lrf.rs

// Copyright (c) 2017-2019, The rav1e contributors. 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.

cfg_if::cfg_if! {
  if #[cfg(nasm_x86_64)] {
    use crate::asm::x86::lrf::*;
  } else {
    use self::rust::*;
  }
}

use crate::context::PLANES;
use crate::context::SB_SIZE;
use crate::encoder::FrameInvariants;
use crate::frame::AsRegion;
use crate::frame::Frame;
use crate::frame::Plane;
use crate::frame::PlaneConfig;
use crate::frame::PlaneOffset;
use crate::frame::PlaneSlice;
use crate::hawktracer::*;
use crate::tiling::{Area, PlaneRegionMut, Rect};
use crate::util::clamp;
use crate::util::CastFromPrimitive;
use crate::util::ILog;
use crate::util::Pixel;

use crate::api::SGRComplexityLevel;
use std::cmp;
use std::iter::FusedIterator;
use std::ops::{Index, IndexMut};

pub const RESTORATION_TILESIZE_MAX_LOG2: usize = 8;

pub const RESTORE_NONE: u8 = 0;
pub const RESTORE_SWITCHABLE: u8 = 1;
pub const RESTORE_WIENER: u8 = 2;
pub const RESTORE_SGRPROJ: u8 = 3;

pub const WIENER_TAPS_MIN: [i8; 3] = [-5, -23, -17];
pub const WIENER_TAPS_MID: [i8; 3] = [3, -7, 15];
pub const WIENER_TAPS_MAX: [i8; 3] = [10, 8, 46];
#[allow(unused)]
pub const WIENER_TAPS_K: [i8; 3] = [1, 2, 3];
pub const WIENER_BITS: usize = 7;

pub const SGRPROJ_XQD_MIN: [i8; 2] = [-96, -32];
pub const SGRPROJ_XQD_MID: [i8; 2] = [-32, 31];
pub const SGRPROJ_XQD_MAX: [i8; 2] = [31, 95];
pub const SGRPROJ_PRJ_SUBEXP_K: u8 = 4;
pub const SGRPROJ_PRJ_BITS: u8 = 7;
pub const SGRPROJ_PARAMS_BITS: u8 = 4;
pub const SGRPROJ_MTABLE_BITS: u8 = 20;
pub const SGRPROJ_SGR_BITS: u8 = 8;
pub const SGRPROJ_RECIP_BITS: u8 = 12;
pub const SGRPROJ_RST_BITS: u8 = 4;
pub const SGRPROJ_PARAMS_S: [[u32; 2]; 1 << SGRPROJ_PARAMS_BITS] = [
  [140, 3236],
  [112, 2158],
  [93, 1618],
  [80, 1438],
  [70, 1295],
  [58, 1177],
  [47, 1079],
  [37, 996],
  [30, 925],
  [25, 863],
  [0, 2589],
  [0, 1618],
  [0, 1177],
  [0, 925],
  [56, 0],
  [22, 0],
];

// List of indices to SGRPROJ_PARAMS_S values that at a given complexity level.
// SGRPROJ_ALL_SETS contains every possible index
const SGRPROJ_ALL_SETS: &[u8] =
  &[0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15];
// SGRPROJ_REDUCED_SETS has half of the values. Using only these values gives
// most of the gains from sgr. The decision of which values to use is somewhat
// arbitrary. The sgr parameters has 3 discontinuous groups. The first has both
// parameters as non-zero. The other two are distinguishable by which of the
// two parameters is zero. There are an even number of each of these groups and
// the non-zero parameters grow as the indices increase. This array uses the
// 1nd, 3rd, ... smallest params of each group.
const SGRPROJ_REDUCED_SETS: &[u8] = &[1, 3, 5, 7, 9, 11, 13, 15];

pub fn get_sgr_sets(complexity: SGRComplexityLevel) -> &'static [u8] {
  match complexity {
    SGRComplexityLevel::Full => SGRPROJ_ALL_SETS,
    SGRComplexityLevel::Reduced => SGRPROJ_REDUCED_SETS,
  }
}

pub const SOLVE_IMAGE_MAX: usize = 1 << RESTORATION_TILESIZE_MAX_LOG2;
pub const SOLVE_IMAGE_STRIDE: usize = SOLVE_IMAGE_MAX + 6 + 2;
pub const SOLVE_IMAGE_HEIGHT: usize = SOLVE_IMAGE_STRIDE;
pub const SOLVE_IMAGE_SIZE: usize = SOLVE_IMAGE_STRIDE * SOLVE_IMAGE_HEIGHT;

pub const STRIPE_IMAGE_MAX: usize = (1 << RESTORATION_TILESIZE_MAX_LOG2)
  + (1 << (RESTORATION_TILESIZE_MAX_LOG2 - 1));
pub const STRIPE_IMAGE_STRIDE: usize = STRIPE_IMAGE_MAX + 6 + 2;
pub const STRIPE_IMAGE_HEIGHT: usize = 64 + 6 + 2;
pub const STRIPE_IMAGE_SIZE: usize = STRIPE_IMAGE_STRIDE * STRIPE_IMAGE_HEIGHT;

pub const IMAGE_WIDTH_MAX: usize = [STRIPE_IMAGE_MAX, SOLVE_IMAGE_MAX]
  [(STRIPE_IMAGE_MAX < SOLVE_IMAGE_MAX) as usize];

/// The buffer used in `sgrproj_stripe_filter()` and `sgrproj_solve()`.
#[derive(Debug)]
pub struct IntegralImageBuffer {
  pub integral_image: Vec<u32>,
  pub sq_integral_image: Vec<u32>,
}

impl IntegralImageBuffer {
  /// Creates a new buffer with the given size, filled with zeros.
  #[inline]
  pub fn zeroed(size: usize) -> Self {
    Self { integral_image: vec![0; size], sq_integral_image: vec![0; size] }
  }
}

#[allow(unused)] // Wiener coming soon!
#[derive(Copy, Clone, Debug, PartialEq, Eq)]
pub enum RestorationFilter {
  None,
  Wiener { coeffs: [[i8; 3]; 2] },
  Sgrproj { set: u8, xqd: [i8; 2] },
}

impl Default for RestorationFilter {
  fn default() -> RestorationFilter {
    RestorationFilter::None
  }
}

impl RestorationFilter {
  pub fn notequal(self, cmp: RestorationFilter) -> bool {
    match self {
      RestorationFilter::None {} => {
        if let RestorationFilter::None {} = cmp {
          false
        } else {
          true
        }
      }
      RestorationFilter::Sgrproj { set, xqd } => {
        if let RestorationFilter::Sgrproj { set: set2, xqd: xqd2 } = cmp {
          !(set == set2 && xqd[0] == xqd2[0] && xqd[1] == xqd2[1])
        } else {
          true
        }
      }
      RestorationFilter::Wiener { coeffs } => {
        if let RestorationFilter::Wiener { coeffs: coeffs2 } = cmp {
          !(coeffs[0][0] == coeffs2[0][0]
            && coeffs[0][1] == coeffs2[0][1]
            && coeffs[0][2] == coeffs2[0][2]
            && coeffs[1][0] == coeffs2[1][0]
            && coeffs[1][1] == coeffs2[1][1]
            && coeffs[1][2] == coeffs2[1][2])
        } else {
          true
        }
      }
    }
  }
}

pub(crate) mod rust {
  use crate::cpu_features::CpuFeatureLevel;
  use crate::frame::PlaneSlice;
  use crate::lrf::{
    get_integral_square, sgrproj_sum_finish, SGRPROJ_RST_BITS,
    SGRPROJ_SGR_BITS,
  };
  use crate::util::CastFromPrimitive;
  use crate::Pixel;

  #[inline(always)]
  pub(crate) fn sgrproj_box_ab_internal(
    r: usize, af: &mut [u32], bf: &mut [u32], iimg: &[u32], iimg_sq: &[u32],
    iimg_stride: usize, start_x: usize, y: usize, stripe_w: usize, s: u32,
    bdm8: usize,
  ) {
    let d: usize = r * 2 + 1;
    let n: usize = d * d;
    let one_over_n = if r == 1 { 455 } else { 164 };

    for x in start_x..stripe_w + 2 {
      let sum = get_integral_square(iimg, iimg_stride, x, y, d);
      let ssq = get_integral_square(iimg_sq, iimg_stride, x, y, d);
      let (reta, retb) =
        sgrproj_sum_finish(ssq, sum, n as u32, one_over_n, s, bdm8);
      af[x] = reta;
      bf[x] = retb;
    }
  }

  // computes an intermediate (ab) row for stripe_w + 2 columns at row y
  pub(crate) fn sgrproj_box_ab_r1(
    af: &mut [u32], bf: &mut [u32], iimg: &[u32], iimg_sq: &[u32],
    iimg_stride: usize, y: usize, stripe_w: usize, s: u32, bdm8: usize,
    _cpu: CpuFeatureLevel,
  ) {
    sgrproj_box_ab_internal(
      1,
      af,
      bf,
      iimg,
      iimg_sq,
      iimg_stride,
      0,
      y,
      stripe_w,
      s,
      bdm8,
    );
  }

  // computes an intermediate (ab) row for stripe_w + 2 columns at row y
  pub(crate) fn sgrproj_box_ab_r2(
    af: &mut [u32], bf: &mut [u32], iimg: &[u32], iimg_sq: &[u32],
    iimg_stride: usize, y: usize, stripe_w: usize, s: u32, bdm8: usize,
    _cpu: CpuFeatureLevel,
  ) {
    sgrproj_box_ab_internal(
      2,
      af,
      bf,
      iimg,
      iimg_sq,
      iimg_stride,
      0,
      y,
      stripe_w,
      s,
      bdm8,
    );
  }

  pub(crate) fn sgrproj_box_f_r0<T: Pixel>(
    f: &mut [u32], y: usize, w: usize, cdeffed: &PlaneSlice<T>,
    _cpu: CpuFeatureLevel,
  ) {
    sgrproj_box_f_r0_internal(f, 0, y, w, cdeffed);
  }

  #[inline(always)]
  pub(crate) fn sgrproj_box_f_r0_internal<T: Pixel>(
    f: &mut [u32], start_x: usize, y: usize, w: usize, cdeffed: &PlaneSlice<T>,
  ) {
    for x in start_x..w {
      f[x] = (u32::cast_from(cdeffed.p(x, y))) << SGRPROJ_RST_BITS;
    }
  }

  pub(crate) fn sgrproj_box_f_r1<T: Pixel>(
    af: &[&[u32]; 3], bf: &[&[u32]; 3], f: &mut [u32], y: usize, w: usize,
    cdeffed: &PlaneSlice<T>, _cpu: CpuFeatureLevel,
  ) {
    sgrproj_box_f_r1_internal(af, bf, f, 0, y, w, cdeffed);
  }

  #[inline(always)]
  pub(crate) fn sgrproj_box_f_r1_internal<T: Pixel>(
    af: &[&[u32]; 3], bf: &[&[u32]; 3], f: &mut [u32], start_x: usize,
    y: usize, w: usize, cdeffed: &PlaneSlice<T>,
  ) {
    let shift = 5 + SGRPROJ_SGR_BITS - SGRPROJ_RST_BITS;
    for x in start_x..w {
      let a = 3 * (af[0][x] + af[2][x] + af[0][x + 2] + af[2][x + 2])
        + 4
          * (af[1][x]
            + af[0][x + 1]
            + af[1][x + 1]
            + af[2][x + 1]
            + af[1][x + 2]);
      let b = 3 * (bf[0][x] + bf[2][x] + bf[0][x + 2] + bf[2][x + 2])
        + 4
          * (bf[1][x]
            + bf[0][x + 1]
            + bf[1][x + 1]
            + bf[2][x + 1]
            + bf[1][x + 2]);
      let v = a * u32::cast_from(cdeffed.p(x, y)) + b;
      f[x] = (v + (1 << shift >> 1)) >> shift;
    }
  }

  pub(crate) fn sgrproj_box_f_r2<T: Pixel>(
    af: &[&[u32]; 2], bf: &[&[u32]; 2], f0: &mut [u32], f1: &mut [u32],
    y: usize, w: usize, cdeffed: &PlaneSlice<T>, _cpu: CpuFeatureLevel,
  ) {
    sgrproj_box_f_r2_internal(af, bf, f0, f1, 0, y, w, cdeffed);
  }

  #[inline(always)]
  pub(crate) fn sgrproj_box_f_r2_internal<T: Pixel>(
    af: &[&[u32]; 2], bf: &[&[u32]; 2], f0: &mut [u32], f1: &mut [u32],
    start_x: usize, y: usize, w: usize, cdeffed: &PlaneSlice<T>,
  ) {
    let shift = 5 + SGRPROJ_SGR_BITS - SGRPROJ_RST_BITS;
    let shifto = 4 + SGRPROJ_SGR_BITS - SGRPROJ_RST_BITS;
    for x in start_x..w {
      let a = 5 * (af[0][x] + af[0][x + 2]) + 6 * (af[0][x + 1]);
      let b = 5 * (bf[0][x] + bf[0][x + 2]) + 6 * (bf[0][x + 1]);
      let ao = 5 * (af[1][x] + af[1][x + 2]) + 6 * (af[1][x + 1]);
      let bo = 5 * (bf[1][x] + bf[1][x + 2]) + 6 * (bf[1][x + 1]);
      let v = (a + ao) * u32::cast_from(cdeffed.p(x, y)) + b + bo;
      f0[x] = (v + (1 << shift >> 1)) >> shift;
      let vo = ao * u32::cast_from(cdeffed.p(x, y + 1)) + bo;
      f1[x] = (vo + (1 << shifto >> 1)) >> shifto;
    }
  }
}

#[inline(always)]
fn sgrproj_sum_finish(
  ssq: u32, sum: u32, n: u32, one_over_n: u32, s: u32, bdm8: usize,
) -> (u32, u32) {
  let scaled_ssq = (ssq + (1 << (2 * bdm8) >> 1)) >> (2 * bdm8);
  let scaled_sum = (sum + (1 << bdm8 >> 1)) >> bdm8;
  let p =
    cmp::max(0, (scaled_ssq * n) as i32 - (scaled_sum * scaled_sum) as i32)
      as u32;
  let z = (p * s + (1 << SGRPROJ_MTABLE_BITS >> 1)) >> SGRPROJ_MTABLE_BITS;
  let a = if z >= 255 {
    256
  } else if z == 0 {
    1
  } else {
    ((z << SGRPROJ_SGR_BITS) + z / 2) / (z + 1)
  };
  let b = ((1 << SGRPROJ_SGR_BITS) - a) * sum * one_over_n;
  (a, (b + (1 << SGRPROJ_RECIP_BITS >> 1)) >> SGRPROJ_RECIP_BITS)
}

// Using an integral image, compute the sum of a square region
fn get_integral_square(
  iimg: &[u32], stride: usize, x: usize, y: usize, size: usize,
) -> u32 {
  // Cancel out overflow in iimg by using wrapping arithmetic
  iimg[y * stride + x]
    .wrapping_add(iimg[(y + size) * stride + x + size])
    .wrapping_sub(iimg[(y + size) * stride + x])
    .wrapping_sub(iimg[y * stride + x + size])
}

struct VertPaddedIter<'a, T: Pixel> {
  // The two sources that can be selected when clipping
  deblocked: &'a Plane<T>,
  cdeffed: &'a Plane<T>,
  // x index to choice where on the row to start
  x: isize,
  // y index that will be mutated
  y: isize,
  // The index at which to terminate. Can be larger than the slice length.
  end: isize,
  // Used for source buffer choice/clipping. May (and regularly will)
  // be negative.
  stripe_begin: isize,
  // Also used for source buffer choice/clipping. May specify a stripe boundary
  // less than, equal to, or larger than the buffers we're accessing.
  stripe_end: isize,
  // Active area cropping is done by specifying a value smaller than the height
  // of the plane.
  crop: isize,
}

impl<'a, 'b, T: Pixel> VertPaddedIter<'a, T> {
  fn new(
    cdeffed: &PlaneSlice<'a, T>, deblocked: &PlaneSlice<'a, T>,
    stripe_h: usize, crop: usize,
  ) -> VertPaddedIter<'a, T> {
    // cdeffed and deblocked must start at the same coordinates from their
    // underlying planes. Since cropping is provided via a separate params, the
    // height of the underlying planes do not need to match.
    assert_eq!(cdeffed.x, deblocked.x);
    assert_eq!(cdeffed.y, deblocked.y);

    // To share integral images, always use the max box filter radius of 2
    let r = 2;

    // The number of rows outside the stripe are needed
    let rows_above = r + 2;
    let rows_below = 2;

    // Offset crop and stripe_h so they are relative to the underlying plane
    // and not the plane slice.
    let crop = crop as isize + deblocked.y;
    let stripe_end = stripe_h as isize + deblocked.y;

    // Move y up the number rows above.
    // If y is negative we repeat the first row
    let y = deblocked.y - rows_above as isize;

    VertPaddedIter {
      deblocked: deblocked.plane,
      cdeffed: cdeffed.plane,
      x: deblocked.x,
      y,
      end: (rows_above + stripe_h + rows_below) as isize + y,
      stripe_begin: deblocked.y,
      stripe_end,
      crop,
    }
  }
}

impl<'a, T: Pixel> Iterator for VertPaddedIter<'a, T> {
  type Item = &'a [T];

  #[inline(always)]
  fn next(&mut self) -> Option<Self::Item> {
    if self.end > self.y {
      // clamp before deciding the source
      // clamp vertically to storage at top and passed-in height at bottom
      let cropped_y = clamp(self.y, 0, self.crop - 1);
      // clamp vertically to stripe limits
      let ly = clamp(cropped_y, self.stripe_begin - 2, self.stripe_end + 1);

      // decide if we're vertically inside or outside the strip
      let src_plane =
        if ly >= self.stripe_begin && ly < self.stripe_end as isize {
          self.cdeffed
        } else {
          self.deblocked
        };
      // cannot directly return self.ps.row(row) due to lifetime issue
      let range = src_plane.row_range(self.x, ly);
      self.y += 1;
      Some(&src_plane.data[range])
    } else {
      None
    }
  }

  fn size_hint(&self) -> (usize, Option<usize>) {
    let remaining = self.end - self.y;
    debug_assert!(remaining >= 0);
    let remaining = remaining as usize;

    (remaining, Some(remaining))
  }
}

impl<T: Pixel> ExactSizeIterator for VertPaddedIter<'_, T> {}
impl<T: Pixel> FusedIterator for VertPaddedIter<'_, T> {}

struct HorzPaddedIter<'a, T: Pixel> {
  // Active area cropping is done using the length of the slice
  slice: &'a [T],
  // x index of the iterator
  // When less than 0, repeat the first element. When greater than end, repeat
  // the last element
  index: isize,
  // The index at which to terminate. Can be larger than the slice length.
  end: usize,
}

impl<'a, T: Pixel> HorzPaddedIter<'a, T> {
  fn new(
    slice: &'a [T], start_index: isize, width: usize,
  ) -> HorzPaddedIter<'a, T> {
    HorzPaddedIter {
      slice,
      index: start_index,
      end: (width as isize + start_index) as usize,
    }
  }
}

impl<'a, T: Pixel> Iterator for HorzPaddedIter<'a, T> {
  type Item = &'a T;

  #[inline(always)]
  fn next(&mut self) -> Option<Self::Item> {
    if self.index < self.end as isize {
      // clamp to the edges of the frame
      let x = clamp(self.index, 0, self.slice.len() as isize - 1) as usize;
      self.index += 1;
      Some(&self.slice[x])
    } else {
      None
    }
  }

  #[inline(always)]
  fn size_hint(&self) -> (usize, Option<usize>) {
    let size: usize = (self.end as isize - self.index) as usize;
    (size, Some(size))
  }
}

impl<T: Pixel> ExactSizeIterator for HorzPaddedIter<'_, T> {}
impl<T: Pixel> FusedIterator for HorzPaddedIter<'_, T> {}

pub fn setup_integral_image<T: Pixel>(
  integral_image_buffer: &mut IntegralImageBuffer,
  integral_image_stride: usize, crop_w: usize, crop_h: usize, stripe_w: usize,
  stripe_h: usize, cdeffed: &PlaneSlice<T>, deblocked: &PlaneSlice<T>,
) {
  let integral_image = &mut integral_image_buffer.integral_image;
  let sq_integral_image = &mut integral_image_buffer.sq_integral_image;

  // Number of elements outside the stripe
  let left_w = 4; // max radius of 2 + 2 padding
  let right_w = 3; // max radius of 2 + 1 padding

  assert_eq!(cdeffed.x, deblocked.x);

  // Find how many unique elements to use to the left and right
  let left_uniques = if cdeffed.x == 0 { 0 } else { left_w };
  let right_uniques = right_w.min(crop_w - stripe_w);

  // Find the total number of unique elements used
  let row_uniques = left_uniques + stripe_w + right_uniques;

  // Negative start indices result in repeating the first element of the row
  let start_index_x = if cdeffed.x == 0 { -(left_w as isize) } else { 0 };

  let mut rows_iter = VertPaddedIter::new(
    // Move left to encompass all the used data
    &cdeffed.go_left(left_uniques),
    &deblocked.go_left(left_uniques),
    // since r2 uses every other row, we need an extra row if stripe_h is odd
    stripe_h + (stripe_h & 1),
    crop_h,
  )
  .map(|row: &[T]| {
    HorzPaddedIter::new(
      // Limit how many unique elements we use
      &row[..row_uniques],
      start_index_x,
      left_w + stripe_w + right_w,
    )
  });

  // Setup the first row
  {
    let mut sum: u32 = 0;
    let mut sq_sum: u32 = 0;
    // Remove the first row and use it outside of the main loop
    let row = rows_iter.next().unwrap();
    for (src, (integral, sq_integral)) in
      row.zip(integral_image.iter_mut().zip(sq_integral_image.iter_mut()))
    {
      let current = u32::cast_from(*src);

      // Wrap adds to prevent undefined behaviour on overflow. Overflow is
      // cancelled out when calculating the sum of a region.
      sum = sum.wrapping_add(current);
      *integral = sum;
      sq_sum = sq_sum.wrapping_add(current * current);
      *sq_integral = sq_sum;
    }
  }
  // Calculate all other rows
  let mut integral_slice = &mut integral_image[..];
  let mut sq_integral_slice = &mut sq_integral_image[..];
  for row in rows_iter {
    let mut sum: u32 = 0;
    let mut sq_sum: u32 = 0;

    // Split the data between the previous row and future rows.
    // This allows us to mutate the current row while accessing the
    // previous row.
    let (integral_row_prev, integral_row) =
      integral_slice.split_at_mut(integral_image_stride);
    let (sq_integral_row_prev, sq_integral_row) =
      sq_integral_slice.split_at_mut(integral_image_stride);
    for (
      src,
      ((integral_above, sq_integral_above), (integral, sq_integral)),
    ) in row.zip(
      integral_row_prev
        .iter()
        .zip(sq_integral_row_prev.iter())
        .zip(integral_row.iter_mut().zip(sq_integral_row.iter_mut())),
    ) {
      let current = u32::cast_from(*src);
      // Wrap adds to prevent undefined behaviour on overflow. Overflow is
      // cancelled out when calculating the sum of a region.
      sum = sum.wrapping_add(current);
      *integral = sum.wrapping_add(*integral_above);
      sq_sum = sq_sum.wrapping_add(current * current);
      *sq_integral = sq_sum.wrapping_add(*sq_integral_above);
    }

    // The current row also contains all future rows. Replacing the slice with
    // it moves down a row.
    integral_slice = integral_row;
    sq_integral_slice = sq_integral_row;
  }
}

pub fn sgrproj_stripe_filter<T: Pixel, U: Pixel>(
  set: u8, xqd: [i8; 2], fi: &FrameInvariants<T>,
  integral_image_buffer: &IntegralImageBuffer, integral_image_stride: usize,
  cdeffed: &PlaneSlice<U>, out: &mut PlaneRegionMut<U>,
) {
  let &Rect { width: stripe_w, height: stripe_h, .. } = out.rect();
  let bdm8 = fi.sequence.bit_depth - 8;
  let mut a_r2: [[u32; IMAGE_WIDTH_MAX + 2]; 2] =
    [[0; IMAGE_WIDTH_MAX + 2]; 2];
  let mut b_r2: [[u32; IMAGE_WIDTH_MAX + 2]; 2] =
    [[0; IMAGE_WIDTH_MAX + 2]; 2];
  let mut f_r2_0: [u32; IMAGE_WIDTH_MAX] = [0; IMAGE_WIDTH_MAX];
  let mut f_r2_1: [u32; IMAGE_WIDTH_MAX] = [0; IMAGE_WIDTH_MAX];
  let mut a_r1: [[u32; IMAGE_WIDTH_MAX + 2]; 3] =
    [[0; IMAGE_WIDTH_MAX + 2]; 3];
  let mut b_r1: [[u32; IMAGE_WIDTH_MAX + 2]; 3] =
    [[0; IMAGE_WIDTH_MAX + 2]; 3];
  let mut f_r1: [u32; IMAGE_WIDTH_MAX] = [0; IMAGE_WIDTH_MAX];

  let s_r2: u32 = SGRPROJ_PARAMS_S[set as usize][0];
  let s_r1: u32 = SGRPROJ_PARAMS_S[set as usize][1];

  /* prime the intermediate arrays */
  // One oddness about the radius=2 intermediate array computations that
  // the spec doesn't make clear: Although the spec defines computation
  // of every row (of a, b and f), only half of the rows (every-other
  // row) are actually used.
  let integral_image = &integral_image_buffer.integral_image;
  let sq_integral_image = &integral_image_buffer.sq_integral_image;
  if s_r2 > 0 {
    sgrproj_box_ab_r2(
      &mut a_r2[0],
      &mut b_r2[0],
      integral_image,
      sq_integral_image,
      integral_image_stride,
      0,
      stripe_w,
      s_r2,
      bdm8,
      fi.cpu_feature_level,
    );
  }
  if s_r1 > 0 {
    let integral_image_offset = integral_image_stride + 1;
    sgrproj_box_ab_r1(
      &mut a_r1[0],
      &mut b_r1[0],
      &integral_image[integral_image_offset..],
      &sq_integral_image[integral_image_offset..],
      integral_image_stride,
      0,
      stripe_w,
      s_r1,
      bdm8,
      fi.cpu_feature_level,
    );
    sgrproj_box_ab_r1(
      &mut a_r1[1],
      &mut b_r1[1],
      &integral_image[integral_image_offset..],
      &sq_integral_image[integral_image_offset..],
      integral_image_stride,
      1,
      stripe_w,
      s_r1,
      bdm8,
      fi.cpu_feature_level,
    );
  }

  /* iterate by row */
  // Increment by two to handle the use of even rows by r=2 and run a nested
  //  loop to handle increments of one.
  for y in (0..stripe_h).step_by(2) {
    // get results to use y and y+1
    let f_r2_ab: [&[u32]; 2] = if s_r2 > 0 {
      sgrproj_box_ab_r2(
        &mut a_r2[(y / 2 + 1) % 2],
        &mut b_r2[(y / 2 + 1) % 2],
        integral_image,
        sq_integral_image,
        integral_image_stride,
        y + 2,
        stripe_w,
        s_r2,
        bdm8,
        fi.cpu_feature_level,
      );
      let ap0: [&[u32]; 2] = [&a_r2[(y / 2) % 2], &a_r2[(y / 2 + 1) % 2]];
      let bp0: [&[u32]; 2] = [&b_r2[(y / 2) % 2], &b_r2[(y / 2 + 1) % 2]];
      sgrproj_box_f_r2(
        &ap0,
        &bp0,
        &mut f_r2_0,
        &mut f_r2_1,
        y,
        stripe_w,
        cdeffed,
        fi.cpu_feature_level,
      );
      [&f_r2_0, &f_r2_1]
    } else {
      sgrproj_box_f_r0(
        &mut f_r2_0,
        y,
        stripe_w,
        cdeffed,
        fi.cpu_feature_level,
      );
      // share results for both rows
      [&f_r2_0, &f_r2_0]
    };
    for dy in 0..(2.min(stripe_h - y)) {
      let y = y + dy;
      if s_r1 > 0 {
        let integral_image_offset = integral_image_stride + 1;
        sgrproj_box_ab_r1(
          &mut a_r1[(y + 2) % 3],
          &mut b_r1[(y + 2) % 3],
          &integral_image[integral_image_offset..],
          &sq_integral_image[integral_image_offset..],
          integral_image_stride,
          y + 2,
          stripe_w,
          s_r1,
          bdm8,
          fi.cpu_feature_level,
        );
        let ap1: [&[u32]; 3] =
          [&a_r1[y % 3], &a_r1[(y + 1) % 3], &a_r1[(y + 2) % 3]];
        let bp1: [&[u32]; 3] =
          [&b_r1[y % 3], &b_r1[(y + 1) % 3], &b_r1[(y + 2) % 3]];
        sgrproj_box_f_r1(
          &ap1,
          &bp1,
          &mut f_r1,
          y,
          stripe_w,
          cdeffed,
          fi.cpu_feature_level,
        );
      } else {
        sgrproj_box_f_r0(
          &mut f_r1,
          y,
          stripe_w,
          cdeffed,
          fi.cpu_feature_level,
        );
      }

      /* apply filter */
      let w0 = xqd[0] as i32;
      let w1 = xqd[1] as i32;
      let w2 = (1 << SGRPROJ_PRJ_BITS) - w0 - w1;

      let line = &cdeffed[y];

      #[inline(always)]
      fn apply_filter<U: Pixel>(
        out: &mut [U], line: &[U], f_r1: &[u32], f_r2_ab: &[u32],
        stripe_w: usize, bit_depth: usize, w0: i32, w1: i32, w2: i32,
      ) {
        let line_it = line[..stripe_w].iter();
        let f_r2_ab_it = f_r2_ab[..stripe_w].iter();
        let f_r1_it = f_r1[..stripe_w].iter();
        let out_it = out[..stripe_w].iter_mut();

        for ((o, &u), (&f_r2_ab, &f_r1)) in
          out_it.zip(line_it).zip(f_r2_ab_it.zip(f_r1_it))
        {
          let u = i32::cast_from(u) << SGRPROJ_RST_BITS;
          let v = w0 * f_r2_ab as i32 + w1 * u + w2 * f_r1 as i32;
          let s = (v + (1 << (SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS) >> 1))
            >> (SGRPROJ_RST_BITS + SGRPROJ_PRJ_BITS);
          *o = U::cast_from(clamp(s, 0, (1 << bit_depth) - 1));
        }
      }

      apply_filter(
        &mut out[y],
        line,
        &f_r1,
        f_r2_ab[dy],
        stripe_w,
        fi.sequence.bit_depth,
        w0,
        w1,
        w2,
      );
    }
  }
}

// Frame inputs below aren't all equal, and will change as work
// continues.  There's no deblocked reconstruction available at this
// point of RDO, so we use the non-deblocked reconstruction, cdef and
// input.  The input can be a full-sized frame. Cdef input is a partial
// frame constructed specifically for RDO.

// For simplicity, this ignores stripe segmentation (it's possible the
// extra complexity isn't worth it and we'll ignore stripes
// permanently during RDO, but that's not been tested yet). Data
// access inside the cdef frame is monolithic and clipped to the cdef
// borders.

// Input params follow the same rules as sgrproj_stripe_filter.
// Inputs are relative to the colocated slice views.
pub fn sgrproj_solve<T: Pixel>(
  set: u8, fi: &FrameInvariants<T>,
  integral_image_buffer: &IntegralImageBuffer, input: &PlaneSlice<u16>,
  cdeffed: &PlaneSlice<u16>, cdef_w: usize, cdef_h: usize,
) -> (i8, i8) {
  let bdm8 = fi.sequence.bit_depth - 8;

  let mut a_r2: [[u32; IMAGE_WIDTH_MAX + 2]; 2] =
    [[0; IMAGE_WIDTH_MAX + 2]; 2];
  let mut b_r2: [[u32; IMAGE_WIDTH_MAX + 2]; 2] =
    [[0; IMAGE_WIDTH_MAX + 2]; 2];
  let mut f_r2_0: [u32; IMAGE_WIDTH_MAX] = [0; IMAGE_WIDTH_MAX];
  let mut f_r2_1: [u32; IMAGE_WIDTH_MAX] = [0; IMAGE_WIDTH_MAX];
  let mut a_r1: [[u32; IMAGE_WIDTH_MAX + 2]; 3] =
    [[0; IMAGE_WIDTH_MAX + 2]; 3];
  let mut b_r1: [[u32; IMAGE_WIDTH_MAX + 2]; 3] =
    [[0; IMAGE_WIDTH_MAX + 2]; 3];
  let mut f_r1: [u32; IMAGE_WIDTH_MAX] = [0; IMAGE_WIDTH_MAX];

  let s_r2: u32 = SGRPROJ_PARAMS_S[set as usize][0];
  let s_r1: u32 = SGRPROJ_PARAMS_S[set as usize][1];

  let mut h: [[f64; 2]; 2] = [[0., 0.], [0., 0.]];
  let mut c: [f64; 2] = [0., 0.];

  /* prime the intermediate arrays */
  // One oddness about the radius=2 intermediate array computations that
  // the spec doesn't make clear: Although the spec defines computation
  // of every row (of a, b and f), only half of the rows (every-other
  // row) are actually used.
  let integral_image = &integral_image_buffer.integral_image;
  let sq_integral_image = &integral_image_buffer.sq_integral_image;
  if s_r2 > 0 {
    sgrproj_box_ab_r2(
      &mut a_r2[0],
      &mut b_r2[0],
      integral_image,
      sq_integral_image,
      SOLVE_IMAGE_STRIDE,
      0,
      cdef_w,
      s_r2,
      bdm8,
      fi.cpu_feature_level,
    );
  }
  if s_r1 > 0 {
    let integral_image_offset = SOLVE_IMAGE_STRIDE + 1;
    sgrproj_box_ab_r1(
      &mut a_r1[0],
      &mut b_r1[0],
      &integral_image[integral_image_offset..],
      &sq_integral_image[integral_image_offset..],
      SOLVE_IMAGE_STRIDE,
      0,
      cdef_w,
      s_r1,
      bdm8,
      fi.cpu_feature_level,
    );
    sgrproj_box_ab_r1(
      &mut a_r1[1],
      &mut b_r1[1],
      &integral_image[integral_image_offset..],
      &sq_integral_image[integral_image_offset..],
      SOLVE_IMAGE_STRIDE,
      1,
      cdef_w,
      s_r1,
      bdm8,
      fi.cpu_feature_level,
    );
  }

  /* iterate by row */
  // Increment by two to handle the use of even rows by r=2 and run a nested
  //  loop to handle increments of one.
  for y in (0..cdef_h).step_by(2) {
    // get results to use y and y+1
    let f_r2_01: [&[u32]; 2] = if s_r2 > 0 {
      sgrproj_box_ab_r2(
        &mut a_r2[(y / 2 + 1) % 2],
        &mut b_r2[(y / 2 + 1) % 2],
        integral_image,
        sq_integral_image,
        SOLVE_IMAGE_STRIDE,
        y + 2,
        cdef_w,
        s_r2,
        bdm8,
        fi.cpu_feature_level,
      );
      let ap0: [&[u32]; 2] = [&a_r2[(y / 2) % 2], &a_r2[(y / 2 + 1) % 2]];
      let bp0: [&[u32]; 2] = [&b_r2[(y / 2) % 2], &b_r2[(y / 2 + 1) % 2]];
      sgrproj_box_f_r2(
        &ap0,
        &bp0,
        &mut f_r2_0,
        &mut f_r2_1,
        y,
        cdef_w,
        cdeffed,
        fi.cpu_feature_level,
      );
      [&f_r2_0, &f_r2_1]
    } else {
      sgrproj_box_f_r0(&mut f_r2_0, y, cdef_w, cdeffed, fi.cpu_feature_level);
      // share results for both rows
      [&f_r2_0, &f_r2_0]
    };
    for dy in 0..(2.min(cdef_h - y)) {
      let y = y + dy;
      if s_r1 > 0 {
        let integral_image_offset = SOLVE_IMAGE_STRIDE + 1;
        sgrproj_box_ab_r1(
          &mut a_r1[(y + 2) % 3],
          &mut b_r1[(y + 2) % 3],
          &integral_image[integral_image_offset..],
          &sq_integral_image[integral_image_offset..],
          SOLVE_IMAGE_STRIDE,
          y + 2,
          cdef_w,
          s_r1,
          bdm8,
          fi.cpu_feature_level,
        );
        let ap1: [&[u32]; 3] =
          [&a_r1[y % 3], &a_r1[(y + 1) % 3], &a_r1[(y + 2) % 3]];
        let bp1: [&[u32]; 3] =
          [&b_r1[y % 3], &b_r1[(y + 1) % 3], &b_r1[(y + 2) % 3]];
        sgrproj_box_f_r1(
          &ap1,
          &bp1,
          &mut f_r1,
          y,
          cdef_w,
          cdeffed,
          fi.cpu_feature_level,
        );
      } else {
        sgrproj_box_f_r0(&mut f_r1, y, cdef_w, cdeffed, fi.cpu_feature_level);
      }

      #[inline(always)]
      fn process_line<T: Pixel>(
        h: &mut [[f64; 2]; 2], c: &mut [f64; 2], cdeffed: &[T], input: &[T],
        f_r1: &[u32], f_r2_ab: &[u32], cdef_w: usize,
      ) {
        let cdeffed_it = cdeffed[..cdef_w].iter();
        let input_it = input[..cdef_w].iter();
        let f_r2_ab_it = f_r2_ab[..cdef_w].iter();
        let f_r1_it = f_r1[..cdef_w].iter();

        #[derive(Debug, Copy, Clone)]
        struct Sums {
          h: [[i64; 2]; 2],
          c: [i64; 2],
        }

        let sums: Sums = cdeffed_it
          .zip(input_it)
          .zip(f_r2_ab_it.zip(f_r1_it))
          .map(|((&u, &i), (&f2, &f1))| {
            let u = i32::cast_from(u) << SGRPROJ_RST_BITS;
            let s = (i32::cast_from(i) << SGRPROJ_RST_BITS) - u;
            let f2 = f2 as i32 - u;
            let f1 = f1 as i32 - u;
            (s as i64, f1 as i64, f2 as i64)
          })
          .fold(Sums { h: [[0; 2]; 2], c: [0; 2] }, |sums, (s, f1, f2)| {
            let mut ret: Sums = sums;
            ret.h[0][0] += f2 * f2;
            ret.h[1][1] += f1 * f1;
            ret.h[0][1] += f1 * f2;
            ret.c[0] += f2 * s;
            ret.c[1] += f1 * s;
            ret
          });

        h[0][0] += sums.h[0][0] as f64;
        h[1][1] += sums.h[1][1] as f64;
        h[0][1] += sums.h[0][1] as f64;
        c[0] += sums.c[0] as f64;
        c[1] += sums.c[1] as f64;
      }

      process_line(
        &mut h,
        &mut c,
        &cdeffed[y],
        &input[y],
        &f_r1,
        f_r2_01[dy],
        cdef_w,
      );
    }
  }

  // this is lifted almost in-tact from libaom
  let n = cdef_w as f64 * cdef_h as f64;
  h[0][0] /= n;
  h[0][1] /= n;
  h[1][1] /= n;
  h[1][0] = h[0][1];
  c[0] *= (1 << SGRPROJ_PRJ_BITS) as f64 / n;
  c[1] *= (1 << SGRPROJ_PRJ_BITS) as f64 / n;
  let (xq0, xq1) = if s_r2 == 0 {
    // H matrix is now only the scalar h[1][1]
    // C vector is now only the scalar c[1]
    if h[1][1] == 0. {
      (0, 0)
    } else {
      (0, (c[1] / h[1][1]).round() as i32)
    }
  } else if s_r1 == 0 {
    // H matrix is now only the scalar h[0][0]
    // C vector is now only the scalar c[0]
    if h[0][0] == 0. {
      (0, 0)
    } else {
      ((c[0] / h[0][0]).round() as i32, 0)
    }
  } else {
    let det = h[0][0] * h[1][1] - h[0][1] * h[1][0];
    if det == 0. {
      (0, 0)
    } else {
      // If scaling up dividend would overflow, instead scale down the divisor
      let div1 = h[1][1] * c[0] - h[0][1] * c[1];
      let div2 = h[0][0] * c[1] - h[1][0] * c[0];
      ((div1 / det).round() as i32, (div2 / det).round() as i32)
    }
  };
  {
    let xqd0 =
      clamp(xq0, SGRPROJ_XQD_MIN[0] as i32, SGRPROJ_XQD_MAX[0] as i32);
    let xqd1 = clamp(
      (1 << SGRPROJ_PRJ_BITS) - xqd0 - xq1,
      SGRPROJ_XQD_MIN[1] as i32,
      SGRPROJ_XQD_MAX[1] as i32,
    );
    (xqd0 as i8, xqd1 as i8)
  }
}

fn wiener_stripe_filter<T: Pixel>(
  coeffs: [[i8; 3]; 2], fi: &FrameInvariants<T>, crop_w: usize, crop_h: usize,
  stripe_w: usize, stripe_h: usize, stripe_x: usize, stripe_y: isize,
  cdeffed: &Plane<T>, deblocked: &Plane<T>, out: &mut Plane<T>,
) {
  let bit_depth = fi.sequence.bit_depth;
  let round_h = if bit_depth == 12 { 5 } else { 3 };
  let round_v = if bit_depth == 12 { 9 } else { 11 };
  let offset = 1 << (bit_depth + WIENER_BITS - round_h - 1);
  let limit = (1 << (bit_depth + 1 + WIENER_BITS - round_h)) - 1;

  let mut coeffs_ = [[0; 3]; 2];
  for i in 0..2 {
    for j in 0..3 {
      coeffs_[i][j] = i32::from(coeffs[i][j]);
    }
  }

  let mut work: [i32; SB_SIZE + 7] = [0; SB_SIZE + 7];
  let vfilter: [i32; 7] = [
    coeffs_[0][0],
    coeffs_[0][1],
    coeffs_[0][2],
    128 - 2 * (coeffs_[0][0] + coeffs_[0][1] + coeffs_[0][2]),
    coeffs_[0][2],
    coeffs_[0][1],
    coeffs_[0][0],
  ];
  let hfilter: [i32; 7] = [
    coeffs_[1][0],
    coeffs_[1][1],
    coeffs_[1][2],
    128 - 2 * (coeffs_[1][0] + coeffs_[1][1] + coeffs_[1][2]),
    coeffs_[1][2],
    coeffs_[1][1],
    coeffs_[1][0],
  ];

  // unlike x, our y can be negative to start as the first stripe
  // starts off the top of the frame by 8 pixels, and can also run off the end of the frame
  let start_wi = if stripe_y < 0 { -stripe_y } else { 0 } as usize;
  let start_yi = if stripe_y < 0 { 0 } else { stripe_y } as usize;
  let end_i = cmp::max(
    0,
    if stripe_h as isize + stripe_y > crop_h as isize {
      crop_h as isize - stripe_y - start_wi as isize
    } else {
      stripe_h as isize - start_wi as isize
    },
  ) as usize;

  let mut out_slice =
    out.mut_slice(PlaneOffset { x: 0, y: start_yi as isize });

  for xi in stripe_x..stripe_x + stripe_w {
    let n = cmp::min(7, crop_w as isize + 3 - xi as isize);
    for yi in stripe_y - 3..stripe_y + stripe_h as isize + 4 {
      let src_plane: &Plane<T>;
      let mut acc = 0;
      let ly;
      if yi < stripe_y {
        ly =
          cmp::max(clamp(yi, 0, crop_h as isize - 1), stripe_y - 2) as usize;
        src_plane = deblocked;
      } else if yi < stripe_y + stripe_h as isize {
        ly = clamp(yi, 0, crop_h as isize - 1) as usize;
        src_plane = cdeffed;
      } else {
        ly = cmp::min(
          clamp(yi, 0, crop_h as isize - 1),
          stripe_y + stripe_h as isize + 1,
        ) as usize;
        src_plane = deblocked;
      }

      for i in 0..3 - xi as isize {
        acc += hfilter[i as usize] * i32::cast_from(src_plane.p(0, ly));
      }
      for i in cmp::max(0, 3 - (xi as isize))..n {
        acc += hfilter[i as usize]
          * i32::cast_from(src_plane.p((xi as isize + i - 3) as usize, ly));
      }
      for i in n..7 {
        acc +=
          hfilter[i as usize] * i32::cast_from(src_plane.p(crop_w - 1, ly));
      }

      acc = (acc + (1 << round_h >> 1)) >> round_h;
      work[(yi - stripe_y + 3) as usize] = clamp(acc, -offset, limit - offset);
    }

    for (wi, dst) in (start_wi..start_wi + end_i)
      .zip(out_slice.rows_iter_mut().map(|row| &mut row[xi]).take(end_i))
    {
      let mut acc = 0;
      for (i, src) in (0..7).zip(work[wi..wi + 7].iter_mut()) {
        acc += vfilter[i] * *src;
      }
      *dst = T::cast_from(clamp(
        (acc + (1 << round_v >> 1)) >> round_v,
        0,
        (1 << bit_depth) - 1,
      ));
    }
  }
}

#[derive(Copy, Clone, Debug)]
pub struct RestorationUnit {
  pub filter: RestorationFilter,
}

impl Default for RestorationUnit {
  fn default() -> RestorationUnit {
    RestorationUnit { filter: RestorationFilter::default() }
  }
}

#[derive(Clone, Debug)]
pub struct FrameRestorationUnits {
  units: Box<[RestorationUnit]>,
  pub cols: usize,
  pub rows: usize,
}

impl FrameRestorationUnits {
  pub fn new(cols: usize, rows: usize) -> Self {
    Self {
      units: vec![RestorationUnit::default(); cols * rows].into_boxed_slice(),
      cols,
      rows,
    }
  }
}

impl Index<usize> for FrameRestorationUnits {
  type Output = [RestorationUnit];
  #[inline(always)]
  fn index(&self, index: usize) -> &Self::Output {
    &self.units[index * self.cols..(index + 1) * self.cols]
  }
}

impl IndexMut<usize> for FrameRestorationUnits {
  #[inline(always)]
  fn index_mut(&mut self, index: usize) -> &mut Self::Output {
    &mut self.units[index * self.cols..(index + 1) * self.cols]
  }
}

#[derive(Clone, Debug)]
pub struct RestorationPlaneConfig {
  pub lrf_type: u8,
  pub unit_size: usize,
  // (1 << sb_x_shift) gives the number of superblocks horizontally or
  // vertically in a restoration unit, not accounting for RU stretching
  pub sb_h_shift: usize,
  pub sb_v_shift: usize,
  pub sb_cols: usize, // actual number of SB cols in this LRU (accounting for stretch and crop)
  pub sb_rows: usize, // actual number of SB rows in this LRU (accounting for stretch and crop)
  // stripe height is 64 in all cases except 4:2:0 chroma planes where
  // it is 32.  This is independent of all other setup parameters
  pub stripe_height: usize,
  pub cols: usize,
  pub rows: usize,
}

#[derive(Clone, Debug)]
pub struct RestorationPlane {
  pub cfg: RestorationPlaneConfig,
  pub units: FrameRestorationUnits,
}

#[derive(Clone, Default)]
pub struct RestorationPlaneOffset {
  pub row: usize,
  pub col: usize,
}

impl RestorationPlane {
  pub fn new(
    lrf_type: u8, unit_size: usize, sb_h_shift: usize, sb_v_shift: usize,
    sb_cols: usize, sb_rows: usize, stripe_decimate: usize, cols: usize,
    rows: usize,
  ) -> RestorationPlane {
    let stripe_height = if stripe_decimate != 0 { 32 } else { 64 };
    RestorationPlane {
      cfg: RestorationPlaneConfig {
        lrf_type,
        unit_size,
        sb_h_shift,
        sb_v_shift,
        sb_cols,
        sb_rows,
        stripe_height,
        cols,
        rows,
      },
      units: FrameRestorationUnits::new(cols, rows),
    }
  }

  // Stripes are always 64 pixels high in a non-subsampled
  // frame, and decimated from 64 pixels in chroma.  When
  // filtering, they are not co-located on Y with superblocks.
  fn restoration_unit_index_by_stripe(
    &self, stripenum: usize, rux: usize,
  ) -> (usize, usize) {
    (
      cmp::min(rux, self.cfg.cols - 1),
      cmp::min(
        stripenum * self.cfg.stripe_height / self.cfg.unit_size,
        self.cfg.rows - 1,
      ),
    )
  }

  pub fn restoration_unit_by_stripe(
    &self, stripenum: usize, rux: usize,
  ) -> &RestorationUnit {
    let (x, y) = self.restoration_unit_index_by_stripe(stripenum, rux);
    &self.units[y][x]
  }
}

#[derive(Clone, Debug)]
pub struct RestorationState {
  pub planes: [RestorationPlane; PLANES],
}

impl RestorationState {
  pub fn new<T: Pixel>(fi: &FrameInvariants<T>, input: &Frame<T>) -> Self {
    let PlaneConfig { xdec, ydec, .. } = input.planes[1].cfg;
    // stripe size is decimated in 4:2:0 (and only 4:2:0)
    let stripe_uv_decimate = if xdec > 0 && ydec > 0 { 1 } else { 0 };
    let y_sb_log2 = if fi.sequence.use_128x128_superblock { 7 } else { 6 };
    let uv_sb_h_log2 = y_sb_log2 - xdec;
    let uv_sb_v_log2 = y_sb_log2 - ydec;

    let (lrf_y_shift, lrf_uv_shift) = if fi.sequence.enable_large_lru
      && fi.sequence.enable_restoration
    {
      assert!(
        fi.width > 1 && fi.height > 1,
        "Width and height must be higher than 1 for LRF setup"
      );

      // Specific content does affect optimal LRU size choice, but the
      // quantizer in use is a surprisingly strong selector.
      let lrf_base_shift = if fi.base_q_idx > 200 {
        0 // big
      } else if fi.base_q_idx > 160 {
        1
      } else {
        2 // small
      };
      let lrf_chroma_shift = if stripe_uv_decimate > 0 {
        // 4:2:0 only
        if lrf_base_shift == 2 {
          1 // smallest chroma LRU is a win at low quant
        } else {
          // Will a down-shifted chroma LRU eliminate stretch in chroma?
          // If so, that's generally a win.
          let lrf_unit_size =
            1 << (RESTORATION_TILESIZE_MAX_LOG2 - lrf_base_shift);
          let unshifted_stretch = ((fi.width >> xdec) - 1) % lrf_unit_size
            <= lrf_unit_size / 2
            || ((fi.height >> ydec) - 1) % lrf_unit_size <= lrf_unit_size / 2;
          let shifted_stretch = ((fi.width >> xdec) - 1)
            % (lrf_unit_size >> 1)
            <= lrf_unit_size / 4
            || ((fi.height >> ydec) - 1) % (lrf_unit_size >> 1)
              <= lrf_unit_size / 4;
          if unshifted_stretch && !shifted_stretch {
            1 // shift to eliminate stretch
          } else {
            0 // don't shift; save the signaling bits
          }
        }
      } else {
        0
      };
      (lrf_base_shift, lrf_base_shift + lrf_chroma_shift)
    } else {
      // Explicit request to tie LRU size to superblock size ==
      // smallest possible LRU size
      let lrf_y_shift = if fi.sequence.use_128x128_superblock { 1 } else { 2 };
      (lrf_y_shift, lrf_y_shift + stripe_uv_decimate)
    };

    let mut y_unit_size = 1 << (RESTORATION_TILESIZE_MAX_LOG2 - lrf_y_shift);
    let mut uv_unit_size = 1 << (RESTORATION_TILESIZE_MAX_LOG2 - lrf_uv_shift);

    // Right now we defer to tiling setup: don't choose an LRU size
    // large enough that a tile is not an integer number of LRUs
    // wide/high.
    if fi.tiling.cols > 1 || fi.tiling.rows > 1 {
      // despite suggestions to the contrary, apparently tiles can be
      // non-powers-of-2.
      let trailing_h_zeros = fi.tiling.tile_width_sb.trailing_zeros() as usize;
      let trailing_v_zeros =
        fi.tiling.tile_height_sb.trailing_zeros() as usize;
      let tile_aligned_y_unit_size =
        1 << (y_sb_log2 + trailing_h_zeros.min(trailing_v_zeros));
      let tile_aligned_uv_h_unit_size = 1 << (uv_sb_h_log2 + trailing_h_zeros);
      let tile_aligned_uv_v_unit_size = 1 << (uv_sb_v_log2 + trailing_v_zeros);
      y_unit_size = y_unit_size.min(tile_aligned_y_unit_size);
      uv_unit_size = uv_unit_size
        .min(tile_aligned_uv_h_unit_size.min(tile_aligned_uv_v_unit_size));

      // But it's actually worse: LRUs can't span tiles (in our design
      // that is, spec allows it).  However, the spec mandates the last
      // LRU stretches forward into any less-than-half-LRU span of
      // superblocks at the right and bottom of a frame.  These
      // superblocks may well be in a different tile!  Even if LRUs are
      // minimum size (one superblock), when the right or bottom edge of
      // the frame is a superblock that's less than half the
      // width/height of a normal superblock, the LRU is forced by the
      // spec to span into it (and thus a different tile).  Tiling is
      // under no such restriction; it could decide the right/left
      // sliver will be in its own tile row/column.  We can't disallow
      // the combination here.  The tiling code will have to either
      // prevent it or tolerate it.  (prayer mechanic == Issue #1629).
    }

    // derive the rest
    let y_unit_log2 = y_unit_size.ilog() - 1;
    let uv_unit_log2 = uv_unit_size.ilog() - 1;
    let y_cols = ((fi.width + (y_unit_size >> 1)) / y_unit_size).max(1);
    let y_rows = ((fi.height + (y_unit_size >> 1)) / y_unit_size).max(1);
    let uv_cols = ((((fi.width + (1 << xdec >> 1)) >> xdec)
      + (uv_unit_size >> 1))
      / uv_unit_size)
      .max(1);
    let uv_rows = ((((fi.height + (1 << ydec >> 1)) >> ydec)
      + (uv_unit_size >> 1))
      / uv_unit_size)
      .max(1);

    RestorationState {
      planes: [
        RestorationPlane::new(
          RESTORE_SWITCHABLE,
          y_unit_size,
          y_unit_log2 - y_sb_log2,
          y_unit_log2 - y_sb_log2,
          fi.sb_width,
          fi.sb_height,
          0,
          y_cols,
          y_rows,
        ),
        RestorationPlane::new(
          RESTORE_SWITCHABLE,
          uv_unit_size,
          uv_unit_log2 - uv_sb_h_log2,
          uv_unit_log2 - uv_sb_v_log2,
          fi.sb_width,
          fi.sb_height,
          stripe_uv_decimate,
          uv_cols,
          uv_rows,
        ),
        RestorationPlane::new(
          RESTORE_SWITCHABLE,
          uv_unit_size,
          uv_unit_log2 - uv_sb_h_log2,
          uv_unit_log2 - uv_sb_v_log2,
          fi.sb_width,
          fi.sb_height,
          stripe_uv_decimate,
          uv_cols,
          uv_rows,
        ),
      ],
    }
  }

  #[hawktracer(lrf_filter_frame)]
  pub fn lrf_filter_frame<T: Pixel>(
    &mut self, out: &mut Frame<T>, pre_cdef: &Frame<T>,
    fi: &FrameInvariants<T>,
  ) {
    let cdeffed = out.clone();

    // unlike the other loop filters that operate over the padded
    // frame dimensions, restoration filtering and source pixel
    // accesses are clipped to the original frame dimensions
    // that's why we use fi.width and fi.height instead of PlaneConfig fields

    // number of stripes (counted according to colocated Y luma position)
    let stripe_n = (fi.height + 7) / 64 + 1;

    // Buffers for the stripe filter.
    let mut stripe_filter_buffer =
      IntegralImageBuffer::zeroed(STRIPE_IMAGE_SIZE);

    for pli in 0..PLANES {
      let rp = &self.planes[pli];
      let xdec = out.planes[pli].cfg.xdec;
      let ydec = out.planes[pli].cfg.ydec;
      let crop_w = (fi.width + (1 << xdec >> 1)) >> xdec;
      let crop_h = (fi.height + (1 << ydec >> 1)) >> ydec;

      for si in 0..stripe_n {
        let (stripe_start_y, stripe_size) = if si == 0 {
          (0, (64 - 8) >> ydec)
        } else {
          let start = (si * 64 - 8) >> ydec;
          (
            start as isize,
            // one past, unlike spec
            (64 >> ydec).min(crop_h - start),
          )
        };

        // horizontally, go rdu-by-rdu
        for rux in 0..rp.cfg.cols {
          // stripe x pixel locations must be clipped to frame, last may need to stretch
          let x = rux * rp.cfg.unit_size;
          let size =
            if rux == rp.cfg.cols - 1 { crop_w - x } else { rp.cfg.unit_size };
          let ru = rp.restoration_unit_by_stripe(si, rux);
          match ru.filter {
            RestorationFilter::Wiener { coeffs } => {
              wiener_stripe_filter(
                coeffs,
                fi,
                crop_w,
                crop_h,
                size,
                stripe_size,
                x,
                stripe_start_y,
                &cdeffed.planes[pli],
                &pre_cdef.planes[pli],
                &mut out.planes[pli],
              );
            }
            RestorationFilter::Sgrproj { set, xqd } => {
              if !fi.sequence.enable_cdef {
                continue;
              }

              setup_integral_image(
                &mut stripe_filter_buffer,
                STRIPE_IMAGE_STRIDE,
                crop_w - x,
                (crop_h as isize - stripe_start_y) as usize,
                size,
                stripe_size,
                &cdeffed.planes[pli]
                  .slice(PlaneOffset { x: x as isize, y: stripe_start_y }),
                &pre_cdef.planes[pli]
                  .slice(PlaneOffset { x: x as isize, y: stripe_start_y }),
              );

              sgrproj_stripe_filter(
                set,
                xqd,
                fi,
                &stripe_filter_buffer,
                STRIPE_IMAGE_STRIDE,
                &cdeffed.planes[pli]
                  .slice(PlaneOffset { x: x as isize, y: stripe_start_y }),
                &mut out.planes[pli].region_mut(Area::Rect {
                  x: x as isize,
                  y: stripe_start_y,
                  width: size,
                  height: stripe_size,
                }),
              );
            }
            RestorationFilter::None => {
              // do nothing
            }
          }
        }
      }
    }
  }
}