Merge pull request #2518 from Silvertree2010/perf/export-speed

Export: encode while the next batch renders, skip transparent pixels when compositing

crates/export/src/encode.rs

@@ -26,9 +26,38 @@ pub(crate) struct Packet {
     pub(crate) cto: i32,
 }
 
-pub(crate) trait VideoEncoder {
+/// A frame for the movie encoder: 8-bit RGBA, or 8-bit 4:2:0 planes already converted on the
+/// render threads (for encoders whose [`VideoEncoder::takes_yuv420`] is true).
+pub(crate) enum Pixels {
+    Rgba(Vec<u8>),
+    Yuv420(Yuv420),
+}
+
+/// BT.709 limited-range 8-bit 4:2:0 planes ([`rgba_to_yuv420`]).
+pub(crate) struct Yuv420 {
+    y: Vec<u8>,
+    u: Vec<u8>,
+    v: Vec<u8>,
+}
+
+impl Yuv420 {
+    fn from_rgba(rgba: &[u8], w: usize, h: usize) -> Yuv420 {
+        let (mut y, mut u, mut v) = (vec![], vec![], vec![]);
+        rgba_to_yuv420(rgba, w, h, &mut y, &mut u, &mut v);
+        Yuv420 { y, u, v }
+    }
+}
+
+pub(crate) trait VideoEncoder: Send {
     fn sample_entry(&self) -> SampleEntry;
     fn encode(&mut self, rgba: &[u8], index: u64) -> Result<Vec<Packet>>;
+    /// Whether frames should come as [`Pixels::Yuv420`] (converted while other frames render).
+    fn takes_yuv420(&self) -> bool {
+        false
+    }
+    fn encode_yuv420(&mut self, _frame: &Yuv420, _index: u64) -> Result<Vec<Packet>> {
+        Err(ExportError::Unsupported("this encoder takes RGBA frames".into()))
+    }
     fn flush(&mut self) -> Result<Vec<Packet>>;
     /// Edit-list media start (B-frame delay) in the track timescale.
     fn media_start(&self) -> Option<i64> {
@@ -73,9 +102,18 @@ impl VideoEncoder for H264 {
         SampleEntry::avc(cfg, self.w as u16, self.h as u16)
     }
     fn encode(&mut self, rgba: &[u8], index: u64) -> Result<Vec<Packet>> {
-        rgba_to_yuv420(rgba, self.w as usize, self.h as usize, &mut self.y, &mut self.u, &mut self.v);
+        let mut planes = Yuv420 { y: std::mem::take(&mut self.y), u: std::mem::take(&mut self.u), v: std::mem::take(&mut self.v) };
+        rgba_to_yuv420(rgba, self.w as usize, self.h as usize, &mut planes.y, &mut planes.u, &mut planes.v);
+        let r = self.encode_yuv420(&planes, index);
+        (self.y, self.u, self.v) = (planes.y, planes.u, planes.v);
+        r
+    }
+    fn takes_yuv420(&self) -> bool {
+        true
+    }
+    fn encode_yuv420(&mut self, p: &Yuv420, index: u64) -> Result<Vec<Packet>> {
         let cw = (self.w as usize).div_ceil(2);
-        let frame = filmcraft_h264enc::YuvFrame { y: &self.y, u: &self.u, v: &self.v, y_stride: self.w as usize, uv_stride: cw };
+        let frame = filmcraft_h264enc::YuvFrame { y: &p.y, u: &p.u, v: &p.v, y_stride: self.w as usize, uv_stride: cw };
         let ps = self.enc.try_encode(&frame, index as i64 * self.rate.den).map_err(enc)?;
         Ok(self.packets(ps))
     }
@@ -88,6 +126,25 @@ impl VideoEncoder for H264 {
     }
 }
 
+/// `v / 255` for every 8-bit `v` (the same values as dividing).
+pub(crate) const UNIT8: [f32; 256] = {
+    let mut t = [0.0; 256];
+    let mut i = 0;
+    while i < 256 {
+        t[i] = i as f32 / 255.0;
+        i += 1;
+    }
+    t
+};
+
+/// `v.round()` for `0 <= v < 2^23`, the range the colour conversions round in: truncation and
+/// the fraction are exact there, so this is `f32::round` without its libm call.
+#[inline(always)]
+pub(crate) fn round_pos(v: f32) -> f32 {
+    let t = v as u32 as f32;
+    if v - t >= 0.5 { t + 1.0 } else { t }
+}
+
 /// BT.709 limited-range 8-bit 4:2:0 from RGBA8 (2×2 chroma average; alpha ignored).
 pub(crate) fn rgba_to_yuv420(rgba: &[u8], w: usize, h: usize, y: &mut Vec<u8>, u: &mut Vec<u8>, v: &mut Vec<u8>) {
     let (cw, ch) = (w.div_ceil(2), h.div_ceil(2));
@@ -95,14 +152,13 @@ pub(crate) fn rgba_to_yuv420(rgba: &[u8], w: usize, h: usize, y: &mut Vec<u8>, u
     u.resize(cw * ch, 0);
     v.resize(cw * ch, 0);
     y.par_chunks_mut(w * 2).zip(u.par_chunks_mut(cw).zip(v.par_chunks_mut(cw))).enumerate().for_each(|(cy, (yr, (ur, vr)))| {
-        let rows = yr.len() / w;
         let mut acc = vec![(0f32, 0f32, 0f32); cw];
-        for dy in 0..rows {
+        for (dy, yrow) in yr.chunks_exact_mut(w).enumerate() {
             let src = &rgba[(cy * 2 + dy) * w * 4..][..w * 4];
-            for x in 0..w {
-                let (r, g, b) = (src[x * 4] as f32 / 255.0, src[x * 4 + 1] as f32 / 255.0, src[x * 4 + 2] as f32 / 255.0);
+            for (x, (s, yo)) in src.as_chunks::<4>().0.iter().zip(yrow.iter_mut()).enumerate() {
+                let (r, g, b) = (UNIT8[s[0] as usize], UNIT8[s[1] as usize], UNIT8[s[2] as usize]);
                 let l = 0.2126 * r + 0.7152 * g + 0.0722 * b;
-                yr[dy * w + x] = (16.0 + 219.0 * l).round().clamp(1.0, 254.0) as u8;
+                *yo = round_pos(16.0 + 219.0 * l).clamp(1.0, 254.0) as u8;
                 let a = &mut acc[x / 2];
                 a.0 += (b - l) / 1.8556;
                 a.1 += (r - l) / 1.5748;
@@ -110,8 +166,8 @@ pub(crate) fn rgba_to_yuv420(rgba: &[u8], w: usize, h: usize, y: &mut Vec<u8>, u
             }
         }
         for (cx, a) in acc.iter().enumerate() {
-            ur[cx] = (128.0 + 224.0 * a.0 / a.2).round().clamp(1.0, 254.0) as u8;
-            vr[cx] = (128.0 + 224.0 * a.1 / a.2).round().clamp(1.0, 254.0) as u8;
+            ur[cx] = round_pos(128.0 + 224.0 * a.0 / a.2).clamp(1.0, 254.0) as u8;
+            vr[cx] = round_pos(128.0 + 224.0 * a.1 / a.2).clamp(1.0, 254.0) as u8;
         }
     });
 }
@@ -265,15 +321,34 @@ pub(crate) async fn movie(job: &Cx<'_>, comp: &Comp, w: u32, h: u32, st: &mut St
     };
     let brand = if matches!(fmt, OutputFormat::ProRes | OutputFormat::Uncompressed) { Brand::Mov } else { Brand::Mp4 };
     let batch = batch_size();
-    let render = |_: u64, img: effectcraft_raster::Image| -> Vec<u8> { job.pixels(&img, comp, channels, w, h) };
+    let yuv = venc.takes_yuv420();
+    let render = |_: u64, img: effectcraft_raster::Image| -> Pixels {
+        let px = job.pixels(&img, comp, channels, w, h);
+        if yuv { Pixels::Yuv420(Yuv420::from_rgba(&px, w as usize, h as usize)) } else { Pixels::Rgba(px) }
+    };
+    let encode = |venc: &mut Box<dyn VideoEncoder>, px: &Pixels, index: u64| match px {
+        Pixels::Rgba(rgba) => venc.encode(rgba, index),
+        Pixels::Yuv420(planes) => venc.encode_yuv420(planes, index),
+    };
+
+    let beside = crate::pipeline::Beside::shared();
+    let total = st.total;
 
     // Encode the first batch before creating tracks: encoders finalise their config on frame 1.
-    let first_end = batch.min(st.total);
-    let first: Vec<Vec<u8>> = job.frames(comp, (0..first_end).collect(), render).await;
-    let mut pending = Vec::new();
-    for (k, px) in first.iter().enumerate() {
-        pending.extend(venc.encode(px, k as u64)?);
-    }
+    // Each batch encodes while the next one renders.
+    let first_end = batch.min(total);
+    let first: Vec<Pixels> = job.frames(comp, (0..first_end).collect(), render).await;
+    let next_end = (first_end + batch).min(total);
+    let (pending, mut ready) = job
+        .frames_beside(beside, comp, (first_end..next_end).collect(), render, || -> Result<Vec<Packet>> {
+            let mut pending = Vec::new();
+            for (k, px) in first.iter().enumerate() {
+                pending.extend(encode(&mut venc, px, k as u64)?);
+            }
+            Ok(pending)
+        })
+        .await?;
+    let pending = pending?;
     drop(first);
     st.advance(first_end)?;
 
@@ -351,19 +426,25 @@ pub(crate) async fn movie(job: &Cx<'_>, comp: &Comp, w: u32, h: u32, st: &mut St
         }
         Ok(())
     };
-    let total = st.total;
     let frame_end_time = |k: u64| -> Tick { if k >= total { span_end } else { job.settings.frame_time(comp, k) } };
     write_video(&mut mux, pending)?;
     write_audio_until(&mut mux, frame_end_time(first_end), &mut cursor)?;
     let mut i = first_end;
     while i < total {
         let end = (i + batch).min(total);
-        let frames: Vec<Vec<u8>> = job.frames(comp, (i..end).collect(), render).await;
-        for (k, px) in frames.iter().enumerate() {
-            let ps = venc.encode(px, i + k as u64)?;
-            write_video(&mut mux, ps)?;
-        }
-        write_audio_until(&mut mux, frame_end_time(end), &mut cursor)?;
+        let next_end = (end + batch).min(total);
+        let frames = std::mem::take(&mut ready);
+        let (done, next) = job
+            .frames_beside(beside, comp, (end..next_end).collect(), render, || -> Result<()> {
+                for (k, px) in frames.iter().enumerate() {
+                    let ps = encode(&mut venc, px, i + k as u64)?;
+                    write_video(&mut mux, ps)?;
+                }
+                write_audio_until(&mut mux, frame_end_time(end), &mut cursor)
+            })
+            .await?;
+        done?;
+        ready = next;
         st.advance(end - i)?;
         i = end;
     }
@@ -384,6 +465,70 @@ pub(crate) async fn movie(job: &Cx<'_>, comp: &Comp, w: u32, h: u32, st: &mut St
 mod tests {
     use super::*;
 
+    /// The conversion as first written (per-pixel division and `f32::round`).
+    fn yuv420_reference(rgba: &[u8], w: usize, h: usize) -> (Vec<u8>, Vec<u8>, Vec<u8>) {
+        let (cw, ch) = (w.div_ceil(2), h.div_ceil(2));
+        let (mut y, mut u, mut v) = (vec![0u8; w * h], vec![0u8; cw * ch], vec![0u8; cw * ch]);
+        for cy in 0..ch {
+            let mut acc = vec![(0f32, 0f32, 0f32); cw];
+            for row in cy * 2..(cy * 2 + 2).min(h) {
+                for x in 0..w {
+                    let s = &rgba[(row * w + x) * 4..];
+                    let (r, g, b) = (s[0] as f32 / 255.0, s[1] as f32 / 255.0, s[2] as f32 / 255.0);
+                    let l = 0.2126 * r + 0.7152 * g + 0.0722 * b;
+                    y[row * w + x] = (16.0 + 219.0 * l).round().clamp(1.0, 254.0) as u8;
+                    let a = &mut acc[x / 2];
+                    a.0 += (b - l) / 1.8556;
+                    a.1 += (r - l) / 1.5748;
+                    a.2 += 1.0;
+                }
+            }
+            for (cx, a) in acc.iter().enumerate() {
+                u[cy * cw + cx] = (128.0 + 224.0 * a.0 / a.2).round().clamp(1.0, 254.0) as u8;
+                v[cy * cw + cx] = (128.0 + 224.0 * a.1 / a.2).round().clamp(1.0, 254.0) as u8;
+            }
+        }
+        (y, u, v)
+    }
+
+    #[test]
+    fn round_pos_is_round() {
+        for i in 0..2_000_000u32 {
+            let v = i as f32 * 0.000_25;
+            assert_eq!(round_pos(v).to_bits(), v.round().to_bits(), "{v}");
+            let n = f32::from_bits(v.to_bits().wrapping_sub(1)).max(0.0);
+            assert_eq!(round_pos(n).to_bits(), n.round().to_bits(), "{n}");
+        }
+        for k in 0..1024u32 {
+            for v in [k as f32 + 0.5, f32::from_bits((k as f32 + 0.5).to_bits() - 1), f32::from_bits((k as f32 + 0.5).to_bits() + 1)] {
+                assert_eq!(round_pos(v), v.round(), "{v}");
+            }
+        }
+        for (i, u) in UNIT8.iter().enumerate() {
+            assert_eq!(u.to_bits(), (i as f32 / 255.0).to_bits());
+        }
+    }
+
+    #[test]
+    fn yuv420_matches_the_reference() {
+        // Every 8-bit level in every channel, odd sizes, and pseudo-random pixels.
+        for (w, h) in [(1, 1), (2, 2), (3, 5), (17, 9), (64, 33)] {
+            let mut rgba = vec![0u8; w * h * 4];
+            let mut seed = (w * 31 + h) as u32;
+            for (i, b) in rgba.iter_mut().enumerate() {
+                seed = seed.wrapping_mul(1_664_525).wrapping_add(1_013_904_223);
+                *b = if i % 7 == 0 { (i / 4 % 256) as u8 } else { (seed >> 24) as u8 };
+            }
+            let (mut y, mut u, mut v) = (vec![], vec![], vec![]);
+            rgba_to_yuv420(&rgba, w, h, &mut y, &mut u, &mut v);
+            assert_eq!((y, u, v), yuv420_reference(&rgba, w, h), "{w}x{h}");
+        }
+        let grey: Vec<u8> = (0..=255u8).flat_map(|g| [g, g, g, 255, g, 255 - g, g / 2, 255]).collect();
+        let (mut y, mut u, mut v) = (vec![], vec![], vec![]);
+        rgba_to_yuv420(&grey, 512, 1, &mut y, &mut u, &mut v);
+        assert_eq!((y, u, v), yuv420_reference(&grey, 512, 1));
+    }
+
     #[test]
     fn yuv_of_primaries() {
         let rgba = [255u8, 255, 255, 255, 0, 0, 0, 255, 255, 0, 0, 255, 0, 255, 0, 255];

crates/export/src/pipeline.rs

@@ -91,6 +91,24 @@ pub(crate) fn map_frames_while<I: Send, T: Send>(
     out
 }
 
+/// Threads that encode one batch of frames while the next renders ([`Cx::frames_beside`]).
+/// Their own pool: the encoder's parallel steps (colour conversion) would otherwise queue behind
+/// the render's work in the global pool. `None` where threads can't be started (a browser
+/// without them): encoding and rendering then take turns.
+pub(crate) struct Beside(Option<rayon::ThreadPool>);
+
+impl Beside {
+    fn new() -> Beside {
+        Beside(rayon::ThreadPoolBuilder::new().thread_name(|i| format!("export-encode-{i}")).build().ok())
+    }
+
+    /// The process's encode threads, started by the first movie export and shared by all.
+    pub fn shared() -> &'static Beside {
+        static SHARED: std::sync::OnceLock<Beside> = std::sync::OnceLock::new();
+        SHARED.get_or_init(Beside::new)
+    }
+}
+
 /// A job being exported: the job, the project with the Render Settings overrides applied, the
 /// job's layer cache and per-frame timing for the render log.
 pub(crate) struct Cx<'a> {
@@ -214,6 +232,40 @@ impl<'a> Cx<'a> {
         map_frames_while(ks, self.gpu_cap(), gpu_live, |k| self.frame(comp, k), f)
     }
 
+    /// [`Cx::frames`] for `ks` while `encode` runs beside them on `beside`'s threads, so the
+    /// render threads don't wait for the (largely sequential) encoder of the previous batch.
+    /// With deferred readbacks, or where no encode thread could be started, `encode` runs first.
+    pub async fn frames_beside<T: Send, A: Send>(
+        &self,
+        beside: &Beside,
+        comp: &Comp,
+        ks: Vec<u64>,
+        f: impl Fn(u64, Image) -> T + Sync + Send,
+        encode: impl FnOnce() -> A + Send,
+    ) -> crate::Result<(A, Vec<T>)> {
+        match &beside.0 {
+            Some(pool) if !self.deferred() => {
+                let mut done = None;
+                let frames = pool.in_place_scope(|s| {
+                    s.spawn(|_| done = Some(std::panic::catch_unwind(std::panic::AssertUnwindSafe(encode))));
+                    effectcraft_render::passes::block_on(self.frames(comp, ks, f))
+                });
+                match done {
+                    Some(Ok(a)) => Ok((a, frames)),
+                    Some(Err(panic)) => {
+                        let why = panic.downcast_ref::<&str>().map(|s| s.to_string()).or_else(|| panic.downcast_ref::<String>().cloned()).unwrap_or_default();
+                        Err(crate::ExportError::Encode(format!("the encoder stopped unexpectedly: {why}")))
+                    }
+                    None => Err(crate::ExportError::Encode("the encoder stopped unexpectedly".into())),
+                }
+            }
+            _ => {
+                let a = encode();
+                Ok((a, self.frames(comp, ks, f).await))
+            }
+        }
+    }
+
     /// GPU frames rendered at once (#879), decided once per batch: with a GPU accelerator every
     /// frame in flight holds its own working textures, so a whole batch rendered at once can
     /// exhaust VRAM (or shared memory, which crashes instead of falling back). `None` when the
@@ -706,6 +758,45 @@ mod tests {
         assert!(stats.hits >= 9, "{stats:?}");
     }
 
+    #[test]
+    fn frames_render_beside_the_encoder() {
+        use effectcraft_project::render_queue::{OutputFormat, OutputModule, RenderSettings};
+        use effectcraft_render::NoFootage;
+        use effectcraft_time::FrameRate;
+        let mut p = Project::default();
+        let comp = Comp::new(32, 16, FrameRate::FPS_25, Tick::from_seconds_f64(1.0));
+        let cid = p.add_item("Comp", effectcraft_color::Label::Sandstone, None, effectcraft_project::ItemKind::Comp(comp.into()));
+        let (settings, output) = (RenderSettings::default(), OutputModule::for_format(OutputFormat::H264));
+        let job = Job {
+            project: &p,
+            footage: &NoFootage,
+            expr: None,
+            accel: None,
+            comp: cid,
+            settings: &settings,
+            output: &output,
+            path: "/c/out.mp4",
+            sink: None,
+            nested_switches: true,
+            options: Default::default(),
+        };
+        let cx = Cx::new(&job);
+        let comp = cx.comp().unwrap();
+        use effectcraft_render::passes::block_on;
+        let want = block_on(cx.frames(comp, (0..6).collect(), |k, img| (k, img.width, img.height)));
+        // On threads of their own, and taking turns (no threads): the same frames, in order.
+        for beside in [Beside::shared(), &Beside(None)] {
+            let (a, frames) = block_on(cx.frames_beside(beside, comp, (0..6).collect(), |k, img| (k, img.width, img.height), || 7)).unwrap();
+            assert_eq!((a, &frames), (7, &want));
+        }
+        // A panicking encoder fails the export instead of taking the render thread down.
+        let r = block_on(cx.frames_beside(Beside::shared(), comp, vec![0], |k, _| k, || -> u32 { panic!("encoder bug") }));
+        match r {
+            Err(crate::ExportError::Encode(e)) => assert!(e.contains("encoder bug"), "{e}"),
+            other => panic!("{:?}", other.map(|(a, f)| (a, f.len()))),
+        }
+    }
+
     #[test]
     fn interleave_takes_alternate_lines() {
         let a = Image::filled(3, 4, [1.0, 0.0, 0.0, 1.0]);

crates/export/tests/it/main.rs

@@ -4,6 +4,7 @@ mod audio_sample_positions;
 mod hevc_av1_roundtrip;
 #[cfg(not(target_arch = "wasm32"))]
 mod mov_audio_sample_positions;
+mod movie_bytes;
 mod ms_rounded_footage;
 #[cfg(not(target_arch = "wasm32"))]
 mod native_audio_sample_positions;

crates/export/tests/it/movie_bytes.rs

@@ -0,0 +1,104 @@
+//! Movie exports pinned byte for byte: a comp longer than one render batch, with a rotating
+//! stroked ring (mostly transparent, bilinear), a fading half-transparent solid at a fractional
+//! position and a Screen-mode shape, in 8 bpc. The comp is at most four macroblock rows high,
+//! so the H.264 encoder writes one slice per frame on any number of cores. The hashes are those of the files written before
+//! frames rendered while the previous batch encoded, before transparent source tiles were
+//! skipped and before blended pixels were quantised as they were drawn: those changes must not
+//! move a single byte.
+
+use effectcraft_color::{BlendMode, Label};
+use effectcraft_export::render_queue::{OutputFormat, OutputModule, RenderSettings, TimeSpan};
+use effectcraft_export::{Job, Sink, export};
+use effectcraft_keyframe::{Keyframe, Value};
+use effectcraft_project::build::{self, Ids};
+use effectcraft_project::{Comp, ItemId, ItemKind, LayerSource, Project, Solid};
+use effectcraft_render::NoFootage;
+use effectcraft_time::{FrameRate, Tick};
+
+fn fnv1a(data: &[u8]) -> u64 {
+    data.iter().fold(0xcbf2_9ce4_8422_2325u64, |h, &b| (h ^ b as u64).wrapping_mul(0x100_0000_01b3))
+}
+
+fn project() -> (Project, ItemId) {
+    let mut p = Project::default();
+    let mut comp = Comp::new(160, 64, FrameRate::new(30, 1), Tick::from_seconds_f64(2.0));
+    comp.background = [0.05, 0.1, 0.3];
+    let t = Tick::from_seconds_f64;
+
+    let sid = p.add_item("Half", Label::Red, None, ItemKind::Solid(Solid { color: [0.9, 0.5, 0.1], width: 70, height: 40, pixel_aspect: 1.0 }));
+    let mut solid = build::layer(&mut p, &comp, "Half", LayerSource::Solid { item: sid }, (70, 40), None);
+    if let Some(pr) = solid.props.prop_mut("transform/position") {
+        pr.value = Value::Vec3([60.25, 36.5, 0.0]);
+    }
+    if let Some(pr) = solid.props.prop_mut("transform/opacity") {
+        pr.keys = vec![Keyframe::new(t(0.0), Value::Scalar(20.0)), Keyframe::new(t(2.0), Value::Scalar(70.0))];
+    }
+
+    let mut ring = build::layer(&mut p, &comp, "Ring", LayerSource::Shape, (160, 64), None);
+    let mut box_ = build::layer(&mut p, &comp, "Box", LayerSource::Shape, (160, 64), None);
+    let mut next = p.next_id;
+    {
+        let mut ids = Ids(&mut next);
+        let e = build::shape_ellipse(&mut ids, [56.0, 56.0], [0.0, 0.0]);
+        let s = build::shape_stroke(&mut ids, [0.3, 0.8, 1.0, 1.0], 3.0);
+        let g = build::shape_group(&mut ids, "Ring", vec![e, s]);
+        if let Some(c) = ring.props.sub_mut("contents") {
+            c.children.push(g.into());
+        }
+        let r = build::shape_rect(&mut ids, [30.0, 20.0], [25.0, -10.0], 4.0);
+        let f = build::shape_fill(&mut ids, [0.2, 1.0, 0.4, 0.8]);
+        let g = build::shape_group(&mut ids, "Box", vec![r, f]);
+        if let Some(c) = box_.props.sub_mut("contents") {
+            c.children.push(g.into());
+        }
+    }
+    p.next_id = next;
+    if let Some(pr) = ring.props.prop_mut("transform/rotation") {
+        pr.keys = vec![Keyframe::new(t(0.0), Value::Scalar(0.0)), Keyframe::new(t(2.0), Value::Scalar(130.0))];
+    }
+    if let Some(pr) = box_.props.prop_mut("transform/position") {
+        pr.keys = vec![Keyframe::new(t(0.0), Value::Vec3([70.0, 30.0, 0.0])), Keyframe::new(t(2.0), Value::Vec3([95.5, 38.25, 0.0]))];
+    }
+    box_.blend_mode = BlendMode::Screen;
+
+    comp.layers = vec![box_, ring, solid];
+    let cid = p.add_item("Comp 1", Label::Sandstone, None, ItemKind::Comp(comp.into()));
+    (p, cid)
+}
+
+fn export_bytes(format: OutputFormat, name: &str) -> Vec<u8> {
+    let (p, cid) = project();
+    let settings = RenderSettings { time_span: TimeSpan::LengthOfComp, ..Default::default() };
+    let om = OutputModule::for_format(format);
+    let out = std::sync::Arc::new(std::sync::Mutex::new(Vec::new()));
+    let file = out.clone();
+    let sink: &Sink = &move |_: &str, data: Vec<u8>| *file.lock().unwrap() = data;
+    let job = Job {
+        project: &p,
+        footage: &NoFootage,
+        expr: None,
+        accel: None,
+        comp: cid,
+        settings: &settings,
+        output: &om,
+        path: name,
+        sink: Some(sink),
+        options: Default::default(),
+        nested_switches: true,
+    };
+    let r = export(&job, &mut |_| true).expect("export");
+    assert_eq!(r.frames, 60);
+    std::mem::take(&mut *out.lock().unwrap())
+}
+
+#[test]
+fn h264_movie_bytes_are_pinned() {
+    let data = export_bytes(OutputFormat::H264, "/out/pinned.mp4");
+    assert_eq!((data.len(), fnv1a(&data)), (104_161, 10_466_446_817_305_548_344), "H.264 MP4");
+}
+
+#[test]
+fn prores_movie_bytes_are_pinned() {
+    let data = export_bytes(OutputFormat::ProRes, "/out/pinned.mov");
+    assert_eq!((data.len(), fnv1a(&data)), (270_094, 8_038_709_743_204_075_800), "ProRes MOV");
+}

crates/raster/src/lib.rs

@@ -18,10 +18,32 @@ use rayon::prelude::*;
 
 pub use blur::{box_blur, directional_blur, gaussian_blur, gaussian_blur_exact, gaussian_is_exact, radial_blur};
 pub use channels3d::AuxChannels;
-pub use warp::{Sampling, WarpOpts, accumulate_warp, composite_warp, composite_warp_reference, resample};
+pub use warp::{Sampling, WarpOpts, accumulate_warp, composite_warp, composite_warp_changed, composite_warp_reference, resample};
 
 pub type Px = [f32; 4];
 
+/// Clamp (colour ≤ alpha ≤ 1, NaN → 0) and round half up to `levels` steps (`inv` = 1 / levels):
+/// one pixel of an 8/16 bpc buffer. 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)]
+pub fn quantize_px(p: &mut Px, 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 }
+}
+
 #[derive(Clone, Debug, Default, PartialEq)]
 pub struct Image {
     pub width: u32,
@@ -199,26 +221,33 @@ impl Image {
     }
     /// Straight-alpha 8-bit RGBA (sRGB-encoded values as stored).
     pub fn to_rgba8(&self) -> Vec<u8> {
-        let mut out = vec![0u8; self.data.len() * 4];
-        out.par_chunks_mut(4).zip(self.data.par_iter()).for_each(|(o, p)| {
+        self.map_rgba8(|o, p| {
             let a = p[3].clamp(0.0, 1.0);
             let inv = if a > 0.0 { 1.0 / p[3] } else { 0.0 };
             for c in 0..3 {
                 o[c] = ((p[c] * inv).clamp(0.0, 1.0) * 255.0 + 0.5) as u8;
             }
             o[3] = (a * 255.0 + 0.5) as u8;
-        });
-        out
+        })
     }
     /// Composite over an opaque background colour and return opaque 8-bit RGBA.
     pub fn to_rgba8_over(&self, bg: [f32; 3]) -> Vec<u8> {
-        let mut out = vec![0u8; self.data.len() * 4];
-        out.par_chunks_mut(4).zip(self.data.par_iter()).for_each(|(o, p)| {
+        self.map_rgba8(|o, p| {
             let k = 1.0 - p[3].clamp(0.0, 1.0);
             for c in 0..3 {
                 o[c] = ((p[c] + bg[c] * k).clamp(0.0, 1.0) * 255.0 + 0.5) as u8;
             }
             o[3] = 255;
+        })
+    }
+    /// 8-bit RGBA with each pixel written by `f`, row-parallel.
+    fn map_rgba8(&self, f: impl Fn(&mut [u8; 4], &Px) + Sync) -> Vec<u8> {
+        let mut out = vec![0u8; self.data.len() * 4];
+        let w = self.width.max(1) as usize;
+        out.par_chunks_mut(w * 4).zip(self.data.par_chunks(w)).for_each(|(o, row)| {
+            for (o, p) in o.as_chunks_mut::<4>().0.iter_mut().zip(row) {
+                f(o, p);
+            }
         });
         out
     }

crates/raster/src/warp.rs

@@ -57,7 +57,15 @@ fn half(src: &Image) -> Image {
 /// bilinear without bounds checks, integer-translation copies, inlined Normal blending); it
 /// matches [`composite_warp_reference`] to float rounding.
 pub fn composite_warp(dst: &mut Image, src: &Image, m: &Mat3, opts: &WarpOpts) {
-    composite_warp_impl(dst, src, m, opts, true, None);
+    composite_warp_impl(dst, src, m, opts, true, None, None);
+}
+
+/// [`composite_warp`], also returning the destination pixels it changed when it knows them:
+/// `Some([x0, y0, x1, y1])` (half-open, empty when nothing was drawn) on the fast affine path,
+/// `None` otherwise. With `quantize` (an 8/16 bpc destination's levels), the fast path also
+/// quantises each pixel it blends ([`crate::quantize_px`]): exactly when this returns `Some`.
+pub fn composite_warp_changed(dst: &mut Image, src: &Image, m: &Mat3, opts: &WarpOpts, quantize: Option<f32>) -> Option<[u32; 4]> {
+    composite_warp_impl(dst, src, m, opts, true, None, quantize)
 }
 
 /// `dst += warp(src) · weight` (plain sum, no blending): motion-blur sub-samples accumulate
@@ -65,7 +73,7 @@ pub fn composite_warp(dst: &mut Image, src: &Image, m: &Mat3, opts: &WarpOpts) {
 pub fn accumulate_warp(dst: &mut Image, src: &Image, m: &Mat3, sampling: Sampling, weight: f32) {
     let opts = WarpOpts { sampling, ..Default::default() };
     if m.is_affine() && m.inverse().is_some_and(|i| i.is_affine()) {
-        composite_warp_impl(dst, src, m, &opts, true, Some(weight));
+        composite_warp_impl(dst, src, m, &opts, true, Some(weight), None);
         return;
     }
     let mut tmp = Image::new(dst.width, dst.height);
@@ -79,7 +87,7 @@ pub fn accumulate_warp(dst: &mut Image, src: &Image, m: &Mat3, sampling: Samplin
 
 /// The straightforward per-pixel implementation (reference for the fast path's tests).
 pub fn composite_warp_reference(dst: &mut Image, src: &Image, m: &Mat3, opts: &WarpOpts) {
-    composite_warp_impl(dst, src, m, opts, false, None);
+    composite_warp_impl(dst, src, m, opts, false, None, None);
 }
 
 /// Bilinear blend of four taps (same operation order as [`Image::sample_bilinear`]).
@@ -94,9 +102,49 @@ fn bilerp(a: Px, b: Px, c: Px, d: Px, tx: f32, ty: f32) -> Px {
     o
 }
 
-/// Bilinear sample with an unchecked interior fast path.
+/// Side of the square source tiles of [`ClearTiles`].
+const TILE: usize = 16;
+
+/// Which `TILE`×`TILE` source tiles are fully transparent, the one-pixel apron to their right
+/// and below included: the four taps of a bilinear sample whose top-left tap lies in such a
+/// tile all have zero alpha, so the sample is transparent and draws nothing.
+struct ClearTiles {
+    cols: usize,
+    clear: Vec<bool>,
+}
+
+impl ClearTiles {
+    fn new(img: &Image) -> ClearTiles {
+        let (w, h) = (img.width as usize, img.height as usize);
+        let (cols, rows) = (w.div_ceil(TILE), h.div_ceil(TILE));
+        let clear = (0..rows)
+            .into_par_iter()
+            .flat_map_iter(|ty| {
+                let mut row_clear = vec![true; cols];
+                for y in ty * TILE..(ty * TILE + TILE + 1).min(h) {
+                    let row = &img.data[y * w..(y + 1) * w];
+                    for (tx, c) in row_clear.iter_mut().enumerate() {
+                        if *c {
+                            *c = row[tx * TILE..(tx * TILE + TILE + 1).min(w)].iter().all(|p| p[3] == 0.0);
+                        }
+                    }
+                }
+                row_clear
+            })
+            .collect();
+        ClearTiles { cols, clear }
+    }
+
+    #[inline(always)]
+    fn is_clear(&self, x: usize, y: usize) -> bool {
+        self.clear.get((y / TILE) * self.cols + x / TILE).copied().unwrap_or(false)
+    }
+}
+
+/// Bilinear sample with an unchecked interior fast path; `None` when every interior tap is in a
+/// [`ClearTiles`] tile (the sample would have zero alpha).
 #[inline(always)]
-fn sample_bilinear_fast(img: &Image, x: f64, y: f64) -> Px {
+fn sample_bilinear_fast(img: &Image, clear: &ClearTiles, x: f64, y: f64) -> Option<Px> {
     let fx = x - 0.5;
     let fy = y - 0.5;
     let x0 = fx.floor();
@@ -104,14 +152,17 @@ fn sample_bilinear_fast(img: &Image, x: f64, y: f64) -> Px {
     let (xi, yi) = (x0 as i64, y0 as i64);
     let (w, h) = (img.width as i64, img.height as i64);
     if xi >= 0 && yi >= 0 && xi + 1 < w && yi + 1 < h {
+        if clear.is_clear(xi as usize, yi as usize) {
+            return None;
+        }
         let tx = (fx - x0) as f32;
         let ty = (fy - y0) as f32;
         let i = yi as usize * w as usize + xi as usize;
         let row0 = &img.data[i..i + 2];
         let row1 = &img.data[i + w as usize..i + w as usize + 2];
-        return bilerp(row0[0], row0[1], row1[0], row1[1], tx, ty);
+        return Some(bilerp(row0[0], row0[1], row1[0], row1[1], tx, ty));
     }
-    img.sample_bilinear(x, y)
+    Some(img.sample_bilinear(x, y))
 }
 
 /// Index range `[lo, hi)` of `i` in `0..n` for which `p0 + i·s` may lie in `[min, max]`
@@ -127,7 +178,33 @@ fn span(p0: f64, s: f64, min: f64, max: f64, n: usize) -> (usize, usize) {
     if hi <= lo { (0, 0) } else { (lo as usize, hi as usize) }
 }
 
-/// Fast affine rows: `blend(dst, src, x, y)` is only called for visible source pixels.
+/// Columns `[a, b)` of one destination row that a warp changed (`a == b`: none).
+#[derive(Clone, Copy)]
+struct RowSpan {
+    a: usize,
+    b: usize,
+}
+
+impl RowSpan {
+    const NONE: RowSpan = RowSpan { a: usize::MAX, b: 0 };
+    #[inline(always)]
+    fn add(&mut self, x: usize) {
+        self.a = self.a.min(x);
+        self.b = x + 1;
+    }
+}
+
+/// Union of changed pixels as `[x0, y0, x1, y1]` (half-open; `x0 >= x1` when empty).
+fn union(r: [usize; 4], q: [usize; 4]) -> [usize; 4] {
+    match (r[0] < r[2], q[0] < q[2]) {
+        (false, _) => q,
+        (_, false) => r,
+        _ => [r[0].min(q[0]), r[1].min(q[1]), r[2].max(q[2]), r[3].max(q[3])],
+    }
+}
+
+/// Fast affine rows: `blend(dst, src, x, y)` is only called for visible source pixels. Returns
+/// the rectangle of the pixels it blended.
 #[allow(clippy::too_many_arguments)]
 fn affine_rows(
     dst: &mut Image,
@@ -135,12 +212,14 @@ fn affine_rows(
     inv: &Mat3,
     sampling: Sampling,
     op: f32,
+    quantize: Option<f32>,
     x0: usize,
     x1: usize,
     y0: usize,
     y1: usize,
     blend: impl Fn(Px, Px, u32, u32) -> Px + Sync,
-) {
+) -> [usize; 4] {
+    let levels = quantize.map(|l| (l, 1.0 / l));
     let w = dst.width as usize;
     let m = inv.0;
     let (iw, ih) = (img.width as f64, img.height as f64);
@@ -148,71 +227,104 @@ fn affine_rows(
     // Pure integer translation: every sample lands on a pixel centre, so all filters return the
     // source pixel itself (bicubic still clamps negatives, as in the reference).
     let int_tx = m[0][0] == 1.0 && m[1][1] == 1.0 && m[0][1] == 0.0 && m[1][0] == 0.0 && m[0][2].fract() == 0.0 && m[1][2].fract() == 0.0;
-    dst.data.par_chunks_mut(w).enumerate().skip(y0).take(y1 - y0).for_each(|(y, row)| {
-        let fy = y as f64 + 0.5;
-        let p0 = vec2(m[0][0] * (x0 as f64 + 0.5) + m[0][1] * fy + m[0][2], m[1][0] * (x0 as f64 + 0.5) + m[1][1] * fy + m[1][2]);
-        let n = x1 - x0;
-        let (ax, bx) = span(p0.x, step.x, -1.0, iw + 1.0, n);
-        let (ay, by) = span(p0.y, step.y, -1.0, ih + 1.0, n);
-        let (lo, hi) = (ax.max(ay), bx.min(by));
-        if lo >= hi {
-            return;
-        }
-        if int_tx {
-            let sy = y as i64 + m[1][2] as i64;
-            if sy < 0 || sy >= img.height as i64 {
-                return;
+    // Bilinear samples skip the transparent parts of the source (most of a thin shape's or a
+    // glow-padded layer's buffer) without reading them.
+    let clear = (sampling == Sampling::Bilinear && !int_tx).then(|| ClearTiles::new(img));
+    let empty = [usize::MAX, usize::MAX, 0, 0];
+    dst.data
+        .par_chunks_mut(w)
+        .enumerate()
+        .skip(y0)
+        .take(y1 - y0)
+        .map(|(y, row)| {
+            let mut changed = RowSpan::NONE;
+            let fy = y as f64 + 0.5;
+            let p0 = vec2(m[0][0] * (x0 as f64 + 0.5) + m[0][1] * fy + m[0][2], m[1][0] * (x0 as f64 + 0.5) + m[1][1] * fy + m[1][2]);
+            let n = x1 - x0;
+            let (ax, bx) = span(p0.x, step.x, -1.0, iw + 1.0, n);
+            let (ay, by) = span(p0.y, step.y, -1.0, ih + 1.0, n);
+            let (lo, hi) = (ax.max(ay), bx.min(by));
+            if lo >= hi {
+                return empty;
             }
-            let dx = m[0][2] as i64;
-            let srow = &img.data[sy as usize * img.width as usize..(sy as usize + 1) * img.width as usize];
-            let xa = (x0 as i64 + lo as i64).max(-dx).max(0) as usize;
-            let xb = ((x0 + hi) as i64).min(img.width as i64 - dx).max(0) as usize;
-            for x in xa..xb.max(xa) {
-                let mut s = srow[(x as i64 + dx) as usize];
-                if sampling == Sampling::Bicubic {
-                    s = [s[0].max(0.0), s[1].max(0.0), s[2].max(0.0), s[3].max(0.0)];
+            if int_tx {
+                let sy = y as i64 + m[1][2] as i64;
+                if sy < 0 || sy >= img.height as i64 {
+                    return empty;
                 }
-                if s[3] <= 0.0 {
-                    continue;
-                }
-                if op < 1.0 {
-                    for c in s.iter_mut() {
-                        *c *= op;
+                let dx = m[0][2] as i64;
+                let srow = &img.data[sy as usize * img.width as usize..(sy as usize + 1) * img.width as usize];
+                let xa = (x0 as i64 + lo as i64).max(-dx).max(0) as usize;
+                let xb = ((x0 + hi) as i64).min(img.width as i64 - dx).max(0) as usize;
+                for x in xa..xb.max(xa) {
+                    let mut s = srow[(x as i64 + dx) as usize];
+                    if sampling == Sampling::Bicubic {
+                        s = [s[0].max(0.0), s[1].max(0.0), s[2].max(0.0), s[3].max(0.0)];
                     }
+                    if s[3] <= 0.0 {
+                        continue;
+                    }
+                    if op < 1.0 {
+                        for c in s.iter_mut() {
+                            *c *= op;
+                        }
+                    }
+                    row[x] = blend(row[x], s, x as u32, y as u32);
+                    if let Some((l, inv)) = levels {
+                        crate::quantize_px(&mut row[x], l, inv);
+                    }
+                    changed.add(x);
                 }
-                row[x] = blend(row[x], s, x as u32, y as u32);
-            }
-            return;
-        }
-        for i in lo..hi {
-            let x = x0 + i;
-            let p = vec2(p0.x + step.x * i as f64, p0.y + step.y * i as f64);
-            if p.x < -1.0 || p.y < -1.0 || p.x > iw + 1.0 || p.y > ih + 1.0 {
-                continue;
-            }
-            let mut s = match sampling {
-                Sampling::Nearest => img.get(p.x.floor() as i64, p.y.floor() as i64),
-                Sampling::Bilinear => sample_bilinear_fast(img, p.x, p.y),
-                Sampling::Bicubic => img.sample_bicubic(p.x, p.y),
-            };
-            if s[3] <= 0.0 {
-                continue;
-            }
-            if op < 1.0 {
-                for c in s.iter_mut() {
-                    *c *= op;
+            } else {
+                for i in lo..hi {
+                    let x = x0 + i;
+                    let p = vec2(p0.x + step.x * i as f64, p0.y + step.y * i as f64);
+                    if p.x < -1.0 || p.y < -1.0 || p.x > iw + 1.0 || p.y > ih + 1.0 {
+                        continue;
+                    }
+                    let mut s = match (sampling, &clear) {
+                        (Sampling::Nearest, _) => img.get(p.x.floor() as i64, p.y.floor() as i64),
+                        (Sampling::Bilinear, Some(clear)) => match sample_bilinear_fast(img, clear, p.x, p.y) {
+                            Some(s) => s,
+                            None => continue,
+                        },
+                        (Sampling::Bilinear, None) => img.sample_bilinear(p.x, p.y),
+                        (Sampling::Bicubic, _) => img.sample_bicubic(p.x, p.y),
+                    };
+                    if s[3] <= 0.0 {
+                        continue;
+                    }
+                    if op < 1.0 {
+                        for c in s.iter_mut() {
+                            *c *= op;
+                        }
+                    }
+                    row[x] = blend(row[x], s, x as u32, y as u32);
+                    if let Some((l, inv)) = levels {
+                        crate::quantize_px(&mut row[x], l, inv);
+                    }
+                    changed.add(x);
                 }
             }
-            row[x] = blend(row[x], s, x as u32, y as u32);
-        }
-    });
+            if changed.a < changed.b { [changed.a, y, changed.b, y + 1] } else { empty }
+        })
+        .reduce(|| empty, union)
 }
 
-fn composite_warp_impl(dst: &mut Image, src: &Image, m: &Mat3, opts: &WarpOpts, allow_fast: bool, accumulate: Option<f32>) {
+fn composite_warp_impl(
+    dst: &mut Image,
+    src: &Image,
+    m: &Mat3,
+    opts: &WarpOpts,
+    allow_fast: bool,
+    accumulate: Option<f32>,
+    quantize: Option<f32>,
+) -> Option<[u32; 4]> {
+    const NOTHING: Option<[u32; 4]> = Some([0; 4]);
     if src.is_empty() || dst.is_empty() || opts.opacity <= 0.0 {
-        return;
+        return NOTHING;
     }
-    let Some(inv) = m.inverse() else { return };
+    let Some(inv) = m.inverse() else { return NOTHING };
     // Minification: pick a pre-filtered level so sampling doesn't alias.
     let scale = m.mean_scale();
     let mut level_img;
@@ -239,7 +351,7 @@ fn composite_warp_impl(dst: &mut Image, src: &Image, m: &Mat3, opts: &WarpOpts,
         area = area.intersect(&c);
     }
     if area.is_empty() {
-        return;
+        return NOTHING;
     }
     let (x0, x1) = (area.x0.max(0.0) as usize, area.x1.min(dst.width as f64) as usize);
     let (y0, y1) = (area.y0.max(0.0) as usize, area.y1.min(dst.height as f64) as usize);
@@ -247,25 +359,28 @@ fn composite_warp_impl(dst: &mut Image, src: &Image, m: &Mat3, opts: &WarpOpts,
     let affine = inv.is_affine();
     let mode = opts.mode;
     let op = opts.opacity;
+    let changed = |r: [usize; 4]| Some(r.map(|v| v.min(u32::MAX as usize) as u32));
     if let Some(k) = accumulate
         && affine
     {
-        affine_rows(dst, img, &inv, opts.sampling, 1.0, x0, x1, y0, y1, |d, s, _, _| [d[0] + s[0] * k, d[1] + s[1] * k, d[2] + s[2] * k, d[3] + s[3] * k]);
-        return;
+        let r = affine_rows(dst, img, &inv, opts.sampling, 1.0, None, x0, x1, y0, y1, |d, s, _, _| {
+            [d[0] + s[0] * k, d[1] + s[1] * k, d[2] + s[2] * k, d[3] + s[3] * k]
+        });
+        return changed(r);
     }
     if allow_fast && affine && !mode.is_stencil() {
-        match mode {
-            BlendMode::Normal => affine_rows(dst, img, &inv, opts.sampling, op, x0, x1, y0, y1, |d, s, _, _| {
+        let r = match mode {
+            BlendMode::Normal => affine_rows(dst, img, &inv, opts.sampling, op, quantize, x0, x1, y0, y1, |d, s, _, _| {
                 let k = 1.0 - s[3];
                 [s[0] + d[0] * k, s[1] + d[1] * k, s[2] + d[2] * k, s[3] + d[3] * k]
             }),
             BlendMode::Dissolve | BlendMode::DancingDissolve => {
                 let seed = opts.seed;
-                affine_rows(dst, img, &inv, opts.sampling, op, x0, x1, y0, y1, |d, s, x, y| blend_pixel(mode, d, s, hash_noise(x, y, seed)))
+                affine_rows(dst, img, &inv, opts.sampling, op, quantize, x0, x1, y0, y1, |d, s, x, y| blend_pixel(mode, d, s, hash_noise(x, y, seed)))
             }
-            _ => affine_rows(dst, img, &inv, opts.sampling, op, x0, x1, y0, y1, |d, s, _, _| blend_pixel(mode, d, s, 0.5)),
-        }
-        return;
+            _ => affine_rows(dst, img, &inv, opts.sampling, op, quantize, x0, x1, y0, y1, |d, s, _, _| blend_pixel(mode, d, s, 0.5)),
+        };
+        return changed(r);
     }
     dst.data.par_chunks_mut(w).enumerate().skip(y0).take(y1 - y0).for_each(|(y, row)| {
         let fy = y as f64 + 0.5;
@@ -312,6 +427,7 @@ fn composite_warp_impl(dst: &mut Image, src: &Image, m: &Mat3, opts: &WarpOpts,
             }
         }
     }
+    None
 }
 
 /// Resize with bilinear filtering (box pre-filter when shrinking).
@@ -394,6 +510,116 @@ mod tests {
         }
     }
 
+    /// A mostly transparent source: a thin ring with a few stray pixels (one with negative zero alpha).
+    fn sparse_image(w: u32, h: u32) -> Image {
+        let mut img = Image::new(w, h);
+        let (cx, cy, r) = (w as f32 / 2.0, h as f32 / 2.0, w.min(h) as f32 * 0.4);
+        for y in 0..h {
+            for x in 0..w {
+                let d = ((x as f32 + 0.5 - cx).hypot(y as f32 + 0.5 - cy) - r).abs();
+                if d < 1.5 {
+                    let a = 1.0 - d / 1.5;
+                    img.set(x, y, [0.2 * a, 0.6 * a, a, a]);
+                }
+            }
+        }
+        img.set(3, 2, [0.5, 0.5, 0.5, 0.5]);
+        img.set(w - 1, h - 1, [1.0; 4]);
+        img.set(w / 2, 1, [0.0, 0.0, 0.0, -0.0]);
+        img
+    }
+
+    #[test]
+    fn clear_tiles_hold_only_transparent_taps() {
+        let img = sparse_image(150, 97);
+        let clear = ClearTiles::new(&img);
+        let mut skipped = 0;
+        for y in 0..img.height as usize - 1 {
+            for x in 0..img.width as usize - 1 {
+                if clear.is_clear(x, y) {
+                    skipped += 1;
+                    for (dx, dy) in [(0, 0), (1, 0), (0, 1), (1, 1)] {
+                        assert_eq!(img.get(x as i64 + dx, y as i64 + dy)[3], 0.0, "tap ({}, {}) of a clear tile", x as i64 + dx, y as i64 + dy);
+                    }
+                }
+            }
+        }
+        assert!(skipped > 100 * 70, "the ring's empty middle is skipped: {skipped}");
+    }
+
+    /// The changed rectangle holds every pixel a warp changed, and the skipped transparent
+    /// samples leave the result as the reference path draws it.
+    #[test]
+    fn changed_rect_covers_the_changed_pixels() {
+        let src = sparse_image(90, 70);
+        let mats = [
+            Mat3::translate(vec2(7.0, 3.0)),
+            Mat3::translate(vec2(10.25, 4.5)),
+            Mat3::translate(vec2(60.0, 50.0)) * Mat3::rotate_deg(33.0) * Mat3::scale(vec2(0.9, 1.2)) * Mat3::translate(vec2(-45.0, -35.0)),
+            Mat3::translate(vec2(-30.0, -20.0)),
+            Mat3::scale(vec2(0.3, 0.3)),
+        ];
+        for (mi, m) in mats.iter().enumerate() {
+            for mode in [BlendMode::Normal, BlendMode::Add, BlendMode::Dissolve] {
+                let opts = WarpOpts { mode, seed: 3, ..Default::default() };
+                let before = test_image(120, 100, 4);
+                let mut a = before.clone();
+                let r = composite_warp_changed(&mut a, &src, m, &opts, None);
+                let Some([x0, y0, x1, y1]) = r else { panic!("mat {mi}: the fast path tracks what it changed") };
+                for (i, (p, q)) in a.data.iter().zip(&before.data).enumerate() {
+                    let (x, y) = (i as u32 % a.width, i as u32 / a.width);
+                    if p != q {
+                        assert!(x >= x0 && x < x1 && y >= y0 && y < y1, "mat {mi} {mode:?}: ({x}, {y}) changed outside {r:?}");
+                    }
+                }
+                let mut b = before.clone();
+                composite_warp_reference(&mut b, &src, m, &opts);
+                for (i, (p, q)) in a.data.iter().zip(&b.data).enumerate() {
+                    for c in 0..4 {
+                        assert!((p[c] - q[c]).abs() < 1e-4, "mat {mi} {mode:?} px {i}: {p:?} vs {q:?}");
+                    }
+                }
+            }
+        }
+        let mut dst = Image::new(10, 10);
+        assert_eq!(composite_warp_changed(&mut dst, &Image::new(4, 4), &Mat3::IDENTITY, &WarpOpts::default(), None), Some([u32::MAX, u32::MAX, 0, 0]));
+        let stencil = WarpOpts { mode: BlendMode::StencilAlpha, ..Default::default() };
+        assert_eq!(composite_warp_changed(&mut dst, &src, &Mat3::IDENTITY, &stencil, Some(255.0)), None, "stencils change the whole destination");
+    }
+
+    /// Quantising while blending equals blending, then quantising the changed rectangle of an
+    /// already quantised destination (what an 8/16 bpc comp did after every layer), bit for bit.
+    #[test]
+    fn quantize_while_blending_matches_quantizing_after() {
+        let src = sparse_image(90, 70);
+        let mats = [Mat3::translate(vec2(7.0, 3.0)), Mat3::translate(vec2(60.0, 50.0)) * Mat3::rotate_deg(33.0) * Mat3::translate(vec2(-45.0, -35.0))];
+        for m in &mats {
+            for mode in [BlendMode::Normal, BlendMode::Screen, BlendMode::Dissolve] {
+                for (levels, opacity) in [(255.0f32, 1.0), (32_768.0, 0.7)] {
+                    let mut before = test_image(120, 100, 8);
+                    for p in &mut before.data {
+                        crate::quantize_px(p, levels, 1.0 / levels);
+                    }
+                    let opts = WarpOpts { mode, opacity, seed: 5, ..Default::default() };
+                    let mut after = before.clone();
+                    let Some([x0, y0, x1, y1]) = composite_warp_changed(&mut after, &src, m, &opts, None) else { panic!("fast path") };
+                    for y in y0..y1 {
+                        for x in x0..x1 {
+                            let i = after.idx(x, y);
+                            crate::quantize_px(&mut after.data[i], levels, 1.0 / levels);
+                        }
+                    }
+                    let mut fused = before.clone();
+                    let r = composite_warp_changed(&mut fused, &src, m, &opts, Some(levels));
+                    assert_eq!(r, Some([x0, y0, x1, y1]));
+                    for (i, (p, q)) in fused.data.iter().zip(&after.data).enumerate() {
+                        assert_eq!(p.map(f32::to_bits), q.map(f32::to_bits), "{mode:?} {levels} px {i}");
+                    }
+                }
+            }
+        }
+    }
+
     /// Accumulating sub-samples equals warping each into its own image and summing.
     #[test]
     fn accumulate_matches_per_sample_images() {

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::*;