decodeframe.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 <stdlib.h>  // qsort()

#include "./aom_config.h"
#include "./aom_dsp_rtcd.h"
#include "./aom_scale_rtcd.h"
#include "./av1_rtcd.h"

#include "aom/aom_codec.h"
#include "aom_dsp/aom_dsp_common.h"
#include "aom_dsp/bitreader.h"
#include "aom_dsp/bitreader_buffer.h"
#include "aom_mem/aom_mem.h"
#include "aom_ports/mem.h"
#include "aom_ports/mem_ops.h"
#include "aom_scale/aom_scale.h"
#include "aom_util/aom_thread.h"

#include "av1/common/alloccommon.h"
#if CONFIG_CLPF
#include "aom/aom_image.h"
#include "av1/common/clpf.h"
#endif
#include "av1/common/common.h"
#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/idct.h"
#include "av1/common/pred_common.h"
#include "av1/common/quant_common.h"
#include "av1/common/reconinter.h"
#include "av1/common/reconintra.h"
#include "av1/common/seg_common.h"
#include "av1/common/thread_common.h"
#include "av1/common/tile_common.h"

#include "av1/decoder/decodeframe.h"
#include "av1/decoder/decodemv.h"
#include "av1/decoder/decoder.h"
#include "av1/decoder/detokenize.h"
#include "av1/decoder/dsubexp.h"

#define MAX_AV1_HEADER_SIZE 80
#define ACCT_STR __func__

#if CONFIG_PVQ
#include "av1/decoder/pvq_decoder.h"
#include "av1/encoder/encodemb.h"

#include "aom_dsp/entdec.h"
#include "av1/common/partition.h"
#include "av1/decoder/decint.h"
#include "av1/encoder/hybrid_fwd_txfm.h"
#endif

static struct aom_read_bit_buffer *init_read_bit_buffer(
    AV1Decoder *pbi, struct aom_read_bit_buffer *rb, const uint8_t *data,
    const uint8_t *data_end, uint8_t clear_data[MAX_AV1_HEADER_SIZE]);
static int read_compressed_header(AV1Decoder *pbi, const uint8_t *data,
                                  size_t partition_size);
static size_t read_uncompressed_header(AV1Decoder *pbi,
                                       struct aom_read_bit_buffer *rb);

static int is_compound_reference_allowed(const AV1_COMMON *cm) {
  int i;
  if (frame_is_intra_only(cm)) return 0;
  for (i = 1; i < INTER_REFS_PER_FRAME; ++i)
    if (cm->ref_frame_sign_bias[i + 1] != cm->ref_frame_sign_bias[1]) return 1;

  return 0;
}

static void setup_compound_reference_mode(AV1_COMMON *cm) {
#if CONFIG_EXT_REFS
  cm->comp_fwd_ref[0] = LAST_FRAME;
  cm->comp_fwd_ref[1] = LAST2_FRAME;
  cm->comp_fwd_ref[2] = LAST3_FRAME;
  cm->comp_fwd_ref[3] = GOLDEN_FRAME;

  cm->comp_bwd_ref[0] = BWDREF_FRAME;
  cm->comp_bwd_ref[1] = ALTREF_FRAME;
#else
  if (cm->ref_frame_sign_bias[LAST_FRAME] ==
      cm->ref_frame_sign_bias[GOLDEN_FRAME]) {
    cm->comp_fixed_ref = ALTREF_FRAME;
    cm->comp_var_ref[0] = LAST_FRAME;
    cm->comp_var_ref[1] = GOLDEN_FRAME;
  } else if (cm->ref_frame_sign_bias[LAST_FRAME] ==
             cm->ref_frame_sign_bias[ALTREF_FRAME]) {
    cm->comp_fixed_ref = GOLDEN_FRAME;
    cm->comp_var_ref[0] = LAST_FRAME;
    cm->comp_var_ref[1] = ALTREF_FRAME;
  } else {
    cm->comp_fixed_ref = LAST_FRAME;
    cm->comp_var_ref[0] = GOLDEN_FRAME;
    cm->comp_var_ref[1] = ALTREF_FRAME;
  }
#endif  // CONFIG_EXT_REFS
}

static int read_is_valid(const uint8_t *start, size_t len, const uint8_t *end) {
  return len != 0 && len <= (size_t)(end - start);
}

static int decode_unsigned_max(struct aom_read_bit_buffer *rb, int max) {
  const int data = aom_rb_read_literal(rb, get_unsigned_bits(max));
  return data > max ? max : data;
}

static TX_MODE read_tx_mode(struct aom_read_bit_buffer *rb) {
#if CONFIG_TX64X64
  TX_MODE tx_mode =
      aom_rb_read_bit(rb) ? TX_MODE_SELECT : aom_rb_read_literal(rb, 2);
  if (tx_mode == ALLOW_32X32) tx_mode += aom_rb_read_bit(rb);
  return tx_mode;
#else
  return aom_rb_read_bit(rb) ? TX_MODE_SELECT : aom_rb_read_literal(rb, 2);
#endif  // CONFIG_TX64X64
}

static void read_tx_size_probs(FRAME_CONTEXT *fc, aom_reader *r) {
  int i, j, k;
  for (i = 0; i < MAX_TX_DEPTH; ++i)
    for (j = 0; j < TX_SIZE_CONTEXTS; ++j)
      for (k = 0; k < i + 1; ++k)
        av1_diff_update_prob(r, &fc->tx_size_probs[i][j][k], ACCT_STR);
}

#if !CONFIG_EC_ADAPT
static void read_switchable_interp_probs(FRAME_CONTEXT *fc, aom_reader *r) {
  int i, j;
  for (j = 0; j < SWITCHABLE_FILTER_CONTEXTS; ++j) {
    for (i = 0; i < SWITCHABLE_FILTERS - 1; ++i)
      av1_diff_update_prob(r, &fc->switchable_interp_prob[j][i], ACCT_STR);
  }
}
#endif

static void read_inter_mode_probs(FRAME_CONTEXT *fc, aom_reader *r) {
#if CONFIG_REF_MV
  int i;
  for (i = 0; i < NEWMV_MODE_CONTEXTS; ++i)
    av1_diff_update_prob(r, &fc->newmv_prob[i], ACCT_STR);
  for (i = 0; i < ZEROMV_MODE_CONTEXTS; ++i)
    av1_diff_update_prob(r, &fc->zeromv_prob[i], ACCT_STR);
  for (i = 0; i < REFMV_MODE_CONTEXTS; ++i)
    av1_diff_update_prob(r, &fc->refmv_prob[i], ACCT_STR);
  for (i = 0; i < DRL_MODE_CONTEXTS; ++i)
    av1_diff_update_prob(r, &fc->drl_prob[i], ACCT_STR);
#if CONFIG_EXT_INTER
  av1_diff_update_prob(r, &fc->new2mv_prob, ACCT_STR);
#endif  // CONFIG_EXT_INTER
#else
#if !CONFIG_EC_ADAPT
  int i, j;
  for (i = 0; i < INTER_MODE_CONTEXTS; ++i) {
    for (j = 0; j < INTER_MODES - 1; ++j)
      av1_diff_update_prob(r, &fc->inter_mode_probs[i][j], ACCT_STR);
  }
#else
  (void)fc;
  (void)r;
#endif
#endif
}

#if CONFIG_EXT_INTER
static void read_inter_compound_mode_probs(FRAME_CONTEXT *fc, aom_reader *r) {
  int i, j;
  if (aom_read(r, GROUP_DIFF_UPDATE_PROB, ACCT_STR)) {
    for (j = 0; j < INTER_MODE_CONTEXTS; ++j) {
      for (i = 0; i < INTER_COMPOUND_MODES - 1; ++i) {
        av1_diff_update_prob(r, &fc->inter_compound_mode_probs[j][i], ACCT_STR);
      }
    }
  }
}
#endif  // CONFIG_EXT_INTER
#if !CONFIG_EC_ADAPT
#if !CONFIG_EXT_TX
static void read_ext_tx_probs(FRAME_CONTEXT *fc, aom_reader *r) {
  int i, j, k;
  if (aom_read(r, GROUP_DIFF_UPDATE_PROB, ACCT_STR)) {
    for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
      for (j = 0; j < TX_TYPES; ++j) {
        for (k = 0; k < TX_TYPES - 1; ++k)
          av1_diff_update_prob(r, &fc->intra_ext_tx_prob[i][j][k], ACCT_STR);
      }
    }
  }
  if (aom_read(r, GROUP_DIFF_UPDATE_PROB, ACCT_STR)) {
    for (i = TX_4X4; i < EXT_TX_SIZES; ++i) {
      for (k = 0; k < TX_TYPES - 1; ++k)
        av1_diff_update_prob(r, &fc->inter_ext_tx_prob[i][k], ACCT_STR);
    }
  }
}
#endif
#endif

static REFERENCE_MODE read_frame_reference_mode(
    const AV1_COMMON *cm, struct aom_read_bit_buffer *rb) {
  if (is_compound_reference_allowed(cm)) {
    return aom_rb_read_bit(rb)
               ? REFERENCE_MODE_SELECT
               : (aom_rb_read_bit(rb) ? COMPOUND_REFERENCE : SINGLE_REFERENCE);
  } else {
    return SINGLE_REFERENCE;
  }
}

static void read_frame_reference_mode_probs(AV1_COMMON *cm, aom_reader *r) {
  FRAME_CONTEXT *const fc = cm->fc;
  int i, j;

  if (cm->reference_mode == REFERENCE_MODE_SELECT)
    for (i = 0; i < COMP_INTER_CONTEXTS; ++i)
      av1_diff_update_prob(r, &fc->comp_inter_prob[i], ACCT_STR);

  if (cm->reference_mode != COMPOUND_REFERENCE) {
    for (i = 0; i < REF_CONTEXTS; ++i) {
      for (j = 0; j < (SINGLE_REFS - 1); ++j) {
        av1_diff_update_prob(r, &fc->single_ref_prob[i][j], ACCT_STR);
      }
    }
  }

  if (cm->reference_mode != SINGLE_REFERENCE) {
    for (i = 0; i < REF_CONTEXTS; ++i) {
#if CONFIG_EXT_REFS
      for (j = 0; j < (FWD_REFS - 1); ++j)
        av1_diff_update_prob(r, &fc->comp_ref_prob[i][j], ACCT_STR);
      for (j = 0; j < (BWD_REFS - 1); ++j)
        av1_diff_update_prob(r, &fc->comp_bwdref_prob[i][j], ACCT_STR);
#else
      for (j = 0; j < (COMP_REFS - 1); ++j)
        av1_diff_update_prob(r, &fc->comp_ref_prob[i][j], ACCT_STR);
#endif  // CONFIG_EXT_REFS
    }
  }
}

static void update_mv_probs(aom_prob *p, int n, aom_reader *r) {
  int i;
  for (i = 0; i < n; ++i) av1_diff_update_prob(r, &p[i], ACCT_STR);
}

static void read_mv_probs(nmv_context *ctx, int allow_hp, aom_reader *r) {
  int i;

#if !CONFIG_EC_ADAPT
  int j;
  update_mv_probs(ctx->joints, MV_JOINTS - 1, r);

  for (i = 0; i < 2; ++i) {
    nmv_component *const comp_ctx = &ctx->comps[i];
    update_mv_probs(&comp_ctx->sign, 1, r);
    update_mv_probs(comp_ctx->classes, MV_CLASSES - 1, r);
    update_mv_probs(comp_ctx->class0, CLASS0_SIZE - 1, r);
    update_mv_probs(comp_ctx->bits, MV_OFFSET_BITS, r);
  }
  for (i = 0; i < 2; ++i) {
    nmv_component *const comp_ctx = &ctx->comps[i];
    for (j = 0; j < CLASS0_SIZE; ++j) {
      update_mv_probs(comp_ctx->class0_fp[j], MV_FP_SIZE - 1, r);
    }
    update_mv_probs(comp_ctx->fp, MV_FP_SIZE - 1, r);
  }
#endif  // !CONFIG_EC_ADAPT

  if (allow_hp) {
    for (i = 0; i < 2; ++i) {
      nmv_component *const comp_ctx = &ctx->comps[i];
      update_mv_probs(&comp_ctx->class0_hp, 1, r);
      update_mv_probs(&comp_ctx->hp, 1, r);
    }
  }
}

static void inverse_transform_block(MACROBLOCKD *xd, int plane,
                                    const TX_TYPE tx_type,
                                    const TX_SIZE tx_size, uint8_t *dst,
                                    int stride, int16_t scan_line, int eob) {
  struct macroblockd_plane *const pd = &xd->plane[plane];
  tran_low_t *const dqcoeff = pd->dqcoeff;
  INV_TXFM_PARAM inv_txfm_param;
  inv_txfm_param.tx_type = tx_type;
  inv_txfm_param.tx_size = tx_size;
  inv_txfm_param.eob = eob;
  inv_txfm_param.lossless = xd->lossless[xd->mi[0]->mbmi.segment_id];

#if CONFIG_AOM_HIGHBITDEPTH
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    inv_txfm_param.bd = xd->bd;
    highbd_inv_txfm_add(dqcoeff, dst, stride, &inv_txfm_param);
  } else {
#endif  // CONFIG_AOM_HIGHBITDEPTH
    inv_txfm_add(dqcoeff, dst, stride, &inv_txfm_param);
#if CONFIG_AOM_HIGHBITDEPTH
  }
#endif  // CONFIG_AOM_HIGHBITDEPTH
  memset(dqcoeff, 0, (scan_line + 1) * sizeof(dqcoeff[0]));
}

#if CONFIG_PVQ
static int av1_pvq_decode_helper(od_dec_ctx *dec, int16_t *ref_coeff,
                                 int16_t *dqcoeff, int16_t *quant, int pli,
                                 int bs, TX_TYPE tx_type, int xdec,
                                 int ac_dc_coded) {
  unsigned int flags;  // used for daala's stream analyzer.
  int off;
  const int is_keyframe = 0;
  const int has_dc_skip = 1;
  int quant_shift = get_tx_scale(bs);
  // DC quantizer for PVQ
  int pvq_dc_quant;
  int lossless = (quant[0] == 0);
  const int blk_size = tx_size_wide[bs];
  int eob = 0;
  int i;
  // TODO(yushin) : To enable activity masking,
  // int use_activity_masking = dec->use_activity_masking;
  int use_activity_masking = 0;

  DECLARE_ALIGNED(16, int16_t, dqcoeff_pvq[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);
  DECLARE_ALIGNED(16, int16_t, ref_coeff_pvq[OD_TXSIZE_MAX * OD_TXSIZE_MAX]);

  od_coeff ref_int32[OD_TXSIZE_MAX * OD_TXSIZE_MAX];
  od_coeff out_int32[OD_TXSIZE_MAX * OD_TXSIZE_MAX];

  od_raster_to_coding_order(ref_coeff_pvq, blk_size, tx_type, ref_coeff,
                            blk_size);

  if (lossless)
    pvq_dc_quant = 1;
  else {
    // TODO(yushin): Enable this for activity masking,
    // when pvq_qm_q4 is available in AOM.
    // pvq_dc_quant = OD_MAXI(1, quant*
    // dec->state.pvq_qm_q4[pli][od_qm_get_index(bs, 0)] >> 4);
    pvq_dc_quant = OD_MAXI(1, quant[0] >> quant_shift);
  }

  off = od_qm_offset(bs, xdec);

  // copy int16 inputs to int32
  for (i = 0; i < blk_size * blk_size; i++) ref_int32[i] = ref_coeff_pvq[i];

  od_pvq_decode(dec, ref_int32, out_int32, (int)quant[1] >> quant_shift, pli,
                bs, OD_PVQ_BETA[use_activity_masking][pli][bs],
                OD_ROBUST_STREAM, is_keyframe, &flags, ac_dc_coded,
                dec->state.qm + off, dec->state.qm_inv + off);

  // copy int32 result back to int16
  for (i = 0; i < blk_size * blk_size; i++) dqcoeff_pvq[i] = out_int32[i];

  if (!has_dc_skip || dqcoeff_pvq[0]) {
    dqcoeff_pvq[0] =
        has_dc_skip + generic_decode(dec->ec, &dec->state.adapt.model_dc[pli],
                                     -1, &dec->state.adapt.ex_dc[pli][bs][0], 2,
                                     "dc:mag");
    if (dqcoeff_pvq[0])
      dqcoeff_pvq[0] *= od_ec_dec_bits(dec->ec, 1, "dc:sign") ? -1 : 1;
  }
  dqcoeff_pvq[0] = dqcoeff_pvq[0] * pvq_dc_quant + ref_coeff_pvq[0];

  od_coding_order_to_raster(dqcoeff, blk_size, tx_type, dqcoeff_pvq, blk_size);

  eob = blk_size * blk_size;

  return eob;
}

static int av1_pvq_decode_helper2(MACROBLOCKD *const xd,
                                  MB_MODE_INFO *const mbmi, int plane, int row,
                                  int col, TX_SIZE tx_size, TX_TYPE tx_type) {
  struct macroblockd_plane *const pd = &xd->plane[plane];
  // transform block size in pixels
  int tx_blk_size = tx_size_wide[tx_size];
  int i, j;
  tran_low_t *pvq_ref_coeff = pd->pvq_ref_coeff;
  const int diff_stride = tx_blk_size;
  int16_t *pred = pd->pred;
  tran_low_t *const dqcoeff = pd->dqcoeff;
  int ac_dc_coded;  // bit0: DC coded, bit1 : AC coded
  uint8_t *dst;
  int eob;

  eob = 0;
  dst = &pd->dst.buf[4 * row * pd->dst.stride + 4 * col];

  // decode ac/dc coded flag. bit0: DC coded, bit1 : AC coded
  // NOTE : we don't use 5 symbols for luma here in aom codebase,
  // since block partition is taken care of by aom.
  // So, only AC/DC skip info is coded
  ac_dc_coded = od_decode_cdf_adapt(
      xd->daala_dec.ec,
      xd->daala_dec.state.adapt.skip_cdf[2 * tx_size + (plane != 0)], 4,
      xd->daala_dec.state.adapt.skip_increment, "skip");

  if (ac_dc_coded) {
    int xdec = pd->subsampling_x;
    int seg_id = mbmi->segment_id;
    int16_t *quant;
    FWD_TXFM_PARAM fwd_txfm_param;
    // ToDo(yaowu): correct this with optimal number from decoding process.
    const int max_scan_line = tx_size_2d[tx_size];

    for (j = 0; j < tx_blk_size; j++)
      for (i = 0; i < tx_blk_size; i++) {
        pred[diff_stride * j + i] = dst[pd->dst.stride * j + i];
      }

    fwd_txfm_param.tx_type = tx_type;
    fwd_txfm_param.tx_size = tx_size;
    fwd_txfm_param.fwd_txfm_opt = FWD_TXFM_OPT_NORMAL;
    fwd_txfm_param.rd_transform = 0;
    fwd_txfm_param.lossless = xd->lossless[seg_id];

    fwd_txfm(pred, pvq_ref_coeff, diff_stride, &fwd_txfm_param);

    quant = &pd->seg_dequant[seg_id][0];  // aom's quantizer

    eob = av1_pvq_decode_helper(&xd->daala_dec, pvq_ref_coeff, dqcoeff, quant,
                                plane, tx_size, tx_type, xdec, ac_dc_coded);

    // Since av1 does not have separate inverse transform
    // but also contains adding to predicted image,
    // pass blank dummy image to av1_inv_txfm_add_*x*(), i.e. set dst as zeros
    for (j = 0; j < tx_blk_size; j++)
      for (i = 0; i < tx_blk_size; i++) dst[j * pd->dst.stride + i] = 0;

    inverse_transform_block(xd, plane, tx_type, tx_size, dst, pd->dst.stride,
                            max_scan_line, eob);
  }

  return eob;
}
#endif

static void predict_and_reconstruct_intra_block(AV1_COMMON *cm,
                                                MACROBLOCKD *const xd,
#if CONFIG_ANS
                                                struct AnsDecoder *const r,
#else
                                                aom_reader *r,
#endif  // CONFIG_ANS
                                                MB_MODE_INFO *const mbmi,
                                                int plane, int row, int col,
                                                TX_SIZE tx_size) {
  struct macroblockd_plane *const pd = &xd->plane[plane];
  PREDICTION_MODE mode = (plane == 0) ? mbmi->mode : mbmi->uv_mode;
  PLANE_TYPE plane_type = (plane == 0) ? PLANE_TYPE_Y : PLANE_TYPE_UV;
  uint8_t *dst;
  int block_idx = (row << 1) + col;
#if CONFIG_PVQ
  (void)cm;
  (void)r;
#endif
  dst = &pd->dst.buf[4 * row * pd->dst.stride + 4 * col];

  if (mbmi->sb_type < BLOCK_8X8)
    if (plane == 0) mode = xd->mi[0]->bmi[(row << 1) + col].as_mode;

  av1_predict_intra_block(xd, pd->width, pd->height, tx_size, mode, dst,
                          pd->dst.stride, dst, pd->dst.stride, col, row, plane);

  if (!mbmi->skip) {
    TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, tx_size);
#if !CONFIG_PVQ
    const SCAN_ORDER *scan_order = get_scan(cm, tx_size, tx_type, 0);
    int16_t max_scan_line = 0;
    const int eob =
        av1_decode_block_tokens(xd, plane, scan_order, col, row, tx_size,
                                tx_type, &max_scan_line, r, mbmi->segment_id);
#if CONFIG_ADAPT_SCAN
    av1_update_scan_count_facade(cm, tx_size, tx_type, pd->dqcoeff, eob);
#endif
    if (eob)
      inverse_transform_block(xd, plane, tx_type, tx_size, dst, pd->dst.stride,
                              max_scan_line, eob);
#else
    av1_pvq_decode_helper2(xd, mbmi, plane, row, col, tx_size, tx_type);
#endif
  }
}

#if CONFIG_VAR_TX
static void decode_reconstruct_tx(AV1_COMMON *cm, MACROBLOCKD *const xd,
                                  aom_reader *r, MB_MODE_INFO *const mbmi,
                                  int plane, BLOCK_SIZE plane_bsize,
                                  int blk_row, int blk_col, TX_SIZE tx_size,
                                  int *eob_total) {
  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);
  const TX_SIZE 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];
  // Scale to match transform block unit.
  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;

  if (tx_size == plane_tx_size) {
    PLANE_TYPE plane_type = (plane == 0) ? PLANE_TYPE_Y : PLANE_TYPE_UV;
    int block_idx = (blk_row << 1) + blk_col;
    TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, plane_tx_size);
    const SCAN_ORDER *sc = get_scan(cm, plane_tx_size, tx_type, 1);
    int16_t max_scan_line = 0;
    const int eob =
        av1_decode_block_tokens(xd, plane, sc, blk_col, blk_row, plane_tx_size,
                                tx_type, &max_scan_line, r, mbmi->segment_id);
    inverse_transform_block(
        xd, plane, tx_type, plane_tx_size,
        &pd->dst.buf[4 * blk_row * pd->dst.stride + 4 * blk_col],
        pd->dst.stride, max_scan_line, eob);
    *eob_total += eob;
  } else {
    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
    const int bsl = tx_size_wide_unit[sub_txs];
    int i;

    assert(bsl > 0);

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

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

      decode_reconstruct_tx(cm, xd, r, mbmi, plane, plane_bsize, offsetr,
                            offsetc, sub_txs, eob_total);
    }
  }
}
#endif  // CONFIG_VAR_TX

#if !CONFIG_VAR_TX || CONFIG_SUPERTX || (CONFIG_EXT_TX && CONFIG_RECT_TX)
static int reconstruct_inter_block(AV1_COMMON *cm, MACROBLOCKD *const xd,
#if CONFIG_ANS
                                   struct AnsDecoder *const r,
#else
                                   aom_reader *r,
#endif
                                   int segment_id, int plane, int row, int col,
                                   TX_SIZE tx_size) {
  PLANE_TYPE plane_type = (plane == 0) ? PLANE_TYPE_Y : PLANE_TYPE_UV;
  int block_idx = (row << 1) + col;
  TX_TYPE tx_type = get_tx_type(plane_type, xd, block_idx, tx_size);
#if CONFIG_PVQ
  int eob;
  (void)cm;
  (void)r;
  (void)segment_id;
#else
  struct macroblockd_plane *const pd = &xd->plane[plane];
#endif

#if !CONFIG_PVQ
  const SCAN_ORDER *scan_order = get_scan(cm, tx_size, tx_type, 1);
  int16_t max_scan_line = 0;
  const int eob =
      av1_decode_block_tokens(xd, plane, scan_order, col, row, tx_size, tx_type,
                              &max_scan_line, r, segment_id);
#if CONFIG_ADAPT_SCAN
  av1_update_scan_count_facade(cm, tx_size, tx_type, pd->dqcoeff, eob);
#endif
  if (eob)
    inverse_transform_block(xd, plane, tx_type, tx_size,
                            &pd->dst.buf[4 * row * pd->dst.stride + 4 * col],
                            pd->dst.stride, max_scan_line, eob);
#else
  eob = av1_pvq_decode_helper2(xd, &xd->mi[0]->mbmi, plane, row, col, tx_size,
                               tx_type);
#endif
  return eob;
}
#endif  // !CONFIG_VAR_TX || CONFIG_SUPER_TX

static INLINE void dec_reset_skip_context(MACROBLOCKD *xd) {
  int i;
  for (i = 0; i < MAX_MB_PLANE; i++) {
    struct macroblockd_plane *const pd = &xd->plane[i];
    memset(pd->above_context, 0, sizeof(ENTROPY_CONTEXT) * pd->n4_w);
    memset(pd->left_context, 0, sizeof(ENTROPY_CONTEXT) * pd->n4_h);
  }
}

static MB_MODE_INFO *set_offsets(AV1_COMMON *const cm, MACROBLOCKD *const xd,
                                 BLOCK_SIZE bsize, int mi_row, int mi_col,
                                 int bw, int bh, int x_mis, int y_mis, int bwl,
                                 int bhl) {
  const int offset = mi_row * cm->mi_stride + mi_col;
  int x, y;
  const TileInfo *const tile = &xd->tile;

  xd->mi = cm->mi_grid_visible + offset;
  xd->mi[0] = &cm->mi[offset];
  // TODO(slavarnway): Generate sb_type based on bwl and bhl, instead of
  // passing bsize from decode_partition().
  xd->mi[0]->mbmi.sb_type = bsize;
#if CONFIG_RD_DEBUG
  xd->mi[0]->mbmi.mi_row = mi_row;
  xd->mi[0]->mbmi.mi_col = mi_col;
#endif
  for (y = 0; y < y_mis; ++y)
    for (x = !y; x < x_mis; ++x) xd->mi[y * cm->mi_stride + x] = xd->mi[0];

  set_plane_n4(xd, bw, bh, bwl, bhl);
  set_skip_context(xd, mi_row, mi_col);

#if CONFIG_VAR_TX
  xd->max_tx_size = max_txsize_lookup[bsize];
#endif

  // Distance of Mb to the various image edges. These are specified to 8th pel
  // as they are always compared to values that are in 1/8th pel units
  set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw, cm->mi_rows, cm->mi_cols);

  av1_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
  return &xd->mi[0]->mbmi;
}

#if CONFIG_SUPERTX
static MB_MODE_INFO *set_offsets_extend(AV1_COMMON *const cm,
                                        MACROBLOCKD *const xd,
                                        const TileInfo *const tile,
                                        BLOCK_SIZE bsize_pred, int mi_row_pred,
                                        int mi_col_pred, int mi_row_ori,
                                        int mi_col_ori) {
  // Used in supertx
  // (mi_row_ori, mi_col_ori): location for mv
  // (mi_row_pred, mi_col_pred, bsize_pred): region to predict
  const int bw = num_8x8_blocks_wide_lookup[bsize_pred];
  const int bh = num_8x8_blocks_high_lookup[bsize_pred];
  const int offset = mi_row_ori * cm->mi_stride + mi_col_ori;
  const int bwl = b_width_log2_lookup[bsize_pred];
  const int bhl = b_height_log2_lookup[bsize_pred];
  xd->mi = cm->mi_grid_visible + offset;
  xd->mi[0] = cm->mi + offset;
  set_mi_row_col(xd, tile, mi_row_pred, bh, mi_col_pred, bw, cm->mi_rows,
                 cm->mi_cols);

  xd->up_available = (mi_row_ori > tile->mi_row_start);
  xd->left_available = (mi_col_ori > tile->mi_col_start);

  set_plane_n4(xd, bw, bh, bwl, bhl);

  return &xd->mi[0]->mbmi;
}

static MB_MODE_INFO *set_mb_offsets(AV1_COMMON *const cm, MACROBLOCKD *const xd,
                                    BLOCK_SIZE bsize, int mi_row, int mi_col,
                                    int bw, int bh, int x_mis, int y_mis) {
  const int offset = mi_row * cm->mi_stride + mi_col;
  const TileInfo *const tile = &xd->tile;
  int x, y;

  xd->mi = cm->mi_grid_visible + offset;
  xd->mi[0] = cm->mi + offset;
  xd->mi[0]->mbmi.sb_type = bsize;
  for (y = 0; y < y_mis; ++y)
    for (x = !y; x < x_mis; ++x) xd->mi[y * cm->mi_stride + x] = xd->mi[0];

  set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw, cm->mi_rows, cm->mi_cols);
  return &xd->mi[0]->mbmi;
}

static void set_offsets_topblock(AV1_COMMON *const cm, MACROBLOCKD *const xd,
                                 const TileInfo *const tile, BLOCK_SIZE bsize,
                                 int mi_row, int mi_col) {
  const int bw = num_8x8_blocks_wide_lookup[bsize];
  const int bh = num_8x8_blocks_high_lookup[bsize];
  const int offset = mi_row * cm->mi_stride + mi_col;
  const int bwl = b_width_log2_lookup[bsize];
  const int bhl = b_height_log2_lookup[bsize];

  xd->mi = cm->mi_grid_visible + offset;
  xd->mi[0] = cm->mi + offset;

  set_plane_n4(xd, bw, bh, bwl, bhl);

  set_mi_row_col(xd, tile, mi_row, bh, mi_col, bw, cm->mi_rows, cm->mi_cols);

  av1_setup_dst_planes(xd->plane, get_frame_new_buffer(cm), mi_row, mi_col);
}

static void set_param_topblock(AV1_COMMON *const cm, MACROBLOCKD *const xd,
                               BLOCK_SIZE bsize, int mi_row, int mi_col,
                               int txfm, int skip) {
  const int bw = num_8x8_blocks_wide_lookup[bsize];
  const int bh = num_8x8_blocks_high_lookup[bsize];
  const int x_mis = AOMMIN(bw, cm->mi_cols - mi_col);
  const int y_mis = AOMMIN(bh, cm->mi_rows - mi_row);
  const int offset = mi_row * cm->mi_stride + mi_col;
  int x, y;

  xd->mi = cm->mi_grid_visible + offset;
  xd->mi[0] = cm->mi + offset;

  for (y = 0; y < y_mis; ++y)
    for (x = 0; x < x_mis; ++x) {
      xd->mi[y * cm->mi_stride + x]->mbmi.skip = skip;
      xd->mi[y * cm->mi_stride + x]->mbmi.tx_type = txfm;
    }
#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 & MAX_MIB_MASK);
  set_txfm_ctxs(xd->mi[0]->mbmi.tx_size, bw, bh, skip, xd);
#endif
}

static void set_ref(AV1_COMMON *const cm, MACROBLOCKD *const xd, int idx,
                    int mi_row, int mi_col) {
  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
  RefBuffer *ref_buffer = &cm->frame_refs[mbmi->ref_frame[idx] - LAST_FRAME];
  xd->block_refs[idx] = ref_buffer;
  if (!av1_is_valid_scale(&ref_buffer->sf))
    aom_internal_error(&cm->error, AOM_CODEC_UNSUP_BITSTREAM,
                       "Invalid scale factors");
  av1_setup_pre_planes(xd, idx, ref_buffer->buf, mi_row, mi_col,
                       &ref_buffer->sf);
  xd->corrupted |= ref_buffer->buf->corrupted;
}

static void dec_predict_b_extend(
    AV1Decoder *const pbi, MACROBLOCKD *const xd, const TileInfo *const tile,
    int block, int mi_row_ori, int mi_col_ori, int mi_row_pred, int mi_col_pred,
    int mi_row_top, int mi_col_top, uint8_t *dst_buf[3], int dst_stride[3],
    BLOCK_SIZE bsize_top, BLOCK_SIZE bsize_pred, int b_sub8x8, int bextend) {
  // Used in supertx
  // (mi_row_ori, mi_col_ori): location for mv
  // (mi_row_pred, mi_col_pred, bsize_pred): region to predict
  // (mi_row_top, mi_col_top, bsize_top): region of the top partition size
  // block: sub location of sub8x8 blocks
  // b_sub8x8: 1: ori is sub8x8; 0: ori is not sub8x8
  // bextend: 1: region to predict is an extension of ori; 0: not
  int r = (mi_row_pred - mi_row_top) * MI_SIZE;
  int c = (mi_col_pred - mi_col_top) * MI_SIZE;
  const int mi_width_top = num_8x8_blocks_wide_lookup[bsize_top];
  const int mi_height_top = num_8x8_blocks_high_lookup[bsize_top];
  MB_MODE_INFO *mbmi;
  AV1_COMMON *const cm = &pbi->common;

  if (mi_row_pred < mi_row_top || mi_col_pred < mi_col_top ||
      mi_row_pred >= mi_row_top + mi_height_top ||
      mi_col_pred >= mi_col_top + mi_width_top || mi_row_pred >= cm->mi_rows ||
      mi_col_pred >= cm->mi_cols)
    return;

  mbmi = set_offsets_extend(cm, xd, tile, bsize_pred, mi_row_pred, mi_col_pred,
                            mi_row_ori, mi_col_ori);
  set_ref(cm, xd, 0, mi_row_pred, mi_col_pred);
  if (has_second_ref(&xd->mi[0]->mbmi))
    set_ref(cm, xd, 1, mi_row_pred, mi_col_pred);

  if (!bextend) {
    mbmi->tx_size = b_width_log2_lookup[bsize_top];
  }

  xd->plane[0].dst.stride = dst_stride[0];
  xd->plane[1].dst.stride = dst_stride[1];
  xd->plane[2].dst.stride = dst_stride[2];
  xd->plane[0].dst.buf = dst_buf[0] +
                         (r >> xd->plane[0].subsampling_y) * dst_stride[0] +
                         (c >> xd->plane[0].subsampling_x);
  xd->plane[1].dst.buf = dst_buf[1] +
                         (r >> xd->plane[1].subsampling_y) * dst_stride[1] +
                         (c >> xd->plane[1].subsampling_x);
  xd->plane[2].dst.buf = dst_buf[2] +
                         (r >> xd->plane[2].subsampling_y) * dst_stride[2] +
                         (c >> xd->plane[2].subsampling_x);

  if (!b_sub8x8)
    av1_build_inter_predictors_sb_extend(xd,
#if CONFIG_EXT_INTER
                                         mi_row_ori, mi_col_ori,
#endif  // CONFIG_EXT_INTER
                                         mi_row_pred, mi_col_pred, bsize_pred);
  else
    av1_build_inter_predictors_sb_sub8x8_extend(xd,
#if CONFIG_EXT_INTER
                                                mi_row_ori, mi_col_ori,
#endif  // CONFIG_EXT_INTER
                                                mi_row_pred, mi_col_pred,
                                                bsize_pred, block);
}

static void dec_extend_dir(AV1Decoder *const pbi, MACROBLOCKD *const xd,
                           const TileInfo *const tile, int block,
                           BLOCK_SIZE bsize, BLOCK_SIZE top_bsize, int mi_row,
                           int mi_col, int mi_row_top, int mi_col_top,
                           uint8_t *dst_buf[3], int dst_stride[3], int dir) {
  // dir: 0-lower, 1-upper, 2-left, 3-right
  //      4-lowerleft, 5-upperleft, 6-lowerright, 7-upperright
  const int mi_width = num_8x8_blocks_wide_lookup[bsize];
  const int mi_height = num_8x8_blocks_high_lookup[bsize];
  int xss = xd->plane[1].subsampling_x;
  int yss = xd->plane[1].subsampling_y;
  int b_sub8x8 = (bsize < BLOCK_8X8) ? 1 : 0;
  BLOCK_SIZE extend_bsize;
  int unit, mi_row_pred, mi_col_pred;

  if (dir == 0 || dir == 1) {
    extend_bsize = (mi_width == 1 || bsize < BLOCK_8X8 || xss < yss)
                       ? BLOCK_8X8
                       : BLOCK_16X8;
    unit = num_8x8_blocks_wide_lookup[extend_bsize];
    mi_row_pred = mi_row + ((dir == 0) ? mi_height : -1);
    mi_col_pred = mi_col;

    dec_predict_b_extend(pbi, xd, tile, block, mi_row, mi_col, mi_row_pred,
                         mi_col_pred, mi_row_top, mi_col_top, dst_buf,
                         dst_stride, top_bsize, extend_bsize, b_sub8x8, 1);

    if (mi_width > unit) {
      int i;
      assert(!b_sub8x8);
      for (i = 0; i < mi_width / unit - 1; i++) {
        mi_col_pred += unit;
        dec_predict_b_extend(pbi, xd, tile, block, mi_row, mi_col, mi_row_pred,
                             mi_col_pred, mi_row_top, mi_col_top, dst_buf,
                             dst_stride, top_bsize, extend_bsize, b_sub8x8, 1);
      }
    }
  } else if (dir == 2 || dir == 3) {
    extend_bsize = (mi_height == 1 || bsize < BLOCK_8X8 || yss < xss)
                       ? BLOCK_8X8
                       : BLOCK_8X16;
    unit = num_8x8_blocks_high_lookup[extend_bsize];
    mi_row_pred = mi_row;
    mi_col_pred = mi_col + ((dir == 3) ? mi_width : -1);

    dec_predict_b_extend(pbi, xd, tile, block, mi_row, mi_col, mi_row_pred,
                         mi_col_pred, mi_row_top, mi_col_top, dst_buf,
                         dst_stride, top_bsize, extend_bsize, b_sub8x8, 1);

    if (mi_height > unit) {
      int i;
      for (i = 0; i < mi_height / unit - 1; i++) {
        mi_row_pred += unit;
        dec_predict_b_extend(pbi, xd, tile, block, mi_row, mi_col, mi_row_pred,
                             mi_col_pred, mi_row_top, mi_col_top, dst_buf,
                             dst_stride, top_bsize, extend_bsize, b_sub8x8, 1);
      }
    }
  } else {
    extend_bsize = BLOCK_8X8;
    mi_row_pred = mi_row + ((dir == 4 || dir == 6) ? mi_height : -1);
    mi_col_pred = mi_col + ((dir == 6 || dir == 7) ? mi_width : -1);
    dec_predict_b_extend(pbi, xd, tile, block, mi_row, mi_col, mi_row_pred,
                         mi_col_pred, mi_row_top, mi_col_top, dst_buf,
                         dst_stride, top_bsize, extend_bsize, b_sub8x8, 1);
  }
}

static void dec_extend_all(AV1Decoder *const pbi, MACROBLOCKD *const xd,
                           const TileInfo *const tile, int block,
                           BLOCK_SIZE bsize, BLOCK_SIZE top_bsize, int mi_row,
                           int mi_col, int mi_row_top, int mi_col_top,
                           uint8_t *dst_buf[3], int dst_stride[3]) {
  dec_extend_dir(pbi, xd, tile, block, bsize, top_bsize, mi_row, mi_col,
                 mi_row_top, mi_col_top, dst_buf, dst_stride, 0);
  dec_extend_dir(pbi, xd, tile, block, bsize, top_bsize, mi_row, mi_col,
                 mi_row_top, mi_col_top, dst_buf, dst_stride, 1);
  dec_extend_dir(pbi, xd, tile, block, bsize, top_bsize, mi_row, mi_col,
                 mi_row_top, mi_col_top, dst_buf, dst_stride, 2);
  dec_extend_dir(pbi, xd, tile, block, bsize, top_bsize, mi_row, mi_col,
                 mi_row_top, mi_col_top, dst_buf, dst_stride, 3);
  dec_extend_dir(pbi, xd, tile, block, bsize, top_bsize, mi_row, mi_col,
                 mi_row_top, mi_col_top, dst_buf, dst_stride, 4);
  dec_extend_dir(pbi, xd, tile, block, bsize, top_bsize, mi_row, mi_col,
                 mi_row_top, mi_col_top, dst_buf, dst_stride, 5);
  dec_extend_dir(pbi, xd, tile, block, bsize, top_bsize, mi_row, mi_col,
                 mi_row_top, mi_col_top, dst_buf, dst_stride, 6);
  dec_extend_dir(pbi, xd, tile, block, bsize, top_bsize, mi_row, mi_col,
                 mi_row_top, mi_col_top, dst_buf, dst_stride, 7);
}

static void dec_predict_sb_complex(AV1Decoder *const pbi, MACROBLOCKD *const xd,
                                   const TileInfo *const tile, int mi_row,
                                   int mi_col, int mi_row_top, int mi_col_top,
                                   BLOCK_SIZE bsize, BLOCK_SIZE top_bsize,
                                   uint8_t *dst_buf[3], int dst_stride[3]) {
  const AV1_COMMON *const cm = &pbi->common;
  const int hbs = num_8x8_blocks_wide_lookup[bsize] / 2;
  const PARTITION_TYPE partition = get_partition(cm, mi_row, mi_col, bsize);
  const BLOCK_SIZE subsize = get_subsize(bsize, partition);
#if CONFIG_EXT_PARTITION_TYPES
  const BLOCK_SIZE bsize2 = get_subsize(bsize, PARTITION_SPLIT);
#endif
  int i;
  const int mi_offset = mi_row * cm->mi_stride + mi_col;
  uint8_t *dst_buf1[3], *dst_buf2[3], *dst_buf3[3];

  DECLARE_ALIGNED(16, uint8_t, tmp_buf1[MAX_MB_PLANE * MAX_TX_SQUARE * 2]);
  DECLARE_ALIGNED(16, uint8_t, tmp_buf2[MAX_MB_PLANE * MAX_TX_SQUARE * 2]);
  DECLARE_ALIGNED(16, uint8_t, tmp_buf3[MAX_MB_PLANE * MAX_TX_SQUARE * 2]);
  int dst_stride1[3] = { MAX_TX_SIZE, MAX_TX_SIZE, MAX_TX_SIZE };
  int dst_stride2[3] = { MAX_TX_SIZE, MAX_TX_SIZE, MAX_TX_SIZE };
  int dst_stride3[3] = { MAX_TX_SIZE, MAX_TX_SIZE, MAX_TX_SIZE };

#if CONFIG_AOM_HIGHBITDEPTH
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    int len = sizeof(uint16_t);
    dst_buf1[0] = CONVERT_TO_BYTEPTR(tmp_buf1);
    dst_buf1[1] = CONVERT_TO_BYTEPTR(tmp_buf1 + MAX_TX_SQUARE * len);
    dst_buf1[2] = CONVERT_TO_BYTEPTR(tmp_buf1 + 2 * MAX_TX_SQUARE * len);
    dst_buf2[0] = CONVERT_TO_BYTEPTR(tmp_buf2);
    dst_buf2[1] = CONVERT_TO_BYTEPTR(tmp_buf2 + MAX_TX_SQUARE * len);
    dst_buf2[2] = CONVERT_TO_BYTEPTR(tmp_buf2 + 2 * MAX_TX_SQUARE * len);
    dst_buf3[0] = CONVERT_TO_BYTEPTR(tmp_buf3);
    dst_buf3[1] = CONVERT_TO_BYTEPTR(tmp_buf3 + MAX_TX_SQUARE * len);
    dst_buf3[2] = CONVERT_TO_BYTEPTR(tmp_buf3 + 2 * MAX_TX_SQUARE * len);
  } else {
#endif
    dst_buf1[0] = tmp_buf1;
    dst_buf1[1] = tmp_buf1 + MAX_TX_SQUARE;
    dst_buf1[2] = tmp_buf1 + 2 * MAX_TX_SQUARE;
    dst_buf2[0] = tmp_buf2;
    dst_buf2[1] = tmp_buf2 + MAX_TX_SQUARE;
    dst_buf2[2] = tmp_buf2 + 2 * MAX_TX_SQUARE;
    dst_buf3[0] = tmp_buf3;
    dst_buf3[1] = tmp_buf3 + MAX_TX_SQUARE;
    dst_buf3[2] = tmp_buf3 + 2 * MAX_TX_SQUARE;
#if CONFIG_AOM_HIGHBITDEPTH
  }
#endif

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

  xd->mi = cm->mi_grid_visible + mi_offset;
  xd->mi[0] = cm->mi + mi_offset;

  for (i = 0; i < MAX_MB_PLANE; i++) {
    xd->plane[i].dst.buf = dst_buf[i];
    xd->plane[i].dst.stride = dst_stride[i];
  }

  switch (partition) {
    case PARTITION_NONE:
      assert(bsize < top_bsize);
      dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
                           mi_row_top, mi_col_top, dst_buf, dst_stride,
                           top_bsize, bsize, 0, 0);
      dec_extend_all(pbi, xd, tile, 0, bsize, top_bsize, mi_row, mi_col,
                     mi_row_top, mi_col_top, dst_buf, dst_stride);
      break;
    case PARTITION_HORZ:
      if (bsize == BLOCK_8X8) {
        // For sub8x8, predict in 8x8 unit
        // First half
        dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
                             mi_row_top, mi_col_top, dst_buf, dst_stride,
                             top_bsize, BLOCK_8X8, 1, 0);
        if (bsize < top_bsize)
          dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
                         mi_row_top, mi_col_top, dst_buf, dst_stride);

        // Second half
        dec_predict_b_extend(pbi, xd, tile, 2, mi_row, mi_col, mi_row, mi_col,
                             mi_row_top, mi_col_top, dst_buf1, dst_stride1,
                             top_bsize, BLOCK_8X8, 1, 1);
        if (bsize < top_bsize)
          dec_extend_all(pbi, xd, tile, 2, subsize, top_bsize, mi_row, mi_col,
                         mi_row_top, mi_col_top, dst_buf1, dst_stride1);

        // weighted average to smooth the boundary
        xd->plane[0].dst.buf = dst_buf[0];
        xd->plane[0].dst.stride = dst_stride[0];
        av1_build_masked_inter_predictor_complex(
            xd, dst_buf[0], dst_stride[0], dst_buf1[0], dst_stride1[0], mi_row,
            mi_col, mi_row_top, mi_col_top, bsize, top_bsize, PARTITION_HORZ,
            0);
      } else {
        // First half
        dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
                             mi_row_top, mi_col_top, dst_buf, dst_stride,
                             top_bsize, subsize, 0, 0);
        if (bsize < top_bsize)
          dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
                         mi_row_top, mi_col_top, dst_buf, dst_stride);
        else
          dec_extend_dir(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
                         mi_row_top, mi_col_top, dst_buf, dst_stride, 0);

        if (mi_row + hbs < cm->mi_rows) {
          // Second half
          dec_predict_b_extend(pbi, xd, tile, 0, mi_row + hbs, mi_col,
                               mi_row + hbs, mi_col, mi_row_top, mi_col_top,
                               dst_buf1, dst_stride1, top_bsize, subsize, 0, 0);
          if (bsize < top_bsize)
            dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row + hbs,
                           mi_col, mi_row_top, mi_col_top, dst_buf1,
                           dst_stride1);
          else
            dec_extend_dir(pbi, xd, tile, 0, subsize, top_bsize, mi_row + hbs,
                           mi_col, mi_row_top, mi_col_top, dst_buf1,
                           dst_stride1, 1);

          // weighted average to smooth the boundary
          for (i = 0; i < MAX_MB_PLANE; i++) {
            xd->plane[i].dst.buf = dst_buf[i];
            xd->plane[i].dst.stride = dst_stride[i];
            av1_build_masked_inter_predictor_complex(
                xd, dst_buf[i], dst_stride[i], dst_buf1[i], dst_stride1[i],
                mi_row, mi_col, mi_row_top, mi_col_top, bsize, top_bsize,
                PARTITION_HORZ, i);
          }
        }
      }
      break;
    case PARTITION_VERT:
      if (bsize == BLOCK_8X8) {
        // First half
        dec_predict_b_extend(pbi, xd, tile, 0, mi_row, mi_col, mi_row, mi_col,
                             mi_row_top, mi_col_top, dst_buf, dst_stride,
                             top_bsize, BLOCK_8X8, 1, 0);
        if (bsize < top_bsize)
          dec_extend_all(pbi, xd, tile, 0, subsize, top_bsize, mi_row, mi_col,
                         mi_row_top, mi_col_top, dst_buf, dst_stride);

        // Second half
        dec_predict_b_extend(pbi, xd, tile, 1, mi_row, mi_col, mi_row, mi_col,
                             mi_row_top, mi_col_top, dst_buf1, dst_stride1,
                             top_bsize, BLOCK_8X8, 1, 1);
        if (bsize