cm, const MB_MODE_INFO *mbmi,
                          const MB_MODE_INFO_EXT *mbmi_ext, aom_writer *w) {
  uint8_t ref_frame_type = av1_ref_frame_type(mbmi->ref_frame);

  assert(mbmi->ref_mv_idx < 3);

  if (mbmi->mode == NEWMV) {
    int idx;
    for (idx = 0; idx < 2; ++idx) {
      if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) {
        uint8_t drl_ctx =
            av1_drl_ctx(mbmi_ext->ref_mv_stack[ref_frame_type], idx);
        aom_prob drl_prob = cm->fc->drl_prob[drl_ctx];

        aom_write(w, mbmi->ref_mv_idx != idx, drl_prob);
        if (mbmi->ref_mv_idx == idx) return;
      }
    }
    return;
  }

  if (mbmi->mode == NEARMV) {
    int idx;
    // TODO(jingning): Temporary solution to compensate the NEARESTMV offset.
    for (idx = 1; idx < 3; ++idx) {
      if (mbmi_ext->ref_mv_count[ref_frame_type] > idx + 1) {
        uint8_t drl_ctx =
            av1_drl_ctx(mbmi_ext->ref_mv_stack[ref_frame_type], idx);
        aom_prob drl_prob = cm->fc->drl_prob[drl_ctx];

        aom_write(w, mbmi->ref_mv_idx != (idx - 1), drl_prob);
        if (mbmi->ref_mv_idx == (idx - 1)) return;
      }
    }
    return;
  }
}
#endif

#if CONFIG_EXT_INTER
static void write_inter_compound_mode(AV1_COMMON *cm, aom_writer *w,
                                      PREDICTION_MODE mode,
                                      const int16_t mode_ctx) {
  const aom_prob *const inter_compound_probs =
      cm->fc->inter_compound_mode_probs[mode_ctx];

  assert(is_inter_compound_mode(mode));
  av1_write_token(w, av1_inter_compound_mode_tree, inter_compound_probs,
                  &inter_compound_mode_encodings[INTER_COMPOUND_OFFSET(mode)]);
}
#endif  // CONFIG_EXT_INTER

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

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

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

  av1_tree_probs_from_distribution(tree, branch_ct, counts);
  for (i = 0; i < n - 1; ++i)
    av1_cond_prob_diff_update(w, &probs[i], branch_ct[i]);
}

static int prob_diff_update_savings(const aom_tree_index *tree,
                                    aom_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);
  av1_tree_probs_from_distribution(tree, branch_ct, counts);
  for (i = 0; i < n - 1; ++i) {
    savings += av1_cond_prob_diff_update_savings(&probs[i], branch_ct[i]);
  }
  return savings;
}

#if CONFIG_VAR_TX
static void write_tx_size_vartx(const AV1_COMMON *cm, const MACROBLOCKD *xd,
                                const MB_MODE_INFO *mbmi, TX_SIZE tx_size,
                                int blk_row, int blk_col, aom_writer *w) {
  const int tx_row = blk_row >> 1;
  const int tx_col = 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 + tx_col,
                                   xd->left_txfm_context + tx_row, 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_row][tx_col]) {
    aom_write(w, 0, cm->fc->txfm_partition_prob[ctx]);
    txfm_partition_update(xd->above_txfm_context + tx_col,
                          xd->left_txfm_context + tx_row, tx_size);
  } else {
    const BLOCK_SIZE bsize = txsize_to_bsize[tx_size];
    int bsl = b_width_log2_lookup[bsize];
    int i;
    aom_write(w, 1, cm->fc->txfm_partition_prob[ctx]);

    if (tx_size == TX_8X8) {
      txfm_partition_update(xd->above_txfm_context + tx_col,
                            xd->left_txfm_context + tx_row, 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_vartx(cm, xd, mbmi, tx_size - 1, offsetr, offsetc, w);
    }
  }
}

static void update_txfm_partition_probs(AV1_COMMON *cm, aom_writer *w,
                                        FRAME_COUNTS *counts) {
  int k;
  for (k = 0; k < TXFM_PARTITION_CONTEXTS; ++k)
    av1_cond_prob_diff_update(w, &cm->fc->txfm_partition_prob[k],
                              counts->txfm_partition[k]);
}
#endif

static void write_selected_tx_size(const AV1_COMMON *cm, const MACROBLOCKD *xd,
                                   aom_writer *w) {
  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
  const BLOCK_SIZE bsize = mbmi->sb_type;
  // For sub8x8 blocks the tx_size symbol does not need to be sent
  if (bsize >= BLOCK_8X8) {
    const TX_SIZE tx_size = mbmi->tx_size;
    const int is_inter = is_inter_block(mbmi);
    const int tx_size_ctx = get_tx_size_context(xd);
    const int tx_size_cat = is_inter ? inter_tx_size_cat_lookup[bsize]
                                     : intra_tx_size_cat_lookup[bsize];
    const TX_SIZE coded_tx_size = txsize_sqr_up_map[tx_size];

#if CONFIG_EXT_TX && CONFIG_RECT_TX
    assert(IMPLIES(is_rect_tx(tx_size), is_rect_tx_allowed(mbmi)));
    assert(
        IMPLIES(is_rect_tx(tx_size), tx_size == max_txsize_rect_lookup[bsize]));
#endif  // CONFIG_EXT_TX && CONFIG_RECT_TX

    av1_write_token(w, av1_tx_size_tree[tx_size_cat],
                    cm->fc->tx_size_probs[tx_size_cat][tx_size_ctx],
                    &tx_size_encodings[tx_size_cat][coded_tx_size]);
  }
}

#if CONFIG_REF_MV
static void update_inter_mode_probs(AV1_COMMON *cm, aom_writer *w,
                                    FRAME_COUNTS *counts) {
  int i;
  for (i = 0; i < NEWMV_MODE_CONTEXTS; ++i)
    av1_cond_prob_diff_update(w, &cm->fc->newmv_prob[i], counts->newmv_mode[i]);
  for (i = 0; i < ZEROMV_MODE_CONTEXTS; ++i)
    av1_cond_prob_diff_update(w, &cm->fc->zeromv_prob[i],
                              counts->zeromv_mode[i]);
  for (i = 0; i < REFMV_MODE_CONTEXTS; ++i)
    av1_cond_prob_diff_update(w, &cm->fc->refmv_prob[i], counts->refmv_mode[i]);
  for (i = 0; i < DRL_MODE_CONTEXTS; ++i)
    av1_cond_prob_diff_update(w, &cm->fc->drl_prob[i], counts->drl_mode[i]);
#if CONFIG_EXT_INTER
  av1_cond_prob_diff_update(w, &cm->fc->new2mv_prob, counts->new2mv_mode);
#endif  // CONFIG_EXT_INTER
}
#endif

#if CONFIG_EXT_INTER
static void update_inter_compound_mode_probs(AV1_COMMON *cm, aom_writer *w) {
  const int savings_thresh = av1_cost_one(GROUP_DIFF_UPDATE_PROB) -
                             av1_cost_zero(GROUP_DIFF_UPDATE_PROB);
  int i;
  int savings = 0;
  int do_update = 0;
  for (i = 0; i < INTER_MODE_CONTEXTS; ++i) {
    savings += prob_diff_update_savings(
        av1_inter_compound_mode_tree, cm->fc->inter_compound_mode_probs[i],
        cm->counts.inter_compound_mode[i], INTER_COMPOUND_MODES);
  }
  do_update = savings > savings_thresh;
  aom_write(w, do_update, GROUP_DIFF_UPDATE_PROB);
  if (do_update) {
    for (i = 0; i < INTER_MODE_CONTEXTS; ++i) {
      prob_diff_update(
          av1_inter_compound_mode_tree, cm->fc->inter_compound_mode_probs[i],
          cm->counts.inter_compound_mode[i], INTER_COMPOUND_MODES, w);
    }
  }
}
#endif  // CONFIG_EXT_INTER

static int write_skip(const AV1_COMMON *cm, const MACROBLOCKD *xd,
                      int segment_id, const MODE_INFO *mi, aom_writer *w) {
  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)) {
    return 1;
  } else {
    const int skip = mi->mbmi.skip;
    aom_write(w, skip, av1_get_skip_prob(cm, xd));
    return skip;
  }
}

static void update_skip_probs(AV1_COMMON *cm, aom_writer *w,
                              FRAME_COUNTS *counts) {
  int k;

  for (k = 0; k < SKIP_CONTEXTS; ++k)
    av1_cond_prob_diff_update(w, &cm->fc->skip_probs[k], counts->skip[k]);
}

static void update_switchable_interp_probs(AV1_COMMON *cm, aom_writer *w,
                                           FRAME_COUNTS *counts) {
  int j;
  for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j)
    prob_diff_update(av1_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(AV1_COMMON *cm, aom_writer *w) {
  const int savings_thresh = av1_cost_one(GROUP_DIFF_UPDATE_PROB) -
                             av1_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(
          av1_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;
    aom_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(
            av1_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(
            av1_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;
    aom_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(
              av1_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);
      }
    }
  }
}

#else

static void update_ext_tx_probs(AV1_COMMON *cm, aom_writer *w) {
  const int savings_thresh = av1_cost_one(GROUP_DIFF_UPDATE_PROB) -
                             av1_cost_zero(GROUP_DIFF_UPDATE_PROB);
  int i, j;

  int savings = 0;
  int do_update = 0;
  for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
    for (j = 0; j < TX_TYPES; ++j)
      savings += prob_diff_update_savings(
          av1_ext_tx_tree, cm->fc->intra_ext_tx_prob[i][j],
          cm->counts.intra_ext_tx[i][j], TX_TYPES);
  }
  do_update = savings > savings_thresh;
  aom_write(w, do_update, GROUP_DIFF_UPDATE_PROB);
  if (do_update) {
    for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
      for (j = 0; j < TX_TYPES; ++j)
        prob_diff_update(av1_ext_tx_tree, cm->fc->intra_ext_tx_prob[i][j],
                         cm->counts.intra_ext_tx[i][j], TX_TYPES, w);
    }
  }
  savings = 0;
  for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
    savings +=
        prob_diff_update_savings(av1_ext_tx_tree, cm->fc->inter_ext_tx_prob[i],
                                 cm->counts.inter_ext_tx[i], TX_TYPES);
  }
  do_update = savings > savings_thresh;
  aom_write(w, do_update, GROUP_DIFF_UPDATE_PROB);
  if (do_update) {
    for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
      prob_diff_update(av1_ext_tx_tree, cm->fc->inter_ext_tx_prob[i],
                       cm->counts.inter_ext_tx[i], TX_TYPES, w);
    }
  }
}
#endif  // CONFIG_EXT_TX

static void pack_palette_tokens(aom_writer *w, const TOKENEXTRA **tp, int n,
                                int num) {
  int i;
  const TOKENEXTRA *p = *tp;

  for (i = 0; i < num; ++i) {
    av1_write_token(w, av1_palette_color_tree[n - 2], p->context_tree,
                    &palette_color_encodings[n - 2][p->token]);
    ++p;
  }

  *tp = p;
}

#if CONFIG_SUPERTX
static void update_supertx_probs(AV1_COMMON *cm, aom_writer *w) {
  const int savings_thresh = av1_cost_one(GROUP_DIFF_UPDATE_PROB) -
                             av1_cost_zero(GROUP_DIFF_UPDATE_PROB);
  int i, j;
  int savings = 0;
  int do_update = 0;
  for (i = 0; i < PARTITION_SUPERTX_CONTEXTS; ++i) {
    for (j = 1; j < TX_SIZES; ++j) {
      savings += av1_cond_prob_diff_update_savings(&cm->fc->supertx_prob[i][j],
                                                   cm->counts.supertx[i][j]);
    }
  }
  do_update = savings > savings_thresh;
  aom_write(w, do_update, GROUP_DIFF_UPDATE_PROB);
  if (do_update) {
    for (i = 0; i < PARTITION_SUPERTX_CONTEXTS; ++i) {
      for (j = 1; j < TX_SIZES; ++j) {
        av1_cond_prob_diff_update(w, &cm->fc->supertx_prob[i][j],
                                  cm->counts.supertx[i][j]);
      }
    }
  }
}
#endif  // CONFIG_SUPERTX

#if !CONFIG_ANS
static void pack_mb_tokens(aom_writer *w, const TOKENEXTRA **tp,
                           const TOKENEXTRA *const stop,
                           aom_bit_depth_t bit_depth, const TX_SIZE tx) {
  const TOKENEXTRA *p = *tp;
#if CONFIG_VAR_TX
  int count = 0;
  const int seg_eob = get_tx2d_size(tx);
#endif

  while (p < stop && p->token != EOSB_TOKEN) {
    const int t = p->token;
    const struct av1_token *const a = &av1_coef_encodings[t];
    int v = a->value;
    int n = a->len;
#if CONFIG_AOM_HIGHBITDEPTH
    const av1_extra_bit *b;
    if (bit_depth == AOM_BITS_12)
      b = &av1_extra_bits_high12[t];
    else if (bit_depth == AOM_BITS_10)
      b = &av1_extra_bits_high10[t];
    else
      b = &av1_extra_bits[t];
#else
    const av1_extra_bit *const b = &av1_extra_bits[t];
    (void)bit_depth;
#endif  // CONFIG_AOM_HIGHBITDEPTH

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

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

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

        if (t != ONE_TOKEN) {
          int len = UNCONSTRAINED_NODES - p->skip_eob_node;
          av1_write_tree(w, av1_coef_con_tree,
                         av1_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 - txsize_sqr_up_map[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 {
            aom_write(w, bb, pb[i >> 1]);
          }
          i = b->tree[i + bb];
        } while (n);
      }

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

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

  *tp = p;
}
#else
// This function serializes the tokens in forward order using a buffered ans
// coder.
static void pack_mb_tokens(struct BufAnsCoder *ans, const TOKENEXTRA **tp,
                           const TOKENEXTRA *const stop,
                           aom_bit_depth_t bit_depth, const TX_SIZE tx) {
  const TOKENEXTRA *p = *tp;
#if CONFIG_VAR_TX
  int count = 0;
  const int seg_eob = 16 << (tx << 1);
#endif  // CONFIG_VAR_TX

  while (p < stop && p->token != EOSB_TOKEN) {
    const int t = p->token;
#if CONFIG_AOM_HIGHBITDEPTH
    const av1_extra_bit *b;
    if (bit_depth == AOM_BITS_12)
      b = &av1_extra_bits_high12[t];
    else if (bit_depth == AOM_BITS_10)
      b = &av1_extra_bits_high10[t];
    else
      b = &av1_extra_bits[t];
#else
    const av1_extra_bit *const b = &av1_extra_bits[t];
    (void)bit_depth;
#endif  // CONFIG_AOM_HIGHBITDEPTH

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

    if (t != EOB_TOKEN) {
      struct rans_sym s;
      const rans_dec_lut *token_cdf = p->token_cdf;
      assert(token_cdf);
      s.cum_prob = (*token_cdf)[t - ZERO_TOKEN];
      s.prob = (*token_cdf)[t - ZERO_TOKEN + 1] - s.cum_prob;
      buf_rans_write(ans, &s);

      if (b->base_val) {
        const int e = p->extra, l = b->len;
        int skip_bits = (b->base_val == CAT6_MIN_VAL)
                            ? TX_SIZES - 1 - txsize_sqr_up_map[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 {
              buf_uabs_write(ans, bb, pb[i >> 1]);
            }
            i = b->tree[i + bb];
          } while (n);
        }

        buf_uabs_write(ans, e & 1, 128);
      }
    }
    ++p;

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

  *tp = p;
}
#endif  // !CONFIG_ANS

#if CONFIG_VAR_TX
static void pack_txb_tokens(aom_writer *w, const TOKENEXTRA **tp,
                            const TOKENEXTRA *const tok_end, MACROBLOCKD *xd,
                            MB_MODE_INFO *mbmi, int plane,
                            BLOCK_SIZE plane_bsize, aom_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];
  const int tx_row = blk_row >> (1 - pd->subsampling_y);
  const int tx_col = blk_col >> (1 - pd->subsampling_x);
  TX_SIZE plane_tx_size;
  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;

  plane_tx_size = plane ? get_uv_tx_size_impl(
                              mbmi->inter_tx_size[tx_row][tx_col], bsize, 0, 0)
                        : mbmi->inter_tx_size[tx_row][tx_col];

  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 = num_4x4_blocks_txsize_lookup[tx_size - 1];

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