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 "./av1_rtcd.h"
#include "./aom_dsp_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"

#include "av1/encoder/cost.h"
#include "av1/encoder/encodemb.h"
#include "av1/encoder/encodemv.h"
#include "av1/encoder/encoder.h"
#include "av1/encoder/hybrid_fwd_txfm.h"
#include "av1/encoder/mcomp.h"
#include "av1/encoder/palette.h"
#include "av1/encoder/quantize.h"
#include "av1/encoder/ratectrl.h"
#include "av1/encoder/rd.h"
#include "av1/encoder/rdopt.h"
#include "av1/encoder/aq_variance.h"

#if CONFIG_DUAL_FILTER
#if CONFIG_EXT_INTERP
static const int filter_sets[25][2] = {
  { 0, 0 }, { 0, 1 }, { 0, 2 }, { 0, 3 }, { 0, 4 }, { 1, 0 }, { 1, 1 },
  { 1, 2 }, { 1, 3 }, { 1, 4 }, { 2, 0 }, { 2, 1 }, { 2, 2 }, { 2, 3 },
  { 2, 4 }, { 3, 0 }, { 3, 1 }, { 3, 2 }, { 3, 3 }, { 3, 4 }, { 4, 0 },
  { 4, 1 }, { 4, 2 }, { 4, 3 }, { 4, 4 },
};
#else
static const int filter_sets[9][2] = {
  { 0, 0 }, { 0, 1 }, { 0, 2 }, { 1, 0 }, { 1, 1 },
  { 1, 2 }, { 2, 0 }, { 2, 1 }, { 2, 2 },
};
#endif
#endif

#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_FAST_SEARCH 1
#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];
  int this_rate;
  int64_t this_dist;
  int64_t this_sse;
  int64_t this_rd;
  int64_t best_rd;
  int exit_early;
  int use_fast_coef_costing;
  const scan_order *so;
  uint8_t skippable;
};

#define LAST_NEW_MV_INDEX 6
static const MODE_DEFINITION av1_mode_order[MAX_MODES] = {
  { NEARESTMV, { LAST_FRAME, NONE } },
#if CONFIG_EXT_REFS
  { NEARESTMV, { LAST2_FRAME, NONE } },
  { NEARESTMV, { LAST3_FRAME, NONE } },
  { NEARESTMV, { BWDREF_FRAME, NONE } },
#endif  // CONFIG_EXT_REFS
  { NEARESTMV, { ALTREF_FRAME, NONE } },
  { NEARESTMV, { GOLDEN_FRAME, NONE } },

  { DC_PRED, { INTRA_FRAME, NONE } },

  { NEWMV, { LAST_FRAME, NONE } },
#if CONFIG_EXT_REFS
  { NEWMV, { LAST2_FRAME, NONE } },
  { NEWMV, { LAST3_FRAME, NONE } },
  { NEWMV, { BWDREF_FRAME, NONE } },
#endif  // CONFIG_EXT_REFS
  { NEWMV, { ALTREF_FRAME, NONE } },
  { NEWMV, { GOLDEN_FRAME, NONE } },

  { NEARMV, { LAST_FRAME, NONE } },
#if CONFIG_EXT_REFS
  { NEARMV, { LAST2_FRAME, NONE } },
  { NEARMV, { LAST3_FRAME, NONE } },
  { NEARMV, { BWDREF_FRAME, NONE } },
#endif  // CONFIG_EXT_REFS
  { NEARMV, { ALTREF_FRAME, NONE } },
  { NEARMV, { GOLDEN_FRAME, NONE } },

#if CONFIG_EXT_INTER
  { NEWFROMNEARMV, { LAST_FRAME, NONE } },
#if CONFIG_EXT_REFS
  { NEWFROMNEARMV, { LAST2_FRAME, NONE } },
  { NEWFROMNEARMV, { LAST3_FRAME, NONE } },
  { NEWFROMNEARMV, { BWDREF_FRAME, NONE } },
#endif  // CONFIG_EXT_REFS
  { NEWFROMNEARMV, { ALTREF_FRAME, NONE } },
  { NEWFROMNEARMV, { GOLDEN_FRAME, NONE } },
#endif  // CONFIG_EXT_INTER

  { ZEROMV, { LAST_FRAME, NONE } },
#if CONFIG_EXT_REFS
  { ZEROMV, { LAST2_FRAME, NONE } },
  { ZEROMV, { LAST3_FRAME, NONE } },
  { ZEROMV, { BWDREF_FRAME, NONE } },
#endif  // CONFIG_EXT_REFS
  { ZEROMV, { GOLDEN_FRAME, NONE } },
  { ZEROMV, { ALTREF_FRAME, NONE } },

// 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 } },

#if CONFIG_EXT_INTER
  { NEAR_NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARMV, { LAST_FRAME, ALTREF_FRAME } },
  { 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_NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARMV, { LAST2_FRAME, ALTREF_FRAME } },
  { 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_NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARMV, { LAST3_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_NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { 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_NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARMV, { LAST_FRAME, BWDREF_FRAME } },
  { 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_NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARMV, { LAST2_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_NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARMV, { LAST3_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_NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARMV, { GOLDEN_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 } },
  { V_PRED, { INTRA_FRAME, NONE } },
  { D135_PRED, { INTRA_FRAME, NONE } },
  { D207_PRED, { INTRA_FRAME, NONE } },
  { D153_PRED, { INTRA_FRAME, NONE } },
  { D63_PRED, { INTRA_FRAME, NONE } },
  { D117_PRED, { INTRA_FRAME, NONE } },
  { D45_PRED, { INTRA_FRAME, NONE } },

#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
};

static const REF_DEFINITION av1_ref_order[MAX_REFS] = {
  { { LAST_FRAME, NONE } },
#if CONFIG_EXT_REFS
  { { LAST2_FRAME, NONE } },          { { LAST3_FRAME, NONE } },
  { { BWDREF_FRAME, NONE } },
#endif  // CONFIG_EXT_REFS
  { { GOLDEN_FRAME, NONE } },         { { ALTREF_FRAME, NONE } },

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

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

  { { INTRA_FRAME, NONE } },
};

static INLINE int write_uniform_cost(int n, int v) {
  int l = get_unsigned_bits(n), 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);
}

// 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

static const TX_TYPE_1D vtx_tab[TX_TYPES] = {
  DCT_1D,      ADST_1D, DCT_1D,      ADST_1D,
#if CONFIG_EXT_TX
  FLIPADST_1D, DCT_1D,  FLIPADST_1D, ADST_1D, FLIPADST_1D, IDTX_1D,
  DCT_1D,      IDTX_1D, ADST_1D,     IDTX_1D, FLIPADST_1D, IDTX_1D,
#endif  // CONFIG_EXT_TX
};

static const TX_TYPE_1D htx_tab[TX_TYPES] = {
  DCT_1D,  DCT_1D,      ADST_1D,     ADST_1D,
#if CONFIG_EXT_TX
  DCT_1D,  FLIPADST_1D, FLIPADST_1D, FLIPADST_1D, ADST_1D, IDTX_1D,
  IDTX_1D, DCT_1D,      IDTX_1D,     ADST_1D,     IDTX_1D, FLIPADST_1D,
#endif  // CONFIG_EXT_TX
};

static void get_energy_distribution_fine(const AV1_COMP *cpi, BLOCK_SIZE bsize,
                                         uint8_t *src, int src_stride,
                                         uint8_t *dst, int dst_stride,
                                         double *hordist, double *verdist) {
  int bw = 4 << (b_width_log2_lookup[bsize]);
  int bh = 4 << (b_height_log2_lookup[bsize]);
  unsigned int esq[16] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 };
  unsigned int var[16];
  double total = 0;

  const int f_index = bsize - BLOCK_16X16;
  if (f_index < 0) {
    int i, j, index;
    int w_shift = bw == 8 ? 1 : 2;
    int h_shift = bh == 8 ? 1 : 2;
#if CONFIG_AOM_HIGHBITDEPTH
    if (cpi->common.use_highbitdepth) {
      uint16_t *src16 = CONVERT_TO_SHORTPTR(src);
      uint16_t *dst16 = CONVERT_TO_SHORTPTR(dst);
      for (i = 0; i < bh; ++i)
        for (j = 0; j < bw; ++j) {
          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_AOM_HIGHBITDEPTH

      for (i = 0; i < bh; ++i)
        for (j = 0; j < bw; ++j) {
          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_AOM_HIGHBITDEPTH
    }
#endif  // CONFIG_AOM_HIGHBITDEPTH
  } else {
    var[0] = cpi->fn_ptr[f_index].vf(src, src_stride, dst, dst_stride, &esq[0]);
    var[1] = cpi->fn_ptr[f_index].vf(src + bw / 4, src_stride, dst + bw / 4,
                                     dst_stride, &esq[1]);
    var[2] = cpi->fn_ptr[f_index].vf(src + bw / 2, src_stride, dst + bw / 2,
                                     dst_stride, &esq[2]);
    var[3] = 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;

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

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

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

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

static int adst_vs_flipadst(const AV1_COMP *cpi, BLOCK_SIZE bsize, uint8_t *src,
                            int src_stride, uint8_t *dst, int dst_stride,
                            double *hdist, double *vdist) {
  int prune_bitmask = 0;
  double svm_proj_h = 0, svm_proj_v = 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(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(int16_t *diff, int stride, int w, int h, double *hcorr,
                double *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, MACROBLOCKD *xd, int adst_flipadst,
                             int dct_idtx) {
  struct macroblock_plane *const p = &x->plane[0];
  struct macroblockd_plane *const pd = &xd->plane[0];
  const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);
  const int bw = 4 << (b_width_log2_lookup[bs]);
  const int bh = 4 << (b_height_log2_lookup[bs]);
  double hdist[3] = { 0, 0, 0 }, vdist[3] = { 0, 0, 0 };
  double hcorr, vcorr;
  int prune = 0;
  av1_subtract_plane(x, bsize, 0);

  if (adst_flipadst)
    prune |= adst_vs_flipadst(cpi, bsize, p->src.buf, p->src.stride,
                              pd->dst.buf, pd->dst.stride, hdist, vdist);
  if (dct_idtx) prune |= dct_vs_idtx(p->src_diff, bw, bw, bh, &hcorr, &vcorr);

  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,
                             MACROBLOCK *x, MACROBLOCKD *xd) {
  struct macroblock_plane *const p = &x->plane[0];
  struct macroblockd_plane *const pd = &xd->plane[0];
  double hdist[3] = { 0, 0, 0 }, vdist[3] = { 0, 0, 0 };
  av1_subtract_plane(x, bsize, 0);
  return adst_vs_flipadst(cpi, bsize, p->src.buf, p->src.stride, pd->dst.buf,
                          pd->dst.stride, hdist, vdist);
}

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

  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
  }
  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
}

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_AOM_HIGHBITDEPTH
      (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd - 5 :
#endif  // CONFIG_AOM_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];
    const BLOCK_SIZE bs = get_plane_block_size(bsize, pd);

    unsigned int sse;
    int rate;
    int64_t dist;

    // 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_AOM_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_AOM_HIGHBITDEPTH

/* The trailing '0' is a terminator which is used inside 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). */
static int cost_coeffs(MACROBLOCK *x, int plane, int block, int coeff_ctx,
                       TX_SIZE tx_size, const int16_t *scan, const int16_t *nb,
                       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 = coeff_ctx;
  int c, cost;
#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

#if !CONFIG_VAR_TX && !CONFIG_SUPERTX
  // Check for consistency of tx_size with mode info
  assert(type == PLANE_TYPE_Y ? mbmi->tx_size == tx_size
                              : get_uv_tx_size(mbmi, pd) == tx_size);
#endif  // !CONFIG_VAR_TX && !CONFIG_SUPERTX

  if (eob == 0) {
    // single eob token
    cost = token_costs[0][0][pt][EOB_TOKEN];
    c = 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_high_cost);
      cost += (*token_costs)[0][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_high_cost);
        cost += (*token_costs)[!prev_t][!prev_t][t];
        prev_t = t;
        if (!--band_left) {
          band_left = *band_count++;
          ++token_costs;
        }
      }

      // eob token
      if (band_left) 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;
      unsigned int(*tok_cost_ptr)[COEFF_CONTEXTS][ENTROPY_TOKENS];
      cost = av1_get_token_cost(v, &tok, cat6_high_cost);
      cost += (*token_costs)[0][pt][tok];

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

      tok_cost_ptr = &((*token_costs)[!tok]);

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

        v = qcoeff[rc];
        cost += av1_get_token_cost(v, &tok, cat6_high_cost);
        pt = get_coef_context(nb, token_cache, c);
        cost += (*tok_cost_ptr)[pt][tok];
        token_cache[rc] = av1_pt_energy_class[tok];
        if (!--band_left) {
          band_left = *band_count++;
          ++token_costs;
        }
        tok_cost_ptr = &((*token_costs)[!tok]);
      }

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

  return cost;
}

static void dist_block(const AV1_COMP *cpi, MACROBLOCK *x, int plane, int block,
                       int blk_row, int blk_col, TX_SIZE tx_size,
                       int64_t *out_dist, int64_t *out_sse) {
  MACROBLOCKD *const xd = &x->e_mbd;
  const struct macroblock_plane *const p = &x->plane[plane];
  const struct macroblockd_plane *const pd = &xd->plane[plane];
  if (cpi->sf.use_transform_domain_distortion) {
    // Transform domain distortion computation is more accurate as it does
    // not involve an inverse transform, but it is less accurate.
    const int ss_txfrm_size = num_4x4_blocks_txsize_log2_lookup[tx_size];
    int64_t this_sse;
    int tx_type = get_tx_type(pd->plane_type, xd, block, tx_size);
    int shift = (MAX_TX_SCALE - get_tx_scale(xd, tx_type, 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_AOM_HIGHBITDEPTH
    const int bd = (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) ? xd->bd : 8;
    *out_dist = av1_highbd_block_error(coeff, dqcoeff, 16 << ss_txfrm_size,
                                       &this_sse, bd) >>
                shift;
#else
    *out_dist =
        av1_block_error(coeff, dqcoeff, 16 << ss_txfrm_size, &this_sse) >>
        shift;
#endif  // CONFIG_AOM_HIGHBITDEPTH
    *out_sse = this_sse >> shift;
  } else {
    const BLOCK_SIZE tx_bsize = txsize_to_bsize[tx_size];
    const int bsw = 4 * num_4x4_blocks_wide_lookup[tx_bsize];
    const int bsh = 4 * num_4x4_blocks_high_lookup[tx_bsize];
    const int src_stride = x->plane[plane].src.stride;
    const int dst_stride = xd->plane[plane].dst.stride;
    const int src_idx = 4 * (blk_row * src_stride + blk_col);
    const int dst_idx = 4 * (blk_row * dst_stride + blk_col);
    const uint8_t *src = &x->plane[plane].src.buf[src_idx];
    const uint8_t *dst = &xd->plane[plane].dst.buf[dst_idx];
    const tran_low_t *dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
    const uint16_t eob = p->eobs[block];

    unsigned int tmp;

    assert(cpi != NULL);

    cpi->fn_ptr[tx_bsize].vf(src, src_stride, dst, dst_stride, &tmp);
    *out_sse = (int64_t)tmp * 16;

    if (eob) {
      const MB_MODE_INFO *mbmi = &xd->mi[0]->mbmi;
#if CONFIG_AOM_HIGHBITDEPTH
      DECLARE_ALIGNED(16, uint16_t, recon16[MAX_TX_SQUARE]);
      uint8_t *recon = (uint8_t *)recon16;
#else
      DECLARE_ALIGNED(16, uint8_t, recon[MAX_TX_SQUARE]);
#endif  // CONFIG_AOM_HIGHBITDEPTH

      const PLANE_TYPE plane_type = plane == 0 ? PLANE_TYPE_Y : PLANE_TYPE_UV;

      INV_TXFM_PARAM inv_txfm_param;

      inv_txfm_param.tx_type = get_tx_type(plane_type, xd, block, tx_size);
      inv_txfm_param.tx_size = tx_size;
      inv_txfm_param.eob = eob;
      inv_txfm_param.lossless = xd->lossless[mbmi->segment_id];

#if CONFIG_AOM_HIGHBITDEPTH
      if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
        recon = CONVERT_TO_BYTEPTR(recon);
        inv_txfm_param.bd = xd->bd;
        aom_highbd_convolve_copy(dst, dst_stride, recon, MAX_TX_SIZE, NULL, 0,
                                 NULL, 0, bsw, bsh, xd->bd);
        highbd_inv_txfm_add(dqcoeff, recon, MAX_TX_SIZE, &inv_txfm_param);
      } else
#endif  // CONFIG_AOM_HIGHBITDEPTH
      {
        aom_convolve_copy(dst, dst_stride, recon, MAX_TX_SIZE, NULL, 0, NULL, 0,
                          bsw, bsh);
        inv_txfm_add(dqcoeff, recon, MAX_TX_SIZE, &inv_txfm_param);
      }

      cpi->fn_ptr[tx_bsize].vf(src, src_stride, recon, MAX_TX_SIZE, &tmp);
    }

    *out_dist = (int64_t)tmp * 16;
  }
}

static int rate_block(int plane, int block, int blk_row, int blk_col,
                      int coeff_ctx, TX_SIZE tx_size,
                      struct rdcost_block_args *args) {
  int coeff_cost =
      cost_coeffs(args->x, plane, block, coeff_ctx, tx_size, args->so->scan,
                  args->so->neighbors, args->use_fast_coef_costing);
  const struct macroblock_plane *p = &args->x->plane[plane];
  *(args->t_above + blk_col) = !(p->eobs[block] == 0);
  *(args->t_left + blk_row) = !(p->eobs[block] == 0);
  return coeff_cost;
}

static uint64_t sum_squares_2d(const int16_t *diff, int diff_stride,
                               TX_SIZE tx_size) {
  uint64_t sse;
  switch (tx_size) {
#if CONFIG_EXT_TX
    case TX_4X8:
      sse = aom_sum_squares_2d_i16(diff, diff_stride, 4) +
            aom_sum_squares_2d_i16(diff + 4 * diff_stride, diff_stride, 4);
      break;
    case TX_8X4:
      sse = aom_sum_squares_2d_i16(diff, diff_stride, 4) +
            aom_sum_squares_2d_i16(diff + 4, diff_stride, 4);