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

ROADMAP.md

@@ -137,7 +137,6 @@ 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) |
@@ -294,7 +293,6 @@ 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

@@ -430,27 +430,4 @@ 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

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

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/hevc/src/slicedec.rs

@@ -1366,106 +1366,84 @@ impl<'a> SliceDecoder<'a> {
 
     /// prediction_unit(): parse, derive motion, store it and predict. Returns merge_flag.
     fn prediction_unit(&mut self, g: PbGeom, part: PartMode, skip: bool) -> Result<bool> {
+        let PbGeom { xc, yc, ncb, xp, yp, w, h, part_idx } = g;
         let merge = if skip { true } else { self.dd(MERGE_FLAG) == 1 };
         let mvf = if merge {
-            let idx = self.merge_idx();
-            self.derive_merge(g, part, idx as usize)
+            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(PbGeom { xc, yc, ncb, xp, yp, w, h, part_idx }, part, idx as usize)
         } else {
-            self.amvp_motion(g)?
+            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(PbGeom { xc, yc, ncb, xp, yp, w, h, part_idx }, l, ref_idx[l] as usize, mvp_flag[l] as usize);
+                // mvLX = (mvp + mvd) wrapped to 16 bits (8-94)
+                f.mv[l] = [(mvp[0] as i32 + mvd[l][0]) as i16, (mvp[1] as i32 + mvd[l][1]) as i16];
+            }
+            f
         };
         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");
             }
         }
-        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) {
+        // store motion
         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]> {

packaging/windows/deckcraft.wxs

@@ -58,15 +58,9 @@
       <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="no" />
+              Description="DeckCraft presentations and slide shows" WorkingDirectory="INSTALLFOLDER" Advertise="yes"
+              Icon="DeckcraftIcon.ico" IconIndex="0" />
         </File>
 
         <!-- .deckcraft: DeckCraft's own format. -->

packaging/windows/package.ps1

@@ -83,15 +83,6 @@ 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