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/aom_timer.h"
#include "aom_ports/mem.h"
#include "aom_ports/system_state.h"

#if CONFIG_CFL
#include "av1/common/cfl.h"
#endif
#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/obmc.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
#include "av1/common/warped_motion.h"

#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"
#include "av1/encoder/palette.h"
#include "av1/encoder/ratectrl.h"
#include "av1/encoder/rd.h"
#include "av1/encoder/rdopt.h"
#include "av1/encoder/tokenize.h"
#include "av1/encoder/tx_prune_model_weights.h"

#if CONFIG_DUAL_FILTER
#define DUAL_FILTER_SET_SIZE (SWITCHABLE_FILTERS * SWITCHABLE_FILTERS)
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  // CONFIG_DUAL_FILTER

#define LAST_FRAME_MODE_MASK                                          \
  ((1 << INTRA_FRAME) | (1 << LAST2_FRAME) | (1 << LAST3_FRAME) |     \
   (1 << GOLDEN_FRAME) | (1 << BWDREF_FRAME) | (1 << ALTREF2_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 << ALTREF2_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 << ALTREF2_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 << ALTREF2_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 << ALTREF2_FRAME) | \
   (1 << ALTREF_FRAME))
#define ALTREF2_FRAME_MODE_MASK                                     \
  ((1 << INTRA_FRAME) | (1 << LAST_FRAME) | (1 << LAST2_FRAME) |    \
   (1 << LAST3_FRAME) | (1 << GOLDEN_FRAME) | (1 << BWDREF_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) | \
   (1 << ALTREF2_FRAME))

#if CONFIG_EXT_COMP_REFS
#define SECOND_REF_FRAME_MASK                                         \
  ((1 << ALTREF_FRAME) | (1 << ALTREF2_FRAME) | (1 << BWDREF_FRAME) | \
   (1 << GOLDEN_FRAME) | (1 << LAST2_FRAME) | 0x01)
#else  // !CONFIG_EXT_COMP_REFS
#define SECOND_REF_FRAME_MASK \
  ((1 << ALTREF_FRAME) | (1 << ALTREF2_FRAME) | (1 << BWDREF_FRAME) | 0x01)
#endif  // CONFIG_EXT_COMP_REFS

#define NEW_MV_DISCOUNT_FACTOR 8

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

static const double ADST_FLIP_SVM[8] = {
  /* vertical */
  -6.6623, -2.8062, -3.2531, 3.1671,
  /* horizontal */
  -7.7051, -3.2234, -3.6193, 3.4533
};

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 } },
  { NEARESTMV, { LAST2_FRAME, NONE_FRAME } },
  { NEARESTMV, { LAST3_FRAME, NONE_FRAME } },
  { NEARESTMV, { BWDREF_FRAME, NONE_FRAME } },
  { NEARESTMV, { ALTREF2_FRAME, NONE_FRAME } },
  { NEARESTMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEARESTMV, { GOLDEN_FRAME, NONE_FRAME } },

  { DC_PRED, { INTRA_FRAME, NONE_FRAME } },

  { NEWMV, { LAST_FRAME, NONE_FRAME } },
  { NEWMV, { LAST2_FRAME, NONE_FRAME } },
  { NEWMV, { LAST3_FRAME, NONE_FRAME } },
  { NEWMV, { BWDREF_FRAME, NONE_FRAME } },
  { NEWMV, { ALTREF2_FRAME, NONE_FRAME } },
  { NEWMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEWMV, { GOLDEN_FRAME, NONE_FRAME } },

  { NEARMV, { LAST_FRAME, NONE_FRAME } },
  { NEARMV, { LAST2_FRAME, NONE_FRAME } },
  { NEARMV, { LAST3_FRAME, NONE_FRAME } },
  { NEARMV, { BWDREF_FRAME, NONE_FRAME } },
  { NEARMV, { ALTREF2_FRAME, NONE_FRAME } },
  { NEARMV, { ALTREF_FRAME, NONE_FRAME } },
  { NEARMV, { GOLDEN_FRAME, NONE_FRAME } },

  { GLOBALMV, { LAST_FRAME, NONE_FRAME } },
  { GLOBALMV, { LAST2_FRAME, NONE_FRAME } },
  { GLOBALMV, { LAST3_FRAME, NONE_FRAME } },
  { GLOBALMV, { BWDREF_FRAME, NONE_FRAME } },
  { GLOBALMV, { ALTREF2_FRAME, NONE_FRAME } },
  { GLOBALMV, { GOLDEN_FRAME, NONE_FRAME } },
  { GLOBALMV, { ALTREF_FRAME, NONE_FRAME } },

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

  { NEAREST_NEARESTMV, { LAST_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST2_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST3_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARESTMV, { GOLDEN_FRAME, ALTREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST2_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST3_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { GOLDEN_FRAME, BWDREF_FRAME } },
  { NEAREST_NEARESTMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEARESTMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEARESTMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEARESTMV, { GOLDEN_FRAME, ALTREF2_FRAME } },

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

  { PAETH_PRED, { INTRA_FRAME, NONE_FRAME } },

  { SMOOTH_PRED, { INTRA_FRAME, NONE_FRAME } },
  { SMOOTH_V_PRED, { INTRA_FRAME, NONE_FRAME } },
  { SMOOTH_H_PRED, { INTRA_FRAME, NONE_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 } },
  { GLOBAL_GLOBALMV, { LAST_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 } },
  { GLOBAL_GLOBALMV, { 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 } },
  { GLOBAL_GLOBALMV, { LAST3_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 } },
  { GLOBAL_GLOBALMV, { GOLDEN_FRAME, ALTREF_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 } },
  { GLOBAL_GLOBALMV, { 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 } },
  { GLOBAL_GLOBALMV, { 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 } },
  { GLOBAL_GLOBALMV, { 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 } },
  { GLOBAL_GLOBALMV, { GOLDEN_FRAME, BWDREF_FRAME } },

  { NEAR_NEARMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEW_NEARESTMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEWMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEW_NEARMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEAR_NEWMV, { LAST_FRAME, ALTREF2_FRAME } },
  { NEW_NEWMV, { LAST_FRAME, ALTREF2_FRAME } },
  { GLOBAL_GLOBALMV, { LAST_FRAME, ALTREF2_FRAME } },

  { NEAR_NEARMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEW_NEARESTMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEWMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEW_NEARMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEAR_NEWMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { NEW_NEWMV, { LAST2_FRAME, ALTREF2_FRAME } },
  { GLOBAL_GLOBALMV, { LAST2_FRAME, ALTREF2_FRAME } },

  { NEAR_NEARMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEW_NEARESTMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEWMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEW_NEARMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEAR_NEWMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { NEW_NEWMV, { LAST3_FRAME, ALTREF2_FRAME } },
  { GLOBAL_GLOBALMV, { LAST3_FRAME, ALTREF2_FRAME } },

  { NEAR_NEARMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
  { NEW_NEARESTMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
  { NEAREST_NEWMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
  { NEW_NEARMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
  { NEAR_NEWMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
  { NEW_NEWMV, { GOLDEN_FRAME, ALTREF2_FRAME } },
  { GLOBAL_GLOBALMV, { GOLDEN_FRAME, ALTREF2_FRAME } },

  { 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_COMP_REFS
  { NEAR_NEARMV, { LAST_FRAME, LAST2_FRAME } },
  { NEW_NEARESTMV, { LAST_FRAME, LAST2_FRAME } },
  { NEAREST_NEWMV, { LAST_FRAME, LAST2_FRAME } },
  { NEW_NEARMV, { LAST_FRAME, LAST2_FRAME } },
  { NEAR_NEWMV, { LAST_FRAME, LAST2_FRAME } },
  { NEW_NEWMV, { LAST_FRAME, LAST2_FRAME } },
  { GLOBAL_GLOBALMV, { LAST_FRAME, LAST2_FRAME } },

  { NEAR_NEARMV, { LAST_FRAME, LAST3_FRAME } },
  { NEW_NEARESTMV, { LAST_FRAME, LAST3_FRAME } },
  { NEAREST_NEWMV, { LAST_FRAME, LAST3_FRAME } },
  { NEW_NEARMV, { LAST_FRAME, LAST3_FRAME } },
  { NEAR_NEWMV, { LAST_FRAME, LAST3_FRAME } },
  { NEW_NEWMV, { LAST_FRAME, LAST3_FRAME } },
  { GLOBAL_GLOBALMV, { LAST_FRAME, LAST3_FRAME } },

  { NEAR_NEARMV, { LAST_FRAME, GOLDEN_FRAME } },
  { NEW_NEARESTMV, { LAST_FRAME, GOLDEN_FRAME } },
  { NEAREST_NEWMV, { LAST_FRAME, GOLDEN_FRAME } },
  { NEW_NEARMV, { LAST_FRAME, GOLDEN_FRAME } },
  { NEAR_NEWMV, { LAST_FRAME, GOLDEN_FRAME } },
  { NEW_NEWMV, { LAST_FRAME, GOLDEN_FRAME } },
  { GLOBAL_GLOBALMV, { LAST_FRAME, GOLDEN_FRAME } },

  { NEAR_NEARMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { NEW_NEARESTMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { NEAREST_NEWMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { NEW_NEARMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { NEAR_NEWMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { NEW_NEWMV, { BWDREF_FRAME, ALTREF_FRAME } },
  { GLOBAL_GLOBALMV, { BWDREF_FRAME, ALTREF_FRAME } },
#endif  // CONFIG_EXT_COMP_REFS
};

static const PREDICTION_MODE intra_rd_search_mode_order[INTRA_MODES] = {
  DC_PRED,       H_PRED,        V_PRED,    SMOOTH_PRED, PAETH_PRED,
  SMOOTH_V_PRED, SMOOTH_H_PRED, D135_PRED, D207_PRED,   D153_PRED,
  D63_PRED,      D117_PRED,     D45_PRED,
};

#if CONFIG_CFL
static const UV_PREDICTION_MODE uv_rd_search_mode_order[UV_INTRA_MODES] = {
  UV_DC_PRED,     UV_CFL_PRED,   UV_H_PRED,        UV_V_PRED,
  UV_SMOOTH_PRED, UV_PAETH_PRED, UV_SMOOTH_V_PRED, UV_SMOOTH_H_PRED,
  UV_D135_PRED,   UV_D207_PRED,  UV_D153_PRED,     UV_D63_PRED,
  UV_D117_PRED,   UV_D45_PRED,
};
#else
#define uv_rd_search_mode_order intra_rd_search_mode_order
#endif  // CONFIG_CFL

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);
}

// 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

int inter_block_yrd(const AV1_COMP *cpi, MACROBLOCK *x, RD_STATS *rd_stats,
                    BLOCK_SIZE bsize, int64_t ref_best_rd, int fast);
int inter_block_uvrd(const AV1_COMP *cpi, MACROBLOCK *x, RD_STATS *rd_stats,
                     BLOCK_SIZE bsize, int64_t ref_best_rd, int fast);

static unsigned pixel_dist_visible_only(
    const AV1_COMP *const cpi, const MACROBLOCK *x, const uint8_t *src,
    const int src_stride, const uint8_t *dst, const int dst_stride,
    const BLOCK_SIZE tx_bsize, int txb_rows, int txb_cols, int visible_rows,
    int visible_cols) {
  unsigned sse;

  if (txb_rows == visible_rows && txb_cols == visible_cols) {
    cpi->fn_ptr[tx_bsize].vf(src, src_stride, dst, dst_stride, &sse);
    return sse;
  }
  const MACROBLOCKD *xd = &x->e_mbd;

  if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
    uint64_t sse64 = aom_highbd_sse_odd_size(src, src_stride, dst, dst_stride,
                                             visible_cols, visible_rows);
    return (unsigned int)ROUND_POWER_OF_TWO(sse64, (xd->bd - 8) * 2);
  }
  sse = aom_sse_odd_size(src, src_stride, dst, dst_stride, visible_cols,
                         visible_rows);
  return sse;
}

#if CONFIG_DIST_8X8
static uint64_t cdef_dist_8x8_16bit(uint16_t *dst, int dstride, uint16_t *src,
                                    int sstride, int coeff_shift) {
  uint64_t svar = 0;
  uint64_t dvar = 0;
  uint64_t sum_s = 0;
  uint64_t sum_d = 0;
  uint64_t sum_s2 = 0;
  uint64_t sum_d2 = 0;
  uint64_t sum_sd = 0;
  uint64_t dist = 0;

  int i, j;
  for (i = 0; i < 8; i++) {
    for (j = 0; j < 8; j++) {
      sum_s += src[i * sstride + j];
      sum_d += dst[i * dstride + j];
      sum_s2 += src[i * sstride + j] * src[i * sstride + j];
      sum_d2 += dst[i * dstride + j] * dst[i * dstride + j];
      sum_sd += src[i * sstride + j] * dst[i * dstride + j];
    }
  }
  /* Compute the variance -- the calculation cannot go negative. */
  svar = sum_s2 - ((sum_s * sum_s + 32) >> 6);
  dvar = sum_d2 - ((sum_d * sum_d + 32) >> 6);

  // Tuning of jm's original dering distortion metric used in CDEF tool,
  // suggested by jm
  const uint64_t a = 4;
  const uint64_t b = 2;
  const uint64_t c1 = (400 * a << 2 * coeff_shift);
  const uint64_t c2 = (b * 20000 * a * a << 4 * coeff_shift);

  dist =
      (uint64_t)floor(.5 +
                      (sum_d2 + sum_s2 - 2 * sum_sd) * .5 * (svar + dvar + c1) /
                          (sqrt(svar * (double)dvar + c2)));

  // Calibrate dist to have similar rate for the same QP with MSE only
  // distortion (as in master branch)
  dist = (uint64_t)((float)dist * 0.75);

  return dist;
}

static int od_compute_var_4x4(uint16_t *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;
    }
  }

  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 use_activity_masking, uint16_t *x,
                                  uint16_t *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;

#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 slightly 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_common(int activity_masking, uint16_t *x,
                                     uint16_t *y, int bsize_w, int bsize_h,
                                     int qindex, od_coeff *tmp,
                                     od_coeff *e_lp) {
  int i, j;
  double sum = 0;
  const int mid = OD_DIST_LP_MID;

  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(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;
}

static double od_compute_dist(uint16_t *x, uint16_t *y, int bsize_w,
                              int bsize_h, int qindex) {
  assert(bsize_w >= 8 && bsize_h >= 8);

  int activity_masking = 0;

  int i, 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]);
  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];
    }
  }
  int mid = OD_DIST_LP_MID;
  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];
    }
  }
  return od_compute_dist_common(activity_masking, x, y, bsize_w, bsize_h,
                                qindex, tmp, e_lp);
}

static double od_compute_dist_diff(uint16_t *x, int16_t *e, int bsize_w,
                                   int bsize_h, int qindex) {
  assert(bsize_w >= 8 && bsize_h >= 8);

  int activity_masking = 0;

  DECLARE_ALIGNED(16, uint16_t, y[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(16, od_coeff, tmp[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(16, od_coeff, e_lp[MAX_TX_SQUARE]);
  int i, j;
  for (i = 0; i < bsize_h; i++) {
    for (j = 0; j < bsize_w; j++) {
      y[i * bsize_w + j] = x[i * bsize_w + j] - e[i * bsize_w + j];
    }
  }
  int mid = OD_DIST_LP_MID;
  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];
    }
  }
  return od_compute_dist_common(activity_masking, x, y, bsize_w, bsize_h,
                                qindex, tmp, e_lp);
}

int64_t av1_dist_8x8(const AV1_COMP *const cpi, const MACROBLOCK *x,
                     const uint8_t *src, int src_stride, const uint8_t *dst,
                     int dst_stride, const BLOCK_SIZE tx_bsize, int bsw,
                     int bsh, int visible_w, int visible_h, int qindex) {
  int64_t d = 0;
  int i, j;
  const MACROBLOCKD *xd = &x->e_mbd;

  DECLARE_ALIGNED(16, uint16_t, orig[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(16, uint16_t, rec[MAX_TX_SQUARE]);

  assert(bsw >= 8);
  assert(bsh >= 8);
  assert((bsw & 0x07) == 0);
  assert((bsh & 0x07) == 0);

  if (x->tune_metric == AOM_TUNE_CDEF_DIST ||
      x->tune_metric == AOM_TUNE_DAALA_DIST) {
    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
      for (j = 0; j < bsh; j++)
        for (i = 0; i < bsw; i++)
          orig[j * bsw + i] = CONVERT_TO_SHORTPTR(src)[j * src_stride + i];

      if ((bsw == visible_w) && (bsh == visible_h)) {
        for (j = 0; j < bsh; j++)
          for (i = 0; i < bsw; i++)
            rec[j * bsw + i] = CONVERT_TO_SHORTPTR(dst)[j * dst_stride + i];
      } else {
        for (j = 0; j < visible_h; j++)
          for (i = 0; i < visible_w; i++)
            rec[j * bsw + i] = CONVERT_TO_SHORTPTR(dst)[j * dst_stride + i];

        if (visible_w < bsw) {
          for (j = 0; j < bsh; j++)
            for (i = visible_w; i < bsw; i++)
              rec[j * bsw + i] = CONVERT_TO_SHORTPTR(src)[j * src_stride + i];
        }

        if (visible_h < bsh) {
          for (j = visible_h; j < bsh; j++)
            for (i = 0; i < bsw; i++)
              rec[j * bsw + i] = CONVERT_TO_SHORTPTR(src)[j * src_stride + i];
        }
      }
    } else {
      for (j = 0; j < bsh; j++)
        for (i = 0; i < bsw; i++) orig[j * bsw + i] = src[j * src_stride + i];

      if ((bsw == visible_w) && (bsh == visible_h)) {
        for (j = 0; j < bsh; j++)
          for (i = 0; i < bsw; i++) rec[j * bsw + i] = dst[j * dst_stride + i];
      } else {
        for (j = 0; j < visible_h; j++)
          for (i = 0; i < visible_w; i++)
            rec[j * bsw + i] = dst[j * dst_stride + i];

        if (visible_w < bsw) {
          for (j = 0; j < bsh; j++)
            for (i = visible_w; i < bsw; i++)
              rec[j * bsw + i] = src[j * src_stride + i];
        }

        if (visible_h < bsh) {
          for (j = visible_h; j < bsh; j++)
            for (i = 0; i < bsw; i++)
              rec[j * bsw + i] = src[j * src_stride + i];
        }
      }
    }
  }

  if (x->tune_metric == AOM_TUNE_DAALA_DIST) {
    d = (int64_t)od_compute_dist(orig, rec, bsw, bsh, qindex);
  } else if (x->tune_metric == AOM_TUNE_CDEF_DIST) {
    int coeff_shift = AOMMAX(xd->bd - 8, 0);

    for (i = 0; i < bsh; i += 8) {
      for (j = 0; j < bsw; j += 8) {
        d += cdef_dist_8x8_16bit(&rec[i * bsw + j], bsw, &orig[i * bsw + j],
                                 bsw, coeff_shift);
      }
    }
    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH)
      d = ((uint64_t)d) >> 2 * coeff_shift;
  } else {
    // Otherwise, MSE by default
    d = pixel_dist_visible_only(cpi, x, src, src_stride, dst, dst_stride,
                                tx_bsize, bsh, bsw, visible_h, visible_w);
  }

  return d;
}

static int64_t dist_8x8_diff(const MACROBLOCK *x, const uint8_t *src,
                             int src_stride, const int16_t *diff,
                             int diff_stride, int bsw, int bsh, int visible_w,
                             int visible_h, int qindex) {
  int64_t d = 0;
  int i, j;
  const MACROBLOCKD *xd = &x->e_mbd;

  DECLARE_ALIGNED(16, uint16_t, orig[MAX_TX_SQUARE]);
  DECLARE_ALIGNED(16, int16_t, diff16[MAX_TX_SQUARE]);

  assert(bsw >= 8);
  assert(bsh >= 8);
  assert((bsw & 0x07) == 0);
  assert((bsh & 0x07) == 0);

  if (x->tune_metric == AOM_TUNE_CDEF_DIST ||
      x->tune_metric == AOM_TUNE_DAALA_DIST) {
    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
      for (j = 0; j < bsh; j++)
        for (i = 0; i < bsw; i++)
          orig[j * bsw + i] = CONVERT_TO_SHORTPTR(src)[j * src_stride + i];
    } else {
      for (j = 0; j < bsh; j++)
        for (i = 0; i < bsw; i++) orig[j * bsw + i] = src[j * src_stride + i];
    }

    if ((bsw == visible_w) && (bsh == visible_h)) {
      for (j = 0; j < bsh; j++)
        for (i = 0; i < bsw; i++)
          diff16[j * bsw + i] = diff[j * diff_stride + i];
    } else {
      for (j = 0; j < visible_h; j++)
        for (i = 0; i < visible_w; i++)
          diff16[j * bsw + i] = diff[j * diff_stride + i];

      if (visible_w < bsw) {
        for (j = 0; j < bsh; j++)
          for (i = visible_w; i < bsw; i++) diff16[j * bsw + i] = 0;
      }

      if (visible_h < bsh) {
        for (j = visible_h; j < bsh; j++)
          for (i = 0; i < bsw; i++) diff16[j * bsw + i] = 0;
      }
    }
  }

  if (x->tune_metric == AOM_TUNE_DAALA_DIST) {
    d = (int64_t)od_compute_dist_diff(orig, diff16, bsw, bsh, qindex);
  } else if (x->tune_metric == AOM_TUNE_CDEF_DIST) {
    int coeff_shift = AOMMAX(xd->bd - 8, 0);
    DECLARE_ALIGNED(16, uint16_t, dst16[MAX_TX_SQUARE]);

    for (i = 0; i < bsh; i++) {
      for (j = 0; j < bsw; j++) {
        dst16[i * bsw + j] = orig[i * bsw + j] - diff16[i * bsw + j];
      }
    }

    for (i = 0; i < bsh; i += 8) {
      for (j = 0; j < bsw; j += 8) {
        d += cdef_dist_8x8_16bit(&dst16[i * bsw + j], bsw, &orig[i * bsw + j],
                                 bsw, coeff_shift);
      }
    }
    // Don't scale 'd' for HBD since it will be done by caller side for diff
    // input
  } else {
    // Otherwise, MSE by default
    d = aom_sum_squares_2d_i16(diff, diff_stride, visible_w, visible_h);
  }

  return d;
}
#endif  // CONFIG_DIST_8X8

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 (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 {
      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]);
        }
    }
  } 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;
}

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;
  }
}

static 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);
  }