refactor(av1): split inverse transform into stages and passes (S3776)

* refactor(h264): split dpb functions (S3776)

* fix(h264): restore the MMCO 6 long-term index range check

crates/av1/src/decoder.rs

@@ -405,49 +405,35 @@ impl Decoder {
         r
     }
 
+    /// Whether an OBU belongs to the selected operating point (spec 7.5: OBUs of other layers are dropped).
+    fn in_operating_point(&self, obu: &ObuUnit) -> bool {
+        if obu.obu_type != OBU_SEQUENCE_HEADER_T
+            && obu.obu_type != OBU_TEMPORAL_DELIMITER_T
+            && obu.ext
+            && let Some(seq) = &self.seq
+            && seq.op_idc != 0
+        {
+            let idc = seq.op_idc;
+            let in_t = (idc >> obu.temporal_id) & 1;
+            let in_s = (idc >> (obu.spatial_id + 8)) & 1;
+            if in_t == 0 || in_s == 0 {
+                return false;
+            }
+        }
+        true
+    }
+
     fn parse_obus(&mut self, data: &[u8], out: &mut Vec<(Output, usize, u32)>) -> Result<()> {
         let mut tu = 0usize;
         let mut pos = 0;
         while pos < data.len() {
-            let h = data[pos];
-            if h & 0x80 != 0 {
-                return Err(Error::Invalid("obu_forbidden_bit"));
+            let obu = parse_obu_unit(data, pos)?;
+            pos = obu.end;
+            if !self.in_operating_point(&obu) {
+                continue;
             }
-            let obu_type = ((h >> 3) & 0xf) as u32;
-            let ext = h & 4 != 0;
-            let has_size = h & 2 != 0;
-            let mut p = pos + 1;
-            let (mut temporal_id, mut spatial_id) = (0, 0);
-            if ext {
-                let e = *data.get(p).ok_or(Error::Truncated)?;
-                temporal_id = (e >> 5) as u32;
-                spatial_id = ((e >> 3) & 3) as u32;
-                p += 1;
-            }
-            let size = if has_size {
-                let (v, n) = leb128(data.get(p..).ok_or(Error::Truncated)?)?;
-                p += n;
-                v as usize
-            } else {
-                data.len().checked_sub(p).ok_or(Error::Truncated)?
-            };
-            let end = p.checked_add(size).filter(|&e| e <= data.len()).ok_or(Error::Truncated)?;
-            let payload = &data[p..end];
-            pos = end;
-            if obu_type != OBU_SEQUENCE_HEADER_T
-                && obu_type != OBU_TEMPORAL_DELIMITER_T
-                && ext
-                && let Some(seq) = &self.seq
-                && seq.op_idc != 0
-            {
-                let idc = seq.op_idc;
-                let in_t = (idc >> temporal_id) & 1;
-                let in_s = (idc >> (spatial_id + 8)) & 1;
-                if in_t == 0 || in_s == 0 {
-                    continue;
-                }
-            }
-            match obu_type {
+            let (payload, spatial_id) = (obu.payload, obu.spatial_id);
+            match obu.obu_type {
                 OBU_SEQUENCE_HEADER_T => {
                     let s = SequenceHeader::parse(payload)?;
                     self.seq = Some(s);
@@ -457,27 +443,11 @@ impl Decoder {
                     tu += 1;
                 }
                 OBU_FRAME_HEADER_T | OBU_REDUNDANT_FRAME_HEADER_T | OBU_FRAME_T => {
-                    if self.seen_frame_header {
-                        // frame_header_copy(): identical to the active header
-                        if obu_type != OBU_FRAME_T {
-                            continue;
-                        }
-                    }
-                    let mut r = BitReader::new(payload);
-                    let shown = self.frame_header(&mut r, temporal_id, spatial_id)?;
-                    if let Some(o) = shown {
+                    if let Some(o) = self.frame_obu(&obu)? {
                         if out.iter().filter(|(o, _, _)| matches!(o, Output::Existing { .. })).count() >= 32 {
                             return Err(Error::Invalid("too many frame outputs in chunk"));
                         }
                         out.push((o, tu, spatial_id));
-                        continue;
-                    }
-                    if obu_type == OBU_FRAME_T {
-                        r.byte_align();
-                        let rest = payload.get(r.byte_pos()..).ok_or(Error::Truncated)?;
-                        if let Some(o) = self.tile_group(rest)? {
-                            out.push((o, tu, spatial_id));
-                        }
                     }
                 }
                 OBU_TILE_GROUP_T => {
@@ -492,6 +462,22 @@ impl Decoder {
         Ok(())
     }
 
+    /// A frame header or frame OBU: returns the picture it completes, if any.
+    fn frame_obu(&mut self, obu: &ObuUnit) -> Result<Option<Output>> {
+        // frame_header_copy(): identical to the active header
+        if self.seen_frame_header && obu.obu_type != OBU_FRAME_T {
+            return Ok(None);
+        }
+        let mut r = BitReader::new(obu.payload);
+        let shown = self.frame_header(&mut r, obu.temporal_id, obu.spatial_id)?;
+        if shown.is_some() || obu.obu_type != OBU_FRAME_T {
+            return Ok(shown);
+        }
+        r.byte_align();
+        let rest = obu.payload.get(r.byte_pos()..).ok_or(Error::Truncated)?;
+        self.tile_group(rest)
+    }
+
     /// frame_header_obu( ): returns the output for show_existing_frame.
     fn frame_header(&mut self, r: &mut BitReader, temporal_id: u32, spatial_id: u32) -> Result<Option<Output>> {
         let seq = self.seq.clone().ok_or(Error::Invalid("frame before sequence header"))?;
@@ -1017,6 +1003,43 @@ mod tests {
     }
 }
 
+/// One OBU: its header fields, payload and the position just past it.
+struct ObuUnit<'a> {
+    obu_type: u32,
+    ext: bool,
+    temporal_id: u32,
+    spatial_id: u32,
+    payload: &'a [u8],
+    end: usize,
+}
+
+fn parse_obu_unit(data: &[u8], pos: usize) -> Result<ObuUnit<'_>> {
+    let h = data[pos];
+    if h & 0x80 != 0 {
+        return Err(Error::Invalid("obu_forbidden_bit"));
+    }
+    let obu_type = ((h >> 3) & 0xf) as u32;
+    let ext = h & 4 != 0;
+    let has_size = h & 2 != 0;
+    let mut p = pos + 1;
+    let (mut temporal_id, mut spatial_id) = (0, 0);
+    if ext {
+        let e = *data.get(p).ok_or(Error::Truncated)?;
+        temporal_id = (e >> 5) as u32;
+        spatial_id = ((e >> 3) & 3) as u32;
+        p += 1;
+    }
+    let size = if has_size {
+        let (v, n) = leb128(data.get(p..).ok_or(Error::Truncated)?)?;
+        p += n;
+        v as usize
+    } else {
+        data.len().checked_sub(p).ok_or(Error::Truncated)?
+    };
+    let end = p.checked_add(size).filter(|&e| e <= data.len()).ok_or(Error::Truncated)?;
+    Ok(ObuUnit { obu_type, ext, temporal_id, spatial_id, payload: &data[p..end], end })
+}
+
 #[cfg(test)]
 mod guard_tests {
     use super::*;

crates/av1/src/grain.rs

@@ -55,21 +55,79 @@ impl Rng {
     }
 }
 
-pub(crate) fn apply(pic: &mut Picture, g: &FilmGrainParams, seq: &SequenceHeader) {
-    let _ = seq;
-    let bd = pic.bit_depth as u32;
-    let mono = pic.mono_chrome;
-    let num_planes = if mono { 1 } else { 3 };
-    let (sub_x, sub_y) = (pic.subsampling_x as usize, pic.subsampling_y as usize);
-    let w = pic.width as usize;
-    let h = pic.height as usize;
-    let grain_center = 128i32 << (bd - 8);
-    let grain_min = -grain_center;
-    let grain_max = (256i32 << (bd - 8)) - 1 - grain_center;
+/// Picture-level parameters shared by the grain synthesis stages.
+struct GrainCtx<'a> {
+    g: &'a FilmGrainParams,
+    bd: u32,
+    mono: bool,
+    sub_x: usize,
+    sub_y: usize,
+    w: usize,
+    h: usize,
+    grain_min: i32,
+    grain_max: i32,
+}
 
-    // Generate grain (7.18.3.3).
-    let mut rng = Rng(g.grain_seed);
-    let shift = 12 - bd + g.grain_scale_shift;
+impl GrainCtx<'_> {
+    fn num_planes(&self) -> usize {
+        if self.mono { 1 } else { 3 }
+    }
+
+    fn grain_shift(&self) -> u32 {
+        12 - self.bd + self.g.grain_scale_shift
+    }
+
+    fn ar_shift(&self) -> u32 {
+        self.g.ar_coeff_shift_minus_6 + 6
+    }
+
+    fn lag(&self) -> i32 {
+        self.g.ar_coeff_lag as i32
+    }
+
+    fn stripe_w(&self) -> usize {
+        self.w + 64
+    }
+}
+
+type GrainTemplate = Vec<[i32; 82]>;
+
+/// The luma, Cb and Cr grain templates (7.18.3.3).
+struct Templates {
+    luma: GrainTemplate,
+    cb: GrainTemplate,
+    cr: GrainTemplate,
+}
+
+impl Templates {
+    fn plane(&self, plane: usize) -> &GrainTemplate {
+        match plane {
+            0 => &self.luma,
+            1 => &self.cb,
+            _ => &self.cr,
+        }
+    }
+}
+
+fn luma_ar_sum(g: &FilmGrainParams, luma: &[[i32; 82]], lag: i32, y: usize, x: usize) -> i32 {
+    let mut s = 0;
+    let mut pos = 0;
+    'outer: for dr in -lag..=0 {
+        for dc in -lag..=lag {
+            if dr == 0 && dc == 0 {
+                break 'outer;
+            }
+            let c = g.ar_coeffs_y_plus_128[pos] as i32 - 128;
+            s += luma[(y as i32 + dr) as usize][(x as i32 + dc) as usize] * c;
+            pos += 1;
+        }
+    }
+    s
+}
+
+fn luma_grain(c: &GrainCtx, rng: &mut Rng) -> GrainTemplate {
+    let g = c.g;
+    let shift = c.grain_shift();
     let mut luma = vec![[0i32; 82]; 73];
     for row in luma.iter_mut() {
         for v in row.iter_mut() {
@@ -77,77 +135,89 @@ pub(crate) fn apply(pic: &mut Picture, g: &FilmGrainParams, seq: &SequenceHeader
             *v = round2(gv, shift);
         }
     }
-    let ar_shift = g.ar_coeff_shift_minus_6 + 6;
-    let lag = g.ar_coeff_lag as i32;
+    let ar_shift = c.ar_shift();
+    let lag = c.lag();
     for y in 3..73 {
         for x in 3..82 - 3 {
-            let mut s = 0;
-            let mut pos = 0;
-            'outer: for dr in -lag..=0 {
-                for dc in -lag..=lag {
-                    if dr == 0 && dc == 0 {
-                        break 'outer;
-                    }
-                    let c = g.ar_coeffs_y_plus_128[pos] as i32 - 128;
-                    s += luma[(y as i32 + dr) as usize][(x as i32 + dc) as usize] * c;
-                    pos += 1;
-                }
-            }
-            luma[y][x] = (luma[y][x] + round2(s, ar_shift)).clamp(grain_min, grain_max);
-        }
-    }
-    let chroma_w = if sub_x != 0 { 44 } else { 82 };
-    let chroma_h = if sub_y != 0 { 38 } else { 73 };
-    let mut cb = vec![[0i32; 82]; 73];
-    let mut cr = vec![[0i32; 82]; 73];
-    if !mono {
-        for (arr, seed, on) in [(&mut cb, 0xb524u32, g.num_cb_points > 0), (&mut cr, 0x49d8, g.num_cr_points > 0)] {
-            rng.0 = g.grain_seed ^ seed;
-            for row in arr.iter_mut().take(chroma_h) {
-                for v in row.iter_mut().take(chroma_w) {
-                    let gv = if on || g.chroma_scaling_from_luma { GAUSSIAN_SEQUENCE[rng.next(11) as usize] as i32 } else { 0 };
-                    *v = round2(gv, shift);
-                }
-            }
-        }
-        for y in 3..chroma_h {
-            for x in 3..chroma_w - 3 {
-                let mut s0 = 0;
-                let mut s1 = 0;
-                let mut pos = 0;
-                'outer: for dr in -lag..=0 {
-                    for dc in -lag..=lag {
-                        let c0 = g.ar_coeffs_cb_plus_128[pos] as i32 - 128;
-                        let c1 = g.ar_coeffs_cr_plus_128[pos] as i32 - 128;
-                        if dr == 0 && dc == 0 {
-                            if g.num_y_points > 0 {
-                                let mut l = 0;
-                                let lx = ((x - 3) << sub_x) + 3;
-                                let ly = ((y - 3) << sub_y) + 3;
-                                for i in 0..=sub_y {
-                                    for j in 0..=sub_x {
-                                        l += luma[ly + i][lx + j];
-                                    }
-                                }
-                                let l = round2(l, (sub_x + sub_y) as u32);
-                                s0 += l * c0;
-                                s1 += l * c1;
-                            }
-                            break 'outer;
-                        }
-                        let (yy, xx) = ((y as i32 + dr) as usize, (x as i32 + dc) as usize);
-                        s0 += cb[yy][xx] * c0;
-                        s1 += cr[yy][xx] * c1;
-                        pos += 1;
-                    }
-                }
-                cb[y][x] = (cb[y][x] + round2(s0, ar_shift)).clamp(grain_min, grain_max);
-                cr[y][x] = (cr[y][x] + round2(s1, ar_shift)).clamp(grain_min, grain_max);
-            }
+            let s = luma_ar_sum(g, &luma, lag, y, x);
+            luma[y][x] = (luma[y][x] + round2(s, ar_shift)).clamp(c.grain_min, c.grain_max);
         }
     }
+    luma
+}
 
-    // Scaling lookup (7.18.3.4).
+/// The luma contribution to the chroma autoregressive filter at chroma position (y, x).
+fn chroma_luma_term(c: &GrainCtx, luma: &[[i32; 82]], y: usize, x: usize) -> i32 {
+    let mut l = 0;
+    let lx = ((x - 3) << c.sub_x) + 3;
+    let ly = ((y - 3) << c.sub_y) + 3;
+    for i in 0..=c.sub_y {
+        for j in 0..=c.sub_x {
+            l += luma[ly + i][lx + j];
+        }
+    }
+    round2(l, (c.sub_x + c.sub_y) as u32)
+}
+
+fn chroma_ar_sums(c: &GrainCtx, t: &Templates, y: usize, x: usize) -> (i32, i32) {
+    let g = c.g;
+    let lag = c.lag();
+    let mut s0 = 0;
+    let mut s1 = 0;
+    let mut pos = 0;
+    'outer: for dr in -lag..=0 {
+        for dc in -lag..=lag {
+            let c0 = g.ar_coeffs_cb_plus_128[pos] as i32 - 128;
+            let c1 = g.ar_coeffs_cr_plus_128[pos] as i32 - 128;
+            if dr == 0 && dc == 0 {
+                if g.num_y_points > 0 {
+                    let l = chroma_luma_term(c, &t.luma, y, x);
+                    s0 += l * c0;
+                    s1 += l * c1;
+                }
+                break 'outer;
+            }
+            let (yy, xx) = ((y as i32 + dr) as usize, (x as i32 + dc) as usize);
+            s0 += t.cb[yy][xx] * c0;
+            s1 += t.cr[yy][xx] * c1;
+            pos += 1;
+        }
+    }
+    (s0, s1)
+}
+
+/// Chroma grain templates; zero for monochrome pictures.
+fn chroma_grain(c: &GrainCtx, rng: &mut Rng, luma: GrainTemplate) -> Templates {
+    let g = c.g;
+    let mut t = Templates { luma, cb: vec![[0i32; 82]; 73], cr: vec![[0i32; 82]; 73] };
+    if c.mono {
+        return t;
+    }
+    let shift = c.grain_shift();
+    let ar_shift = c.ar_shift();
+    let chroma_w = if c.sub_x != 0 { 44 } else { 82 };
+    let chroma_h = if c.sub_y != 0 { 38 } else { 73 };
+    for (arr, seed, on) in [(&mut t.cb, 0xb524u32, g.num_cb_points > 0), (&mut t.cr, 0x49d8, g.num_cr_points > 0)] {
+        rng.0 = g.grain_seed ^ seed;
+        for row in arr.iter_mut().take(chroma_h) {
+            for v in row.iter_mut().take(chroma_w) {
+                let gv = if on || g.chroma_scaling_from_luma { GAUSSIAN_SEQUENCE[rng.next(11) as usize] as i32 } else { 0 };
+                *v = round2(gv, shift);
+            }
+        }
+    }
+    for y in 3..chroma_h {
+        for x in 3..chroma_w - 3 {
+            let (s0, s1) = chroma_ar_sums(c, &t, y, x);
+            t.cb[y][x] = (t.cb[y][x] + round2(s0, ar_shift)).clamp(c.grain_min, c.grain_max);
+            t.cr[y][x] = (t.cr[y][x] + round2(s1, ar_shift)).clamp(c.grain_min, c.grain_max);
+        }
+    }
+    t
+}
+
+/// Scaling lookup tables (7.18.3.4), one per plane.
+fn scaling_luts(g: &FilmGrainParams, num_planes: usize) -> [[i32; 256]; 3] {
     let mut lut = [[0i32; 256]; 3];
     for (plane, l) in lut.iter_mut().enumerate().take(num_planes) {
         let (n, xs, ys) = if plane == 0 || g.chroma_scaling_from_luma {
@@ -159,101 +229,141 @@ pub(crate) fn apply(pic: &mut Picture, g: &FilmGrainParams, seq: &SequenceHeader
         };
         *l = scaling_lut(n, xs, ys);
     }
-    let scale_lut = |plane: usize, index: i32| -> i32 {
-        let sh = bd - 8;
-        let x = index >> sh;
-        let rem = index - (x << sh);
-        if bd == 8 || x == 255 {
-            lut[plane][x as usize]
-        } else {
-            let start = lut[plane][x as usize];
-            let end = lut[plane][x as usize + 1];
-            start + round2((end - start) * rem, sh)
-        }
-    };
+    lut
+}
 
-    // Noise stripes (7.18.3.5).
-    let stripe_w = w + 64;
-    let n_stripes = (h.div_ceil(2)).div_ceil(16).max(1);
-    let mut stripes: Vec<[Vec<i32>; 3]> = (0..n_stripes).map(|_| [vec![0; 34 * stripe_w], vec![0; 34 * stripe_w], vec![0; 34 * stripe_w]]).collect();
-    let overlap = g.overlap_flag;
-    for (luma_num, stripe) in stripes.iter_mut().enumerate() {
-        let mut r = g.grain_seed;
-        r ^= (((luma_num * 37 + 178) & 255) << 8) as u32;
-        r ^= ((luma_num * 173 + 105) & 255) as u32;
-        rng.0 = r;
-        let mut x = 0;
-        while x < w.div_ceil(2) {
-            let rand = rng.next(8) as usize;
-            let off_x = rand >> 4;
-            let off_y = rand & 15;
-            for (plane, s) in stripe.iter_mut().enumerate().take(num_planes) {
-                let psx = if plane > 0 { sub_x } else { 0 };
-                let psy = if plane > 0 { sub_y } else { 0 };
-                let pox = if psx != 0 { 6 + off_x } else { 9 + off_x * 2 };
-                let poy = if psy != 0 { 6 + off_y } else { 9 + off_y * 2 };
-                let src = match plane {
-                    0 => &luma,
-                    1 => &cb,
-                    _ => &cr,
+fn scale_lookup(lut: &[[i32; 256]; 3], bd: u32, plane: usize, index: i32) -> i32 {
+    let sh = bd - 8;
+    let x = index >> sh;
+    let rem = index - (x << sh);
+    if bd == 8 || x == 255 {
+        lut[plane][x as usize]
+    } else {
+        let start = lut[plane][x as usize];
+        let end = lut[plane][x as usize + 1];
+        start + round2((end - start) * rem, sh)
+    }
+}
+
+/// Fills one plane of a noise stripe from the grain template at stripe column `x`.
+fn fill_stripe_plane(c: &GrainCtx, src: &GrainTemplate, s: &mut [i32], plane: usize, offsets: (usize, usize), x: usize) {
+    let (off_x, off_y) = offsets;
+    let stripe_w = c.stripe_w();
+    let overlap = c.g.overlap_flag;
+    let psx = if plane > 0 { c.sub_x } else { 0 };
+    let psy = if plane > 0 { c.sub_y } else { 0 };
+    let pox = if psx != 0 { 6 + off_x } else { 9 + off_x * 2 };
+    let poy = if psy != 0 { 6 + off_y } else { 9 + off_y * 2 };
+    for i in 0..(34 >> psy) {
+        for j in 0..(34 >> psx) {
+            let mut gv = src[poy + i][pox + j];
+            let (idx, blend) = if psx == 0 { (i * stripe_w + x * 2 + j, j < 2) } else { (i * stripe_w + x + j, j == 0) };
+            if blend && overlap && x > 0 {
+                let old = s[idx];
+                gv = if psx != 0 {
+                    old * 23 + gv * 22
+                } else if j == 0 {
+                    old * 27 + gv * 17
+                } else {
+                    old * 17 + gv * 27
                 };
-                for i in 0..(34 >> psy) {
-                    for j in 0..(34 >> psx) {
-                        let mut gv = src[poy + i][pox + j];
-                        if psx == 0 {
-                            let idx = i * stripe_w + x * 2 + j;
-                            if j < 2 && overlap && x > 0 {
-                                let old = s[idx];
-                                gv = if j == 0 { old * 27 + gv * 17 } else { old * 17 + gv * 27 };
-                                gv = round2(gv, 5).clamp(grain_min, grain_max);
-                            }
-                            s[idx] = gv;
-                        } else {
-                            let idx = i * stripe_w + x + j;
-                            if j == 0 && overlap && x > 0 {
-                                let old = s[idx];
-                                gv = round2(old * 23 + gv * 22, 5).clamp(grain_min, grain_max);
-                            }
-                            s[idx] = gv;
-                        }
-                    }
-                }
+                gv = round2(gv, 5).clamp(c.grain_min, c.grain_max);
             }
-            x += 16;
+            s[idx] = gv;
         }
     }
-    let noise_at = |plane: usize, y: usize, x: usize| -> i32 {
-        let psy = if plane > 0 { sub_y } else { 0 };
+}
+
+/// Noise stripes (7.18.3.5).
+struct Stripes {
+    data: Vec<[Vec<i32>; 3]>,
+    stripe_w: usize,
+    overlap: bool,
+    sub_y: usize,
+    grain_min: i32,
+    grain_max: i32,
+}
+
+impl Stripes {
+    fn build(c: &GrainCtx, t: &Templates, rng: &mut Rng) -> Stripes {
+        let g = c.g;
+        let stripe_w = c.stripe_w();
+        let n_stripes = (c.h.div_ceil(2)).div_ceil(16).max(1);
+        let mut data: Vec<[Vec<i32>; 3]> = (0..n_stripes).map(|_| [vec![0; 34 * stripe_w], vec![0; 34 * stripe_w], vec![0; 34 * stripe_w]]).collect();
+        for (luma_num, stripe) in data.iter_mut().enumerate() {
+            let mut r = g.grain_seed;
+            r ^= (((luma_num * 37 + 178) & 255) << 8) as u32;
+            r ^= ((luma_num * 173 + 105) & 255) as u32;
+            rng.0 = r;
+            let mut x = 0;
+            while x < c.w.div_ceil(2) {
+                let rand = rng.next(8) as usize;
+                let offsets = (rand >> 4, rand & 15);
+                for (plane, s) in stripe.iter_mut().enumerate().take(c.num_planes()) {
+                    fill_stripe_plane(c, t.plane(plane), s, plane, offsets, x);
+                }
+                x += 16;
+            }
+        }
+        Stripes { data, stripe_w, overlap: g.overlap_flag, sub_y: c.sub_y, grain_min: c.grain_min, grain_max: c.grain_max }
+    }
+
+    fn noise_at(&self, plane: usize, y: usize, x: usize) -> i32 {
+        let psy = if plane > 0 { self.sub_y } else { 0 };
         let luma_num = y >> (5 - psy);
         let i = y - (luma_num << (5 - psy));
-        let mut gv = stripes[luma_num][plane][i * stripe_w + x];
+        let mut gv = self.data[luma_num][plane][i * self.stripe_w + x];
         if psy == 0 {
-            if i < 2 && luma_num > 0 && overlap {
-                let old = stripes[luma_num - 1][plane][(i + 32) * stripe_w + x];
+            if i < 2 && luma_num > 0 && self.overlap {
+                let old = self.data[luma_num - 1][plane][(i + 32) * self.stripe_w + x];
                 gv = if i == 0 { old * 27 + gv * 17 } else { old * 17 + gv * 27 };
-                gv = round2(gv, 5).clamp(grain_min, grain_max);
+                gv = round2(gv, 5).clamp(self.grain_min, self.grain_max);
             }
-        } else if i < 1 && luma_num > 0 && overlap {
-            let old = stripes[luma_num - 1][plane][(i + 16) * stripe_w + x];
-            gv = round2(old * 23 + gv * 22, 5).clamp(grain_min, grain_max);
+        } else if i < 1 && luma_num > 0 && self.overlap {
+            let old = self.data[luma_num - 1][plane][(i + 16) * self.stripe_w + x];
+            gv = round2(old * 23 + gv * 22, 5).clamp(self.grain_min, self.grain_max);
         }
         gv
-    };
+    }
+}
 
-    // Blend (7.18.3.5).
-    let (min_value, max_luma, max_chroma) = if g.clip_to_restricted_range {
-        let ml = 235 << (bd - 8);
-        (16 << (bd - 8), ml, if pic.matrix_coefficients as usize == MC_IDENTITY { ml } else { 240 << (bd - 8) })
-    } else {
-        let m = (256 << (bd - 8)) - 1;
-        (0, m, m)
-    };
-    let pix_max = (1i32 << bd) - 1;
-    let scaling_shift = g.grain_scaling_minus_8 + 8;
-    if !mono {
+/// Output clipping range and scaling shift for the blend step (7.18.3.5).
+struct BlendRange {
+    min_value: i32,
+    max_luma: i32,
+    max_chroma: i32,
+    pix_max: i32,
+    scaling_shift: u32,
+}
+
+impl BlendRange {
+    fn new(c: &GrainCtx, matrix_coefficients: usize) -> BlendRange {
+        let bd = c.bd;
+        let (min_value, max_luma, max_chroma) = if c.g.clip_to_restricted_range {
+            let ml = 235 << (bd - 8);
+            (16 << (bd - 8), ml, if matrix_coefficients == MC_IDENTITY { ml } else { 240 << (bd - 8) })
+        } else {
+            let m = (256 << (bd - 8)) - 1;
+            (0, m, m)
+        };
+        BlendRange { min_value, max_luma, max_chroma, pix_max: (1i32 << bd) - 1, scaling_shift: c.g.grain_scaling_minus_8 + 8 }
+    }
+}
+
+struct BlendEnv<'a> {
+    c: &'a GrainCtx<'a>,
+    lut: &'a [[i32; 256]; 3],
+    stripes: &'a Stripes,
+    range: &'a BlendRange,
+}
+
+impl BlendEnv<'_> {
+    fn blend_chroma(&self, py: &[u16], pu: &mut [u16], pv: &mut [u16]) {
+        let c = self.c;
+        let g = c.g;
+        let (w, h, sub_x, sub_y, bd) = (c.w, c.h, c.sub_x, c.sub_y, c.bd);
         let cw = (w + sub_x) >> sub_x;
         let ch = (h + sub_y) >> sub_y;
-        let [py, pu, pv] = &mut pic.planes;
         for y in 0..ch {
             for x in 0..cw {
                 let lx = x << sub_x;
@@ -272,26 +382,62 @@ pub(crate) fn apply(pic: &mut Picture, g: &FilmGrainParams, seq: &SequenceHeader
                         avg
                     } else {
                         let combined = avg * (luma_mult as i32 - 128) + orig * (mult as i32 - 128);
-                        ((combined >> 6) + ((offset as i32 - 256) << (bd - 8))).clamp(0, pix_max)
+                        ((combined >> 6) + ((offset as i32 - 256) << (bd - 8))).clamp(0, self.range.pix_max)
                     };
-                    let noise = round2(scale_lut(plane, merged) * noise_at(plane, y, x), scaling_shift);
-                    out[y * cw + x] = (orig + noise).clamp(min_value, max_chroma) as u16;
+                    let noise = round2(scale_lookup(self.lut, bd, plane, merged) * self.stripes.noise_at(plane, y, x), self.range.scaling_shift);
+                    out[y * cw + x] = (orig + noise).clamp(self.range.min_value, self.range.max_chroma) as u16;
                 }
             }
         }
     }
-    if g.num_y_points > 0 {
-        let py = &mut pic.planes[0];
+
+    fn blend_luma(&self, py: &mut [u16]) {
+        let (w, h) = (self.c.w, self.c.h);
         for y in 0..h {
             for x in 0..w {
                 let orig = py[y * w + x] as i32;
-                let noise = round2(scale_lut(0, orig) * noise_at(0, y, x), scaling_shift);
-                py[y * w + x] = (orig + noise).clamp(min_value, max_luma) as u16;
+                let noise = round2(scale_lookup(self.lut, self.c.bd, 0, orig) * self.stripes.noise_at(0, y, x), self.range.scaling_shift);
+                py[y * w + x] = (orig + noise).clamp(self.range.min_value, self.range.max_luma) as u16;
             }
         }
     }
 }
 
+pub(crate) fn apply(pic: &mut Picture, g: &FilmGrainParams, seq: &SequenceHeader) {
+    let _ = seq;
+    let bd = pic.bit_depth as u32;
+    let grain_center = 128i32 << (bd - 8);
+    let c = GrainCtx {
+        g,
+        bd,
+        mono: pic.mono_chrome,
+        sub_x: pic.subsampling_x as usize,
+        sub_y: pic.subsampling_y as usize,
+        w: pic.width as usize,
+        h: pic.height as usize,
+        grain_min: -grain_center,
+        grain_max: (256i32 << (bd - 8)) - 1 - grain_center,
+    };
+
+    // Generate grain (7.18.3.3).
+    let mut rng = Rng(g.grain_seed);
+    let luma = luma_grain(&c, &mut rng);
+    let templates = chroma_grain(&c, &mut rng, luma);
+    let lut = scaling_luts(g, c.num_planes());
+    let stripes = Stripes::build(&c, &templates, &mut rng);
+
+    // Blend (7.18.3.5).
+    let range = BlendRange::new(&c, pic.matrix_coefficients as usize);
+    let env = BlendEnv { c: &c, lut: &lut, stripes: &stripes, range: &range };
+    if !c.mono {
+        let [py, pu, pv] = &mut pic.planes;
+        env.blend_chroma(py, pu, pv);
+    }
+    if g.num_y_points > 0 {
+        env.blend_luma(&mut pic.planes[0]);
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

crates/av1/src/transform.rs

@@ -165,6 +165,15 @@ pub(crate) fn inverse_dct<T: Lane>(t: &mut [T], n: u32, r: u32) {
     for i in 0..len {
         t[i] = copy[brev(n, i)];
     }
+    inverse_dct_stage1(t, n, r);
+    inverse_dct_stage2(t, n, r);
+    inverse_dct_stage3(t, n, r);
+    inverse_dct_stage4(t, n, r);
+    inverse_dct_stage5(t, n, r);
+    inverse_dct_stage6(t, n, r);
+}
+
+fn inverse_dct_stage1<T: Lane>(t: &mut [T], n: u32, r: u32) {
     if n == 6 {
         for i in 0..16 {
             bf(t, 32 + i, 63 - i, 63 - 4 * brev(4, i) as i32, false);
@@ -190,6 +199,9 @@ pub(crate) fn inverse_dct<T: Lane>(t: &mut [T], n: u32, r: u32) {
             hd(t, 16 + 2 * i, 17 + 2 * i, i & 1 == 1, r);
         }
     }
+}
+
+fn inverse_dct_stage2<T: Lane>(t: &mut [T], n: u32, r: u32) {
     if n == 6 {
         for i in 0..4 {
             for j in 0..2 {
@@ -221,6 +233,9 @@ pub(crate) fn inverse_dct<T: Lane>(t: &mut [T], n: u32, r: u32) {
             }
         }
     }
+}
+
+fn inverse_dct_stage3<T: Lane>(t: &mut [T], n: u32, r: u32) {
     for i in 0..2 {
         bf(t, 2 * i, 2 * i + 1, 32 + 16 * i as i32, i == 0);
     }
@@ -248,6 +263,9 @@ pub(crate) fn inverse_dct<T: Lane>(t: &mut [T], n: u32, r: u32) {
             }
         }
     }
+}
+
+fn inverse_dct_stage4<T: Lane>(t: &mut [T], n: u32, r: u32) {
     for i in 0..2 {
         hd(t, i, 3 - i, false, r);
     }
@@ -273,6 +291,9 @@ pub(crate) fn inverse_dct<T: Lane>(t: &mut [T], n: u32, r: u32) {
             }
         }
     }
+}
+
+fn inverse_dct_stage5<T: Lane>(t: &mut [T], n: u32, r: u32) {
     if n >= 3 {
         for i in 0..4 {
             hd(t, i, 7 - i, false, r);
@@ -300,6 +321,9 @@ pub(crate) fn inverse_dct<T: Lane>(t: &mut [T], n: u32, r: u32) {
             hd(t, i, 15 - i, false, r);
         }
     }
+}
+
+fn inverse_dct_stage6<T: Lane>(t: &mut [T], n: u32, r: u32) {
     if n >= 5 {
         for i in 0..4 {
             bf(t, 27 - i, 20 + i, 32, true);
@@ -531,10 +555,9 @@ pub(crate) fn inverse_transform_2d(dequant: &[i32], residual: &mut [i32], tx_sz:
     let col_clamp = (bit_depth + 6).max(16);
     let rk = row_kind(tx_type);
     let ck = col_kind(tx_type);
-    let mut t = [0i32; 64];
     let rect2 = log2w.abs_diff(log2h) == 1;
+    let p = Tx2d { w, h, log2w, log2h, row_shift, col_shift, row_clamp, col_clamp, rk, ck, rect2 };
     if !lossless && bit_depth <= 10 {
-        let p = Tx2d { w, h, log2w, log2h, row_shift, col_shift, row_clamp, col_clamp, rk, ck, rect2 };
         if w >= 8 && h >= 8 {
             transform_lanes::<8>(dequant, residual, &p);
         } else {
@@ -542,6 +565,13 @@ pub(crate) fn inverse_transform_2d(dequant: &[i32], residual: &mut [i32], tx_sz:
         }
         return;
     }
+    scalar_row_pass(dequant, residual, &p, lossless);
+    scalar_col_pass(residual, &p, lossless);
+}
+
+fn scalar_row_pass(dequant: &[i32], residual: &mut [i32], p: &Tx2d, lossless: bool) {
+    let (w, h, log2w, row_shift, row_clamp, col_clamp, rk, rect2) = (p.w, p.h, p.log2w, p.row_shift, p.row_clamp, p.col_clamp, p.rk, p.rect2);
+    let mut t = [0i32; 64];
     let in_w = w.min(32);
     for i in 0..h {
         // Every 1D transform maps an all-zero row to zeros (each step is linear with
@@ -572,6 +602,11 @@ pub(crate) fn inverse_transform_2d(dequant: &[i32], residual: &mut [i32], tx_sz:
             residual[i * w + j] = round2(t[j], row_shift).clamp(lo, hi);
         }
     }
+}
+
+fn scalar_col_pass(residual: &mut [i32], p: &Tx2d, lossless: bool) {
+    let (w, h, log2h, col_shift, col_clamp, ck) = (p.w, p.h, p.log2h, p.col_shift, p.col_clamp, p.ck);
+    let mut t = [0i32; 64];
     for j in 0..w {
         for i in 0..h {
             t[i] = residual[i * w + j];
@@ -617,11 +652,16 @@ fn transform_1d<T: Lane>(t: &mut [T], kind: Kind1d, n: u32, r: u32) {
 /// The 2D inverse transform with `L` rows (row pass) or columns (column pass) at a time;
 /// bit-identical to the one-at-a-time process for bit depths up to 10.
 fn transform_lanes<const L: usize>(dequant: &[i32], residual: &mut [i32], p: &Tx2d) {
+    let mut t = [[0i32; L]; 64];
+    lanes_row_pass(dequant, residual, p, &mut t);
+    lanes_col_pass(residual, p, &mut t);
+}
+
+fn lanes_row_pass<const L: usize>(dequant: &[i32], residual: &mut [i32], p: &Tx2d, t: &mut [[i32; L]; 64]) {
     let (w, h) = (p.w, p.h);
     let in_w = w.min(32);
     let lo = -(1i32 << (p.col_clamp - 1));
     let hi = (1i32 << (p.col_clamp - 1)) - 1;
-    let mut t = [[0i32; L]; 64];
     let mut i0 = 0;
     while i0 < h {
         // Rows without coefficients (and rows 32..63) transform to zeros.
@@ -649,6 +689,10 @@ fn transform_lanes<const L: usize>(dequant: &[i32], residual: &mut [i32], p: &Tx
         }
         i0 += L;
     }
+}
+
+fn lanes_col_pass<const L: usize>(residual: &mut [i32], p: &Tx2d, t: &mut [[i32; L]; 64]) {
+    let (w, h) = (p.w, p.h);
     let mut j0 = 0;
     while j0 < w {
         for i in 0..h {

crates/h264/src/dpb.rs

@@ -3,7 +3,7 @@
 
 use crate::error::{Result, ensure};
 use crate::picture::{FrameRef, RefPic};
-use crate::slice::{Mmco, SliceHeader, SliceType};
+use crate::slice::{Mmco, RefPicListMod, SliceHeader, SliceType};
 use std::sync::Arc;
 
 #[derive(Clone, Copy, Debug, PartialEq, Eq)]
@@ -174,6 +174,103 @@ impl Dpb {
         }
     }
 
+    /// Initial reference picture lists (8.2.4.2) as DPB entries.
+    fn initial_entry_lists<'a>(
+        sh: &SliceHeader,
+        cur_poc: i32,
+        short: &[&'a DpbEntry],
+        long: &[&'a DpbEntry],
+        pic_num: impl Fn(&DpbEntry) -> i32,
+    ) -> [Vec<&'a DpbEntry>; 2] {
+        let mut entry_lists: [Vec<&'a DpbEntry>; 2] = [Vec::new(), Vec::new()];
+        if sh.slice_type != SliceType::B {
+            let mut s = short.to_vec();
+            s.sort_by_key(|e| std::cmp::Reverse(pic_num(e)));
+            entry_lists[0] = s.into_iter().chain(long.iter().copied()).collect();
+        } else {
+            let mut before: Vec<&'a DpbEntry> = short.iter().copied().filter(|e| e.poc < cur_poc).collect();
+            let mut after: Vec<&'a DpbEntry> = short.iter().copied().filter(|e| e.poc > cur_poc).collect();
+            before.sort_by_key(|e| std::cmp::Reverse(e.poc));
+            after.sort_by_key(|e| e.poc);
+            entry_lists[0] = before.iter().chain(after.iter()).chain(long.iter()).copied().collect();
+            entry_lists[1] = after.iter().chain(before.iter()).chain(long.iter()).copied().collect();
+            if entry_lists[1].len() > 1 && entry_lists[0].len() == entry_lists[1].len() {
+                let same = entry_lists[0].iter().zip(entry_lists[1].iter()).all(|(a, b)| std::ptr::eq(*a, *b));
+                if same {
+                    entry_lists[1].swap(0, 1);
+                }
+            }
+        }
+        entry_lists
+    }
+
+    /// Reference picture list modification (8.2.4.3) of one list; `n` is num_ref_idx_active.
+    fn modify_ref_list<'a>(
+        list: &mut Vec<Option<&'a DpbEntry>>,
+        mods: &[RefPicListMod],
+        n: usize,
+        cur_fn: u32,
+        max_frame_num: u32,
+        short: &[&'a DpbEntry],
+        long: &[&'a DpbEntry],
+    ) -> Result<()> {
+        let max_pic_num = max_frame_num as i32;
+        let curr_pic_num = cur_fn as i32;
+        let mut pred = curr_pic_num;
+        let mut ref_idx = 0usize;
+        for m in mods {
+            ensure!(ref_idx < n, "too many ref_pic_list_modification entries");
+            list.push(None); // temporarily one longer
+            let (target, is_long): (Option<&DpbEntry>, bool) = match m.idc {
+                0 | 1 => {
+                    let d = m.value as i32 + 1;
+                    let no_wrap = if m.idc == 0 {
+                        if pred - d < 0 { pred - d + max_pic_num } else { pred - d }
+                    } else if pred + d >= max_pic_num {
+                        pred + d - max_pic_num
+                    } else {
+                        pred + d
+                    };
+                    pred = no_wrap;
+                    let pn = if no_wrap > curr_pic_num { no_wrap - max_pic_num } else { no_wrap };
+                    (short.iter().copied().find(|e| Self::frame_num_wrap(e, cur_fn, max_frame_num) == pn), false)
+                }
+                2 => (long.iter().copied().find(|e| e.long_term_frame_idx == m.value), true),
+                _ => continue,
+            };
+            let Some(target) = target else {
+                // Missing picture: keep the list unchanged (error concealment).
+                list.pop();
+                ref_idx += 1;
+                continue;
+            };
+            for c in (ref_idx + 1..=n).rev() {
+                list[c] = list[c - 1];
+            }
+            list[ref_idx] = Some(target);
+            ref_idx += 1;
+            let mut nidx = ref_idx;
+            for c in ref_idx..=n {
+                let keep = match list[c] {
+                    None => true,
+                    Some(e) => {
+                        if is_long {
+                            !(e.mark == RefMark::Long && e.long_term_frame_idx == target.long_term_frame_idx)
+                        } else {
+                            !(e.mark == RefMark::Short && std::ptr::eq(e, target))
+                        }
+                    }
+                };
+                if keep {
+                    list[nidx] = list[c];
+                    nidx += 1;
+                }
+            }
+            list.truncate(n);
+        }
+        Ok(())
+    }
+
     /// Initial reference picture lists (8.2.4.2) + modification (8.2.4.3) for a frame slice.
     pub fn build_ref_lists(&self, sh: &SliceHeader, cur_poc: i32, max_frame_num: u32) -> Result<[Vec<RefPic>; 2]> {
         let mut lists: [Vec<RefPic>; 2] = [Vec::new(), Vec::new()];
@@ -186,25 +283,7 @@ impl Dpb {
         let mut long: Vec<&DpbEntry> = self.entries.iter().filter(|e| e.mark == RefMark::Long).collect();
         long.sort_by_key(|e| e.long_term_frame_idx);
         let to_ref = |e: &&DpbEntry| RefPic { frame: e.frame.clone(), long_term: e.mark == RefMark::Long };
-        let mut entry_lists: [Vec<&DpbEntry>; 2] = [Vec::new(), Vec::new()];
-        if sh.slice_type != SliceType::B {
-            let mut s = short.clone();
-            s.sort_by_key(|e| std::cmp::Reverse(pic_num(e)));
-            entry_lists[0] = s.into_iter().chain(long.iter().copied()).collect();
-        } else {
-            let mut before: Vec<&DpbEntry> = short.iter().copied().filter(|e| e.poc < cur_poc).collect();
-            let mut after: Vec<&DpbEntry> = short.iter().copied().filter(|e| e.poc > cur_poc).collect();
-            before.sort_by_key(|e| std::cmp::Reverse(e.poc));
-            after.sort_by_key(|e| e.poc);
-            entry_lists[0] = before.iter().chain(after.iter()).chain(long.iter()).copied().collect();
-            entry_lists[1] = after.iter().chain(before.iter()).chain(long.iter()).copied().collect();
-            if entry_lists[1].len() > 1 && entry_lists[0].len() == entry_lists[1].len() {
-                let same = entry_lists[0].iter().zip(entry_lists[1].iter()).all(|(a, b)| std::ptr::eq(*a, *b));
-                if same {
-                    entry_lists[1].swap(0, 1);
-                }
-            }
-        }
+        let entry_lists = Self::initial_entry_lists(sh, cur_poc, &short, &long, pic_num);
         let nlists = if sh.slice_type == SliceType::B { 2 } else { 1 };
         for l in 0..nlists {
             let n = sh.num_ref_idx_active[l] as usize;
@@ -212,60 +291,7 @@ impl Dpb {
             list.truncate(n);
             list.resize(n, None);
             if !sh.ref_pic_list_mod[l].is_empty() {
-                let max_pic_num = max_frame_num as i32;
-                let curr_pic_num = cur_fn as i32;
-                let mut pred = curr_pic_num;
-                let mut ref_idx = 0usize;
-                for m in &sh.ref_pic_list_mod[l] {
-                    ensure!(ref_idx < n, "too many ref_pic_list_modification entries");
-                    list.push(None); // temporarily one longer
-                    let (target, is_long): (Option<&DpbEntry>, bool) = match m.idc {
-                        0 | 1 => {
-                            let d = m.value as i32 + 1;
-                            let no_wrap = if m.idc == 0 {
-                                if pred - d < 0 { pred - d + max_pic_num } else { pred - d }
-                            } else if pred + d >= max_pic_num {
-                                pred + d - max_pic_num
-                            } else {
-                                pred + d
-                            };
-                            pred = no_wrap;
-                            let pn = if no_wrap > curr_pic_num { no_wrap - max_pic_num } else { no_wrap };
-                            (short.iter().copied().find(|e| pic_num(e) == pn), false)
-                        }
-                        2 => (long.iter().copied().find(|e| e.long_term_frame_idx == m.value), true),
-                        _ => continue,
-                    };
-                    let Some(target) = target else {
-                        // Missing picture: keep the list unchanged (error concealment).
-                        list.pop();
-                        ref_idx += 1;
-                        continue;
-                    };
-                    for c in (ref_idx + 1..=n).rev() {
-                        list[c] = list[c - 1];
-                    }
-                    list[ref_idx] = Some(target);
-                    ref_idx += 1;
-                    let mut nidx = ref_idx;
-                    for c in ref_idx..=n {
-                        let keep = match list[c] {
-                            None => true,
-                            Some(e) => {
-                                if is_long {
-                                    !(e.mark == RefMark::Long && e.long_term_frame_idx == target.long_term_frame_idx)
-                                } else {
-                                    !(e.mark == RefMark::Short && std::ptr::eq(e, target))
-                                }
-                            }
-                        };
-                        if keep {
-                            list[nidx] = list[c];
-                            nidx += 1;
-                        }
-                    }
-                    list.truncate(n);
-                }
+                Self::modify_ref_list(&mut list, &sh.ref_pic_list_mod[l], n, cur_fn, max_frame_num, &short, &long)?;
             }
             // Fill missing entries (should not happen in conforming streams) with the first valid entry.
             let fallback = list.iter().flatten().next().copied().or_else(|| self.entries.last());
@@ -275,91 +301,88 @@ impl Dpb {
         Ok(lists)
     }
 
-    /// Adaptive memory control (8.2.5.4). Returns true if MMCO 5 was executed.
-    fn apply_mmcos(&mut self, mmcos: &[Mmco], cur_frame_num: u32, max_frame_num: u32, cur_long: &mut Option<u32>) -> bool {
+    /// One memory management control operation; returns true for MMCO 5.
+    fn apply_mmco(&mut self, m: &Mmco, cur_frame_num: u32, max_frame_num: u32, cur_long: &mut Option<u32>) -> bool {
         let mut had5 = false;
         let curr_pic_num = cur_frame_num as i32;
-        for m in mmcos {
-            match m.op {
-                1 => {
-                    let pn = curr_pic_num - (m.difference_of_pic_nums_minus1 as i32 + 1);
-                    for e in &mut self.entries {
-                        if e.mark == RefMark::Short && Self::frame_num_wrap(e, cur_frame_num, max_frame_num) == pn {
-                            e.mark = RefMark::Unused;
-                        }
+        match m.op {
+            1 => {
+                let pn = curr_pic_num - (m.difference_of_pic_nums_minus1 as i32 + 1);
+                for e in &mut self.entries {
+                    if e.mark == RefMark::Short && Self::frame_num_wrap(e, cur_frame_num, max_frame_num) == pn {
+                        e.mark = RefMark::Unused;
                     }
                 }
-                2 => {
-                    for e in &mut self.entries {
-                        if e.mark == RefMark::Long && e.long_term_frame_idx == m.long_term_pic_num {
-                            e.mark = RefMark::Unused;
-                        }
+            }
+            2 => {
+                for e in &mut self.entries {
+                    if e.mark == RefMark::Long && e.long_term_frame_idx == m.long_term_pic_num {
+                        e.mark = RefMark::Unused;
                     }
                 }
-                3 => {
-                    let pn = curr_pic_num - (m.difference_of_pic_nums_minus1 as i32 + 1);
+            }
+            3 => {
+                let pn = curr_pic_num - (m.difference_of_pic_nums_minus1 as i32 + 1);
+                for e in &mut self.entries {
+                    if e.mark == RefMark::Long && e.long_term_frame_idx == m.long_term_frame_idx {
+                        e.mark = RefMark::Unused;
+                    }
+                }
+                for e in &mut self.entries {
+                    if e.mark == RefMark::Short && Self::frame_num_wrap(e, cur_frame_num, max_frame_num) == pn {
+                        e.mark = RefMark::Long;
+                        e.long_term_frame_idx = m.long_term_frame_idx;
+                    }
+                }
+            }
+            4 => {
+                let max = m.max_long_term_frame_idx_plus1;
+                for e in &mut self.entries {
+                    if e.mark == RefMark::Long && (max == 0 || e.long_term_frame_idx > max - 1) {
+                        e.mark = RefMark::Unused;
+                    }
+                }
+                self.max_long_term_frame_idx = if max == 0 { None } else { Some(max - 1) };
+            }
+            5 => {
+                for e in &mut self.entries {
+                    e.mark = RefMark::Unused;
+                }
+                self.max_long_term_frame_idx = None;
+                had5 = true;
+            }
+            6 => {
+                // Only an index within the allowed range may be assigned (8.2.5.4.6).
+                let allowed = match self.max_long_term_frame_idx {
+                    Some(max) => m.long_term_frame_idx <= max,
+                    None => false,
+                };
+                if allowed {
                     for e in &mut self.entries {
                         if e.mark == RefMark::Long && e.long_term_frame_idx == m.long_term_frame_idx {
                             e.mark = RefMark::Unused;
                         }
                     }
-                    for e in &mut self.entries {
-                        if e.mark == RefMark::Short && Self::frame_num_wrap(e, cur_frame_num, max_frame_num) == pn {
-                            e.mark = RefMark::Long;
-                            e.long_term_frame_idx = m.long_term_frame_idx;
-                        }
-                    }
+                    *cur_long = Some(m.long_term_frame_idx);
                 }
-                4 => {
-                    let max = m.max_long_term_frame_idx_plus1;
-                    for e in &mut self.entries {
-                        if e.mark == RefMark::Long && (max == 0 || e.long_term_frame_idx > max - 1) {
-                            e.mark = RefMark::Unused;
-                        }
-                    }
-                    self.max_long_term_frame_idx = if max == 0 { None } else { Some(max - 1) };
-                }
-                5 => {
-                    for e in &mut self.entries {
-                        e.mark = RefMark::Unused;
-                    }
-                    self.max_long_term_frame_idx = None;
-                    had5 = true;
-                }
-                6 => {
-                    let allowed = match self.max_long_term_frame_idx {
-                        Some(max) => m.long_term_frame_idx <= max,
-                        None => false,
-                    };
-                    if allowed {
-                        for e in &mut self.entries {
-                            if e.mark == RefMark::Long && e.long_term_frame_idx == m.long_term_frame_idx {
-                                e.mark = RefMark::Unused;
-                            }
-                        }
-                        *cur_long = Some(m.long_term_frame_idx);
-                    }
-                }
-                _ => {}
             }
+            _ => {}
+        }
+        had5
+    }
+
+    /// Adaptive memory control (8.2.5.4). Returns true if MMCO 5 was executed.
+    fn apply_mmcos(&mut self, mmcos: &[Mmco], cur_frame_num: u32, max_frame_num: u32, cur_long: &mut Option<u32>) -> bool {
+        let mut had5 = false;
+        for m in mmcos {
+            had5 |= self.apply_mmco(m, cur_frame_num, max_frame_num, cur_long);
         }
         self.prune();
         had5
     }
 
-    /// Reference marking for the just-decoded picture and insertion into the DPB with output bumping.
-    /// Returns pictures ready for output.
-    #[allow(clippy::too_many_arguments)]
-    pub fn store_picture(
-        &mut self,
-        sh: &SliceHeader,
-        frame: FrameRef,
-        poc: i32,
-        max_frame_num: u32,
-        max_num_ref_frames: usize,
-        meta: Arc<OutputMeta>,
-        out: &mut Vec<Output>,
-    ) {
+    /// Reference marking of the current picture (8.2.5); returns (mark, long-term idx, MMCO 5 executed).
+    fn mark_current(&mut self, sh: &SliceHeader, max_frame_num: u32, max_num_ref_frames: usize) -> (RefMark, u32, bool) {
         let mut mark = RefMark::Unused;
         let mut long_idx = 0;
         let mut had5 = false;
@@ -391,12 +414,12 @@ impl Dpb {
                 }
             }
         }
-        if had5 {
-            // All earlier pictures are output before the current one (C.4.4).
-            self.flush(out);
-        }
-        let frame_num = if had5 { 0 } else { sh.frame_num };
-        // Bumping until there is space (C.4.5.1 / C.4.5.2).
+        (mark, long_idx, had5)
+    }
+
+    /// Bumping until there is space (C.4.5.1 / C.4.5.2). Returns true when the (non-reference) current
+    /// picture has the smallest POC and is to be output directly instead of being stored.
+    fn bump_for_space(&mut self, mark: RefMark, poc: i32, out: &mut Vec<Output>) -> bool {
         if mark == RefMark::Unused {
             loop {
                 let waiting = self.entries.len();
@@ -406,8 +429,7 @@ impl Dpb {
                 // non-reference picture with the smallest POC is output directly
                 let min_poc = self.entries.iter().filter(|e| e.needed_for_output).map(|e| e.poc).min();
                 if min_poc.is_none_or(|m| poc < m) {
-                    out.push(Output { frame, poc, meta });
-                    return;
+                    return true;
                 }
                 if !self.bump(out) {
                     break;
@@ -425,6 +447,32 @@ impl Dpb {
                 }
             }
         }
+        false
+    }
+
+    /// Reference marking for the just-decoded picture and insertion into the DPB with output bumping.
+    /// Returns pictures ready for output.
+    #[allow(clippy::too_many_arguments)]
+    pub fn store_picture(
+        &mut self,
+        sh: &SliceHeader,
+        frame: FrameRef,
+        poc: i32,
+        max_frame_num: u32,
+        max_num_ref_frames: usize,
+        meta: Arc<OutputMeta>,
+        out: &mut Vec<Output>,
+    ) {
+        let (mark, long_idx, had5) = self.mark_current(sh, max_frame_num, max_num_ref_frames);
+        if had5 {
+            // All earlier pictures are output before the current one (C.4.4).
+            self.flush(out);
+        }
+        let frame_num = if had5 { 0 } else { sh.frame_num };
+        if self.bump_for_space(mark, poc, out) {
+            out.push(Output { frame, poc, meta });
+            return;
+        }
         self.entries.push(DpbEntry {
             frame,
             poc,

crates/h264/src/picture.rs

@@ -250,7 +250,7 @@ impl Frame {
 
     /// Copy the w x h window of chroma component `c` at (x0, y0) (clamped) into `out`.
     #[inline]
-    pub fn chroma_window(&self, c: usize, x0: i32, y0: i32, w: usize, h: usize, out: &mut [u8], os: usize) {
+    pub fn chroma_window(&self, c: usize, (x0, y0): (i32, i32), (w, h): (usize, usize), out: &mut [u8], os: usize) {
         let (cw, ch) = (self.cwidth, self.cheight);
         for r in 0..h {
             let yy = (y0 + r as i32).clamp(0, ch as i32 - 1) as usize;

crates/h264/src/slicedec.rs

@@ -232,7 +232,7 @@ fn predict_block(f: &Frame, mv: [i16; 2], px: usize, py: usize, bw: usize, bh: u
     let (fx, fy) = ((mv[0] & 7) as u32, (mv[1] & 7) as u32);
     for (comp, out) in c.into_iter().enumerate() {
         let mut w = [0u8; 16 * 9];
-        f.chroma_window(comp, cx, cy, cw + 1, ch + 1, &mut w, 16);
+        f.chroma_window(comp, (cx, cy), (cw + 1, ch + 1), &mut w, 16);
         inter::mc_chroma_win(&w, 16, fx, fy, cw, ch, out, cs);
     }
 }
@@ -665,33 +665,38 @@ impl<'a> SliceDecoder<'a> {
             return Ok(());
         }
         for p in 0..4 {
-            let sub = self.cur.sub[p];
-            let (x8, y8) = ((p & 1) * 2, (p >> 1) * 2);
-            if sub.direct {
-                self.derive_direct(1 << p)?;
-                continue;
+            self.derive_sub_mb_motion(p, n)?;
+        }
+        Ok(())
+    }
+
+    /// Motion of one 8x8 sub-macroblock `p` of a P8x8/B8x8 MB.
+    fn derive_sub_mb_motion(&mut self, p: usize, n: [usize; 2]) -> Result<()> {
+        let sub = self.cur.sub[p];
+        let (x8, y8) = ((p & 1) * 2, (p >> 1) * 2);
+        if sub.direct {
+            return self.derive_direct(1 << p);
+        }
+        let mut refs = [-1i8; 2];
+        for (l, r) in refs.iter_mut().enumerate() {
+            if sub.pred & (1 << l) != 0 {
+                let v = self.cur.ref_idx[l][p];
+                ensure!((v as usize) < n[l], "ref_idx {v} out of range (list {l}, size {})", n[l]);
+                *r = v;
             }
-            let mut refs = [-1i8; 2];
-            for (l, r) in refs.iter_mut().enumerate() {
-                if sub.pred & (1 << l) != 0 {
-                    let v = self.cur.ref_idx[l][p];
-                    ensure!((v as usize) < n[l], "ref_idx {v} out of range (list {l}, size {})", n[l]);
-                    *r = v;
+        }
+        for j in 0..sub.shape.num_parts() {
+            let (sx, sy, w, h) = sub.shape.rect(j);
+            let (x, y) = (x8 + sx, y8 + sy);
+            let mut mvs = [[0i16; 2]; 2];
+            for l in 0..2 {
+                if refs[l] >= 0 {
+                    let mvp = self.mvp(l, x, y, w, refs[l], 0, 0);
+                    let d = self.cur.mvd[l][p * 4 + j];
+                    mvs[l] = [mvp[0].wrapping_add(d[0]), mvp[1].wrapping_add(d[1])];
                 }
             }
-            for j in 0..sub.shape.num_parts() {
-                let (sx, sy, w, h) = sub.shape.rect(j);
-                let (x, y) = (x8 + sx, y8 + sy);
-                let mut mvs = [[0i16; 2]; 2];
-                for l in 0..2 {
-                    if refs[l] >= 0 {
-                        let mvp = self.mvp(l, x, y, w, refs[l], 0, 0);
-                        let d = self.cur.mvd[l][p * 4 + j];
-                        mvs[l] = [mvp[0].wrapping_add(d[0]), mvp[1].wrapping_add(d[1])];
-                    }
-                }
-                self.set_motion(x, y, w, h, refs, mvs);
-            }
+            self.set_motion(x, y, w, h, refs, mvs);
         }
         Ok(())
     }
@@ -1380,6 +1385,11 @@ impl<'a> SliceDecoder<'a> {
             cavlc::residual_block(r, &mut lv, 0, 15, 16, nc)?;
             self.put_luma_dc(&lv);
         }
+        self.residual_cavlc_luma(r, cbp, kind, t8)?;
+        self.residual_cavlc_chroma(r, cbp)
+    }
+
+    fn residual_cavlc_luma(&mut self, r: &mut BitReader, cbp: u8, kind: MbKind, t8: bool) -> Result<()> {
         for b8 in 0..4 {
             let (bx8, by8) = ((b8 & 1) * 2, (b8 >> 1) * 2);
             if cbp & (1 << b8) == 0 {
@@ -1413,6 +1423,10 @@ impl<'a> SliceDecoder<'a> {
                 }
             }
         }
+        Ok(())
+    }
+
+    fn residual_cavlc_chroma(&mut self, r: &mut BitReader, cbp: u8) -> Result<()> {
         let cbp_c = cbp >> 4;
         if cbp_c & 3 != 0 {
             for c in 0..2 {

crates/hevc/src/inter.rs

@@ -128,7 +128,7 @@ pub fn mc_chroma(f: &Frame, c: usize, x: i32, y: i32, mv: [i16; 2], w: usize, h:
     let yi = y + (mv[1] as i32 >> 3);
     let shift1 = bd.min(12) - 8;
     let shift3 = 14 - bd;
-    f.chroma_window(c, xi - 1, yi - 1, w + 3, h + 3, &mut sc.win, WIN);
+    f.chroma_window(c, (xi - 1, yi - 1), (w + 3, h + 3), &mut sc.win, WIN);
     interpolate::<4>(&sc.win, &mut sc.tmp, fx, fy, &CHROMA_FILTER[fx], &CHROMA_FILTER[fy], w, h, shift1, shift3, out);
 }
 
@@ -158,7 +158,7 @@ pub fn put_bi(p0: &[i16], p1: &[i16], w: usize, h: usize, bd: u32, dst: &mut [u1
 }
 
 /// Explicit weighted prediction, one list (8-265). `o` is the offset already scaled to the bit depth.
-pub fn put_weighted_uni(p: &[i16], w: usize, h: usize, bd: u32, log2wd: u32, wt: i32, o: i32, dst: &mut [u16], ds: usize) {
+pub fn put_weighted_uni(p: &[i16], (w, h): (usize, usize), bd: u32, (log2wd, wt, o): (u32, i32, i32), dst: &mut [u16], ds: usize) {
     let max = (1i32 << bd) - 1;
     for r in 0..h {
         for c in 0..w {
@@ -173,14 +173,9 @@ pub fn put_weighted_uni(p: &[i16], w: usize, h: usize, bd: u32, log2wd: u32, wt:
 pub fn put_weighted_bi(
     p0: &[i16],
     p1: &[i16],
-    w: usize,
-    h: usize,
+    (w, h): (usize, usize),
     bd: u32,
-    log2wd: u32,
-    w0: i32,
-    w1: i32,
-    o0: i32,
-    o1: i32,
+    (log2wd, w0, w1, o0, o1): (u32, i32, i32, i32, i32),
     dst: &mut [u16],
     ds: usize,
 ) {

crates/hevc/src/mvpred.rs

@@ -36,9 +36,23 @@ fn same_motion(a: &MvField, b: &MvField) -> bool {
     true
 }
 
+/// Coding block and prediction block geometry (6.4.2): coding block (xc, yc, ncb), prediction block (xp, yp, w, h, part_idx).
+#[derive(Clone, Copy)]
+pub struct PbGeom {
+    pub xc: i32,
+    pub yc: i32,
+    pub ncb: i32,
+    pub xp: i32,
+    pub yp: i32,
+    pub w: i32,
+    pub h: i32,
+    pub part_idx: usize,
+}
+
 impl<'a> SliceDecoder<'a> {
     /// Prediction block availability (6.4.2).
-    fn pb_avail(&self, xc: i32, yc: i32, ncb: i32, xp: i32, yp: i32, w: i32, h: i32, part_idx: usize, xn: i32, yn: i32) -> bool {
+    fn pb_avail(&self, g: &PbGeom, xn: i32, yn: i32) -> bool {
+        let PbGeom { xc, yc, ncb, xp, yp, w, h, part_idx } = *g;
         let same_cb = xc <= xn && yc <= yn && xc + ncb > xn && yc + ncb > yn;
         let avail = if !same_cb {
             self.pic.zavail(xp, yp, xn, yn, self.slice_addr)
@@ -59,23 +73,13 @@ impl<'a> SliceDecoder<'a> {
     }
 
     /// Merge mode motion (8.5.3.2.2).
-    pub fn derive_merge(
-        &self,
-        xc: i32,
-        yc: i32,
-        ncb: i32,
-        xp0: i32,
-        yp0: i32,
-        w0: i32,
-        h0: i32,
-        part_idx0: usize,
-        part: PartMode,
-        merge_idx: usize,
-    ) -> MvField {
+    pub fn derive_merge(&self, g: PbGeom, part: PartMode, merge_idx: usize) -> MvField {
+        let PbGeom { xc, yc, ncb, xp: xp0, yp: yp0, w: w0, h: h0, part_idx: part_idx0 } = g;
         let pml = self.pps.log2_parallel_merge_level;
         let (xp, yp, w, h, part_idx) = if pml > 2 && ncb == 8 { (xc, yc, ncb, ncb, 0) } else { (xp0, yp0, w0, h0, part_idx0) };
+        let g = PbGeom { xc, yc, ncb, xp, yp, w, h, part_idx };
         let same_mer = |xn: i32, yn: i32| (xp >> pml) == (xn >> pml) && (yp >> pml) == (yn >> pml);
-        let avail = |xn: i32, yn: i32| self.pb_avail(xc, yc, ncb, xp, yp, w, h, part_idx, xn, yn) && !same_mer(xn, yn);
+        let avail = |xn: i32, yn: i32| self.pb_avail(&g, xn, yn) && !same_mer(xn, yn);
         let mut list: [MvField; 5] = [MvField::default(); 5];
         let mut n = 0;
         let max = self.sh.max_num_merge_cand as usize;
@@ -237,20 +241,8 @@ impl<'a> SliceDecoder<'a> {
     }
 
     /// AMVP predictor (8.5.3.2.6, 8.5.3.2.7).
-    pub fn derive_amvp(
-        &self,
-        xc: i32,
-        yc: i32,
-        ncb: i32,
-        xp: i32,
-        yp: i32,
-        w: i32,
-        h: i32,
-        part_idx: usize,
-        l: usize,
-        ref_idx: usize,
-        mvp_flag: usize,
-    ) -> [i16; 2] {
+    pub fn derive_amvp(&self, g: PbGeom, l: usize, ref_idx: usize, mvp_flag: usize) -> [i16; 2] {
+        let PbGeom { xp, yp, w, h, .. } = g;
         let target = &self.refs[l][ref_idx];
         let (tpoc, tlt) = (target.poc, target.long_term);
         let cur_poc = self.pic.poc;
@@ -277,10 +269,7 @@ impl<'a> SliceDecoder<'a> {
             pick.map(|(mv, r)| if r.poc != tpoc && !r.long_term && !tlt { scale_mv(mv, cur_poc - r.poc, cur_poc - tpoc) } else { mv })
         };
         let pa = [(xp - 1, yp + h), (xp - 1, yp + h - 1)];
-        let av_a = [
-            self.pb_avail(xc, yc, ncb, xp, yp, w, h, part_idx, pa[0].0, pa[0].1),
-            self.pb_avail(xc, yc, ncb, xp, yp, w, h, part_idx, pa[1].0, pa[1].1),
-        ];
+        let av_a = [self.pb_avail(&g, pa[0].0, pa[0].1), self.pb_avail(&g, pa[1].0, pa[1].1)];
         let is_scaled = av_a[0] || av_a[1];
         let mut a = None;
         for k in 0..2 {
@@ -296,7 +285,7 @@ impl<'a> SliceDecoder<'a> {
             }
         }
         let pb = [(xp + w, yp - 1), (xp + w - 1, yp - 1), (xp - 1, yp - 1)];
-        let av_b = [0, 1, 2].map(|k| self.pb_avail(xc, yc, ncb, xp, yp, w, h, part_idx, pb[k].0, pb[k].1));
+        let av_b = [0, 1, 2].map(|k| self.pb_avail(&g, pb[k].0, pb[k].1));
         let mut b = None;
         for k in 0..3 {
             if av_b[k] && b.is_none() {

crates/hevc/src/picture.rs

@@ -125,7 +125,7 @@ impl Frame {
     }
 
     /// Copy a w x h window of chroma component `c` (0 = Cb, 1 = Cr) with edge replication.
-    pub fn chroma_window(&self, c: usize, x0: i32, y0: i32, w: usize, h: usize, out: &mut [i16], os: usize) {
+    pub fn chroma_window(&self, c: usize, (x0, y0): (i32, i32), (w, h): (usize, usize), out: &mut [i16], os: usize) {
         let (cw, ch) = (self.cwidth as i32, self.cheight as i32);
         let sh = self.log2_ctb - 1;
         let mask = (1i32 << sh) - 1;

crates/hevc/src/slicedec.rs

@@ -7,6 +7,7 @@ use crate::cabac::{Cabac, Contexts, init_contexts};
 use crate::error::{Result, ensure, invalid};
 use crate::inter;
 use crate::intra::{self, Refs};
+use crate::mvpred::PbGeom;
 use crate::params::{Layout, Pps, Sps};
 use crate::picture::{FrameRef, RefPic};
 use crate::slice::{SliceHeader, SliceType};
@@ -678,7 +679,7 @@ impl<'a> SliceDecoder<'a> {
         self.mark_edges(x0, y0, n, n, E_TU_V | E_PU_V, E_TU_H | E_PU_H);
         if skip {
             self.cu_intra = false;
-            self.prediction_unit(x0, y0, n, x0, y0, n, n, 0, PartMode::P2Nx2N, true)?;
+            self.prediction_unit(PbGeom { xc: x0, yc: y0, ncb: n, xp: x0, yp: y0, w: n, h: n, part_idx: 0 }, PartMode::P2Nx2N, true)?;
             self.finish_cu(x0, y0, n);
             return Ok(());
         }
@@ -707,7 +708,7 @@ impl<'a> SliceDecoder<'a> {
             }
             self.parse_intra_modes(x0, y0, n, part)?;
             self.max_trafo_depth = self.sps.max_th_depth_intra + self.intra_split as u32;
-            self.transform_tree(x0, y0, x0, y0, log2, 0, 0, [true, true])?;
+            self.transform_tree((x0, y0), (x0, y0), log2, 0, 0, [true, true])?;
         } else {
             self.intra_split = false;
             let mut merge2nx2n = false;
@@ -724,7 +725,8 @@ impl<'a> SliceDecoder<'a> {
                 PartMode::PNxN => &[(0, 0, h, h), (h, 0, h, h), (0, h, h, h), (h, h, h, h)],
             };
             for (i, &(px, py, pw, ph)) in parts.iter().enumerate() {
-                let merge = self.prediction_unit(x0, y0, n, x0 + px, y0 + py, pw, ph, i, part, false)?;
+                let merge =
+                    self.prediction_unit(PbGeom { xc: x0, yc: y0, ncb: n, xp: x0 + px, yp: y0 + py, w: pw, h: ph, part_idx: i }, part, false)?;
                 if part == PartMode::P2Nx2N {
                     merge2nx2n = merge;
                 }
@@ -738,7 +740,7 @@ impl<'a> SliceDecoder<'a> {
             let root_cbf = if !merge2nx2n { self.dd(RQT_ROOT_CBF) == 1 } else { true };
             if root_cbf {
                 self.max_trafo_depth = self.sps.max_th_depth_inter;
-                self.transform_tree(x0, y0, x0, y0, log2, 0, 0, [true, true])?;
+                self.transform_tree((x0, y0), (x0, y0), log2, 0, 0, [true, true])?;
             }
         }
         self.finish_cu(x0, y0, n);
@@ -912,7 +914,7 @@ impl<'a> SliceDecoder<'a> {
         Ok(())
     }
 
-    fn transform_tree(&mut self, x0: i32, y0: i32, xb: i32, yb: i32, log2: u32, depth: u32, blk_idx: u32, parent_cbf_c: [bool; 2]) -> Result<()> {
+    fn transform_tree(&mut self, (x0, y0): (i32, i32), base: (i32, i32), log2: u32, depth: u32, blk_idx: u32, parent_cbf_c: [bool; 2]) -> Result<()> {
         let split =
             if log2 <= self.sps.log2_max_tb && log2 > self.sps.log2_min_tb && depth < self.max_trafo_depth && !(self.intra_split && depth == 0) {
                 self.dd(SPLIT_TRANSFORM + (5 - log2 as usize)) == 1
@@ -937,10 +939,10 @@ impl<'a> SliceDecoder<'a> {
         }
         if split {
             let h = 1 << (log2 - 1);
-            self.transform_tree(x0, y0, x0, y0, log2 - 1, depth + 1, 0, cbf_c)?;
-            self.transform_tree(x0 + h, y0, x0, y0, log2 - 1, depth + 1, 1, cbf_c)?;
-            self.transform_tree(x0, y0 + h, x0, y0, log2 - 1, depth + 1, 2, cbf_c)?;
-            self.transform_tree(x0 + h, y0 + h, x0, y0, log2 - 1, depth + 1, 3, cbf_c)?;
+            self.transform_tree((x0, y0), (x0, y0), log2 - 1, depth + 1, 0, cbf_c)?;
+            self.transform_tree((x0 + h, y0), (x0, y0), log2 - 1, depth + 1, 1, cbf_c)?;
+            self.transform_tree((x0, y0 + h), (x0, y0), log2 - 1, depth + 1, 2, cbf_c)?;
+            self.transform_tree((x0 + h, y0 + h), (x0, y0), log2 - 1, depth + 1, 3, cbf_c)?;
             return Ok(());
         }
         let cbf_luma =
@@ -949,7 +951,7 @@ impl<'a> SliceDecoder<'a> {
         if depth > 0 {
             self.mark_edges(x0, y0, n, n, E_TU_V, E_TU_H);
         }
-        self.transform_unit(x0, y0, xb, yb, log2, blk_idx, cbf_luma, cbf_c)
+        self.transform_unit((x0, y0), base, log2, blk_idx, cbf_luma, cbf_c)
     }
 
     /// cu_qp_delta_abs / sign and QP update (7.3.8.14).
@@ -979,7 +981,15 @@ impl<'a> SliceDecoder<'a> {
         Ok(())
     }
 
-    fn transform_unit(&mut self, x0: i32, y0: i32, xb: i32, yb: i32, log2: u32, blk_idx: u32, cbf_luma: bool, cbf_c: [bool; 2]) -> Result<()> {
+    fn transform_unit(
+        &mut self,
+        (x0, y0): (i32, i32),
+        (xb, yb): (i32, i32),
+        log2: u32,
+        blk_idx: u32,
+        cbf_luma: bool,
+        cbf_c: [bool; 2],
+    ) -> Result<()> {
         let n = 1i32 << log2;
         let chroma_here = log2 > 2 || blk_idx == 3;
         let cbf_chroma = cbf_c[0] || cbf_c[1];
@@ -1284,7 +1294,7 @@ impl<'a> SliceDecoder<'a> {
                 max_y = max_y.max(yc);
             }
         }
-        self.reconstruct(x0, y0, log2, c, qp, ts, max_x, max_y)
+        self.reconstruct((x0, y0), log2, c, qp, ts, (max_x, max_y))
     }
 
     fn coeff_abs_level_remaining(&mut self, rice: u32) -> Result<i32> {
@@ -1303,7 +1313,7 @@ impl<'a> SliceDecoder<'a> {
     }
 
     /// Scaling, transform and addition to the prediction.
-    fn reconstruct(&mut self, x0: i32, y0: i32, log2: u32, c: usize, qp: i32, ts: bool, max_x: usize, max_y: usize) -> Result<()> {
+    fn reconstruct(&mut self, (x0, y0): (i32, i32), log2: u32, c: usize, qp: i32, ts: bool, (max_x, max_y): (usize, usize)) -> Result<()> {
         let n = 1usize << log2;
         let bd = if c == 0 { self.pic.bd_y } else { self.pic.bd_c };
         let coeffs = &mut self.coeffs[..n * n];
@@ -1355,19 +1365,8 @@ impl<'a> SliceDecoder<'a> {
     }
 
     /// prediction_unit(): parse, derive motion, store it and predict. Returns merge_flag.
-    fn prediction_unit(
-        &mut self,
-        xc: i32,
-        yc: i32,
-        ncb: i32,
-        xp: i32,
-        yp: i32,
-        w: i32,
-        h: i32,
-        part_idx: usize,
-        part: PartMode,
-        skip: bool,
-    ) -> Result<bool> {
+    fn prediction_unit(&mut self, g: PbGeom, part: PartMode, skip: bool) -> Result<bool> {
+        let PbGeom { xc, yc, ncb, xp, yp, w, h, part_idx } = g;
         let merge = if skip { true } else { self.dd(MERGE_FLAG) == 1 };
         let mvf = if merge {
             let idx = if self.sh.max_num_merge_cand > 1 {
@@ -1382,7 +1381,7 @@ impl<'a> SliceDecoder<'a> {
             } else {
                 0
             };
-            self.derive_merge(xc, yc, ncb, xp, yp, w, h, part_idx, part, idx as usize)
+            self.derive_merge(PbGeom { xc, yc, ncb, xp, yp, w, h, part_idx }, part, idx as usize)
         } else {
             let pred_idc = if self.sh.is_b() {
                 if w + h != 12 {
@@ -1426,7 +1425,7 @@ impl<'a> SliceDecoder<'a> {
                 if ref_idx[l] < 0 {
                     continue;
                 }
-                let mvp = self.derive_amvp(xc, yc, ncb, xp, yp, w, h, part_idx, l, ref_idx[l] as usize, mvp_flag[l] as usize);
+                let mvp = self.derive_amvp(PbGeom { xc, yc, ncb, xp, yp, w, h, part_idx }, l, ref_idx[l] as usize, mvp_flag[l] as usize);
                 // mvLX = (mvp + mvd) wrapped to 16 bits (8-94)
                 f.mv[l] = [(mvp[0] as i32 + mvd[l][0]) as i16, (mvp[1] as i32 + mvd[l][1]) as i16];
             }
@@ -1524,11 +1523,11 @@ impl<'a> SliceDecoder<'a> {
             if bi {
                 let (w0, o0) = get(0);
                 let (w1, o1) = get(1);
-                inter::put_weighted_bi(&self.pred0, &self.pred1, bw, bh, bd, log2wd, w0, w1, o0, o1, dst, stride);
+                inter::put_weighted_bi(&self.pred0, &self.pred1, (bw, bh), bd, (log2wd, w0, w1, o0, o1), dst, stride);
             } else {
                 let l = if lists[0] { 0 } else { 1 };
                 let (wt, o) = get(l);
-                inter::put_weighted_uni(&self.pred0, bw, bh, bd, log2wd, wt, o, dst, stride);
+                inter::put_weighted_uni(&self.pred0, (bw, bh), bd, (log2wd, wt, o), dst, stride);
             }
         }
     }