rdopt.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 <math.h>

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

#include "aom_dsp/aom_dsp_common.h"
#include "aom_dsp/blend.h"
#include "aom_mem/aom_mem.h"
#include "aom_ports/mem.h"
#include "aom_ports/system_state.h"

#include "av1/common/common.h"
#include "av1/common/common_data.h"
#include "av1/common/entropy.h"
#include "av1/common/entropymode.h"
#include "av1/common/idct.h"
#include "av1/common/mvref_common.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/scan.h"
#include "av1/common/seg_common.h"
#if CONFIG_LV_MAP
#include "av1/common/txb_common.h"
#endif
#if CONFIG_WARPED_MOTION
#include "av1/common/warped_motion.h"
#endif  // CONFIG_WARPED_MOTION

#include "av1/encoder/aq_variance.h"
#include "av1/encoder/av1_quantize.h"
#include "av1/encoder/cost.h"
#include "av1/encoder/encodemb.h"
#include "av1/encoder/encodemv.h"
#include "av1/encoder/encoder.h"
#if CONFIG_LV_MAP
#include "av1/encoder/encodetxb.h"
#endif
#include "av1/encoder/hybrid_fwd_txfm.h"
#include "av1/encoder/mcomp.h"
#if CONFIG_PALETTE
#include "av1/encoder/palette.h"
#endif  // CONFIG_PALETTE
#include "av1/encoder/ratectrl.h"
#include "av1/encoder/rd.h"
#include "av1/encoder/rdopt.h"
#include "av1/encoder/tokenize.h"
#if CONFIG_PVQ
#include "av1/encoder/pvq_encoder.h"
#endif  // CONFIG_PVQ
#if CONFIG_PVQ || CONFIG_DAALA_DIST
#include "av1/common/pvq.h"
#endif  // CONFIG_PVQ || CONFIG_DAALA_DIST
#if CONFIG_DUAL_FILTER
#define DUAL_FILTER_SET_SIZE (SWITCHABLE_FILTERS * SWITCHABLE_FILTERS)
#if USE_EXTRA_FILTER
static const int filter_sets[DUAL_FILTER_SET_SIZE][2] = {
  { 0, 0 }, { 0, 1 }, { 0, 2 }, { 0, 3 }, { 1, 0 }, { 1, 1 },
  { 1, 2 }, { 1, 3 }, { 2, 0 }, { 2, 1 }, { 2, 2 }, { 2, 3 },
  { 3, 0 }, { 3, 1 }, { 3, 2 }, { 3, 3 },
};
#else   // USE_EXTRA_FILTER
static const int filter_sets[DUAL_FILTER_SET_SIZE][2] = {
  { 0, 0 }, { 0, 1 }, { 0, 2 }, { 1, 0 }, { 1, 1 },
  { 1, 2 }, { 2, 0 }, { 2, 1 }, { 2, 2 },
};
#endif  // USE_EXTRA_FILTER
#endif  // CONFIG_DUAL_FILTER

#if CONFIG_EXT_REFS

#define LAST_FRAME_MODE_MASK                                      \
  ((1 << INTRA_FRAME) | (1 << LAST2_FRAME) | (1 << LAST3_FRAME) | \
   (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF_FRAME))
#define LAST2_FRAME_MODE_MASK                                    \
  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST3_FRAME) | \
   (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF_FRAME))
#define LAST3_FRAME_MODE_MASK                                    \
  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) | \
   (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF_FRAME))
#define GOLDEN_FRAME_MODE_MASK                                   \
  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) | \
   (1 << LAST3_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF_FRAME))
#define BWDREF_FRAME_MODE_MASK                                   \
  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) | \
   (1 << LAST3_FRAME) | (1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME))
#define ALTREF_FRAME_MODE_MASK                                   \
  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) | \
   (1 << LAST3_FRAME) | (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME))

#else

#define LAST_FRAME_MODE_MASK \
  ((1 << GOLDEN_FRAME) | (1 << ALTREF_FRAME) | (1 << INTRA_FRAME))
#define GOLDEN_FRAME_MODE_MASK \
  ((1 << LAST_FRAME) | (1 << ALTREF_FRAME) | (1 << INTRA_FRAME))
#define ALTREF_FRAME_MODE_MASK \
  ((1 << LAST_FRAME) | (1 << GOLDEN_FRAME) | (1 << INTRA_FRAME))

#endif  // CONFIG_EXT_REFS

#if CONFIG_EXT_REFS
#define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | (1 << BWDREF_FRAME) | 0x01)
#else
#define SECOND_REF_FRAME_MASK ((1 << ALTREF_FRAME) | 0x01)
#endif  // CONFIG_EXT_REFS

#define MIN_EARLY_TERM_INDEX 3
#define NEW_MV_DISCOUNT_FACTOR 8

#if CONFIG_EXT_INTRA
#define ANGLE_SKIP_THRESH 10
#define FILTER_FAST_SEARCH 1
#endif  // CONFIG_EXT_INTRA

const double ADST_FLIP_SVM[8] = { -6.6623, -2.8062, -3.2531, 3.1671,    // vert
                                  -7.7051, -3.2234, -3.6193, 3.4533 };  // horz

typedef struct {
  PREDICTION_MODE mode;
  MV_REFERENCE_FRAME ref_frame[2];
} MODE_DEFINITION;

typedef struct { MV_REFERENCE_FRAME ref_frame[2]; } REF_DEFINITION;

struct rdcost_block_args {
  const AV1_COMP *cpi;
  MACROBLOCK *x;
  ENTROPY_CONTEXT t_above[2 * MAX_MIB_SIZE];
  ENTROPY_CONTEXT t_left[2 * MAX_MIB_SIZE];
  RD_STATS rd_stats;
  int64_t this_rd;
  int64_t best_rd;
  int exit_early;
  int use_fast_coef_costing;
};

#define LAST_NEW_MV_INDEX 6
static const MODE_DEFINITION av1_mode_order[MAX_MODES] = {
  { NEARESTMV, { LAST_FRAME, NONE_FRAME } },
#if CONFIG_EXT_REFS
  { NEARESTMV, { LAST2_FRAME, NONE_FRAME } },
  { NEARESTMV, { LAST3_FRAME, NONE_FRAME } },
  { NEARESTMV, { BWDREF_FRAME, NONE_FRAME } },
#endif  // CONFIG_EXT_REFS
  { NEARESTMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEARESTMV, { GOLDEN_FRAME, NONE_FRAME } },

  { DC_PRED, { INTRA_FRAME, NONE_FRAME } },

  { NEWMV, { LAST_FRAME, NONE_FRAME } },
#if CONFIG_EXT_REFS
  { NEWMV, { LAST2_FRAME, NONE_FRAME } },
  { NEWMV, { LAST3_FRAME, NONE_FRAME } },
  { NEWMV, { BWDREF_FRAME, NONE_FRAME } },
#endif  // CONFIG_EXT_REFS
  { NEWMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEWMV, { GOLDEN_FRAME, NONE_FRAME } },

  { NEARMV, { LAST_FRAME, NONE_FRAME } },
#if CONFIG_EXT_REFS
  { NEARMV, { LAST2_FRAME, NONE_FRAME } },
  { NEARMV, { LAST3_FRAME, NONE_FRAME } },
  { NEARMV, { BWDREF_FRAME, NONE_FRAME } },
#endif  // CONFIG_EXT_REFS
  { NEARMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEARMV, { GOLDEN_FRAME, NONE_FRAME } },

  { ZEROMV, { LAST_FRAME, NONE_FRAME } },
#if CONFIG_EXT_REFS
  { ZEROMV, { LAST2_FRAME, NONE_FRAME } },
  { ZEROMV, { LAST3_FRAME, NONE_FRAME } },
  { ZEROMV, { BWDREF_FRAME, NONE_FRAME } },
#endif  // CONFIG_EXT_REFS
  { ZEROMV, { GOLDEN_FRAME, NONE_FRAME } },
  { ZEROMV, { ALTREF_FRAME, NONE_FRAME } },

// TODO(zoeliu): May need to reconsider the order on the modes to check

#if CONFIG_EXT_INTER
  { NEAREST_NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
#if CONFIG_EXT_REFS
  { NEAREST_NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
#endif  // CONFIG_EXT_REFS
  { NEAREST_NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
#if CONFIG_EXT_REFS
  { NEAREST_NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
#endif  // CONFIG_EXT_REFS

#else  // CONFIG_EXT_INTER

  { NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
#if CONFIG_EXT_REFS
  { NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
#endif  // CONFIG_EXT_REFS
  { NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
#if CONFIG_EXT_REFS
  { NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
#endif  // CONFIG_EXT_REFS
#endif  // CONFIG_EXT_INTER

  { TM_PRED, { INTRA_FRAME, NONE_FRAME } },

#if CONFIG_ALT_INTRA
  { SMOOTH_PRED, { INTRA_FRAME, NONE_FRAME } },
#if CONFIG_SMOOTH_HV
  { SMOOTH_V_PRED, { INTRA_FRAME, NONE_FRAME } },
  { SMOOTH_H_PRED, { INTRA_FRAME, NONE_FRAME } },
#endif  // CONFIG_SMOOTH_HV
#endif  // CONFIG_ALT_INTRA

#if CONFIG_EXT_INTER
  { NEAR_NEARMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEW_NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEAREST_NEWMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEW_NEARMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEAR_NEWMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEW_NEWMV, { LAST_FRAME, ALTREF_FRAME } },
  { ZERO_ZEROMV, { LAST_FRAME, ALTREF_FRAME } },

#if CONFIG_EXT_REFS
  { NEAR_NEARMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEW_NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEAREST_NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEW_NEARMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEAR_NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEW_NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
  { ZERO_ZEROMV, { LAST2_FRAME, ALTREF_FRAME } },

  { NEAR_NEARMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEW_NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEAREST_NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEW_NEARMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEAR_NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEW_NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
  { ZERO_ZEROMV, { LAST3_FRAME, ALTREF_FRAME } },
#endif  // CONFIG_EXT_REFS

  { NEAR_NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEW_NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEAREST_NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEW_NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEAR_NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEW_NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { ZERO_ZEROMV, { GOLDEN_FRAME, ALTREF_FRAME } },

#if CONFIG_EXT_REFS
  { NEAR_NEARMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEW_NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEAREST_NEWMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEW_NEARMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEAR_NEWMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEW_NEWMV, { LAST_FRAME, BWDREF_FRAME } },
  { ZERO_ZEROMV, { LAST_FRAME, BWDREF_FRAME } },

  { NEAR_NEARMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEW_NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEAREST_NEWMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEW_NEARMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEAR_NEWMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEW_NEWMV, { LAST2_FRAME, BWDREF_FRAME } },
  { ZERO_ZEROMV, { LAST2_FRAME, BWDREF_FRAME } },

  { NEAR_NEARMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEW_NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEAREST_NEWMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEW_NEARMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEAR_NEWMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEW_NEWMV, { LAST3_FRAME, BWDREF_FRAME } },
  { ZERO_ZEROMV, { LAST3_FRAME, BWDREF_FRAME } },

  { NEAR_NEARMV, { GOLDEN_FRAME, BWDREF_FRAME } },
  { NEW_NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
  { NEAREST_NEWMV, { GOLDEN_FRAME, BWDREF_FRAME } },
  { NEW_NEARMV, { GOLDEN_FRAME, BWDREF_FRAME } },
  { NEAR_NEWMV, { GOLDEN_FRAME, BWDREF_FRAME } },
  { NEW_NEWMV, { GOLDEN_FRAME, BWDREF_FRAME } },
  { ZERO_ZEROMV, { GOLDEN_FRAME, BWDREF_FRAME } },
#endif  // CONFIG_EXT_REFS

#else  // CONFIG_EXT_INTER

  { NEARMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEWMV, { LAST_FRAME, ALTREF_FRAME } },
#if CONFIG_EXT_REFS
  { NEARMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEWMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEARMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEWMV, { LAST3_FRAME, ALTREF_FRAME } },
#endif  // CONFIG_EXT_REFS
  { NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEWMV, { GOLDEN_FRAME, ALTREF_FRAME } },

#if CONFIG_EXT_REFS
  { NEARMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEWMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEARMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEWMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEARMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEWMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEARMV, { GOLDEN_FRAME, BWDREF_FRAME } },
  { NEWMV, { GOLDEN_FRAME, BWDREF_FRAME } },
#endif  // CONFIG_EXT_REFS

  { ZEROMV, { LAST_FRAME, ALTREF_FRAME } },
#if CONFIG_EXT_REFS
  { ZEROMV, { LAST2_FRAME, ALTREF_FRAME } },
  { ZEROMV, { LAST3_FRAME, ALTREF_FRAME } },
#endif  // CONFIG_EXT_REFS
  { ZEROMV, { GOLDEN_FRAME, ALTREF_FRAME } },

#if CONFIG_EXT_REFS
  { ZEROMV, { LAST_FRAME, BWDREF_FRAME } },
  { ZEROMV, { LAST2_FRAME, BWDREF_FRAME } },
  { ZEROMV, { LAST3_FRAME, BWDREF_FRAME } },
  { ZEROMV, { GOLDEN_FRAME, BWDREF_FRAME } },
#endif  // CONFIG_EXT_REFS

#endif  // CONFIG_EXT_INTER

  { H_PRED, { INTRA_FRAME, NONE_FRAME } },
  { V_PRED, { INTRA_FRAME, NONE_FRAME } },
  { D135_PRED, { INTRA_FRAME, NONE_FRAME } },
  { D207_PRED, { INTRA_FRAME, NONE_FRAME } },
  { D153_PRED, { INTRA_FRAME, NONE_FRAME } },
  { D63_PRED, { INTRA_FRAME, NONE_FRAME } },
  { D117_PRED, { INTRA_FRAME, NONE_FRAME } },
  { D45_PRED, { INTRA_FRAME, NONE_FRAME } },

#if CONFIG_EXT_INTER
  { ZEROMV, { LAST_FRAME, INTRA_FRAME } },
  { NEARESTMV, { LAST_FRAME, INTRA_FRAME } },
  { NEARMV, { LAST_FRAME, INTRA_FRAME } },
  { NEWMV, { LAST_FRAME, INTRA_FRAME } },

#if CONFIG_EXT_REFS
  { ZEROMV, { LAST2_FRAME, INTRA_FRAME } },
  { NEARESTMV, { LAST2_FRAME, INTRA_FRAME } },
  { NEARMV, { LAST2_FRAME, INTRA_FRAME } },
  { NEWMV, { LAST2_FRAME, INTRA_FRAME } },

  { ZEROMV, { LAST3_FRAME, INTRA_FRAME } },
  { NEARESTMV, { LAST3_FRAME, INTRA_FRAME } },
  { NEARMV, { LAST3_FRAME, INTRA_FRAME } },
  { NEWMV, { LAST3_FRAME, INTRA_FRAME } },
#endif  // CONFIG_EXT_REFS

  { ZEROMV, { GOLDEN_FRAME, INTRA_FRAME } },
  { NEARESTMV, { GOLDEN_FRAME, INTRA_FRAME } },
  { NEARMV, { GOLDEN_FRAME, INTRA_FRAME } },
  { NEWMV, { GOLDEN_FRAME, INTRA_FRAME } },

#if CONFIG_EXT_REFS
  { ZEROMV, { BWDREF_FRAME, INTRA_FRAME } },
  { NEARESTMV, { BWDREF_FRAME, INTRA_FRAME } },
  { NEARMV, { BWDREF_FRAME, INTRA_FRAME } },
  { NEWMV, { BWDREF_FRAME, INTRA_FRAME } },
#endif  // CONFIG_EXT_REFS

  { ZEROMV, { ALTREF_FRAME, INTRA_FRAME } },
  { NEARESTMV, { ALTREF_FRAME, INTRA_FRAME } },
  { NEARMV, { ALTREF_FRAME, INTRA_FRAME } },
  { NEWMV, { ALTREF_FRAME, INTRA_FRAME } },
#endif  // CONFIG_EXT_INTER
};

#if CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA || CONFIG_PALETTE
static INLINE int write_uniform_cost(int n, int v) {
  const int l = get_unsigned_bits(n);
  const int m = (1 << l) - n;
  if (l == 0) return 0;
  if (v < m)
    return (l - 1) * av1_cost_bit(128, 0);
  else
    return l * av1_cost_bit(128, 0);
}
#endif  // CONFIG_EXT_INTRA || CONFIG_FILTER_INTRA || CONFIG_PALETTE

// constants for prune 1 and prune 2 decision boundaries
#define FAST_EXT_TX_CORR_MID 0.0
#define FAST_EXT_TX_EDST_MID 0.1
#define FAST_EXT_TX_CORR_MARGIN 0.5
#define FAST_EXT_TX_EDST_MARGIN 0.3

#if CONFIG_DAALA_DIST
static int od_compute_var_4x4(od_coeff *x, int stride) {
  int sum;
  int s2;
  int i;
  sum = 0;
  s2 = 0;
  for (i = 0; i < 4; i++) {
    int j;
    for (j = 0; j < 4; j++) {
      int t;

      t = x[i * stride + j];
      sum += t;
      s2 += t * t;
    }
  }
  // TODO(yushin) : Check wheter any changes are required for high bit depth.
  return (s2 - (sum * sum >> 4)) >> 4;
}

/* OD_DIST_LP_MID controls the frequency weighting filter used for computing
   the distortion. For a value X, the filter is [1 X 1]/(X + 2) and
   is applied both horizontally and vertically. For X=5, the filter is
   a good approximation for the OD_QM8_Q4_HVS quantization matrix. */
#define OD_DIST_LP_MID (5)
#define OD_DIST_LP_NORM (OD_DIST_LP_MID + 2)

static double od_compute_dist_8x8(int qm, int use_activity_masking, od_coeff *x,
                                  od_coeff *y, od_coeff *e_lp, int stride) {
  double sum;
  int min_var;
  double mean_var;
  double var_stat;
  double activity;
  double calibration;
  int i;
  int j;
  double vardist;

  vardist = 0;
  OD_ASSERT(qm != OD_FLAT_QM);
  (void)qm;
#if 1
  min_var = INT_MAX;
  mean_var = 0;
  for (i = 0; i < 3; i++) {
    for (j = 0; j < 3; j++) {
      int varx;
      int vary;
      varx = od_compute_var_4x4(x + 2 * i * stride + 2 * j, stride);
      vary = od_compute_var_4x4(y + 2 * i * stride + 2 * j, stride);
      min_var = OD_MINI(min_var, varx);
      mean_var += 1. / (1 + varx);
      /* The cast to (double) is to avoid an overflow before the sqrt.*/
      vardist += varx - 2 * sqrt(varx * (double)vary) + vary;
    }
  }
  /* We use a different variance statistic depending on whether activity
     masking is used, since the harmonic mean appeared slghtly worse with
     masking off. The calibration constant just ensures that we preserve the
     rate compared to activity=1. */
  if (use_activity_masking) {
    calibration = 1.95;
    var_stat = 9. / mean_var;
  } else {
    calibration = 1.62;
    var_stat = min_var;
  }
  /* 1.62 is a calibration constant, 0.25 is a noise floor and 1/6 is the
     activity masking constant. */
  activity = calibration * pow(.25 + var_stat, -1. / 6);
#else
  activity = 1;
#endif  // 1
  sum = 0;
  for (i = 0; i < 8; i++) {
    for (j = 0; j < 8; j++)
      sum += e_lp[i * stride + j] * (double)e_lp[i * stride + j];
  }
  /* Normalize the filter to unit DC response. */
  sum *= 1. / (OD_DIST_LP_NORM * OD_DIST_LP_NORM * OD_DIST_LP_NORM *
               OD_DIST_LP_NORM);
  return activity * activity * (sum + vardist);
}

// Note : Inputs x and y are in a pixel domain
static double od_compute_dist(int qm, int activity_masking, od_coeff *x,
                              od_coeff *y, int bsize_w, int bsize_h,
                              int qindex) {
  int i;
  double sum;
  sum = 0;

  assert(bsize_w >= 8 && bsize_h >= 8);

  if (qm == OD_FLAT_QM) {
    for (i = 0; i < bsize_w * bsize_h; i++) {
      double tmp;
      tmp = x[i] - y[i];
      sum += tmp * tmp;
    }
  } else {
    int j;
    DECLARE_ALIGNED(16, od_coeff, e[MAX_TX_SQUARE]);
    DECLARE_ALIGNED(16, od_coeff, tmp[MAX_TX_SQUARE]);
    DECLARE_ALIGNED(16, od_coeff, e_lp[MAX_TX_SQUARE]);
    int mid = OD_DIST_LP_MID;
    for (i = 0; i < bsize_h; i++) {
      for (j = 0; j < bsize_w; j++) {
        e[i * bsize_w + j] = x[i * bsize_w + j] - y[i * bsize_w + j];
      }
    }
    for (i = 0; i < bsize_h; i++) {
      tmp[i * bsize_w] = mid * e[i * bsize_w] + 2 * e[i * bsize_w + 1];
      tmp[i * bsize_w + bsize_w - 1] =
          mid * e[i * bsize_w + bsize_w - 1] + 2 * e[i * bsize_w + bsize_w - 2];
      for (j = 1; j < bsize_w - 1; j++) {
        tmp[i * bsize_w + j] = mid * e[i * bsize_w + j] +
                               e[i * bsize_w + j - 1] + e[i * bsize_w + j + 1];
      }
    }
    for (j = 0; j < bsize_w; j++) {
      e_lp[j] = mid * tmp[j] + 2 * tmp[bsize_w + j];
      e_lp[(bsize_h - 1) * bsize_w + j] =
          mid * tmp[(bsize_h - 1) * bsize_w + j] +
          2 * tmp[(bsize_h - 2) * bsize_w + j];
    }
    for (i = 1; i < bsize_h - 1; i++) {
      for (j = 0; j < bsize_w; j++) {
        e_lp[i * bsize_w + j] = mid * tmp[i * bsize_w + j] +
                                tmp[(i - 1) * bsize_w + j] +
                                tmp[(i + 1) * bsize_w + j];
      }
    }
    for (i = 0; i < bsize_h; i += 8) {
      for (j = 0; j < bsize_w; j += 8) {
        sum += od_compute_dist_8x8(qm, activity_masking, &x[i * bsize_w + j],
                                   &y[i * bsize_w + j], &e_lp[i * bsize_w + j],
                                   bsize_w);
      }
    }
    /* Scale according to linear regression against SSE, for 8x8 blocks. */
    if (activity_masking) {
      sum *= 2.2 + (1.7 - 2.2) * (qindex - 99) / (210 - 99) +
             (qindex < 99 ? 2.5 * (qindex - 99) / 99 * (qindex - 99) / 99 : 0);
    } else {
      sum *= qindex >= 128
                 ? 1.4 + (0.9 - 1.4) * (qindex - 128) / (209 - 128)
                 : qindex <= 43
                       ? 1.5 + (2.0 - 1.5) * (qindex - 43) / (16 - 43)
                       : 1.5 + (1.4 - 1.5) * (qindex - 43) / (128 - 43);
    }
  }
  return sum;
}

int64_t av1_daala_dist(const uint8_t *src, int src_stride, const uint8_t *dst,
                       int dst_stride, int bsw, int bsh, int qm,
                       int use_activity_masking, int qindex) {
  int i, j;
  int64_t d;
  DECLARE_ALIGNED(16, od_coeff, orig[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(16, od_coeff, rec[MAX_TX_SQUARE]);

  assert(qm == OD_HVS_QM);

  for (j = 0; j < bsh; j++)
    for (i = 0; i < bsw; i++) orig[j * bsw + i] = src[j * src_stride + i];

  for (j = 0; j < bsh; j++)
    for (i = 0; i < bsw; i++) rec[j * bsw + i] = dst[j * dst_stride + i];

  d = (int64_t)od_compute_dist(qm, use_activity_masking, orig, rec, bsw, bsh,
                               qindex);
  return d;
}
#endif  // CONFIG_DAALA_DIST

static void get_energy_distribution_fine(const AV1_COMP *cpi, BLOCK_SIZE bsize,
                                         const uint8_t *src, int src_stride,
                                         const uint8_t *dst, int dst_stride,
                                         double *hordist, double *verdist) {
  const int bw = block_size_wide[bsize];
  const int bh = block_size_high[bsize];
  unsigned int esq[16] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };

  const int f_index = bsize - BLOCK_16X16;
  if (f_index < 0) {
    const int w_shift = bw == 8 ? 1 : 2;
    const int h_shift = bh == 8 ? 1 : 2;
#if CONFIG_HIGHBITDEPTH
    if (cpi->common.use_highbitdepth) {
      const uint16_t *src16 = CONVERT_TO_SHORTPTR(src);
      const uint16_t *dst16 = CONVERT_TO_SHORTPTR(dst);
      for (int i = 0; i < bh; ++i)
        for (int j = 0; j < bw; ++j) {
          const int index = (j >> w_shift) + ((i >> h_shift) << 2);
          esq[index] +=
              (src16[j + i * src_stride] - dst16[j + i * dst_stride]) *
              (src16[j + i * src_stride] - dst16[j + i * dst_stride]);
        }
    } else {
#endif  // CONFIG_HIGHBITDEPTH

      for (int i = 0; i < bh; ++i)
        for (int j = 0; j < bw; ++j) {
          const int index = (j >> w_shift) + ((i >> h_shift) << 2);
          esq[index] += (src[j + i * src_stride] - dst[j + i * dst_stride]) *
                        (src[j + i * src_stride] - dst[j + i * dst_stride]);
        }
#if CONFIG_HIGHBITDEPTH
    }
#endif  // CONFIG_HIGHBITDEPTH
  } else {
    cpi->fn_ptr[f_index].vf(src, src_stride, dst, dst_stride, &esq[0]);
    cpi->fn_ptr[f_index].vf(src + bw / 4, src_stride, dst + bw / 4, dst_stride,
                            &esq[1]);
    cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, dst + bw / 2, dst_stride,
                            &esq[2]);
    cpi->fn_ptr[f_index].vf(src + 3 * bw / 4, src_stride, dst + 3 * bw / 4,
                            dst_stride, &esq[3]);
    src += bh / 4 * src_stride;
    dst += bh / 4 * dst_stride;

    cpi->fn_ptr[f_index].vf(src, src_stride, dst, dst_stride, &esq[4]);
    cpi->fn_ptr[f_index].vf(src + bw / 4, src_stride, dst + bw / 4, dst_stride,
                            &esq[5]);
    cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, dst + bw / 2, dst_stride,
                            &esq[6]);
    cpi->fn_ptr[f_index].vf(src + 3 * bw / 4, src_stride, dst + 3 * bw / 4,
                            dst_stride, &esq[7]);
    src += bh / 4 * src_stride;
    dst += bh / 4 * dst_stride;

    cpi->fn_ptr[f_index].vf(src, src_stride, dst, dst_stride, &esq[8]);
    cpi->fn_ptr[f_index].vf(src + bw / 4, src_stride, dst + bw / 4, dst_stride,
                            &esq[9]);
    cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, dst + bw / 2, dst_stride,
                            &esq[10]);
    cpi->fn_ptr[f_index].vf(src + 3 * bw / 4, src_stride, dst + 3 * bw / 4,
                            dst_stride, &esq[11]);
    src += bh / 4 * src_stride;
    dst += bh / 4 * dst_stride;

    cpi->fn_ptr[f_index].vf(src, src_stride, dst, dst_stride, &esq[12]);
    cpi->fn_ptr[f_index].vf(src + bw / 4, src_stride, dst + bw / 4, dst_stride,
                            &esq[13]);
    cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, dst + bw / 2, dst_stride,
                            &esq[14]);
    cpi->fn_ptr[f_index].vf(src + 3 * bw / 4, src_stride, dst + 3 * bw / 4,
                            dst_stride, &esq[15]);
  }

  double total = (double)esq[0] + esq[1] + esq[2] + esq[3] + esq[4] + esq[5] +
                 esq[6] + esq[7] + esq[8] + esq[9] + esq[10] + esq[11] +
                 esq[12] + esq[13] + esq[14] + esq[15];
  if (total > 0) {
    const double e_recip = 1.0 / total;
    hordist[0] = ((double)esq[0] + esq[4] + esq[8] + esq[12]) * e_recip;
    hordist[1] = ((double)esq[1] + esq[5] + esq[9] + esq[13]) * e_recip;
    hordist[2] = ((double)esq[2] + esq[6] + esq[10] + esq[14]) * e_recip;
    verdist[0] = ((double)esq[0] + esq[1] + esq[2] + esq[3]) * e_recip;
    verdist[1] = ((double)esq[4] + esq[5] + esq[6] + esq[7]) * e_recip;
    verdist[2] = ((double)esq[8] + esq[9] + esq[10] + esq[11]) * e_recip;
  } else {
    hordist[0] = verdist[0] = 0.25;
    hordist[1] = verdist[1] = 0.25;
    hordist[2] = verdist[2] = 0.25;
  }
}

static int adst_vs_flipadst(const AV1_COMP *cpi, BLOCK_SIZE bsize,
                            const uint8_t *src, int src_stride,
                            const uint8_t *dst, int dst_stride) {
  int prune_bitmask = 0;
  double svm_proj_h = 0, svm_proj_v = 0;
  double hdist[3] = { 0, 0, 0 }, vdist[3] = { 0, 0, 0 };
  get_energy_distribution_fine(cpi, bsize, src, src_stride, dst, dst_stride,
                               hdist, vdist);

  svm_proj_v = vdist[0] * ADST_FLIP_SVM[0] + vdist[1] * ADST_FLIP_SVM[1] +
               vdist[2] * ADST_FLIP_SVM[2] + ADST_FLIP_SVM[3];
  svm_proj_h = hdist[0] * ADST_FLIP_SVM[4] + hdist[1] * ADST_FLIP_SVM[5] +
               hdist[2] * ADST_FLIP_SVM[6] + ADST_FLIP_SVM[7];
  if (svm_proj_v > FAST_EXT_TX_EDST_MID + FAST_EXT_TX_EDST_MARGIN)
    prune_bitmask |= 1 << FLIPADST_1D;
  else if (svm_proj_v < FAST_EXT_TX_EDST_MID - FAST_EXT_TX_EDST_MARGIN)
    prune_bitmask |= 1 << ADST_1D;

  if (svm_proj_h > FAST_EXT_TX_EDST_MID + FAST_EXT_TX_EDST_MARGIN)
    prune_bitmask |= 1 << (FLIPADST_1D + 8);
  else if (svm_proj_h < FAST_EXT_TX_EDST_MID - FAST_EXT_TX_EDST_MARGIN)
    prune_bitmask |= 1 << (ADST_1D + 8);

  return prune_bitmask;
}

#if CONFIG_EXT_TX
static void get_horver_correlation(const int16_t *diff, int stride, int w,
                                   int h, double *hcorr, double *vcorr) {
  // Returns hor/ver correlation coefficient
  const int num = (h - 1) * (w - 1);
  double num_r;
  int i, j;
  int64_t xy_sum = 0, xz_sum = 0;
  int64_t x_sum = 0, y_sum = 0, z_sum = 0;
  int64_t x2_sum = 0, y2_sum = 0, z2_sum = 0;
  double x_var_n, y_var_n, z_var_n, xy_var_n, xz_var_n;
  *hcorr = *vcorr = 1;

  assert(num > 0);
  num_r = 1.0 / num;
  for (i = 1; i < h; ++i) {
    for (j = 1; j < w; ++j) {
      const int16_t x = diff[i * stride + j];
      const int16_t y = diff[i * stride + j - 1];
      const int16_t z = diff[(i - 1) * stride + j];
      xy_sum += x * y;
      xz_sum += x * z;
      x_sum += x;
      y_sum += y;
      z_sum += z;
      x2_sum += x * x;
      y2_sum += y * y;
      z2_sum += z * z;
    }
  }
  x_var_n = x2_sum - (x_sum * x_sum) * num_r;
  y_var_n = y2_sum - (y_sum * y_sum) * num_r;
  z_var_n = z2_sum - (z_sum * z_sum) * num_r;
  xy_var_n = xy_sum - (x_sum * y_sum) * num_r;
  xz_var_n = xz_sum - (x_sum * z_sum) * num_r;
  if (x_var_n > 0 && y_var_n > 0) {
    *hcorr = xy_var_n / sqrt(x_var_n * y_var_n);
    *hcorr = *hcorr < 0 ? 0 : *hcorr;
  }
  if (x_var_n > 0 && z_var_n > 0) {
    *vcorr = xz_var_n / sqrt(x_var_n * z_var_n);
    *vcorr = *vcorr < 0 ? 0 : *vcorr;
  }
}

int dct_vs_idtx(const int16_t *diff, int stride, int w, int h) {
  double hcorr, vcorr;
  int prune_bitmask = 0;
  get_horver_correlation(diff, stride, w, h, &hcorr, &vcorr);

  if (vcorr > FAST_EXT_TX_CORR_MID + FAST_EXT_TX_CORR_MARGIN)
    prune_bitmask |= 1 << IDTX_1D;
  else if (vcorr < FAST_EXT_TX_CORR_MID - FAST_EXT_TX_CORR_MARGIN)
    prune_bitmask |= 1 << DCT_1D;

  if (hcorr > FAST_EXT_TX_CORR_MID + FAST_EXT_TX_CORR_MARGIN)
    prune_bitmask |= 1 << (IDTX_1D + 8);
  else if (hcorr < FAST_EXT_TX_CORR_MID - FAST_EXT_TX_CORR_MARGIN)
    prune_bitmask |= 1 << (DCT_1D + 8);
  return prune_bitmask;
}

// Performance drop: 0.5%, Speed improvement: 24%
static int prune_two_for_sby(const AV1_COMP *cpi, BLOCK_SIZE bsize,
                             MACROBLOCK *x, const MACROBLOCKD *xd,
                             int adst_flipadst, int dct_idtx) {
  int prune = 0;

  if (adst_flipadst) {
    const struct macroblock_plane *const p = &x->plane[0];
    const struct macroblockd_plane *const pd = &xd->plane[0];
    prune |= adst_vs_flipadst(cpi, bsize, p->src.buf, p->src.stride,
                              pd->dst.buf, pd->dst.stride);
  }
  if (dct_idtx) {
    av1_subtract_plane(x, bsize, 0);
    const struct macroblock_plane *const p = &x->plane[0];
    const int bw = 4 << (b_width_log2_lookup[bsize]);
    const int bh = 4 << (b_height_log2_lookup[bsize]);
    prune |= dct_vs_idtx(p->src_diff, bw, bw, bh);
  }

  return prune;
}
#endif  // CONFIG_EXT_TX

// Performance drop: 0.3%, Speed improvement: 5%
static int prune_one_for_sby(const AV1_COMP *cpi, BLOCK_SIZE bsize,
                             const MACROBLOCK *x, const MACROBLOCKD *xd) {
  const struct macroblock_plane *const p = &x->plane[0];
  const struct macroblockd_plane *const pd = &xd->plane[0];
  return adst_vs_flipadst(cpi, bsize, p->src.buf, p->src.stride, pd->dst.buf,
                          pd->dst.stride);
}

static int prune_tx_types(const AV1_COMP *cpi, BLOCK_SIZE bsize, MACROBLOCK *x,
                          const MACROBLOCKD *const xd, int tx_set) {
#if CONFIG_EXT_TX
  const int *tx_set_1D = tx_set >= 0 ? ext_tx_used_inter_1D[tx_set] : NULL;
#else
  const int tx_set_1D[TX_TYPES_1D] = { 0 };
#endif  // CONFIG_EXT_TX

  switch (cpi->sf.tx_type_search.prune_mode) {
    case NO_PRUNE: return 0; break;
    case PRUNE_ONE:
      if ((tx_set >= 0) && !(tx_set_1D[FLIPADST_1D] & tx_set_1D[ADST_1D]))
        return 0;
      return prune_one_for_sby(cpi, bsize, x, xd);
      break;
#if CONFIG_EXT_TX
    case PRUNE_TWO:
      if ((tx_set >= 0) && !(tx_set_1D[FLIPADST_1D] & tx_set_1D[ADST_1D])) {
        if (!(tx_set_1D[DCT_1D] & tx_set_1D[IDTX_1D])) return 0;
        return prune_two_for_sby(cpi, bsize, x, xd, 0, 1);
      }
      if ((tx_set >= 0) && !(tx_set_1D[DCT_1D] & tx_set_1D[IDTX_1D]))
        return prune_two_for_sby(cpi, bsize, x, xd, 1, 0);
      return prune_two_for_sby(cpi, bsize, x, xd, 1, 1);
      break;
#endif  // CONFIG_EXT_TX
  }
  assert(0);
  return 0;
}

static int do_tx_type_search(TX_TYPE tx_type, int prune) {
// TODO(sarahparker) implement for non ext tx
#if CONFIG_EXT_TX
  return !(((prune >> vtx_tab[tx_type]) & 1) |
           ((prune >> (htx_tab[tx_type] + 8)) & 1));
#else
  // temporary to avoid compiler warnings
  (void)vtx_tab;
  (void)htx_tab;
  (void)tx_type;
  (void)prune;
  return 1;
#endif  // CONFIG_EXT_TX
}

static void model_rd_from_sse(const AV1_COMP *const cpi,
                              const MACROBLOCKD *const xd, BLOCK_SIZE bsize,
                              int plane, int64_t sse, int *rate,
                              int64_t *dist) {
  const struct macroblockd_plane *const pd = &xd->plane[plane];
  const int dequant_shift =
#if CONFIG_HIGHBITDEPTH
      (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd - 5 :
#endif  // CONFIG_HIGHBITDEPTH
                                                    3;

  // Fast approximate the modelling function.
  if (cpi->sf.simple_model_rd_from_var) {
    const int64_t square_error = sse;
    int quantizer = (pd->dequant[1] >> dequant_shift);

    if (quantizer < 120)
      *rate = (int)((square_error * (280 - quantizer)) >>
                    (16 - AV1_PROB_COST_SHIFT));
    else
      *rate = 0;
    *dist = (square_error * quantizer) >> 8;
  } else {
    av1_model_rd_from_var_lapndz(sse, num_pels_log2_lookup[bsize],
                                 pd->dequant[1] >> dequant_shift, rate, dist);
  }

  *dist <<= 4;
}

static void model_rd_for_sb(const AV1_COMP *const cpi, BLOCK_SIZE bsize,
                            MACROBLOCK *x, MACROBLOCKD *xd, int plane_from,
                            int plane_to, int *out_rate_sum,
                            int64_t *out_dist_sum, int *skip_txfm_sb,
                            int64_t *skip_sse_sb) {
  // Note our transform coeffs are 8 times an orthogonal transform.
  // Hence quantizer step is also 8 times. To get effective quantizer
  // we need to divide by 8 before sending to modeling function.
  int plane;
  const int ref = xd->mi[0]->mbmi.ref_frame[0];

  int64_t rate_sum = 0;
  int64_t dist_sum = 0;
  int64_t total_sse = 0;

  x->pred_sse[ref] = 0;

  for (plane = plane_from; plane <= plane_to; ++plane) {
    struct macroblock_plane *const p = &x->plane[plane];
    struct macroblockd_plane *const pd = &xd->plane[plane];
#if CONFIG_CB4X4 && !CONFIG_CHROMA_2X2
    const BLOCK_SIZE bs = AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
#else
    const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
#endif  // CONFIG_CB4X4 && !CONFIG_CHROMA_2X2

    unsigned int sse;
    int rate;
    int64_t dist;

#if CONFIG_CB4X4
    if (x->skip_chroma_rd && plane) continue;
#endif  // CONFIG_CB4X4

    // TODO(geza): Write direct sse functions that do not compute
    // variance as well.
    cpi->fn_ptr[bs].vf(p->src.buf, p->src.stride, pd->dst.buf, pd->dst.stride,
                       &sse);

    if (plane == 0) x->pred_sse[ref] = sse;

    total_sse += sse;

    model_rd_from_sse(cpi, xd, bs, plane, sse, &rate, &dist);

    rate_sum += rate;
    dist_sum += dist;
  }

  *skip_txfm_sb = total_sse == 0;
  *skip_sse_sb = total_sse << 4;
  *out_rate_sum = (int)rate_sum;
  *out_dist_sum = dist_sum;
}

int64_t av1_block_error_c(const tran_low_t *coeff, const tran_low_t *dqcoeff,
                          intptr_t block_size, int64_t *ssz) {
  int i;
  int64_t error = 0, sqcoeff = 0;

  for (i = 0; i < block_size; i++) {
    const int diff = coeff[i] - dqcoeff[i];
    error += diff * diff;
    sqcoeff += coeff[i] * coeff[i];
  }

  *ssz = sqcoeff;
  return error;
}

int64_t av1_block_error_fp_c(const int16_t *coeff, const int16_t *dqcoeff,
                             int block_size) {
  int i;
  int64_t error = 0;

  for (i = 0; i < block_size; i++) {
    const int diff = coeff[i] - dqcoeff[i];
    error += diff * diff;
  }

  return error;
}

#if CONFIG_HIGHBITDEPTH
int64_t av1_highbd_block_error_c(const tran_low_t *coeff,
                                 const tran_low_t *dqcoeff, intptr_t block_size,
                                 int64_t *ssz, int bd) {
  int i;
  int64_t error = 0, sqcoeff = 0;
  int shift = 2 * (bd - 8);
  int rounding = shift > 0 ? 1 << (shift - 1) : 0;

  for (i = 0; i < block_size; i++) {
    const int64_t diff = coeff[i] - dqcoeff[i];
    error += diff * diff;
    sqcoeff += (int64_t)coeff[i] * (int64_t)coeff[i];
  }
  assert(error >= 0 && sqcoeff >= 0);
  error = (error + rounding) >> shift;
  sqcoeff = (sqcoeff + rounding) >> shift;

  *ssz = sqcoeff;
  return error;
}
#endif  // CONFIG_HIGHBITDEPTH

#if CONFIG_PVQ
// Without PVQ, av1_block_error_c() return two kind of errors,
// 1) reconstruction (i.e. decoded) error and
// 2) Squared sum of transformed residue (i.e. 'coeff')
// However, if PVQ is enabled, coeff does not keep the transformed residue
// but instead a transformed original is kept.
// Hence, new parameter ref vector (i.e. transformed predicted signal)
// is required to derive the residue signal,
// i.e. coeff - ref = residue (all transformed).

#if CONFIG_HIGHBITDEPTH
static int64_t av1_highbd_block_error2_c(const tran_low_t *coeff,
                                         const tran_low_t *dqcoeff,
                                         const tran_low_t *ref,
                                         intptr_t block_size, int64_t *ssz,
                                         int bd) {
  int64_t error;
  int64_t sqcoeff;
  int shift = 2 * (bd - 8);
  int rounding = shift > 0 ? 1 << (shift - 1) : 0;
  // Use the existing sse codes for calculating distortion of decoded signal:
  // i.e. (orig - decoded)^2
  // For high bit depth, throw away ssz until a 32-bit version of
  // av1_block_error_fp is written.
  int64_t ssz_trash;
  error = av1_block_error(coeff, dqcoeff, block_size, &ssz_trash);
  // prediction residue^2 = (orig - ref)^2
  sqcoeff = av1_block_error(coeff, ref, block_size, &ssz_trash);
  error = (error + rounding) >> shift;
  sqcoeff = (sqcoeff + rounding) >> shift;
  *ssz = sqcoeff;
  return error;
}
#else
// TODO(yushin) : Since 4x4 case does not need ssz, better to refactor into
// a separate function that does not do the extra computations for ssz.
static int64_t av1_block_error2_c(const tran_low_t *coeff,
                                  const tran_low_t *dqcoeff,
                                  const tran_low_t *ref, intptr_t block_size,
                                  int64_t *ssz) {
  int64_t error;
  // Use the existing sse codes for calculating distortion of decoded signal:
  // i.e. (orig - decoded)^2
  error = av1_block_error_fp(coeff, dqcoeff, block_size);
  // prediction residue^2 = (orig - ref)^2
  *ssz = av1_block_error_fp(coeff, ref, block_size);
  return error;
}
#endif  // CONFIG_HIGHBITDEPTH
#endif  // CONFIG_PVQ

#if !CONFIG_PVQ || CONFIG_VAR_TX
/* The trailing '0' is a terminator which is used inside av1_cost_coeffs() to
 * decide whether to include cost of a trailing EOB node or not (i.e. we
 * can skip this if the last coefficient in this transform block, e.g. the
 * 16th coefficient in a 4x4 block or the 64th coefficient in a 8x8 block,
 * were non-zero). */
#if !CONFIG_LV_MAP
static int cost_coeffs(const AV1_COMMON *const cm, MACROBLOCK *x, int plane,
                       int block, TX_SIZE tx_size, const SCAN_ORDER *scan_order,
                       const ENTROPY_CONTEXT *a, const ENTROPY_CONTEXT *l,
                       int use_fast_coef_costing) {
  MACROBLOCKD *const xd = &x->e_mbd;
  MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
  const struct macroblock_plane *p = &x->plane[plane];
  const struct macroblockd_plane *pd = &xd->plane[plane];
  const PLANE_TYPE type = pd->plane_type;
  const uint16_t *band_count = &band_count_table[tx_size][1];
  const int eob = p->eobs[block];
  const tran_low_t *const qcoeff = BLOCK_OFFSET(p->qcoeff, block);
  const int tx_size_ctx = txsize_sqr_map[tx_size];
  unsigned int(*token_costs)[2][COEFF_CONTEXTS][ENTROPY_TOKENS] =
      x->token_costs[tx_size_ctx][type][is_inter_block(mbmi)];
  uint8_t token_cache[MAX_TX_SQUARE];
  int pt = combine_entropy_contexts(*a, *l);
  int c, cost;
  const int16_t *scan = scan_order->scan;
  const int16_t *nb = scan_order->neighbors;
  const int ref = is_inter_block(mbmi);
  aom_prob *blockz_probs =
      cm->fc->blockzero_probs[txsize_sqr_map[tx_size]][type][ref];

#if CONFIG_HIGHBITDEPTH
  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, xd->bd);
#else
  const int cat6_bits = av1_get_cat6_extrabits_size(tx_size, 8);
#endif  // CONFIG_HIGHBITDEPTH

#if !CONFIG_VAR_TX && !CONFIG_SUPERTX
  // Check for consistency of tx_size with mode info
  assert(tx_size == get_tx_size(plane, xd));
#endif  // !CONFIG_VAR_TX && !CONFIG_SUPERTX
  (void)cm;

  if (eob == 0) {
    // single eob token
    cost = av1_cost_bit(blockz_probs[pt], 0);
  } else {
    if (use_fast_coef_costing) {
      int band_left = *band_count++;

      // dc token
      int v = qcoeff[0];
      int16_t prev_t;
      cost = av1_get_token_cost(v, &prev_t, cat6_bits);
      cost += (*token_costs)[!prev_t][pt][prev_t];

      token_cache[0] = av1_pt_energy_class[prev_t];
      ++token_costs;

      // ac tokens
      for (c = 1; c < eob; c++) {
        const int rc = scan[c];
        int16_t t;

        v = qcoeff[rc];
        cost += av1_get_token_cost(v, &t, cat6_bits);
        cost += (*token_costs)[!t][!prev_t][t];
        prev_t = t;
        if (!--band_left) {
          band_left = *band_count++;
          ++token_costs;
        }
      }

      // eob token
      cost += (*token_costs)[0][!prev_t][EOB_TOKEN];

    } else {  // !use_fast_coef_costing
      int band_left = *band_count++;

      // dc token
      int v = qcoeff[0];
      int16_t tok;
      cost = av1_get_token_cost(v, &tok, cat6_bits);
      cost += (*token_costs)[!tok][pt][tok];

      token_cache[0] = av1_pt_energy_class[tok];
      ++token_costs;

      // ac tokens
      for (c = 1; c < eob; c++) {
        const int rc = scan[c];

        v = qcoeff[rc];
        cost += av1_get_token_cost(v, &tok, cat6_bits);
        pt = get_coef_context(nb, token_cache, c);
        cost += (*token_costs)[!tok][pt][tok];
        token_cache[rc] = av1_pt_energy_class[tok];
        if (!--band_left) {
          band_left = *band_count++;
          ++token_costs;
        }
      }

      // eob token
      pt = get_coef_context(nb, token_cache, c);
      cost += (*token_costs)[0][pt][EOB_TOKEN];
    }
  }

  return cost;
}
#endif  // !CONFIG_LV_MAP

int av1_cost_coeffs(const AV1_COMP *const cpi, MACROBLOCK *x, int plane,
                    int block, TX_SIZE tx_size, const SCAN_ORDER *scan_order,
                    const ENTROPY_CONTEXT *a, const ENTROPY_CONTEXT *l,
                    int use_fast_coef_costing) {
#if !CONFIG_LV_MAP
  const AV1_COMMON *const cm = &cpi->common;
  return cost_coeffs(cm, x, plane, block, tx_size, scan_order, a, l,
                     use_fast_coef_costing);
#else  // !CONFIG_LV_MAP
  (void)scan_order;
  (void)use_fast_coef_costing;
  const MACROBLOCKD *xd = &x->e_mbd;
  const MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
  const struct macroblockd_plane *pd = &xd->plane[plane];
  const BLOCK_SIZE bsize = mbmi->sb_type;
#if CONFIG_CB4X4
#if CONFIG_CHROMA_2X2
  const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize, pd);
#else
  const BLOCK_SIZE plane_bsize =
      AOMMAX(BLOCK_4X4, get_plane_block_size(bsize, pd));
#endif  // CONFIG_CHROMA_2X2
#else   // CONFIG_CB4X4
  const BLOCK_SIZE plane_bsize =
      get_plane_block_size(AOMMAX(BLOCK_8X8, bsize), pd);
#endif  // CONFIG_CB4X4

  TXB_CTX txb_ctx;
  get_txb_ctx(plane_bsize, tx_size, plane, a, l, &txb_ctx);
  return av1_cost_coeffs_txb(cpi, x, plane, block, &txb_ctx);
#endif  // !CONFIG_LV_MAP
}
#endif  // !CONFIG_PVQ || CONFIG_VAR_TX

// Get transform block visible dimensions cropped to the MI units.
static void get_txb_dimensions(const MACROBLOCKD *xd, int plane,
                               BLOCK_SIZE plane_bsize, int blk_row, int blk_col,
                               BLOCK_SIZE tx_bsize, int *width, int *height,
                               int *visible_width, int *visible_height) {
#if !(CONFIG_EXT_TX && CONFIG_RECT_TX && CONFIG_RECT_TX_EXT)
  assert(tx_bsize <= plane_bsize);
#endif  // !(CONFIG_EXT_TX && CONFIG_RECT_TX && CONFIG_RECT_TX_EXT)
  int txb_height = block_size_high[tx_bsize];
  int txb_width = block_size_wide[tx_bsize];
  const int block_height = block_size_high[plane_bsize];
  const int block_width = block_size_wide[plane_bsize];
  const struct macroblockd_plane *const pd = &xd->plane[plane];
  // TODO(aconverse@google.com): Investigate using crop_width/height here rather
  // than the MI size
  const int block_rows =
      (xd->mb_to_bottom_edge >= 0)
          ? block_height
          : (xd->mb_to_bottom_edge >> (3 + pd->subsampling_y)) + block_height;
  const int block_cols =
      (xd->mb_to_right_edge >= 0)
          ? block_width
          : (xd->mb_to_right_edge >> (3 + pd->subsampling_x)) + block_width;
  const int tx_unit_size = tx_size_wide_log2[0];
  if (width) *width = txb_width;
  if (height) *height = txb_height;
  *visible_width = clamp(block_cols - (blk_col << tx_unit_size), 0, txb_width);
  *visible_height =
      clamp(block_rows - (blk_row << tx_unit_size), 0, txb_height);
}

// Compute the pixel domain sum square error on all visible 4x4s in the
// transform block.
static unsigned pixel_sse(const AV1_COMP *const cpi, const MACROBLOCKD *xd,
                          int plane, const uint8_t *src, const int src_stride,
                          const uint8_t *dst, const int dst_stride, int blk_row,
                          int blk_col, const BLOCK_SIZE plane_bsize,
                          const BLOCK_SIZE tx_bsize) {
  int txb_rows, txb_cols, visible_rows, visible_cols;
  get_txb_dimensions(xd, plane, plane_bsize, blk_row, blk_col, tx_bsize,
                     &txb_cols, &txb_rows, &visible_cols, &visible_rows);
  assert(visible_rows > 0);
  assert(visible_cols > 0);
#if CONFIG_EXT_TX && CONFIG_RECT_TX && CONFIG_RECT_TX_EXT
  if ((txb_rows == visible_rows && txb_cols == visible_cols) &&
      tx_bsize < BLOCK_SIZES) {
#else
  if (txb_rows == visible_rows && txb_cols == visible_cols) {
#endif
    unsigned sse;
    cpi->fn_ptr[tx_bsize].vf(src, src_stride, dst, dst_stride, &sse);
    return sse;
  }
#if CONFIG_HIGHBITDEPTH
  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    uint64_t sse = aom_highbd_sse_odd_size(src, src_stride, dst, dst_stride,
                                           visible_cols, visible_rows);
    return (unsigned int)ROUND_POWER_OF_TWO(sse, (xd->bd - 8) * 2);
  }
#endif  // CONFIG_HIGHBITDEPTH
  unsigned sse = aom_sse_odd_size(src, src_stride, dst, dst_stride,
                                  visible_cols, visible_rows);
  return sse;
}

// Compute the squares sum squares on all visible 4x4s in the transform block.
static int64_t sum_squares_visible(const MACROBLOCKD *xd, int plane,
                                   const int16_t *diff, const int diff_stride,
                                   int blk_row, int blk_col,
                                   const BLOCK_SIZE plane_bsize,
                                   const BLOCK_SIZE tx_bsize) {
  int visible_rows, visible_cols;
  get_txb_dimensions(xd, plane, plane_bsize, blk_row, blk_col, tx_bsize, NULL,
                     NULL, &visible_cols, &visible_rows);
  return aom_sum_squares_2d_i16(diff, diff_stride, visible_cols, visible_rows);
}

void av1_dist_block(const AV1_COMP *cpi, MACROBLOCK *x, int plane,
                    BLOCK_SIZE plane_bsize, int block, int blk_row, int blk_col,
                    TX_SIZE tx_size, int64_t *out_dist, int64_t *out_sse,
                    OUTPUT_STATUS output_status) {
  MACROBLOCKD *const xd = &x->e_mbd;
  const struct macroblock_plane *const p = &x->plane[plane];
#if CONFIG_DAALA_DIST
  int qm = OD_HVS_QM;
  int use_activity_masking = 0;
#if CONFIG_PVQ
  use_activity_masking = x->daala_enc.use_activity_masking;
#endif  // CONFIG_PVQ
  struct macroblockd_plane *const pd = &xd->plane[plane];
#else   // CONFIG_DAALA_DIST
  const struct macroblockd_plane *const pd = &xd->plane[plane];
#endif  // CONFIG_DAALA_DIST

  if (cpi->sf.use_transform_domain_distortion && !CONFIG_DAALA_DIST) {
    // Transform domain distortion computation is more efficient as it does
    // not involve an inverse transform, but it is less accurate.
    const int buffer_length = tx_size_2d[tx_size];
    int64_t this_sse;
    int shift = (MAX_TX_SCALE - av1_get_tx_scale(tx_size)) * 2;
    tran_low_t *const coeff = BLOCK_OFFSET(p->coeff, block);
    tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
#if CONFIG_PVQ
    tran_low_t *ref_coeff = BLOCK_OFFSET(pd->pvq_ref_coeff, block);

#if CONFIG_HIGHBITDEPTH
    const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;