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

#include "aom_dsp/quantize.h"
#include "aom_mem/aom_mem.h"
#include "aom_ports/mem.h"

#include "av1/common/idct.h"
#include "av1/common/reconinter.h"
#include "av1/common/reconintra.h"
#include "av1/common/scan.h"

#include "av1/encoder/encodemb.h"
#include "av1/encoder/hybrid_fwd_txfm.h"
#include "av1/encoder/quantize.h"
#include "av1/encoder/rd.h"
#include "av1/encoder/tokenize.h"

#if CONFIG_PVQ
#include "av1/encoder/encint.h"
#include "av1/common/partition.h"
#include "av1/encoder/pvq_encoder.h"
#endif

// Check if one needs to use c version subtraction.
static int check_subtract_block_size(int w, int h) { return w < 4 || h < 4; }

void av1_subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
  struct macroblock_plane *const p = &x->plane[plane];
  const struct macroblockd_plane *const pd = &x->e_mbd.plane[plane];
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
  const int bw = block_size_wide[plane_bsize];
  const int bh = block_size_high[plane_bsize];

  if (check_subtract_block_size(bw, bh)) {
#if CONFIG_AOM_HIGHBITDEPTH
    if (x->e_mbd.cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
      aom_highbd_subtract_block_c(bh, bw, p->src_diff, bw, p->src.buf,
                                  p->src.stride, pd->dst.buf, pd->dst.stride,
                                  x->e_mbd.bd);
      return;
    }
#endif  // CONFIG_AOM_HIGHBITDEPTH
    aom_subtract_block_c(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
                         pd->dst.buf, pd->dst.stride);

    return;
  }

#if CONFIG_AOM_HIGHBITDEPTH
  if (x->e_mbd.cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    aom_highbd_subtract_block(bh, bw, p->src_diff, bw, p->src.buf,
                              p->src.stride, pd->dst.buf, pd->dst.stride,
                              x->e_mbd.bd);
    return;
  }
#endif  // CONFIG_AOM_HIGHBITDEPTH
  aom_subtract_block(bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
                     pd->dst.buf, pd->dst.stride);
}

typedef struct av1_token_state {
  int rate;
  int64_t error;
  int next;
  int16_t token;
  tran_low_t qc;
  tran_low_t dqc;
} av1_token_state;

// These numbers are empirically obtained.
static const int plane_rd_mult[REF_TYPES][PLANE_TYPES] = {
  { 10, 6 }, { 8, 5 },
};

#define UPDATE_RD_COST()                             \
  {                                                  \
    rd_cost0 = RDCOST(rdmult, rddiv, rate0, error0); \
    rd_cost1 = RDCOST(rdmult, rddiv, rate1, error1); \
  }

int av1_optimize_b(const AV1_COMMON *cm, MACROBLOCK *mb, int plane, int block,
                   TX_SIZE tx_size, int ctx) {
  MACROBLOCKD *const xd = &mb->e_mbd;
  struct macroblock_plane *const p = &mb->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
  const int ref = is_inter_block(&xd->mi[0]->mbmi);
  av1_token_state tokens[MAX_TX_SQUARE + 1][2];
  unsigned best_index[MAX_TX_SQUARE + 1][2];
  uint8_t token_cache[MAX_TX_SQUARE];
  const tran_low_t *const coeff = BLOCK_OFFSET(mb->plane[plane].coeff, block);
  tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
  const int eob = p->eobs[block];
  const PLANE_TYPE plane_type = pd->plane_type;
  const int default_eob = tx_size_2d[tx_size];
  const int16_t *const dequant_ptr = pd->dequant;
  const uint8_t *const band_translate = get_band_translate(tx_size);
  TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
  const SCAN_ORDER *const scan_order =
      get_scan(cm, tx_size, tx_type, is_inter_block(&xd->mi[0]->mbmi));
  const int16_t *const scan = scan_order->scan;
  const int16_t *const nb = scan_order->neighbors;
  int dqv;
  const int shift = get_tx_scale(tx_size);
#if CONFIG_AOM_QM
  int seg_id = xd->mi[0]->mbmi.segment_id;
  const qm_val_t *iqmatrix = pd->seg_iqmatrix[seg_id][!ref][tx_size];
#endif
#if CONFIG_NEW_QUANT
  int dq = get_dq_profile_from_ctx(mb->qindex, ctx, ref, plane_type);
  const dequant_val_type_nuq *dequant_val = pd->dequant_val_nuq[dq];
#elif !CONFIG_AOM_QM
  const int dq_step[2] = { dequant_ptr[0] >> shift, dequant_ptr[1] >> shift };
#endif  // CONFIG_NEW_QUANT
  int next = eob, sz = 0;
  const int64_t rdmult = (mb->rdmult * plane_rd_mult[ref][plane_type]) >> 1;
  const int64_t rddiv = mb->rddiv;
  int64_t rd_cost0, rd_cost1;
  int rate0, rate1;
  int64_t error0, error1;
  int16_t t0, t1;
  int best, band = (eob < default_eob) ? band_translate[eob]
                                       : band_translate[eob - 1];
  int pt, i, final_eob;
#if CONFIG_AOM_HIGHBITDEPTH
  const int *cat6_high_cost = av1_get_high_cost_table(xd->bd);
#else
  const int *cat6_high_cost = av1_get_high_cost_table(8);
#endif
  unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
      mb->token_costs[txsize_sqr_map[tx_size]][plane_type][ref];
  const uint16_t *band_counts = &band_count_table[tx_size][band];
  uint16_t band_left = eob - band_cum_count_table[tx_size][band] + 1;
  int shortcut = 0;
  int next_shortcut = 0;

  assert((mb->qindex == 0) ^ (xd->lossless[xd->mi[0]->mbmi.segment_id] == 0));

  token_costs += band;

  assert((!plane_type && !plane) || (plane_type && plane));
  assert(eob <= default_eob);

  /* Now set up a Viterbi trellis to evaluate alternative roundings. */
  /* Initialize the sentinel node of the trellis. */
  tokens[eob][0].rate = 0;
  tokens[eob][0].error = 0;
  tokens[eob][0].next = default_eob;
  tokens[eob][0].token = EOB_TOKEN;
  tokens[eob][0].qc = 0;
  tokens[eob][1] = tokens[eob][0];

  for (i = 0; i < eob; i++) {
    const int rc = scan[i];
    tokens[i][0].rate = av1_get_token_cost(qcoeff[rc], &t0, cat6_high_cost);
    tokens[i][0].token = t0;
    token_cache[rc] = av1_pt_energy_class[t0];
  }

  for (i = eob; i-- > 0;) {
    int base_bits, dx;
    int64_t d2;
    const int rc = scan[i];
    int x = qcoeff[rc];
#if CONFIG_AOM_QM
    int iwt = iqmatrix[rc];
    dqv = dequant_ptr[rc != 0];
    dqv = ((iwt * (int)dqv) + (1 << (AOM_QM_BITS - 1))) >> AOM_QM_BITS;
#else
    dqv = dequant_ptr[rc != 0];
#endif
    next_shortcut = shortcut;

    /* Only add a trellis state for non-zero coefficients. */
    if (UNLIKELY(x)) {
      error0 = tokens[next][0].error;
      error1 = tokens[next][1].error;
      /* Evaluate the first possibility for this state. */
      rate0 = tokens[next][0].rate;
      rate1 = tokens[next][1].rate;

      if (next_shortcut) {
        /* Consider both possible successor states. */
        if (next < default_eob) {
          pt = get_coef_context(nb, token_cache, i + 1);
          rate0 += (*token_costs)[0][pt][tokens[next][0].token];
          rate1 += (*token_costs)[0][pt][tokens[next][1].token];
        }
        UPDATE_RD_COST();
        /* And pick the best. */
        best = rd_cost1 < rd_cost0;
      } else {
        if (next < default_eob) {
          pt = get_coef_context(nb, token_cache, i + 1);
          rate0 += (*token_costs)[0][pt][tokens[next][0].token];
        }
        best = 0;
      }

      dx = (dqcoeff[rc] - coeff[rc]) * (1 << shift);
#if CONFIG_AOM_HIGHBITDEPTH
      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
        dx >>= xd->bd - 8;
      }
#endif  // CONFIG_AOM_HIGHBITDEPTH
      d2 = (int64_t)dx * dx;
      tokens[i][0].rate += (best ? rate1 : rate0);
      tokens[i][0].error = d2 + (best ? error1 : error0);
      tokens[i][0].next = next;
      tokens[i][0].qc = x;
      tokens[i][0].dqc = dqcoeff[rc];
      best_index[i][0] = best;

      /* Evaluate the second possibility for this state. */
      rate0 = tokens[next][0].rate;
      rate1 = tokens[next][1].rate;

      // The threshold of 3 is empirically obtained.
      if (UNLIKELY(abs(x) > 3)) {
        shortcut = 0;
      } else {
#if CONFIG_NEW_QUANT
        shortcut = ((av1_dequant_abscoeff_nuq(abs(x), dqv,
                                              dequant_val[band_translate[i]]) >
                     (abs(coeff[rc]) << shift)) &&
                    (av1_dequant_abscoeff_nuq(abs(x) - 1, dqv,
                                              dequant_val[band_translate[i]]) <
                     (abs(coeff[rc]) << shift)));
#else  // CONFIG_NEW_QUANT
#if CONFIG_AOM_QM
        if ((abs(x) * dequant_ptr[rc != 0] * iwt >
             ((abs(coeff[rc]) << shift) << AOM_QM_BITS)) &&
            (abs(x) * dequant_ptr[rc != 0] * iwt <
             (((abs(coeff[rc]) << shift) + dequant_ptr[rc != 0])
              << AOM_QM_BITS)))
#else
        if ((abs(x) * dequant_ptr[rc != 0] > (abs(coeff[rc]) << shift)) &&
            (abs(x) * dequant_ptr[rc != 0] <
             (abs(coeff[rc]) << shift) + dequant_ptr[rc != 0]))
#endif  // CONFIG_AOM_QM
          shortcut = 1;
        else
          shortcut = 0;
#endif  // CONFIG_NEW_QUANT
      }

      if (shortcut) {
        sz = -(x < 0);
        x -= 2 * sz + 1;
      } else {
        tokens[i][1] = tokens[i][0];
        best_index[i][1] = best_index[i][0];
        next = i;

        if (UNLIKELY(!(--band_left))) {
          --band_counts;
          band_left = *band_counts;
          --token_costs;
        }
        continue;
      }

      /* Consider both possible successor states. */
      if (!x) {
        /* If we reduced this coefficient to zero, check to see if
         *  we need to move the EOB back here.
         */
        t0 = tokens[next][0].token == EOB_TOKEN ? EOB_TOKEN : ZERO_TOKEN;
        t1 = tokens[next][1].token == EOB_TOKEN ? EOB_TOKEN : ZERO_TOKEN;
        base_bits = 0;
      } else {
        base_bits = av1_get_token_cost(x, &t0, cat6_high_cost);
        t1 = t0;
      }

      if (next_shortcut) {
        if (LIKELY(next < default_eob)) {
          if (t0 != EOB_TOKEN) {
            token_cache[rc] = av1_pt_energy_class[t0];
            pt = get_coef_context(nb, token_cache, i + 1);
            rate0 += (*token_costs)[!x][pt][tokens[next][0].token];
          }
          if (t1 != EOB_TOKEN) {
            token_cache[rc] = av1_pt_energy_class[t1];
            pt = get_coef_context(nb, token_cache, i + 1);
            rate1 += (*token_costs)[!x][pt][tokens[next][1].token];
          }
        }

        UPDATE_RD_COST();
        /* And pick the best. */
        best = rd_cost1 < rd_cost0;
      } else {
        // The two states in next stage are identical.
        if (next < default_eob && t0 != EOB_TOKEN) {
          token_cache[rc] = av1_pt_energy_class[t0];
          pt = get_coef_context(nb, token_cache, i + 1);
          rate0 += (*token_costs)[!x][pt][tokens[next][0].token];
        }
        best = 0;
      }

#if CONFIG_NEW_QUANT
      dx = av1_dequant_coeff_nuq(x, dqv, dequant_val[band_translate[i]]) -
           (coeff[rc] << shift);
#if CONFIG_AOM_HIGHBITDEPTH
      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
        dx >>= xd->bd - 8;
      }
#endif  // CONFIG_AOM_HIGHBITDEPTH
#else   // CONFIG_NEW_QUANT
#if CONFIG_AOM_HIGHBITDEPTH
      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
        dx -= ((dqv >> (xd->bd - 8)) + sz) ^ sz;
      } else {
        dx -= (dqv + sz) ^ sz;
      }
#else
      dx -= (dqv + sz) ^ sz;
#endif  // CONFIG_AOM_HIGHBITDEPTH
#endif  // CONFIG_NEW_QUANT
      d2 = (int64_t)dx * dx;

      tokens[i][1].rate = base_bits + (best ? rate1 : rate0);
      tokens[i][1].error = d2 + (best ? error1 : error0);
      tokens[i][1].next = next;
      tokens[i][1].token = best ? t1 : t0;
      tokens[i][1].qc = x;

      if (x) {
#if CONFIG_NEW_QUANT
        tokens[i][1].dqc = av1_dequant_abscoeff_nuq(
            abs(x), dqv, dequant_val[band_translate[i]]);
        tokens[i][1].dqc = shift ? ROUND_POWER_OF_TWO(tokens[i][1].dqc, shift)
                                 : tokens[i][1].dqc;
        if (sz) tokens[i][1].dqc = -tokens[i][1].dqc;
#else
// The 32x32 transform coefficient uses half quantization step size.
// Account for the rounding difference in the dequantized coefficeint
// value when the quantization index is dropped from an even number
// to an odd number.

#if CONFIG_AOM_QM
        tran_low_t offset = dqv >> shift;
#else
        tran_low_t offset = dq_step[rc != 0];
#endif
        if (shift & x) offset += (dqv & 0x01);

        if (sz == 0)
          tokens[i][1].dqc = dqcoeff[rc] - offset;
        else
          tokens[i][1].dqc = dqcoeff[rc] + offset;
#endif  // CONFIG_NEW_QUANT
      } else {
        tokens[i][1].dqc = 0;
      }

      best_index[i][1] = best;
      /* Finally, make this the new head of the trellis. */
      next = i;
    } else {
      /* There's no choice to make for a zero coefficient, so we don't
       *  add a new trellis node, but we do need to update the costs.
       */
      t0 = tokens[next][0].token;
      t1 = tokens[next][1].token;
      pt = get_coef_context(nb, token_cache, i + 1);
      /* Update the cost of each path if we're past the EOB token. */
      if (t0 != EOB_TOKEN) {
        tokens[next][0].rate += (*token_costs)[1][pt][t0];
        tokens[next][0].token = ZERO_TOKEN;
      }
      if (t1 != EOB_TOKEN) {
        tokens[next][1].rate += (*token_costs)[1][pt][t1];
        tokens[next][1].token = ZERO_TOKEN;
      }
      best_index[i][0] = best_index[i][1] = 0;
      shortcut = (tokens[next][0].rate != tokens[next][1].rate);
      /* Don't update next, because we didn't add a new node. */
    }

    if (UNLIKELY(!(--band_left))) {
      --band_counts;
      band_left = *band_counts;
      --token_costs;
    }
  }

  /* Now pick the best path through the whole trellis. */
  rate0 = tokens[next][0].rate;
  rate1 = tokens[next][1].rate;
  error0 = tokens[next][0].error;
  error1 = tokens[next][1].error;
  t0 = tokens[next][0].token;
  t1 = tokens[next][1].token;
  rate0 += (*token_costs)[0][ctx][t0];
  rate1 += (*token_costs)[0][ctx][t1];
  UPDATE_RD_COST();
  best = rd_cost1 < rd_cost0;

  final_eob = -1;

  for (i = next; i < eob; i = next) {
    const int x = tokens[i][best].qc;
    const int rc = scan[i];
    if (x) final_eob = i;
    qcoeff[rc] = x;
    dqcoeff[rc] = tokens[i][best].dqc;

    next = tokens[i][best].next;
    best = best_index[i][best];
  }
  final_eob++;

  mb->plane[plane].eobs[block] = final_eob;
  assert(final_eob <= default_eob);
  return final_eob;
}

#if CONFIG_AOM_HIGHBITDEPTH
typedef enum QUANT_FUNC {
  QUANT_FUNC_LOWBD = 0,
  QUANT_FUNC_HIGHBD = 1,
  QUANT_FUNC_TYPES = 2
} QUANT_FUNC;

static AV1_QUANT_FACADE
    quant_func_list[AV1_XFORM_QUANT_TYPES][QUANT_FUNC_TYPES] = {
      { av1_quantize_fp_facade, av1_highbd_quantize_fp_facade },
      { av1_quantize_b_facade, av1_highbd_quantize_b_facade },
      { av1_quantize_dc_facade, av1_highbd_quantize_dc_facade },
#if CONFIG_NEW_QUANT
      { av1_quantize_fp_nuq_facade, av1_highbd_quantize_fp_nuq_facade },
      { av1_quantize_b_nuq_facade, av1_highbd_quantize_b_nuq_facade },
      { av1_quantize_dc_nuq_facade, av1_highbd_quantize_dc_nuq_facade },
#endif  // CONFIG_NEW_QUANT
      { NULL, NULL }
    };

#elif !CONFIG_PVQ

typedef enum QUANT_FUNC {
  QUANT_FUNC_LOWBD = 0,
  QUANT_FUNC_TYPES = 1
} QUANT_FUNC;

static AV1_QUANT_FACADE quant_func_list[AV1_XFORM_QUANT_TYPES]
                                       [QUANT_FUNC_TYPES] = {
                                         { av1_quantize_fp_facade },
                                         { av1_quantize_b_facade },
                                         { av1_quantize_dc_facade },
#if CONFIG_NEW_QUANT
                                         { av1_quantize_fp_nuq_facade },
                                         { av1_quantize_b_nuq_facade },
                                         { av1_quantize_dc_nuq_facade },
#endif  // CONFIG_NEW_QUANT
                                         { NULL }
                                       };
#endif

static FWD_TXFM_OPT fwd_txfm_opt_list[AV1_XFORM_QUANT_TYPES] = {
  FWD_TXFM_OPT_NORMAL, FWD_TXFM_OPT_NORMAL, FWD_TXFM_OPT_DC,
#if CONFIG_NEW_QUANT
  FWD_TXFM_OPT_NORMAL, FWD_TXFM_OPT_NORMAL, FWD_TXFM_OPT_DC,
#endif  // CONFIG_NEW_QUANT
  FWD_TXFM_OPT_NORMAL
};

void av1_xform_quant(const AV1_COMMON *cm, MACROBLOCK *x, int plane, int block,
                     int blk_row, int blk_col, BLOCK_SIZE plane_bsize,
                     TX_SIZE tx_size, int ctx,
                     AV1_XFORM_QUANT xform_quant_idx) {
  MACROBLOCKD *const xd = &x->e_mbd;
#if !CONFIG_PVQ
  const struct macroblock_plane *const p = &x->plane[plane];
  const struct macroblockd_plane *const pd = &xd->plane[plane];
#else
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
#endif
  PLANE_TYPE plane_type = (plane == 0) ? PLANE_TYPE_Y : PLANE_TYPE_UV;
  TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
  const int is_inter = is_inter_block(&xd->mi[0]->mbmi);
  const SCAN_ORDER *const scan_order = get_scan(cm, tx_size, tx_type, is_inter);
  tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
  tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
  uint16_t *const eob = &p->eobs[block];
  const int diff_stride = block_size_wide[plane_bsize];
#if CONFIG_AOM_QM
  int seg_id = xd->mi[0]->mbmi.segment_id;
  const qm_val_t *qmatrix = pd->seg_qmatrix[seg_id][!is_inter][tx_size];
  const qm_val_t *iqmatrix = pd->seg_iqmatrix[seg_id][!is_inter][tx_size];
#endif

  FWD_TXFM_PARAM fwd_txfm_param;

#if !CONFIG_PVQ
  const int tx2d_size = tx_size_2d[tx_size];
  QUANT_PARAM qparam;
  const int16_t *src_diff;

  src_diff = &p->src_diff[4 * (blk_row * diff_stride + blk_col)];
  qparam.log_scale = get_tx_scale(tx_size);
#if CONFIG_NEW_QUANT
  qparam.tx_size = tx_size;
  qparam.dq = get_dq_profile_from_ctx(x->qindex, ctx, is_inter, plane_type);
#endif  // CONFIG_NEW_QUANT
#if CONFIG_AOM_QM
  qparam.qmatrix = qmatrix;
  qparam.iqmatrix = iqmatrix;
#endif  // CONFIG_AOM_QM
#else
  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
  tran_low_t *ref_coeff = BLOCK_OFFSET(pd->pvq_ref_coeff, block);
  uint8_t *src, *dst;
  int16_t *src_int16, *pred;
  const int src_stride = p->src.stride;
  const int dst_stride = pd->dst.stride;
  int tx_blk_size;
  int i, j;
  int skip = 1;
  PVQ_INFO *pvq_info = NULL;

  (void)scan_order;
  (void)qcoeff;

  if (x->pvq_coded) {
    assert(block < MAX_PVQ_BLOCKS_IN_SB);
    pvq_info = &x->pvq[block][plane];
  }
  dst = &pd->dst.buf[4 * (blk_row * dst_stride + blk_col)];
  src = &p->src.buf[4 * (blk_row * src_stride + blk_col)];
  src_int16 = &p->src_int16[4 * (blk_row * diff_stride + blk_col)];
  pred = &pd->pred[4 * (blk_row * diff_stride + blk_col)];

  // transform block size in pixels
  tx_blk_size = tx_size_wide[tx_size];

  // copy uint8 orig and predicted block to int16 buffer
  // in order to use existing VP10 transform functions
  for (j = 0; j < tx_blk_size; j++)
    for (i = 0; i < tx_blk_size; i++) {
      src_int16[diff_stride * j + i] = src[src_stride * j + i];
      pred[diff_stride * j + i] = dst[dst_stride * j + i];
    }
#endif
  (void)ctx;

  fwd_txfm_param.tx_type = tx_type;
  fwd_txfm_param.tx_size = tx_size;
  fwd_txfm_param.fwd_txfm_opt = fwd_txfm_opt_list[xform_quant_idx];
  fwd_txfm_param.rd_transform = x->use_lp32x32fdct;
  fwd_txfm_param.lossless = xd->lossless[xd->mi[0]->mbmi.segment_id];

#if CONFIG_AOM_HIGHBITDEPTH
  fwd_txfm_param.bd = xd->bd;
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    highbd_fwd_txfm(src_diff, coeff, diff_stride, &fwd_txfm_param);
    if (xform_quant_idx != AV1_XFORM_QUANT_SKIP_QUANT) {
      if (LIKELY(!x->skip_block)) {
        quant_func_list[xform_quant_idx][QUANT_FUNC_HIGHBD](
            coeff, tx2d_size, p, qcoeff, pd, dqcoeff, eob, scan_order, &qparam);
      } else {
        av1_quantize_skip(tx2d_size, qcoeff, dqcoeff, eob);
      }
    }
    return;
  }
#endif  // CONFIG_AOM_HIGHBITDEPTH

#if !CONFIG_PVQ
  fwd_txfm(src_diff, coeff, diff_stride, &fwd_txfm_param);
  if (xform_quant_idx != AV1_XFORM_QUANT_SKIP_QUANT) {
    if (LIKELY(!x->skip_block)) {
      quant_func_list[xform_quant_idx][QUANT_FUNC_LOWBD](
          coeff, tx2d_size, p, qcoeff, pd, dqcoeff, eob, scan_order, &qparam);
    } else {
      av1_quantize_skip(tx2d_size, qcoeff, dqcoeff, eob);
    }
  }
#else   // #if !CONFIG_PVQ
  fwd_txfm_param.rd_transform = 0;

  fwd_txfm(src_int16, coeff, diff_stride, &fwd_txfm_param);
  fwd_txfm(pred, ref_coeff, diff_stride, &fwd_txfm_param);

  // PVQ for inter mode block
  if (!x->skip_block)
    skip = av1_pvq_encode_helper(&x->daala_enc,
                                 coeff,        // target original vector
                                 ref_coeff,    // reference vector
                                 dqcoeff,      // de-quantized vector
                                 eob,          // End of Block marker
                                 pd->dequant,  // aom's quantizers
                                 plane,        // image plane
                                 tx_size,      // block size in log_2 - 2
                                 tx_type,
                                 &x->rate,  // rate measured
                                 x->pvq_speed,
                                 pvq_info);  // PVQ info for a block

  x->pvq_skip[plane] = skip;

  if (!skip) mbmi->skip = 0;
#endif  // #if !CONFIG_PVQ
}

static void encode_block(int plane, int block, int blk_row, int blk_col,
                         BLOCK_SIZE plane_bsize, TX_SIZE tx_size, void *arg) {
  struct encode_b_args *const args = arg;
  AV1_COMMON *cm = args->cm;
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
  int ctx;
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
  uint8_t *dst;
  ENTROPY_CONTEXT *a, *l;
  INV_TXFM_PARAM inv_txfm_param;
#if CONFIG_PVQ
  int tx_blk_size;
  int i, j;
#endif
#if CONFIG_VAR_TX
  int i;
  const int bwl = b_width_log2_lookup[plane_bsize];
#endif
  dst = &pd->dst.buf[4 * blk_row * pd->dst.stride + 4 * blk_col];
  a = &args->ta[blk_col];
  l = &args->tl[blk_row];
#if CONFIG_VAR_TX
  ctx = get_entropy_context(tx_size, a, l);
#else
  ctx = combine_entropy_contexts(*a, *l);
#endif

#if CONFIG_VAR_TX
  // Assert not magic number (uninitialized).
  assert(x->blk_skip[plane][(blk_row << bwl) + blk_col] != 234);

  if (x->blk_skip[plane][(blk_row << bwl) + blk_col] == 0) {
#else
  {
#endif
#if CONFIG_NEW_QUANT
    av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
                    ctx, AV1_XFORM_QUANT_FP_NUQ);
#else
    av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
                    ctx, AV1_XFORM_QUANT_FP);
#endif  // CONFIG_NEW_QUANT
  }
#if CONFIG_VAR_TX
  else {
    p->eobs[block] = 0;
  }
#endif
#if !CONFIG_PVQ
  if (p->eobs[block] && !xd->lossless[xd->mi[0]->mbmi.segment_id]) {
    *a = *l = av1_optimize_b(cm, x, plane, block, tx_size, ctx) > 0;
  } else {
    *a = *l = p->eobs[block] > 0;
  }

#if CONFIG_VAR_TX
  for (i = 0; i < tx_size_wide_unit[tx_size]; ++i) a[i] = a[0];

  for (i = 0; i < tx_size_high_unit[tx_size]; ++i) l[i] = l[0];
#endif

  if (p->eobs[block]) *(args->skip) = 0;

  if (p->eobs[block] == 0) return;
#else
  (void)ctx;
  *a = *l = !x->pvq_skip[plane];

  if (!x->pvq_skip[plane]) *(args->skip) = 0;

  if (x->pvq_skip[plane]) return;

  // transform block size in pixels
  tx_blk_size = tx_size_wide[tx_size];

  // Since av1 does not have separate function which does inverse transform
  // but av1_inv_txfm_add_*x*() also does addition of predicted image to
  // inverse transformed 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;
#endif

  // inverse transform parameters
  inv_txfm_param.tx_type = get_tx_type(pd->plane_type, xd, block, tx_size);
  inv_txfm_param.tx_size = tx_size;
  inv_txfm_param.eob = p->eobs[block];
  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, pd->dst.stride, &inv_txfm_param);
    return;
  }
#endif  // CONFIG_AOM_HIGHBITDEPTH
  inv_txfm_add(dqcoeff, dst, pd->dst.stride, &inv_txfm_param);
}

#if CONFIG_VAR_TX
static void encode_block_inter(int plane, int block, int blk_row, int blk_col,
                               BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
                               void *arg) {
  struct encode_b_args *const args = arg;
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
  const BLOCK_SIZE bsize = txsize_to_bsize[tx_size];
  const struct macroblockd_plane *const pd = &xd->plane[plane];
  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) {
    encode_block(plane, block, blk_row, blk_col, plane_bsize, tx_size, arg);
  } else {
    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
    // This is the square transform block partition entry point.
    int bsl = tx_size_wide_unit[sub_txs];
    int i;
    assert(bsl > 0);
    assert(tx_size < TX_SIZES_ALL);

    for (i = 0; i < 4; ++i) {
      const int offsetr = blk_row + ((i >> 1) * bsl);
      const int offsetc = blk_col + ((i & 0x01) * bsl);
      int step = tx_size_wide_unit[sub_txs] * tx_size_high_unit[sub_txs];

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

      encode_block_inter(plane, block, offsetr, offsetc, plane_bsize, sub_txs,
                         arg);
      block += step;
    }
  }
}
#endif

typedef struct encode_block_pass1_args {
  AV1_COMMON *cm;
  MACROBLOCK *x;
} encode_block_pass1_args;

static void encode_block_pass1(int plane, int block, int blk_row, int blk_col,
                               BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
                               void *arg) {
  encode_block_pass1_args *args = (encode_block_pass1_args *)arg;
  AV1_COMMON *cm = args->cm;
  MACROBLOCK *const x = args->x;
  MACROBLOCKD *const xd = &x->e_mbd;
  struct macroblock_plane *const p = &x->plane[plane];
  struct macroblockd_plane *const pd = &xd->plane[plane];
  tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
  uint8_t *dst;
  int ctx = 0;
  dst = &pd->dst.buf[4 * blk_row * pd->dst.stride + 4 * blk_col];

#if CONFIG_NEW_QUANT
  av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
                  ctx, AV1_XFORM_QUANT_B_NUQ);
#else
  av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
                  ctx, AV1_XFORM_QUANT_B);
#endif  // CONFIG_NEW_QUANT
#if !CONFIG_PVQ
  if (p->eobs[block] > 0) {
#else
  if (!x->pvq_skip[plane]) {
    {
      int tx_blk_size;
      int i, j;
      // transform block size in pixels
      tx_blk_size = tx_size_wide[tx_size];

      // Since av1 does not have separate function which does inverse transform
      // but av1_inv_txfm_add_*x*() also does addition of predicted image to
      // inverse transformed 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;
    }
#endif  // !CONFIG_PVQ
#if CONFIG_AOM_HIGHBITDEPTH
    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
      if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {
        av1_highbd_iwht4x4_add(dqcoeff, dst, pd->dst.stride, p->eobs[block],
                               xd->bd);
      } else {
        av1_highbd_idct4x4_add(dqcoeff, dst, pd->dst.stride, p->eobs[block],
                               xd->bd);
      }
      return;
    }
#endif  //  CONFIG_AOM_HIGHBITDEPTH
    if (xd->lossless[xd->mi[0]->mbmi.segment_id]) {
      av1_iwht4x4_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
    } else {
      av1_idct4x4_add(dqcoeff, dst, pd->dst.stride, p->eobs[block]);
    }
  }
}

void av1_encode_sby_pass1(AV1_COMMON *cm, MACROBLOCK *x, BLOCK_SIZE bsize) {
  encode_block_pass1_args args = { cm, x };
  av1_subtract_plane(x, bsize, 0);
  av1_foreach_transformed_block_in_plane(&x->e_mbd, bsize, 0,
                                         encode_block_pass1, &args);
}

void av1_encode_sb(AV1_COMMON *cm, MACROBLOCK *x, BLOCK_SIZE bsize) {
  MACROBLOCKD *const xd = &x->e_mbd;
  struct optimize_ctx ctx;
  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
  struct encode_b_args arg = { cm, x, &ctx, &mbmi->skip, NULL, NULL, 1 };
  int plane;

  mbmi->skip = 1;

  if (x->skip) return;

  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
#if CONFIG_VAR_TX
    // TODO(jingning): Clean this up.
    const struct macroblockd_plane *const pd = &xd->plane[plane];
    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
    const int mi_width = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
    const int mi_height = block_size_high[plane_bsize] >> tx_size_wide_log2[0];
    const TX_SIZE max_tx_size = max_txsize_rect_lookup[plane_bsize];
    const BLOCK_SIZE txb_size = txsize_to_bsize[max_tx_size];
    const int bw = block_size_wide[txb_size] >> tx_size_wide_log2[0];
    const int bh = block_size_high[txb_size] >> tx_size_wide_log2[0];
    int idx, idy;
    int block = 0;
    int step = tx_size_wide_unit[max_tx_size] * tx_size_high_unit[max_tx_size];
    av1_get_entropy_contexts(bsize, TX_4X4, pd, ctx.ta[plane], ctx.tl[plane]);
#else
    const struct macroblockd_plane *const pd = &xd->plane[plane];
    const TX_SIZE tx_size = plane ? get_uv_tx_size(mbmi, pd) : mbmi->tx_size;
    av1_get_entropy_contexts(bsize, tx_size, pd, ctx.ta[plane], ctx.tl[plane]);
#endif
#if !CONFIG_PVQ
    av1_subtract_plane(x, bsize, plane);
#endif
    arg.ta = ctx.ta[plane];
    arg.tl = ctx.tl[plane];

#if CONFIG_VAR_TX
    for (idy = 0; idy < mi_height; idy += bh) {
      for (idx = 0; idx < mi_width; idx += bw) {
        encode_block_inter(plane, block, idy, idx, plane_bsize, max_tx_size,
                           &arg);
        block += step;
      }
    }
#else
    av1_foreach_transformed_block_in_plane(xd, bsize, plane, encode_block,
                                           &arg);
#endif
  }
}

#if CONFIG_SUPERTX
void av1_encode_sb_supertx(AV1_COMMON *cm, MACROBLOCK *x, BLOCK_SIZE bsize) {
  MACROBLOCKD *const xd = &x->e_mbd;
  struct optimize_ctx ctx;
  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
  struct encode_b_args arg = { cm, x, &ctx, &mbmi->skip, NULL, NULL, 1 };
  int plane;

  mbmi->skip = 1;
  if (x->skip) return;

  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {