Skip transparent pixels and quantise while compositing

Much of a layer buffer is often transparent (a stroked ring, a glow-padded
text layer, a precomp canvas), yet every destination pixel sampled it
bilinearly, and 8/16 bpc comps then re-quantised the layer's whole rotated
bounding box.

- Bilinear warps skip 16x16 source tiles whose pixels (and a one-pixel
  apron) all have zero alpha: their samples are transparent and were never
  drawn.
- The fast affine warp reports the rectangle it changed and, given the
  comp's levels, quantises each pixel as it blends it, so draw_comp has
  nothing left to re-quantise for that layer. quantize_px moves to
  effectcraft-raster for this.
- quantize_rect skips runs of four all-zero (+0.0) pixels, which quantise
  to themselves.
- to_rgba8 / to_rgba8_over convert row by row instead of per 4-byte item.

Every pixel stays bit-identical: tests compare the warp against the
reference path, fused quantising against quantising after blending, and
the zero-run skip against per-pixel quantising.

With the previous commit, demo comp to H.264 (`cargo xtask bench --scenario
demo --format h264 --frames 300`, 11 interleaved runs on a 16-thread Ryzen):
wall 9.26 s -> 6.97 s median (-24.8%), 8.55 s -> 6.55 s min (-23.3%); CPU
time -23%. Output byte-identical.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

crates/render/src/color.rs

@@ -21,7 +21,7 @@
 
 use effectcraft_color::{ColorSpace, Conversion};
 use effectcraft_project::{Footage, HdrMode, ProjectSettings};
-use effectcraft_raster::Image;
+use effectcraft_raster::{Image, quantize_px};
 use rayon::prelude::*;
 
 /// The colour pipeline of a project (see the module docs).
@@ -218,27 +218,6 @@ pub fn quantize(img: &mut Image, levels: f32) {
     quantize_rect(img, PxRect { x0: 0, y0: 0, x1: img.width, y1: img.height }, levels);
 }
 
-/// Clamp (colour ≤ alpha ≤ 1, NaN → 0) and round half up; values are non-negative after the
-/// clamp, so the integer conversion rounds like `round()` and vectorises.
-#[allow(clippy::manual_clamp)] // max/min map NaN to 0; clamp would keep NaN
-#[inline(always)]
-fn quantize_px(p: &mut [f32; 4], levels: f32, inv: f32) {
-    let a = round_half_up(p[3].max(0.0).min(1.0) * levels) * inv;
-    let hi = [a, a, a, a];
-    for c in 0..3 {
-        p[c] = round_half_up(p[c].max(0.0).min(hi[c]) * levels) * inv;
-    }
-    p[3] = a;
-}
-
-/// `v` (0..=65 535) rounded half up. The same either way; the fastest on each target: x86-64's
-/// baseline converts to integers in one instruction but has no `floor` (SSE4.1), while in wasm
-/// `floor` is one instruction and converting back from `u32` several (about 1.5× faster there).
-#[inline(always)]
-fn round_half_up(v: f32) -> f32 {
-    if cfg!(target_arch = "wasm32") { (v + 0.5).floor() } else { (v + 0.5) as u32 as f32 }
-}
-
 /// [`quantize`] restricted to a rectangle (already clamped to the image), row-parallel.
 fn quantize_rect(img: &mut Image, r: PxRect, levels: f32) {
     if r.x0 >= r.x1 || r.y0 >= r.y1 {
@@ -248,7 +227,17 @@ fn quantize_rect(img: &mut Image, r: PxRect, levels: f32) {
     let w = img.width as usize;
     let (x0, x1) = (r.x0 as usize, r.x1 as usize);
     img.data[r.y0 as usize * w..r.y1 as usize * w].par_chunks_mut(w).for_each(|row| {
-        for p in &mut row[x0..x1] {
+        // Runs of all-zero pixels (+0.0 in every channel) quantise to themselves: skip them a
+        // cache line at a time (sparse shape, precomp and glow buffers are mostly empty).
+        let (blocks, rest) = row[x0..x1].as_chunks_mut::<4>();
+        for b in blocks {
+            if b.as_flattened().iter().fold(0, |a, v| a | v.to_bits()) != 0 {
+                for p in b {
+                    quantize_px(p, levels, inv);
+                }
+            }
+        }
+        for p in rest {
             quantize_px(p, levels, inv);
         }
     });
@@ -307,6 +296,39 @@ pub fn output_hdr(img: &mut Image, to_linear: Option<Conversion>, mode: HdrMode,
 
 #[cfg(test)]
 mod tests {
+    use super::*;
+
+    /// Skipping all-zero runs leaves every pixel as quantising it one by one does, bit for bit
+    /// (negative zeros and lone pixels in empty runs included).
+    #[test]
+    fn quantize_skips_only_zero_runs() {
+        let mut img = Image::new(23, 7);
+        for (i, p) in img.data.iter_mut().enumerate() {
+            *p = match i % 11 {
+                0 => [0.3, -0.2, 0.71, 0.9],
+                3 => [-0.0, 0.0, 0.0, 0.0],
+                5 => [0.0, 0.0, 0.0, -0.0],
+                7 => [1.7, 0.4, f32::NAN, 0.5],
+                _ => [0.0; 4],
+            };
+        }
+        for levels in [255.0f32, 32_768.0] {
+            let mut want = img.clone();
+            for p in &mut want.data {
+                quantize_px(p, levels, 1.0 / levels);
+            }
+            for r in [PxRect { x0: 0, y0: 0, x1: 23, y1: 7 }, PxRect { x0: 3, y0: 1, x1: 20, y1: 6 }] {
+                let mut got = img.clone();
+                quantize_rect(&mut got, r, levels);
+                for (i, (g, p)) in got.data.iter().zip(&img.data).enumerate() {
+                    let (x, y) = (i as u32 % 23, i as u32 / 23);
+                    let w = if x >= r.x0 && x < r.x1 && y >= r.y0 && y < r.y1 { want.data[i] } else { *p };
+                    assert_eq!(g.map(f32::to_bits), w.map(f32::to_bits), "{levels} {r:?} px ({x}, {y})");
+                }
+            }
+        }
+    }
+
     /// Both forms `round_half_up` picks between agree on everything quantising feeds it:
     /// 0..=levels, exact halves included.
     #[test]

crates/render/src/layers.rs

@@ -342,11 +342,26 @@ impl<'a> Renderer<'a> {
 
     /// Draw a processed layer into `target` (transform, motion blur) with `mode`/`opacity`.
     pub(crate) fn place(&self, ctx: &EvalCtx, layer: &Layer, buf: &Buf, target: &mut Image, mode: BlendMode, opacity: f32) {
+        self.place_changed(ctx, layer, buf, target, mode, opacity, None);
+    }
+
+    /// [`Self::place`], returning the target pixels it changed when the warp tracked them (then
+    /// also quantised to `quantize` levels; see [`composite_warp_changed`]).
+    #[allow(clippy::too_many_arguments)]
+    fn place_changed(
+        &self,
+        ctx: &EvalCtx,
+        layer: &Layer,
+        buf: &Buf,
+        target: &mut Image,
+        mode: BlendMode,
+        opacity: f32,
+        quantize: Option<f32>,
+    ) -> Option<[u32; 4]> {
         let pl = self.placement(ctx, layer, buf);
         let opts = WarpOpts { sampling: pl.sampling, opacity, mode, seed: pl.seed, clip: None };
         if let [m] = pl.matrices.as_slice() {
-            composite_warp(target, &buf.img, m, &opts);
-            return;
+            return composite_warp_changed(target, &buf.img, m, &opts, quantize);
         }
         // Sub-samples sum straight into one buffer (each warp is row-parallel).
         let mut acc = Image::new(target.width, target.height);
@@ -355,6 +370,7 @@ impl<'a> Renderer<'a> {
             effectcraft_raster::accumulate_warp(&mut acc, &buf.img, m, opts.sampling, k);
         }
         target.blend_from(&acc, mode, opacity, pl.seed);
+        None
     }
 
     /// Draw a 2D layer into `canvas`; returns the region it may have changed.
@@ -377,8 +393,14 @@ impl<'a> Renderer<'a> {
         let st = self.styled_to_blend(st);
         let t1 = web_time::Instant::now();
         let changed = if st.plain {
-            self.composite_layer(ctx, layer, &st.body, canvas, opacity);
-            self.footprint(ctx, layer, &st.body, opacity, blank)
+            // Pixels the warp tracks are quantised as it blends them (nothing left to do); the
+            // footprint is still needed when it doesn't, or when no pixel needs quantising.
+            let footprint = self.footprint(ctx, layer, &st.body, opacity, blank);
+            let quantize = if footprint == Region::Empty { None } else { self.pipe.levels };
+            match self.composite_layer(ctx, layer, &st.body, canvas, opacity, quantize) {
+                Some(_) if quantize.is_some() => Region::Empty,
+                drawn => narrow(footprint, drawn),
+            }
         } else {
             self.composite_styled(ctx, layer, &st, canvas, opacity);
             Region::Full
@@ -432,16 +454,18 @@ impl<'a> Renderer<'a> {
         });
     }
 
-    fn composite_layer(&self, ctx: &EvalCtx, layer: &Layer, buf: &Buf, canvas: &mut Image, opacity: f32) {
+    /// Composite a plain layer; returns the canvas pixels it changed when they are known (then
+    /// also quantised to `quantize` levels).
+    fn composite_layer(&self, ctx: &EvalCtx, layer: &Layer, buf: &Buf, canvas: &mut Image, opacity: f32, quantize: Option<f32>) -> Option<[u32; 4]> {
         let matte = layer.track_matte.is_some_and(|tm| tm.layer != layer.id && ctx.comp.layer(tm.layer).is_some());
         if !matte && !layer.preserve_transparency {
-            self.place(ctx, layer, buf, canvas, layer.blend_mode, opacity);
-            return;
+            return self.place_changed(ctx, layer, buf, canvas, layer.blend_mode, opacity, quantize);
         }
         // Render in isolation, then matte / preserve transparency, then blend.
         let mut iso = Image::new(canvas.width, canvas.height);
         self.place(ctx, layer, buf, &mut iso, BlendMode::Normal, 1.0);
         self.composite_iso(ctx, layer, iso, canvas, opacity);
+        None
     }
 
     /// Apply the track matte / preserve transparency to an isolated layer render, then blend.
@@ -556,3 +580,21 @@ pub struct Placement {
     /// Dissolve noise seed.
     pub seed: u32,
 }
+
+/// The part of `footprint` a warp actually changed (`drawn`, see [`composite_warp_changed`]):
+/// pixels it left alone are still quantised.
+fn narrow(footprint: Region, drawn: Option<[u32; 4]>) -> Region {
+    let Some([x0, y0, x1, y1]) = drawn else { return footprint };
+    if x0 >= x1 || y0 >= y1 {
+        return Region::Empty;
+    }
+    let r = color::PxRect { x0, y0, x1, y1 };
+    match footprint {
+        Region::Empty => Region::Empty,
+        Region::Full => Region::Rect(r),
+        Region::Rect(f) => {
+            let q = color::PxRect { x0: r.x0.max(f.x0), y0: r.y0.max(f.y0), x1: r.x1.min(f.x1), y1: r.y1.min(f.y1) };
+            if q.x0 >= q.x1 || q.y0 >= q.y1 { Region::Empty } else { Region::Rect(q) }
+        }
+    }
+}

crates/render/src/lib.rs

@@ -32,7 +32,7 @@ use effectcraft_effects::{Buf, EffectCtx, EffectEnv, EffectHost, LayerPixels, Pa
 use effectcraft_geom::{Mat3, Mat4, vec2};
 use effectcraft_project::{Comp, Footage, FootageKind, FrameBlend, GroupKind, ItemId, ItemKind, Layer, LayerSource, MatteKind, Project, Quality, Sampling};
 pub use effectcraft_raster::Image;
-use effectcraft_raster::{WarpOpts, composite_warp};
+use effectcraft_raster::{WarpOpts, composite_warp_changed};
 use effectcraft_time::{FrameRate, TICKS_PER_SECOND, Tick};
 pub use eval::{EvalCtx, ExprHost, effect_bounds, source_size};
 use rayon::prelude::*;