Drop the bind group ordering comment

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>

core/src/primitive/layer.rs

@@ -1,7 +1,7 @@
 use std::ops::{Index, IndexMut};
 
 use crate::{
-    render::{MaskIdx, Primitive, PrimitiveHandle, PrimitiveInst, Primitives},
+    render::{LayerDraws, MaskIdx, Primitive, PrimitiveHandle, PrimitiveInst},
     util::to_mut,
 };
 
@@ -39,7 +39,7 @@ struct Child {
     tail: usize,
 }
 
-pub type PrimitiveLayers = Layers<Primitives>;
+pub type DrawLayers = Layers<LayerDraws>;
 
 impl<T: Default> Layers<T> {
     pub fn new() -> Layers<T> {
@@ -119,7 +119,7 @@ impl<T: Default> Layers<T> {
     }
 }
 
-impl PrimitiveLayers {
+impl DrawLayers {
     pub fn write<P: Primitive>(
         &mut self,
         layer: LayerId,

core/src/primitive/text.rs

@@ -1,6 +1,5 @@
 use crate::{
-    Align, GlyphAtlas, GlyphEntry, GlyphKey, PlacedGlyph, RegionAlign, Textures, UiColor,
+    Align, GlyphAtlas, GlyphEntry, GlyphKey, PlacedGlyph, RegionAlign, UiColor, util::Vec2,
-    util::Vec2,
 };
 use parley::{
     Alignment, AlignmentOptions, FontContext, FontFamily, FontFamilyName, GenericFamily, Layout,
@@ -159,7 +158,7 @@ impl TextBuffer {
 }
 
 impl TextData {
-    pub fn place(&mut self, buffer: &TextBuffer, textures: &mut Textures) -> Vec<PlacedGlyph> {
+    pub fn place(&mut self, buffer: &TextBuffer) -> Vec<PlacedGlyph> {
         let mut placed = Vec::new();
         for line in buffer.layout.lines() {
             for item in line.items() {
@@ -185,17 +184,14 @@ impl TextData {
                         subpixel,
                         coords: coords_hash,
                     };
-                    let Some(entry) = self.glyph_entry(
+                    let Some(entry) = self.glyph_entry(GlyphRaster {
-                        GlyphRaster {
                         key,
                         font: font_ref,
                         font_size,
                         coords,
                         subpixel,
                         glyph_id: glyph.id,
-                        },
+                    }) else {
-                        textures,
-                    ) else {
                         continue;
                     };
                     placed.push(PlacedGlyph {
@@ -211,11 +207,7 @@ impl TextData {
         placed
     }
 
-    fn glyph_entry(
+    fn glyph_entry(&mut self, glyph: GlyphRaster<'_>) -> Option<GlyphEntry> {
-        &mut self,
-        glyph: GlyphRaster<'_>,
-        textures: &mut Textures,
-    ) -> Option<GlyphEntry> {
         if let Some(entry) = self.atlas.get(&glyph.key) {
             return entry;
         }
@@ -237,7 +229,7 @@ impl TextData {
         .render(&mut scaler, glyph.glyph_id as u16);
 
         if let Some(image) = image {
-            self.atlas.insert(glyph.key, &image, textures)
+            self.atlas.insert(glyph.key, &image)
         } else {
             self.atlas.insert_empty(glyph.key);
             None
@@ -278,10 +270,9 @@ impl TextData {
         buffer: &mut TextBuffer,
         attrs: &TextAttrs,
         width: Option<f32>,
-        textures: &mut Textures,
     ) -> RenderedText {
         buffer.shape(self, attrs, width);
-        let glyphs = self.place(buffer, textures);
+        let glyphs = self.place(buffer);
         RenderedText {
             glyphs,
             size: buffer.size(),

core/src/primitive/texture.rs

@@ -1,7 +1,4 @@
-use crate::{
+use crate::util::{RefCounter, Vec2};
-    render::TexturePrimitive,
-    util::{RefCounter, Vec2},
-};
 use image::{DynamicImage, GenericImageView};
 use std::{
     ops::Index,
@@ -10,7 +7,7 @@ use std::{
 
 #[derive(Debug, Clone)]
 pub struct TextureHandle {
-    inner: TexturePrimitive,
+    slot: u32,
     size: Vec2,
     counter: RefCounter,
     send: Sender<u32>,
@@ -31,6 +28,8 @@ pub enum TextureUpdate<'a> {
     Set(u32, &'a DynamicImage),
     Patch(u32, PatchRect, &'a DynamicImage),
     Free(u32),
+    /// Added and freed before the renderer drained either update. It still has
+    /// to push a slot to stay lined up with `images`; `Free` then empties it.
     PushFree,
     SetFree,
 }
@@ -64,14 +63,8 @@ impl Textures {
     pub fn add(&mut self, image: impl Into<DynamicImage>) -> TextureHandle {
         let image = image.into();
         let size = image.dimensions().into();
-        let view_idx = self.push(image);
-        // 0 == default in renderer; TODO: actually create samplers here
-        let sampler_idx = 0;
         TextureHandle {
-            inner: TexturePrimitive {
+            slot: self.push(image),
-                view_idx,
-                sampler_idx,
-            },
             size,
             counter: RefCounter::new(),
             send: self.send.clone(),
@@ -92,15 +85,20 @@ impl Textures {
     }
 
     pub fn image_mut(&mut self, handle: &TextureHandle) -> &mut DynamicImage {
-        self.images[handle.inner.view_idx as usize]
+        self.images[handle.slot as usize]
             .as_mut()
             .expect("texture was freed while still held")
     }
 
     /// Queue an upload of just `rect`, after writing it with `image_mut`.
     pub fn patch(&mut self, handle: &TextureHandle, rect: PatchRect) {
-        self.updates
+        self.updates.push(Update::Patch(handle.slot, rect));
-            .push(Update::Patch(handle.inner.view_idx, rect));
+    }
+
+    /// How many textures are live, which is what a ui can ask; the renderer's
+    /// copies follow from the updates it drains.
+    pub fn count(&self) -> usize {
+        self.images.iter().flatten().count()
     }
 
     pub fn free(&mut self) {
@@ -131,8 +129,9 @@ impl Textures {
 }
 
 impl TextureHandle {
-    pub fn primitive(&self) -> TexturePrimitive {
+    /// Index into `Textures`, and into the renderer's parallel slots.
-        self.inner
+    pub fn slot(&self) -> u32 {
+        self.slot
     }
     pub fn size(&self) -> Vec2 {
         self.size
@@ -142,7 +141,7 @@ impl TextureHandle {
 impl Drop for TextureHandle {
     fn drop(&mut self) {
         if self.counter.drop() {
-            let _ = self.send.send(self.inner.view_idx);
+            let _ = self.send.send(self.slot);
         }
     }
 }
@@ -151,7 +150,7 @@ impl Index<&TextureHandle> for Textures {
     type Output = DynamicImage;
 
     fn index(&self, index: &TextureHandle) -> &Self::Output {
-        self.images[index.inner.view_idx as usize].as_ref().unwrap()
+        self.images[index.slot as usize].as_ref().unwrap()
     }
 }
 

core/src/render/atlas.rs

@@ -1,11 +1,12 @@
 use crate::{
-    PatchRect, TextureHandle, Textures,
+    PatchRect,
     util::{HashMap, Vec2},
 };
 use image::RgbaImage;
 use swash::scale::image::{Content, Image};
 
-const PAGE: u32 = 1024;
+/// Side of one page, and so of every layer of `render::page`'s array texture.
+pub(crate) const PAGE: u32 = 1024;
 
 /// Transparent margin kept around every glyph, so that sampling one cannot
 /// pick up its neighbour along a shared edge.
@@ -35,8 +36,8 @@ pub struct GlyphEntry {
     pub width: u32,
     pub height: u32,
     pub is_colored: bool,
-    pub view_idx: u32,
+    /// Which atlas array layer this glyph is on.
-    pub sampler_idx: u32,
+    pub layer: u32,
 }
 
 impl GlyphEntry {
@@ -48,18 +49,26 @@ impl GlyphEntry {
 }
 
 struct Page {
-    handle: TextureHandle,
+    image: RgbaImage,
     x: u32,
     y: u32,
     shelf_height: u32,
 }
 
+/// A rectangle of one page the renderer has not uploaded yet.
+#[derive(Clone, Copy)]
+pub struct PageUpload {
+    pub layer: u32,
+    pub rect: PatchRect,
+}
+
 #[derive(Default)]
 pub struct GlyphAtlas {
     pages: Vec<Page>,
     /// `None` for a glyph that rasterised to nothing -- a space, say. Cached
     /// too, so it is not re-rasterised on every layout.
     entries: HashMap<GlyphKey, Option<GlyphEntry>>,
+    uploads: Vec<PageUpload>,
 }
 
 impl GlyphAtlas {
@@ -67,12 +76,7 @@ impl GlyphAtlas {
         self.entries.get(key).copied()
     }
 
-    pub fn insert(
+    pub fn insert(&mut self, key: GlyphKey, image: &Image) -> Option<GlyphEntry> {
-        &mut self,
-        key: GlyphKey,
-        image: &Image,
-        textures: &mut Textures,
-    ) -> Option<GlyphEntry> {
         let w = image.placement.width;
         let h = image.placement.height;
         if w == 0 || h == 0 {
@@ -94,23 +98,11 @@ impl GlyphAtlas {
             return None;
         }
 
-        let (page_idx, x, y) = self.allocate(w, h, textures);
+        let upload = self.allocate(w, h);
-        let page = &self.pages[page_idx];
+        let PatchRect { x, y, .. } = upload.rect;
+        write_glyph(&mut self.pages[upload.layer as usize].image, image, x, y);
+        self.uploads.push(upload);
 
-        let img = textures.image_mut(&page.handle);
-        let rgba = img.as_mut_rgba8().expect("atlas page is rgba8");
-        write_glyph(rgba, image, x, y);
-
-        let handle = page.handle.clone();
-        let rect = PatchRect {
-            x,
-            y,
-            width: w,
-            height: h,
-        };
-        textures.patch(&handle, rect);
-
-        let page = &self.pages[page_idx];
         let scale = 1.0 / PAGE as f32;
         let entry = GlyphEntry {
             uv_min: Vec2::new(x as f32 * scale, y as f32 * scale),
@@ -120,39 +112,60 @@ impl GlyphAtlas {
             width: w,
             height: h,
             is_colored: matches!(image.content, Content::Color),
-            view_idx: page.handle.primitive().view_idx,
+            layer: upload.layer,
-            sampler_idx: page.handle.primitive().sampler_idx,
         };
         self.entries.insert(key, Some(entry));
         Some(entry)
     }
 
-    fn allocate(&mut self, w: u32, h: u32, textures: &mut Textures) -> (usize, u32, u32) {
+    /// Reserves room for a `w` by `h` glyph, adding a page if none has it.
+    fn allocate(&mut self, w: u32, h: u32) -> PageUpload {
+        let rect = |x, y| PatchRect {
+            x,
+            y,
+            width: w,
+            height: h,
+        };
         if let Some((i, (x, y))) = self
             .pages
             .iter_mut()
             .enumerate()
             .find_map(|(i, page)| page.allocate(w, h).map(|position| (i, position)))
         {
-            return (i, x, y);
+            return PageUpload {
+                layer: i as u32,
+                rect: rect(x, y),
+            };
         }
 
-        let handle = textures.add(RgbaImage::new(PAGE, PAGE));
         self.pages.push(Page {
-            handle,
+            image: RgbaImage::new(PAGE, PAGE),
             x: PAD + w + PAD,
             y: PAD,
             shelf_height: h + PAD,
         });
-        (self.pages.len() - 1, PAD, PAD)
+        PageUpload {
+            layer: self.pages.len() as u32 - 1,
+            rect: rect(PAD, PAD),
+        }
+    }
+
+    /// Drains what has been written since the last call, for the renderer to
+    /// upload. A new page needs nothing more: wgpu leaves the rest of a fresh
+    /// layer transparent, which is what an atlas wants.
+    pub fn uploads(&mut self) -> impl Iterator<Item = (PageUpload, &RgbaImage)> {
+        let pages = &self.pages;
+        self.uploads
+            .drain(..)
+            .map(|upload| (upload, &pages[upload.layer as usize].image))
     }
 
     pub fn insert_empty(&mut self, key: GlyphKey) {
         self.entries.insert(key, None);
     }
 
-    pub fn page_count(&self) -> usize {
+    pub fn page_count(&self) -> u32 {
-        self.pages.len()
+        self.pages.len() as u32
     }
 
     pub fn glyph_count(&self) -> usize {

core/src/render/data.rs

@@ -12,20 +12,16 @@ pub struct WindowUniform {
 #[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
 pub struct PrimitiveInstance {
     pub region: UiRegion,
-    pub binding: u32,
-    pub idx: u32,
     pub mask_idx: MaskIdx,
 }
 
 impl PrimitiveInstance {
-    const ATTRIBS: [VertexAttribute; 7] = vertex_attr_array![
+    const ATTRIBS: [VertexAttribute; 5] = vertex_attr_array![
         0 => Float32x2,
         1 => Float32x2,
         2 => Float32x2,
         3 => Float32x2,
         4 => Uint32,
-        5 => Uint32,
-        6 => Uint32,
     ];
 
     pub fn desc() -> VertexBufferLayout<'static> {

core/src/render/mod.rs

@@ -1,8 +1,6 @@
-use std::num::NonZero;
-
 use crate::{
     UiData, UiRenderState,
-    render::{data::PrimitiveInstance, texture::GpuTextures, util::ArrBuf},
+    render::{data::PrimitiveInstance, util::ArrBuf},
     util::{HashMap, Vec2},
 };
 use data::WindowUniform;
@@ -13,6 +11,7 @@ use wgpu::{
 
 mod atlas;
 mod data;
+mod page;
 mod primitive;
 mod texture;
 mod util;
@@ -21,42 +20,63 @@ pub use atlas::*;
 pub use data::{Mask, MaskIdx};
 pub use primitive::*;
 
-const SHAPE_SHADER: &str = include_str!("./shader.wgsl");
+const PRELUDE: &str = include_str!("./shader/prelude.wgsl");
 
 pub struct UiRenderNode {
-    uniform_group: BindGroup,
+    shared_layout: BindGroupLayout,
-    primitive_layout: BindGroupLayout,
+    shared_group: BindGroup,
-    rsc_layout: BindGroupLayout,
+    format: TextureFormat,
-    rsc_group: BindGroup,
 
-    pipeline: RenderPipeline,
+    /// One per registered primitive, in id order.
+    primitives: Vec<PrimitivePipeline>,
 
     layers: HashMap<usize, RenderLayer>,
     active: Vec<usize>,
     window_buffer: Buffer,
-    textures: GpuTextures,
     masks: ArrBuf<Mask>,
 }
 
 struct RenderLayer {
+    /// One per registered primitive, `None` where this layer draws none.
+    primitives: Vec<Option<ListBuffers>>,
+}
+
+/// What draws one registered primitive.
+struct PrimitivePipeline {
+    data_layout: BindGroupLayout,
+    pipeline: RenderPipeline,
+    render: Box<dyn PrimitiveRender>,
+}
+
+/// One list's vertex buffer and the data its shader reads.
+struct ListBuffers {
     instance: ArrBuf<PrimitiveInstance>,
-    primitives: PrimitiveBuffers,
+    data: ArrBuf<u8>,
-    primitive_group: BindGroup,
+    group: Option<BindGroup>,
+    /// What the primitive asked to keep per instance, if anything.
+    bindings: Vec<u32>,
 }
 
 impl UiRenderNode {
     pub fn draw<'a>(&'a self, pass: &mut RenderPass<'a>) {
-        pass.set_pipeline(&self.pipeline);
+        pass.set_bind_group(0, &self.shared_group, &[]);
-        pass.set_bind_group(0, &self.uniform_group, &[]);
-        pass.set_bind_group(2, &self.rsc_group, &[]);
         for i in &self.active {
             let layer = &self.layers[i];
-            if layer.instance.len() == 0 {
+            for (id, list) in layer.primitives.iter().enumerate() {
-                continue;
+                let Some(list) = list else { continue };
+                let Some(group) = &list.group else { continue };
+                let primitive = &self.primitives[id];
+                pass.set_pipeline(&primitive.pipeline);
+                pass.set_bind_group(1, group, &[]);
+                pass.set_vertex_buffer(0, list.instance.buffer.slice(..));
+                primitive.render.draw(
+                    pass,
+                    ListDraw {
+                        instances: list.instance.len() as u32,
+                        bindings: &list.bindings,
+                    },
+                );
             }
-            pass.set_bind_group(1, &layer.primitive_group, &[]);
-            pass.set_vertex_buffer(0, layer.instance.buffer.slice(..));
-            pass.draw(0..4, 0..layer.instance.len() as u32);
         }
     }
 
@@ -67,55 +87,56 @@ impl UiRenderNode {
         ui: &mut UiData,
         ui_render: &mut UiRenderState,
     ) {
+        // Before the layers: each list is given its pipeline's data layout.
+        self.build_pipelines(device, queue, &ui.primitives);
         self.active.clear();
-        for (i, primitives) in ui_render.layers.iter_mut() {
+        for (i, draws) in ui_render.layers.iter_mut() {
             self.active.push(i);
-            for change in primitives.apply_free() {
+            for change in draws.apply_free() {
                 if let Some(inst) = ui_render.active.get_mut(&change.id) {
                     for h in &mut inst.primitives {
-                        if h.layer == i && h.inst_idx == change.old {
+                        if h.layer == i && h.kind == change.kind && h.inst_idx == change.old {
                             h.inst_idx = change.new;
                             break;
                         }
                     }
                 }
             }
-            let rlayer = self.layers.entry(i).or_insert_with(|| {
+            let rlayer = self.layers.entry(i).or_insert_with(RenderLayer::new);
-                let primitives = PrimitiveBuffers::new(device);
+            if draws.updated {
-                let primitive_group =
+                let lists = draws.primitives();
-                    Self::primitive_group(device, &self.primitive_layout, primitives.buffers());
+                // The zip would otherwise skip a list with no pipeline.
-                RenderLayer {
+                assert!(lists.len() <= self.primitives.len());
-                    instance: ArrBuf::new(
+                rlayer.primitives.resize_with(lists.len(), || None);
-                        device,
+                for ((buffers, list), primitive) in rlayer
-                        BufferUsages::VERTEX | BufferUsages::COPY_DST,
+                    .primitives
-                        "instance",
+                    .iter_mut()
-                    ),
+                    .zip(lists)
-                    primitives,
+                    .zip(&self.primitives)
-                    primitive_group,
+                {
+                    let Some(list) = list else {
+                        continue;
+                    };
+                    buffers
+                        .get_or_insert_with(|| ListBuffers::new(device))
+                        .update(device, queue, primitive, list);
                 }
-            });
+                draws.updated = false;
-            if primitives.updated {
-                rlayer
-                    .instance
-                    .update(device, queue, primitives.instances());
-                rlayer.primitives.update(device, queue, primitives.data());
-                rlayer.primitive_group = Self::primitive_group(
-                    device,
-                    &self.primitive_layout,
-                    rlayer.primitives.buffers(),
-                );
-                primitives.updated = false;
             }
         }
-        let mut changed = false;
+        for primitive in &mut self.primitives {
-        changed |= self.textures.update(&mut ui.textures);
+            primitive.render.update(ui);
+        }
         if ui.masks.changed {
             ui.masks.changed = false;
-            self.masks.update(device, queue, &ui.masks[..]);
+            if self.masks.update(device, queue, &ui.masks[..]) {
-            changed = true;
+                self.shared_group = Self::shared_group(
+                    device,
+                    &self.shared_layout,
+                    &self.window_buffer,
+                    &self.masks,
+                );
             }
-        if changed {
-            self.rsc_group = Self::rsc_group(device, &self.rsc_layout, &self.textures, &self.masks);
         }
     }
 
@@ -128,17 +149,7 @@ impl UiRenderNode {
         queue.write_buffer(&self.window_buffer, 0, bytemuck::cast_slice(slice));
     }
 
-    pub fn new(
+    pub fn new(device: &Device, config: &SurfaceConfiguration) -> Self {
-        device: &Device,
-        queue: &Queue,
-        config: &SurfaceConfiguration,
-        limits: UiLimits,
-    ) -> Self {
-        let shader = device.create_shader_module(ShaderModuleDescriptor {
-            label: Some("UI Shape Shader"),
-            source: ShaderSource::Wgsl(SHAPE_SHADER.into()),
-        });
-
         let window_uniform = WindowUniform {
             width: config.width as f32,
             height: config.height as f32,
@@ -149,67 +160,73 @@ impl UiRenderNode {
             usage: BufferUsages::UNIFORM | BufferUsages::COPY_DST,
         });
 
-        let uniform_layout = device.create_bind_group_layout(&BindGroupLayoutDescriptor {
+        let shared_layout = Self::shared_layout(device);
-            entries: &[BindGroupLayoutEntry {
-                binding: 0,
-                visibility: ShaderStages::VERTEX | ShaderStages::FRAGMENT,
-                ty: BindingType::Buffer {
-                    ty: BufferBindingType::Uniform,
-                    has_dynamic_offset: false,
-                    min_binding_size: None,
-                },
-                count: None,
-            }],
-            label: Some("window"),
-        });
-
-        let uniform_group = Self::bind_group_0(device, &uniform_layout, &window_buffer);
-
-        let primitive_layout = device.create_bind_group_layout(&BindGroupLayoutDescriptor {
-            entries: &core::array::from_fn::<_, { PrimitiveBuffers::LEN }, _>(|i| {
-                BindGroupLayoutEntry {
-                    binding: i as u32,
-                    visibility: ShaderStages::FRAGMENT,
-                    ty: BindingType::Buffer {
-                        ty: BufferBindingType::Storage { read_only: true },
-                        has_dynamic_offset: false,
-                        min_binding_size: None,
-                    },
-                    count: None,
-                }
-            }),
-            label: Some("primitive"),
-        });
-
-        let tex_manager = GpuTextures::new(device, queue);
         let masks = ArrBuf::new(
             device,
             BufferUsages::STORAGE | BufferUsages::COPY_DST,
             "ui masks",
         );
+        let shared_group = Self::shared_group(device, &shared_layout, &window_buffer, &masks);
 
-        let rsc_layout = Self::rsc_layout(device, &limits);
+        Self {
-        let rsc_group = Self::rsc_group(device, &rsc_layout, &tex_manager, &masks);
+            shared_layout,
+            shared_group,
+            format: config.format,
+            primitives: Vec::new(),
+            window_buffer,
+            layers: HashMap::default(),
+            active: Vec::new(),
+            masks,
+        }
+    }
 
-        let pipeline_layout = device.create_pipeline_layout(&PipelineLayoutDescriptor {
+    /// Compiles a pipeline for every primitive registered since the last call.
-            label: Some("UI Shape Pipeline Layout"),
+    /// Sources only ever arrive at the end, so an id keeps its pipeline.
-            bind_group_layouts: &[&uniform_layout, &primitive_layout, &rsc_layout],
+    fn build_pipelines(&mut self, device: &Device, queue: &Queue, registry: &PrimitiveRegistry) {
+        for source in &registry.sources()[self.primitives.len()..] {
+            let render = (source.render)(device, queue);
+            let data_layout = Self::data_layout(device, source.stride);
+            let mut groups = vec![&self.shared_layout, &data_layout];
+            groups.extend(render.layout());
+            let layout = device.create_pipeline_layout(&PipelineLayoutDescriptor {
+                label: Some(source.label),
+                bind_group_layouts: &groups,
                 immediate_size: 0,
             });
-        let pipeline = device.create_render_pipeline(&RenderPipelineDescriptor {
+            let pipeline = Self::pipeline(device, &layout, self.format, source.wgsl, source.label);
-            label: Some("UI Shape Pipeline"),
+            self.primitives.push(PrimitivePipeline {
-            layout: Some(&pipeline_layout),
+                data_layout,
+                pipeline,
+                render,
+            });
+        }
+    }
+
+    fn pipeline(
+        device: &Device,
+        layout: &PipelineLayout,
+        format: TextureFormat,
+        wgsl: &str,
+        label: &str,
+    ) -> RenderPipeline {
+        let module = device.create_shader_module(ShaderModuleDescriptor {
+            label: Some(label),
+            source: ShaderSource::Wgsl(format!("{PRELUDE}\n{wgsl}").into()),
+        });
+        device.create_render_pipeline(&RenderPipelineDescriptor {
+            label: Some(label),
+            layout: Some(layout),
             vertex: VertexState {
-                module: &shader,
+                module: &module,
                 entry_point: Some("vs_main"),
                 buffers: &[PrimitiveInstance::desc()],
                 compilation_options: Default::default(),
             },
             fragment: Some(FragmentState {
-                module: &shader,
+                module: &module,
                 entry_point: Some("fs_main"),
                 targets: &[Some(ColorTargetState {
-                    format: config.format,
+                    format,
                     blend: Some(BlendState::ALPHA_BLENDING),
                     write_mask: ColorWrites::ALL,
                 })],
@@ -232,90 +249,42 @@ impl UiRenderNode {
             },
             multiview_mask: None,
             cache: None,
-        });
-
-        Self {
-            uniform_group,
-            primitive_layout,
-            rsc_layout,
-            rsc_group,
-            pipeline,
-            window_buffer,
-            layers: HashMap::default(),
-            active: Vec::new(),
-            textures: tex_manager,
-            masks,
-        }
-    }
-
-    fn bind_group_0(
-        device: &Device,
-        layout: &BindGroupLayout,
-        window_buffer: &Buffer,
-    ) -> BindGroup {
-        device.create_bind_group(&BindGroupDescriptor {
-            layout,
-            entries: &[BindGroupEntry {
-                binding: 0,
-                resource: window_buffer.as_entire_binding(),
-            }],
-            label: Some("ui window"),
         })
     }
 
-    fn primitive_group(
+    /// What every draw in the ui is given: the window and the masks.
-        device: &Device,
+    fn shared_layout(device: &Device) -> BindGroupLayout {
-        layout: &BindGroupLayout,
-        buffers: [(u32, &Buffer); PrimitiveBuffers::LEN],
-    ) -> BindGroup {
-        device.create_bind_group(&BindGroupDescriptor {
-            layout,
-            entries: &buffers.map(|(binding, buf)| BindGroupEntry {
-                binding,
-                resource: buf.as_entire_binding(),
-            }),
-            label: Some("ui primitives"),
-        })
-    }
-
-    fn rsc_layout(device: &Device, limits: &UiLimits) -> BindGroupLayout {
         device.create_bind_group_layout(&BindGroupLayoutDescriptor {
             entries: &[
                 BindGroupLayoutEntry {
                     binding: 0,
-                    visibility: ShaderStages::FRAGMENT,
+                    visibility: ShaderStages::VERTEX | ShaderStages::FRAGMENT,
-                    ty: BindingType::Texture {
+                    ty: BindingType::Buffer {
-                        sample_type: TextureSampleType::Float { filterable: false },
+                        ty: BufferBindingType::Uniform,
-                        view_dimension: TextureViewDimension::D2,
+                        has_dynamic_offset: false,
-                        multisampled: false,
+                        min_binding_size: BufferSize::new(size_of::<WindowUniform>() as u64),
                     },
-                    count: Some(NonZero::new(limits.max_textures).unwrap()),
+                    count: None,
                 },
                 BindGroupLayoutEntry {
                     binding: 1,
                     visibility: ShaderStages::FRAGMENT,
-                    ty: BindingType::Sampler(SamplerBindingType::NonFiltering),
-                    count: Some(NonZero::new(limits.max_samplers).unwrap()),
-                },
-                BindGroupLayoutEntry {
-                    binding: 2,
-                    visibility: ShaderStages::FRAGMENT,
                     ty: BindingType::Buffer {
                         ty: BufferBindingType::Storage { read_only: true },
                         has_dynamic_offset: false,
-                        min_binding_size: None,
+                        min_binding_size: BufferSize::new(size_of::<Mask>() as u64),
                     },
                     count: None,
                 },
             ],
-            label: Some("ui rsc"),
+            label: Some("ui shared"),
         })
     }
 
-    fn rsc_group(
+    fn shared_group(
         device: &Device,
         layout: &BindGroupLayout,
-        tex_manager: &GpuTextures,
+        window: &Buffer,
         masks: &ArrBuf<Mask>,
     ) -> BindGroup {
         device.create_bind_group(&BindGroupDescriptor {
@@ -323,45 +292,85 @@ impl UiRenderNode {
             entries: &[
                 BindGroupEntry {
                     binding: 0,
-                    resource: BindingResource::TextureViewArray(&tex_manager.views()),
+                    resource: window.as_entire_binding(),
                 },
                 BindGroupEntry {
                     binding: 1,
-                    resource: BindingResource::SamplerArray(&tex_manager.samplers()),
-                },
-                BindGroupEntry {
-                    binding: 2,
                     resource: masks.buffer.as_entire_binding(),
                 },
             ],
-            label: Some("ui rsc"),
+            label: Some("ui shared"),
         })
     }
 
-    pub fn view_count(&self) -> usize {
+    /// Layout for a list of one primitive's data. Every size in the ui is
-        self.textures.view_count()
+    /// stated, so "is the buffer big enough for one entry?" is answered when
+    /// the bind group is made; a `None` size is wgpu's to check on every draw.
+    fn data_layout(device: &Device, stride: u64) -> BindGroupLayout {
+        device.create_bind_group_layout(&BindGroupLayoutDescriptor {
+            entries: &[BindGroupLayoutEntry {
+                binding: 0,
+                visibility: ShaderStages::FRAGMENT,
+                ty: BindingType::Buffer {
+                    ty: BufferBindingType::Storage { read_only: true },
+                    has_dynamic_offset: false,
+                    min_binding_size: BufferSize::new(stride),
+                },
+                count: None,
+            }],
+            label: Some("ui primitive data"),
+        })
     }
 }
 
-pub struct UiLimits {
+impl RenderLayer {
-    max_textures: u32,
+    fn new() -> Self {
-    max_samplers: u32,
-}
-
-impl Default for UiLimits {
-    fn default() -> Self {
         Self {
-            max_textures: 100000,
+            primitives: Vec::new(),
-            max_samplers: 1000,
         }
     }
 }
 
-impl UiLimits {
+impl ListBuffers {
-    pub fn max_binding_array_elements_per_shader_stage(&self) -> u32 {
+    fn new(device: &Device) -> Self {
-        self.max_textures + self.max_samplers
+        Self {
+            instance: ArrBuf::new(
+                device,
+                BufferUsages::VERTEX | BufferUsages::COPY_DST,
+                "instance",
+            ),
+            data: ArrBuf::new(
+                device,
+                BufferUsages::STORAGE | BufferUsages::COPY_DST,
+                "primitive data",
+            ),
+            group: None,
+            bindings: Vec::new(),
         }
-    pub fn max_binding_array_sampler_elements_per_shader_stage(&self) -> u32 {
+    }
-        self.max_samplers
+
+    fn update(
+        &mut self,
+        device: &Device,
+        queue: &Queue,
+        primitive: &PrimitivePipeline,
+        list: &InstanceList,
+    ) {
+        self.bindings.clear();
+        primitive.render.instance_bindings(list, &mut self.bindings);
+        self.instance.update(device, queue, list.instances());
+        let resized = self.data.update(device, queue, list.data());
+        if list.instances().is_empty() {
+            self.group = None;
+        } else if resized || self.group.is_none() {
+            self.group = Some(device.create_bind_group(&BindGroupDescriptor {
+                layout: &primitive.data_layout,
+                entries: &[BindGroupEntry {
+                    binding: 0,
+                    resource: self.data.buffer.as_entire_binding(),
+                }],
+                label: Some("ui primitive data"),
+            }));
+        }
     }
 }

core/src/render/page.rs

@@ -0,0 +1,156 @@
+use wgpu::*;
+
+use crate::{GlyphAtlas, UiData};
+
+use super::{
+    atlas::PAGE,
+    primitive::{ListDraw, PrimitiveRender},
+    texture::{default_sampler, sampled_group, sampled_layout, write_region},
+};
+
+/// Draws glyphs from the atlas, which it owns: one array texture bound once
+/// for a whole list, since every glyph in it reads the same pages.
+pub struct GlyphRender {
+    pages: GpuPages,
+    layout: BindGroupLayout,
+    sampler: Sampler,
+}
+
+impl GlyphRender {
+    pub fn new(device: &Device, queue: &Queue) -> Self {
+        let layout = sampled_layout(device, TextureViewDimension::D2Array, "ui atlas");
+        let sampler = default_sampler(device);
+        Self {
+            pages: GpuPages::new(device, queue, &layout, &sampler),
+            layout,
+            sampler,
+        }
+    }
+}
+
+impl PrimitiveRender for GlyphRender {
+    fn layout(&self) -> Option<&BindGroupLayout> {
+        Some(&self.layout)
+    }
+
+    fn update(&mut self, ui: &mut UiData) {
+        self.pages
+            .update(&mut ui.text.atlas, &self.layout, &self.sampler);
+    }
+
+    fn draw<'a>(&'a self, pass: &mut RenderPass<'a>, list: ListDraw<'a>) {
+        pass.set_bind_group(2, self.pages.group(), &[]);
+        pass.draw(0..4, 0..list.instances);
+    }
+}
+
+/// The glyph atlas on the GPU: one array texture whose layers are the pages
+/// `GlyphAtlas` packs.
+///
+/// One array rather than a texture per page because a layer index is ordinary
+/// Vulkan 1.0 / GLES sampling, where a `binding_array` would need
+/// `VK_EXT_descriptor_indexing`, which a real share of Android GPUs lack.
+pub struct GpuPages {
+    device: Device,
+    queue: Queue,
+    texture: Texture,
+    group: BindGroup,
+}
+
+impl GpuPages {
+    pub fn new(
+        device: &Device,
+        queue: &Queue,
+        layout: &BindGroupLayout,
+        sampler: &Sampler,
+    ) -> Self {
+        let texture = create_array(device, 1);
+        Self {
+            device: device.clone(),
+            queue: queue.clone(),
+            group: atlas_group(device, layout, &texture, sampler),
+            texture,
+        }
+    }
+
+    pub fn update(&mut self, atlas: &mut GlyphAtlas, layout: &BindGroupLayout, sampler: &Sampler) {
+        if atlas.page_count() > self.texture.depth_or_array_layers() {
+            self.grow(atlas.page_count(), layout, sampler);
+        }
+        for (upload, page) in atlas.uploads() {
+            let dst = TexelCopyTextureInfo {
+                texture: &self.texture,
+                mip_level: 0,
+                origin: Origin3d {
+                    x: upload.rect.x,
+                    y: upload.rect.y,
+                    z: upload.layer,
+                },
+                aspect: TextureAspect::All,
+            };
+            write_region(&self.queue, dst, page, upload.rect);
+        }
+    }
+
+    pub fn group(&self) -> &BindGroup {
+        &self.group
+    }
+
+    /// Doubles until `needed` fits and copies the old layers across GPU side.
+    /// The new texture stales the group, so that is rebuilt here.
+    fn grow(&mut self, needed: u32, layout: &BindGroupLayout, sampler: &Sampler) {
+        let old = self.texture.depth_or_array_layers();
+        let mut layers = old;
+        while layers < needed {
+            layers *= 2;
+        }
+        let texture = create_array(&self.device, layers);
+        let mut encoder = self
+            .device
+            .create_command_encoder(&CommandEncoderDescriptor {
+                label: Some("atlas grow"),
+            });
+        encoder.copy_texture_to_texture(
+            self.texture.as_image_copy(),
+            texture.as_image_copy(),
+            Extent3d {
+                width: PAGE,
+                height: PAGE,
+                depth_or_array_layers: old,
+            },
+        );
+        self.queue.submit(std::iter::once(encoder.finish()));
+        self.group = atlas_group(&self.device, layout, &texture, sampler);
+        self.texture = texture;
+    }
+}
+
+fn atlas_group(
+    device: &Device,
+    layout: &BindGroupLayout,
+    texture: &Texture,
+    sampler: &Sampler,
+) -> BindGroup {
+    let view = texture.create_view(&TextureViewDescriptor {
+        dimension: Some(TextureViewDimension::D2Array),
+        ..Default::default()
+    });
+    sampled_group(device, layout, &view, sampler, "ui atlas")
+}
+
+fn create_array(device: &Device, layers: u32) -> Texture {
+    device.create_texture(&TextureDescriptor {
+        label: Some("glyph atlas"),
+        size: Extent3d {
+            width: PAGE,
+            height: PAGE,
+            depth_or_array_layers: layers,
+        },
+        mip_level_count: 1,
+        sample_count: 1,
+        dimension: TextureDimension::D2,
+        format: TextureFormat::Rgba8Unorm,
+        usage: TextureUsages::TEXTURE_BINDING | TextureUsages::COPY_DST | TextureUsages::COPY_SRC,
+        view_formats: &[],
+    })
+}

core/src/render/primitive.rs

@@ -1,115 +1,247 @@
-use std::ops::{Deref, DerefMut};
+use std::{any::TypeId, marker::PhantomData};
 
 use crate::{
-    Color, UiRegion, WidgetId,
+    Color, TextureHandle, UiData, UiRegion, WidgetId,
     render::{
-        ArrBuf,
         data::{MaskIdx, PrimitiveInstance},
+        page::GlyphRender,
+        texture::ImageRender,
     },
-    util::Vec2,
+    util::{HashMap, Vec2},
 };
 use bytemuck::Pod;
-use wgpu::*;
+use wgpu::{BindGroupLayout, Device, Queue, RenderPass};
 
-pub struct Primitives {
+/// One instance of a primitive, laid out as the struct its shader reads.
+///
+/// The type carries its own shader, so drawing one is all the wiring it needs:
+/// its list, free list, buffers and pipeline follow from being registered.
+pub trait Primitive: Pod + 'static {
+    /// Compiled after `prelude.wgsl`, which states what it declares and what
+    /// it is given.
+    const WGSL: &'static str;
+
+    /// Made once, the first time the renderer sees this primitive. It owns
+    /// whatever the shader samples and records the primitive's own draws; the
+    /// default owns nothing and draws every instance in one call.
+    fn render(device: &Device, queue: &Queue) -> Box<dyn PrimitiveRender>
+    where
+        Self: Sized,
+    {
+        let _ = (device, queue);
+        Box::new(Instanced)
+    }
+}
+
+/// The renderer's half of a primitive: what it samples, what it uploads, and
+/// what draws it records.
+///
+/// Everything a draw shares -- the pipeline, the window and masks, the list's
+/// own data and instance buffer -- is set before this is called. What is left
+/// is what only this primitive knows: its group 2, and how many draws its
+/// instances are.
+pub trait PrimitiveRender {
+    /// The layout its shader reads at group 2. `None` for a primitive whose
+    /// shader samples nothing, whose pipeline then has no group 2 at all.
+    fn layout(&self) -> Option<&BindGroupLayout> {
+        None
+    }
+
+    /// Uploads whatever this primitive owns, once a frame, before any draw.
+    fn update(&mut self, ui: &mut UiData) {
+        let _ = ui;
+    }
+
+    /// Keeps what the primitive needs per instance at draw time, read from
+    /// the list's own data. A primitive that binds nothing per instance --
+    /// most of them -- leaves this empty and draws in one call.
+    fn instance_bindings(&self, list: &InstanceList, out: &mut Vec<u32>) {
+        let _ = (list, out);
+    }
+
+    fn draw<'a>(&'a self, pass: &mut RenderPass<'a>, list: ListDraw<'a>);
+}
+
+/// What a `PrimitiveRender` draws: this list's instances, and whatever
+/// `instance_bindings` kept for them.
+pub struct ListDraw<'a> {
+    pub instances: u32,
+    pub bindings: &'a [u32],
+}
+
+/// The default: nothing sampled, every instance in one call.
+pub struct Instanced;
+
+impl PrimitiveRender for Instanced {
+    fn draw<'a>(&'a self, pass: &mut RenderPass<'a>, list: ListDraw<'a>) {
+        pass.draw(0..4, 0..list.instances);
+    }
+}
+
+/// Which registered primitive an instance is.
+pub struct PrimitiveKind<P> {
+    id: u32,
+    _p: PhantomData<fn(P)>,
+}
+
+impl<P> PrimitiveKind<P> {
+    fn new(id: u32) -> Self {
+        Self {
+            id,
+            _p: PhantomData,
+        }
+    }
+}
+
+impl<P> Clone for PrimitiveKind<P> {
+    fn clone(&self) -> Self {
+        *self
+    }
+}
+
+impl<P> Copy for PrimitiveKind<P> {}
+
+/// Every primitive a ui can draw, in the order they were first drawn.
+#[derive(Default)]
+pub struct PrimitiveRegistry {
+    kinds: Vec<PrimitiveSource>,
+    ids: HashMap<TypeId, u32>,
+}
+
+pub struct PrimitiveSource {
+    pub wgsl: &'static str,
+    pub label: &'static str,
+    /// Size of one instance's entry, stated as the data binding's minimum.
+    pub stride: u64,
+    pub render: fn(&Device, &Queue) -> Box<dyn PrimitiveRender>,
+}
+
+impl PrimitiveRegistry {
+    /// Registers `P` if this is the first time it has been drawn.
+    pub fn kind<P: Primitive>(&mut self) -> PrimitiveKind<P> {
+        let Self { kinds, ids } = self;
+        let id = *ids.entry(TypeId::of::<P>()).or_insert_with(|| {
+            kinds.push(PrimitiveSource {
+                wgsl: P::WGSL,
+                label: std::any::type_name::<P>(),
+                stride: size_of::<P>() as u64,
+                render: P::render,
+            });
+            kinds.len() as u32 - 1
+        });
+        PrimitiveKind::new(id)
+    }
+
+    pub fn sources(&self) -> &[PrimitiveSource] {
+        &self.kinds
+    }
+}
+
+/// One registered primitive's instances in one layer. Everything per-instance
+/// rides here, so it stays in step through a `swap_remove`.
+pub struct InstanceList {
     instances: Vec<PrimitiveInstance>,
+    /// The widget each instance belongs to, for renumbering its handles.
     assoc: Vec<WidgetId>,
-    data: PrimitiveData,
     free: Vec<usize>,
+    /// `stride` bytes of the primitive's own data per instance.
+    data: Vec<u8>,
+    /// From the type the list was made for, so a write is never checked.
+    stride: usize,
+}
+
+impl InstanceList {
+    fn new<P: Primitive>() -> Self {
+        Self {
+            instances: Vec::new(),
+            assoc: Vec::new(),
+            free: Vec::new(),
+            data: Vec::new(),
+            stride: size_of::<P>(),
+        }
+    }
+
+    pub fn instances(&self) -> &[PrimitiveInstance] {
+        &self.instances
+    }
+
+    pub fn data(&self) -> &[u8] {
+        &self.data
+    }
+
+    pub fn stride(&self) -> usize {
+        self.stride
+    }
+
+    fn push(&mut self, id: WidgetId, inst: PrimitiveInstance, data: &[u8]) -> usize {
+        if let Some(i) = self.free.pop() {
+            self.instances[i] = inst;
+            self.assoc[i] = id;
+            self.data[i * self.stride..][..self.stride].copy_from_slice(data);
+            i
+        } else {
+            let i = self.instances.len();
+            self.instances.push(inst);
+            self.assoc.push(id);
+            self.data.extend_from_slice(data);
+            i
+        }
+    }
+
+    fn free(&mut self, i: usize) -> MaskIdx {
+        self.free.push(i);
+        self.instances[i].mask_idx
+    }
+
+    fn apply_free(&mut self, kind: u32) -> impl Iterator<Item = PrimitiveChange> {
+        self.free.sort_by(|a, b| b.cmp(a));
+        let instances = &mut self.instances;
+        let assoc = &mut self.assoc;
+        let data = &mut self.data;
+        let stride = self.stride;
+        self.free.drain(..).filter_map(move |i| {
+            instances.swap_remove(i);
+            assoc.swap_remove(i);
+            let last = instances.len();
+            data.copy_within(last * stride..(last + 1) * stride, i * stride);
+            data.truncate(last * stride);
+            if i == last {
+                return None;
+            }
+            let id = assoc[i];
+            Some(PrimitiveChange {
+                id,
+                kind,
+                old: last,
+                new: i,
+            })
+        })
+    }
+}
+
+/// Everything one layer draws, one list per registered primitive.
+pub struct LayerDraws {
+    /// `None` until this layer draws that primitive, because only the write
+    /// knows the type the list is for.
+    primitives: Vec<Option<InstanceList>>,
     pub updated: bool,
 }
 
-impl Default for Primitives {
+impl Default for LayerDraws {
     fn default() -> Self {
         Self {
-            instances: Default::default(),
+            primitives: Vec::new(),
-            assoc: Default::default(),
-            data: Default::default(),
-            free: Vec::new(),
             updated: true,
         }
     }
 }
 
-pub trait Primitive: Pod {
+impl LayerDraws {
-    const BINDING: u32;
-    fn vec(data: &mut PrimitiveData) -> &mut PrimitiveVec<Self>;
-}
-
-macro_rules! primitives {
-    ($($name:ident: $ty:ty => $binding:expr,)*) => {
-        #[derive(Default)]
-        pub struct PrimitiveData {
-            $(pub(crate) $name: PrimitiveVec<$ty>,)*
-        }
-
-        pub struct PrimitiveBuffers {
-            $($name: ArrBuf<$ty>,)*
-        }
-
-        impl PrimitiveBuffers {
-            pub fn update(&mut self, device: &Device, queue: &Queue, data: &PrimitiveData) {
-                $(self.$name.update(device, queue, &data.$name);)*
-            }
-        }
-
-        impl PrimitiveBuffers {
-            pub const LEN: usize = primitives!(@count $($name)*);
-            pub fn buffers(&self) -> [(u32, &Buffer); Self::LEN] {
-                [
-                    $((<$ty>::BINDING, &self.$name.buffer),)*
-                ]
-            }
-            pub fn new(device: &Device) -> Self {
-                Self {
-                    $($name: ArrBuf::new(
-                        device,
-                        BufferUsages::STORAGE | BufferUsages::COPY_DST,
-                        stringify!($name),
-                    ),)*
-                }
-            }
-        }
-
-        impl PrimitiveData {
-            pub fn clear(&mut self) {
-                $(self.$name.clear();)*
-            }
-            pub fn free(&mut self, binding: u32, idx: usize) {
-                match binding {
-                    $(<$ty>::BINDING => self.$name.free(idx),)*
-                    _ => unreachable!()
-                }
-            }
-        }
-
-        $(
-            unsafe impl bytemuck::Pod for $ty {}
-            unsafe impl bytemuck::Zeroable for $ty {}
-            impl Primitive for $ty {
-                const BINDING: u32 = $binding;
-                fn vec(data: &mut PrimitiveData) -> &mut PrimitiveVec<Self> {
-                    &mut data.$name
-                }
-            }
-        )*
-    };
-    (@count $t1:tt $($t:tt)+) => { 1 + primitives!(@count $($t)+) };
-    (@count $t:tt) => { 1 };
-}
-
-pub struct PrimitiveInst<P> {
-    pub id: WidgetId,
-    pub primitive: P,
-    pub region: UiRegion,
-    pub mask_idx: MaskIdx,
-}
-
-impl Primitives {
     pub fn write<P: Primitive>(
         &mut self,
         layer: usize,
         PrimitiveInst {
+            kind,
             id,
             primitive,
             region,
@@ -117,65 +249,67 @@ impl Primitives {
         }: PrimitiveInst<P>,
     ) -> PrimitiveHandle {
         self.updated = true;
-        let vec = P::vec(&mut self.data);
+        // Grown on first use rather than sized from the registry, which a
-        let i = vec.add(primitive);
+        // layer cannot see.
-        let inst = PrimitiveInstance {
+        if self.primitives.len() <= kind.id as usize {
-            region,
+            self.primitives.resize_with(kind.id as usize + 1, || None);
-            idx: i as u32,
+        }
-            mask_idx,
+        let inst_idx = self.primitives[kind.id as usize]
-            binding: P::BINDING,
+            .get_or_insert_with(InstanceList::new::<P>)
-        };
+            .push(
-        let inst_i = if let Some(i) = self.free.pop() {
+                id,
-            self.instances[i] = inst;
+                PrimitiveInstance { region, mask_idx },
-            self.assoc[i] = id;
+                bytemuck::bytes_of(&primitive),
-            i
+            );
-        } else {
+        PrimitiveHandle {
-            let i = self.instances.len();
+            layer,
-            self.instances.push(inst);
+            kind: kind.id,
-            self.assoc.push(id);
+            inst_idx,
-            i
+        }
-        };
-        PrimitiveHandle::new::<P>(layer, inst_i, i)
     }
 
-    /// returns (old index, new index)
+    pub fn primitives(&self) -> &[Option<InstanceList>] {
-    pub fn apply_free(&mut self) -> impl Iterator<Item = PrimitiveChange> {
+        &self.primitives
-        self.free.sort_by(|a, b| b.cmp(a));
-        self.free.drain(..).filter_map(|i| {
-            self.instances.swap_remove(i);
-            self.assoc.swap_remove(i);
-            if i == self.instances.len() {
-                return None;
     }
-            let id = self.assoc[i];
+
-            let old = self.instances.len();
+    pub fn apply_free(&mut self) -> impl Iterator<Item = PrimitiveChange> {
-            Some(PrimitiveChange { id, old, new: i })
+        self.primitives
-        })
+            .iter_mut()
+            .enumerate()
+            .filter_map(|(kind, list)| Some((kind as u32, list.as_mut()?)))
+            .flat_map(|(kind, list)| list.apply_free(kind))
     }
 
     pub fn free(&mut self, h: &PrimitiveHandle) -> MaskIdx {
         self.updated = true;
-        self.data.free(h.binding, h.data_idx);
+        self.list(h).free(h.inst_idx)
-        self.free.push(h.inst_idx);
-        self.instances[h.inst_idx].mask_idx
-    }
-
-    pub fn data(&self) -> &PrimitiveData {
-        &self.data
-    }
-
-    pub fn instances(&self) -> &Vec<PrimitiveInstance> {
-        &self.instances
     }
 
     pub fn region_mut(&mut self, h: &PrimitiveHandle) -> &mut UiRegion {
         self.updated = true;
-        &mut self.instances[h.inst_idx].region
+        &mut self.list(h).instances[h.inst_idx].region
     }
+
+    /// A handle is only ever made by `write`, which is what created the list.
+    fn list(&mut self, h: &PrimitiveHandle) -> &mut InstanceList {
+        self.primitives[h.kind as usize]
+            .as_mut()
+            .expect("handle names a primitive this layer never drew")
+    }
+}
+
+pub struct PrimitiveInst<P> {
+    pub kind: PrimitiveKind<P>,
+    pub id: WidgetId,
+    pub primitive: P,
+    pub region: UiRegion,
+    pub mask_idx: MaskIdx,
 }
 
 pub struct PrimitiveChange {
     pub id: WidgetId,
+    /// Which registered primitive's list moved, since they index separately.
+    pub kind: u32,
     pub old: usize,
     pub new: usize,
 }
@@ -183,30 +317,12 @@ pub struct PrimitiveChange {
 #[derive(Debug)]
 pub struct PrimitiveHandle {
     pub layer: usize,
+    pub kind: u32,
     pub inst_idx: usize,
-    pub data_idx: usize,
-    pub binding: u32,
 }
 
-impl PrimitiveHandle {
-    fn new<P: Primitive>(layer: usize, inst_idx: usize, data_idx: usize) -> Self {
-        Self {
-            layer,
-            inst_idx,
-            data_idx,
-            binding: P::BINDING,
-        }
-    }
-}
-
-primitives!(
-    rects: RectPrimitive => 0,
-    textures: TexturePrimitive => 1,
-    glyphs: GlyphPrimitive => 2,
-);
-
 #[repr(C)]
-#[derive(Copy, Clone)]
+#[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
 pub struct RectPrimitive {
     pub color: Color<u8>,
     pub radius: f32,
@@ -214,6 +330,10 @@ pub struct RectPrimitive {
     pub inner_radius: f32,
 }
 
+impl Primitive for RectPrimitive {
+    const WGSL: &'static str = include_str!("shader/rect.wgsl");
+}
+
 impl RectPrimitive {
     pub fn color(color: Color<u8>) -> Self {
         Self {
@@ -225,71 +345,51 @@ impl RectPrimitive {
     }
 }
 
-#[repr(C)]
+/// `color` is multiplied by the atlas alpha for a mask glyph; a colour glyph
-#[derive(Debug, Copy, Clone)]
+/// takes the texel unchanged, which `GlyphEntry::IS_COLORED` selects.
-pub struct TexturePrimitive {
+#[repr(C, align(8))]
-    pub view_idx: u32,
-    pub sampler_idx: u32,
-}
-
-#[repr(C)]
 #[derive(Debug, Copy, Clone)]
 pub struct GlyphPrimitive {
     pub uv_min: Vec2,
     pub uv_max: Vec2,
-    pub view_idx: u32,
+    /// Which atlas array layer this glyph is on.
-    pub sampler_idx: u32,
+    pub layer: u32,
     pub color: Color<u8>,
     pub flags: u32,
 }
 
-pub struct PrimitiveVec<T> {
+// Manual rather than derived: the align(8) leaves four bytes of padding, which
-    vec: Vec<T>,
+// is how WGSL lays the struct out.
-    free: Vec<usize>,
+unsafe impl bytemuck::Pod for GlyphPrimitive {}
+unsafe impl bytemuck::Zeroable for GlyphPrimitive {}
+impl Primitive for GlyphPrimitive {
+    const WGSL: &'static str = include_str!("shader/glyph.wgsl");
+
+    fn render(device: &Device, queue: &Queue) -> Box<dyn PrimitiveRender> {
+        Box::new(GlyphRender::new(device, queue))
+    }
 }
 
-impl<T> PrimitiveVec<T> {
+/// One drawn image. Its shader reads nothing per instance; the slot names the
-    pub fn new() -> Self {
+/// texture to bind for it.
+#[repr(C)]
+#[derive(Debug, Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
+pub struct TexturePrimitive {
+    pub slot: u32,
+}
+
+impl Primitive for TexturePrimitive {
+    const WGSL: &'static str = include_str!("shader/texture.wgsl");
+
+    fn render(device: &Device, queue: &Queue) -> Box<dyn PrimitiveRender> {
+        Box::new(ImageRender::new(device, queue))
+    }
+}
+
+impl From<&TextureHandle> for TexturePrimitive {
+    fn from(handle: &TextureHandle) -> Self {
         Self {
-            vec: Vec::new(),
+            slot: handle.slot(),
-            free: Vec::new(),
         }
     }
-    pub fn add(&mut self, t: T) -> usize {
-        if let Some(i) = self.free.pop() {
-            self.vec[i] = t;
-            i
-        } else {
-            let i = self.vec.len();
-            self.vec.push(t);
-            i
-        }
-    }
-    pub fn free(&mut self, i: usize) {
-        self.free.push(i);
-    }
-    pub fn clear(&mut self) {
-        self.free.clear();
-        self.vec.clear();
-    }
-}
-
-impl<T> Default for PrimitiveVec<T> {
-    fn default() -> Self {
-        Self::new()
-    }
-}
-
-impl<T> Deref for PrimitiveVec<T> {
-    type Target = Vec<T>;
-
-    fn deref(&self) -> &Self::Target {
-        &self.vec
-    }
-}
-
-impl<T> DerefMut for PrimitiveVec<T> {
-    fn deref_mut(&mut self) -> &mut Self::Target {
-        &mut self.vec
-    }
 }

core/src/render/shader.wgsl

@@ -1,204 +0,0 @@
-const RECT: u32 = 0u;
-const TEXTURE: u32 = 1u;
-const GLYPH: u32 = 2u;
-
-@group(0) @binding(0)
-var<uniform> window: WindowUniform;
-@group(1) @binding(RECT)
-var<storage> rects: array<Rect>;
-@group(1) @binding(TEXTURE)
-var<storage> textures: array<TextureInfo>;
-@group(1) @binding(GLYPH)
-var<storage> glyphs: array<GlyphInfo>;
-
-struct Rect {
-    color: u32,
-    radius: f32,
-    thickness: f32,
-    inner_radius: f32,
-}
-
-struct TextureInfo {
-    view_idx: u32,
-    sampler_idx: u32,
-}
-
-struct GlyphInfo {
-    uv_min: vec2<f32>,
-    uv_max: vec2<f32>,
-    view_idx: u32,
-    sampler_idx: u32,
-    color: u32,
-    flags: u32,
-}
-
-struct Mask {
-    x: UiSpan,
-    y: UiSpan,
-}
-
-struct UiSpan {
-    start: UiScalar,
-    end: UiScalar,
-}
-
-struct UiScalar {
-    rel: f32,
-    abs: f32,
-}
-
-struct UiVec2 {
-    rel: vec2<f32>,
-    abs: vec2<f32>,
-}
-
-@group(2) @binding(0)
-var views: binding_array<texture_2d<f32>>;
-@group(2) @binding(1)
-var samplers: binding_array<sampler>;
-@group(2) @binding(2)
-var<storage> masks: array<Mask>;
-
-struct WindowUniform {
-    dim: vec2<f32>,
-};
-
-struct InstanceInput {
-    @location(0) x_start: vec2<f32>,
-    @location(1) x_end: vec2<f32>,
-    @location(2) y_start: vec2<f32>,
-    @location(3) y_end: vec2<f32>,
-    @location(4) binding: u32,
-    @location(5) idx: u32,
-    @location(6) mask_idx: u32,
-}
-
-struct VertexOutput {
-    @location(0) top_left: vec2<f32>,
-    @location(1) bot_right: vec2<f32>,
-    @location(2) uv: vec2<f32>,
-    @location(3) binding: u32,
-    @location(4) idx: u32,
-    @location(5) mask_idx: u32,
-    @builtin(position) clip_position: vec4<f32>,
-};
-
-struct Region {
-    pos: vec2<f32>,
-    uv: vec2<f32>,
-    top_left: vec2<f32>,
-    bot_right: vec2<f32>,
-}
-
-@vertex
-fn vs_main(
-    @builtin(vertex_index) vi: u32,
-    in: InstanceInput,
-) -> VertexOutput {
-    var out: VertexOutput;
-
-    let top_left_rel = vec2(in.x_start.x, in.y_start.x);
-    let top_left_abs = vec2(in.x_start.y, in.y_start.y);
-    let bot_right_rel = vec2(in.x_end.x, in.y_end.x);
-    let bot_right_abs = vec2(in.x_end.y, in.y_end.y);
-
-    let top_left = floor(top_left_rel * window.dim) + floor(top_left_abs);
-    let bot_right = floor(bot_right_rel * window.dim) + floor(bot_right_abs);
-    let size = bot_right - top_left;
-
-    let uv = vec2<f32>(
-        f32(vi % 2u),
-        f32(vi / 2u)
-    );
-    let pos = (top_left + uv * size) / window.dim * 2.0 - 1.0;
-    out.clip_position = vec4<f32>(pos.x, -pos.y, 0.0, 1.0);
-    out.uv = uv;
-    out.binding = in.binding;
-    out.idx = in.idx;
-    out.top_left = top_left;
-    out.bot_right = bot_right;
-    out.mask_idx = in.mask_idx;
-
-    return out;
-}
-
-@fragment
-fn fs_main(
-    in: VertexOutput
-) -> @location(0) vec4<f32> {
-    let pos = in.clip_position.xy;
-    let region = Region(pos, in.uv, in.top_left, in.bot_right);
-    let i = in.idx;
-    var color: vec4<f32>;
-    switch in.binding {
-        case RECT: {
-            color = draw_rounded_rect(region, rects[i]);
-        }
-        case TEXTURE: {
-            color = draw_texture(region, textures[i]);
-        }
-        case GLYPH: {
-            color = draw_glyph(region, glyphs[i]);
-        }
-        default: {
-            color = vec4(1.0, 0.0, 1.0, 1.0);
-        }
-    }
-    if in.mask_idx != 4294967295u {
-        let mask = masks[in.mask_idx];
-        let tl = UiVec2(vec2(mask.x.start.rel, mask.y.start.rel), vec2(mask.x.start.abs, mask.y.start.abs));
-        let br = UiVec2(vec2(mask.x.end.rel, mask.y.end.rel), vec2(mask.x.end.abs, mask.y.end.abs));
-
-        let top_left = floor(tl.rel * window.dim) + floor(tl.abs);
-        let bot_right = floor(br.rel * window.dim) + floor(br.abs);
-        if pos.x < top_left.x || pos.x > bot_right.x || pos.y < top_left.y || pos.y > bot_right.y {
-            color *= 0.0;
-        }
-    }
-    return color;
-}
-
-// TODO: this seems really inefficient (per frag indexing)?
-fn draw_texture(region: Region, info: TextureInfo) -> vec4<f32> {
-    return textureSample(views[info.view_idx], samplers[info.sampler_idx], region.uv);
-}
-
-fn draw_glyph(region: Region, g: GlyphInfo) -> vec4<f32> {
-    let uv = mix(g.uv_min, g.uv_max, region.uv);
-    let texel = textureSample(views[g.view_idx], samplers[g.sampler_idx], uv);
-    if (g.flags & 1u) != 0u {
-        return texel;
-    }
-    var color = unpack4x8unorm(g.color);
-    color.a *= texel.a;
-    return color;
-}
-
-fn draw_rounded_rect(region: Region, rect: Rect) -> vec4<f32> {
-    var color = unpack4x8unorm(rect.color);
-
-    let edge = 0.5;
-
-    let size = region.bot_right - region.top_left;
-    let corner = size / 2.0;
-    let center = region.top_left + corner;
-
-    let dist = distance_from_rect(region.pos, center, corner, rect.radius);
-    color.a *= 1.0 - smoothstep(-min(edge, rect.radius), edge, dist);
-
-    if rect.thickness > 0.0 {
-        let dist2 = distance_from_rect(region.pos, center, corner - rect.thickness, rect.inner_radius);
-        color.a *= smoothstep(-min(edge, rect.inner_radius), edge, dist2);
-    }
-
-    return color;
-}
-
-fn distance_from_rect(pixel_pos: vec2<f32>, rect_center: vec2<f32>, rect_corner: vec2<f32>, radius: f32) -> f32 {
-    // vec from center to pixel
-    let p = pixel_pos - rect_center;
-    // vec from inner rect corner to pixel
-    let q = abs(p) - (rect_corner - radius);
-    return length(max(q, vec2(0.0))) - radius;
-}
-

core/src/render/shader/glyph.wgsl

@@ -0,0 +1,33 @@
+// Matches `GlyphEntry::IS_COLORED`.
+const COLORED: u32 = 1u;
+
+// The glyph atlas, whose array layers are its pages.
+@group(2) @binding(0)
+var atlas: texture_2d_array<f32>;
+@group(2) @binding(1)
+var samp: sampler;
+
+struct GlyphInfo {
+    uv_min: vec2<f32>,
+    uv_max: vec2<f32>,
+    // Which layer of the atlas array this glyph's page is.
+    layer: u32,
+    color: u32,
+    flags: u32,
+}
+
+@group(1) @binding(0)
+var<storage> glyphs: array<GlyphInfo>;
+
+@fragment
+fn fs_main(in: VertexOutput) -> @location(0) vec4<f32> {
+    let g = glyphs[in.idx];
+    let uv = mix(g.uv_min, g.uv_max, in.uv);
+    let texel = textureSample(atlas, samp, uv, i32(g.layer));
+    if (g.flags & COLORED) != 0u {
+        return masked(in, texel);
+    }
+    var color = unpack4x8unorm(g.color);
+    color.a *= texel.a;
+    return masked(in, color);
+}

core/src/render/shader/prelude.wgsl

@@ -0,0 +1,96 @@
+// Prepended to every primitive's shader, which declares its own instance data
+// as `var<storage> <name>: array<T>` at group 1 binding 0, and an `fs_main`
+// shading one instance of it. What it samples, if anything, is bound at group
+// 2: the texture at binding 0 and the sampler at binding 1.
+
+@group(0) @binding(0)
+var<uniform> window: WindowUniform;
+@group(0) @binding(1)
+var<storage> masks: array<Mask>;
+
+struct WindowUniform {
+    dim: vec2<f32>,
+};
+
+struct Mask {
+    x: UiSpan,
+    y: UiSpan,
+}
+
+struct UiSpan {
+    start: UiScalar,
+    end: UiScalar,
+}
+
+struct UiScalar {
+    rel: f32,
+    abs: f32,
+}
+
+struct InstanceInput {
+    @location(0) x_start: vec2<f32>,
+    @location(1) x_end: vec2<f32>,
+    @location(2) y_start: vec2<f32>,
+    @location(3) y_end: vec2<f32>,
+    @location(4) mask_idx: u32,
+}
+
+struct VertexOutput {
+    @location(0) top_left: vec2<f32>,
+    @location(1) bot_right: vec2<f32>,
+    @location(2) uv: vec2<f32>,
+    @location(3) @interpolate(flat) mask_idx: u32,
+    @location(4) @interpolate(flat) idx: u32,
+    @builtin(position) clip_position: vec4<f32>,
+};
+
+@vertex
+fn vs_main(
+    @builtin(vertex_index) vi: u32,
+    @builtin(instance_index) ii: u32,
+    in: InstanceInput,
+) -> VertexOutput {
+    var out: VertexOutput;
+
+    let top_left_rel = vec2(in.x_start.x, in.y_start.x);
+    let top_left_abs = vec2(in.x_start.y, in.y_start.y);
+    let bot_right_rel = vec2(in.x_end.x, in.y_end.x);
+    let bot_right_abs = vec2(in.x_end.y, in.y_end.y);
+
+    let top_left = floor(top_left_rel * window.dim) + floor(top_left_abs);
+    let bot_right = floor(bot_right_rel * window.dim) + floor(bot_right_abs);
+    let size = bot_right - top_left;
+
+    let uv = vec2<f32>(
+        f32(vi % 2u),
+        f32(vi / 2u)
+    );
+    let pos = (top_left + uv * size) / window.dim * 2.0 - 1.0;
+    out.clip_position = vec4<f32>(pos.x, -pos.y, 0.0, 1.0);
+    out.uv = uv;
+    out.top_left = top_left;
+    out.bot_right = bot_right;
+    out.mask_idx = in.mask_idx;
+    out.idx = ii;
+
+    return out;
+}
+
+fn masked(in: VertexOutput, color: vec4<f32>) -> vec4<f32> {
+    if in.mask_idx == 4294967295u {
+        return color;
+    }
+    let mask = masks[in.mask_idx];
+    let tl = vec2(mask.x.start.rel, mask.y.start.rel);
+    let tl_abs = vec2(mask.x.start.abs, mask.y.start.abs);
+    let br = vec2(mask.x.end.rel, mask.y.end.rel);
+    let br_abs = vec2(mask.x.end.abs, mask.y.end.abs);
+
+    let top_left = floor(tl * window.dim) + floor(tl_abs);
+    let bot_right = floor(br * window.dim) + floor(br_abs);
+    let pos = in.clip_position.xy;
+    if pos.x < top_left.x || pos.x > bot_right.x || pos.y < top_left.y || pos.y > bot_right.y {
+        return color * 0.0;
+    }
+    return color;
+}

core/src/render/shader/rect.wgsl

@@ -0,0 +1,39 @@
+struct Rect {
+    color: u32,
+    radius: f32,
+    thickness: f32,
+    inner_radius: f32,
+}
+
+@group(1) @binding(0)
+var<storage> rects: array<Rect>;
+
+@fragment
+fn fs_main(in: VertexOutput) -> @location(0) vec4<f32> {
+    let rect = rects[in.idx];
+    var color = unpack4x8unorm(rect.color);
+
+    let edge = 0.5;
+    let size = in.bot_right - in.top_left;
+    let corner = size / 2.0;
+    let center = in.top_left + corner;
+    let pos = in.clip_position.xy;
+
+    let dist = distance_from_rect(pos, center, corner, rect.radius);
+    color.a *= 1.0 - smoothstep(-min(edge, rect.radius), edge, dist);
+
+    if rect.thickness > 0.0 {
+        let dist2 = distance_from_rect(pos, center, corner - rect.thickness, rect.inner_radius);
+        color.a *= smoothstep(-min(edge, rect.inner_radius), edge, dist2);
+    }
+
+    return masked(in, color);
+}
+
+fn distance_from_rect(pixel_pos: vec2<f32>, rect_center: vec2<f32>, rect_corner: vec2<f32>, radius: f32) -> f32 {
+    // vec from center to pixel
+    let p = pixel_pos - rect_center;
+    // vec from inner rect corner to pixel
+    let q = abs(p) - (rect_corner - radius);
+    return length(max(q, vec2(0.0))) - radius;
+}

core/src/render/shader/texture.wgsl

@@ -0,0 +1,10 @@
+// The image this instance draws, bound for it alone.
+@group(2) @binding(0)
+var image: texture_2d<f32>;
+@group(2) @binding(1)
+var samp: sampler;
+
+@fragment
+fn fs_main(in: VertexOutput) -> @location(0) vec4<f32> {
+    return masked(in, textureSample(image, samp, in.uv));
+}

core/src/render/texture.rs

@@ -1,119 +1,119 @@
-use image::{DynamicImage, EncodableLayout, GenericImageView};
+use image::{DynamicImage, EncodableLayout, GenericImageView, RgbaImage};
 use wgpu::{util::DeviceExt, *};
 
-use crate::{PatchRect, TextureUpdate, Textures};
+use crate::{
+    PatchRect, TextureUpdate, Textures, UiData,
+    render::{
+        TexturePrimitive,
+        primitive::{ListDraw, PrimitiveRender},
+    },
+};
 
+/// Draws standalone images, which it owns. Each is its own texture, so each
+/// instance binds its own and is a draw of its own.
+pub struct ImageRender {
+    textures: GpuTextures,
+    layout: BindGroupLayout,
+    sampler: Sampler,
+}
+
+impl ImageRender {
+    pub fn new(device: &Device, queue: &Queue) -> Self {
+        Self {
+            textures: GpuTextures::new(device, queue),
+            layout: sampled_layout(device, TextureViewDimension::D2, "ui image"),
+            sampler: default_sampler(device),
+        }
+    }
+}
+
+impl PrimitiveRender for ImageRender {
+    fn layout(&self) -> Option<&BindGroupLayout> {
+        Some(&self.layout)
+    }
+
+    fn update(&mut self, ui: &mut UiData) {
+        self.textures
+            .update(&mut ui.textures, &self.layout, &self.sampler);
+    }
+
+    fn instance_bindings(&self, list: &super::InstanceList, out: &mut Vec<u32>) {
+        let slots = list
+            .data()
+            .chunks_exact(list.stride())
+            .map(|data| bytemuck::pod_read_unaligned::<TexturePrimitive>(data).slot);
+        out.extend(slots);
+    }
+
+    fn draw<'a>(&'a self, pass: &mut RenderPass<'a>, list: ListDraw<'a>) {
+        for (i, &slot) in list.bindings.iter().enumerate() {
+            let Some(image) = self.textures.group(slot) else {
+                continue;
+            };
+            pass.set_bind_group(2, image, &[]);
+            pass.draw(0..4, i as u32..i as u32 + 1);
+        }
+    }
+}
+
+/// The standalone images a ui draws, each its own texture and bind group --
+/// unlike the glyph atlas in `super::page`, which is one array they share.
 pub struct GpuTextures {
     device: Device,
     queue: Queue,
-    /// Parallel to `views`; patches require textures rather than views.
+    slots: Vec<Option<ImageGpu>>,
-    textures: Vec<Option<Texture>>,
+}
-    views: Vec<TextureView>,
+
-    view_count: usize,
+struct ImageGpu {
-    samplers: Vec<Sampler>,
+    /// Kept for `patch`, which needs the texture rather than the view.
-    null_view: TextureView,
+    texture: Texture,
-    no_views: Vec<TextureView>,
+    group: BindGroup,
 }
 
 impl GpuTextures {
-    pub fn update(&mut self, textures: &mut Textures) -> bool {
+    pub fn new(device: &Device, queue: &Queue) -> Self {
-        let mut bindings_changed = false;
+        Self {
+            device: device.clone(),
+            queue: queue.clone(),
+            slots: Vec::new(),
+        }
+    }
+
+    pub fn update(&mut self, textures: &mut Textures, layout: &BindGroupLayout, sampler: &Sampler) {
         for update in textures.updates() {
-            bindings_changed |= match update {
+            match update {
                 TextureUpdate::Push(image) => {
-                    self.push(image);
+                    let image = self.create(image, layout, sampler);
-                    true
+                    self.slots.push(Some(image));
                 }
                 TextureUpdate::Set(i, image) => {
-                    self.set(i, image);
+                    let image = self.create(image, layout, sampler);
-                    true
+                    self.slots[i as usize] = Some(image);
                 }
-                TextureUpdate::Patch(i, rect, image) => {
+                TextureUpdate::Patch(i, rect, image) => self.patch(i, rect, image),
-                    self.patch(i, rect, image);
+                TextureUpdate::PushFree => self.slots.push(None),
-                    false
+                TextureUpdate::SetFree => {}
+                TextureUpdate::Free(i) => self.slots[i as usize] = None,
             }
-                TextureUpdate::SetFree => {
-                    self.view_count += 1;
-                    true
         }
-                TextureUpdate::Free(i) => {
-                    self.free(i);
-                    true
-                }
-                TextureUpdate::PushFree => {
-                    self.push_free();
-                    true
-                }
-            };
-        }
-        bindings_changed
-    }
-    fn set(&mut self, i: u32, image: &DynamicImage) {
-        self.view_count += 1;
-        let (texture, view) = self.create(image);
-        self.textures[i as usize] = Some(texture);
-        self.views[i as usize] = view;
-    }
-    fn free(&mut self, i: u32) {
-        self.view_count -= 1;
-        self.textures[i as usize] = None;
-        self.views[i as usize] = self.null_view.clone();
-    }
-    fn push(&mut self, image: &DynamicImage) {
-        self.view_count += 1;
-        let (texture, view) = self.create(image);
-        self.textures.push(Some(texture));
-        self.views.push(view);
-    }
-    fn push_free(&mut self) {
-        self.view_count += 1;
-        self.textures.push(None);
-        self.views.push(self.null_view.clone());
     }
 
-    fn patch(&mut self, i: u32, rect: PatchRect, image: &DynamicImage) {
+    pub fn group(&self, slot: u32) -> Option<&BindGroup> {
-        let Some(texture) = &self.textures[i as usize] else {
+        self.slots.get(slot as usize)?.as_ref().map(|i| &i.group)
-            return;
-        };
-        if rect.width == 0 || rect.height == 0 {
-            return;
-        }
-        // `write_texture` requires tightly packed rows, unlike the atlas image.
-        let sub = image
-            .view(rect.x, rect.y, rect.width, rect.height)
-            .to_image();
-        self.queue.write_texture(
-            TexelCopyTextureInfo {
-                texture,
-                mip_level: 0,
-                origin: Origin3d {
-                    x: rect.x,
-                    y: rect.y,
-                    z: 0,
-                },
-                aspect: TextureAspect::All,
-            },
-            sub.as_bytes(),
-            TexelCopyBufferLayout {
-                offset: 0,
-                bytes_per_row: Some(rect.width * 4),
-                rows_per_image: Some(rect.height),
-            },
-            Extent3d {
-                width: rect.width,
-                height: rect.height,
-                depth_or_array_layers: 1,
-            },
-        );
     }
 
-    fn create(&self, image: &DynamicImage) -> (Texture, TextureView) {
+    fn create(
-        let image = image.to_rgba8();
+        &self,
-        let (width, height) = image.dimensions();
+        image: &DynamicImage,
+        layout: &BindGroupLayout,
+        sampler: &Sampler,
+    ) -> ImageGpu {
+        let rgba = image.to_rgba8();
+        let (width, height) = rgba.dimensions();
         let texture = self.device.create_texture_with_data(
             &self.queue,
             &TextureDescriptor {
-                label: None,
+                label: Some("image"),
                 size: Extent3d {
                     width,
                     height,
@@ -127,63 +127,115 @@ impl GpuTextures {
                 view_formats: &[],
             },
             wgt::TextureDataOrder::MipMajor,
-            image.as_bytes(),
+            rgba.as_bytes(),
         );
         let view = texture.create_view(&TextureViewDescriptor::default());
-        (texture, view)
+        let group = sampled_group(&self.device, layout, &view, sampler, "ui image");
+        ImageGpu { texture, group }
     }
 
-    pub fn new(device: &Device, queue: &Queue) -> Self {
+    fn patch(&mut self, i: u32, rect: PatchRect, image: &DynamicImage) {
-        let null_view = null_texture_view(device);
+        let Some(Some(slot)) = self.slots.get(i as usize) else {
-        Self {
+            return;
-            device: device.clone(),
+        };
-            queue: queue.clone(),
+        let dst = TexelCopyTextureInfo {
-            textures: Vec::new(),
+            texture: &slot.texture,
-            views: Vec::new(),
+            mip_level: 0,
-            samplers: vec![default_sampler(device)],
+            origin: Origin3d {
-            no_views: vec![null_view.clone()],
+                x: rect.x,
-            null_view,
+                y: rect.y,
-            view_count: 0,
+                z: 0,
+            },
+            aspect: TextureAspect::All,
+        };
+        match image.as_rgba8() {
+            Some(rgba) => write_region(&self.queue, dst, rgba, rect),
+            // The texture is rgba8, so any other layout has to be converted --
+            // and converting the rectangle is cheaper than the whole image.
+            None => {
+                let sub = image
+                    .view(rect.x, rect.y, rect.width, rect.height)
+                    .to_image();
+                write_region(&self.queue, dst, &sub, PatchRect { x: 0, y: 0, ..rect });
+            }
+        }
     }
 }
 
-    pub fn views(&self) -> Vec<&TextureView> {
+pub fn write_region(queue: &Queue, dst: TexelCopyTextureInfo, src: &RgbaImage, rect: PatchRect) {
-        if self.views.is_empty() {
+    if rect.width == 0 || rect.height == 0 {
-            &self.no_views
+        return;
-        } else {
-            &self.views
     }
-        .iter()
+    let stride = src.width() * 4;
-        .by_ref()
+    queue.write_texture(
-        .collect()
+        dst,
-    }
+        src.as_bytes(),
-
+        TexelCopyBufferLayout {
-    pub fn samplers(&self) -> Vec<&Sampler> {
+            offset: (rect.y * stride + rect.x * 4) as u64,
-        self.samplers.iter().by_ref().collect()
+            bytes_per_row: Some(stride),
-    }
+            rows_per_image: Some(rect.height),
-
+        },
-    pub fn view_count(&self) -> usize {
+        Extent3d {
-        self.view_count
+            width: rect.width,
-    }
+            height: rect.height,
-}
-
-pub fn null_texture_view(device: &Device) -> TextureView {
-    device
-        .create_texture(&TextureDescriptor {
-            label: Some("null"),
-            size: Extent3d {
-                width: 1,
-                height: 1,
             depth_or_array_layers: 1,
         },
-            mip_level_count: 1,
+    );
-            sample_count: 1,
+}
-            dimension: TextureDimension::D2,
+
-            format: TextureFormat::Rgba8Unorm,
+/// What a primitive that samples binds: a texture, and the sampler that reads
-            usage: TextureUsages::TEXTURE_BINDING,
+/// it.
-            view_formats: &[],
+pub fn sampled_group(
+    device: &Device,
+    layout: &BindGroupLayout,
+    view: &TextureView,
+    sampler: &Sampler,
+    label: &'static str,
+) -> BindGroup {
+    device.create_bind_group(&BindGroupDescriptor {
+        layout,
+        entries: &[
+            BindGroupEntry {
+                binding: 0,
+                resource: BindingResource::TextureView(view),
+            },
+            BindGroupEntry {
+                binding: 1,
+                resource: BindingResource::Sampler(sampler),
+            },
+        ],
+        label: Some(label),
+    })
+}
+
+/// The layout for one of those. The dimension differs -- the atlas is an
+/// array of pages and an image is not -- and nothing else does.
+pub fn sampled_layout(
+    device: &Device,
+    dimension: TextureViewDimension,
+    label: &'static str,
+) -> BindGroupLayout {
+    device.create_bind_group_layout(&BindGroupLayoutDescriptor {
+        entries: &[
+            BindGroupLayoutEntry {
+                binding: 0,
+                visibility: ShaderStages::FRAGMENT,
+                ty: BindingType::Texture {
+                    sample_type: TextureSampleType::Float { filterable: false },
+                    view_dimension: dimension,
+                    multisampled: false,
+                },
+                count: None,
+            },
+            BindGroupLayoutEntry {
+                binding: 1,
+                visibility: ShaderStages::FRAGMENT,
+                ty: BindingType::Sampler(SamplerBindingType::NonFiltering),
+                count: None,
+            },
+        ],
+        label: Some(label),
     })
-        .create_view(&TextureViewDescriptor::default())
 }
 
 pub fn default_sampler(device: &Device) -> Sampler {

core/src/render/util/mod.rs

@@ -21,17 +21,25 @@ impl<T: Pod> ArrBuf<T> {
             _pd: PhantomData,
         }
     }
-    pub fn update(&mut self, device: &Device, queue: &Queue, data: &[T]) {
+    /// Returns whether the `Buffer` was recreated, which stales any cached
-        if self.len != data.len() {
+    /// `BindGroup` holding it.
+    pub fn update(&mut self, device: &Device, queue: &Queue, data: &[T]) -> bool {
+        let resized = self.len != data.len();
+        if resized {
             self.len = data.len();
             self.buffer =
                 Self::init_buf(device, std::mem::size_of_val(data), self.usage, self.label);
         }
         queue.write_buffer(&self.buffer, 0, bytemuck::cast_slice(data));
+        resized
+    }
+    pub fn len(&self) -> usize {
+        self.len
     }
     fn init_buf(device: &Device, size: usize, usage: BufferUsages, label: &'static str) -> Buffer {
         let mut size = size as u64;
         if usage.contains(BufferUsages::STORAGE) {
+            // A binding cannot be empty or under the layout's minimum.
             size = size.max(std::mem::size_of::<T>() as u64);
         }
         device.create_buffer(&BufferDescriptor {
@@ -41,8 +49,4 @@ impl<T: Pod> ArrBuf<T> {
             usage,
         })
     }
-    #[allow(clippy::len_without_is_empty)]
-    pub fn len(&self) -> usize {
-        self.len
-    }
 }

core/src/ui/mod.rs

@@ -1,4 +1,6 @@
-use crate::{Mask, TextData, Textures, WeakWidget, WidgetId, Widgets, util::TrackedArena};
+use crate::{
+    Mask, PrimitiveRegistry, TextData, Textures, WeakWidget, WidgetId, Widgets, util::TrackedArena,
+};
 
 mod active;
 mod cache;
@@ -7,13 +9,15 @@ mod render_state;
 mod size;
 
 pub use active::*;
-pub use painter::Painter;
+pub use painter::{Painter, PrimitiveLike};
 pub use render_state::*;
 pub use size::*;
 
 #[derive(Default)]
 pub struct UiData {
     pub widgets: Widgets,
+    /// Every primitive this ui can draw.
+    pub primitives: PrimitiveRegistry,
     pub textures: Textures,
     pub text: TextData,
     pub masks: TrackedArena<Mask, u32>,

core/src/ui/painter.rs

@@ -1,7 +1,10 @@
 use crate::{
     Axis, Len, RenderedText, Size, SizeCtx, StrongWidget, TextAttrs, TextBuffer, TextData,
     TextureHandle, UiRegion, UiRenderState, UiRsc, UiScalar, UiVec2, Widget, WidgetId,
-    render::{GlyphPrimitive, Mask, MaskIdx, Primitive, PrimitiveHandle, PrimitiveInst},
+    render::{
+        GlyphPrimitive, Mask, MaskIdx, Primitive, PrimitiveHandle, PrimitiveInst, PrimitiveKind,
+        TexturePrimitive,
+    },
     util::Vec2,
 };
 
@@ -21,15 +24,26 @@ pub struct Painter<'a> {
 
 impl<'a> Painter<'a> {
     fn primitive_at<P: Primitive>(&mut self, primitive: P, region: UiRegion) {
+        let kind = self.rsc.ui_mut().primitives.kind::<P>();
+        self.write(kind, primitive, region);
+    }
+
+    /// Takes the kind, for a caller writing many of one primitive.
+    fn write<P: Primitive>(&mut self, kind: PrimitiveKind<P>, primitive: P, region: UiRegion) {
         let h = self.state.layers.write(
             self.layer,
             PrimitiveInst {
+                kind,
                 id: self.id,
                 primitive,
                 region,
                 mask_idx: self.mask,
             },
         );
+        self.push_primitive(h);
+    }
+
+    fn push_primitive(&mut self, h: PrimitiveHandle) {
         if self.mask != MaskIdx::NONE {
             // TODO: I have no clue if this works at all :joy:
             self.rsc.ui_mut().masks.push_ref(self.mask);
@@ -38,11 +52,13 @@ impl<'a> Painter<'a> {
     }
 
     /// Writes a primitive to be rendered
-    pub fn primitive<P: Primitive>(&mut self, primitive: P) {
+    pub fn primitive(&mut self, primitive: impl PrimitiveLike) {
+        let primitive = primitive.into_primitive(self);
         self.primitive_at(primitive, self.region)
     }
 
-    pub fn primitive_within<P: Primitive>(&mut self, primitive: P, region: UiRegion) {
+    pub fn primitive_within(&mut self, primitive: impl PrimitiveLike, region: UiRegion) {
+        let primitive = primitive.into_primitive(self);
         self.primitive_at(primitive, region.within(&self.region));
     }
 
@@ -75,21 +91,6 @@ impl<'a> Painter<'a> {
         );
     }
 
-    pub fn texture_within(&mut self, handle: &TextureHandle, region: UiRegion) {
-        self.textures.push(handle.clone());
-        self.primitive_at(handle.primitive(), region.within(&self.region));
-    }
-
-    pub fn texture(&mut self, handle: &TextureHandle) {
-        self.textures.push(handle.clone());
-        self.primitive(handle.primitive());
-    }
-
-    pub fn texture_at(&mut self, handle: &TextureHandle, region: UiRegion) {
-        self.textures.push(handle.clone());
-        self.primitive_at(handle.primitive(), region);
-    }
-
     pub fn render_text(
         &mut self,
         buffer: &mut TextBuffer,
@@ -97,10 +98,12 @@ impl<'a> Painter<'a> {
         width: Option<f32>,
     ) -> RenderedText {
         let ui = self.rsc.ui_mut();
-        ui.text.render(buffer, attrs, width, &mut ui.textures)
+        ui.text.render(buffer, attrs, width)
     }
 
+    // TODO: merge the text methods into the primitive ones.
     pub fn glyphs(&mut self, text: &RenderedText, origin: UiRegion) {
+        let kind = self.rsc.ui_mut().primitives.kind::<GlyphPrimitive>();
         for glyph in text.glyphs.iter() {
             let mut region = origin;
             region.x.end = region.x.start;
@@ -108,12 +111,12 @@ impl<'a> Painter<'a> {
             let mut region = region.offset(UiVec2::abs(glyph.offset));
             region.x.end = region.x.start + UiScalar::abs(glyph.entry.width as f32);
             region.y.end = region.y.start + UiScalar::abs(glyph.entry.height as f32);
-            self.primitive_at(
+            self.write(
+                kind,
                 GlyphPrimitive {
                     uv_min: glyph.entry.uv_min,
                     uv_max: glyph.entry.uv_max,
-                    view_idx: glyph.entry.view_idx,
+                    layer: glyph.entry.layer,
-                    sampler_idx: glyph.entry.sampler_idx,
                     color: text.color,
                     flags: glyph.entry.flags(),
                 },
@@ -169,3 +172,28 @@ impl<'a> Painter<'a> {
         self.state.size_ctx(self.id, self.region.size(), self.rsc)
     }
 }
+
+/// What `Painter::primitive` takes: a primitive, or something that yields one
+/// and does whatever else drawing it needs.
+pub trait PrimitiveLike {
+    type Primitive: Primitive;
+    fn into_primitive(self, painter: &mut Painter) -> Self::Primitive;
+}
+
+impl<P: Primitive> PrimitiveLike for P {
+    type Primitive = P;
+    fn into_primitive(self, _: &mut Painter) -> P {
+        self
+    }
+}
+
+impl PrimitiveLike for &TextureHandle {
+    type Primitive = TexturePrimitive;
+
+    /// Retains a share of the handle, so the slot the primitive names cannot
+    /// be freed and reused while it is still drawn.
+    fn into_primitive(self, painter: &mut Painter) -> TexturePrimitive {
+        painter.textures.push(self.clone());
+        self.into()
+    }
+}

core/src/ui/render_state.rs

@@ -1,13 +1,13 @@
 use crate::{
-    ActiveData, Axis, IdLike, MaskIdx, Painter, PixelRegion, PrimitiveLayers, SizeCtx,
+    ActiveData, Axis, DrawLayers, IdLike, MaskIdx, Painter, PixelRegion, SizeCtx, StrongWidget,
-    StrongWidget, UiRegion, UiRsc, UiVec2, WidgetId, Widgets,
+    UiRegion, UiRsc, UiVec2, WidgetId, Widgets,
     ui::cache::Cache,
     util::{HashMap, HashSet, Vec2, forget_ref},
 };
 
 pub struct UiRenderState {
     pub active: HashMap<WidgetId, ActiveData>,
-    pub layers: PrimitiveLayers,
+    pub layers: DrawLayers,
     pub(super) output_size: Vec2,
     pub cache: Cache,
 
@@ -243,14 +243,14 @@ impl UiRenderState {
     }
 
     pub fn debug_layers(&self) {
-        for ((idx, depth), primitives) in self.layers.iter_depth() {
+        for ((idx, depth), draws) in self.layers.iter_depth() {
             let indent = " ".repeat(depth * 2);
-            let len = primitives.instances().len();
+            let counts: Vec<String> = draws
-            print!("{indent}{idx}: {len} primitives");
+                .primitives()
-            if len >= 1 {
+                .iter()
-                print!(" ({})", primitives.instances()[0].binding);
+                .map(|l| l.as_ref().map_or(0, |l| l.instances().len()).to_string())
-            }
+                .collect();
-            println!();
+            println!("{indent}{idx}: [{}]", counts.join(", "));
         }
     }
 
@@ -308,7 +308,6 @@ impl UiRenderState {
             source,
             cache: &mut self.cache,
             text: &mut ui.text,
-            textures: &mut ui.textures,
             widgets: &ui.widgets,
             outer,
             output_size: self.output_size,

core/src/ui/size.rs

@@ -1,11 +1,10 @@
 use crate::{
-    Axis, AxisT, IdLike, Len, RenderedText, Size, TextAttrs, TextBuffer, TextData, Textures,
+    Axis, AxisT, IdLike, Len, RenderedText, Size, TextAttrs, TextBuffer, TextData, UiVec2,
-    UiVec2, WidgetAxisFns, WidgetId, Widgets, XAxis, YAxis, ui::cache::Cache, util::Vec2,
+    WidgetAxisFns, WidgetId, Widgets, XAxis, YAxis, ui::cache::Cache, util::Vec2,
 };
 
 pub struct SizeCtx<'a> {
     pub text: &'a mut TextData,
-    pub textures: &'a mut Textures,
     pub(super) source: WidgetId,
     pub(super) widgets: &'a Widgets,
     pub(super) cache: &'a mut Cache,
@@ -33,7 +32,6 @@ impl SizeCtx<'_> {
             .get_dyn_dynamic(id)
             .desired_len::<A>(&mut SizeCtx {
                 text: self.text,
-                textures: self.textures,
                 source: self.source,
                 widgets: self.widgets,
                 cache: self.cache,
@@ -82,7 +80,7 @@ impl SizeCtx<'_> {
         attrs: &TextAttrs,
         width: Option<f32>,
     ) -> RenderedText {
-        self.text.render(buffer, attrs, width, self.textures)
+        self.text.render(buffer, attrs, width)
     }
 
     pub fn label(&self, id: WidgetId) -> &String {

examples/tabs/main.rs

@@ -212,10 +212,10 @@ impl DefaultAppState for Client {
         render: &mut UiRenderState,
     ) {
         let new = format!(
-            "widgets: {}\nactive: {}\nviews: {}",
+            "widgets: {}\nactive: {}\ntextures: {}",
             rsc.widgets().len(),
             render.active_widgets(),
-            self.ui_state.renderer.ui.view_count(),
+            rsc.ui().textures.count(),
         );
         if new != *rsc.widgets()[self.info].content {
             *rsc.widgets_mut()[self.info].content = new;

src/default/render.rs

@@ -1,4 +1,4 @@
-use iris_core::{UiData, UiLimits, UiRenderNode, UiRenderState};
+use iris_core::{UiData, UiRenderNode, UiRenderState};
 use pollster::FutureExt;
 use std::sync::Arc;
 use wgpu::*;
@@ -83,18 +83,9 @@ impl UiRenderer {
             .block_on()
             .expect("Could not get adapter!");
 
-        let ui_limits = UiLimits::default();
-
         let (device, queue) = adapter
             .request_device(&DeviceDescriptor {
-                required_features: Features::TEXTURE_BINDING_ARRAY
-                    | Features::PARTIALLY_BOUND_BINDING_ARRAY
-                    | Features::SAMPLED_TEXTURE_AND_STORAGE_BUFFER_ARRAY_NON_UNIFORM_INDEXING,
                 required_limits: Limits {
-                    max_binding_array_elements_per_shader_stage: ui_limits
-                        .max_binding_array_elements_per_shader_stage(),
-                    max_binding_array_sampler_elements_per_shader_stage: ui_limits
-                        .max_binding_array_sampler_elements_per_shader_stage(),
                     max_buffer_size: 1 << 30,
                     ..Default::default()
                 },
@@ -126,7 +117,7 @@ impl UiRenderer {
 
         let encoder = Self::create_encoder(&device);
 
-        let ui = UiRenderNode::new(&device, &queue, &config, ui_limits);
+        let ui = UiRenderNode::new(&device, &config);
 
         Self {
             surface,

src/widget/image.rs

@@ -7,7 +7,7 @@ pub struct Image {
 
 impl Widget for Image {
     fn draw(&mut self, painter: &mut Painter) {
-        painter.texture(&self.handle);
+        painter.primitive(&self.handle);
     }
 
     fn desired_width(&mut self, _: &mut SizeCtx) -> Len {

tests/draw_cost.rs

@@ -0,0 +1,205 @@
+//! What one frame of `UiRenderNode::draw` costs on the CPU, against the number
+//! of layers it walks. Recording only: the pass is built and dropped without
+//! being submitted, so this is the loop's cost and not the GPU's.
+//!
+//!     cargo test --release --test draw_cost -- --ignored --nocapture
+//!
+//! **Read the instruction count, not the clock.** Wall time here swings by 2x
+//! between runs of one binary on this machine -- more under `cargo test` than
+//! run directly -- while instructions retired are stable to 0.1%:
+//!
+//!     perf stat -e instructions:u target/release/.../draw_cost-* --ignored
+//!
+//! Measured that way on 2026-09-13, drawing each primitive through its own
+//! `PrimitiveRender` rather than a match in the renderer costs **6
+//! instructions per list drawn**, which is 0.1% of a frame at both 256 and
+//! 1024 layers. Recording one list into the pass costs wgpu ~5,400.
+//!
+//! The instance is leaked on purpose. Dropping the last one makes the Vulkan
+//! loader unload Mesa's ICD, which faults when a thread that touched Vulkan
+//! exits -- and libtest runs every test on a spawned thread.
+
+use std::time::Instant;
+
+use iris::prelude::*;
+use iris_core::{
+    GlyphPrimitive, MaskIdx, PrimitiveInst, RectPrimitive, TextureHandle, TexturePrimitive, UiData,
+    UiRegion, UiRenderNode, UiRenderState,
+};
+use wgpu::{Color as GpuColor, *};
+
+const SIZE: u32 = 1024;
+const FRAMES: u32 = 200;
+/// Reported as the best of this many batches. The mean moves by 15% between
+/// runs on this machine, which is more than the thing being measured.
+const BATCHES: u32 = 8;
+
+fn gpu() -> Option<(Device, Queue)> {
+    // Probed rather than assumed: this machine's Vulkan device comes and goes,
+    // and GL is what is left when it is gone.
+    let all = Instance::new(&InstanceDescriptor::default());
+    let instance = match pollster::block_on(all.request_adapter(&RequestAdapterOptions::default()))
+    {
+        Ok(_) => all,
+        Err(_) => Instance::new(&InstanceDescriptor {
+            backends: Backends::GL,
+            ..Default::default()
+        }),
+    };
+    // Leaked rather than dropped: see the note at the top of the file.
+    let instance: &'static Instance = Box::leak(Box::new(instance));
+    let adapter =
+        pollster::block_on(instance.request_adapter(&RequestAdapterOptions::default())).ok()?;
+    println!("adapter: {:?}", adapter.get_info());
+    pollster::block_on(adapter.request_device(&DeviceDescriptor::default())).ok()
+}
+
+fn config(format: TextureFormat) -> SurfaceConfiguration {
+    SurfaceConfiguration {
+        usage: TextureUsages::RENDER_ATTACHMENT,
+        format,
+        width: SIZE,
+        height: SIZE,
+        present_mode: PresentMode::Fifo,
+        desired_maximum_frame_latency: 2,
+        alpha_mode: CompositeAlphaMode::Auto,
+        view_formats: vec![],
+    }
+}
+
+/// Every layer draws all three primitives, so the renderer takes a different
+/// path for each list it walks -- which is the case a single-primitive layer
+/// would never exercise. Images are bound per instance, so there are few.
+fn fill(
+    ui: &mut UiData,
+    render: &mut UiRenderState,
+    layers: usize,
+    per_layer: usize,
+) -> Vec<TextureHandle> {
+    let rect = ui.primitives.kind::<RectPrimitive>();
+    let glyph = ui.primitives.kind::<GlyphPrimitive>();
+    let texture = ui.primitives.kind::<TexturePrimitive>();
+    let id = ui.widgets.add_strong(Rect::new(UiColor::WHITE)).id();
+    let handles: Vec<_> = (0..4)
+        .map(|_| ui.textures.add(image::RgbaImage::new(4, 4)))
+        .collect();
+
+    let mut layer = 0;
+    for _ in 0..layers {
+        for _ in 0..per_layer {
+            render.layers.write(
+                layer,
+                PrimitiveInst {
+                    kind: rect,
+                    id,
+                    primitive: RectPrimitive::color(UiColor::WHITE),
+                    region: UiRegion::FULL,
+                    mask_idx: MaskIdx::NONE,
+                },
+            );
+            render.layers.write(
+                layer,
+                PrimitiveInst {
+                    kind: glyph,
+                    id,
+                    primitive: GlyphPrimitive {
+                        uv_min: vec2(0.0, 0.0),
+                        uv_max: vec2(1.0, 1.0),
+                        layer: 0,
+                        color: UiColor::WHITE,
+                        flags: 0,
+                    },
+                    region: UiRegion::FULL,
+                    mask_idx: MaskIdx::NONE,
+                },
+            );
+        }
+        for h in &handles[..2] {
+            render.layers.write(
+                layer,
+                PrimitiveInst {
+                    kind: texture,
+                    id,
+                    primitive: TexturePrimitive::from(h),
+                    region: UiRegion::FULL,
+                    mask_idx: MaskIdx::NONE,
+                },
+            );
+        }
+        layer = render.layers.next(layer);
+    }
+    handles
+}
+
+fn frame_cost(device: &Device, queue: &Queue, layers: usize, per_layer: usize) -> f64 {
+    let format = TextureFormat::Bgra8Unorm;
+    let mut node = UiRenderNode::new(device, &config(format));
+    let mut ui = UiData::default();
+    let mut render = UiRenderState::new();
+    let _handles = fill(&mut ui, &mut render, layers, per_layer);
+    node.update(device, queue, &mut ui, &mut render);
+
+    let target = device.create_texture(&TextureDescriptor {
+        label: Some("draw cost"),
+        size: Extent3d {
+            width: SIZE,
+            height: SIZE,
+            depth_or_array_layers: 1,
+        },
+        mip_level_count: 1,
+        sample_count: 1,
+        dimension: TextureDimension::D2,
+        format,
+        usage: TextureUsages::RENDER_ATTACHMENT,
+        view_formats: &[],
+    });
+    let view = target.create_view(&TextureViewDescriptor::default());
+
+    let record = |frames: u32| {
+        let start = Instant::now();
+        for _ in 0..frames {
+            let mut encoder = device.create_command_encoder(&CommandEncoderDescriptor::default());
+            {
+                let pass = &mut encoder.begin_render_pass(&RenderPassDescriptor {
+                    color_attachments: &[Some(RenderPassColorAttachment {
+                        view: &view,
+                        resolve_target: None,
+                        ops: Operations {
+                            load: LoadOp::Clear(GpuColor::BLACK),
+                            store: StoreOp::Store,
+                        },
+                        depth_slice: None,
+                    })],
+                    ..Default::default()
+                });
+                node.draw(pass);
+            }
+            drop(encoder.finish());
+        }
+        start.elapsed().as_secs_f64() / frames as f64
+    };
+    record(FRAMES / 4);
+    (0..BATCHES)
+        .map(|_| record(FRAMES))
+        .fold(f64::MAX, f64::min)
+}
+
+#[test]
+#[ignore = "measurement, not a check"]
+fn draw_cost_by_layer_count() {
+    let Some((device, queue)) = gpu() else {
+        panic!("no wgpu device; see the this-machine-graphics notes");
+    };
+    println!(
+        "layers, each 8 rects + 8 glyphs + 2 images: us/frame (us per layer), best of {BATCHES}"
+    );
+    let base = frame_cost(&device, &queue, 1, 8) * 1e6;
+    for layers in [8, 64, 256, 1024] {
+        let per_frame = frame_cost(&device, &queue, layers, 8) * 1e6;
+        // Net of the empty pass, which is the same in any version of this.
+        println!(
+            "{layers:>5}: {per_frame:8.1} us ({:.3} us)",
+            (per_frame - base).max(0.0) / layers as f64
+        );
+    }
+}