clippy: reorder doc list lead-in before bullets in axis_data_range

git push origin fix/sonar-s3776-complex-b 2>&1 | tail -2

ROADMAP.md

@@ -137,6 +137,7 @@ app's JSON control channel (`deckcraft --control PORT`).
 
 | Date | What landed |
 |---|---|
+| 2026-10-25 | Windows installer: the Start menu shortcut takes the icon embedded in `deckcraft.exe` instead of an MSI `<Icon>`, so a taskbar pin keeps its icon after an update (#1114); a pin that already lost it has to be unpinned and pinned again once |
 | 2026-10-24 | AutoFormat as you type: a paragraph that starts with `*`, `-`, `--`, `>`, `<>`, `1.`, `1)`, `I.`, `i.`, `A.`, `a.` or `a)` and a space or tab becomes a bulleted or numbered list in that style; Backspace right after puts the typed text back, and Tools ▸ AutoCorrect and `text.autocorrect {lists}` turn it off |
 | 2026-10-24 | Tables: while the caret is in a cell, dragging into another cell selects that range of cells (shaded, merged cells taken in whole), and Shift+arrow grows or shrinks it; formatting applies to the selected cells only, and Delete clears their text and keeps the table |
 | 2026-10-24 | Preferences ▸ Measurement units: inches or centimetres. The Format pane's Size and Position fields, Shape Format ▸ Size on the ribbon, the Slide Size dialog and the ruler show and take the chosen unit; the deck still stores points (#1326) |
@@ -293,6 +294,7 @@ app's JSON control channel (`deckcraft --control PORT`).
 
 | Date | Change | Summary |
 |---|---|---|
+| 2026-10-25 | trivial | Progress log: Windows shortcut icon survives updates (#1114). Scores and milestone names unchanged |
 | 2026-10-24 | trivial | Progress log: selecting a range of table cells by dragging. Scores and milestone names unchanged |
 | 2026-10-24 | trivial | Progress log: measurement units in centimetres (#1326). Scores unchanged |
 | 2026-10-22 | trivial | Progress log: Animation Pane ▸ Play From. Scores unchanged |

apps/deckcraft/src/main.rs

@@ -430,4 +430,27 @@ mod tests {
         assert_eq!(mac, win, "macOS and Windows offer different file types");
         assert_eq!(mac, linux, "macOS and Linux offer different file types");
     }
+
+    /// #1114: a shortcut that names an MSI `<Icon>` shows a blank icon once an update deletes the
+    /// cached copy of it, and a taskbar pin (a copy of the shortcut) is never repaired. Shortcuts
+    /// take the icon embedded in `deckcraft.exe`; only `ARPPRODUCTICON` uses the MSI `<Icon>`.
+    #[test]
+    fn windows_shortcuts_use_the_icon_embedded_in_the_exe() {
+        let wxs = include_str!("../../../packaging/windows/deckcraft.wxs");
+        let shortcuts: Vec<&str> = wxs.split("<Shortcut").skip(1).filter_map(|s| s.split("/>").next()).collect();
+        assert!(!shortcuts.is_empty(), "deckcraft.wxs has no shortcut");
+        for shortcut in shortcuts {
+            assert!(!shortcut.contains("Icon="), "a shortcut names an MSI <Icon>:{shortcut}");
+        }
+        assert!(wxs.contains("<Property Id=\"ARPPRODUCTICON\""), "Apps and features keeps its icon");
+    }
+
+    /// The shortcuts take the exe's own icon, and `package.ps1 -SkipBuild` packages an existing
+    /// exe without `DECKCRAFT_REQUIRE_WINRES`, so the script has to refuse one with no icon.
+    #[test]
+    fn the_windows_package_script_rejects_an_exe_without_an_icon() {
+        let script = include_str!("../../../packaging/windows/package.ps1");
+        assert!(script.contains("ExtractIconEx"), "package.ps1 does not count the exe's icons");
+        assert!(script.contains("has no embedded icon"), "package.ps1 does not reject an iconless exe");
+    }
 }

contributors/people.toml

@@ -65,3 +65,6 @@ sync_github_name = true
 
 [people.Katacc]
 sync_github_name = true
+
+[people.SofiaAFigueredo]
+real_name = "Sofia Alves Figueredo"

crates/av1/src/cdef.rs

@@ -45,7 +45,7 @@ fn cdef_block(fs: &FrameState, out: &mut FrameBuf, r: usize, c: usize, idx: usiz
     let var_str = if (var >> 6) != 0 { (31 - ((var >> 6) as u32).leading_zeros()).min(12) as i32 } else { 0 };
     pri = if var != 0 { (pri * (4 + var_str) + 8) >> 4 } else { 0 };
     let damping = cp.damping as i32 + coeff_shift as i32;
-    filter(fs, out, 0, r, c, pri, sec, damping, dir);
+    filter(fs, out, 0, (r, c), (pri, sec), damping, dir);
     if fs.num_planes == 1 {
         return true;
     }
@@ -53,8 +53,8 @@ fn cdef_block(fs: &FrameState, out: &mut FrameBuf, r: usize, c: usize, idx: usiz
     let sec = (cp.uv_sec[idx] << coeff_shift) as i32;
     let dir = if pri == 0 { 0 } else { CDEF_UV_DIR[fs.ssx][fs.ssy][y_dir] as usize };
     let damping = cp.damping as i32 + coeff_shift as i32 - 1;
-    filter(fs, out, 1, r, c, pri, sec, damping, dir);
-    filter(fs, out, 2, r, c, pri, sec, damping, dir);
+    filter(fs, out, 1, (r, c), (pri, sec), damping, dir);
+    filter(fs, out, 2, (r, c), (pri, sec), damping, dir);
     true
 }
 
@@ -124,33 +124,74 @@ fn constrain(diff: i32, threshold: i32, adj: u32) -> i32 {
     if diff < 0 { -mag } else { mag }
 }
 
-#[allow(clippy::too_many_arguments)]
-fn filter(fs: &FrameState, out: &mut FrameBuf, plane: usize, r: usize, c: usize, pri: i32, sec: i32, damping: i32, dir: usize) {
-    if pri == 0 && sec == 0 {
-        // Every tap contributes 0 and the clamp to [min, max] keeps x: CdefFrame = CurrFrame.
-        return;
+/// Tap offsets, weights and strengths of one block.
+struct Taps {
+    pri: i32,
+    sec: i32,
+    pri_adj: u32,
+    sec_adj: u32,
+    pri_off: [isize; 2],
+    sec_off: [[isize; 2]; 2],
+    pri_taps: [i32; 2],
+    sec_taps: [i32; 2],
+}
+
+impl Taps {
+    fn new(coeff_shift: u32, (pri, sec): (i32, i32), damping: i32, dir: usize) -> Taps {
+        let pri_tap = ((pri >> coeff_shift) & 1) as usize;
+        let off = |d: usize, k: usize| CDEF_DIRECTIONS[d][k][0] as isize * PAD_STRIDE as isize + CDEF_DIRECTIONS[d][k][1] as isize;
+        let pri_off = [off(dir, 0), off(dir, 1)];
+        let d_lo = (dir + 6) & 7;
+        let d_hi = (dir + 2) & 7;
+        let sec_off = [[off(d_lo, 0), off(d_hi, 0)], [off(d_lo, 1), off(d_hi, 1)]];
+        let pri_taps = [CDEF_PRI_TAPS[pri_tap][0] as i32, CDEF_PRI_TAPS[pri_tap][1] as i32];
+        let sec_taps = [CDEF_SEC_TAPS[pri_tap][0] as i32, CDEF_SEC_TAPS[pri_tap][1] as i32];
+        let adj = |t: i32| if t == 0 { 0 } else { (damping - (31 - (t as u32).leading_zeros()) as i32).max(0) as u32 };
+        Taps { pri, sec, pri_adj: adj(pri), sec_adj: adj(sec), pri_off, sec_off, pri_taps, sec_taps }
     }
-    let coeff_shift = fs.bit_depth - 8;
+
+    /// The 12 (offset, weight, strength, adjustment) taps in filtering order.
+    fn tap_list(&self) -> [(isize, i32, i32, u32); 12] {
+        let mut taps = [(0isize, 0i32, 0i32, 0u32); 12];
+        let mut n = 0;
+        for k in 0..2 {
+            for sign in [-1isize, 1] {
+                taps[n] = (sign * self.pri_off[k], self.pri_taps[k], self.pri, self.pri_adj);
+                taps[n + 1] = (sign * self.sec_off[k][0], self.sec_taps[k], self.sec, self.sec_adj);
+                taps[n + 2] = (sign * self.sec_off[k][1], self.sec_taps[k], self.sec, self.sec_adj);
+                n += 3;
+            }
+        }
+        taps
+    }
+}
+
+/// Position and size of a block within its plane.
+#[derive(Clone, Copy)]
+struct BlockRect {
+    x0: usize,
+    y0: usize,
+    w: usize,
+    h: usize,
+}
+
+/// The block with two samples of padding on every side, plus whether every sample is available
+/// (CdefAvailable: its 4x4 block is inside the frame).
+fn padded_block(fs: &FrameState, plane: usize, src: &Plane, rect: BlockRect) -> ([i32; PAD_STRIDE * PAD_STRIDE], bool) {
     let (sub_x, sub_y) = if plane > 0 { (fs.ssx, fs.ssy) } else { (0, 0) };
-    let x0 = (c * 4) >> sub_x;
-    let y0 = (r * 4) >> sub_y;
-    let w = 8 >> sub_x;
-    let h = 8 >> sub_y;
-    let src = &fs.cur.planes[plane];
     let (mi_rows, mi_cols) = (fs.fh.mi_rows as isize, fs.fh.mi_cols as isize);
-    // CdefAvailable per sample: its 4x4 block is inside the frame.
     let mut buf = [CDEF_NA; PAD_STRIDE * PAD_STRIDE];
     let mut all_avail = true;
-    for i in 0..h + 4 {
-        let y = (y0 + i) as isize - 2;
+    for i in 0..rect.h + 4 {
+        let y = (rect.y0 + i) as isize - 2;
         let cand_r = (y << sub_y) >> 2;
         if y < 0 || cand_r >= mi_rows {
             all_avail = false;
             continue;
         }
         let row = src.row(y as usize);
-        for j in 0..w + 4 {
-            let x = (x0 + j) as isize - 2;
+        for j in 0..rect.w + 4 {
+            let x = (rect.x0 + j) as isize - 2;
             let cand_c = (x << sub_x) >> 2;
             if x >= 0 && cand_c < mi_cols {
                 buf[i * PAD_STRIDE + j] = row[x as usize] as i32;
@@ -159,70 +200,70 @@ fn filter(fs: &FrameState, out: &mut FrameBuf, plane: usize, r: usize, c: usize,
             }
         }
     }
-    let pri_tap = ((pri >> coeff_shift) & 1) as usize;
-    let off = |d: usize, k: usize| CDEF_DIRECTIONS[d][k][0] as isize * PAD_STRIDE as isize + CDEF_DIRECTIONS[d][k][1] as isize;
-    let pri_off = [off(dir, 0), off(dir, 1)];
-    let d_lo = (dir + 6) & 7;
-    let d_hi = (dir + 2) & 7;
-    let sec_off = [[off(d_lo, 0), off(d_hi, 0)], [off(d_lo, 1), off(d_hi, 1)]];
-    let pri_taps = [CDEF_PRI_TAPS[pri_tap][0] as i32, CDEF_PRI_TAPS[pri_tap][1] as i32];
-    let sec_taps = [CDEF_SEC_TAPS[pri_tap][0] as i32, CDEF_SEC_TAPS[pri_tap][1] as i32];
-    let adj = |t: i32| if t == 0 { 0 } else { (damping - (31 - (t as u32).leading_zeros()) as i32).max(0) as u32 };
-    let (pri_adj, sec_adj) = (adj(pri), adj(sec));
+    (buf, all_avail)
+}
+
+/// Filters one sample whose taps may be unavailable (CDEF_NA).
+fn filter_sample(buf: &[i32; PAD_STRIDE * PAD_STRIDE], taps: &Taps, ci: isize, x: i32) -> u16 {
+    let mut sum = 0i32;
+    let mut max = x;
+    let mut min = x;
+    for k in 0..2 {
+        for sign in [-1isize, 1] {
+            let p = buf[(ci + sign * taps.pri_off[k]) as usize];
+            if p != CDEF_NA {
+                if taps.pri != 0 {
+                    sum += taps.pri_taps[k] * constrain(p - x, taps.pri, taps.pri_adj);
+                }
+                max = max.max(p);
+                min = min.min(p);
+            }
+            for s in 0..2 {
+                let q = buf[(ci + sign * taps.sec_off[k][s]) as usize];
+                if q != CDEF_NA {
+                    if taps.sec != 0 {
+                        sum += taps.sec_taps[k] * constrain(q - x, taps.sec, taps.sec_adj);
+                    }
+                    max = max.max(q);
+                    min = min.min(q);
+                }
+            }
+        }
+    }
+    (x + ((8 + sum - (sum < 0) as i32) >> 4)).clamp(min, max) as u16
+}
+
+fn filter(fs: &FrameState, out: &mut FrameBuf, plane: usize, (r, c): (usize, usize), strengths: (i32, i32), damping: i32, dir: usize) {
+    let (pri, sec) = strengths;
+    if pri == 0 && sec == 0 {
+        // Every tap contributes 0 and the clamp to [min, max] keeps x: CdefFrame = CurrFrame.
+        return;
+    }
+    let coeff_shift = fs.bit_depth - 8;
+    let (sub_x, sub_y) = if plane > 0 { (fs.ssx, fs.ssy) } else { (0, 0) };
+    let rect = BlockRect { x0: (c * 4) >> sub_x, y0: (r * 4) >> sub_y, w: 8 >> sub_x, h: 8 >> sub_y };
+    let src = &fs.cur.planes[plane];
+    let (buf, all_avail) = padded_block(fs, plane, src, rect);
+    let taps = Taps::new(coeff_shift, strengths, damping, dir);
     let dst = &mut out.planes[plane];
     if all_avail {
         // Every tap is available: whole rows at a time (constrain() with threshold 0 is 0).
-        let mut taps = [(0isize, 0i32, 0i32, 0u32); 12];
-        let mut n = 0;
-        for k in 0..2 {
-            for sign in [-1isize, 1] {
-                taps[n] = (sign * pri_off[k], pri_taps[k], pri, pri_adj);
-                taps[n + 1] = (sign * sec_off[k][0], sec_taps[k], sec, sec_adj);
-                taps[n + 2] = (sign * sec_off[k][1], sec_taps[k], sec, sec_adj);
-                n += 3;
-            }
-        }
-        let rows = (y0, x0, h);
-        match w {
-            8 => filter_rows::<8>(&buf, &taps, dst, rows),
-            _ => filter_rows::<4>(&buf, &taps, dst, rows),
+        let list = taps.tap_list();
+        let rows = (rect.y0, rect.x0, rect.h);
+        match rect.w {
+            8 => filter_rows::<8>(&buf, &list, dst, rows),
+            _ => filter_rows::<4>(&buf, &list, dst, rows),
         }
         return;
     }
-    for i in 0..h {
+    for i in 0..rect.h {
         // The centre sample is read even where CdefAvailable is 0 (8x8 blocks straddling an odd
         // MiRows / MiCols edge).
-        let srow = &src.row(y0 + i)[x0..x0 + w];
-        let drow = &mut dst.row_mut(y0 + i)[x0..x0 + w];
-        for j in 0..w {
+        let srow = &src.row(rect.y0 + i)[rect.x0..rect.x0 + rect.w];
+        let drow = &mut dst.row_mut(rect.y0 + i)[rect.x0..rect.x0 + rect.w];
+        for j in 0..rect.w {
             let ci = ((i + 2) * PAD_STRIDE + j + 2) as isize;
-            let x = srow[j] as i32;
-            let mut sum = 0i32;
-            let mut max = x;
-            let mut min = x;
-            for k in 0..2 {
-                for sign in [-1isize, 1] {
-                    let p = buf[(ci + sign * pri_off[k]) as usize];
-                    if p != CDEF_NA {
-                        if pri != 0 {
-                            sum += pri_taps[k] * constrain(p - x, pri, pri_adj);
-                        }
-                        max = max.max(p);
-                        min = min.min(p);
-                    }
-                    for s in 0..2 {
-                        let q = buf[(ci + sign * sec_off[k][s]) as usize];
-                        if q != CDEF_NA {
-                            if sec != 0 {
-                                sum += sec_taps[k] * constrain(q - x, sec, sec_adj);
-                            }
-                            max = max.max(q);
-                            min = min.min(q);
-                        }
-                    }
-                }
-            }
-            drow[j] = (x + ((8 + sum - (sum < 0) as i32) >> 4)).clamp(min, max) as u16;
+            drow[j] = filter_sample(&buf, &taps, ci, srow[j] as i32);
         }
     }
 }

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/restoration.rs

@@ -73,57 +73,70 @@ impl LrCtx<'_> {
     }
 }
 
+/// Position and size of a (at most 4x4) restoration block within its plane.
+#[derive(Clone, Copy)]
+struct LrBlock {
+    x: i32,
+    y: i32,
+    w: i32,
+    h: i32,
+}
+
 /// Loop restoration (7.17): returns LrFrame given UpscaledCurrFrame and UpscaledCdefFrame.
 pub(crate) fn loop_restoration(fs: &FrameState, cur: &FrameBuf, cdef: &FrameBuf) -> FrameBuf {
-    let fh = &fs.fh;
     let mut lr = cdef.clone();
-    let up_w = fh.upscaled_width as i32;
-    let fr_h = fh.frame_height as i32;
-    let bd = fs.bit_depth;
     for plane in 0..fs.num_planes {
-        let rtype_frame = fh.lr.frame_restoration_type[plane];
-        if rtype_frame == RESTORE_NONE as u8 {
+        if fs.fh.lr.frame_restoration_type[plane] == RESTORE_NONE as u8 {
             continue;
         }
-        let (sub_x, sub_y) = if plane > 0 { (fs.ssx as i32, fs.ssy as i32) } else { (0, 0) };
-        let unit_size = fh.lr.loop_restoration_size[plane] as i32;
-        let unit_rows = fs.lr_unit_rows[plane] as i32;
-        let unit_cols = fs.lr_unit_cols[plane] as i32;
-        let plane_end_x = ((up_w + sub_x) >> sub_x) - 1;
-        let plane_end_y = ((fr_h + sub_y) >> sub_y) - 1;
-        let mut row = 0;
-        while row * 4 < fr_h {
-            let luma_y = row * 4;
-            let stripe_num = (luma_y + 8) / 64;
-            let stripe_start_y = (-8 + stripe_num * 64) >> sub_y;
-            let stripe_end_y = stripe_start_y + (64 >> sub_y) - 1;
-            let ctx = LrCtx { cur: &cur.planes[plane], cdef: &cdef.planes[plane], plane_end_x, plane_end_y, stripe_start_y, stripe_end_y };
-            let unit_row = (unit_rows - 1).min(((row * 4 + 8) >> sub_y) / unit_size);
-            let mut col = 0;
-            while col * 4 < up_w {
-                let unit_col = (unit_cols - 1).min(((col * 4) >> sub_x) / unit_size);
-                let x = (col * 4) >> sub_x;
-                let y = (row * 4) >> sub_y;
-                let w = (4 >> sub_x).min(plane_end_x - x + 1);
-                let h = (4 >> sub_y).min(plane_end_y - y + 1);
-                if w > 0 && h > 0 {
-                    let ui = (unit_row * unit_cols + unit_col) as usize;
-                    match fs.lr.lr_type[plane][ui] {
-                        t if t == RESTORE_WIENER as u8 => wiener(fs, &ctx, &mut lr.planes[plane], plane, ui, x, y, w, h, bd),
-                        t if t == RESTORE_SGRPROJ as u8 => self_guided(fs, &ctx, &mut lr.planes[plane], plane, ui, x, y, w, h, bd),
-                        _ => {}
-                    }
-                }
-                col += 1;
-            }
-            row += 1;
-        }
+        restore_plane(fs, cur, cdef, &mut lr.planes[plane], plane);
     }
     lr
 }
 
-#[allow(clippy::too_many_arguments)]
-fn wiener(fs: &FrameState, ctx: &LrCtx, out: &mut Plane, plane: usize, ui: usize, x: i32, y: i32, w: i32, h: i32, bd: u32) {
+fn restore_plane(fs: &FrameState, cur: &FrameBuf, cdef: &FrameBuf, out: &mut Plane, plane: usize) {
+    let fh = &fs.fh;
+    let up_w = fh.upscaled_width as i32;
+    let fr_h = fh.frame_height as i32;
+    let bd = fs.bit_depth;
+    let (sub_x, sub_y) = if plane > 0 { (fs.ssx as i32, fs.ssy as i32) } else { (0, 0) };
+    let unit_size = fh.lr.loop_restoration_size[plane] as i32;
+    let unit_rows = fs.lr_unit_rows[plane] as i32;
+    let unit_cols = fs.lr_unit_cols[plane] as i32;
+    let plane_end_x = ((up_w + sub_x) >> sub_x) - 1;
+    let plane_end_y = ((fr_h + sub_y) >> sub_y) - 1;
+    let mut row = 0;
+    while row * 4 < fr_h {
+        let luma_y = row * 4;
+        let stripe_num = (luma_y + 8) / 64;
+        let stripe_start_y = (-8 + stripe_num * 64) >> sub_y;
+        let stripe_end_y = stripe_start_y + (64 >> sub_y) - 1;
+        let ctx = LrCtx { cur: &cur.planes[plane], cdef: &cdef.planes[plane], plane_end_x, plane_end_y, stripe_start_y, stripe_end_y };
+        let unit_row = (unit_rows - 1).min(((row * 4 + 8) >> sub_y) / unit_size);
+        let mut col = 0;
+        while col * 4 < up_w {
+            let unit_col = (unit_cols - 1).min(((col * 4) >> sub_x) / unit_size);
+            let x = (col * 4) >> sub_x;
+            let y = (row * 4) >> sub_y;
+            let w = (4 >> sub_x).min(plane_end_x - x + 1);
+            let h = (4 >> sub_y).min(plane_end_y - y + 1);
+            if w > 0 && h > 0 {
+                let ui = (unit_row * unit_cols + unit_col) as usize;
+                let blk = LrBlock { x, y, w, h };
+                match fs.lr.lr_type[plane][ui] {
+                    t if t == RESTORE_WIENER as u8 => wiener(fs, &ctx, out, (plane, ui), blk, bd),
+                    t if t == RESTORE_SGRPROJ as u8 => self_guided(fs, &ctx, out, (plane, ui), blk, bd),
+                    _ => {}
+                }
+            }
+            col += 1;
+        }
+        row += 1;
+    }
+}
+
+fn wiener(fs: &FrameState, ctx: &LrCtx, out: &mut Plane, (plane, ui): (usize, usize), blk: LrBlock, bd: u32) {
+    let LrBlock { x, y, w, h } = blk;
     let round0: u32 = if bd == 12 { 5 } else { 3 };
     let round1: u32 = if bd == 12 { 9 } else { 11 };
     let coef = &fs.lr.lr_wiener[plane][ui];
@@ -166,13 +179,13 @@ fn wiener(fs: &FrameState, ctx: &LrCtx, out: &mut Plane, plane: usize, ui: usize
     }
 }
 
-#[allow(clippy::too_many_arguments)]
-fn self_guided(fs: &FrameState, ctx: &LrCtx, out: &mut Plane, plane: usize, ui: usize, x: i32, y: i32, w: i32, h: i32, bd: u32) {
+fn self_guided(fs: &FrameState, ctx: &LrCtx, out: &mut Plane, (plane, ui): (usize, usize), blk: LrBlock, bd: u32) {
+    let LrBlock { x, y, w, h } = blk;
     let set = fs.lr.lr_sgr_set[plane][ui] as usize;
     let mut flt0 = [[0i32; 4]; 4];
     let mut flt1 = [[0i32; 4]; 4];
-    box_filter(ctx, x, y, w, h, set, 0, bd, &mut flt0);
-    box_filter(ctx, x, y, w, h, set, 1, bd, &mut flt1);
+    box_filter(ctx, blk, (set, 0), bd, &mut flt0);
+    box_filter(ctx, blk, (set, 1), bd, &mut flt1);
     let xqd = fs.lr.lr_sgr_xqd[plane][ui];
     let w0 = xqd[0] as i32;
     let w1 = xqd[1] as i32;
@@ -198,8 +211,56 @@ fn round2_64(x: i64, n: u32) -> i64 {
     if n == 0 { x } else { (x + (1i64 << (n - 1))) >> n }
 }
 
-#[allow(clippy::too_many_arguments)]
-fn box_filter(ctx: &LrCtx, x: i32, y: i32, w: i32, h: i32, set: usize, pass: usize, bd: u32, f: &mut [[i32; 4]; 4]) {
+/// Box sums of the samples within `r` of (x, y): (sum of squares, sum).
+fn box_sums(ctx: &LrCtx, x: i32, y: i32, r: i32) -> (i64, i64) {
+    let mut a = 0i64;
+    let mut b = 0i64;
+    for dy in -r..=r {
+        for dx in -r..=r {
+            let c = ctx.sample(x + dx, y + dy) as i64;
+            a += c * c;
+            b += c;
+        }
+    }
+    (a, b)
+}
+
+/// Self-guided filter constants of one pass.
+struct SgrConsts {
+    n: i64,
+    s: i64,
+    one_over_n: i64,
+}
+
+/// (A, B) coefficients of one sample from its box sums.
+fn sgr_coefficients(a: i64, b: i64, k: &SgrConsts, bd: u32) -> (i64, i64) {
+    let a = round2_64(a, 2 * (bd - 8));
+    let d = round2_64(b, bd - 8);
+    let p = (a * k.n - d * d).max(0);
+    let z = round2_64(p * k.s, SGRPROJ_MTABLE_BITS as u32);
+    let a2 = if z >= 255 {
+        256
+    } else if z == 0 {
+        1
+    } else {
+        ((z << SGRPROJ_SGR_BITS) + z / 2) / (z + 1)
+    };
+    let b2 = ((1i64 << SGRPROJ_SGR_BITS) - a2) * b * k.one_over_n;
+    (a2, round2_64(b2, SGRPROJ_RECIP_BITS as u32))
+}
+
+fn box_weight(pass: usize, i: i32, dx: i32, dy: i32) -> i64 {
+    if pass == 0 {
+        if ((i + dy) & 1) != 0 { if dx == 0 { 6 } else { 5 } } else { 0 }
+    } else if dx == 0 || dy == 0 {
+        4
+    } else {
+        3
+    }
+}
+
+fn box_filter(ctx: &LrCtx, blk: LrBlock, (set, pass): (usize, usize), bd: u32, f: &mut [[i32; 4]; 4]) {
+    let LrBlock { x, y, w, h } = blk;
     let r = SGR_PARAMS[set][pass * 2] as i32;
     if r == 0 {
         return;
@@ -207,35 +268,15 @@ fn box_filter(ctx: &LrCtx, x: i32, y: i32, w: i32, h: i32, set: usize, pass: usi
     let eps = SGR_PARAMS[set][pass * 2 + 1] as i64;
     let n = ((2 * r + 1) * (2 * r + 1)) as i64;
     let n2e = n * n * eps;
-    let s = ((1i64 << SGRPROJ_MTABLE_BITS) + n2e / 2) / n2e;
-    let one_over_n = ((1i64 << SGRPROJ_RECIP_BITS) + n / 2) / n;
+    let k = SgrConsts { n, s: ((1i64 << SGRPROJ_MTABLE_BITS) + n2e / 2) / n2e, one_over_n: ((1i64 << SGRPROJ_RECIP_BITS) + n / 2) / n };
     let mut a_arr = [[0i64; 6]; 6];
     let mut b_arr = [[0i64; 6]; 6];
     for i in -1..h + 1 {
         for j in -1..w + 1 {
-            let mut a = 0i64;
-            let mut b = 0i64;
-            for dy in -r..=r {
-                for dx in -r..=r {
-                    let c = ctx.sample(x + j + dx, y + i + dy) as i64;
-                    a += c * c;
-                    b += c;
-                }
-            }
-            let a = round2_64(a, 2 * (bd - 8));
-            let d = round2_64(b, bd - 8);
-            let p = (a * n - d * d).max(0);
-            let z = round2_64(p * s, SGRPROJ_MTABLE_BITS as u32);
-            let a2 = if z >= 255 {
-                256
-            } else if z == 0 {
-                1
-            } else {
-                ((z << SGRPROJ_SGR_BITS) + z / 2) / (z + 1)
-            };
-            let b2 = ((1i64 << SGRPROJ_SGR_BITS) - a2) * b * one_over_n;
+            let (a, b) = box_sums(ctx, x + j, y + i, r);
+            let (a2, b2) = sgr_coefficients(a, b, &k, bd);
             a_arr[(i + 1) as usize][(j + 1) as usize] = a2;
-            b_arr[(i + 1) as usize][(j + 1) as usize] = round2_64(b2, SGRPROJ_RECIP_BITS as u32);
+            b_arr[(i + 1) as usize][(j + 1) as usize] = b2;
         }
     }
     for i in 0..h {
@@ -245,13 +286,7 @@ fn box_filter(ctx: &LrCtx, x: i32, y: i32, w: i32, h: i32, set: usize, pass: usi
             let mut b = 0i64;
             for dy in -1..=1i32 {
                 for dx in -1..=1i32 {
-                    let weight = if pass == 0 {
-                        if ((i + dy) & 1) != 0 { if dx == 0 { 6 } else { 5 } } else { 0 }
-                    } else if dx == 0 || dy == 0 {
-                        4
-                    } else {
-                        3
-                    };
+                    let weight = box_weight(pass, i, dx, dy);
                     a += weight * a_arr[(i + dy + 1) as usize][(j + dx + 1) as usize];
                     b += weight * b_arr[(i + dy + 1) as usize][(j + dx + 1) as usize];
                 }

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/engine/src/cmd/shape.rs

@@ -5,7 +5,7 @@ use std::collections::HashMap;
 
 use deckcraft_color::{ColorTransform, Rgba, SchemeSlot};
 use deckcraft_model::style::{Dash, Glow, Gradient, GradientShape, GradientStop, LineEnd, PatternFill, PictureFill, PictureMode, Reflection, Shadow};
-use deckcraft_model::{ColorRef, Effects, Fill, Geom, Line, ShapeStyle};
+use deckcraft_model::{ColorRef, Effects, Fill, Geom, Line, ShapeId, ShapeStyle};
 use serde_json::{Value, json};
 
 use super::{
@@ -556,81 +556,102 @@ fn edit_group_appearance(s: &mut Session, p: &Value, cmd: &str, f: impl Fn(&mut
     edit_shapes(s, p, cmd, |sh| apply(sh, 0, &f))
 }
 
-fn fill(s: &mut Session, p: &Value) -> Result<Value> {
-    let new_fill: Option<Option<Fill>> = if bool_or(p, "reset", false) {
-        Some(None)
-    } else if bool_or(p, "none", false) {
-        Some(Some(Fill::None))
-    } else if let Some(c) = color_param(p, "color") {
-        Some(Some(Fill::solid(c)))
-    } else if let Some(g) = p.get("gradient") {
-        Some(Some(Fill::Gradient(
-            gradient_param(g).ok_or_else(|| bad("shape.fill", "gradient needs at least two `stops`: [[0, \"#fff\"], [1, \"#000\"]]"))?,
-        )))
-    } else if let Some(pt) = p.get("pattern") {
-        Some(Some(Fill::Pattern(PatternFill {
+/// The fill a `shape.fill` request sets: `Some(None)` clears it, `Some(Some(_))` sets it, and `None`
+/// leaves it as it is.
+fn requested_fill(s: &mut Session, p: &Value) -> Result<Option<Option<Fill>>> {
+    if bool_or(p, "reset", false) {
+        return Ok(Some(None));
+    }
+    if bool_or(p, "none", false) {
+        return Ok(Some(Some(Fill::None)));
+    }
+    if let Some(c) = color_param(p, "color") {
+        return Ok(Some(Some(Fill::solid(c))));
+    }
+    if let Some(g) = p.get("gradient") {
+        let gradient = gradient_param(g).ok_or_else(|| bad("shape.fill", "gradient needs at least two `stops`: [[0, \"#fff\"], [1, \"#000\"]]"))?;
+        return Ok(Some(Some(Fill::Gradient(gradient))));
+    }
+    if let Some(pt) = p.get("pattern") {
+        return Ok(Some(Some(Fill::Pattern(PatternFill {
             preset: pt.get("preset").and_then(Value::as_str).unwrap_or("pct50").to_string(),
             fg: pt.get("fg").and_then(color_value).unwrap_or(ColorRef::scheme(SchemeSlot::Accent1)),
             bg: pt.get("bg").and_then(color_value).unwrap_or(ColorRef::scheme(SchemeSlot::Bg1)),
-        })))
-    } else if p.get("picture").is_some() {
+        }))));
+    }
+    if p.get("picture").is_some() {
         let (name, bytes) = super::insert::media_bytes(&with_param(p, "data", p.get("picture").cloned().unwrap_or_default()), "shape.fill")?;
         let ct = super::insert::content_type(&name, &bytes);
         let media = s.edit(|doc, _| doc.add_media(&name, ct, bytes).ok_or_else(|| bad("shape.fill", NO_ID_SPACE)))?;
-        Some(Some(Fill::Picture(PictureFill { media, ..Default::default() })))
-    } else {
-        None
-    };
+        return Ok(Some(Some(Fill::Picture(PictureFill { media, ..Default::default() }))));
+    }
+    Ok(None)
+}
+
+/// The effective fill of each targeted shape, for a transparency change on an inherited (style) fill.
+fn inherited_fills(s: &Session, p: &Value) -> Result<HashMap<ShapeId, Option<Fill>>> {
+    let st = s.doc()?;
+    let ctx = super::text::ctx_for(&st.doc, &st.selection);
+    let mut fills = HashMap::new();
+    for id in targets(s, p)? {
+        if let Some(sh) = st.shape(id) {
+            deckcraft_model::walk(std::slice::from_ref(sh), &mut |sh, _| {
+                let fill = ctx.as_ref().and_then(|c| {
+                    let (f, ph) = deckcraft_model::resolve::fill(c, sh);
+                    f.map(|f| match f {
+                        Fill::Solid { color } => Fill::solid(ColorRef::rgb(c.color(&color, ph))),
+                        other => other,
+                    })
+                });
+                fills.insert(sh.id, fill);
+            });
+        }
+    }
+    Ok(fills)
+}
+
+/// Applies a `shape.fill` request to one shape: the fill, then its picture options, then transparency.
+fn apply_fill(
+    sh: &mut deckcraft_model::Shape,
+    new_fill: Option<&Option<Fill>>,
+    pic_props: Option<&Value>,
+    transparency: Option<f64>,
+    inherited: &HashMap<ShapeId, Option<Fill>>,
+) -> Result<()> {
+    if let Some(f) = new_fill {
+        sh.fill.clone_from(f);
+    }
+    if let (Some(pp), Some(Fill::Picture(pf))) = (pic_props, sh.fill.as_mut()) {
+        picture_props(pf, pp);
+    }
+    if let (Some(t), Some(Fill::Picture(pf))) = (transparency, sh.fill.as_mut()) {
+        let a = 1.0 - t.clamp(0.0, 1.0);
+        pf.alpha = (a < 1.0).then_some(a);
+        return Ok(());
+    }
+    if let Some(t) = transparency {
+        let base = sh.fill.clone().or_else(|| inherited.get(&sh.id).cloned().flatten());
+        if let Some(Fill::Solid { mut color }) = base {
+            color.mods.retain(|m| !matches!(m, ColorTransform::Alpha(_)));
+            if t > 0.0 {
+                color.mods.push(ColorTransform::Alpha(((1.0 - t.clamp(0.0, 1.0)) * 100_000.0).round() as i32));
+            }
+            sh.fill = Some(Fill::Solid { color });
+        }
+    }
+    Ok(())
+}
+
+fn fill(s: &mut Session, p: &Value) -> Result<Value> {
+    let new_fill = requested_fill(s, p)?;
     let transparency = f64_param(p, "transparency");
     let pic_props = p.get("pictureProps").filter(|v| v.is_object());
     if new_fill.is_none() && transparency.is_none() && pic_props.is_none() {
         return Err(bad("shape.fill", "give `color`, `none`, `gradient`, `pattern`, `picture`, `pictureProps`, `transparency` or `reset`"));
     }
     // Transparency on an inherited (style) fill needs the effective colour.
-    let st = s.doc()?;
-    let ctx_fill = {
-        let ctx = super::text::ctx_for(&st.doc, &st.selection);
-        let mut fills = HashMap::new();
-        for id in targets(s, p)? {
-            if let Some(sh) = st.shape(id) {
-                deckcraft_model::walk(std::slice::from_ref(sh), &mut |sh, _| {
-                    let fill = ctx.as_ref().and_then(|c| {
-                        let (f, ph) = deckcraft_model::resolve::fill(c, sh);
-                        f.map(|f| match f {
-                            Fill::Solid { color } => Fill::solid(ColorRef::rgb(c.color(&color, ph))),
-                            other => other,
-                        })
-                    });
-                    fills.insert(sh.id, fill);
-                });
-            }
-        }
-        fills
-    };
-    edit_group_appearance(s, p, "shape.fill", |sh| {
-        if let Some(f) = &new_fill {
-            sh.fill.clone_from(f);
-        }
-        if let (Some(pp), Some(Fill::Picture(pf))) = (pic_props, sh.fill.as_mut()) {
-            picture_props(pf, pp);
-        }
-        if let (Some(t), Some(Fill::Picture(pf))) = (transparency, sh.fill.as_mut()) {
-            let a = 1.0 - t.clamp(0.0, 1.0);
-            pf.alpha = (a < 1.0).then_some(a);
-            return Ok(());
-        }
-        if let Some(t) = transparency {
-            let base = sh.fill.clone().or_else(|| ctx_fill.get(&sh.id).cloned().flatten());
-            if let Some(Fill::Solid { mut color }) = base {
-                color.mods.retain(|m| !matches!(m, ColorTransform::Alpha(_)));
-                if t > 0.0 {
-                    color.mods.push(ColorTransform::Alpha(((1.0 - t.clamp(0.0, 1.0)) * 100_000.0).round() as i32));
-                }
-                sh.fill = Some(Fill::Solid { color });
-            }
-        }
-        Ok(())
-    })
+    let inherited = inherited_fills(s, p)?;
+    edit_group_appearance(s, p, "shape.fill", |sh| apply_fill(sh, new_fill.as_ref(), pic_props, transparency, &inherited))
 }
 
 /// Applies Format Shape ▸ Picture or texture fill options to a picture fill.

crates/format/src/rtf.rs

@@ -126,155 +126,216 @@ fn advance(rtf: &str, i: usize, n: usize) -> usize {
     if rtf.is_char_boundary(i + n) { i + n } else { i + rtf.get(i..).and_then(|r| r.chars().next()).map_or(1, char::len_utf8) }
 }
 
+/// Group state: whether the text is being skipped, and the `\uc` fallback count.
+#[derive(Clone, Copy)]
+struct St {
+    skip: bool,
+    uc: usize,
+}
+
+/// Parser state for [`rtf_to_outline`]: the cursor, the group stack and the outline so far.
+struct Outline<'a> {
+    rtf: &'a str,
+    b: &'a [u8],
+    i: usize,
+    stack: Vec<St>,
+    st: St,
+    level: Option<u32>,
+    indent: i64,
+    pending_skip: usize,
+    line: String,
+    out: String,
+}
+
 /// Read an RTF outline into tab-indented outline text: a paragraph's outline level (or its left
 /// indent in 0.25" steps) becomes its number of tabs.
 pub fn rtf_to_outline(rtf: &str) -> String {
-    #[derive(Clone, Copy)]
-    struct St {
-        skip: bool,
-        uc: usize,
-    }
-    let b = rtf.as_bytes();
-    let mut stack: Vec<St> = vec![];
-    let mut st = St { skip: false, uc: 1 };
-    let (mut level, mut indent): (Option<u32>, i64) = (None, 0);
-    let mut line = String::new();
-    let mut out = String::new();
-    let mut pending_skip = 0usize;
-    let mut i = 0;
-    let flush = |line: &mut String, level: Option<u32>, indent: i64, out: &mut String| {
-        let text = line.trim();
-        if !text.is_empty() {
-            let n = level.map(|l| l as i64).unwrap_or(indent.saturating_add(180) / 360).clamp(0, 9) as usize;
-            out.push_str(&"\t".repeat(n));
-            out.push_str(text);
-            out.push('\n');
-        }
-        line.clear();
+    let mut p = Outline {
+        rtf,
+        b: rtf.as_bytes(),
+        i: 0,
+        stack: vec![],
+        st: St { skip: false, uc: 1 },
+        level: None,
+        indent: 0,
+        pending_skip: 0,
+        line: String::new(),
+        out: String::new(),
     };
-    while i < b.len() {
-        let c = b[i];
-        match c {
-            b'{' => {
-                stack.push(st);
-                i += 1;
-                // `{\*\dest …}`: an ignorable destination.
-                if b[i..].starts_with(b"\\*") {
-                    st.skip = true;
-                }
-            }
-            b'}' => {
-                st = stack.pop().unwrap_or(St { skip: false, uc: 1 });
-                i += 1;
-            }
-            b'\\' => {
-                i += 1;
-                let Some(&n) = b.get(i) else { break };
-                if n.is_ascii_alphabetic() {
-                    let s = i;
-                    while i < b.len() && b[i].is_ascii_alphabetic() {
-                        i += 1;
-                    }
-                    let word = &rtf[s..i];
-                    let ns = i;
-                    if i < b.len() && (b[i] == b'-' || b[i].is_ascii_digit()) {
-                        i += 1;
-                        while i < b.len() && b[i].is_ascii_digit() {
-                            i += 1;
-                        }
-                    }
-                    let arg: Option<i64> = rtf[ns..i].parse().ok();
-                    if i < b.len() && b[i] == b' ' {
-                        i += 1;
-                    }
-                    if pending_skip > 0 {
-                        pending_skip -= 1;
-                        continue;
-                    }
-                    if SKIP.contains(&word) {
-                        st.skip = true;
-                    }
-                    if st.skip {
-                        continue;
-                    }
-                    match word {
-                        "par" => {
-                            flush(&mut line, level, indent, &mut out);
-                        }
-                        "pard" => {
-                            level = None;
-                            indent = 0;
-                        }
-                        "outlinelevel" => level = arg.map(|a| a.clamp(0, 9) as u32),
-                        "li" => indent = arg.unwrap_or(0),
-                        "tab" | "line" => line.push(' '),
-                        "uc" => st.uc = arg.unwrap_or(1).clamp(0, 8) as usize,
-                        "u" => {
-                            let v = arg.unwrap_or(0);
-                            // RTF writes UTF-16 units as signed 16-bit values.
-                            push_utf16(&mut line, (v & 0xffff) as u32);
-                            pending_skip = 0;
-                            // Skip the fallback characters.
-                            let mut k = st.uc;
-                            while k > 0 && i < b.len() {
-                                if b[i] == b'\\' && b.get(i + 1) == Some(&b'\'') {
-                                    i = advance(rtf, i, 4);
-                                } else if b[i] == b'{' || b[i] == b'}' || b[i] == b'\\' {
-                                    break;
-                                } else {
-                                    i = advance(rtf, i, 1);
-                                }
-                                k -= 1;
-                            }
-                        }
-                        "emdash" => line.push('—'),
-                        "endash" => line.push('–'),
-                        "bullet" => line.push('•'),
-                        "lquote" => line.push('‘'),
-                        "rquote" => line.push('’'),
-                        "ldblquote" => line.push('“'),
-                        "rdblquote" => line.push('”'),
-                        _ => {}
-                    }
-                } else if n == b'\'' {
-                    let hex = rtf.get(i + 1..i + 3).filter(|h| h.bytes().all(|b| b.is_ascii_hexdigit())).and_then(|h| u8::from_str_radix(h, 16).ok());
-                    i = advance(rtf, i, 3);
-                    if !st.skip
-                        && let Some(h) = hex
-                        && h != 0
-                    {
-                        line.push(cp1252(h));
-                    }
-                } else {
-                    i += 1;
-                    if st.skip {
-                        continue;
-                    }
-                    match n {
-                        b'\\' | b'{' | b'}' => line.push(n as char),
-                        b'~' => line.push('\u{a0}'),
-                        b'_' => line.push('\u{2011}'),
-                        b'\n' | b'\r' => flush(&mut line, level, indent, &mut out),
-                        _ => {}
+    p.run();
+    p.flush();
+    finish_surrogates(p.out)
+}
+
+impl Outline<'_> {
+    fn run(&mut self) {
+        while self.i < self.b.len() {
+            match self.b[self.i] {
+                b'{' => self.open_group(),
+                b'}' => self.close_group(),
+                b'\\' => {
+                    self.i += 1;
+                    if !self.escape() {
+                        break;
                     }
                 }
-            }
-            b'\r' | b'\n' => i += 1,
-            _ => {
-                let s = i;
-                while i < b.len() && !matches!(b[i], b'{' | b'}' | b'\\' | b'\r' | b'\n') {
-                    i += 1;
-                }
-                if !st.skip
-                    && let Some(t) = rtf.get(s..i)
-                {
-                    line.push_str(&t.replace('\u{0}', ""));
-                }
+                b'\r' | b'\n' => self.i += 1,
+                _ => self.text(),
             }
         }
     }
-    flush(&mut line, level, indent, &mut out);
-    finish_surrogates(out)
+
+    /// `{`: push the group; `{\*\dest …}` is an ignorable destination.
+    fn open_group(&mut self) {
+        self.stack.push(self.st);
+        self.i += 1;
+        if self.b[self.i..].starts_with(b"\\*") {
+            self.st.skip = true;
+        }
+    }
+
+    fn close_group(&mut self) {
+        self.st = self.stack.pop().unwrap_or(St { skip: false, uc: 1 });
+        self.i += 1;
+    }
+
+    /// The character after a backslash: a control word, a hex escape or a symbol. False at the
+    /// end of the input.
+    fn escape(&mut self) -> bool {
+        let Some(&n) = self.b.get(self.i) else { return false };
+        if n.is_ascii_alphabetic() {
+            self.control_word();
+        } else if n == b'\'' {
+            self.hex_escape();
+        } else {
+            self.symbol(n);
+        }
+        true
+    }
+
+    fn control_word(&mut self) {
+        let rtf = self.rtf;
+        let s = self.i;
+        while self.i < self.b.len() && self.b[self.i].is_ascii_alphabetic() {
+            self.i += 1;
+        }
+        let word = &rtf[s..self.i];
+        let ns = self.i;
+        if self.i < self.b.len() && (self.b[self.i] == b'-' || self.b[self.i].is_ascii_digit()) {
+            self.i += 1;
+            while self.i < self.b.len() && self.b[self.i].is_ascii_digit() {
+                self.i += 1;
+            }
+        }
+        let arg: Option<i64> = rtf[ns..self.i].parse().ok();
+        if self.i < self.b.len() && self.b[self.i] == b' ' {
+            self.i += 1;
+        }
+        if self.pending_skip > 0 {
+            self.pending_skip -= 1;
+            return;
+        }
+        if SKIP.contains(&word) {
+            self.st.skip = true;
+        }
+        if !self.st.skip {
+            self.apply(word, arg);
+        }
+    }
+
+    fn apply(&mut self, word: &str, arg: Option<i64>) {
+        match word {
+            "par" => self.flush(),
+            "pard" => {
+                self.level = None;
+                self.indent = 0;
+            }
+            "outlinelevel" => self.level = arg.map(|a| a.clamp(0, 9) as u32),
+            "li" => self.indent = arg.unwrap_or(0),
+            "tab" | "line" => self.line.push(' '),
+            "uc" => self.st.uc = arg.unwrap_or(1).clamp(0, 8) as usize,
+            "u" => self.unicode(arg.unwrap_or(0)),
+            "emdash" => self.line.push('—'),
+            "endash" => self.line.push('–'),
+            "bullet" => self.line.push('•'),
+            "lquote" => self.line.push('‘'),
+            "rquote" => self.line.push('’'),
+            "ldblquote" => self.line.push('“'),
+            "rdblquote" => self.line.push('”'),
+            _ => {}
+        }
+    }
+
+    /// `\uN`: one UTF-16 unit, then the `\uc` fallback characters, which are skipped.
+    fn unicode(&mut self, v: i64) {
+        // RTF writes UTF-16 units as signed 16-bit values.
+        push_utf16(&mut self.line, (v & 0xffff) as u32);
+        self.pending_skip = 0;
+        let mut k = self.st.uc;
+        while k > 0 && self.i < self.b.len() {
+            if self.b[self.i] == b'\\' && self.b.get(self.i + 1) == Some(&b'\'') {
+                self.i = advance(self.rtf, self.i, 4);
+            } else if matches!(self.b[self.i], b'{' | b'}' | b'\\') {
+                break;
+            } else {
+                self.i = advance(self.rtf, self.i, 1);
+            }
+            k -= 1;
+        }
+    }
+
+    /// `\'hh`: one cp1252 byte.
+    fn hex_escape(&mut self) {
+        let hex =
+            self.rtf.get(self.i + 1..self.i + 3).filter(|h| h.bytes().all(|b| b.is_ascii_hexdigit())).and_then(|h| u8::from_str_radix(h, 16).ok());
+        self.i = advance(self.rtf, self.i, 3);
+        if !self.st.skip
+            && let Some(h) = hex
+            && h != 0
+        {
+            self.line.push(cp1252(h));
+        }
+    }
+
+    /// A one-character control symbol (`\\`, `\{`, `\~`, ...).
+    fn symbol(&mut self, n: u8) {
+        self.i += 1;
+        if self.st.skip {
+            return;
+        }
+        match n {
+            b'\\' | b'{' | b'}' => self.line.push(n as char),
+            b'~' => self.line.push('\u{a0}'),
+            b'_' => self.line.push('\u{2011}'),
+            b'\n' | b'\r' => self.flush(),
+            _ => {}
+        }
+    }
+
+    /// Plain text up to the next special character.
+    fn text(&mut self) {
+        let s = self.i;
+        while self.i < self.b.len() && !matches!(self.b[self.i], b'{' | b'}' | b'\\' | b'\r' | b'\n') {
+            self.i += 1;
+        }
+        if !self.st.skip
+            && let Some(t) = self.rtf.get(s..self.i)
+        {
+            self.line.push_str(&t.replace('\u{0}', ""));
+        }
+    }
+
+    /// Ends the current line: a non-empty line becomes one outline entry, indented by its level.
+    fn flush(&mut self) {
+        let text = self.line.trim();
+        if !text.is_empty() {
+            let n = self.level.map(|l| l as i64).unwrap_or(self.indent.saturating_add(180) / 360).clamp(0, 9) as usize;
+            self.out.push_str(&"\t".repeat(n));
+            self.out.push_str(text);
+            self.out.push('\n');
+        }
+        self.line.clear();
+    }
 }
 
 /// Push a UTF-16 unit; surrogate halves are parked as private markers (NUL plus four hex digits)

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 +174,116 @@ 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
+    }
+
+    /// picNumLXNoWrap (8-34, 8-35) of a short-term modification from the predictor `pred`, and the
+    /// picNum it selects.
+    fn short_term_pic_nums(pred: i32, idc: u32, value: u32, curr_pic_num: i32, max_pic_num: i32) -> (i32, i32) {
+        let d = value as i32 + 1;
+        let no_wrap = if 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
+        };
+        let pn = if no_wrap > curr_pic_num { no_wrap - max_pic_num } else { no_wrap };
+        (no_wrap, pn)
+    }
+
+    /// Puts `target` at `ref_idx` and shifts the rest of the list up, then drops the later entries
+    /// that repeat it (8.2.4.3.1, 8.2.4.3.2) and keeps the list `n` long.
+    fn insert_ref<'a>(list: &mut Vec<Option<&'a DpbEntry>>, ref_idx: usize, n: usize, target: &'a DpbEntry, is_long: bool) {
+        for c in (ref_idx + 1..=n).rev() {
+            list[c] = list[c - 1];
+        }
+        list[ref_idx] = Some(target);
+        let mut nidx = ref_idx + 1;
+        for c in ref_idx + 1..=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);
+    }
+
+    /// 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 (no_wrap, pn) = Self::short_term_pic_nums(pred, m.idc, m.value, curr_pic_num, max_pic_num);
+                    pred = 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;
+            };
+            Self::insert_ref(list, ref_idx, n, target, is_long);
+            ref_idx += 1;
+        }
+        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 +296,7 @@ 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 +304,7 @@ 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 +314,88 @@ 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 +427,12 @@ 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 +442,7 @@ 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 +460,32 @@ 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/params.rs

@@ -575,7 +575,7 @@ impl Pps {
         ensure!(sps_id < 32, "seq_parameter_set_id out of range");
         let sps = spss
             .get(sps_id as usize)
-            .and_then(|s| s.as_ref())
+            .and_then(Option::as_ref)
             .ok_or_else(|| crate::Error::MissingParameterSet(format!("SPS {sps_id} for PPS {id}")))?;
         let entropy_coding_mode = r.read_flag()?;
         let bottom_field_pic_order_in_frame_present = r.read_flag()?;

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)
@@ -59,27 +73,40 @@ 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 (w0, h0) = (g.w, g.h);
+        // With a parallel merge level above 4x4, an 8x8 coding block shares one candidate set.
+        let g = if self.pps.log2_parallel_merge_level > 2 && g.ncb == 8 {
+            PbGeom { xp: g.xc, yp: g.yc, w: g.ncb, h: g.ncb, part_idx: 0, ..g }
+        } else {
+            g
+        };
+        let max = self.sh.max_num_merge_cand as usize;
+        let (list, n) = self.spatial_merge_candidates(&g, part);
+        if merge_idx < n {
+            return self.finish_merge(list[merge_idx], w0, h0);
+        }
+        let mut cands: Vec<MvField> = list[..n].to_vec();
+        cands.extend(self.temporal_merge_candidate(g.xp, g.yp, g.w, g.h));
+        if merge_idx < cands.len() {
+            return self.finish_merge(cands[merge_idx], w0, h0);
+        }
+        self.combined_merge_candidates(&mut cands, max);
+        if merge_idx < cands.len() {
+            return self.finish_merge(cands[merge_idx], w0, h0);
+        }
+        self.zero_merge_candidates(&mut cands, merge_idx);
+        self.finish_merge(cands[merge_idx], w0, h0)
+    }
+
+    /// The spatial candidates A1, B1, B0, A0 and B2 (8.5.3.2.3): the candidates, and how many.
+    fn spatial_merge_candidates(&self, g: &PbGeom, part: PartMode) -> ([MvField; 5], usize) {
+        let PbGeom { xp, yp, w, h, part_idx, .. } = *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 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;
-        // spatial candidates (8.5.3.2.3)
         let pa1 = (xp - 1, yp + h - 1);
         let av_a1 = avail(pa1.0, pa1.1) && !(matches!(part, PartMode::PNx2N | PartMode::PnLx2N | PartMode::PnRx2N) && part_idx == 1);
         let m_a1 = if av_a1 { self.mv_at(pa1.0, pa1.1) } else { MvField::default() };
@@ -126,10 +153,11 @@ impl<'a> SliceDecoder<'a> {
             list[n] = self.mv_at(pb2.0, pb2.1);
             n += 1;
         }
-        if merge_idx < n {
-            return self.finish_merge(list[merge_idx], w0, h0);
-        }
-        // temporal candidate
+        (list, n)
+    }
+
+    /// The temporal candidate (8.5.3.2.8): list 0 from the collocated picture, and list 1 for B slices.
+    fn temporal_merge_candidate(&self, xp: i32, yp: i32, w: i32, h: i32) -> Option<MvField> {
         let mut col = MvField::default();
         let mut col_avail = false;
         if let Some(mv) = self.temporal(xp, yp, w, h, 0, 0) {
@@ -144,30 +172,28 @@ impl<'a> SliceDecoder<'a> {
             col.ref_idx[1] = 0;
             col_avail = true;
         }
-        let mut cands: Vec<MvField> = list[..n].to_vec();
-        if col_avail {
-            cands.push(col);
-        }
-        if merge_idx < cands.len() {
-            return self.finish_merge(cands[merge_idx], w0, h0);
-        }
-        // combined bi-predictive candidates (8.5.3.2.4)
+        col_avail.then_some(col)
+    }
+
+    /// Combined bi-predictive candidates (8.5.3.2.4) from pairs of the original candidates.
+    fn combined_merge_candidates(&self, cands: &mut Vec<MvField>, max: usize) {
         let num_orig = cands.len();
-        if self.sh.is_b() && num_orig > 1 && num_orig < max {
-            let mut comb = 0;
-            while comb < num_orig * (num_orig - 1) && cands.len() < max {
-                let (i0, i1) = COMB[comb];
-                let (c0, c1) = (cands[i0], cands[i1]);
-                if c0.pred(0) && c1.pred(1) && (self.ref_pic(0, c0.ref_idx[0]).poc != self.ref_pic(1, c1.ref_idx[1]).poc || c0.mv[0] != c1.mv[1]) {
-                    cands.push(MvField { mv: [c0.mv[0], c1.mv[1]], ref_idx: [c0.ref_idx[0], c1.ref_idx[1]] });
-                }
-                comb += 1;
+        if !(self.sh.is_b() && num_orig > 1 && num_orig < max) {
+            return;
+        }
+        let mut comb = 0;
+        while comb < num_orig * (num_orig - 1) && cands.len() < max {
+            let (i0, i1) = COMB[comb];
+            let (c0, c1) = (cands[i0], cands[i1]);
+            if c0.pred(0) && c1.pred(1) && (self.ref_pic(0, c0.ref_idx[0]).poc != self.ref_pic(1, c1.ref_idx[1]).poc || c0.mv[0] != c1.mv[1]) {
+                cands.push(MvField { mv: [c0.mv[0], c1.mv[1]], ref_idx: [c0.ref_idx[0], c1.ref_idx[1]] });
             }
+            comb += 1;
         }
-        if merge_idx < cands.len() {
-            return self.finish_merge(cands[merge_idx], w0, h0);
-        }
-        // zero candidates (8.5.3.2.5)
+    }
+
+    /// Zero candidates (8.5.3.2.5) up to and including `merge_idx`.
+    fn zero_merge_candidates(&self, cands: &mut Vec<MvField>, merge_idx: usize) {
         let num_ref = if self.sh.is_b() { self.sh.num_ref_idx[0].min(self.sh.num_ref_idx[1]) } else { self.sh.num_ref_idx[0] } as usize;
         let mut zero_idx = 0;
         while cands.len() <= merge_idx {
@@ -175,7 +201,6 @@ impl<'a> SliceDecoder<'a> {
             cands.push(MvField { mv: [[0; 2]; 2], ref_idx: [r, if self.sh.is_b() { r } else { -1 }] });
             zero_idx += 1;
         }
-        self.finish_merge(cands[merge_idx], w0, h0)
     }
 
     fn finish_merge(&self, mut f: MvField, w: i32, h: i32) -> MvField {
@@ -237,20 +262,8 @@ 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 +290,7 @@ 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 +306,7 @@ 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};
@@ -276,6 +277,30 @@ pub struct SliceDecoder<'a> {
     pub ctb_y: i32,
 }
 
+/// Scaling process (8.6.3): dequantises the top-left `max_x` × `max_y` coefficients of the `n` × `n`
+/// block `coeffs` with scaling matrix `m` (flat when `None`).
+fn dequantize(coeffs: &mut [i32], n: usize, (max_x, max_y): (usize, usize), m: Option<&[u8]>, scale: i64, round: i64, bd_shift: u32) {
+    for y in 0..=max_y {
+        for x in 0..=max_x {
+            let v = coeffs[y * n + x];
+            if v != 0 {
+                let f = m.map_or(16, |m| m[y * n + x] as i64);
+                coeffs[y * n + x] = ((v as i64 * f * scale + round) >> bd_shift).clamp(-32768, 32767) as i32;
+            }
+        }
+    }
+}
+
+/// Adds the `n` × `n` residual `coeffs` to the samples of `plane` at (x0, y0), clipped to 0..=`max`.
+fn add_residual(plane: &mut [u16], stride: usize, (x0, y0): (usize, usize), n: usize, coeffs: &[i32], max: i32) {
+    for y in 0..n {
+        let row = &mut plane[(y0 + y) * stride + x0..][..n];
+        for (x, s) in row.iter_mut().enumerate() {
+            *s = (*s as i32 + coeffs[y * n + x]).clamp(0, max) as u16;
+        }
+    }
+}
+
 impl<'a> SliceDecoder<'a> {
     #[inline(always)]
     fn dd(&mut self, ctx: usize) -> u32 {
@@ -678,7 +703,7 @@ 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 +732,7 @@ 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 +749,8 @@ 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 +764,7 @@ 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 +938,7 @@ 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 +963,10 @@ 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 +975,7 @@ 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 +1005,15 @@ 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 +1318,7 @@ 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 +1337,7 @@ 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];
@@ -1321,15 +1355,7 @@ impl<'a> SliceDecoder<'a> {
                 }
                 _ => None,
             };
-            for y in 0..=max_y {
-                for x in 0..=max_x {
-                    let v = coeffs[y * n + x];
-                    if v != 0 {
-                        let f = m.map_or(16, |m| m[y * n + x] as i64);
-                        coeffs[y * n + x] = ((v as i64 * f * scale + round) >> bd_shift).clamp(-32768, 32767) as i32;
-                    }
-                }
-            }
+            dequantize(coeffs, n, (max_x, max_y), m, scale, round, bd_shift);
             let bd_shift = 20 - bd;
             if ts {
                 transform::transform_skip(coeffs, n, bd_shift);
@@ -1345,106 +1371,112 @@ impl<'a> SliceDecoder<'a> {
         let _ = dc_only;
         let (stride, max) = (if c == 0 { self.pic.width } else { self.pic.cwidth }, (1i32 << bd) - 1);
         let plane = &mut self.pic.planes[c];
-        for y in 0..n {
-            let row = &mut plane[(y0 as usize + y) * stride + x0 as usize..][..n];
-            for (x, s) in row.iter_mut().enumerate() {
-                *s = (*s as i32 + coeffs[y * n + x]).clamp(0, max) as u16;
-            }
-        }
+        add_residual(plane, stride, (x0 as usize, y0 as usize), n, coeffs, max);
         Ok(())
     }
 
     /// 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 merge = if skip { true } else { self.dd(MERGE_FLAG) == 1 };
         let mvf = if merge {
-            let idx = if self.sh.max_num_merge_cand > 1 {
-                let mut i = 0;
-                if self.dd(MERGE_IDX) == 1 {
-                    i = 1;
-                    while i < self.sh.max_num_merge_cand - 1 && self.bypass() == 1 {
-                        i += 1;
-                    }
-                }
-                i
-            } else {
-                0
-            };
-            self.derive_merge(xc, yc, ncb, xp, yp, w, h, part_idx, part, idx as usize)
+            let idx = self.merge_idx();
+            self.derive_merge(g, part, idx as usize)
         } else {
-            let pred_idc = if self.sh.is_b() {
-                if w + h != 12 {
-                    let depth = self.pic.blk_at(xc, yc).depth as usize;
-                    if self.dd(INTER_PRED_IDC + depth) == 1 { 2 } else { self.dd(INTER_PRED_IDC + 4) }
-                } else {
-                    self.dd(INTER_PRED_IDC + 4)
-                }
-            } else {
-                0
-            };
-            let mut ref_idx = [-1i8; 2];
-            let mut mvd = [[0i32; 2]; 2];
-            let mut mvp_flag = [0u32; 2];
-            for l in 0..2 {
-                let used = if l == 0 { pred_idc != 1 } else { pred_idc != 0 };
-                if !used {
-                    continue;
-                }
-                let nref = self.sh.num_ref_idx[l];
-                let mut ri = 0;
-                if nref > 1 {
-                    while ri < nref - 1 {
-                        let bin = if ri < 2 { self.dd(REF_IDX + ri as usize) } else { self.bypass() };
-                        if bin == 0 {
-                            break;
-                        }
-                        ri += 1;
-                    }
-                }
-                ref_idx[l] = ri as i8;
-                if l == 1 && self.sh.mvd_l1_zero && pred_idc == 2 {
-                    mvd[1] = [0, 0];
-                } else {
-                    mvd[l] = self.mvd_coding()?;
-                }
-                mvp_flag[l] = self.dd(MVP_FLAG);
-            }
-            let mut f = MvField { mv: [[0; 2]; 2], ref_idx };
-            for l in 0..2 {
-                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);
-                // 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];
-            }
-            f
+            self.amvp_motion(g)?
         };
         for l in 0..2 {
             if mvf.ref_idx[l] >= 0 {
                 ensure!((mvf.ref_idx[l] as usize) < self.refs[l].len(), "reference index out of range");
             }
         }
-        // store motion
+        self.store_motion(g.xp, g.yp, g.w, g.h, mvf);
+        self.motion_compensate(g.xp, g.yp, g.w as usize, g.h as usize, &mvf);
+        Ok(merge)
+    }
+
+    /// merge_idx(): the merge candidate index, truncated unary (0 when there is one candidate).
+    fn merge_idx(&mut self) -> u32 {
+        if self.sh.max_num_merge_cand <= 1 {
+            return 0;
+        }
+        let mut i = 0;
+        if self.dd(MERGE_IDX) == 1 {
+            i = 1;
+            while i < self.sh.max_num_merge_cand - 1 && self.bypass() == 1 {
+                i += 1;
+            }
+        }
+        i
+    }
+
+    /// inter_pred_idc: 0 = list 0, 1 = list 1, 2 = both.
+    fn inter_pred_idc(&mut self, xc: i32, yc: i32, w: i32, h: i32) -> u32 {
+        if !self.sh.is_b() {
+            return 0;
+        }
+        if w + h != 12 {
+            let depth = self.pic.blk_at(xc, yc).depth as usize;
+            if self.dd(INTER_PRED_IDC + depth) == 1 { 2 } else { self.dd(INTER_PRED_IDC + 4) }
+        } else {
+            self.dd(INTER_PRED_IDC + 4)
+        }
+    }
+
+    /// ref_idx_lX: truncated unary, the first two bins context-coded.
+    fn ref_idx(&mut self, l: usize) -> u32 {
+        let nref = self.sh.num_ref_idx[l];
+        let mut ri = 0;
+        if nref > 1 {
+            while ri < nref - 1 {
+                let bin = if ri < 2 { self.dd(REF_IDX + ri as usize) } else { self.bypass() };
+                if bin == 0 {
+                    break;
+                }
+                ri += 1;
+            }
+        }
+        ri
+    }
+
+    /// The AMVP branch: parse each used list's reference index, mvd and predictor flag, then add
+    /// every mvd to its predictor.
+    fn amvp_motion(&mut self, g: PbGeom) -> Result<MvField> {
+        let pred_idc = self.inter_pred_idc(g.xc, g.yc, g.w, g.h);
+        let mut ref_idx = [-1i8; 2];
+        let mut mvd = [[0i32; 2]; 2];
+        let mut mvp_flag = [0u32; 2];
+        for l in 0..2 {
+            let used = if l == 0 { pred_idc != 1 } else { pred_idc != 0 };
+            if !used {
+                continue;
+            }
+            ref_idx[l] = self.ref_idx(l) as i8;
+            if l == 1 && self.sh.mvd_l1_zero && pred_idc == 2 {
+                mvd[1] = [0, 0];
+            } else {
+                mvd[l] = self.mvd_coding()?;
+            }
+            mvp_flag[l] = self.dd(MVP_FLAG);
+        }
+        let mut f = MvField { mv: [[0; 2]; 2], ref_idx };
+        for l in 0..2 {
+            if ref_idx[l] < 0 {
+                continue;
+            }
+            let mvp = self.derive_amvp(g, 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];
+        }
+        Ok(f)
+    }
+
+    /// Writes a prediction block's motion into the picture's 4x4 motion field.
+    fn store_motion(&mut self, xp: i32, yp: i32, w: i32, h: i32, mvf: MvField) {
         let w4 = self.pic.w4;
         for by in (yp >> 2)..((yp + h) >> 2) {
             let o = by as usize * w4;
             self.pic.mvf[o + (xp >> 2) as usize..o + ((xp + w) >> 2) as usize].fill(mvf);
         }
-        self.motion_compensate(xp, yp, w as usize, h as usize, &mvf);
-        Ok(merge)
     }
 
     fn mvd_coding(&mut self) -> Result<[i32; 2]> {
@@ -1524,11 +1556,11 @@ 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);
             }
         }
     }

crates/model/src/chart.rs

@@ -802,6 +802,7 @@ impl Chart {
 /// The (min, max, major unit) of an axis over the data range [lo, hi]: `fmin`/`fmax`/`major`
 /// where the chart fixes them; otherwise Office's automatic limits, rounded out to the smallest
 /// 1-2-5 unit giving at most `max_intervals` intervals:
+/// The data range [lo, hi] an axis covers, clamped to ±1e100; (0, 1) without finite values.
 ///
 /// - **Maximum:** `hi + (hi − lo)/20` (with `headroom`), rounded up to the unit.
 /// - **Minimum:** zero when every value is ≥ 0 and `lo` is below 5/6 of `hi`; otherwise
@@ -809,22 +810,14 @@ impl Chart {
 ///   near them.
 /// - All-negative values mirror this: the axis ends at zero when `hi` is above 5/6 of `lo`.
 ///   Mixed signs get headroom at both ends.
-pub fn axis_scale(
-    lo: f64,
-    hi: f64,
-    fmin: Option<f64>,
-    fmax: Option<f64>,
-    major: Option<f64>,
-    max_intervals: usize,
-    headroom: bool,
-) -> (f64, f64, f64) {
-    // Values past ±1e100 aren't axis values; clamping keeps the arithmetic finite.
-    let fin = |v: Option<f64>| v.filter(|v| v.is_finite()).map(|v| v.clamp(-1e100, 1e100));
-    let (fmin, fmax) = (fin(fmin), fin(fmax));
-    let (lo, hi) = if lo.is_finite() && hi.is_finite() { (lo.min(hi).clamp(-1e100, 1e100), lo.max(hi).clamp(-1e100, 1e100)) } else { (0.0, 1.0) };
-    // Office's automatic limits ("How chart axis limits are determined", as Jon Peltier
-    // summarises it): 5 % of the data range past the data, and zero when the values don't
-    // sit within the top sixth of their largest magnitude. A single value is measured from zero.
+fn axis_data_range(lo: f64, hi: f64) -> (f64, f64) {
+    if lo.is_finite() && hi.is_finite() { (lo.min(hi).clamp(-1e100, 1e100), lo.max(hi).clamp(-1e100, 1e100)) } else { (0.0, 1.0) }
+}
+
+/// Office's automatic (min, max) of the data range [lo, hi]. "How chart axis limits are determined",
+/// as Jon Peltier summarises it: 5 % of the data range past the data, and zero when the values don't
+/// sit within the top sixth of their largest magnitude. A single value is measured from zero.
+fn auto_limits(lo: f64, hi: f64, headroom: bool) -> (f64, f64) {
     let flat = hi - lo <= 1e-12 * lo.abs().max(hi.abs());
     let pad = if !headroom {
         0.0
@@ -835,21 +828,23 @@ pub fn axis_scale(
     };
     let auto_lo = if lo >= 0.0 && (flat || lo < hi * 5.0 / 6.0) { 0.0 } else { lo - pad };
     let auto_hi = if hi <= 0.0 && (flat || hi > lo * 5.0 / 6.0) { 0.0 } else { hi + pad };
-    let mut lo = fmin.unwrap_or(auto_lo);
-    let mut hi = fmax.unwrap_or(auto_hi);
-    let mag = lo.abs().max(hi.abs());
-    // Relative to the data's magnitude, so tiny values don't count as flat; all-zero falls back to 1.
-    let unit = if mag == 0.0 { 1.0 } else { mag };
-    if hi - lo <= 1e-12 * mag {
-        // A fixed limit on the far side of the data: the automatic end gives way.
-        match (fmin, fmax) {
-            (None, Some(_)) => lo = if hi > 0.0 { 0.0 } else { hi - hi.abs().max(unit) },
-            (Some(_), None) if lo < 0.0 => hi = 0.0,
-            _ => hi = lo + unit,
-        }
+    (auto_lo, auto_hi)
+}
+
+/// The ends of a span too small to show: with one fixed limit on the far side of the data, the
+/// automatic end gives way; otherwise the span is one `unit` wide.
+fn flat_span_ends(fmin: Option<f64>, fmax: Option<f64>, lo: f64, hi: f64, unit: f64) -> (f64, f64) {
+    match (fmin, fmax) {
+        (None, Some(_)) => (if hi > 0.0 { 0.0 } else { hi - hi.abs().max(unit) }, hi),
+        (Some(_), None) if lo < 0.0 => (lo, 0.0),
+        _ => (lo, lo + unit),
     }
-    let n = max_intervals.clamp(1, 50) as f64;
-    let step = match fin(major).filter(|m| *m > 0.0 && (hi - lo) / m <= 1000.0) {
+}
+
+/// The major step: `major` when it gives at most 1000 intervals, else the smallest 1-2-5 step
+/// that fits `n` intervals over [lo, hi].
+fn axis_step(major: Option<f64>, lo: f64, hi: f64, n: f64, unit: f64) -> f64 {
+    match major.filter(|m| *m > 0.0 && (hi - lo) / m <= 1000.0) {
         Some(m) => m,
         None => {
             let mut step = 10f64.powf(((hi - lo) / n).log10().floor());
@@ -865,7 +860,32 @@ pub fn axis_scale(
             }
             step
         }
-    };
+    }
+}
+
+pub fn axis_scale(
+    lo: f64,
+    hi: f64,
+    fmin: Option<f64>,
+    fmax: Option<f64>,
+    major: Option<f64>,
+    max_intervals: usize,
+    headroom: bool,
+) -> (f64, f64, f64) {
+    // Values past ±1e100 aren't axis values; clamping keeps the arithmetic finite.
+    let fin = |v: Option<f64>| v.filter(|v| v.is_finite()).map(|v| v.clamp(-1e100, 1e100));
+    let (fmin, fmax) = (fin(fmin), fin(fmax));
+    let (lo, hi) = axis_data_range(lo, hi);
+    let (auto_lo, auto_hi) = auto_limits(lo, hi, headroom);
+    let (mut lo, mut hi) = (fmin.unwrap_or(auto_lo), fmax.unwrap_or(auto_hi));
+    let mag = lo.abs().max(hi.abs());
+    // Relative to the data's magnitude, so tiny values don't count as flat; all-zero falls back to 1.
+    let unit = if mag == 0.0 { 1.0 } else { mag };
+    if hi - lo <= 1e-12 * mag {
+        (lo, hi) = flat_span_ends(fmin, fmax, lo, hi, unit);
+    }
+    let n = max_intervals.clamp(1, 50) as f64;
+    let step = axis_step(fin(major), lo, hi, n, unit);
     let min = fmin.unwrap_or((lo / step + 1e-9).floor() * step);
     let max = fmax.unwrap_or((hi / step - 1e-9).ceil() * step).max(min + step);
     (min, max, step)

crates/pptx/src/read/dml.rs

@@ -576,6 +576,44 @@ pub fn text_body(imp: &mut Imp, part: &Part, tx: &El) -> TextBody {
 
 const MAX_PARAGRAPH_RUNS: usize = 10_000;
 
+/// The run properties of `e`'s `rPr` child, or the defaults.
+fn run_props(imp: &mut Imp, part: &Part, e: &El) -> RunProps {
+    e.child("rPr").map(|r| rpr(imp, part, r)).unwrap_or_default()
+}
+
+/// Appends the runs of an `a:r` text, where a vertical tab or line feed is a line break, as it
+/// lays out and as export writes it.
+fn push_text_runs(para: &mut Paragraph, props: RunProps, text: &str) {
+    for (i, piece) in text.split(['\u{b}', '\n']).enumerate() {
+        if para.runs.len() >= MAX_PARAGRAPH_RUNS {
+            break;
+        }
+        if i > 0 {
+            para.runs.push(Run { text: String::new(), props: props.clone(), kind: RunKind::Break });
+        }
+        if !piece.is_empty() {
+            if para.runs.len() >= MAX_PARAGRAPH_RUNS {
+                break;
+            }
+            para.runs.push(Run { text: piece.to_string(), props: props.clone(), kind: RunKind::Text });
+        }
+    }
+}
+
+/// Appends an Office Math (a14:m) run: its plain text and its OMML.
+fn push_math_run(para: &mut Paragraph, part: &Part, e: &El) {
+    if para.runs.len() < MAX_PARAGRAPH_RUNS {
+        let mut text = String::new();
+        let mut ts = vec![];
+        e.find_all("t", &mut ts);
+        for t in ts {
+            text.push_str(&t.text());
+        }
+        let omml = e.elements().map(|c| part.keep_inner(c)).collect::<String>();
+        para.runs.push(Run { text, props: RunProps::default(), kind: RunKind::Math { omml } });
+    }
+}
+
 pub fn paragraph(imp: &mut Imp, part: &Part, p: &El) -> Paragraph {
     let mut para = Paragraph::default();
     for e in p.elements() {
@@ -589,50 +627,24 @@ pub fn paragraph(imp: &mut Imp, part: &Part, p: &El) -> Paragraph {
                 }
             }
             "r" => {
-                let props = e.child("rPr").map(|r| rpr(imp, part, r)).unwrap_or_default();
+                let props = run_props(imp, part, e);
                 let text: String = e.child("t").map(|t| t.text()).unwrap_or_default();
-                // A vertical tab or line feed inside text is a line break, as it lays out and as export writes it.
-                for (i, piece) in text.split(['\u{b}', '\n']).enumerate() {
-                    if para.runs.len() >= MAX_PARAGRAPH_RUNS {
-                        break;
-                    }
-                    if i > 0 {
-                        para.runs.push(Run { text: String::new(), props: props.clone(), kind: RunKind::Break });
-                    }
-                    if !piece.is_empty() {
-                        if para.runs.len() >= MAX_PARAGRAPH_RUNS {
-                            break;
-                        }
-                        para.runs.push(Run { text: piece.to_string(), props: props.clone(), kind: RunKind::Text });
-                    }
-                }
+                push_text_runs(&mut para, props, &text);
             }
             "br" => {
                 if para.runs.len() < MAX_PARAGRAPH_RUNS {
-                    let props = e.child("rPr").map(|r| rpr(imp, part, r)).unwrap_or_default();
+                    let props = run_props(imp, part, e);
                     para.runs.push(Run { text: String::new(), props, kind: RunKind::Break });
                 }
             }
             "fld" => {
                 if para.runs.len() < MAX_PARAGRAPH_RUNS {
-                    let props = e.child("rPr").map(|r| rpr(imp, part, r)).unwrap_or_default();
+                    let props = run_props(imp, part, e);
                     let text = e.child("t").map(|t| t.text()).unwrap_or_default();
                     para.runs.push(Run { text, props, kind: RunKind::Field { field: e.attr("type").unwrap_or("").to_string() } });
                 }
             }
-            "m" => {
-                if para.runs.len() < MAX_PARAGRAPH_RUNS {
-                    // Office Math (a14:m) — keep the OMML and its plain text.
-                    let mut text = String::new();
-                    let mut ts = vec![];
-                    e.find_all("t", &mut ts);
-                    for t in ts {
-                        text.push_str(&t.text());
-                    }
-                    let omml = e.elements().map(|c| part.keep_inner(c)).collect::<String>();
-                    para.runs.push(Run { text, props: RunProps::default(), kind: RunKind::Math { omml } });
-                }
-            }
+            "m" => push_math_run(&mut para, part, e),
             "endParaRPr" => para.end_props = rpr(imp, part, e),
             _ => {}
         }

crates/ui-egui/src/show.rs

@@ -739,84 +739,19 @@ fn keyboard(ctx: &egui::Context, show: &mut Show, doc: &Presentation, now: f64,
     let mut keep = true;
     let mut actions = vec![];
     for (k, m) in keys {
-        use egui::Key::*;
-        let presenter_grid = presenter && show.presenter_grid;
-        let grid_open = if presenter { show.presenter_grid } else { show.grid };
-        if grid_open && !matches!(k, Escape | Minus) {
+        let handled = if grid_open(show, presenter) && !matches!(k, egui::Key::Escape | egui::Key::Minus) {
             // See All Slides takes no other keys until a slide is chosen.
-            if k == G && plain(m) {
-                if presenter_grid {
-                    show.presenter_grid = false;
-                } else {
-                    show.grid = false;
-                }
+            if k == egui::Key::G && plain(m) {
+                close_grid(show, presenter);
+                true
             } else {
-                continue;
+                false
             }
         } else {
-            match k {
-                Escape | Minus => {
-                    let cancelled = k == Escape && show.clear_goto();
-                    if !cancelled && !show.back_out(presenter) {
-                        keep = false;
-                    }
-                }
-                ArrowRight | ArrowDown | Space | PageDown | N => {
-                    if show.ended {
-                        keep = false;
-                    } else {
-                        actions.push(show.state.next(doc));
-                    }
-                }
-                Enter => {
-                    if !show.goto.is_empty() {
-                        if let Some(i) = show.take_goto()
-                            && i < doc.slides.len()
-                        {
-                            actions.push(show.state.goto(doc, i));
-                        }
-                    } else if show.ended {
-                        keep = false;
-                    } else {
-                        actions.push(show.state.next(doc));
-                    }
-                }
-                Backspace if !show.goto.is_empty() => {
-                    show.goto.pop();
-                }
-                ArrowLeft | ArrowUp | Backspace | PageUp | P if !(k == P && m.command) => {
-                    show.ended = false;
-                    actions.push(show.state.prev(doc));
-                }
-                P if m.command => show.set_pen(!show.pen),
-                L if m.command => show.set_laser(!show.laser),
-                A if m.command => show.arrow(),
-                Home => {
-                    if let Some(i) = show.state.first_index(doc) {
-                        actions.push(show.state.goto(doc, i));
-                    }
-                }
-                End => {
-                    if let Some(i) = show.state.last_index(doc) {
-                        actions.push(show.state.goto(doc, i));
-                    }
-                }
-                B | Period => toggle_blank(show, Color32::BLACK),
-                W | Comma => toggle_blank(show, Color32::WHITE),
-                E => show.strokes.clear(),
-                G if plain(m) => show.open_grid(presenter),
-                Z if plain(m) => show.zoom_key(),
-                T if plain(m) => show.timer.toggle(now),
-                R if plain(m) => show.timer.reset(now),
-                OpenBracket if plain(m) => show.nudge_notes(-NOTES_STEP),
-                CloseBracket if plain(m) => show.nudge_notes(NOTES_STEP),
-                Num0 | Num1 | Num2 | Num3 | Num4 | Num5 | Num6 | Num7 | Num8 | Num9 => {
-                    if let Some(d) = k.name().chars().next() {
-                        show.push_slide_number(d, now);
-                    }
-                }
-                _ => continue,
-            }
+            navigation_key(show, doc, k, m, presenter, &mut keep, &mut actions) || tool_key(show, k, m, now, presenter)
+        };
+        if !handled {
+            continue;
         }
         // The same key must not also activate a focused presenter button or be handled again
         // by an embedded presenter viewport.
@@ -831,6 +766,111 @@ fn keyboard(ctx: &egui::Context, show: &mut Show, doc: &Presentation, now: f64,
     keep
 }
 
+/// Whether the grid that the presenter key G opens is showing in this viewport.
+fn grid_open(show: &Show, presenter: bool) -> bool {
+    if presenter { show.presenter_grid } else { show.grid }
+}
+
+/// Closes the grid that G opens in this viewport.
+fn close_grid(show: &mut Show, presenter: bool) {
+    if presenter && show.presenter_grid {
+        show.presenter_grid = false;
+    } else {
+        show.grid = false;
+    }
+}
+
+/// Keys that move through the show or end it. False when `k` isn't one of them.
+fn navigation_key(
+    show: &mut Show,
+    doc: &Presentation,
+    k: egui::Key,
+    m: egui::Modifiers,
+    presenter: bool,
+    keep: &mut bool,
+    actions: &mut Vec<ShowAction>,
+) -> bool {
+    use egui::Key::*;
+    match k {
+        Escape | Minus => {
+            let cancelled = k == Escape && show.clear_goto();
+            if !cancelled && !show.back_out(presenter) {
+                *keep = false;
+            }
+        }
+        ArrowRight | ArrowDown | Space | PageDown | N => {
+            if show.ended {
+                *keep = false;
+            } else {
+                actions.push(show.state.next(doc));
+            }
+        }
+        Enter => {
+            if !show.goto.is_empty() {
+                if let Some(i) = show.take_goto()
+                    && i < doc.slides.len()
+                {
+                    actions.push(show.state.goto(doc, i));
+                }
+            } else if show.ended {
+                *keep = false;
+            } else {
+                actions.push(show.state.next(doc));
+            }
+        }
+        Backspace if !show.goto.is_empty() => {
+            show.goto.pop();
+        }
+        ArrowLeft | ArrowUp | Backspace | PageUp => {
+            show.ended = false;
+            actions.push(show.state.prev(doc));
+        }
+        P if m.command => show.set_pen(!show.pen),
+        P => {
+            show.ended = false;
+            actions.push(show.state.prev(doc));
+        }
+        Home => {
+            if let Some(i) = show.state.first_index(doc) {
+                actions.push(show.state.goto(doc, i));
+            }
+        }
+        End => {
+            if let Some(i) = show.state.last_index(doc) {
+                actions.push(show.state.goto(doc, i));
+            }
+        }
+        _ => return false,
+    }
+    true
+}
+
+/// Pen, laser, blanking, zoom, the timer, note size and typed slide numbers. False when `k`
+/// isn't one of them.
+fn tool_key(show: &mut Show, k: egui::Key, m: egui::Modifiers, now: f64, presenter: bool) -> bool {
+    use egui::Key::*;
+    match k {
+        L if m.command => show.set_laser(!show.laser),
+        A if m.command => show.arrow(),
+        B | Period => toggle_blank(show, Color32::BLACK),
+        W | Comma => toggle_blank(show, Color32::WHITE),
+        E => show.strokes.clear(),
+        G if plain(m) => show.open_grid(presenter),
+        Z if plain(m) => show.zoom_key(),
+        T if plain(m) => show.timer.toggle(now),
+        R if plain(m) => show.timer.reset(now),
+        OpenBracket if plain(m) => show.nudge_notes(-NOTES_STEP),
+        CloseBracket if plain(m) => show.nudge_notes(NOTES_STEP),
+        Num0 | Num1 | Num2 | Num3 | Num4 | Num5 | Num6 | Num7 | Num8 | Num9 => {
+            if let Some(d) = k.name().chars().next() {
+                show.push_slide_number(d, now);
+            }
+        }
+        _ => return false,
+    }
+    true
+}
+
 /// Draw and advance the active show, handling hidden frames, timing, input, transitions, and
 /// presenter/reading presentation.
 pub fn ui(app: &mut SlideApp, ui: &mut Ui) {

packaging/windows/deckcraft.wxs

@@ -58,9 +58,15 @@
       <Component Id="DeckcraftApp" Guid="D87EA444-41A4-4C19-922E-66D1A29461FF">
         <File Id="DeckcraftNotice" Source="$(var.BinDir)\NOTICE" KeyPath="no" />
         <File Id="DeckcraftExe" Source="$(var.BinDir)\deckcraft.exe" KeyPath="yes">
+          <!-- No Icon here, so the shortcut shows the icon embedded in deckcraft.exe (build.rs). An
+               MSI Icon is cached under %APPDATA%\Microsoft\Installer\{ProductCode}\, and every
+               update is a major upgrade with a new ProductCode that deletes the old folder. A
+               taskbar pin is a copy of this shortcut that the update never touches, so it kept
+               pointing at the deleted file and lost its icon (#1114). The exe's path survives
+               updates; the shortcut is not advertised, so it targets that path and its icon is the
+               exe's own. The Icon above stays for ARPPRODUCTICON (Apps and features). -->
           <Shortcut Id="StartMenuShortcut" Directory="ProgramMenuFolder" Name="DeckCraft"
-              Description="DeckCraft presentations and slide shows" WorkingDirectory="INSTALLFOLDER" Advertise="yes"
-              Icon="DeckcraftIcon.ico" IconIndex="0" />
+              Description="DeckCraft presentations and slide shows" WorkingDirectory="INSTALLFOLDER" Advertise="no" />
         </File>
 
         <!-- .deckcraft: DeckCraft's own format. -->

packaging/windows/package.ps1

@@ -83,6 +83,15 @@ foreach ($check in @(@('deckcraft.exe', 2), @('deckcraft-cli.exe', 3))) {
   if ($h.Subsystem -ne $check[1]) { throw "$($check[0]) has PE subsystem $($h.Subsystem), expected $($check[1])" }
   Write-Output "ok $($check[0]): $Arch, PE subsystem $($h.Subsystem)"
 }
+
+# The Start menu shortcut is not advertised and names no MSI icon, so it shows the icon embedded in
+# deckcraft.exe (#1114). -SkipBuild takes whatever binary is already there, which may come from a
+# build without DECKCRAFT_REQUIRE_WINRES=1 whose resource compilation failed: refuse an exe with no
+# icon instead of shipping shortcuts with a generic one. ExtractIconEx with index -1 counts them.
+Add-Type -Namespace DeckCraft -Name Shell -MemberDefinition '[DllImport("shell32.dll", CharSet = CharSet.Unicode)] public static extern uint ExtractIconEx(string file, int index, IntPtr[] large, IntPtr[] small, uint icons);'
+$Icons = [DeckCraft.Shell]::ExtractIconEx((Join-Path $Bin 'deckcraft.exe'), -1, $null, $null, 0)
+if ($Icons -eq 0) { throw 'deckcraft.exe has no embedded icon: rebuild it with DECKCRAFT_REQUIRE_WINRES=1 (omit -SkipBuild)' }
+Write-Output "ok deckcraft.exe: $Icons embedded icon group(s)"
 $Stage = Join-Path $TargetDir "windows-package\$Arch"
 Remove-Item -Recurse -Force $Stage -ErrorAction SilentlyContinue
 New-Item -ItemType Directory -Force -Path $Stage | Out-Null