av1/common/restoration.c

/*
 * Copyright (c) 2021, Alliance for Open Media. All rights reserved
 *
 * This source code is subject to the terms of the BSD 3-Clause Clear License
 * and the Alliance for Open Media Patent License 1.0. If the BSD 3-Clause Clear
 * License was not distributed with this source code in the LICENSE file, you
 * can obtain it at aomedia.org/license/software-license/bsd-3-c-c/.  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
 * aomedia.org/license/patent-license/.
 *
 */

#include <math.h>

#include "config/aom_config.h"
#include "config/aom_dsp_rtcd.h"
#include "config/aom_scale_rtcd.h"

#include "aom_mem/aom_mem.h"
#include "av1/common/av1_common_int.h"
#include "av1/common/resize.h"
#include "av1/common/restoration.h"
#include "aom_dsp/aom_dsp_common.h"
#include "aom_mem/aom_mem.h"

#include "aom_ports/mem.h"

// The 's' values are calculated based on original 'r' and 'e' values in the
// spec using GenSgrprojVtable().
// Note: Setting r = 0 skips the filter; with corresponding s = -1 (invalid).
const sgr_params_type av1_sgr_params[SGRPROJ_PARAMS] = {
  { { 2, 1 }, { 140, 3236 } }, { { 2, 1 }, { 112, 2158 } },
  { { 2, 1 }, { 93, 1618 } },  { { 2, 1 }, { 80, 1438 } },
  { { 2, 1 }, { 70, 1295 } },  { { 2, 1 }, { 58, 1177 } },
  { { 2, 1 }, { 47, 1079 } },  { { 2, 1 }, { 37, 996 } },
  { { 2, 1 }, { 30, 925 } },   { { 2, 1 }, { 25, 863 } },
  { { 0, 1 }, { -1, 2589 } },  { { 0, 1 }, { -1, 1618 } },
  { { 0, 1 }, { -1, 1177 } },  { { 0, 1 }, { -1, 925 } },
  { { 2, 0 }, { 56, -1 } },    { { 2, 0 }, { 22, -1 } },
};

AV1PixelRect av1_whole_frame_rect(const AV1_COMMON *cm, int is_uv) {
  AV1PixelRect rect;

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

  rect.top = 0;
  rect.bottom = ROUND_POWER_OF_TWO(cm->height, ss_y);
  rect.left = 0;
  rect.right = ROUND_POWER_OF_TWO(cm->superres_upscaled_width, ss_x);
  return 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.
int av1_lr_count_units_in_tile(int unit_size, int tile_size) {
  return AOMMAX((tile_size + (unit_size >> 1)) / unit_size, 1);
}

// Finds a pixel rectangle for a RU, given the limits in ru domain
// (i.e. ru_start_row, ru_end_row, ru_start_col, ru_end_col)
// and the ru size (ru_height and ru_width).
// Note that offset RUs vertically by RESTORATION_UNIT_OFFSET for luma,
// and RESTORATION_UNIT_OFFSET >> ss_y for chroma, so
// that the first RU in col is shorter than the rest.
// Note the limits of the last RU in row or col is simply the size
// of the image, which makes the last RU either bigger or smaller
// than the other RUs.
AV1PixelRect av1_get_rutile_rect(const AV1_COMMON *cm, int plane,
                                 int ru_start_row, int ru_end_row,
                                 int ru_start_col, int ru_end_col,
                                 int ru_height, int ru_width) {
  AV1PixelRect rect;
  const RestorationInfo *rsi = &cm->rst_info[plane];

  int ss_x = plane && cm->seq_params.subsampling_x;
  int ss_y = plane && cm->seq_params.subsampling_y;
  const int plane_height =
      ROUND_POWER_OF_TWO(cm->superres_upscaled_height, ss_y);
  const int plane_width = ROUND_POWER_OF_TWO(cm->superres_upscaled_width, ss_x);

  const int runit_offset = RESTORATION_UNIT_OFFSET >> ss_y;
  // Top limit is a multiple of RU height minus the offset, clamped to be
  // non-negative. So the first RU vertically is shorter than the rest.
  // The bottom limit is similar except for the apecial case for the last RU.
  rect.top = AOMMAX(ru_start_row * ru_height - runit_offset, 0);
  rect.bottom = rsi->vert_units_per_tile == ru_end_row
                    ? plane_height
                    : AOMMAX(ru_end_row * ru_height - runit_offset, 0);

  // Left limit is a multiple of RU width.
  // The right limit is similar except for the apecial case for the last RU.
  rect.left = ru_start_col * ru_width;
  rect.right = rsi->horz_units_per_tile == ru_end_col ? plane_width
                                                      : ru_end_col * ru_width;

  return rect;
}

void av1_alloc_restoration_struct(AV1_COMMON *cm, RestorationInfo *rsi,
                                  int is_uv) {
  // 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 av1_get_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.
  const AV1PixelRect tile_rect = av1_whole_frame_rect(cm, is_uv);
  const int max_tile_w = tile_rect.right - tile_rect.left;
  const int max_tile_h = tile_rect.bottom - tile_rect.top;

  // 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 = av1_lr_count_units_in_tile(unit_size, max_tile_w);
  const int vpertile = av1_lr_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 = 1;
  const int nunits = ntiles * rsi->units_per_tile;

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

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

#if 0
// Pair of values for each sgrproj parameter:
// Index 0 corresponds to r[0], e[0]
// Index 1 corresponds to r[1], e[1]
int sgrproj_mtable[SGRPROJ_PARAMS][2];

static void GenSgrprojVtable() {
  for (int i = 0; i < SGRPROJ_PARAMS; ++i) {
    const sgr_params_type *const params = &av1_sgr_params[i];
    for (int j = 0; j < 2; ++j) {
      const int e = params->e[j];
      const int r = params->r[j];
      if (r == 0) {                 // filter is disabled
        sgrproj_mtable[i][j] = -1;  // mark invalid
      } else {                      // filter is enabled
        const int n = (2 * r + 1) * (2 * r + 1);
        const int n2e = n * n * e;
        assert(n2e != 0);
        sgrproj_mtable[i][j] = (((1 << SGRPROJ_MTABLE_BITS) + n2e / 2) / n2e);
      }
    }
  }
}
#endif

void av1_loop_restoration_precal() {
#if 0
  GenSgrprojVtable();
#endif
}

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));
  }
}

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));
}

void av1_extend_frame(uint16_t *data, int width, int height, int stride,
                      int border_horz, int border_vert) {
  extend_frame_highbd(data, width, height, stride, border_horz, border_vert);
}

static void copy_tile(int width, int height, const uint16_t *src,
                      int src_stride, uint16_t *dst, int dst_stride) {
  copy_tile_highbd(width, height, src, src_stride, dst, dst_stride);
}

// 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 av1_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,
                                     int *copy_above, int *copy_below) {
  *copy_above = 1;
  *copy_below = 1;

  const int full_stripe_height = RESTORATION_PROC_UNIT_SIZE >> ss_y;
  const int runit_offset = RESTORATION_UNIT_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 ? runit_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;
}

// 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 h, uint16_t *data, int data_stride,
    RestorationLineBuffers *rlbs, int copy_above, int copy_below, int opt) {
  // 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 << 1;

  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 (!opt) {
    if (copy_above) {
      uint16_t *data_tl = data + 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 uint16_t *buf = rsb->stripe_boundary_above + buf_off;
        uint16_t *dst = data_tl + i * data_stride;
        // Save old pixels, then replace with data from stripe_boundary_above
        memcpy(rlbs->tmp_save_above[i + RESTORATION_BORDER], dst, line_size);
        memcpy(dst, 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;
      uint16_t *data_bl = data + 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 uint16_t *src = rsb->stripe_boundary_below + buf_off;

        uint16_t *dst = data_bl + i * data_stride;
        // Save old pixels, then replace with data from stripe_boundary_below
        memcpy(rlbs->tmp_save_below[i], dst, line_size);
        memcpy(dst, src, line_size);
      }
    }
  } else {
    if (copy_above) {
      uint16_t *data_tl = data + data_x0 + limits->v_start * data_stride;

      // Only save and overwrite i=-RESTORATION_BORDER line.
      uint16_t *dst = data_tl + (-RESTORATION_BORDER) * data_stride;
      // Save old pixels, then replace with data from stripe_boundary_above
      memcpy(rlbs->tmp_save_above[0], dst, line_size);
      memcpy(dst, data_tl + (-RESTORATION_BORDER + 1) * data_stride, line_size);
    }

    if (copy_below) {
      const int stripe_end = limits->v_start + h;
      uint16_t *data_bl = data + data_x0 + stripe_end * data_stride;

      // Only save and overwrite i=2 line.
      uint16_t *dst = data_bl + 2 * data_stride;
      // Save old pixels, then replace with data from stripe_boundary_below
      memcpy(rlbs->tmp_save_below[2], dst, line_size);
      memcpy(dst, data_bl + (2 - 1) * data_stride, line_size);
    }
  }
}

// This function restores the boundary lines modified by
// setup_processing_stripe_boundary.
//
// Note: We need to be careful when handling the corners of the processing
// unit, because (eg.) the top-left corner is considered to be part of
// both the left and top borders. This means that, depending on the
// loop_filter_across_tiles_enabled flag, the corner pixels might get
// overwritten twice, once as part of the "top" border and once as part
// of the "left" border (or similar for other corners).
//
// Everything works out fine as long as we make sure to reverse the order
// when restoring, ie. we need to restore the left/right borders followed
// by the top/bottom borders.
static void restore_processing_stripe_boundary(
    const RestorationTileLimits *limits, const RestorationLineBuffers *rlbs,
    int h, uint16_t *data, int data_stride, int copy_above, int copy_below,
    int opt) {
  const int line_width =
      (limits->h_end - limits->h_start) + 2 * RESTORATION_EXTRA_HORZ;
  const int line_size = line_width << 1;

  const int data_x0 = limits->h_start - RESTORATION_EXTRA_HORZ;

  if (!opt) {
    if (copy_above) {
      uint16_t *data_tl = data + data_x0 + limits->v_start * data_stride;
      for (int i = -RESTORATION_BORDER; i < 0; ++i) {
        uint16_t *dst = data_tl + i * data_stride;
        memcpy(dst, rlbs->tmp_save_above[i + RESTORATION_BORDER], line_size);
      }
    }

    if (copy_below) {
      const int stripe_bottom = limits->v_start + h;
      uint16_t *data_bl = data + data_x0 + stripe_bottom * data_stride;

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

        uint16_t *dst = data_bl + i * data_stride;
        memcpy(dst, rlbs->tmp_save_below[i], line_size);
      }
    }
  } else {
    if (copy_above) {
      uint16_t *data_tl = data + data_x0 + limits->v_start * data_stride;

      // Only restore i=-RESTORATION_BORDER line.
      uint16_t *dst = data_tl + (-RESTORATION_BORDER) * data_stride;
      memcpy(dst, rlbs->tmp_save_above[0], line_size);
    }

    if (copy_below) {
      const int stripe_bottom = limits->v_start + h;
      uint16_t *data_bl = data + data_x0 + stripe_bottom * data_stride;

      // Only restore i=2 line.
      if (stripe_bottom + 2 < limits->v_end + RESTORATION_BORDER) {
        uint16_t *dst = data_bl + 2 * data_stride;
        memcpy(dst, rlbs->tmp_save_below[2], line_size);
      }
    }
  }
}

/* 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;
  assert(width > 2 * SGRPROJ_BORDER_HORZ);
  assert(height > 2 * SGRPROJ_BORDER_VERT);

  // 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;
  assert(width > 2 * SGRPROJ_BORDER_HORZ);
  assert(height > 2 * SGRPROJ_BORDER_VERT);

  // 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;
  }
}

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);
  else
    assert(0 && "Invalid value of r in self-guided filter");
}

void av1_decode_xq(const int *xqd, int *xq, const sgr_params_type *params) {
  if (params->r[0] == 0) {
    xq[0] = 0;
    xq[1] = (1 << SGRPROJ_PRJ_BITS) - xqd[1];
  } else if (params->r[1] == 0) {
    xq[0] = xqd[0];
    xq[1] = 0;
  } else {
    xq[0] = xqd[0];
    xq[1] = (1 << SGRPROJ_PRJ_BITS) - xq[0] - xqd[1];
  }
}

const int32_t av1_x_by_xplus1[256] = {
  // Special case: Map 0 -> 1 (corresponding to a value of 1/256)
  // instead of 0. See comments in 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 av1_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,
};

static void calculate_intermediate_result(int32_t *dgd, int width, int height,
                                          int dgd_stride, int bit_depth,
                                          int sgr_params_idx, int radius_idx,
                                          int pass, int32_t *A, int32_t *B) {
  const sgr_params_type *const params = &av1_sgr_params[sgr_params_idx];
  const int r = params->r[radius_idx];
  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;
  const int step = pass == 0 ? 1 : 2;
  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);
  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 += step) {
    for (j = -1; j < width + 1; ++j) {
      const int k = i * buf_stride + j;
      const int n = (2 * r + 1) * (2 * r + 1);

      // a < 2^16 * n < 2^22 regardless of bit depth
      uint32_t a = ROUND_POWER_OF_TWO(A[k], 2 * (bit_depth - 8));
      // b < 2^8 * n < 2^14 regardless of bit depth
      uint32_t b = ROUND_POWER_OF_TWO(B[k], bit_depth - 8);

      // Each term in calculating p = a * n - b * b is < 2^16 * n^2 < 2^28,
      // and p itself satisfies p < 2^14 * n^2 < 2^26.
      // This bound on p is due to:
      // https://en.wikipedia.org/wiki/Popoviciu's_inequality_on_variances
      //
      // Note: Sometimes, in high bit depth, we can end up with a*n < b*b.
      // This is an artefact of rounding, and can only happen if all pixels
      // are (almost) identical, so in this case we saturate to p=0.
      uint32_t p = (a * n < b * b) ? 0 : a * n - b * b;

      const uint32_t s = params->s[radius_idx];

      // p * s < (2^14 * n^2) * round(2^20 / n^2 eps) < 2^34 / eps < 2^32
      // as long as eps >= 4. So p * s fits into a uint32_t, and z < 2^12
      // (this holds even after accounting for the rounding in s)
      const uint32_t z = ROUND_POWER_OF_TWO(p * s, SGRPROJ_MTABLE_BITS);

      // Note: We have to be quite careful about the value of A[k].
      // This is used as a blend factor between individual pixel values and the
      // local mean. So it logically has a range of [0, 256], including both
      // endpoints.
      //
      // This is a pain for hardware, as we'd like something which can be stored
      // in exactly 8 bits.
      // Further, in the calculation of B[k] below, if z == 0 and r == 2,
      // then A[k] "should be" 0. But then we can end up setting B[k] to a value
      // slightly above 2^(8 + bit depth), due to rounding in the value of
      // av1_one_by_x[25-1].
      //
      // Thus we saturate so that, when z == 0, A[k] is set to 1 instead of 0.
      // This fixes the above issues (256 - A[k] fits in a uint8, and we can't
      // overflow), without significantly affecting the final result: z == 0
      // implies that the image is essentially "flat", so the local mean and
      // individual pixel values are very similar.
      //
      // Note that saturating on the other side, ie. requring A[k] <= 255,
      // would be a bad idea, as that corresponds to the case where the image
      // is very variable, when we want to preserve the local pixel value as
      // much as possible.
      A[k] = av1_x_by_xplus1[AOMMIN(z, 255)];  // in range [1, 256]

      // SGRPROJ_SGR - A[k] < 2^8 (from above), B[k] < 2^(bit_depth) * n,
      // av1_one_by_x[n - 1] = round(2^12 / n)
      // => the product here is < 2^(20 + bit_depth) <= 2^32,
      // and B[k] is set to a value < 2^(8 + bit depth)
      // This holds even with the rounding in av1_one_by_x and in the overall
      // result, as long as SGRPROJ_SGR - A[k] is strictly less than 2^8.
      B[k] = (int32_t)ROUND_POWER_OF_TWO((uint32_t)(SGRPROJ_SGR - A[k]) *
                                             (uint32_t)B[k] *
                                             (uint32_t)av1_one_by_x[n - 1],
                                         SGRPROJ_RECIP_BITS);
    }
  }
}

static void selfguided_restoration_fast_internal(
    int32_t *dgd, int width, int height, int dgd_stride, int32_t *dst,
    int dst_stride, int bit_depth, int sgr_params_idx, int radius_idx) {
  const sgr_params_type *const params = &av1_sgr_params[sgr_params_idx];
  const int r = params->r[radius_idx];
  const int width_ext = width + 2 * SGRPROJ_BORDER_HORZ;
  // 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_;
  int i, j;
  calculate_intermediate_result(dgd, width, height, dgd_stride, bit_depth,
                                sgr_params_idx, radius_idx, 1, A, B);
  A += SGRPROJ_BORDER_VERT * buf_stride + SGRPROJ_BORDER_HORZ;
  B += SGRPROJ_BORDER_VERT * buf_stride + SGRPROJ_BORDER_HORZ;

  // Use the A[] and B[] arrays to calculate the filtered image
  (void)r;
  assert(r == 2);
  for (i = 0; i < height; ++i) {
    if (!(i & 1)) {  // even row
      for (j = 0; j < width; ++j) {
        const int k = i * buf_stride + j;
        const int l = i * dgd_stride + j;
        const int m = i * dst_stride + j;
        const int nb = 5;
        const int32_t a = (A[k - buf_stride] + A[k + buf_stride]) * 6 +
                          (A[k - 1 - buf_stride] + A[k - 1 + buf_stride] +
                           A[k + 1 - buf_stride] + A[k + 1 + buf_stride]) *
                              5;
        const int32_t b = (B[k - buf_stride] + B[k + buf_stride]) * 6 +
                          (B[k - 1 - buf_stride] + B[k - 1 + buf_stride] +
                           B[k + 1 - buf_stride] + B[k + 1 + buf_stride]) *
                              5;
        const int32_t v = a * dgd[l] + b;
        dst[m] =
            ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
      }
    } else {  // odd row
      for (j = 0; j < width; ++j) {
        const int k = i * buf_stride + j;
        const int l = i * dgd_stride + j;
        const int m = i * dst_stride + j;
        const int nb = 4;
        const int32_t a = A[k] * 6 + (A[k - 1] + A[k + 1]) * 5;
        const int32_t b = B[k] * 6 + (B[k - 1] + B[k + 1]) * 5;
        const int32_t v = a * dgd[l] + b;
        dst[m] =
            ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
      }
    }
  }
}

static void selfguided_restoration_internal(int32_t *dgd, int width, int height,
                                            int dgd_stride, int32_t *dst,
                                            int dst_stride, int bit_depth,
                                            int sgr_params_idx,
                                            int radius_idx) {
  const int width_ext = width + 2 * SGRPROJ_BORDER_HORZ;
  // 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_;
  int i, j;
  calculate_intermediate_result(dgd, width, height, dgd_stride, bit_depth,
                                sgr_params_idx, radius_idx, 0, A, B);
  A += SGRPROJ_BORDER_VERT * buf_stride + SGRPROJ_BORDER_HORZ;
  B += SGRPROJ_BORDER_VERT * buf_stride + SGRPROJ_BORDER_HORZ;

  // Use the A[] and B[] arrays to calculate the filtered image
  for (i = 0; i < height; ++i) {
    for (j = 0; j < width; ++j) {
      const int k = i * buf_stride + j;
      const int l = i * dgd_stride + j;
      const int m = i * dst_stride + j;
      const int nb = 5;
      const int32_t a =
          (A[k] + A[k - 1] + A[k + 1] + A[k - buf_stride] + A[k + buf_stride]) *
              4 +
          (A[k - 1 - buf_stride] + A[k - 1 + buf_stride] +
           A[k + 1 - buf_stride] + A[k + 1 + buf_stride]) *
              3;
      const int32_t b =
          (B[k] + B[k - 1] + B[k + 1] + B[k - buf_stride] + B[k + buf_stride]) *
              4 +
          (B[k - 1 - buf_stride] + B[k - 1 + buf_stride] +
           B[k + 1 - buf_stride] + B[k + 1 + buf_stride]) *
              3;
      const int32_t v = a * dgd[l] + b;
      dst[m] = ROUND_POWER_OF_TWO(v, SGRPROJ_SGR_BITS + nb - SGRPROJ_RST_BITS);
    }
  }
}

int av1_selfguided_restoration_c(const uint16_t *dgd, int width, int height,
                                 int dgd_stride, int32_t *flt0, int32_t *flt1,
                                 int flt_stride, int sgr_params_idx,
                                 int bit_depth) {
  int32_t dgd32_[RESTORATION_PROC_UNIT_PELS];
  const int dgd32_stride = width + 2 * SGRPROJ_BORDER_HORZ;
  int32_t *dgd32 =
      dgd32_ + dgd32_stride * SGRPROJ_BORDER_VERT + SGRPROJ_BORDER_HORZ;

  for (int i = -SGRPROJ_BORDER_VERT; i < height + SGRPROJ_BORDER_VERT; ++i) {
    for (int j = -SGRPROJ_BORDER_HORZ; j < width + SGRPROJ_BORDER_HORZ; ++j) {
      dgd32[i * dgd32_stride + j] = dgd[i * dgd_stride + j];
    }
  }

  const sgr_params_type *const params = &av1_sgr_params[sgr_params_idx];
  // If params->r == 0 we skip the corresponding filter. We only allow one of
  // the radii to be 0, as having both equal to 0 would be equivalent to
  // skipping SGR entirely.
  assert(!(params->r[0] == 0 && params->r[1] == 0));

  if (params->r[0] > 0)
    selfguided_restoration_fast_internal(dgd32, width, height, dgd32_stride,
                                         flt0, flt_stride, bit_depth,
                                         sgr_params_idx, 0);
  if (params->r[1] > 0)
    selfguided_restoration_internal(dgd32, width, height, dgd32_stride, flt1,
                                    flt_stride, bit_depth, sgr_params_idx, 1);
  return 0;
}

void av1_apply_selfguided_restoration_c(const uint16_t *dat, int width,
                                        int height, int stride, int eps,
                                        const int *xqd, uint16_t *dst,
                                        int dst_stride, int32_t *tmpbuf,
                                        int bit_depth) {
  int32_t *flt0 = tmpbuf;
  int32_t *flt1 = flt0 + RESTORATION_UNITPELS_MAX;
  assert(width * height <= RESTORATION_UNITPELS_MAX);

  const int ret = av1_selfguided_restoration_c(dat, width, height, stride, flt0,
                                               flt1, width, eps, bit_depth);
  (void)ret;
  assert(!ret);
  const sgr_params_type *const params = &av1_sgr_params[eps];
  int xq[2];
  av1_decode_xq(xqd, xq, params);
  for (int i = 0; i < height; ++i) {
    for (int j = 0; j < width; ++j) {
      const int k = i * width + j;
      uint16_t *dstij = dst + i * dst_stride + j;
      const uint16_t *datij = dat + i * stride + j;

      const uint16_t pre_u = *datij;
      const int32_t u = (int32_t)pre_u << SGRPROJ_RST_BITS;
      int32_t v = u << SGRPROJ_PRJ_BITS;
      // If params->r == 0 then we skipped the filtering in
      // av1_selfguided_restoration_c, i.e. flt[k] == u
      if (params->r[0] > 0) v += xq[0] * (flt0[k] - u);
      if (params->r[1] > 0) v += xq[1] * (flt1[k] - u);
      const int16_t w =
          (int16_t)ROUND_POWER_OF_TWO(v, SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);

      const uint16_t out = clip_pixel_highbd(w, bit_depth);
      *dstij = out;
    }
  }
}

static void wiener_filter_stripe_highbd(const RestorationUnitInfo *rui,
                                        int stripe_width, int stripe_height,
                                        int procunit_width, const uint16_t *src,
                                        int src_stride, uint16_t *dst,
                                        int dst_stride, int32_t *tmpbuf,
                                        int bit_depth) {
  (void)tmpbuf;
  const ConvolveParams conv_params = get_conv_params_wiener(bit_depth);

  for (int j = 0; j < stripe_width; j += procunit_width) {
    int w = AOMMIN(procunit_width, (stripe_width - j + 15) & ~15);
    const uint16_t *src_p = src + j;
    uint16_t *dst_p = dst + j;
    av1_highbd_wiener_convolve_add_src(src_p, src_stride, dst_p, dst_stride,
                                       rui->wiener_info.hfilter, 16,
                                       rui->wiener_info.vfilter, 16, w,
                                       stripe_height, &conv_params, bit_depth);
  }
}

static void sgrproj_filter_stripe_highbd(const RestorationUnitInfo *rui,
                                         int stripe_width, int stripe_height,
                                         int procunit_width,
                                         const uint16_t *src, int src_stride,
                                         uint16_t *dst, int dst_stride,
                                         int32_t *tmpbuf, int bit_depth) {
  for (int j = 0; j < stripe_width; j += procunit_width) {
    int w = AOMMIN(procunit_width, stripe_width - j);
    av1_apply_selfguided_restoration(
        src + j, w, stripe_height, src_stride, rui->sgrproj_info.ep,
        rui->sgrproj_info.xqd, dst + j, dst_stride, tmpbuf, bit_depth);
  }
}

typedef void (*stripe_filter_fun)(const RestorationUnitInfo *rui,
                                  int stripe_width, int stripe_height,
                                  int procunit_width, const uint16_t *src,
                                  int src_stride, uint16_t *dst, int dst_stride,
                                  int32_t *tmpbuf, int bit_depth);

#define NUM_STRIPE_FILTERS 2
static const stripe_filter_fun stripe_filters[NUM_STRIPE_FILTERS] = {
  wiener_filter_stripe_highbd, sgrproj_filter_stripe_highbd
};

// Filter one restoration unit
void av1_loop_restoration_filter_unit(
    const RestorationTileLimits *limits, const RestorationUnitInfo *rui,
    const RestorationStripeBoundaries *rsb, RestorationLineBuffers *rlbs,
    const AV1PixelRect *tile_rect, int tile_stripe0, int ss_x, int ss_y,
    int bit_depth, uint16_t *data, int stride, uint16_t *dst, int dst_stride,
    int32_t *tmpbuf, int optimized_lr) {
  RestorationType unit_rtype = rui->restoration_type;

  int unit_h = limits->v_end - limits->v_start;
  int unit_w = limits->h_end - limits->h_start;
  uint16_t *data_tl = data + limits->v_start * stride + limits->h_start;
  uint16_t *dst_tl = dst + limits->v_start * dst_stride + limits->h_start;

  if (unit_rtype == RESTORE_NONE) {
    copy_tile(unit_w, unit_h, data_tl, stride, dst_tl, dst_stride);
    return;
  }

  const int filter_idx = (unit_rtype == RESTORE_SGRPROJ);
  assert(filter_idx < NUM_STRIPE_FILTERS);
  const stripe_filter_fun stripe_filter = stripe_filters[filter_idx];

  const int procunit_width = RESTORATION_PROC_UNIT_SIZE >> ss_x;

  // Convolve the whole tile one stripe at a time
  RestorationTileLimits remaining_stripes = *limits;
  int i = 0;
  while (i < unit_h) {
    int copy_above, copy_below;
    remaining_stripes.v_start = limits->v_start + i;

    get_stripe_boundary_info(&remaining_stripes, tile_rect, ss_y, &copy_above,
                             &copy_below);

    const int full_stripe_height = RESTORATION_PROC_UNIT_SIZE >> ss_y;
    const int runit_offset = RESTORATION_UNIT_OFFSET >> ss_y;

    // Work out where this stripe's boundaries are within
    // rsb->stripe_boundary_{above,below}
    const int tile_stripe =
        (remaining_stripes.v_start - tile_rect->top + runit_offset) /
        full_stripe_height;
    const int frame_stripe = tile_stripe0 + tile_stripe;
    const int rsb_row = RESTORATION_CTX_VERT * frame_stripe;

    // Calculate this stripe's height, based on two rules:
    // * The topmost stripe in each tile is 8 luma pixels shorter than usual.
    // * We can't extend past the end of the current restoration unit
    const int nominal_stripe_height =
        full_stripe_height - ((tile_stripe == 0) ? runit_offset : 0);
    const int h = AOMMIN(nominal_stripe_height,
                         remaining_stripes.v_end - remaining_stripes.v_start);

    setup_processing_stripe_boundary(&remaining_stripes, rsb, rsb_row, h, data,
                                     stride, rlbs, copy_above, copy_below,
                                     optimized_lr);

    stripe_filter(rui, unit_w, h, procunit_width, data_tl + i * stride, stride,
                  dst_tl + i * dst_stride, dst_stride, tmpbuf, bit_depth);

    restore_processing_stripe_boundary(&remaining_stripes, rlbs, h, data,
                                       stride, copy_above, copy_below,
                                       optimized_lr);

    i += h;
  }
}

static void filter_frame_on_unit(const RestorationTileLimits *limits,
                                 const AV1PixelRect *tile_rect,
                                 int rest_unit_idx, int rest_unit_idx_seq,
                                 void *priv, int32_t *tmpbuf,
                                 RestorationLineBuffers *rlbs) {
  (void)rest_unit_idx_seq;
  FilterFrameCtxt *ctxt = (FilterFrameCtxt *)priv;
  const RestorationInfo *rsi = ctxt->rsi;

  av1_loop_restoration_filter_unit(
      limits, &rsi->unit_info[rest_unit_idx], &rsi->boundaries, rlbs, tile_rect,
      ctxt->tile_stripe0, ctxt->ss_x, ctxt->ss_y, ctxt->bit_depth, ctxt->data8,
      ctxt->data_stride, ctxt->dst8, ctxt->dst_stride, tmpbuf,
      rsi->optimized_lr);
}

void av1_loop_restoration_filter_frame_init(AV1LrStruct *lr_ctxt,
                                            YV12_BUFFER_CONFIG *frame,
                                            AV1_COMMON *cm, int optimized_lr,
                                            int num_planes) {
  const SequenceHeader *const seq_params = &cm->seq_params;
  const int bit_depth = seq_params->bit_depth;
  lr_ctxt->dst = &cm->rst_frame;

  const int frame_width = frame->crop_widths[0];
  const int frame_height = frame->crop_heights[0];
  if (aom_realloc_frame_buffer(
          lr_ctxt->dst, frame_width, frame_height, seq_params->subsampling_x,
          seq_params->subsampling_y, AOM_RESTORATION_FRAME_BORDER,
          cm->features.byte_alignment, NULL, NULL, NULL) < 0)
    aom_internal_error(&cm->error, AOM_CODEC_MEM_ERROR,
                       "Failed to allocate restoration dst buffer");

  lr_ctxt->on_rest_unit = filter_frame_on_unit;
  lr_ctxt->frame = frame;
  for (int plane = 0; plane < num_planes; ++plane) {
    RestorationInfo *rsi = &cm->rst_info[plane];
    RestorationType rtype = rsi->frame_restoration_type;
    rsi->optimized_lr = optimized_lr;

    if (rtype == RESTORE_NONE) {
      continue;
    }

    const int is_uv = plane > 0;
    const int plane_width = frame->crop_widths[is_uv];
    const int plane_height = frame->crop_heights[is_uv];
    FilterFrameCtxt *lr_plane_ctxt = &lr_ctxt->ctxt[plane];

    av1_extend_frame(frame->buffers[plane], plane_width, plane_height,
                     frame->strides[is_uv], RESTORATION_BORDER,
                     RESTORATION_BORDER);

    lr_plane_ctxt->rsi = rsi;
    lr_plane_ctxt->ss_x = is_uv && seq_params->subsampling_x;
    lr_plane_ctxt->ss_y = is_uv && seq_params->subsampling_y;
    lr_plane_ctxt->bit_depth = bit_depth;
    lr_plane_ctxt->data8 = frame->buffers[plane];
    lr_plane_ctxt->dst8 = lr_ctxt->dst->buffers[plane];
    lr_plane_ctxt->data_stride = frame->strides[is_uv];
    lr_plane_ctxt->dst_stride = lr_ctxt->dst->strides[is_uv];
    lr_plane_ctxt->tile_rect = av1_whole_frame_rect(cm, is_uv);
    lr_plane_ctxt->tile_stripe0 = 0;
  }
}

void av1_loop_restoration_copy_planes(AV1LrStruct *loop_rest_ctxt,
                                      AV1_COMMON *cm, int num_planes) {
  typedef void (*copy_fun)(const YV12_BUFFER_CONFIG *src_ybc,
                           YV12_BUFFER_CONFIG *dst_ybc, int hstart, int hend,
                           int vstart, int vend);
  static const copy_fun copy_funs[3] = { aom_yv12_partial_coloc_copy_y,
                                         aom_yv12_partial_coloc_copy_u,
                                         aom_yv12_partial_coloc_copy_v };
  assert(num_planes <= 3);
  for (int plane = 0; plane < num_planes; ++plane) {
    if (cm->rst_info[plane].frame_restoration_type == RESTORE_NONE) continue;
    AV1PixelRect tile_rect = loop_rest_ctxt->ctxt[plane].tile_rect;
    copy_funs[plane](loop_rest_ctxt->dst, loop_rest_ctxt->frame, tile_rect.left,
                     tile_rect.right, tile_rect.top, tile_rect.bottom);
  }
}

static void foreach_rest_unit_in_planes(AV1LrStruct *lr_ctxt, AV1_COMMON *cm,
                                        int num_planes) {
  FilterFrameCtxt *ctxt = lr_ctxt->ctxt;

  for (int plane = 0; plane < num_planes; ++plane) {
    if (cm->rst_info[plane].frame_restoration_type == RESTORE_NONE) {
      continue;
    }

    av1_foreach_rest_unit_in_plane(cm, plane, lr_ctxt->on_rest_unit,
                                   &ctxt[plane], &ctxt[plane].tile_rect,
                                   cm->rst_tmpbuf, cm->rlbs);
  }
}

void av1_loop_restoration_filter_frame(YV12_BUFFER_CONFIG *frame,
                                       AV1_COMMON *cm, int optimized_lr,
                                       void *lr_ctxt) {
  assert(!cm->features.all_lossless);
  const int num_planes = av1_num_planes(cm);

  AV1LrStruct *loop_rest_ctxt = (AV1LrStruct *)lr_ctxt;

  av1_loop_restoration_filter_frame_init(loop_rest_ctxt, frame, cm,
                                         optimized_lr, num_planes);

  foreach_rest_unit_in_planes(loop_rest_ctxt, cm, num_planes);

  av1_loop_restoration_copy_planes(loop_rest_ctxt, cm, num_planes);
}

void av1_foreach_rest_unit_in_row(
    RestorationTileLimits *limits, const AV1PixelRect *tile_rect,
    rest_unit_visitor_t on_rest_unit, int row_number, int unit_size,
    int unit_idx0, int hunits_per_tile, int vunits_per_tile, int unit_stride,
    int plane, void *priv, int32_t *tmpbuf, RestorationLineBuffers *rlbs,
    sync_read_fn_t on_sync_read, sync_write_fn_t on_sync_write,
    struct AV1LrSyncData *const lr_sync, int *processed) {
  const int tile_w = tile_rect->right - tile_rect->left;
  const int ext_size = unit_size * 3 / 2;
  int x0 = 0, j = 0;
  while (x0 < tile_w) {
    int remaining_w = tile_w - x0;
    int w = (remaining_w < ext_size) ? remaining_w : unit_size;

    limits->h_start = tile_rect->left + x0;
    limits->h_end = tile_rect->left + x0 + w;
    assert(limits->h_end <= tile_rect->right);

    // Note that the hunits_per_tile is for the number of horz RUs in the
    // rutile, but unit_stride is the stride for RU info for the full frame.
    // If the tile is the full frame, then unit_stride will be the same as
    // hunits_per_tile, but not always.
    const int unit_idx = unit_idx0 + row_number * unit_stride + j;

    // No sync for even numbered rows
    // For odd numbered rows, Loop Restoration of current block requires the LR
    // of top-right and bottom-right blocks to be completed

    // top-right sync
    on_sync_read(lr_sync, row_number, j, plane);
    if ((row_number + 1) < vunits_per_tile)
      // bottom-right sync
      on_sync_read(lr_sync, row_number + 2, j, plane);

    // Note *processed is an index that if provided, is passed down to the
    // visitor function on_rest_unit(), and is then incremented by 1.
    // This can be used by the visitor function as a sequential index.
    on_rest_unit(limits, tile_rect, unit_idx, (processed ? *processed : -1),
                 priv, tmpbuf, rlbs);
    if (processed) (*processed)++;

    on_sync_write(lr_sync, row_number, j, hunits_per_tile, plane);

    x0 += w;
    ++j;
  }
}

void av1_lr_sync_read_dummy(void *const lr_sync, int r, int c, int plane) {
  (void)lr_sync;
  (void)r;
  (void)c;
  (void)plane;
}

void av1_lr_sync_write_dummy(void *const lr_sync, int r, int c,
                             const int sb_cols, int plane) {
  (void)lr_sync;
  (void)r;
  (void)c;
  (void)sb_cols;
  (void)plane;
}

// This is meant to be called when the RUs in an entire coded tile are to
// be processed. The tile_rect passed in is the RU-domain rectangle covering
// all the RUs that are signaled as part of  coded tile. The first RU row is
// expected to be offset. In AV1 syntax, the offsetting only happens for the
// first row in the frame and all other tile boundaries are ignored for the
// purpose of filtering. So whenever this is called make sure that the
// tile_rect passed in is for the entire frame or at least a vertical tile in
// the frame. However we still preserve the generic functionality here in this
// function. In the future if we allow filtering to be conducted independently
// within each tile, this function could be more useful.
void av1_foreach_rest_unit_in_tile(const AV1PixelRect *tile_rect, int unit_idx0,
                                   int hunits_per_tile, int vunits_per_tile,
                                   int unit_stride, int unit_size, int ss_y,
                                   int plane, rest_unit_visitor_t on_rest_unit,
                                   void *priv, int32_t *tmpbuf,
                                   RestorationLineBuffers *rlbs,
                                   int *processed) {
  const int tile_h = tile_rect->bottom - tile_rect->top;
  const int ext_size = unit_size * 3 / 2;

  int y0 = 0, i = 0;
  while (y0 < tile_h) {
    int remaining_h = tile_h - y0;
    int h = (remaining_h < ext_size) ? remaining_h : unit_size;

    RestorationTileLimits limits;
    limits.v_start = tile_rect->top + y0;
    limits.v_end = tile_rect->top + y0 + h;
    assert(limits.v_end <= tile_rect->bottom);
    // Offset the tile upwards to align with the restoration processing stripe
    const int voffset = RESTORATION_UNIT_OFFSET >> ss_y;
    limits.v_start = AOMMAX(tile_rect->top, limits.v_start - voffset);
    if (limits.v_end < tile_rect->bottom) limits.v_end -= voffset;

    assert(i < vunits_per_tile);
    av1_foreach_rest_unit_in_row(
        &limits, tile_rect, on_rest_unit, i, unit_size, unit_idx0,
        hunits_per_tile, vunits_per_tile, unit_stride, plane, priv, tmpbuf,
        rlbs, av1_lr_sync_read_dummy, av1_lr_sync_write_dummy, NULL, processed);

    y0 += h;
    ++i;
  }
}

// This is meant to be called when the RUs in a single coded SB are to be
// processed. The tile_rect passed in is the RU-domain rectangle covering
// all the RUs that are signaled as part of  coded SB. The first RU row is
// expected to be offset only if the tile_rect starts at row 0. Note that
// this is a simple variation of the function above and could have been
// combined, but they are kept distinct to avoid confusion in the future.
void av1_foreach_rest_unit_in_sb(const AV1PixelRect *tile_rect, int unit_idx0,
                                 int hunits_per_tile, int vunits_per_tile,
                                 int unit_stride, int unit_size, int ss_y,
                                 int plane, rest_unit_visitor_t on_rest_unit,
                                 void *priv, int32_t *tmpbuf,
                                 RestorationLineBuffers *rlbs, int *processed) {
  const int tile_h = tile_rect->bottom - tile_rect->top;
  const int ext_size = unit_size * 3 / 2 + RESTORATION_UNIT_OFFSET;

  int y0 = 0, i = 0;
  while (y0 < tile_h) {
    int remaining_h = tile_h - y0;
    int h = (remaining_h < ext_size) ? remaining_h : unit_size;

    RestorationTileLimits limits;
    limits.v_start = tile_rect->top + y0;
    limits.v_end = tile_rect->top + y0 + h;
    assert(limits.v_end <= tile_rect->bottom);
    // Offset the tile upwards to align with the restoration processing stripe
    // if the SB that iuncludes the RUs in this group are the top row
    if (tile_rect->top == 0) {
      const int voffset = RESTORATION_UNIT_OFFSET >> ss_y;
      limits.v_start = AOMMAX(tile_rect->top, limits.v_start - voffset);
      if (limits.v_end < tile_rect->bottom) limits.v_end -= voffset;
      h = limits.v_end - limits.v_start;
    }

    assert(i < vunits_per_tile);
    av1_foreach_rest_unit_in_row(
        &limits, tile_rect, on_rest_unit, i, unit_size, unit_idx0,
        hunits_per_tile, vunits_per_tile, unit_stride, plane, priv, tmpbuf,
        rlbs, av1_lr_sync_read_dummy, av1_lr_sync_write_dummy, NULL, processed);

    y0 += h;
    ++i;
  }
}

void av1_foreach_rest_unit_in_plane(const struct AV1Common *cm, int plane,
                                    rest_unit_visitor_t on_rest_unit,
                                    void *priv, AV1PixelRect *tile_rect,
                                    int32_t *tmpbuf,
                                    RestorationLineBuffers *rlbs) {
  const int is_uv = plane > 0;
  const int ss_y = is_uv && cm->seq_params.subsampling_y;

  const RestorationInfo *rsi = &cm->rst_info[plane];

  const int unit_idx0 =
      (LR_TILE_ROW * LR_TILE_COLS + LR_TILE_COL) * rsi->units_per_tile;
  int processed = 0;
  av1_foreach_rest_unit_in_tile(
      tile_rect, unit_idx0, rsi->horz_units_per_tile, rsi->vert_units_per_tile,
      rsi->horz_units_per_tile, rsi->restoration_unit_size, ss_y, plane,
      on_rest_unit, priv, tmpbuf, rlbs, &processed);
}

int av1_loop_restoration_corners_in_sb(const struct AV1Common *cm, int plane,
                                       int mi_row, int mi_col, BLOCK_SIZE bsize,
                                       int *rcol0, int *rcol1, int *rrow0,
                                       int *rrow1) {
  assert(rcol0 && rcol1 && rrow0 && rrow1);

  if (bsize != cm->seq_params.sb_size) return 0;

  assert(!cm->features.all_lossless);

  const int is_uv = plane > 0;

  const AV1PixelRect tile_rect = av1_whole_frame_rect(cm, is_uv);
  const int tile_w = tile_rect.right - tile_rect.left;
  const int tile_h = tile_rect.bottom - tile_rect.top;

  const int mi_top = 0;
  const int mi_left = 0;

  // Compute the mi-unit corners of the superblock relative to the top-left of
  // the tile
  const int mi_rel_row0 = mi_row - mi_top;
  const int mi_rel_col0 = mi_col - mi_left;
  const int mi_rel_row1 = mi_rel_row0 + mi_size_high[bsize];
  const int mi_rel_col1 = mi_rel_col0 + mi_size_wide[bsize];

  const RestorationInfo *rsi = &cm->rst_info[plane];
  const int size = rsi->restoration_unit_size;

  // Calculate the number of restoration units in this tile (which might be
  // strictly less than rsi->horz_units_per_tile and rsi->vert_units_per_tile)
  const int horz_units = av1_lr_count_units_in_tile(size, tile_w);
  const int vert_units = av1_lr_count_units_in_tile(size, tile_h);

  // The size of an MI-unit on this plane of the image
  const int ss_x = is_uv && cm->seq_params.subsampling_x;
  const int ss_y = is_uv && cm->seq_params.subsampling_y;
  const int mi_size_x = MI_SIZE >> ss_x;
  const int mi_size_y = MI_SIZE >> ss_y;

  // Write m for the relative mi column or row, D for the superres denominator
  // and N for the superres numerator. If u is the upscaled pixel offset then
  // we can write the downscaled pixel offset in two ways as:
  //
  //   MI_SIZE * m = N / D u
  //
  // from which we get u = D * MI_SIZE * m / N
  const int mi_to_num_x = av1_superres_scaled(cm)
                              ? mi_size_x * cm->superres_scale_denominator
                              : mi_size_x;
  const int mi_to_num_y = mi_size_y;
  const int denom_x = av1_superres_scaled(cm) ? size * SCALE_NUMERATOR : size;
  const int denom_y = size;

  const int rnd_x = denom_x - 1;
  const int rnd_y = denom_y - 1;

  // rcol0/rrow0 should be the first column/row of restoration units (relative
  // to the top-left of the tile) that doesn't start left/below of
  // mi_col/mi_row. For this calculation, we need to round up the division (if
  // the sb starts at runit column 10.1, the first matching runit has column
  // index 11)
  *rcol0 = (mi_rel_col0 * mi_to_num_x + rnd_x) / denom_x;
  *rrow0 = (mi_rel_row0 * mi_to_num_y + rnd_y) / denom_y;

  // rel_col1/rel_row1 is the equivalent calculation, but for the superblock
  // below-right. If we're at the bottom or right of the tile, this restoration
  // unit might not exist, in which case we'll clamp accordingly.
  *rcol1 = AOMMIN((mi_rel_col1 * mi_to_num_x + rnd_x) / denom_x, horz_units);
  *rrow1 = AOMMIN((mi_rel_row1 * mi_to_num_y + rnd_y) / denom_y, vert_units);

  return *rcol0 < *rcol1 && *rrow0 < *rrow1;
}

// Extend to left and right
static void extend_lines(uint16_t *buf, int width, int height, int stride,
                         int extend) {
  for (int i = 0; i < height; ++i) {
    aom_memset16(buf - extend, buf[0], extend);
    aom_memset16(buf + width, buf[width - 1], extend);
    buf += stride;
  }
}

static void save_deblock_boundary_lines(
    const YV12_BUFFER_CONFIG *frame, const AV1_COMMON *cm, int plane, int row,
    int stripe, int is_above, RestorationStripeBoundaries *boundaries) {
  const int is_uv = plane > 0;
  const uint16_t *src_buf = frame->buffers[plane];
  const int src_stride = frame->strides[is_uv];
  const uint16_t *src_rows = src_buf + row * src_stride;

  uint16_t *bdry_buf = is_above ? boundaries->stripe_boundary_above
                                : boundaries->stripe_boundary_below;
  uint16_t *bdry_start = bdry_buf + (RESTORATION_EXTRA_HORZ);
  const int bdry_stride = boundaries->stripe_boundary_stride;
  uint16_t *bdry_rows =
      bdry_start + RESTORATION_CTX_VERT * stripe * bdry_stride;

  // There is a rare case in which a processing stripe can end 1px above the
  // crop border. In this case, we do want to use deblocked pixels from below
  // the stripe (hence why we ended up in this function), but instead of
  // fetching 2 "below" rows we need to fetch one and duplicate it.
  // This is equivalent to clamping the sample locations against the crop border
  const int lines_to_save =
      AOMMIN(RESTORATION_CTX_VERT, frame->crop_heights[is_uv] - row);
  assert(lines_to_save == 1 || lines_to_save == 2);

  int upscaled_width;
  int line_bytes;
  if (av1_superres_scaled(cm)) {
    const int ss_x = is_uv && cm->seq_params.subsampling_x;
    upscaled_width = (cm->superres_upscaled_width + ss_x) >> ss_x;
    line_bytes = upscaled_width << 1;
    av1_upscale_normative_rows(cm, src_rows, frame->strides[is_uv], bdry_rows,
                               boundaries->stripe_boundary_stride, plane,
                               lines_to_save);
  } else {
    upscaled_width = frame->crop_widths[is_uv];
    line_bytes = upscaled_width << 1;
    for (int i = 0; i < lines_to_save; i++) {
      memcpy(bdry_rows + i * bdry_stride, src_rows + i * src_stride,
             line_bytes);
    }
  }
  // If we only saved one line, then copy it into the second line buffer
  if (lines_to_save == 1)
    memcpy(bdry_rows + bdry_stride, bdry_rows, line_bytes);

  extend_lines(bdry_rows, upscaled_width, RESTORATION_CTX_VERT, bdry_stride,
               RESTORATION_EXTRA_HORZ);
}

static void save_cdef_boundary_lines(const YV12_BUFFER_CONFIG *frame,
                                     const AV1_COMMON *cm, int plane, int row,
                                     int stripe, int is_above,
                                     RestorationStripeBoundaries *boundaries) {
  const int is_uv = plane > 0;
  const uint16_t *src_buf = frame->buffers[plane];
  const int src_stride = frame->strides[is_uv];
  const uint16_t *src_rows = src_buf + row * src_stride;

  uint16_t *bdry_buf = is_above ? boundaries->stripe_boundary_above
                                : boundaries->stripe_boundary_below;
  uint16_t *bdry_start = bdry_buf + RESTORATION_EXTRA_HORZ;
  const int bdry_stride = boundaries->stripe_boundary_stride;
  uint16_t *bdry_rows =
      bdry_start + RESTORATION_CTX_VERT * stripe * bdry_stride;
  const int src_width = frame->crop_widths[is_uv];

  // At the point where this function is called, we've already applied
  // superres. So we don't need to extend the lines here, we can just
  // pull directly from the topmost row of the upscaled frame.
  const int ss_x = is_uv && cm->seq_params.subsampling_x;
  const int upscaled_width = av1_superres_scaled(cm)
                                 ? (cm->superres_upscaled_width + ss_x) >> ss_x
                                 : src_width;
  const int line_bytes = upscaled_width << 1;
  for (int i = 0; i < RESTORATION_CTX_VERT; i++) {
    // Copy the line at 'row' into both context lines. This is because
    // we want to (effectively) extend the outermost row of CDEF data
    // from this tile to produce a border, rather than using deblocked
    // pixels from the tile above/below.
    memcpy(bdry_rows + i * bdry_stride, src_rows, line_bytes);
  }
  extend_lines(bdry_rows, upscaled_width, RESTORATION_CTX_VERT, bdry_stride,
               RESTORATION_EXTRA_HORZ);
}

static void save_tile_row_boundary_lines(const YV12_BUFFER_CONFIG *frame,
                                         int plane, AV1_COMMON *cm,
                                         int after_cdef) {
  const int is_uv = plane > 0;
  const int ss_y = is_uv && cm->seq_params.subsampling_y;
  const int stripe_height = RESTORATION_PROC_UNIT_SIZE >> ss_y;
  const int stripe_off = RESTORATION_UNIT_OFFSET >> ss_y;

  // Get the tile rectangle, with height rounded up to the next multiple of 8
  // luma pixels (only relevant for the bottom tile of the frame)
  const AV1PixelRect tile_rect = av1_whole_frame_rect(cm, is_uv);
  const int stripe0 = 0;

  RestorationStripeBoundaries *boundaries = &cm->rst_info[plane].boundaries;

  const int plane_height =
      ROUND_POWER_OF_TWO(cm->superres_upscaled_height, ss_y);

  int tile_stripe;
  for (tile_stripe = 0;; ++tile_stripe) {
    const int rel_y0 = AOMMAX(0, tile_stripe * stripe_height - stripe_off);
    const int y0 = tile_rect.top + rel_y0;
    if (y0 >= tile_rect.bottom) break;

    const int rel_y1 = (tile_stripe + 1) * stripe_height - stripe_off;
    const int y1 = AOMMIN(tile_rect.top + rel_y1, tile_rect.bottom);

    const int frame_stripe = stripe0 + tile_stripe;

    // In this case, we should only use CDEF pixels at the top
    // and bottom of the frame as a whole; internal tile boundaries
    // can use deblocked pixels from adjacent tiles for context.
    const int use_deblock_above = (frame_stripe > 0);
    const int use_deblock_below = (y1 < plane_height);

    if (!after_cdef) {
      // Save deblocked context where needed.
      if (use_deblock_above) {
        save_deblock_boundary_lines(frame, cm, plane, y0 - RESTORATION_CTX_VERT,
                                    frame_stripe, 1, boundaries);
      }
      if (use_deblock_below) {
        save_deblock_boundary_lines(frame, cm, plane, y1, frame_stripe, 0,
                                    boundaries);
      }
    } else {
      // Save CDEF context where needed. Note that we need to save the CDEF
      // context for a particular boundary iff we *didn't* save deblocked
      // context for that boundary.
      //
      // In addition, we need to save copies of the outermost line within
      // the tile, rather than using data from outside the tile.
      if (!use_deblock_above) {
        save_cdef_boundary_lines(frame, cm, plane, y0, frame_stripe, 1,
                                 boundaries);
      }
      if (!use_deblock_below) {
        save_cdef_boundary_lines(frame, cm, plane, y1 - 1, frame_stripe, 0,
                                 boundaries);
      }
    }
  }
}

// For each RESTORATION_PROC_UNIT_SIZE pixel high stripe, save 4 scan
// lines to be used as boundary in the loop restoration process. The
// lines are saved in rst_internal.stripe_boundary_lines
void av1_loop_restoration_save_boundary_lines(const YV12_BUFFER_CONFIG *frame,
                                              AV1_COMMON *cm, int after_cdef) {
  const int num_planes = av1_num_planes(cm);
  for (int p = 0; p < num_planes; ++p) {
    save_tile_row_boundary_lines(frame, p, cm, after_cdef);
  }
}