bitstream.c

/*
 * Copyright (c) 2016, Alliance for Open Media. All rights reserved
 *
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
 */

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

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

#include "av1/common/cdef.h"
#if CONFIG_CFL
#include "av1/common/cfl.h"
#endif
#include "av1/common/entropy.h"
#include "av1/common/entropymode.h"
#include "av1/common/entropymv.h"
#include "av1/common/mvref_common.h"
#include "av1/common/odintrin.h"
#include "av1/common/pred_common.h"
#include "av1/common/reconinter.h"
#if CONFIG_EXT_INTRA
#include "av1/common/reconintra.h"
#endif  // CONFIG_EXT_INTRA
#include "av1/common/seg_common.h"
#include "av1/common/tile_common.h"

#if CONFIG_LV_MAP
#include "av1/encoder/encodetxb.h"
#endif  // CONFIG_LV_MAP
#include "av1/encoder/bitstream.h"
#include "av1/encoder/cost.h"
#include "av1/encoder/encodemv.h"
#include "av1/encoder/mcomp.h"
#if CONFIG_PALETTE_DELTA_ENCODING
#include "av1/encoder/palette.h"
#endif  // CONFIG_PALETTE_DELTA_ENCODING
#include "av1/encoder/segmentation.h"
#include "av1/encoder/subexp.h"
#include "av1/encoder/tokenize.h"

#define ENC_MISMATCH_DEBUG 0

static INLINE void write_uniform(aom_writer *w, int n, int v) {
  const int l = get_unsigned_bits(n);
  const 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);
  }
}

static struct av1_token interintra_mode_encodings[INTERINTRA_MODES];
static struct av1_token compound_type_encodings[COMPOUND_TYPES];
#if CONFIG_LOOP_RESTORATION
static void loop_restoration_write_sb_coeffs(const AV1_COMMON *const cm,
                                             MACROBLOCKD *xd,
                                             const RestorationUnitInfo *rui,
                                             aom_writer *const w, int plane);
#endif  // CONFIG_LOOP_RESTORATION
#if CONFIG_OBU
static void write_uncompressed_header_obu(AV1_COMP *cpi,
                                          struct aom_write_bit_buffer *wb);
#else
static void write_uncompressed_header_frame(AV1_COMP *cpi,
                                            struct aom_write_bit_buffer *wb);
#endif

static uint32_t write_compressed_header(AV1_COMP *cpi, uint8_t *data);

#if !CONFIG_OBU || CONFIG_EXT_TILE
static int remux_tiles(const AV1_COMMON *const cm, uint8_t *dst,
                       const uint32_t data_size, const uint32_t max_tile_size,
                       const uint32_t max_tile_col_size,
                       int *const tile_size_bytes,
                       int *const tile_col_size_bytes);
#endif
void av1_encode_token_init(void) {
  av1_tokens_from_tree(interintra_mode_encodings, av1_interintra_mode_tree);
  av1_tokens_from_tree(compound_type_encodings, av1_compound_type_tree);
}

static void write_intra_mode_kf(const AV1_COMMON *cm, FRAME_CONTEXT *frame_ctx,
                                const MODE_INFO *mi, const MODE_INFO *above_mi,
                                const MODE_INFO *left_mi, int block,
                                PREDICTION_MODE mode, aom_writer *w) {
#if CONFIG_INTRABC
  assert(!is_intrabc_block(&mi->mbmi));
#endif  // CONFIG_INTRABC
  aom_write_symbol(w, mode,
                   get_y_mode_cdf(frame_ctx, mi, above_mi, left_mi, block),
                   INTRA_MODES);
  (void)cm;
}

static void write_inter_mode(aom_writer *w, PREDICTION_MODE mode,
                             FRAME_CONTEXT *ec_ctx, const int16_t mode_ctx) {
  const int16_t newmv_ctx = mode_ctx & NEWMV_CTX_MASK;

  aom_write_symbol(w, mode != NEWMV, ec_ctx->newmv_cdf[newmv_ctx], 2);

  if (mode != NEWMV) {
    if (mode_ctx & (1 << ALL_ZERO_FLAG_OFFSET)) {
      assert(mode == GLOBALMV);
      return;
    }

    const int16_t zeromv_ctx =
        (mode_ctx >> GLOBALMV_OFFSET) & GLOBALMV_CTX_MASK;
    aom_write_symbol(w, mode != GLOBALMV, ec_ctx->zeromv_cdf[zeromv_ctx], 2);

    if (mode != GLOBALMV) {
      int16_t refmv_ctx = (mode_ctx >> REFMV_OFFSET) & REFMV_CTX_MASK;

      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;
      aom_write_symbol(w, mode != NEARESTMV, ec_ctx->refmv_cdf[refmv_ctx], 2);
    }
  }
}

static void write_drl_idx(FRAME_CONTEXT *ec_ctx, 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);

  const int new_mv = mbmi->mode == NEWMV || mbmi->mode == NEW_NEWMV;
  if (new_mv) {
    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_write_symbol(w, mbmi->ref_mv_idx != idx, ec_ctx->drl_cdf[drl_ctx],
                         2);
        if (mbmi->ref_mv_idx == idx) return;
      }
    }
    return;
  }

  if (have_nearmv_in_inter_mode(mbmi->mode)) {
    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_write_symbol(w, mbmi->ref_mv_idx != (idx - 1),
                         ec_ctx->drl_cdf[drl_ctx], 2);
        if (mbmi->ref_mv_idx == (idx - 1)) return;
      }
    }
    return;
  }
}

static void write_inter_compound_mode(AV1_COMMON *cm, MACROBLOCKD *xd,
                                      aom_writer *w, PREDICTION_MODE mode,
                                      const int16_t mode_ctx) {
  assert(is_inter_compound_mode(mode));
  (void)cm;
  aom_write_symbol(w, INTER_COMPOUND_OFFSET(mode),
                   xd->tile_ctx->inter_compound_mode_cdf[mode_ctx],
                   INTER_COMPOUND_MODES);
}

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 write_tx_size_vartx(const AV1_COMMON *cm, MACROBLOCKD *xd,
                                const MB_MODE_INFO *mbmi, TX_SIZE tx_size,
                                int depth, int blk_row, int blk_col,
                                aom_writer *w) {
  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
  (void)cm;
  const int tx_row = blk_row >> 1;
  const int tx_col = blk_col >> 1;
  const int max_blocks_high = max_block_high(xd, mbmi->sb_type, 0);
  const int max_blocks_wide = max_block_wide(xd, mbmi->sb_type, 0);

  int ctx = txfm_partition_context(xd->above_txfm_context + blk_col,
                                   xd->left_txfm_context + blk_row,
                                   mbmi->sb_type, tx_size);

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

  if (depth == MAX_VARTX_DEPTH) {
    txfm_partition_update(xd->above_txfm_context + blk_col,
                          xd->left_txfm_context + blk_row, tx_size, tx_size);
    return;
  }

  const int write_txfm_partition =
      tx_size == mbmi->inter_tx_size[tx_row][tx_col];
  if (write_txfm_partition) {
    aom_write_symbol(w, 0, ec_ctx->txfm_partition_cdf[ctx], 2);

    txfm_partition_update(xd->above_txfm_context + blk_col,
                          xd->left_txfm_context + blk_row, tx_size, tx_size);
    // TODO(yuec): set correct txfm partition update for qttx
  } else {
    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
    const int bsw = tx_size_wide_unit[sub_txs];
    const int bsh = tx_size_high_unit[sub_txs];

    aom_write_symbol(w, 1, ec_ctx->txfm_partition_cdf[ctx], 2);

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

    assert(bsw > 0 && bsh > 0);
    for (int row = 0; row < tx_size_high_unit[tx_size]; row += bsh)
      for (int col = 0; col < tx_size_wide_unit[tx_size]; col += bsw) {
        int offsetr = blk_row + row;
        int offsetc = blk_col + col;
        write_tx_size_vartx(cm, xd, mbmi, sub_txs, depth + 1, offsetr, offsetc,
                            w);
      }
  }
}

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;
  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
  (void)cm;
  if (block_signals_txsize(bsize)) {
    const TX_SIZE tx_size = mbmi->tx_size;
    const int tx_size_ctx = get_tx_size_context(xd);
    const int32_t tx_size_cat = intra_tx_size_cat_lookup[bsize];
    const TX_SIZE coded_tx_size = txsize_sqr_up_map[tx_size];
    const int depth = tx_size_to_depth(coded_tx_size, tx_size_cat);
    const int max_depths = tx_size_cat_to_max_depth(tx_size_cat);

    assert(coded_tx_size <= tx_size_cat + 1);
    assert(depth >= 0 && depth <= max_depths);

    assert(!is_inter_block(mbmi));
    assert(IMPLIES(is_rect_tx(tx_size), is_rect_tx_allowed(xd, mbmi)));

    aom_write_symbol(w, depth, ec_ctx->tx_size_cdf[tx_size_cat][tx_size_ctx],
                     max_depths + 1);
  }
}

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;
    const int ctx = av1_get_skip_context(xd);
    FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
    aom_write_symbol(w, skip, ec_ctx->skip_cdfs[ctx], 2);
    return skip;
  }
}

#if CONFIG_EXT_SKIP
static int write_skip_mode(const AV1_COMMON *cm, const MACROBLOCKD *xd,
                           int segment_id, const MODE_INFO *mi, aom_writer *w) {
  if (!cm->skip_mode_flag) return 0;
  if (segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP)) {
    return 0;
  }
  const int skip_mode = mi->mbmi.skip_mode;
  if (!is_comp_ref_allowed(mi->mbmi.sb_type)) {
    assert(!skip_mode);
    return 0;
  }
  const int ctx = av1_get_skip_mode_context(xd);
  aom_write_symbol(w, skip_mode, xd->tile_ctx->skip_mode_cdfs[ctx], 2);
  return skip_mode;
}
#endif  // CONFIG_EXT_SKIP

static void write_is_inter(const AV1_COMMON *cm, const MACROBLOCKD *xd,
                           int segment_id, aom_writer *w, const int is_inter) {
  if (!segfeature_active(&cm->seg, segment_id, SEG_LVL_REF_FRAME)) {
    const int ctx = av1_get_intra_inter_context(xd);
    FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
    aom_write_symbol(w, is_inter, ec_ctx->intra_inter_cdf[ctx], 2);
  }
}

static void write_motion_mode(const AV1_COMMON *cm, MACROBLOCKD *xd,
                              const MODE_INFO *mi, aom_writer *w) {
  const MB_MODE_INFO *mbmi = &mi->mbmi;

  MOTION_MODE last_motion_mode_allowed =
      motion_mode_allowed(0, cm->global_motion, xd, mi);
  switch (last_motion_mode_allowed) {
    case SIMPLE_TRANSLATION: break;
    case OBMC_CAUSAL:
      aom_write_symbol(w, mbmi->motion_mode == OBMC_CAUSAL,
                       xd->tile_ctx->obmc_cdf[mbmi->sb_type], 2);
      break;
    default:
      aom_write_symbol(w, mbmi->motion_mode,
                       xd->tile_ctx->motion_mode_cdf[mbmi->sb_type],
                       MOTION_MODES);
  }
}

static void write_delta_qindex(const AV1_COMMON *cm, const MACROBLOCKD *xd,
                               int delta_qindex, aom_writer *w) {
  int sign = delta_qindex < 0;
  int abs = sign ? -delta_qindex : delta_qindex;
  int rem_bits, thr;
  int smallval = abs < DELTA_Q_SMALL ? 1 : 0;
  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
  (void)cm;

  aom_write_symbol(w, AOMMIN(abs, DELTA_Q_SMALL), ec_ctx->delta_q_cdf,
                   DELTA_Q_PROBS + 1);

  if (!smallval) {
    rem_bits = OD_ILOG_NZ(abs - 1) - 1;
    thr = (1 << rem_bits) + 1;
    aom_write_literal(w, rem_bits - 1, 3);
    aom_write_literal(w, abs - thr, rem_bits);
  }
  if (abs > 0) {
    aom_write_bit(w, sign);
  }
}

#if CONFIG_EXT_DELTA_Q
static void write_delta_lflevel(const AV1_COMMON *cm, const MACROBLOCKD *xd,
#if CONFIG_LOOPFILTER_LEVEL
                                int lf_id,
#endif
                                int delta_lflevel, aom_writer *w) {
  int sign = delta_lflevel < 0;
  int abs = sign ? -delta_lflevel : delta_lflevel;
  int rem_bits, thr;
  int smallval = abs < DELTA_LF_SMALL ? 1 : 0;
  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;
  (void)cm;

#if CONFIG_LOOPFILTER_LEVEL
  if (cm->delta_lf_multi) {
    assert(lf_id >= 0 && lf_id < FRAME_LF_COUNT);
    aom_write_symbol(w, AOMMIN(abs, DELTA_LF_SMALL),
                     ec_ctx->delta_lf_multi_cdf[lf_id], DELTA_LF_PROBS + 1);
  } else {
    aom_write_symbol(w, AOMMIN(abs, DELTA_LF_SMALL), ec_ctx->delta_lf_cdf,
                     DELTA_LF_PROBS + 1);
  }
#else
  aom_write_symbol(w, AOMMIN(abs, DELTA_LF_SMALL), ec_ctx->delta_lf_cdf,
                   DELTA_LF_PROBS + 1);
#endif  // CONFIG_LOOPFILTER_LEVEL

  if (!smallval) {
    rem_bits = OD_ILOG_NZ(abs - 1) - 1;
    thr = (1 << rem_bits) + 1;
    aom_write_literal(w, rem_bits - 1, 3);
    aom_write_literal(w, abs - thr, rem_bits);
  }
  if (abs > 0) {
    aom_write_bit(w, sign);
  }
}
#endif  // CONFIG_EXT_DELTA_Q

static void pack_map_tokens(aom_writer *w, const TOKENEXTRA **tp, int n,
                            int num) {
  const TOKENEXTRA *p = *tp;
  write_uniform(w, n, p->token);  // The first color index.
  ++p;
  --num;
  for (int i = 0; i < num; ++i) {
    aom_write_symbol(w, p->token, p->color_map_cdf, n);
    ++p;
  }
  *tp = p;
}

#if !CONFIG_LV_MAP
static INLINE void write_coeff_extra(const aom_cdf_prob *const *cdf, int val,
                                     int n, aom_writer *w) {
  // Code the extra bits from LSB to MSB in groups of 4
  int i = 0;
  int count = 0;
  while (count < n) {
    const int size = AOMMIN(n - count, 4);
    const int mask = (1 << size) - 1;
    aom_write_cdf(w, val & mask, cdf[i++], 1 << size);
    val >>= size;
    count += size;
  }
}

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_size,
                           TOKEN_STATS *token_stats) {
  const TOKENEXTRA *p = *tp;
  int count = 0;
  const int seg_eob = av1_get_max_eob(tx_size);

  while (p < stop && p->token != EOSB_TOKEN) {
    const int token = p->token;
    const int8_t eob_val = p->eob_val;
    if (token == BLOCK_Z_TOKEN) {
      aom_write_symbol(w, 0, *p->head_cdf, HEAD_TOKENS + 1);
      p++;
      break;
      continue;
    }

    const av1_extra_bit *const extra_bits = &av1_extra_bits[token];
    if (eob_val == LAST_EOB) {
      // Just code a flag indicating whether the value is >1 or 1.
      aom_write_bit(w, token != ONE_TOKEN);
    } else {
      int comb_symb = 2 * AOMMIN(token, TWO_TOKEN) - eob_val + p->first_val;
      aom_write_symbol(w, comb_symb, *p->head_cdf, HEAD_TOKENS + p->first_val);
    }
    if (token > ONE_TOKEN) {
      aom_write_symbol(w, token - TWO_TOKEN, *p->tail_cdf, TAIL_TOKENS);
    }

    if (extra_bits->base_val) {
      const int bit_string = p->extra;
      const int bit_string_length = extra_bits->len;  // Length of extra bits to
      const int is_cat6 = (extra_bits->base_val == CAT6_MIN_VAL);
      // be written excluding
      // the sign bit.
      int skip_bits = is_cat6
                          ? (int)sizeof(av1_cat6_prob) -
                                av1_get_cat6_extrabits_size(tx_size, bit_depth)
                          : 0;

      assert(!(bit_string >> (bit_string_length - skip_bits + 1)));
      if (bit_string_length > 0)
        write_coeff_extra(extra_bits->cdf, bit_string >> 1,
                          bit_string_length - skip_bits, w);

      aom_write_bit_record(w, bit_string & 1, token_stats);
    }
    ++p;

    ++count;
    if (eob_val == EARLY_EOB || count == seg_eob) break;
  }

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

#if CONFIG_LV_MAP
static void pack_txb_tokens(aom_writer *w, AV1_COMMON *cm, MACROBLOCK *const x,
                            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, TOKEN_STATS *token_stats) {
  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;
  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);

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

  plane_tx_size =
      plane ? uv_txsize_lookup[bsize][mbmi->inter_tx_size[tx_row][tx_col]][0][0]
            : mbmi->inter_tx_size[tx_row][tx_col];

  if (tx_size == plane_tx_size
#if DISABLE_VARTX_FOR_CHROMA
      || pd->subsampling_x || pd->subsampling_y
#endif  // DISABLE_VARTX_FOR_CHROMA
      ) {
    TOKEN_STATS tmp_token_stats;
    init_token_stats(&tmp_token_stats);

    tran_low_t *tcoeff = BLOCK_OFFSET(x->mbmi_ext->tcoeff[plane], block);
    uint16_t eob = x->mbmi_ext->eobs[plane][block];
    TXB_CTX txb_ctx = { x->mbmi_ext->txb_skip_ctx[plane][block],
                        x->mbmi_ext->dc_sign_ctx[plane][block] };
    av1_write_coeffs_txb(cm, xd, w, blk_row, blk_col, block, plane, tx_size,
                         tcoeff, eob, &txb_ctx);
#if CONFIG_RD_DEBUG
    token_stats->txb_coeff_cost_map[blk_row][blk_col] = tmp_token_stats.cost;
    token_stats->cost += tmp_token_stats.cost;
#endif
  } else {
    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
    const int bsw = tx_size_wide_unit[sub_txs];
    const int bsh = tx_size_high_unit[sub_txs];

    assert(bsw > 0 && bsh > 0);

    for (int r = 0; r < tx_size_high_unit[tx_size]; r += bsh) {
      for (int c = 0; c < tx_size_wide_unit[tx_size]; c += bsw) {
        const int offsetr = blk_row + r;
        const int offsetc = blk_col + c;
        const int step = bsh * bsw;

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

        pack_txb_tokens(w, cm, x, tp, tok_end, xd, mbmi, plane, plane_bsize,
                        bit_depth, block, offsetr, offsetc, sub_txs,
                        token_stats);
        block += step;
      }
    }
  }
}
#else  // CONFIG_LV_MAP
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, TOKEN_STATS *token_stats) {
  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;
  const int max_blocks_high = max_block_high(xd, plane_bsize, plane);
  const int max_blocks_wide = max_block_wide(xd, plane_bsize, plane);

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

  plane_tx_size =
      plane ? uv_txsize_lookup[bsize][mbmi->inter_tx_size[tx_row][tx_col]][0][0]
            : mbmi->inter_tx_size[tx_row][tx_col];

  if (tx_size == plane_tx_size
#if DISABLE_VARTX_FOR_CHROMA
      || pd->subsampling_x || pd->subsampling_y
#endif  // DISABLE_VARTX_FOR_CHROMA
      ) {
    TOKEN_STATS tmp_token_stats;
    init_token_stats(&tmp_token_stats);
    pack_mb_tokens(w, tp, tok_end, bit_depth, tx_size, &tmp_token_stats);
#if CONFIG_RD_DEBUG
    token_stats->txb_coeff_cost_map[blk_row][blk_col] = tmp_token_stats.cost;
    token_stats->cost += tmp_token_stats.cost;
#endif
  } else {
    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
    const int bsw = tx_size_wide_unit[sub_txs];
    const int bsh = tx_size_high_unit[sub_txs];

    assert(bsw > 0 && bsh > 0);

    for (int r = 0; r < tx_size_high_unit[tx_size]; r += bsh) {
      for (int c = 0; c < tx_size_wide_unit[tx_size]; c += bsw) {
        const int offsetr = blk_row + r;
        const int offsetc = blk_col + c;
        const int step = bsh * bsw;

        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, offsetr, offsetc, sub_txs, token_stats);
        block += step;
      }
    }
  }
}
#endif  // CONFIG_LV_MAP

#if CONFIG_Q_SEGMENTATION
static int neg_interleave(int x, int ref, int max) {
  const int diff = x - ref;
  if (!ref) return x;
  if (ref >= (max - 1)) return -diff;
  if (2 * ref < max) {
    if (abs(diff) <= ref) {
      if (diff > 0)
        return (diff << 1) - 1;
      else
        return ((-diff) << 1);
    }
    return x;
  } else {
    if (abs(diff) < (max - ref)) {
      if (diff > 0)
        return (diff << 1) - 1;
      else
        return ((-diff) << 1);
    }
    return (max - x) - 1;
  }
}

static void write_q_segment_id(const AV1_COMMON *cm, int skip,
                               const MB_MODE_INFO *const mbmi, aom_writer *w,
                               const struct segmentation *seg,
                               struct segmentation_probs *segp,
                               BLOCK_SIZE bsize, int mi_row, int mi_col) {
  int prev_ul = 0; /* Top left segment_id */
  int prev_l = 0;  /* Current left segment_id */
  int prev_u = 0;  /* Current top segment_id */

  if (!seg->q_lvls) return;

  MODE_INFO *const mi = cm->mi + mi_row * cm->mi_stride + mi_col;
  int tinfo = mi->mbmi.boundary_info;
  int above = (!(tinfo & TILE_ABOVE_BOUNDARY)) && ((mi_row - 1) >= 0);
  int left = (!(tinfo & TILE_LEFT_BOUNDARY)) && ((mi_col - 1) >= 0);

  if (above && left)
    prev_ul =
        get_segment_id(cm, cm->q_seg_map, BLOCK_4X4, mi_row - 1, mi_col - 1);

  if (above)
    prev_u = get_segment_id(cm, cm->q_seg_map, BLOCK_4X4, mi_row - 1, mi_col);

  if (left)
    prev_l = get_segment_id(cm, cm->q_seg_map, BLOCK_4X4, mi_row, mi_col - 1);

  int cdf_num = pick_q_seg_cdf(prev_ul, prev_u, prev_l);
  int pred = pick_q_seg_pred(prev_ul, prev_u, prev_l);

  if (skip) {
    set_q_segment_id(cm, cm->q_seg_map, mbmi->sb_type, mi_row, mi_col, pred);
    return;
  }

  int coded_id = neg_interleave(mbmi->q_segment_id, pred, seg->q_lvls);

  aom_cdf_prob *pred_cdf = segp->q_seg_cdf[cdf_num];
  aom_write_symbol(w, coded_id, pred_cdf, 8);

  set_q_segment_id(cm, cm->q_seg_map, bsize, mi_row, mi_col,
                   mbmi->q_segment_id);
}
#endif

static void write_segment_id(aom_writer *w, const struct segmentation *seg,
                             struct segmentation_probs *segp, int segment_id) {
  if (seg->enabled && seg->update_map) {
    aom_write_symbol(w, segment_id, segp->tree_cdf, MAX_SEGMENTS);
  }
}

#define WRITE_REF_BIT(bname, pname) \
  aom_write_symbol(w, bname, av1_get_pred_cdf_##pname(cm, xd), 2)
#define WRITE_REF_BIT2(bname, pname) \
  aom_write_symbol(w, bname, av1_get_pred_cdf_##pname(xd), 2)

// This function encodes the reference frame
static void write_ref_frames(const AV1_COMMON *cm, const MACROBLOCKD *xd,
                             aom_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));
  }
#if CONFIG_SEGMENT_GLOBALMV
  else if (segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP) ||
           segfeature_active(&cm->seg, segment_id, SEG_LVL_GLOBALMV))
#else
  else if (segfeature_active(&cm->seg, segment_id, SEG_LVL_SKIP))
#endif
  {
    assert(!is_compound);
    assert(mbmi->ref_frame[0] == LAST_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) {
      if (is_comp_ref_allowed(mbmi->sb_type))
        aom_write_symbol(w, is_compound, av1_get_reference_mode_cdf(cm, xd), 2);
    } else {
      assert((!is_compound) == (cm->reference_mode == SINGLE_REFERENCE));
    }

    if (is_compound) {
#if CONFIG_EXT_COMP_REFS
      const COMP_REFERENCE_TYPE comp_ref_type = has_uni_comp_refs(mbmi)
                                                    ? UNIDIR_COMP_REFERENCE
                                                    : BIDIR_COMP_REFERENCE;
      aom_write_symbol(w, comp_ref_type, av1_get_comp_reference_type_cdf(xd),
                       2);

      if (comp_ref_type == UNIDIR_COMP_REFERENCE) {
        const int bit = mbmi->ref_frame[0] == BWDREF_FRAME;
        WRITE_REF_BIT2(bit, uni_comp_ref_p);

        if (!bit) {
          assert(mbmi->ref_frame[0] == LAST_FRAME);
          const int bit1 = mbmi->ref_frame[1] == LAST3_FRAME ||
                           mbmi->ref_frame[1] == GOLDEN_FRAME;
          WRITE_REF_BIT2(bit1, uni_comp_ref_p1);
          if (bit1) {
            const int bit2 = mbmi->ref_frame[1] == GOLDEN_FRAME;
            WRITE_REF_BIT2(bit2, uni_comp_ref_p2);
          }
        } else {
          assert(mbmi->ref_frame[1] == ALTREF_FRAME);
        }

        return;
      }

      assert(comp_ref_type == BIDIR_COMP_REFERENCE);
#endif  // CONFIG_EXT_COMP_REFS

      const int bit = (mbmi->ref_frame[0] == GOLDEN_FRAME ||
                       mbmi->ref_frame[0] == LAST3_FRAME);
      WRITE_REF_BIT(bit, comp_ref_p);

      if (!bit) {
        const int bit1 = mbmi->ref_frame[0] == LAST2_FRAME;
        WRITE_REF_BIT(bit1, comp_ref_p1);
      } else {
        const int bit2 = mbmi->ref_frame[0] == GOLDEN_FRAME;
        WRITE_REF_BIT(bit2, comp_ref_p2);
      }

      const int bit_bwd = mbmi->ref_frame[1] == ALTREF_FRAME;
      WRITE_REF_BIT(bit_bwd, comp_bwdref_p);

      if (!bit_bwd) {
        WRITE_REF_BIT(mbmi->ref_frame[1] == ALTREF2_FRAME, comp_bwdref_p1);
      }

    } else {
      const int bit0 = (mbmi->ref_frame[0] <= ALTREF_FRAME &&
                        mbmi->ref_frame[0] >= BWDREF_FRAME);
      WRITE_REF_BIT(bit0, single_ref_p1);

      if (bit0) {
        const int bit1 = mbmi->ref_frame[0] == ALTREF_FRAME;
        WRITE_REF_BIT(bit1, single_ref_p2);

        if (!bit1) {
          WRITE_REF_BIT(mbmi->ref_frame[0] == ALTREF2_FRAME, single_ref_p6);
        }
      } else {
        const int bit2 = (mbmi->ref_frame[0] == LAST3_FRAME ||
                          mbmi->ref_frame[0] == GOLDEN_FRAME);
        WRITE_REF_BIT(bit2, single_ref_p3);

        if (!bit2) {
          const int bit3 = mbmi->ref_frame[0] != LAST_FRAME;
          WRITE_REF_BIT(bit3, single_ref_p4);
        } else {
          const int bit4 = mbmi->ref_frame[0] != LAST3_FRAME;
          WRITE_REF_BIT(bit4, single_ref_p5);
        }
      }
    }
  }
}

#if CONFIG_FILTER_INTRA
static void write_filter_intra_mode_info(const MACROBLOCKD *xd,
                                         const MB_MODE_INFO *const mbmi,
                                         aom_writer *w) {
  if (mbmi->mode == DC_PRED && mbmi->palette_mode_info.palette_size[0] == 0 &&
      av1_filter_intra_allowed_txsize(mbmi->tx_size)) {
    aom_write_symbol(w, mbmi->filter_intra_mode_info.use_filter_intra_mode[0],
                     xd->tile_ctx->filter_intra_cdfs[mbmi->tx_size], 2);
    if (mbmi->filter_intra_mode_info.use_filter_intra_mode[0]) {
      const FILTER_INTRA_MODE mode =
          mbmi->filter_intra_mode_info.filter_intra_mode[0];
      aom_write_symbol(w, mode, xd->tile_ctx->filter_intra_mode_cdf[0],
                       FILTER_INTRA_MODES);
    }
  }
}
#endif  // CONFIG_FILTER_INTRA

#if CONFIG_EXT_INTRA
static void write_intra_angle_info(const MACROBLOCKD *xd,
                                   FRAME_CONTEXT *const ec_ctx, aom_writer *w) {
  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
  const BLOCK_SIZE bsize = mbmi->sb_type;
  if (!av1_use_angle_delta(bsize)) return;

  if (av1_is_directional_mode(mbmi->mode, bsize)) {
#if CONFIG_EXT_INTRA_MOD
    aom_write_symbol(w, mbmi->angle_delta[0] + MAX_ANGLE_DELTA,
                     ec_ctx->angle_delta_cdf[mbmi->mode - V_PRED],
                     2 * MAX_ANGLE_DELTA + 1);
#else
    (void)ec_ctx;
    write_uniform(w, 2 * MAX_ANGLE_DELTA + 1,
                  MAX_ANGLE_DELTA + mbmi->angle_delta[0]);
#endif  // CONFIG_EXT_INTRA_MOD
  }

  if (av1_is_directional_mode(get_uv_mode(mbmi->uv_mode), bsize)) {
#if CONFIG_EXT_INTRA_MOD
    aom_write_symbol(w, mbmi->angle_delta[1] + MAX_ANGLE_DELTA,
                     ec_ctx->angle_delta_cdf[mbmi->uv_mode - V_PRED],
                     2 * MAX_ANGLE_DELTA + 1);
#else
    write_uniform(w, 2 * MAX_ANGLE_DELTA + 1,
                  MAX_ANGLE_DELTA + mbmi->angle_delta[1]);
#endif
  }
}
#endif  // CONFIG_EXT_INTRA

static void write_mb_interp_filter(AV1_COMP *cpi, const MACROBLOCKD *xd,
                                   aom_writer *w) {
  AV1_COMMON *const cm = &cpi->common;
  const MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
  FRAME_CONTEXT *ec_ctx = xd->tile_ctx;

  if (!av1_is_interp_needed(xd)) {
    assert(mbmi->interp_filters ==
           av1_broadcast_interp_filter(
               av1_unswitchable_filter(cm->interp_filter)));
    return;
  }
  if (cm->interp_filter == SWITCHABLE) {
#if CONFIG_DUAL_FILTER
    int dir;
    for (dir = 0; dir < 2; ++dir) {
      if (has_subpel_mv_component(xd->mi[0], xd, dir) ||
          (mbmi->ref_frame[1] > INTRA_FRAME &&
           has_subpel_mv_component(xd->mi[0], xd, dir + 2))) {
        const int ctx = av1_get_pred_context_switchable_interp(xd, dir);
        InterpFilter filter =
            av1_extract_interp_filter(mbmi->interp_filters, dir);
        aom_write_symbol(w, filter, ec_ctx->switchable_interp_cdf[ctx],
                         SWITCHABLE_FILTERS);
        ++cpi->interp_filter_selected[0][filter];
      } else {
        assert(av1_extract_interp_filter(mbmi->interp_filters, dir) ==
               EIGHTTAP_REGULAR);
      }
    }
#else
    {
      const int ctx = av1_get_pred_context_switchable_interp(xd);
      InterpFilter filter = av1_extract_interp_filter(mbmi->interp_filters, 0);
      aom_write_symbol(w, filter, ec_ctx->switchable_interp_cdf[ctx],
                       SWITCHABLE_FILTERS);
      ++cpi->interp_filter_selected[0][filter];
    }
#endif  // CONFIG_DUAL_FILTER
  }
}

#if CONFIG_PALETTE_DELTA_ENCODING
// Transmit color values with delta encoding. Write the first value as
// literal, and the deltas between each value and the previous one. "min_val" is
// the smallest possible value of the deltas.
static void delta_encode_palette_colors(const int *colors, int num,
                                        int bit_depth, int min_val,
                                        aom_writer *w) {
  if (num <= 0) return;
  assert(colors[0] < (1 << bit_depth));
  aom_write_literal(w, colors[0], bit_depth);
  if (num == 1) return;
  int max_delta = 0;
  int deltas[PALETTE_MAX_SIZE];
  memset(deltas, 0, sizeof(deltas));
  for (int i = 1; i < num; ++i) {
    assert(colors[i] < (1 << bit_depth));
    const int delta = colors[i] - colors[i - 1];
    deltas[i - 1] = delta;
    assert(delta >= min_val);
    if (delta > max_delta) max_delta = delta;
  }
  const int min_bits = bit_depth - 3;
  int bits = AOMMAX(av1_ceil_log2(max_delta + 1 - min_val), min_bits);
  assert(bits <= bit_depth);
  int range = (1 << bit_depth) - colors[0] - min_val;
  aom_write_literal(w, bits - min_bits, 2);
  for (int i = 0; i < num - 1; ++i) {
    aom_write_literal(w, deltas[i] - min_val, bits);
    range -= deltas[i];
    bits = AOMMIN(bits, av1_ceil_log2(range));
  }
}

// Transmit luma palette color values. First signal if each color in the color
// cache is used. Those colors that are not in the cache are transmitted with
// delta encoding.
static void write_palette_colors_y(const MACROBLOCKD *const xd,
                                   const PALETTE_MODE_INFO *const pmi,
                                   int bit_depth, aom_writer *w) {
  const int n = pmi->palette_size[0];
  uint16_t color_cache[2 * PALETTE_MAX_SIZE];
  const int n_cache = av1_get_palette_cache(xd, 0, color_cache);
  int out_cache_colors[PALETTE_MAX_SIZE];
  uint8_t cache_color_found[2 * PALETTE_MAX_SIZE];
  const int n_out_cache =
      av1_index_color_cache(color_cache, n_cache, pmi->palette_colors, n,
                            cache_color_found, out_cache_colors);
  int n_in_cache = 0;
  for (int i = 0; i < n_cache && n_in_cache < n; ++i) {
    const int found = cache_color_found[i];
    aom_write_bit(w, found);
    n_in_cache += found;
  }
  assert(n_in_cache + n_out_cache == n);
  delta_encode_palette_colors(out_cache_colors, n_out_cache, bit_depth, 1, w);
}

// Write chroma palette color values. U channel is handled similarly to the luma
// channel. For v channel, either use delta encoding or transmit raw values
// directly, whichever costs less.
static void write_palette_colors_uv(const MACROBLOCKD *const xd,
                                    const PALETTE_MODE_INFO *const pmi,
                                    int bit_depth, aom_writer *w) {
  const int n = pmi->palette_size[1];
  const uint16_t *colors_u = pmi->palette_colors + PALETTE_MAX_SIZE;
  const uint16_t *colors_v = pmi->palette_colors + 2 * PALETTE_MAX_SIZE;
  // U channel colors.
  uint16_t color_cache[2 * PALETTE_MAX_SIZE];