od_dering.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.
 */
#ifdef HAVE_CONFIG_H
#include "config.h"
#endif

// clang-format off

#include <stdlib.h>
#include <math.h>
#include "dering.h"
#include "./av1_rtcd.h"

/* Generated from gen_filter_tables.c. */
const int OD_DIRECTION_OFFSETS_TABLE[8][3] = {
  { -1 * OD_FILT_BSTRIDE + 1, -2 * OD_FILT_BSTRIDE + 2,
    -3 * OD_FILT_BSTRIDE + 3 },
  { 0 * OD_FILT_BSTRIDE + 1, -1 * OD_FILT_BSTRIDE + 2,
    -1 * OD_FILT_BSTRIDE + 3 },
  { 0 * OD_FILT_BSTRIDE + 1, 0 * OD_FILT_BSTRIDE + 2, 0 * OD_FILT_BSTRIDE + 3 },
  { 0 * OD_FILT_BSTRIDE + 1, 1 * OD_FILT_BSTRIDE + 2, 1 * OD_FILT_BSTRIDE + 3 },
  { 1 * OD_FILT_BSTRIDE + 1, 2 * OD_FILT_BSTRIDE + 2, 3 * OD_FILT_BSTRIDE + 3 },
  { 1 * OD_FILT_BSTRIDE + 0, 2 * OD_FILT_BSTRIDE + 1, 3 * OD_FILT_BSTRIDE + 1 },
  { 1 * OD_FILT_BSTRIDE + 0, 2 * OD_FILT_BSTRIDE + 0, 3 * OD_FILT_BSTRIDE + 0 },
  { 1 * OD_FILT_BSTRIDE + 0, 2 * OD_FILT_BSTRIDE - 1, 3 * OD_FILT_BSTRIDE - 1 },
};

/* Detect direction. 0 means 45-degree up-right, 2 is horizontal, and so on.
   The search minimizes the weighted variance along all the lines in a
   particular direction, i.e. the squared error between the input and a
   "predicted" block where each pixel is replaced by the average along a line
   in a particular direction. Since each direction have the same sum(x^2) term,
   that term is never computed. See Section 2, step 2, of:
   http://jmvalin.ca/notes/intra_paint.pdf */
int od_dir_find8_c(const od_dering_in *img, int stride, int32_t *var,
                   int coeff_shift) {
  int i;
  int32_t cost[8] = { 0 };
  int partial[8][15] = { { 0 } };
  int32_t best_cost = 0;
  int best_dir = 0;
  /* Instead of dividing by n between 2 and 8, we multiply by 3*5*7*8/n.
     The output is then 840 times larger, but we don't care for finding
     the max. */
  static const int div_table[] = { 0, 840, 420, 280, 210, 168, 140, 120, 105 };
  for (i = 0; i < 8; i++) {
    int j;
    for (j = 0; j < 8; j++) {
      int x;
      /* We subtract 128 here to reduce the maximum range of the squared
         partial sums. */
      x = (img[i * stride + j] >> coeff_shift) - 128;
      partial[0][i + j] += x;
      partial[1][i + j / 2] += x;
      partial[2][i] += x;
      partial[3][3 + i - j / 2] += x;
      partial[4][7 + i - j] += x;
      partial[5][3 - i / 2 + j] += x;
      partial[6][j] += x;
      partial[7][i / 2 + j] += x;
    }
  }
  for (i = 0; i < 8; i++) {
    cost[2] += partial[2][i] * partial[2][i];
    cost[6] += partial[6][i] * partial[6][i];
  }
  cost[2] *= div_table[8];
  cost[6] *= div_table[8];
  for (i = 0; i < 7; i++) {
    cost[0] += (partial[0][i] * partial[0][i] +
                partial[0][14 - i] * partial[0][14 - i]) *
               div_table[i + 1];
    cost[4] += (partial[4][i] * partial[4][i] +
                partial[4][14 - i] * partial[4][14 - i]) *
               div_table[i + 1];
  }
  cost[0] += partial[0][7] * partial[0][7] * div_table[8];
  cost[4] += partial[4][7] * partial[4][7] * div_table[8];
  for (i = 1; i < 8; i += 2) {
    int j;
    for (j = 0; j < 4 + 1; j++) {
      cost[i] += partial[i][3 + j] * partial[i][3 + j];
    }
    cost[i] *= div_table[8];
    for (j = 0; j < 4 - 1; j++) {
      cost[i] += (partial[i][j] * partial[i][j] +
                  partial[i][10 - j] * partial[i][10 - j]) *
                 div_table[2 * j + 2];
    }
  }
  for (i = 0; i < 8; i++) {
    if (cost[i] > best_cost) {
      best_cost = cost[i];
      best_dir = i;
    }
  }
  /* Difference between the optimal variance and the variance along the
     orthogonal direction. Again, the sum(x^2) terms cancel out. */
  *var = best_cost - cost[(best_dir + 4) & 7];
  /* We'd normally divide by 840, but dividing by 1024 is close enough
     for what we're going to do with this. */
  *var >>= 10;
  return best_dir;
}

#define OD_DERING_VERY_LARGE (30000)
#define OD_DERING_INBUF_SIZE \
  ((OD_BSIZE_MAX + 2 * OD_FILT_BORDER) * (OD_BSIZE_MAX + 2 * OD_FILT_BORDER))

/* Smooth in the direction detected. */
int od_filter_dering_direction_8x8_c(int16_t *y, int ystride, const int16_t *in,
                                     int threshold, int dir) {
  int i;
  int j;
  int k;
  static const int taps[3] = { 3, 2, 1 };
  int total_abs = 0;
  for (i = 0; i < 8; i++) {
    for (j = 0; j < 8; j++) {
      int16_t sum;
      int16_t xx;
      int16_t yy;
      xx = in[i * OD_FILT_BSTRIDE + j];
      sum = 0;
      for (k = 0; k < 3; k++) {
        int16_t p0;
        int16_t p1;
        p0 = in[i * OD_FILT_BSTRIDE + j + OD_DIRECTION_OFFSETS_TABLE[dir][k]] -
             xx;
        p1 = in[i * OD_FILT_BSTRIDE + j - OD_DIRECTION_OFFSETS_TABLE[dir][k]] -
             xx;
        if (abs(p0) < threshold) sum += taps[k] * p0;
        if (abs(p1) < threshold) sum += taps[k] * p1;
      }
      sum = (sum + 8) >> 4;
      total_abs += abs(sum);
      yy = xx + sum;
      y[i * ystride + j] = yy;
    }
  }
  return (total_abs + 8) >> 4;
}

/* Smooth in the direction detected. */
int od_filter_dering_direction_4x4_c(int16_t *y, int ystride, const int16_t *in,
                                     int threshold, int dir) {
  int i;
  int j;
  int k;
  static const int taps[2] = { 4, 1 };
  int total_abs = 0;
  for (i = 0; i < 4; i++) {
    for (j = 0; j < 4; j++) {
      int16_t sum;
      int16_t xx;
      int16_t yy;
      xx = in[i * OD_FILT_BSTRIDE + j];
      sum = 0;
      for (k = 0; k < 2; k++) {
        int16_t p0;
        int16_t p1;
        p0 = in[i * OD_FILT_BSTRIDE + j + OD_DIRECTION_OFFSETS_TABLE[dir][k]] -
             xx;
        p1 = in[i * OD_FILT_BSTRIDE + j - OD_DIRECTION_OFFSETS_TABLE[dir][k]] -
             xx;
        if (abs(p0) < threshold) sum += taps[k] * p0;
        if (abs(p1) < threshold) sum += taps[k] * p1;
      }
      sum = (sum + 8) >> 4;
      total_abs += abs(sum);
      yy = xx + sum;
      y[i * ystride + j] = yy;
    }
  }
  return (total_abs + 2) >> 2;
}

/* Smooth in the direction orthogonal to what was detected. */
void od_filter_dering_orthogonal_8x8_c(int16_t *y, int ystride,
                                       const int16_t *in, int threshold,
                                       int dir) {
  int i;
  int j;
  int offset;
  if (dir > 0 && dir < 4)
    offset = OD_FILT_BSTRIDE;
  else
    offset = 1;
  for (i = 0; i < 8; i++) {
    for (j = 0; j < 8; j++) {
      int16_t yy;
      int16_t sum;
      int16_t p;
      yy = in[i * OD_FILT_BSTRIDE + j];
      sum = 0;
      p = in[i * OD_FILT_BSTRIDE + j + offset] - yy;
      if (abs(p) < threshold) sum += p;
      p = in[i * OD_FILT_BSTRIDE + j - offset] - yy;
      if (abs(p) < threshold) sum += p;
      p = in[i * OD_FILT_BSTRIDE + j + 2 * offset] - yy;
      if (abs(p) < threshold) sum += p;
      p = in[i * OD_FILT_BSTRIDE + j - 2 * offset] - yy;
      if (abs(p) < threshold) sum += p;
      y[i * ystride + j] = yy + ((3 * sum + 8) >> 4);
    }
  }
}

/* Smooth in the direction orthogonal to what was detected. */
void od_filter_dering_orthogonal_4x4_c(int16_t *y, int ystride,
                                       const int16_t *in, int threshold,
                                       int dir) {
  int i;
  int j;
  int offset;
  if (dir > 0 && dir < 4)
    offset = OD_FILT_BSTRIDE;
  else
    offset = 1;
  for (i = 0; i < 4; i++) {
    for (j = 0; j < 4; j++) {
      int16_t yy;
      int16_t sum;
      int16_t p;
      yy = in[i * OD_FILT_BSTRIDE + j];
      sum = 0;
      p = in[i * OD_FILT_BSTRIDE + j + offset] - yy;
      if (abs(p) < threshold) sum += p;
      p = in[i * OD_FILT_BSTRIDE + j - offset] - yy;
      if (abs(p) < threshold) sum += p;
      y[i * ystride + j] = yy + ((5 * sum + 8) >> 4);
    }
  }
}

/* This table approximates x^0.16 with the index being log2(x). It is clamped
   to [-.5, 3]. The table is computed as:
   round(256*min(3, max(.5, 1.08*(sqrt(2)*2.^([0:17]+8)/256/256).^.16))) */
static const int16_t OD_THRESH_TABLE_Q8[18] = {
  128, 134, 150, 168, 188, 210, 234, 262, 292,
  327, 365, 408, 455, 509, 569, 635, 710, 768,
};

/* Compute deringing filter threshold for an 8x8 block based on the
   directional variance difference. A high variance difference means that we
   have a highly directional pattern (e.g. a high contrast edge), so we can
   apply more deringing. A low variance means that we either have a low
   contrast edge, or a non-directional texture, so we want to be careful not
   to blur. */
static INLINE int od_adjust_thresh(int threshold, int32_t var) {
  int v1;
  /* We use the variance of 8x8 blocks to adjust the threshold. */
  v1 = OD_MINI(32767, var >> 6);
  return (threshold * OD_THRESH_TABLE_Q8[OD_ILOG(v1)] + 128) >> 8;
}

static INLINE void copy_8x8_16bit(int16_t *dst, int dstride, int16_t *src, int sstride) {
  int i, j;
  for (i = 0; i < 8; i++)
    for (j = 0; j < 8; j++)
      dst[i * dstride + j] = src[i * sstride + j];
}

static INLINE void copy_4x4_16bit(int16_t *dst, int dstride, int16_t *src, int sstride) {
  int i, j;
  for (i = 0; i < 4; i++)
    for (j = 0; j < 4; j++)
      dst[i * dstride + j] = src[i * sstride + j];
}

/* TODO: Optimize this function for SSE. */
void copy_blocks_16bit(int16_t *dst, int dstride, int16_t *src, int sstride,
    unsigned char (*bskip)[2], int dering_count, int bsize)
{
  int bi, bx, by;
  if (bsize == 3) {
    for (bi = 0; bi < dering_count; bi++) {
      by = bskip[bi][0];
      bx = bskip[bi][1];
      copy_8x8_16bit(&dst[(by << 3) * dstride + (bx << 3)],
                     dstride,
                     &src[(by << 3) * sstride + (bx << 3)], sstride);
    }
  } else {
    for (bi = 0; bi < dering_count; bi++) {
      by = bskip[bi][0];
      bx = bskip[bi][1];
      copy_4x4_16bit(&dst[(by << 2) * dstride + (bx << 2)],
                     dstride,
                     &src[(by << 2) * sstride + (bx << 2)], sstride);
    }
  }
}

void od_dering(int16_t *y, int ystride, const od_dering_in *x, int xstride,
               int nhb, int nvb, int sbx, int sby, int nhsb, int nvsb, int xdec,
               int dir[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS], int pli,
               unsigned char (*bskip)[2], int dering_count, int threshold,
               int coeff_shift) {
  int i;
  int j;
  int bi;
  int bx;
  int by;
  int16_t inbuf[OD_DERING_INBUF_SIZE];
  int16_t *in;
  int bsize;
  int filter2_thresh[OD_DERING_NBLOCKS][OD_DERING_NBLOCKS];
  od_filter_dering_direction_func filter_dering_direction[OD_DERINGSIZES] = {
    od_filter_dering_direction_4x4, od_filter_dering_direction_8x8
  };
  od_filter_dering_orthogonal_func filter_dering_orthogonal[OD_DERINGSIZES] = {
    od_filter_dering_orthogonal_4x4, od_filter_dering_orthogonal_8x8
  };
  bsize = 3 - xdec;
  in = inbuf + OD_FILT_BORDER * OD_FILT_BSTRIDE + OD_FILT_BORDER;
  /* We avoid filtering the pixels for which some of the pixels to average
     are outside the frame. We could change the filter instead, but it would
     add special cases for any future vectorization. */
  for (i = 0; i < OD_DERING_INBUF_SIZE; i++) inbuf[i] = OD_DERING_VERY_LARGE;
  for (i = -OD_FILT_BORDER * (sby != 0);
       i < (nvb << bsize) + OD_FILT_BORDER * (sby != nvsb - 1); i++) {
    for (j = -OD_FILT_BORDER * (sbx != 0);
         j < (nhb << bsize) + OD_FILT_BORDER * (sbx != nhsb - 1); j++) {
      in[i * OD_FILT_BSTRIDE + j] = x[i * xstride + j];
    }
  }
  if (pli == 0) {
    for (bi = 0; bi < dering_count; bi++) {
      int32_t var;
      by = bskip[bi][0];
      bx = bskip[bi][1];
      dir[by][bx] = od_dir_find8(&x[8 * by * xstride + 8 * bx], xstride,
                                 &var, coeff_shift);
      /* Deringing orthogonal to the direction uses a tighter threshold
         because we want to be conservative. We've presumably already
         achieved some deringing, so the amount of change is expected
         to be low. Also, since we might be filtering across an edge, we
         want to make sure not to blur it. That being said, we might want
         to be a little bit more aggressive on pure horizontal/vertical
         since the ringing there tends to be directional, so it doesn't
         get removed by the directional filtering. */
      filter2_thresh[by][bx] = (filter_dering_direction[bsize - OD_LOG_BSIZE0])(
          &y[(by * ystride << bsize) + (bx << bsize)], ystride,
          &in[(by * OD_FILT_BSTRIDE << bsize) + (bx << bsize)],
          od_adjust_thresh(threshold, var), dir[by][bx]);
    }
  } else {
    for (bi = 0; bi < dering_count; bi++) {
      by = bskip[bi][0];
      bx = bskip[bi][1];
      filter2_thresh[by][bx] = (filter_dering_direction[bsize - OD_LOG_BSIZE0])(
          &y[(by * ystride << bsize) + (bx << bsize)], ystride,
          &in[(by * OD_FILT_BSTRIDE << bsize) + (bx << bsize)], threshold,
          dir[by][bx]);
    }
  }
  copy_blocks_16bit(in, OD_FILT_BSTRIDE, y, ystride, bskip, dering_count,
      bsize);
  for (bi = 0; bi < dering_count; bi++) {
    by = bskip[bi][0];
    bx = bskip[bi][1];
    if (filter2_thresh[by][bx] == 0) continue;
    (filter_dering_orthogonal[bsize - OD_LOG_BSIZE0])(
        &y[(by * ystride << bsize) + (bx << bsize)], ystride,
        &in[(by * OD_FILT_BSTRIDE << bsize) + (bx << bsize)], filter2_thresh[by][bx],
        dir[by][bx]);
  }
}