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/bitwriter.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/av1_quantize.h"
#include "av1/encoder/encodemb.h"
#if CONFIG_LV_MAP
#include "av1/encoder/encodetxb.h"
#endif
#include "av1/encoder/hybrid_fwd_txfm.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

#if CONFIG_CFL
#include "av1/common/cfl.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; }

static void subtract_block(const MACROBLOCKD *xd, int rows, int cols,
                           int16_t *diff, ptrdiff_t diff_stride,
                           const uint8_t *src8, ptrdiff_t src_stride,
                           const uint8_t *pred8, ptrdiff_t pred_stride) {
#if !CONFIG_HIGHBITDEPTH
  (void)xd;
#endif

  if (check_subtract_block_size(rows, cols)) {
#if CONFIG_HIGHBITDEPTH
    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
      aom_highbd_subtract_block_c(rows, cols, diff, diff_stride, src8,
                                  src_stride, pred8, pred_stride, xd->bd);
      return;
    }
#endif  // CONFIG_HIGHBITDEPTH
    aom_subtract_block_c(rows, cols, diff, diff_stride, src8, src_stride, pred8,
                         pred_stride);

    return;
  }

#if CONFIG_HIGHBITDEPTH
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    aom_highbd_subtract_block(rows, cols, diff, diff_stride, src8, src_stride,
                              pred8, pred_stride, xd->bd);
    return;
  }
#endif  // CONFIG_HIGHBITDEPTH
  aom_subtract_block(rows, cols, diff, diff_stride, src8, src_stride, pred8,
                     pred_stride);
}

void av1_subtract_txb(MACROBLOCK *x, int plane, BLOCK_SIZE plane_bsize,
                      int blk_col, int blk_row, TX_SIZE tx_size) {
  MACROBLOCKD *const xd = &x->e_mbd;
  struct macroblock_plane *const p = &x->plane[plane];
  const struct macroblockd_plane *const pd = &x->e_mbd.plane[plane];
  const int diff_stride = block_size_wide[plane_bsize];
  const int src_stride = p->src.stride;
  const int dst_stride = pd->dst.stride;
  const int tx1d_width = tx_size_wide[tx_size];
  const int tx1d_height = tx_size_high[tx_size];
  uint8_t *dst =
      &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
  uint8_t *src =
      &p->src.buf[(blk_row * src_stride + blk_col) << tx_size_wide_log2[0]];
  int16_t *src_diff =
      &p->src_diff[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
  subtract_block(xd, tx1d_height, tx1d_width, src_diff, diff_stride, src,
                 src_stride, dst, dst_stride);
}

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];
  const MACROBLOCKD *xd = &x->e_mbd;

  subtract_block(xd, bh, bw, p->src_diff, bw, p->src.buf, p->src.stride,
                 pd->dst.buf, pd->dst.stride);
}

// These numbers are empirically obtained.
static const int plane_rd_mult[REF_TYPES][PLANE_TYPES] = {
#if CONFIG_EC_ADAPT
  { 10, 7 }, { 8, 5 },
#else
  { 10, 6 }, { 8, 6 },
#endif
};

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

static INLINE unsigned int get_token_bit_costs(
    unsigned int token_costs[2][COEFF_CONTEXTS][ENTROPY_TOKENS], int skip_eob,
    int ctx, int token) {
  (void)skip_eob;
  return token_costs[token == ZERO_TOKEN || token == EOB_TOKEN][ctx][token];
}

#if !CONFIG_LV_MAP

typedef struct av1_token_state_greedy {
  int16_t token;
  tran_low_t qc;
  tran_low_t dqc;
} av1_token_state_greedy;

static int optimize_b_greedy(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_greedy tokens[MAX_TX_SQUARE + 1][2];
  uint8_t token_cache[MAX_TX_SQUARE];
  const 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);
  const int eob = p->eobs[block];
  const PLANE_TYPE plane_type = pd->plane_type;
  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 = av1_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];
#endif  // CONFIG_NEW_QUANT
  int sz = 0;
  const int64_t rddiv = mb->rddiv;
  int64_t rd_cost0, rd_cost1;
  int16_t t0, t1;
  int i, final_eob;
  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, xd->bd);
  unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
      mb->token_costs[txsize_sqr_map[tx_size]][plane_type][ref];
  const int default_eob = tx_size_2d[tx_size];

  assert(mb->qindex > 0);

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

  int64_t rdmult = (mb->rdmult * plane_rd_mult[ref][plane_type]) >> 1;

  int64_t rate0, rate1;
  for (i = 0; i < eob; i++) {
    const int rc = scan[i];
    int x = qcoeff[rc];
    t0 = av1_get_token(x);

    tokens[i][0].qc = x;
    tokens[i][0].token = t0;
    tokens[i][0].dqc = dqcoeff[rc];

    token_cache[rc] = av1_pt_energy_class[t0];
  }
  tokens[eob][0].token = EOB_TOKEN;
  tokens[eob][0].qc = 0;
  tokens[eob][0].dqc = 0;
  tokens[eob][1] = tokens[eob][0];

  unsigned int(*token_costs_ptr)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
      token_costs;

  final_eob = 0;

  int64_t eob_cost0, eob_cost1;

  const int ctx0 = ctx;
  /* Record the r-d cost */
  int64_t accu_rate = 0;
  int64_t accu_error = 0;

  rate0 = get_token_bit_costs(*(token_costs_ptr + band_translate[0]), 0, ctx0,
                              EOB_TOKEN);
  int64_t best_block_rd_cost = RDCOST(rdmult, rddiv, rate0, accu_error);

  // int64_t best_block_rd_cost_all0 = best_block_rd_cost;

  int x_prev = 1;

  for (i = 0; i < eob; i++) {
    const int rc = scan[i];
    int x = qcoeff[rc];
    sz = -(x < 0);

    int band_cur = band_translate[i];
    int ctx_cur = (i == 0) ? ctx : get_coef_context(nb, token_cache, i);
    int token_tree_sel_cur = (x_prev == 0);

    if (x == 0) {
      // no need to search when x == 0
      rate0 =
          get_token_bit_costs(*(token_costs_ptr + band_cur), token_tree_sel_cur,
                              ctx_cur, tokens[i][0].token);
      accu_rate += rate0;
      x_prev = 0;
      // accu_error does not change when x==0
    } else {
      /*  Computing distortion
       */
      // compute the distortion for the first candidate
      // and the distortion for quantizing to 0.
      int dx0 = abs(coeff[rc]) * (1 << shift);
#if CONFIG_HIGHBITDEPTH
      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
        dx0 >>= xd->bd - 8;
      }
#endif
      int64_t d0 = (int64_t)dx0 * dx0;

      int x_a = x - 2 * sz - 1;
      int64_t d2, d2_a;

      int dx;

#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

      dx = (dqcoeff[rc] - coeff[rc]) * (1 << shift);
#if CONFIG_HIGHBITDEPTH
      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
        int dx_sign = dx < 0 ? 1 : 0;
        dx = abs(dx) >> (xd->bd - 8);
        if (dx_sign) dx = -dx;
      }
#endif  // CONFIG_HIGHBITDEPTH
      d2 = (int64_t)dx * dx;

      /* compute the distortion for the second candidate
       * x_a = x - 2 * sz + 1;
       */
      if (x_a != 0) {
#if CONFIG_NEW_QUANT
        dx = av1_dequant_coeff_nuq(x, dqv, dequant_val[band_translate[i]]) -
             (coeff[rc] << shift);
#if CONFIG_HIGHBITDEPTH
        if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
          dx >>= xd->bd - 8;
        }
#endif  // CONFIG_HIGHBITDEPTH
#else   // CONFIG_NEW_QUANT
#if CONFIG_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_HIGHBITDEPTH
#endif  // CONFIG_NEW_QUANT
        d2_a = (int64_t)dx * dx;
      } else {
        d2_a = d0;
      }
      /*  Computing rates and r-d cost
       */

      int best_x, best_eob_x;
      int64_t base_bits, next_bits0, next_bits1;
      int64_t next_eob_bits0, next_eob_bits1;

      // rate cost of x
      base_bits = av1_get_token_cost(x, &t0, cat6_bits);
      rate0 = base_bits + get_token_bit_costs(*(token_costs_ptr + band_cur),
                                              token_tree_sel_cur, ctx_cur, t0);

      base_bits = av1_get_token_cost(x_a, &t1, cat6_bits);
      rate1 = base_bits + get_token_bit_costs(*(token_costs_ptr + band_cur),
                                              token_tree_sel_cur, ctx_cur, t1);

      next_bits0 = 0;
      next_bits1 = 0;
      next_eob_bits0 = 0;
      next_eob_bits1 = 0;

      if (i < default_eob - 1) {
        int ctx_next, token_tree_sel_next;
        int band_next = band_translate[i + 1];

        token_cache[rc] = av1_pt_energy_class[t0];
        ctx_next = get_coef_context(nb, token_cache, i + 1);
        token_tree_sel_next = (x == 0);

        next_bits0 = get_token_bit_costs(*(token_costs_ptr + band_next),
                                         token_tree_sel_next, ctx_next,
                                         tokens[i + 1][0].token);
        next_eob_bits0 =
            get_token_bit_costs(*(token_costs_ptr + band_next),
                                token_tree_sel_next, ctx_next, EOB_TOKEN);

        token_cache[rc] = av1_pt_energy_class[t1];
        ctx_next = get_coef_context(nb, token_cache, i + 1);
        token_tree_sel_next = (x_a == 0);

        next_bits1 = get_token_bit_costs(*(token_costs_ptr + band_next),
                                         token_tree_sel_next, ctx_next,
                                         tokens[i + 1][0].token);

        if (x_a != 0) {
          next_eob_bits1 =
              get_token_bit_costs(*(token_costs_ptr + band_next),
                                  token_tree_sel_next, ctx_next, EOB_TOKEN);
        }
      }

      rd_cost0 = RDCOST(rdmult, rddiv, (rate0 + next_bits0), d2);
      rd_cost1 = RDCOST(rdmult, rddiv, (rate1 + next_bits1), d2_a);

      best_x = (rd_cost1 < rd_cost0);

      eob_cost0 = RDCOST(rdmult, rddiv, (accu_rate + rate0 + next_eob_bits0),
                         (accu_error + d2 - d0));
      eob_cost1 = eob_cost0;
      if (x_a != 0) {
        eob_cost1 = RDCOST(rdmult, rddiv, (accu_rate + rate1 + next_eob_bits1),
                           (accu_error + d2_a - d0));
        best_eob_x = (eob_cost1 < eob_cost0);
      } else {
        best_eob_x = 0;
      }

      int dqc, dqc_a = 0;

      dqc = dqcoeff[rc];
      if (best_x + best_eob_x) {
        if (x_a != 0) {
#if CONFIG_NEW_QUANT
          dqc_a = av1_dequant_abscoeff_nuq(abs(x_a), dqv,
                                           dequant_val[band_translate[i]]);
          dqc_a = shift ? ROUND_POWER_OF_TWO(dqc_a, shift) : dqc_a;
          if (sz) dqc_a = -dqc_a;
#else
          if (x_a < 0)
            dqc_a = -((-x_a * dqv) >> shift);
          else
            dqc_a = (x_a * dqv) >> shift;
#endif  // CONFIG_NEW_QUANT
        } else {
          dqc_a = 0;
        }  // if (x_a != 0)
      }

      // record the better quantized value
      if (best_x) {
        qcoeff[rc] = x_a;
        dqcoeff[rc] = dqc_a;

        accu_rate += rate1;
        accu_error += d2_a - d0;
        assert(d2_a <= d0);

        token_cache[rc] = av1_pt_energy_class[t1];
      } else {
        accu_rate += rate0;
        accu_error += d2 - d0;
        assert(d2 <= d0);

        token_cache[rc] = av1_pt_energy_class[t0];
      }

      x_prev = qcoeff[rc];

      // determine whether to move the eob position to i+1
      int64_t best_eob_cost_i = eob_cost0;

      tokens[i][1].token = t0;
      tokens[i][1].qc = x;
      tokens[i][1].dqc = dqc;

      if ((x_a != 0) && (best_eob_x)) {
        best_eob_cost_i = eob_cost1;

        tokens[i][1].token = t1;
        tokens[i][1].qc = x_a;
        tokens[i][1].dqc = dqc_a;
      }

      if (best_eob_cost_i < best_block_rd_cost) {
        best_block_rd_cost = best_eob_cost_i;
        final_eob = i + 1;
      }
    }  // if (x==0)
  }    // for (i)

  assert(final_eob <= eob);
  if (final_eob > 0) {
    assert(tokens[final_eob - 1][1].qc != 0);
    i = final_eob - 1;
    int rc = scan[i];
    qcoeff[rc] = tokens[i][1].qc;
    dqcoeff[rc] = tokens[i][1].dqc;
  }

  for (i = final_eob; i < eob; i++) {
    int rc = scan[i];
    qcoeff[rc] = 0;
    dqcoeff[rc] = 0;
  }

  mb->plane[plane].eobs[block] = final_eob;
  return final_eob;
}

#endif  // !CONFIG_LV_MAP

int av1_optimize_b(const AV1_COMMON *cm, MACROBLOCK *mb, int plane, int block,
                   BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
                   const ENTROPY_CONTEXT *a, const ENTROPY_CONTEXT *l) {
  MACROBLOCKD *const xd = &mb->e_mbd;
  struct macroblock_plane *const p = &mb->plane[plane];
  const int eob = p->eobs[block];
  assert((mb->qindex == 0) ^ (xd->lossless[xd->mi[0]->mbmi.segment_id] == 0));
  if (eob == 0) return eob;
  if (xd->lossless[xd->mi[0]->mbmi.segment_id]) return eob;
#if CONFIG_PVQ
  (void)cm;
  (void)tx_size;
  (void)a;
  (void)l;
  return eob;
#endif

#if !CONFIG_LV_MAP
  (void)plane_bsize;
#if CONFIG_VAR_TX
  int ctx = get_entropy_context(tx_size, a, l);
#else
  int ctx = combine_entropy_contexts(*a, *l);
#endif  // CONFIG_VAR_TX
  return optimize_b_greedy(cm, mb, plane, block, tx_size, ctx);
#else   // !CONFIG_LV_MAP
  TXB_CTX txb_ctx;
  get_txb_ctx(plane_bsize, tx_size, plane, a, l, &txb_ctx);
  return av1_optimize_txb(cm, mb, plane, block, tx_size, &txb_ctx);
#endif  // !CONFIG_LV_MAP
}

#if !CONFIG_PVQ
#if CONFIG_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] = {
#if !CONFIG_NEW_QUANT
      { 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 },
#else   // !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 }
    };

#else

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] = {
#if !CONFIG_NEW_QUANT
                                         { av1_quantize_fp_facade },
                                         { av1_quantize_b_facade },
                                         { av1_quantize_dc_facade },
#else   // !CONFIG_NEW_QUANT
                                         { av1_quantize_fp_nuq_facade },
                                         { av1_quantize_b_nuq_facade },
                                         { av1_quantize_dc_nuq_facade },
#endif  // !CONFIG_NEW_QUANT
                                         { NULL }
                                       };
#endif  // CONFIG_HIGHBITDEPTH
#endif  // CONFIG_PVQ

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;
  MB_MODE_INFO *const mbmi = &xd->mi[0]->mbmi;
#if !(CONFIG_PVQ || CONFIG_DAALA_DIST)
  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 = get_plane_type(plane);
  TX_TYPE tx_type = get_tx_type(plane_type, xd, block, tx_size);
  const int is_inter = is_inter_block(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 = 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 || CONFIG_DAALA_DIST
  uint8_t *dst;
  int16_t *pred;
  const int dst_stride = pd->dst.stride;
  int tx_blk_size;
  int i, j;
#endif

#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[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];
  qparam.log_scale = av1_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
  tran_low_t *ref_coeff = BLOCK_OFFSET(pd->pvq_ref_coeff, block);
  int skip = 1;
  PVQ_INFO *pvq_info = NULL;
  uint8_t *src;
  int16_t *src_int16;
  const int src_stride = p->src.stride;

  (void)ctx;
  (void)scan_order;
  (void)qcoeff;

  if (x->pvq_coded) {
    assert(block < MAX_PVQ_BLOCKS_IN_SB);
    pvq_info = &x->pvq[block][plane];
  }
  src = &p->src.buf[(blk_row * src_stride + blk_col) << tx_size_wide_log2[0]];
  src_int16 =
      &p->src_int16[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];

  // transform block size in pixels
  tx_blk_size = tx_size_wide[tx_size];
#if CONFIG_HIGHBITDEPTH
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    for (j = 0; j < tx_blk_size; j++)
      for (i = 0; i < tx_blk_size; i++)
        src_int16[diff_stride * j + i] =
            CONVERT_TO_SHORTPTR(src)[src_stride * j + i];
  } else {
#endif  // CONFIG_HIGHBITDEPTH
    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];
#if CONFIG_HIGHBITDEPTH
  }
#endif  // CONFIG_HIGHBITDEPTH
#endif

#if CONFIG_PVQ || CONFIG_DAALA_DIST
  dst = &pd->dst.buf[(blk_row * dst_stride + blk_col) << tx_size_wide_log2[0]];
  pred = &pd->pred[(blk_row * diff_stride + blk_col) << tx_size_wide_log2[0]];

  // 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
#if CONFIG_HIGHBITDEPTH
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    for (j = 0; j < tx_blk_size; j++)
      for (i = 0; i < tx_blk_size; i++)
        pred[diff_stride * j + i] =
            CONVERT_TO_SHORTPTR(dst)[dst_stride * j + i];
  } else {
#endif  // CONFIG_HIGHBITDEPTH
    for (j = 0; j < tx_blk_size; j++)
      for (i = 0; i < tx_blk_size; i++)
        pred[diff_stride * j + i] = dst[dst_stride * j + i];
#if CONFIG_HIGHBITDEPTH
  }
#endif  // CONFIG_HIGHBITDEPTH
#endif

  (void)ctx;

  fwd_txfm_param.tx_type = tx_type;
  fwd_txfm_param.tx_size = tx_size;
  fwd_txfm_param.lossless = xd->lossless[mbmi->segment_id];

#if !CONFIG_PVQ
#if CONFIG_HIGHBITDEPTH
  fwd_txfm_param.bd = xd->bd;
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    av1_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);
      }
    }
#if CONFIG_LV_MAP
    p->txb_entropy_ctx[block] =
        (uint8_t)av1_get_txb_entropy_context(qcoeff, scan_order, *eob);
#endif  // CONFIG_LV_MAP
    return;
  }
#endif  // CONFIG_HIGHBITDEPTH
  av1_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);
    }
  }
#if CONFIG_LV_MAP
  p->txb_entropy_ctx[block] =
      (uint8_t)av1_get_txb_entropy_context(qcoeff, scan_order, *eob);
#endif  // CONFIG_LV_MAP
#else   // #if !CONFIG_PVQ
  (void)xform_quant_idx;
#if CONFIG_HIGHBITDEPTH
  fwd_txfm_param.bd = xd->bd;
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    av1_highbd_fwd_txfm(src_int16, coeff, diff_stride, &fwd_txfm_param);
    av1_highbd_fwd_txfm(pred, ref_coeff, diff_stride, &fwd_txfm_param);
  } else {
#endif
    av1_fwd_txfm(src_int16, coeff, diff_stride, &fwd_txfm_param);
    av1_fwd_txfm(pred, ref_coeff, diff_stride, &fwd_txfm_param);
#if CONFIG_HIGHBITDEPTH
  }
#endif

  // PVQ for inter mode block
  if (!x->skip_block) {
    PVQ_SKIP_TYPE ac_dc_coded =
        av1_pvq_encode_helper(x,
                              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
    skip = ac_dc_coded == PVQ_SKIP;
  }
  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;
#if !CONFIG_PVQ
  ENTROPY_CONTEXT *a, *l;
#endif
#if CONFIG_VAR_TX
  int bw = block_size_wide[plane_bsize] >> tx_size_wide_log2[0];
#endif
  dst = &pd->dst
             .buf[(blk_row * pd->dst.stride + blk_col) << tx_size_wide_log2[0]];

#if !CONFIG_PVQ
  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
#else
  ctx = 0;
#endif  // CONFIG_PVQ

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

  if (x->blk_skip[plane][blk_row * bw + blk_col] == 0) {
#else
  {
#endif
    av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
                    ctx, AV1_XFORM_QUANT_FP);
  }
#if CONFIG_VAR_TX
  else {
    p->eobs[block] = 0;
  }
#endif

#if !CONFIG_PVQ
  av1_optimize_b(cm, x, plane, block, plane_bsize, tx_size, a, l);

  av1_set_txb_context(x, plane, block, tx_size, a, l);

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

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

  if (x->pvq_skip[plane]) return;
#endif
  TX_TYPE tx_type = get_tx_type(pd->plane_type, xd, block, tx_size);
  av1_inverse_transform_block(xd, dqcoeff, tx_type, tx_size, dst,
                              pd->dst.stride, p->eobs[block]);
}

#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 {
    assert(tx_size < TX_SIZES_ALL);
    const TX_SIZE sub_txs = sub_tx_size_map[tx_size];
    assert(sub_txs < tx_size);
    // This is the square transform block partition entry point.
    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);
      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[(blk_row * pd->dst.stride + blk_col) << tx_size_wide_log2[0]];

  av1_xform_quant(cm, x, plane, block, blk_row, blk_col, plane_bsize, tx_size,
                  ctx, AV1_XFORM_QUANT_B);
#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
#if CONFIG_HIGHBITDEPTH
      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
        for (j = 0; j < tx_blk_size; j++)
          for (i = 0; i < tx_blk_size; i++)
            CONVERT_TO_SHORTPTR(dst)[j * pd->dst.stride + i] = 0;
      } else {
#endif  // CONFIG_HIGHBITDEPTH
        for (j = 0; j < tx_blk_size; j++)
          for (i = 0; i < tx_blk_size; i++) dst[j * pd->dst.stride + i] = 0;
#if CONFIG_HIGHBITDEPTH
      }
#endif  // CONFIG_HIGHBITDEPTH
    }
#endif  // !CONFIG_PVQ
#if CONFIG_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_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, int mi_row,
                   int mi_col) {
  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_CB4X4 && !CONFIG_CHROMA_2X2
    const int subsampling_x = xd->plane[plane].subsampling_x;
    const int subsampling_y = xd->plane[plane].subsampling_y;

    if (!is_chroma_reference(mi_row, mi_col, bsize, subsampling_x,
                             subsampling_y))
      continue;

    bsize = scale_chroma_bsize(bsize, subsampling_x, subsampling_y);
#else
    (void)mi_row;
    (void)mi_col;
#endif

#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 = get_vartx_max_txsize(mbmi, 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, 0, pd, ctx.ta[plane], ctx.tl[plane]);
#else
    const struct macroblockd_plane *const pd = &xd->plane[plane];
    const TX_SIZE tx_size = get_tx_size(plane, xd);
    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;