restoration.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 <math.h>

#include "./aom_config.h"
#include "./aom_dsp_rtcd.h"
#include "./aom_scale_rtcd.h"
#include "av1/common/onyxc_int.h"
#if CONFIG_FRAME_SUPERRES
#include "av1/common/resize.h"
#endif
#include "av1/common/restoration.h"
#include "aom_dsp/aom_dsp_common.h"
#include "aom_mem/aom_mem.h"

#include "aom_ports/mem.h"

const sgr_params_type sgr_params[SGRPROJ_PARAMS] = {
// r1, eps1, r2, eps2
#if MAX_RADIUS == 2
  { 2, 12, 1, 4 },  { 2, 15, 1, 6 },  { 2, 18, 1, 8 },  { 2, 20, 1, 9 },
  { 2, 22, 1, 10 }, { 2, 25, 1, 11 }, { 2, 35, 1, 12 }, { 2, 45, 1, 13 },
  { 2, 55, 1, 14 }, { 2, 65, 1, 15 }, { 2, 75, 1, 16 }, { 2, 30, 1, 6 },
  { 2, 50, 1, 12 }, { 2, 60, 1, 13 }, { 2, 70, 1, 14 }, { 2, 80, 1, 15 },
#else
  { 2, 12, 1, 4 },  { 2, 15, 1, 6 },  { 2, 18, 1, 8 },  { 2, 20, 1, 9 },
  { 2, 22, 1, 10 }, { 2, 25, 1, 11 }, { 2, 35, 1, 12 }, { 2, 45, 1, 13 },
  { 2, 55, 1, 14 }, { 2, 65, 1, 15 }, { 2, 75, 1, 16 }, { 3, 30, 1, 10 },
  { 3, 50, 1, 12 }, { 3, 50, 2, 25 }, { 3, 60, 2, 35 }, { 3, 70, 2, 45 },
#endif  // MAX_RADIUS == 2
};

#if CONFIG_MAX_TILE
static void tile_width_and_height(const AV1_COMMON *cm, int is_uv, int sb_w,
                                  int sb_h, int *px_w, int *px_h) {
  const int scaled_sb_w = sb_w << MAX_MIB_SIZE_LOG2;
  const int scaled_sb_h = sb_h << MAX_MIB_SIZE_LOG2;

  const int ss_x = is_uv && cm->subsampling_x;
  const int ss_y = is_uv && cm->subsampling_y;

  *px_w = (scaled_sb_w + ss_x) >> ss_x;
  *px_h = (scaled_sb_h + ss_y) >> ss_y;
#if CONFIG_FRAME_SUPERRES
  if (!av1_superres_unscaled(cm)) {
    av1_calculate_unscaled_superres_size(px_w, px_h,
                                         cm->superres_scale_denominator);
  }
#endif  // CONFIG_FRAME_SUPERRES
}
#endif  // CONFIG_MAX_TILE

// Similar to av1_get_tile_rect(), except that we extend the bottommost tile in
// each frame to a multiple of 8 luma pixels.
// This is done to help simplify the implementation of striped-loop-restoration,
// by avoiding nasty edge cases which would otherwise appear when the (cropped)
// frame height is 57 or 63 (mod 64).
static AV1PixelRect get_ext_tile_rect(const TileInfo *tile_info,
                                      const AV1_COMMON *cm, int is_uv) {
  int ss_y = is_uv && cm->subsampling_y;
  AV1PixelRect tile_rect = av1_get_tile_rect(tile_info, cm, is_uv);
  tile_rect.bottom = ALIGN_POWER_OF_TWO(tile_rect.bottom, 3 - ss_y);
  return tile_rect;
}

// Count horizontal or vertical units per tile (use a width or height for
// tile_size, respectively). We basically want to divide the tile size by the
// size of a restoration unit. Rather than rounding up unconditionally as you
// might expect, we round to nearest, which models the way a right or bottom
// restoration unit can extend to up to 150% its normal width or height. The
// max with 1 is to deal with tiles that are smaller than half of a restoration
// unit.
static int count_units_in_tile(int unit_size, int tile_size) {
  return AOMMAX((tile_size + (unit_size >> 1)) / unit_size, 1);
}

void av1_alloc_restoration_struct(AV1_COMMON *cm, RestorationInfo *rsi,
                                  int is_uv) {
#if CONFIG_MAX_TILE
  // We need to allocate enough space for restoration units to cover the
  // largest tile. Without CONFIG_MAX_TILE, this is always the tile at the
  // top-left and we can use get_ext_tile_rect(). With CONFIG_MAX_TILE, we have
  // to do the computation ourselves, iterating over the tiles and keeping
  // track of the largest width and height, then upscaling.
  int max_sb_w = 0;
  int max_sb_h = 0;
  for (int i = 0; i < cm->tile_cols; ++i) {
    const int sb_w = cm->tile_col_start_sb[i + 1] - cm->tile_col_start_sb[i];
    max_sb_w = AOMMAX(max_sb_w, sb_w);
  }
  for (int i = 0; i < cm->tile_rows; ++i) {
    const int sb_h = cm->tile_row_start_sb[i + 1] - cm->tile_row_start_sb[i];
    max_sb_h = AOMMAX(max_sb_h, sb_h);
  }

  int max_tile_w, max_tile_h;
  tile_width_and_height(cm, is_uv, max_sb_w, max_sb_h, &max_tile_w,
                        &max_tile_h);
#else
  TileInfo tile_info;
  av1_tile_init(&tile_info, cm, 0, 0);

  const AV1PixelRect tile_rect = get_ext_tile_rect(&tile_info, cm, is_uv);
  assert(tile_rect.left == 0 && tile_rect.top == 0);

  const int max_tile_w = tile_rect.right;
  const int max_tile_h = tile_rect.bottom;
#endif  // CONFIG_MAX_TILE

  // To calculate hpertile and vpertile (horizontal and vertical units per
  // tile), we basically want to divide the largest tile width or height by the
  // size of a restoration unit. Rather than rounding up unconditionally as you
  // might expect, we round to nearest, which models the way a right or bottom
  // restoration unit can extend to up to 150% its normal width or height. The
  // max with 1 is to deal with tiles that are smaller than half of a
  // restoration unit.
  const int unit_size = rsi->restoration_unit_size;
  const int hpertile = count_units_in_tile(unit_size, max_tile_w);
  const int vpertile = count_units_in_tile(unit_size, max_tile_h);

  rsi->units_per_tile = hpertile * vpertile;
  rsi->horz_units_per_tile = hpertile;
  rsi->vert_units_per_tile = vpertile;

  const int ntiles = cm->tile_rows * cm->tile_cols;
  const int nunits = ntiles * rsi->units_per_tile;

  aom_free(rsi->unit_info);
  CHECK_MEM_ERROR(cm, rsi->unit_info, (RestorationUnitInfo *)aom_malloc(
                                          sizeof(*rsi->unit_info) * nunits));
}

void av1_free_restoration_struct(RestorationInfo *rst_info) {
  aom_free(rst_info->unit_info);
  rst_info->unit_info = NULL;
}

// TODO(debargha): This table can be substantially reduced since only a few
// values are actually used.
int sgrproj_mtable[MAX_EPS][MAX_NELEM];

static void GenSgrprojVtable() {
  int e, n;
  for (e = 1; e <= MAX_EPS; ++e)
    for (n = 1; n <= MAX_NELEM; ++n) {
      const int n2e = n * n * e;
      sgrproj_mtable[e - 1][n - 1] =
          (((1 << SGRPROJ_MTABLE_BITS) + n2e / 2) / n2e);
    }
}

void av1_loop_restoration_precal() { GenSgrprojVtable(); }

static void extend_frame_lowbd(uint8_t *data, int width, int height, int stride,
                               int border_horz, int border_vert) {
  uint8_t *data_p;
  int i;
  for (i = 0; i < height; ++i) {
    data_p = data + i * stride;
    memset(data_p - border_horz, data_p[0], border_horz);
    memset(data_p + width, data_p[width - 1], border_horz);
  }
  data_p = data - border_horz;
  for (i = -border_vert; i < 0; ++i) {
    memcpy(data_p + i * stride, data_p, width + 2 * border_horz);
  }
  for (i = height; i < height + border_vert; ++i) {
    memcpy(data_p + i * stride, data_p + (height - 1) * stride,
           width + 2 * border_horz);
  }
}

#if CONFIG_HIGHBITDEPTH
static void extend_frame_highbd(uint16_t *data, int width, int height,
                                int stride, int border_horz, int border_vert) {
  uint16_t *data_p;
  int i, j;
  for (i = 0; i < height; ++i) {
    data_p = data + i * stride;
    for (j = -border_horz; j < 0; ++j) data_p[j] = data_p[0];
    for (j = width; j < width + border_horz; ++j) data_p[j] = data_p[width - 1];
  }
  data_p = data - border_horz;
  for (i = -border_vert; i < 0; ++i) {
    memcpy(data_p + i * stride, data_p,
           (width + 2 * border_horz) * sizeof(uint16_t));
  }
  for (i = height; i < height + border_vert; ++i) {
    memcpy(data_p + i * stride, data_p + (height - 1) * stride,
           (width + 2 * border_horz) * sizeof(uint16_t));
  }
}
#endif

void extend_frame(uint8_t *data, int width, int height, int stride,
                  int border_horz, int border_vert, int highbd) {
#if !CONFIG_HIGHBITDEPTH
  assert(highbd == 0);
  (void)highbd;
#else
  if (highbd)
    extend_frame_highbd(CONVERT_TO_SHORTPTR(data), width, height, stride,
                        border_horz, border_vert);
  else
#endif
  extend_frame_lowbd(data, width, height, stride, border_horz, border_vert);
}

static void copy_tile_lowbd(int width, int height, const uint8_t *src,
                            int src_stride, uint8_t *dst, int dst_stride) {
  for (int i = 0; i < height; ++i)
    memcpy(dst + i * dst_stride, src + i * src_stride, width);
}

#if CONFIG_HIGHBITDEPTH
static void copy_tile_highbd(int width, int height, const uint16_t *src,
                             int src_stride, uint16_t *dst, int dst_stride) {
  for (int i = 0; i < height; ++i)
    memcpy(dst + i * dst_stride, src + i * src_stride, width * sizeof(*dst));
}
#endif

static void copy_tile(int width, int height, const uint8_t *src, int src_stride,
                      uint8_t *dst, int dst_stride, int highbd) {
#if !CONFIG_HIGHBITDEPTH
  assert(highbd == 0);
  (void)highbd;
#else
  if (highbd)
    copy_tile_highbd(width, height, CONVERT_TO_SHORTPTR(src), src_stride,
                     CONVERT_TO_SHORTPTR(dst), dst_stride);
  else
#endif
  copy_tile_lowbd(width, height, src, src_stride, dst, dst_stride);
}

#if CONFIG_STRIPED_LOOP_RESTORATION
#define REAL_PTR(hbd, d) ((hbd) ? (uint8_t *)CONVERT_TO_SHORTPTR(d) : (d))

// Helper function: Save one column of left/right context to the appropriate
// column buffers, then extend the edge of the current tile into that column.
//
// Note: The code to deal with above/below boundaries may have filled out
// the corners of the border with data from the tiles to our left or right,
// which isn't allowed. To fix that up, we need to include the top and
// bottom context regions in the area which we extend.
// But note that we don't need to store the pixels we overwrite in the
// corners of the context area - those have already been overwritten once,
// so their original values are already in rlbs->tmp_save_{above,below}.
#if CONFIG_LOOPFILTERING_ACROSS_TILES
static void setup_boundary_column(const uint8_t *src8, int src_stride,
                                  uint8_t *dst8, int dst_stride, uint16_t *buf,
                                  int h, int use_highbd) {
  if (use_highbd) {
    const uint16_t *src16 = CONVERT_TO_SHORTPTR(src8);
    uint16_t *dst16 = CONVERT_TO_SHORTPTR(dst8);
    for (int i = -RESTORATION_BORDER; i < 0; i++)
      dst16[i * dst_stride] = src16[i * src_stride];
    for (int i = 0; i < h; i++) {
      buf[i] = dst16[i * dst_stride];
      dst16[i * dst_stride] = src16[i * src_stride];
    }
    for (int i = h; i < h + RESTORATION_BORDER; i++)
      dst16[i * dst_stride] = src16[i * src_stride];
  } else {
    for (int i = -RESTORATION_BORDER; i < 0; i++)
      dst8[i * dst_stride] = src8[i * src_stride];
    for (int i = 0; i < h; i++) {
      buf[i] = dst8[i * dst_stride];
      dst8[i * dst_stride] = src8[i * src_stride];
    }
    for (int i = h; i < h + RESTORATION_BORDER; i++)
      dst8[i * dst_stride] = src8[i * src_stride];
  }
}

static void restore_boundary_column(uint8_t *dst8, int dst_stride,
                                    const uint16_t *buf, int h,
                                    int use_highbd) {
  if (use_highbd) {
    uint16_t *dst16 = CONVERT_TO_SHORTPTR(dst8);
    for (int i = 0; i < h; i++) dst16[i * dst_stride] = buf[i];
  } else {
    for (int i = 0; i < h; i++) dst8[i * dst_stride] = buf[i];
  }
}
#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES

// With striped loop restoration, the filtering for each 64-pixel stripe gets
// most of its input from the output of CDEF (stored in data8), but we need to
// fill out a border of 3 pixels above/below the stripe according to the
// following
// rules:
//
// * At a frame boundary, we copy the outermost row of CDEF pixels three times.
//   This extension is done by a call to extend_frame() at the start of the loop
//   restoration process, so the value of copy_above/copy_below doesn't strictly
//   matter.
//   However, by setting *copy_above = *copy_below = 1 whenever loop filtering
//   across tiles is disabled, we can allow
//   {setup,restore}_processing_stripe_boundary to assume that the top/bottom
//   data has always been copied, simplifying the behaviour at the left and
//   right edges of tiles.
//
// * If we're at a tile boundary and loop filtering across tiles is enabled,
//   then there is a logical stripe which is 64 pixels high, but which is split
//   into an 8px high and a 56px high stripe so that the processing (and
//   coefficient set usage) can be aligned to tiles.
//   In this case, we use the 3 rows of CDEF output across the boundary for
//   context; this corresponds to leaving the frame buffer as-is.
//
// * If we're at a tile boundary and loop filtering across tiles is disabled,
//   then we take the outermost row of CDEF pixels *within the current tile*
//   and copy it three times. Thus we behave exactly as if the tile were a full
//   frame.
//
// * Otherwise, we're at a stripe boundary within a tile. In that case, we
//   take 2 rows of deblocked pixels and extend them to 3 rows of context.
//
// The distinction between the latter two cases is handled by the
// av1_loop_restoration_save_boundary_lines() function, so here we just need
// to decide if we're overwriting the above/below boundary pixels or not.
static void get_stripe_boundary_info(const RestorationTileLimits *limits,
                                     const AV1PixelRect *tile_rect, int ss_y,
#if CONFIG_LOOPFILTERING_ACROSS_TILES
                                     int loop_filter_across_tiles_enabled,
#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
                                     int *copy_above, int *copy_below) {
  *copy_above = 1;
  *copy_below = 1;

#if CONFIG_LOOPFILTERING_ACROSS_TILES
  if (loop_filter_across_tiles_enabled) {
#endif
    const int full_stripe_height = RESTORATION_PROC_UNIT_SIZE >> ss_y;
    const int rtile_offset = RESTORATION_TILE_OFFSET >> ss_y;

    const int first_stripe_in_tile = (limits->v_start == tile_rect->top);
    const int this_stripe_height =
        full_stripe_height - (first_stripe_in_tile ? rtile_offset : 0);
    const int last_stripe_in_tile =
        (limits->v_start + this_stripe_height >= tile_rect->bottom);

    if (first_stripe_in_tile) *copy_above = 0;
    if (last_stripe_in_tile) *copy_below = 0;
#if CONFIG_LOOPFILTERING_ACROSS_TILES
  }
#endif
}

// Overwrite the border pixels around a processing stripe so that the conditions
// listed above get_stripe_boundary_info() are preserved.
// We save the pixels which get overwritten into a temporary buffer, so that
// they can be restored by restore_processing_stripe_boundary() after we've
// processed the stripe.
//
// limits gives the rectangular limits of the remaining stripes for the current
// restoration unit. rsb is the stored stripe boundaries (taken from either
// deblock or CDEF output as necessary).
//
// tile_rect is the limits of the current tile and tile_stripe0 is the index of
// the first stripe in this tile (needed to convert the tile-relative stripe
// index we get from limits into something we can look up in rsb).
static void setup_processing_stripe_boundary(
    const RestorationTileLimits *limits, const RestorationStripeBoundaries *rsb,
    int rsb_row, int use_highbd, int h,
#if CONFIG_LOOPFILTERING_ACROSS_TILES
    const AV1PixelRect *tile_rect, int loop_filter_across_tiles_enabled,
#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
    uint8_t *data8, int data_stride, RestorationLineBuffers *rlbs,
    int copy_above, int copy_below) {
  assert(CONFIG_HIGHBITDEPTH || !use_highbd);

  // Offsets within the line buffers. The buffer logically starts at column
  // -RESTORATION_EXTRA_HORZ so the 1st column (at x0 - RESTORATION_EXTRA_HORZ)
  // has column x0 in the buffer.
  const int buf_stride = rsb->stripe_boundary_stride;
  const int buf_x0_off = limits->h_start;
  const int line_width =
      (limits->h_end - limits->h_start) + 2 * RESTORATION_EXTRA_HORZ;
  const int line_size = line_width << use_highbd;

  const int data_x0 = limits->h_start - RESTORATION_EXTRA_HORZ;

  // Replace RESTORATION_BORDER pixels above the top of the stripe
  // We expand RESTORATION_CTX_VERT=2 lines from rsb->stripe_boundary_above
  // to fill RESTORATION_BORDER=3 lines of above pixels. This is done by
  // duplicating the topmost of the 2 lines (see the AOMMAX call when
  // calculating src_row, which gets the values 0, 0, 1 for i = -3, -2, -1).
  //
  // Special case: If we're at the top of a tile, which isn't on the topmost
  // tile row, and we're allowed to loop filter across tiles, then we have a
  // logical 64-pixel-high stripe which has been split into an 8-pixel high
  // stripe and a 56-pixel high stripe (the current one). So, in this case,
  // we want to leave the boundary alone!
  if (copy_above) {
    uint8_t *data8_tl = data8 + data_x0 + limits->v_start * data_stride;

    for (int i = -RESTORATION_BORDER; i < 0; ++i) {
      const int buf_row = rsb_row + AOMMAX(i + RESTORATION_CTX_VERT, 0);
      const int buf_off = buf_x0_off + buf_row * buf_stride;
      const uint8_t *buf = rsb->stripe_boundary_above + (buf_off << use_highbd);
      uint8_t *dst8 = data8_tl + i * data_stride;
      // Save old pixels, then replace with data from stripe_boundary_above
      memcpy(rlbs->tmp_save_above[i + RESTORATION_BORDER],
             REAL_PTR(use_highbd, dst8), line_size);
      memcpy(REAL_PTR(use_highbd, dst8), buf, line_size);
    }
  }

  // Replace RESTORATION_BORDER pixels below the bottom of the stripe.
  // The second buffer row is repeated, so src_row gets the values 0, 1, 1
  // for i = 0, 1, 2.
  if (copy_below) {
    const int stripe_end = limits->v_start + h;
    uint8_t *data8_bl = data8 + data_x0 + stripe_end * data_stride;

    for (int i = 0; i < RESTORATION_BORDER; ++i) {
      const int buf_row = rsb_row + AOMMIN(i, RESTORATION_CTX_VERT - 1);
      const int buf_off = buf_x0_off + buf_row * buf_stride;
      const uint8_t *src = rsb->stripe_boundary_below + (buf_off << use_highbd);

      uint8_t *dst8 = data8_bl + i * data_stride;
      // Save old pixels, then replace with data from stripe_boundary_below
      memcpy(rlbs->tmp_save_below[i], REAL_PTR(use_highbd, dst8), line_size);
      memcpy(REAL_PTR(use_highbd, dst8), src, line_size);
    }
  }

#if CONFIG_LOOPFILTERING_ACROSS_TILES
  if (!loop_filter_across_tiles_enabled) {
    // If loopfiltering across tiles is disabled, we need to check if we're at
    // the edge of the current tile column. If we are, we need to extend the
    // leftmost/rightmost column within the tile by 3 pixels, so that the output
    // doesn't depend on pixels from the next column over.
    // This applies to the top and bottom borders too, since those may have
    // been filled out with data from the tile to the top-left (etc.) of us.
    const int at_tile_left_border = (limits->h_start == tile_rect->left);
    const int at_tile_right_border = (limits->h_end == tile_rect->right);

    if (at_tile_left_border) {
      uint8_t *dst8 = data8 + limits->h_start + limits->v_start * data_stride;
      for (int j = -RESTORATION_BORDER; j < 0; j++)
        setup_boundary_column(dst8, data_stride, dst8 + j, data_stride,
                              rlbs->tmp_save_left[j + RESTORATION_BORDER], h,
                              use_highbd);
    }

    if (at_tile_right_border) {
      uint8_t *dst8 = data8 + limits->h_end + limits->v_start * data_stride;
      for (int j = 0; j < RESTORATION_BORDER; j++)
        setup_boundary_column(dst8 - 1, data_stride, dst8 + j, data_stride,
                              rlbs->tmp_save_right[j], h, use_highbd);
    }
  }
#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
}

// This function restores the boundary lines modified by
// setup_processing_stripe_boundary.
static void restore_processing_stripe_boundary(
    const RestorationTileLimits *limits, const RestorationLineBuffers *rlbs,
    int use_highbd, int h,
#if CONFIG_LOOPFILTERING_ACROSS_TILES
    const AV1PixelRect *tile_rect, int loop_filter_across_tiles_enabled,
#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
    uint8_t *data8, int data_stride, int copy_above, int copy_below) {
  assert(CONFIG_HIGHBITDEPTH || !use_highbd);

  const int line_width =
      (limits->h_end - limits->h_start) + 2 * RESTORATION_EXTRA_HORZ;
  const int line_size = line_width << use_highbd;

  const int data_x0 = limits->h_start - RESTORATION_EXTRA_HORZ;

  if (copy_above) {
    uint8_t *data8_tl = data8 + data_x0 + limits->v_start * data_stride;
    for (int i = -RESTORATION_BORDER; i < 0; ++i) {
      uint8_t *dst8 = data8_tl + i * data_stride;
      memcpy(REAL_PTR(use_highbd, dst8),
             rlbs->tmp_save_above[i + RESTORATION_BORDER], line_size);
    }
  }

  if (copy_below) {
    const int stripe_bottom = limits->v_start + h;
    uint8_t *data8_bl = data8 + data_x0 + stripe_bottom * data_stride;

    for (int i = 0; i < RESTORATION_BORDER; ++i) {
      if (stripe_bottom + i >= limits->v_end + RESTORATION_BORDER) break;

      uint8_t *dst8 = data8_bl + i * data_stride;
      memcpy(REAL_PTR(use_highbd, dst8), rlbs->tmp_save_below[i], line_size);
    }
  }

#if CONFIG_LOOPFILTERING_ACROSS_TILES
  if (!loop_filter_across_tiles_enabled) {
    // Restore any pixels we overwrote at the left/right edge of this
    // processing unit
    // Note: We don't need to restore the corner pixels, even if we overwrote
    // them in the equivalent place in setup_processing_stripe_boundary:
    // Because !loop_filter_across_tiles_enabled => copy_above = copy_below = 1,
    // the corner pixels will already have been restored before we get here.
    const int at_tile_left_border = (limits->h_start == tile_rect->left);
    const int at_tile_right_border = (limits->h_end == tile_rect->right);

    if (at_tile_left_border) {
      uint8_t *dst8 = data8 + limits->h_start + limits->v_start * data_stride;
      for (int j = -RESTORATION_BORDER; j < 0; j++)
        restore_boundary_column(dst8 + j, data_stride,
                                rlbs->tmp_save_left[j + RESTORATION_BORDER], h,
                                use_highbd);
    }

    if (at_tile_right_border) {
      uint8_t *dst8 = data8 + limits->h_end + limits->v_start * data_stride;
      for (int j = 0; j < RESTORATION_BORDER; j++)
        restore_boundary_column(dst8 + j, data_stride, rlbs->tmp_save_right[j],
                                h, use_highbd);
    }
  }
#endif  // CONFIG_LOOPFILTERING_ACROSS_TILES
}
#endif

#if USE_WIENER_HIGH_INTERMEDIATE_PRECISION
#define wiener_convolve8_add_src aom_convolve8_add_src_hip
#else
#define wiener_convolve8_add_src aom_convolve8_add_src
#endif

static void wiener_filter_stripe(const RestorationUnitInfo *rui,
                                 int stripe_width, int stripe_height,
                                 int procunit_width, const uint8_t *src,
                                 int src_stride, uint8_t *dst, int dst_stride,
                                 int32_t *tmpbuf, int bit_depth) {
  (void)tmpbuf;
  (void)bit_depth;
  assert(bit_depth == 8);

  for (int j = 0; j < stripe_width; j += procunit_width) {
    int w = AOMMIN(procunit_width, (stripe_width - j + 15) & ~15);
    const uint8_t *src_p = src + j;
    uint8_t *dst_p = dst + j;
    wiener_convolve8_add_src(src_p, src_stride, dst_p, dst_stride,
                             rui->wiener_info.hfilter, 16,
                             rui->wiener_info.vfilter, 16, w, stripe_height);
  }
}

/* Calculate windowed sums (if sqr=0) or sums of squares (if sqr=1)
   over the input. The window is of size (2r + 1)x(2r + 1), and we
   specialize to r = 1, 2, 3. A default function is used for r > 3.

   Each loop follows the same format: We keep a window's worth of input
   in individual variables and select data out of that as appropriate.
*/
static void boxsum1(int32_t *src, int width, int height, int src_stride,
                    int sqr, int32_t *dst, int dst_stride) {
  int i, j, a, b, c;

  // Vertical sum over 3-pixel regions, from src into dst.
  if (!sqr) {
    for (j = 0; j < width; ++j) {
      a = src[j];
      b = src[src_stride + j];
      c = src[2 * src_stride + j];

      dst[j] = a + b;
      for (i = 1; i < height - 2; ++i) {
        // Loop invariant: At the start of each iteration,
        // a = src[(i - 1) * src_stride + j]
        // b = src[(i    ) * src_stride + j]
        // c = src[(i + 1) * src_stride + j]
        dst[i * dst_stride + j] = a + b + c;
        a = b;
        b = c;
        c = src[(i + 2) * src_stride + j];
      }
      dst[i * dst_stride + j] = a + b + c;
      dst[(i + 1) * dst_stride + j] = b + c;
    }
  } else {
    for (j = 0; j < width; ++j) {
      a = src[j] * src[j];
      b = src[src_stride + j] * src[src_stride + j];
      c = src[2 * src_stride + j] * src[2 * src_stride + j];

      dst[j] = a + b;
      for (i = 1; i < height - 2; ++i) {
        dst[i * dst_stride + j] = a + b + c;
        a = b;
        b = c;
        c = src[(i + 2) * src_stride + j] * src[(i + 2) * src_stride + j];
      }
      dst[i * dst_stride + j] = a + b + c;
      dst[(i + 1) * dst_stride + j] = b + c;
    }
  }

  // Horizontal sum over 3-pixel regions of dst
  for (i = 0; i < height; ++i) {
    a = dst[i * dst_stride];
    b = dst[i * dst_stride + 1];
    c = dst[i * dst_stride + 2];

    dst[i * dst_stride] = a + b;
    for (j = 1; j < width - 2; ++j) {
      // Loop invariant: At the start of each iteration,
      // a = src[i * src_stride + (j - 1)]
      // b = src[i * src_stride + (j    )]
      // c = src[i * src_stride + (j + 1)]
      dst[i * dst_stride + j] = a + b + c;
      a = b;
      b = c;
      c = dst[i * dst_stride + (j + 2)];
    }
    dst[i * dst_stride + j] = a + b + c;
    dst[i * dst_stride + (j + 1)] = b + c;
  }
}

static void boxsum2(int32_t *src, int width, int height, int src_stride,
                    int sqr, int32_t *dst, int dst_stride) {
  int i, j, a, b, c, d, e;

  // Vertical sum over 5-pixel regions, from src into dst.
  if (!sqr) {
    for (j = 0; j < width; ++j) {
      a = src[j];
      b = src[src_stride + j];
      c = src[2 * src_stride + j];
      d = src[3 * src_stride + j];
      e = src[4 * src_stride + j];

      dst[j] = a + b + c;
      dst[dst_stride + j] = a + b + c + d;
      for (i = 2; i < height - 3; ++i) {
        // Loop invariant: At the start of each iteration,
        // a = src[(i - 2) * src_stride + j]
        // b = src[(i - 1) * src_stride + j]
        // c = src[(i    ) * src_stride + j]
        // d = src[(i + 1) * src_stride + j]
        // e = src[(i + 2) * src_stride + j]
        dst[i * dst_stride + j] = a + b + c + d + e;
        a = b;
        b = c;
        c = d;
        d = e;
        e = src[(i + 3) * src_stride + j];
      }
      dst[i * dst_stride + j] = a + b + c + d + e;
      dst[(i + 1) * dst_stride + j] = b + c + d + e;
      dst[(i + 2) * dst_stride + j] = c + d + e;
    }
  } else {
    for (j = 0; j < width; ++j) {
      a = src[j] * src[j];
      b = src[src_stride + j] * src[src_stride + j];
      c = src[2 * src_stride + j] * src[2 * src_stride + j];
      d = src[3 * src_stride + j] * src[3 * src_stride + j];
      e = src[4 * src_stride + j] * src[4 * src_stride + j];

      dst[j] = a + b + c;
      dst[dst_stride + j] = a + b + c + d;
      for (i = 2; i < height - 3; ++i) {
        dst[i * dst_stride + j] = a + b + c + d + e;
        a = b;
        b = c;
        c = d;
        d = e;
        e = src[(i + 3) * src_stride + j] * src[(i + 3) * src_stride + j];
      }
      dst[i * dst_stride + j] = a + b + c + d + e;
      dst[(i + 1) * dst_stride + j] = b + c + d + e;
      dst[(i + 2) * dst_stride + j] = c + d + e;
    }
  }

  // Horizontal sum over 5-pixel regions of dst
  for (i = 0; i < height; ++i) {
    a = dst[i * dst_stride];
    b = dst[i * dst_stride + 1];
    c = dst[i * dst_stride + 2];
    d = dst[i * dst_stride + 3];
    e = dst[i * dst_stride + 4];

    dst[i * dst_stride] = a + b + c;
    dst[i * dst_stride + 1] = a + b + c + d;
    for (j = 2; j < width - 3; ++j) {
      // Loop invariant: At the start of each iteration,
      // a = src[i * src_stride + (j - 2)]
      // b = src[i * src_stride + (j - 1)]
      // c = src[i * src_stride + (j    )]
      // d = src[i * src_stride + (j + 1)]
      // e = src[i * src_stride + (j + 2)]
      dst[i * dst_stride + j] = a + b + c + d + e;
      a = b;
      b = c;
      c = d;
      d = e;
      e = dst[i * dst_stride + (j + 3)];
    }
    dst[i * dst_stride + j] = a + b + c + d + e;
    dst[i * dst_stride + (j + 1)] = b + c + d + e;
    dst[i * dst_stride + (j + 2)] = c + d + e;
  }
}

#if MAX_RADIUS > 2
static void boxsum3(int32_t *src, int width, int height, int src_stride,
                    int sqr, int32_t *dst, int dst_stride) {
  int i, j, a, b, c, d, e, f, g;

  // Vertical sum over 7-pixel regions, from src into dst.
  if (!sqr) {
    for (j = 0; j < width; ++j) {
      a = src[j];
      b = src[1 * src_stride + j];
      c = src[2 * src_stride + j];
      d = src[3 * src_stride + j];
      e = src[4 * src_stride + j];
      f = src[5 * src_stride + j];
      g = src[6 * src_stride + j];

      dst[j] = a + b + c + d;
      dst[dst_stride + j] = a + b + c + d + e;
      dst[2 * dst_stride + j] = a + b + c + d + e + f;
      for (i = 3; i < height - 4; ++i) {
        dst[i * dst_stride + j] = a + b + c + d + e + f + g;
        a = b;
        b = c;
        c = d;
        d = e;
        e = f;
        f = g;
        g = src[(i + 4) * src_stride + j];
      }
      dst[i * dst_stride + j] = a + b + c + d + e + f + g;
      dst[(i + 1) * dst_stride + j] = b + c + d + e + f + g;
      dst[(i + 2) * dst_stride + j] = c + d + e + f + g;
      dst[(i + 3) * dst_stride + j] = d + e + f + g;
    }
  } else {
    for (j = 0; j < width; ++j) {
      a = src[j] * src[j];
      b = src[1 * src_stride + j] * src[1 * src_stride + j];
      c = src[2 * src_stride + j] * src[2 * src_stride + j];
      d = src[3 * src_stride + j] * src[3 * src_stride + j];
      e = src[4 * src_stride + j] * src[4 * src_stride + j];
      f = src[5 * src_stride + j] * src[5 * src_stride + j];
      g = src[6 * src_stride + j] * src[6 * src_stride + j];

      dst[j] = a + b + c + d;
      dst[dst_stride + j] = a + b + c + d + e;
      dst[2 * dst_stride + j] = a + b + c + d + e + f;
      for (i = 3; i < height - 4; ++i) {
        dst[i * dst_stride + j] = a + b + c + d + e + f + g;
        a = b;
        b = c;
        c = d;
        d = e;
        e = f;
        f = g;
        g = src[(i + 4) * src_stride + j] * src[(i + 4) * src_stride + j];
      }
      dst[i * dst_stride + j] = a + b + c + d + e + f + g;
      dst[(i + 1) * dst_stride + j] = b + c + d + e + f + g;
      dst[(i + 2) * dst_stride + j] = c + d + e + f + g;
      dst[(i + 3) * dst_stride + j] = d + e + f + g;
    }
  }

  // Horizontal sum over 7-pixel regions of dst
  for (i = 0; i < height; ++i) {
    a = dst[i * dst_stride];
    b = dst[i * dst_stride + 1];
    c = dst[i * dst_stride + 2];
    d = dst[i * dst_stride + 3];
    e = dst[i * dst_stride + 4];
    f = dst[i * dst_stride + 5];
    g = dst[i * dst_stride + 6];

    dst[i * dst_stride] = a + b + c + d;
    dst[i * dst_stride + 1] = a + b + c + d + e;
    dst[i * dst_stride + 2] = a + b + c + d + e + f;
    for (j = 3; j < width - 4; ++j) {
      dst[i * dst_stride + j] = a + b + c + d + e + f + g;
      a = b;
      b = c;
      c = d;
      d = e;
      e = f;
      f = g;
      g = dst[i * dst_stride + (j + 4)];
    }
    dst[i * dst_stride + j] = a + b + c + d + e + f + g;
    dst[i * dst_stride + (j + 1)] = b + c + d + e + f + g;
    dst[i * dst_stride + (j + 2)] = c + d + e + f + g;
    dst[i * dst_stride + (j + 3)] = d + e + f + g;
  }
}

// Generic version for any r. To be removed after experiments are done.
static void boxsumr(int32_t *src, int width, int height, int src_stride, int r,
                    int sqr, int32_t *dst, int dst_stride) {
  int32_t *tmp = aom_malloc(width * height * sizeof(*tmp));
  int tmp_stride = width;
  int i, j;
  if (sqr) {
    for (j = 0; j < width; ++j) tmp[j] = src[j] * src[j];
    for (j = 0; j < width; ++j)
      for (i = 1; i < height; ++i)
        tmp[i * tmp_stride + j] =
            tmp[(i - 1) * tmp_stride + j] +
            src[i * src_stride + j] * src[i * src_stride + j];
  } else {
    memcpy(tmp, src, sizeof(*tmp) * width);
    for (j = 0; j < width; ++j)
      for (i = 1; i < height; ++i)
        tmp[i * tmp_stride + j] =
            tmp[(i - 1) * tmp_stride + j] + src[i * src_stride + j];
  }
  for (i = 0; i <= r; ++i)
    memcpy(&dst[i * dst_stride], &tmp[(i + r) * tmp_stride],
           sizeof(*tmp) * width);
  for (i = r + 1; i < height - r; ++i)
    for (j = 0; j < width; ++j)
      dst[i * dst_stride + j] =
          tmp[(i + r) * tmp_stride + j] - tmp[(i - r - 1) * tmp_stride + j];
  for (i = height - r; i < height; ++i)
    for (j = 0; j < width; ++j)
      dst[i * dst_stride + j] = tmp[(height - 1) * tmp_stride + j] -
                                tmp[(i - r - 1) * tmp_stride + j];

  for (i = 0; i < height; ++i) tmp[i * tmp_stride] = dst[i * dst_stride];
  for (i = 0; i < height; ++i)
    for (j = 1; j < width; ++j)
      tmp[i * tmp_stride + j] =
          tmp[i * tmp_stride + j - 1] + dst[i * src_stride + j];

  for (j = 0; j <= r; ++j)
    for (i = 0; i < height; ++i)
      dst[i * dst_stride + j] = tmp[i * tmp_stride + j + r];
  for (j = r + 1; j < width - r; ++j)
    for (i = 0; i < height; ++i)
      dst[i * dst_stride + j] =
          tmp[i * tmp_stride + j + r] - tmp[i * tmp_stride + j - r - 1];
  for (j = width - r; j < width; ++j)
    for (i = 0; i < height; ++i)
      dst[i * dst_stride + j] =
          tmp[i * tmp_stride + width - 1] - tmp[i * tmp_stride + j - r - 1];
  aom_free(tmp);
}
#endif  // MAX_RADIUS > 2

static void boxsum(int32_t *src, int width, int height, int src_stride, int r,
                   int sqr, int32_t *dst, int dst_stride) {
  if (r == 1)
    boxsum1(src, width, height, src_stride, sqr, dst, dst_stride);
  else if (r == 2)
    boxsum2(src, width, height, src_stride, sqr, dst, dst_stride);
#if MAX_RADIUS > 2
  else if (r == 3)
    boxsum3(src, width, height, src_stride, sqr, dst, dst_stride);
  else if (r > 3)
    boxsumr(src, width, height, src_stride, r, sqr, dst, dst_stride);
#endif  // MAX_RADIUS > 2
  else
    assert(0 && "Invalid value of r in self-guided filter");
}

#if MAX_RADIUS > 2
static void boxnum(int width, int height, int r, int8_t *num, int num_stride) {
  int i, j;
  for (i = 0; i <= r; ++i) {
    for (j = 0; j <= r; ++j) {
      num[i * num_stride + j] = (r + 1 + i) * (r + 1 + j);
      num[i * num_stride + (width - 1 - j)] = num[i * num_stride + j];
      num[(height - 1 - i) * num_stride + j] = num[i * num_stride + j];
      num[(height - 1 - i) * num_stride + (width - 1 - j)] =
          num[i * num_stride + j];
    }
  }
  for (j = 0; j <= r; ++j) {
    const int val = (2 * r + 1) * (r + 1 + j);
    for (i = r + 1; i < height - r; ++i) {
      num[i * num_stride + j] = val;
      num[i * num_stride + (width - 1 - j)] = val;
    }
  }
  for (i = 0; i <= r; ++i) {
    const int val = (2 * r + 1) * (r + 1 + i);
    for (j = r + 1; j < width - r; ++j) {
      num[i * num_stride + j] = val;
      num[(height - 1 - i) * num_stride + j] = val;
    }
  }
  for (i = r + 1; i < height - r; ++i) {
    for (j = r + 1; j < width - r; ++j) {
      num[i * num_stride + j] = (2 * r + 1) * (2 * r + 1);
    }
  }
}
#endif  // MAX_RADIUS > 2

void decode_xq(const int *xqd, int *xq) {
  xq[0] = xqd[0];
  xq[1] = (1 << SGRPROJ_PRJ_BITS) - xq[0] - xqd[1];
}

const int32_t x_by_xplus1[256] = {
  // Special case: Map 0 -> 1 (corresponding to a value of 1/256)
  // instead of 0. See comments in av1_selfguided_restoration_internal() for why
  1,   128, 171, 192, 205, 213, 219, 224, 228, 230, 233, 235, 236, 238, 239,
  240, 241, 242, 243, 243, 244, 244, 245, 245, 246, 246, 247, 247, 247, 247,
  248, 248, 248, 248, 249, 249, 249, 249, 249, 250, 250, 250, 250, 250, 250,
  250, 251, 251, 251, 251, 251, 251, 251, 251, 251, 251, 252, 252, 252, 252,
  252, 252, 252, 252, 252, 252, 252, 252, 252, 252, 252, 252, 252, 253, 253,
  253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253,
  253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 253, 254, 254, 254,
  254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254,
  254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254,
  254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254,
  254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254, 254,
  254, 254, 254, 254, 254, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
  255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
  256,
};

const int32_t one_by_x[MAX_NELEM] = {
  4096, 2048, 1365, 1024, 819, 683, 585, 512, 455, 410, 372, 341, 315,
  293,  273,  256,  241,  228, 216, 205, 195, 186, 178, 171, 164,
#if MAX_RADIUS > 2
  158,  152,  146,  141,  137, 132, 128, 124, 120, 117, 114, 111, 108,
  105,  102,  100,  98,   95,  93,  91,  89,  87,  85,  84
#endif  // MAX_RADIUS > 2
};

static void av1_selfguided_restoration_internal(int32_t *dgd, int width,
                                                int height, int dgd_stride,
                                                int32_t *dst, int dst_stride,
                                                int bit_depth, int r, int eps) {
  const int width_ext = width + 2 * SGRPROJ_BORDER_HORZ;
  const int height_ext = height + 2 * SGRPROJ_BORDER_VERT;
  // Adjusting the stride of A and B here appears to avoid bad cache effects,
  // leading to a significant speed improvement.
  // We also align the stride to a multiple of 16 bytes, for consistency
  // with the SIMD version of this function.
  int buf_stride = ((width_ext + 3) & ~3) + 16;
  int32_t A_[RESTORATION_PROC_UNIT_PELS];
  int32_t B_[RESTORATION_PROC_UNIT_PELS];
  int32_t *A = A_;
  int32_t *B = B_;
#if MAX_RADIUS > 2
  const int num_stride = width_ext;
  int8_t num_[RESTORATION_PROC_UNIT_PELS];
  int8_t *num = num_ + SGRPROJ_BORDER_VERT * num_stride + SGRPROJ_BORDER_HORZ;
#endif
  int i, j;

  assert(r <= MAX_RADIUS && "Need MAX_RADIUS >= r");
  assert(r <= SGRPROJ_BORDER_VERT - 1 && r <= SGRPROJ_BORDER_HORZ - 1 &&
         "Need SGRPROJ_BORDER_* >= r+1");

  boxsum(dgd - dgd_stride * SGRPROJ_BORDER_VERT - SGRPROJ_BORDER_HORZ,
         width_ext, height_ext, dgd_stride, r, 0, B, buf_stride);
  boxsum(dgd - dgd_stride * SGRPROJ_BORDER_VERT - SGRPROJ_BORDER_HORZ,
         width_ext, height_ext, dgd_stride, r, 1, A, buf_stride);
#if MAX_RADIUS > 2
  boxnum(width_ext, height_ext, r, num_, num_stride);
#endif
  A += SGRPROJ_BORDER_VERT * buf_stride + SGRPROJ_BORDER_HORZ;
  B += SGRPROJ_BORDER_VERT * buf_stride + SGRPROJ_BORDER_HORZ;
  // Calculate the eventual A[] and B[] arrays. Include a 1-pixel border - ie,
  // for a 64x64 processing unit, we calculate 66x66 pixels of A[] and B[].
  for (i = -1; i < height + 1; ++i) {
    for (j = -1; j < width + 1; ++j) {
      const int k = i * buf_stride + j;
#if MAX_RADIUS > 2
      const int n = num[i * num_stride + j];
#else
      const int n = (2 * r + 1) * (2 * r + 1);
#endif