perf: keep heavy work off UI thread

.gitignore

@@ -5,6 +5,7 @@ target
 
 # Trunk web build output (wasm/js bundle)
 dist
+web/worker_pkg
 
 # These are backup files generated by rustfmt
 **/*.rs.bk

Cargo.lock

@@ -17,6 +17,7 @@ name = "OpenCADStudio"
 version = "0.8.7"
 dependencies = [
  "acadrust",
+ "bincode",
  "bytemuck",
  "clap",
  "console_error_panic_hook",
@@ -74,7 +75,7 @@ checksum = "366ffbaa4442f4684d91e2cd7c5ea7c4ed8add41959a31447066e279e432b618"
 [[package]]
 name = "acadrust"
 version = "0.4.0"
-source = "git+https://github.com/OpenAEC-Foundation/acadifc?rev=cd90c256a5e7d115f6e995275d6b3117d5775da9#cd90c256a5e7d115f6e995275d6b3117d5775da9"
+source = "git+https://github.com/OpenAEC-Foundation/acadifc.git?rev=bee1a58#bee1a5857d444a32b67f98e6babc1ad6483f6865"
 dependencies = [
  "ahash 0.8.12",
  "anyhow",
@@ -2305,8 +2306,6 @@ dependencies = [
 [[package]]
 name = "iced_wgpu"
 version = "0.14.0"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ff144a999b0ca0f8a10257934500060240825c42e950ec0ebee9c8ae30561c13"
 dependencies = [
  "bitflags 2.13.1",
  "bytemuck",
@@ -3667,6 +3666,17 @@ dependencies = [
  "thiserror 1.0.69",
 ]
 
+[[package]]
+name = "ocs_web_worker"
+version = "0.1.0"
+dependencies = [
+ "acadrust",
+ "bincode",
+ "getrandom 0.3.4",
+ "js-sys",
+ "wasm-bindgen",
+]
+
 [[package]]
 name = "once_cell"
 version = "1.21.4"

Cargo.toml

@@ -11,6 +11,7 @@ build = "build.rs"
 # verifies it.
 members = [
     "crates/ocs_plugin_api",
+    "crates/ocs_web_worker",
     "crates/dwg-thumbnailer",
     "crates/dwg-thumbnailer-win",
 ]
@@ -57,7 +58,7 @@ rfd = "0.17"
 clap = { version = "4", features = ["derive"] }
 # Opt-in logging via --log / RUST_LOG (surfaces wgpu / iced / winit diagnostics).
 env_logger = "0.11"
-acadrust = "0.4"
+acadrust = { version = "0.4", features = ["serde"] }
 # Shared DWG embedded-preview extraction (Start-page + file-manager thumbnails).
 dwg-thumbnailer = { path = "crates/dwg-thumbnailer" }
 flate2 = "1"
@@ -90,7 +91,11 @@ windows-sys = { version = "0.61", features = ["Win32_UI_Shell", "Win32_UI_Window
 
 [patch.crates-io]
 # Track the verified DWG round-trip, I/O, and unified PERF fixes.
-acadrust = { git = "https://github.com/OpenAEC-Foundation/acadifc", rev = "cd90c256a5e7d115f6e995275d6b3117d5775da9" }
+acadrust = { git = "https://github.com/OpenAEC-Foundation/acadifc.git", rev = "bee1a58" }
+# iced_wgpu 0.14 requests WebGL2 limits on every wasm adapter, including
+# BrowserWebGpu. Keep the release source local with the one adapter-aware limit
+# fix so WebGPU can expose storage buffers while WebGL2 remains the fallback.
+iced_wgpu = { path = "crates/iced_wgpu" }
 
 [target.'cfg(not(target_arch = "wasm32"))'.dependencies]
 # Native enables the plugin host runtime (out-of-process plugins).
@@ -119,6 +124,7 @@ wasm-bindgen = "0.2"
 # web font loader (#141).
 wasm-bindgen-futures = "0.4"
 js-sys = "0.3"
+bincode = "1.3"
 # window.open (external URLs) + Blob/anchor file downloads (Save) + fetch()
 # (Response) for the lazy per-script web font loader (#141).
 web-sys = { version = "0.3", features = [
@@ -130,6 +136,11 @@ web-sys = { version = "0.3", features = [
     "Blob",
     "Url",
     "Response",
+    "Worker",
+    "WorkerOptions",
+    "WorkerType",
+    "MessageEvent",
+    "ErrorEvent",
     # Async clipboard read for paste into the text/MText editors (iced's own
     # clipboard read is a no-op on the web).
     "Clipboard",

Trunk.toml

@@ -12,6 +12,11 @@ target = "index.html"
 # on the rust <link> in index.html (truck-meshalgo → lzma-sys C deps cannot
 # cross-compile to wasm).
 
+[[hooks]]
+stage = "pre_build"
+command = "sh"
+command_arguments = ["scripts/build-web-worker.sh"]
+
 [serve]
 # Cross-origin isolation headers enable SharedArrayBuffer, which wasm threads
 # (rayon) need. GitHub Pages cannot set these; a self-hosted server (e.g. the

crates/iced_wgpu/Cargo.toml

@@ -0,0 +1,146 @@
+# THIS FILE IS AUTOMATICALLY GENERATED BY CARGO
+#
+# When uploading crates to the registry Cargo will automatically
+# "normalize" Cargo.toml files for maximal compatibility
+# with all versions of Cargo and also rewrite `path` dependencies
+# to registry (e.g., crates.io) dependencies.
+#
+# If you are reading this file be aware that the original Cargo.toml
+# will likely look very different (and much more reasonable).
+# See Cargo.toml.orig for the original contents.
+
+[package]
+edition = "2024"
+name = "iced_wgpu"
+version = "0.14.0"
+authors = ["Héctor Ramón Jiménez <hector@hecrj.dev>"]
+build = false
+autolib = false
+autobins = false
+autoexamples = false
+autotests = false
+autobenches = false
+description = "A renderer for iced on top of wgpu"
+homepage = "https://iced.rs"
+readme = "README.md"
+keywords = [
+    "gui",
+    "ui",
+    "graphics",
+    "interface",
+    "widgets",
+]
+categories = ["gui"]
+license = "MIT"
+repository = "https://github.com/iced-rs/iced"
+
+[package.metadata.docs.rs]
+rustdoc-args = [
+    "--cfg",
+    "docsrs",
+]
+all-features = true
+
+[features]
+default = ["wgpu/default"]
+geometry = [
+    "iced_graphics/geometry",
+    "lyon",
+]
+image = ["iced_graphics/image"]
+strict-assertions = []
+svg = [
+    "iced_graphics/svg",
+    "resvg/text",
+]
+web-colors = ["iced_graphics/web-colors"]
+webgl = ["wgpu/webgl"]
+
+[lib]
+name = "iced_wgpu"
+path = "src/lib.rs"
+
+[dependencies.bitflags]
+version = "2.0"
+
+[dependencies.bytemuck]
+version = "1.0"
+features = ["derive"]
+
+[dependencies.cryoglyph]
+version = "0.1"
+
+[dependencies.futures]
+version = "0.3"
+features = [
+    "std",
+    "async-await",
+]
+default-features = false
+
+[dependencies.glam]
+version = "0.25"
+
+[dependencies.guillotiere]
+version = "0.6"
+
+[dependencies.iced_debug]
+version = "0.14.0"
+
+[dependencies.iced_graphics]
+version = "0.14.0"
+
+[dependencies.log]
+version = "0.4"
+
+[dependencies.lyon]
+version = "1.0"
+optional = true
+
+[dependencies.resvg]
+version = "0.45"
+optional = true
+
+[dependencies.rustc-hash]
+version = "2.0"
+
+[dependencies.thiserror]
+version = "2"
+
+[dependencies.wgpu]
+version = "27.0"
+features = [
+    "std",
+    "wgsl",
+]
+default-features = false
+
+[lints.clippy]
+default_trait_access = "deny"
+filter_map_next = "deny"
+from_over_into = "deny"
+large-enum-variant = "allow"
+manual_let_else = "deny"
+map-entry = "allow"
+match-wildcard-for-single-variants = "deny"
+needless_borrow = "deny"
+new_without_default = "deny"
+redundant-closure-for-method-calls = "deny"
+result_large_err = "allow"
+semicolon_if_nothing_returned = "deny"
+trivially-copy-pass-by-ref = "deny"
+type-complexity = "allow"
+unused_async = "deny"
+useless_conversion = "deny"
+
+[lints.rust]
+missing_docs = "deny"
+unsafe_code = "deny"
+unused_results = "deny"
+
+[lints.rust.rust_2018_idioms]
+level = "deny"
+priority = -1
+
+[lints.rustdoc]
+broken_intra_doc_links = "forbid"

crates/iced_wgpu/LICENSE

@@ -0,0 +1,19 @@
+Copyright 2019 Héctor Ramón, Iced contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy of
+this software and associated documentation files (the "Software"), to deal in
+the Software without restriction, including without limitation the rights to
+use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of
+the Software, and to permit persons to whom the Software is furnished to do so,
+subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

crates/iced_wgpu/OCS_PATCH.md

@@ -0,0 +1,8 @@
+# OpenCADStudio patch
+
+This directory vendors `iced_wgpu` 0.14.0 under its MIT license.
+
+OpenCADStudio changes only the WASM device-limit selection in
+`src/window/compositor.rs`: Browser WebGPU first requests normal WebGPU limits
+so storage-buffer pipelines are available, then falls back to WebGL2 limits.
+An actual WebGL adapter continues to request WebGL2 limits directly.

crates/iced_wgpu/README.md

@@ -0,0 +1,20 @@
+# `iced_wgpu`
+[![Documentation](https://docs.rs/iced_wgpu/badge.svg)][documentation]
+[![Crates.io](https://img.shields.io/crates/v/iced_wgpu.svg)](https://crates.io/crates/iced_wgpu)
+[![License](https://img.shields.io/crates/l/iced_wgpu.svg)](https://github.com/iced-rs/iced/blob/master/LICENSE)
+[![Discord Server](https://img.shields.io/discord/628993209984614400?label=&labelColor=6A7EC2&logo=discord&logoColor=ffffff&color=7389D8)](https://discord.gg/3xZJ65GAhd)
+
+`iced_wgpu` is a [`wgpu`] renderer for [`iced_runtime`]. For now, it is the default renderer of Iced on [native platforms].
+
+[`wgpu`] supports most modern graphics backends: Vulkan, Metal, DX12, OpenGL, and WebGPU.
+
+<p align="center">
+  <img alt="The native target" src="../docs/graphs/native.png" width="80%">
+</p>
+
+[documentation]: https://docs.rs/iced_wgpu
+[`iced_runtime`]: ../runtime
+[`wgpu`]: https://github.com/gfx-rs/wgpu
+[native platforms]: https://github.com/gfx-rs/wgpu#supported-platforms
+[WebGPU API]: https://gpuweb.github.io/gpuweb/
+[`wgpu_glyph`]: https://github.com/hecrj/wgpu_glyph

crates/iced_wgpu/src/buffer.rs

@@ -0,0 +1,132 @@
+use std::marker::PhantomData;
+use std::num::NonZeroU64;
+use std::ops::RangeBounds;
+
+pub const MAX_WRITE_SIZE: usize = 100 * 1024;
+
+const MAX_WRITE_SIZE_U64: NonZeroU64 = NonZeroU64::new(MAX_WRITE_SIZE as u64)
+    .expect("MAX_WRITE_SIZE must be non-zero");
+
+#[derive(Debug)]
+pub struct Buffer<T> {
+    label: &'static str,
+    size: u64,
+    usage: wgpu::BufferUsages,
+    pub(crate) raw: wgpu::Buffer,
+    type_: PhantomData<T>,
+}
+
+impl<T: bytemuck::Pod> Buffer<T> {
+    pub fn new(
+        device: &wgpu::Device,
+        label: &'static str,
+        amount: usize,
+        usage: wgpu::BufferUsages,
+    ) -> Self {
+        let size = next_copy_size::<T>(amount);
+
+        let raw = device.create_buffer(&wgpu::BufferDescriptor {
+            label: Some(label),
+            size,
+            usage,
+            mapped_at_creation: false,
+        });
+
+        Self {
+            label,
+            size,
+            usage,
+            raw,
+            type_: PhantomData,
+        }
+    }
+
+    pub fn resize(&mut self, device: &wgpu::Device, new_count: usize) -> bool {
+        let new_size = next_copy_size::<T>(new_count);
+
+        if self.size < new_size {
+            self.raw = device.create_buffer(&wgpu::BufferDescriptor {
+                label: Some(self.label),
+                size: new_size,
+                usage: self.usage,
+                mapped_at_creation: false,
+            });
+
+            self.size = new_size;
+
+            true
+        } else {
+            false
+        }
+    }
+
+    /// Returns the size of the written bytes.
+    pub fn write(
+        &mut self,
+        device: &wgpu::Device,
+        encoder: &mut wgpu::CommandEncoder,
+        belt: &mut wgpu::util::StagingBelt,
+        offset: usize,
+        contents: &[T],
+    ) -> usize {
+        let bytes: &[u8] = bytemuck::cast_slice(contents);
+        let mut bytes_written = 0;
+
+        // Split write into multiple chunks if necessary
+        while bytes_written + MAX_WRITE_SIZE < bytes.len() {
+            belt.write_buffer(
+                encoder,
+                &self.raw,
+                (offset + bytes_written) as u64,
+                MAX_WRITE_SIZE_U64,
+                device,
+            )
+            .copy_from_slice(
+                &bytes[bytes_written..bytes_written + MAX_WRITE_SIZE],
+            );
+
+            bytes_written += MAX_WRITE_SIZE;
+        }
+
+        // There will always be some bytes left, since the previous
+        // loop guarantees `bytes_written < bytes.len()`
+        let bytes_left = ((bytes.len() - bytes_written) as u64)
+            .try_into()
+            .expect("non-empty write");
+
+        // Write them
+        belt.write_buffer(
+            encoder,
+            &self.raw,
+            (offset + bytes_written) as u64,
+            bytes_left,
+            device,
+        )
+        .copy_from_slice(&bytes[bytes_written..]);
+
+        bytes.len()
+    }
+
+    pub fn slice(
+        &self,
+        bounds: impl RangeBounds<wgpu::BufferAddress>,
+    ) -> wgpu::BufferSlice<'_> {
+        self.raw.slice(bounds)
+    }
+
+    pub fn range(&self, start: usize, end: usize) -> wgpu::BufferSlice<'_> {
+        self.slice(
+            start as u64 * std::mem::size_of::<T>() as u64
+                ..end as u64 * std::mem::size_of::<T>() as u64,
+        )
+    }
+}
+
+fn next_copy_size<T>(amount: usize) -> u64 {
+    let align_mask = wgpu::COPY_BUFFER_ALIGNMENT - 1;
+
+    (((std::mem::size_of::<T>() * amount).next_power_of_two() as u64
+        + align_mask)
+        & !align_mask)
+        .max(wgpu::COPY_BUFFER_ALIGNMENT)
+}

crates/iced_wgpu/src/color.rs

@@ -0,0 +1,205 @@
+use std::borrow::Cow;
+
+use wgpu::util::DeviceExt;
+
+pub fn convert(
+    device: &wgpu::Device,
+    encoder: &mut wgpu::CommandEncoder,
+    source: wgpu::Texture,
+    format: wgpu::TextureFormat,
+) -> wgpu::Texture {
+    if source.format() == format {
+        return source;
+    }
+
+    let sampler = device.create_sampler(&wgpu::SamplerDescriptor {
+        label: Some("iced_wgpu.offscreen.sampler"),
+        ..wgpu::SamplerDescriptor::default()
+    });
+
+    #[derive(Debug, Clone, Copy, bytemuck::Zeroable, bytemuck::Pod)]
+    #[repr(C)]
+    struct Ratio {
+        u: f32,
+        v: f32,
+        // Padding field for 16-byte alignment.
+        // See https://docs.rs/wgpu/latest/wgpu/struct.DownlevelFlags.html#associatedconstant.BUFFER_BINDINGS_NOT_16_BYTE_ALIGNED
+        _padding: [f32; 2],
+    }
+
+    let ratio = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
+        label: Some("iced-wgpu::triangle::msaa ratio"),
+        contents: bytemuck::bytes_of(&Ratio {
+            u: 1.0,
+            v: 1.0,
+            _padding: [0.0; 2],
+        }),
+        usage: wgpu::BufferUsages::COPY_DST | wgpu::BufferUsages::UNIFORM,
+    });
+
+    let constant_layout =
+        device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
+            label: Some("iced_wgpu.offscreen.blit.sampler_layout"),
+            entries: &[
+                wgpu::BindGroupLayoutEntry {
+                    binding: 0,
+                    visibility: wgpu::ShaderStages::FRAGMENT,
+                    ty: wgpu::BindingType::Sampler(
+                        wgpu::SamplerBindingType::NonFiltering,
+                    ),
+                    count: None,
+                },
+                wgpu::BindGroupLayoutEntry {
+                    binding: 1,
+                    visibility: wgpu::ShaderStages::VERTEX,
+                    ty: wgpu::BindingType::Buffer {
+                        ty: wgpu::BufferBindingType::Uniform,
+                        has_dynamic_offset: false,
+                        min_binding_size: None,
+                    },
+                    count: None,
+                },
+            ],
+        });
+
+    let constant_bind_group =
+        device.create_bind_group(&wgpu::BindGroupDescriptor {
+            label: Some("iced_wgpu.offscreen.sampler.bind_group"),
+            layout: &constant_layout,
+            entries: &[
+                wgpu::BindGroupEntry {
+                    binding: 0,
+                    resource: wgpu::BindingResource::Sampler(&sampler),
+                },
+                wgpu::BindGroupEntry {
+                    binding: 1,
+                    resource: ratio.as_entire_binding(),
+                },
+            ],
+        });
+
+    let texture_layout =
+        device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
+            label: Some("iced_wgpu.offscreen.blit.texture_layout"),
+            entries: &[wgpu::BindGroupLayoutEntry {
+                binding: 0,
+                visibility: wgpu::ShaderStages::FRAGMENT,
+                ty: wgpu::BindingType::Texture {
+                    sample_type: wgpu::TextureSampleType::Float {
+                        filterable: false,
+                    },
+                    view_dimension: wgpu::TextureViewDimension::D2,
+                    multisampled: false,
+                },
+                count: None,
+            }],
+        });
+
+    let pipeline_layout =
+        device.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
+            label: Some("iced_wgpu.offscreen.blit.pipeline_layout"),
+            bind_group_layouts: &[&constant_layout, &texture_layout],
+            push_constant_ranges: &[],
+        });
+
+    let shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
+        label: Some("iced_wgpu.offscreen.blit.shader"),
+        source: wgpu::ShaderSource::Wgsl(Cow::Borrowed(include_str!(
+            "shader/blit.wgsl"
+        ))),
+    });
+
+    let pipeline =
+        device.create_render_pipeline(&wgpu::RenderPipelineDescriptor {
+            label: Some("iced_wgpu.offscreen.blit.pipeline"),
+            layout: Some(&pipeline_layout),
+            vertex: wgpu::VertexState {
+                module: &shader,
+                entry_point: Some("vs_main"),
+                buffers: &[],
+                compilation_options: wgpu::PipelineCompilationOptions::default(
+                ),
+            },
+            fragment: Some(wgpu::FragmentState {
+                module: &shader,
+                entry_point: Some("fs_main"),
+                targets: &[Some(wgpu::ColorTargetState {
+                    format,
+                    blend: Some(wgpu::BlendState {
+                        color: wgpu::BlendComponent {
+                            src_factor: wgpu::BlendFactor::SrcAlpha,
+                            dst_factor: wgpu::BlendFactor::OneMinusSrcAlpha,
+                            operation: wgpu::BlendOperation::Add,
+                        },
+                        alpha: wgpu::BlendComponent {
+                            src_factor: wgpu::BlendFactor::One,
+                            dst_factor: wgpu::BlendFactor::OneMinusSrcAlpha,
+                            operation: wgpu::BlendOperation::Add,
+                        },
+                    }),
+                    write_mask: wgpu::ColorWrites::ALL,
+                })],
+                compilation_options: wgpu::PipelineCompilationOptions::default(
+                ),
+            }),
+            primitive: wgpu::PrimitiveState {
+                topology: wgpu::PrimitiveTopology::TriangleList,
+                front_face: wgpu::FrontFace::Cw,
+                ..wgpu::PrimitiveState::default()
+            },
+            depth_stencil: None,
+            multisample: wgpu::MultisampleState::default(),
+            multiview: None,
+            cache: None,
+        });
+
+    let texture = device.create_texture(&wgpu::TextureDescriptor {
+        label: Some("iced_wgpu.offscreen.conversion.source_texture"),
+        size: source.size(),
+        mip_level_count: 1,
+        sample_count: 1,
+        dimension: wgpu::TextureDimension::D2,
+        format,
+        usage: wgpu::TextureUsages::RENDER_ATTACHMENT
+            | wgpu::TextureUsages::COPY_SRC,
+        view_formats: &[],
+    });
+
+    let view = &texture.create_view(&wgpu::TextureViewDescriptor::default());
+
+    let texture_bind_group =
+        device.create_bind_group(&wgpu::BindGroupDescriptor {
+            label: Some("iced_wgpu.offscreen.blit.texture_bind_group"),
+            layout: &texture_layout,
+            entries: &[wgpu::BindGroupEntry {
+                binding: 0,
+                resource: wgpu::BindingResource::TextureView(
+                    &source
+                        .create_view(&wgpu::TextureViewDescriptor::default()),
+                ),
+            }],
+        });
+
+    let mut pass = encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
+        label: Some("iced_wgpu.offscreen.blit.render_pass"),
+        color_attachments: &[Some(wgpu::RenderPassColorAttachment {
+            view,
+            depth_slice: None,
+            resolve_target: None,
+            ops: wgpu::Operations {
+                load: wgpu::LoadOp::Load,
+                store: wgpu::StoreOp::Store,
+            },
+        })],
+        depth_stencil_attachment: None,
+        timestamp_writes: None,
+        occlusion_query_set: None,
+    });
+
+    pass.set_pipeline(&pipeline);
+    pass.set_bind_group(0, &constant_bind_group, &[]);
+    pass.set_bind_group(1, &texture_bind_group, &[]);
+    pass.draw(0..6, 0..1);
+
+    texture
+}

crates/iced_wgpu/src/engine.rs

@@ -0,0 +1,78 @@
+use crate::graphics::{Antialiasing, Shell};
+use crate::primitive;
+use crate::quad;
+use crate::text;
+use crate::triangle;
+
+use std::sync::{Arc, RwLock};
+
+#[derive(Clone)]
+pub struct Engine {
+    pub(crate) device: wgpu::Device,
+    pub(crate) queue: wgpu::Queue,
+    pub(crate) format: wgpu::TextureFormat,
+
+    pub(crate) quad_pipeline: quad::Pipeline,
+    pub(crate) text_pipeline: text::Pipeline,
+    pub(crate) triangle_pipeline: triangle::Pipeline,
+    #[cfg(any(feature = "image", feature = "svg"))]
+    pub(crate) image_pipeline: crate::image::Pipeline,
+    pub(crate) primitive_storage: Arc<RwLock<primitive::Storage>>,
+    _shell: Shell,
+}
+
+impl Engine {
+    pub fn new(
+        _adapter: &wgpu::Adapter,
+        device: wgpu::Device,
+        queue: wgpu::Queue,
+        format: wgpu::TextureFormat,
+        antialiasing: Option<Antialiasing>, // TODO: Initialize AA pipelines lazily
+        shell: Shell,
+    ) -> Self {
+        Self {
+            format,
+
+            quad_pipeline: quad::Pipeline::new(&device, format),
+            text_pipeline: text::Pipeline::new(&device, &queue, format),
+            triangle_pipeline: triangle::Pipeline::new(
+                &device,
+                format,
+                antialiasing,
+            ),
+
+            #[cfg(any(feature = "image", feature = "svg"))]
+            image_pipeline: {
+                let backend = _adapter.get_info().backend;
+
+                crate::image::Pipeline::new(&device, format, backend)
+            },
+
+            primitive_storage: Arc::new(RwLock::new(
+                primitive::Storage::default(),
+            )),
+
+            device,
+            queue,
+            _shell: shell,
+        }
+    }
+
+    #[cfg(any(feature = "image", feature = "svg"))]
+    pub fn create_image_cache(&self) -> crate::image::Cache {
+        self.image_pipeline.create_cache(
+            &self.device,
+            &self.queue,
+            &self._shell,
+        )
+    }
+
+    pub fn trim(&mut self) {
+        self.text_pipeline.trim();
+
+        self.primitive_storage
+            .write()
+            .expect("primitive storage should be writable")
+            .trim();
+    }
+}

crates/iced_wgpu/src/geometry.rs

@@ -0,0 +1,785 @@
+//! Build and draw geometry.
+use crate::core::text::LineHeight;
+use crate::core::{
+    self, Pixels, Point, Radians, Rectangle, Size, Svg, Transformation, Vector,
+};
+use crate::graphics::cache::{self, Cached};
+use crate::graphics::color;
+use crate::graphics::geometry::fill::{self, Fill};
+use crate::graphics::geometry::{
+    self, LineCap, LineDash, LineJoin, Path, Stroke, Style,
+};
+use crate::graphics::gradient::{self, Gradient};
+use crate::graphics::mesh::{self, Mesh};
+use crate::graphics::{Image, Text};
+use crate::text;
+
+use lyon::geom::euclid;
+use lyon::tessellation;
+
+use std::borrow::Cow;
+use std::sync::Arc;
+
+#[derive(Debug)]
+pub enum Geometry {
+    Live {
+        meshes: Vec<Mesh>,
+        images: Vec<Image>,
+        text: Vec<Text>,
+    },
+    Cached(Cache),
+}
+
+#[derive(Debug, Clone, Default)]
+pub struct Cache {
+    pub meshes: Option<mesh::Cache>,
+    pub images: Option<Arc<[Image]>>,
+    pub text: Option<text::Cache>,
+}
+
+impl Cached for Geometry {
+    type Cache = Cache;
+
+    fn load(cache: &Self::Cache) -> Self {
+        Geometry::Cached(cache.clone())
+    }
+
+    fn cache(
+        self,
+        group: cache::Group,
+        previous: Option<Self::Cache>,
+    ) -> Self::Cache {
+        match self {
+            Self::Live {
+                meshes,
+                images,
+                text,
+            } => {
+                let images = if images.is_empty() {
+                    None
+                } else {
+                    Some(Arc::from(images))
+                };
+
+                let meshes = Arc::from(meshes);
+
+                if let Some(mut previous) = previous {
+                    if let Some(cache) = &mut previous.meshes {
+                        cache.update(meshes);
+                    } else {
+                        previous.meshes = if meshes.is_empty() {
+                            None
+                        } else {
+                            Some(mesh::Cache::new(meshes))
+                        };
+                    }
+
+                    if let Some(cache) = &mut previous.text {
+                        cache.update(text);
+                    } else {
+                        previous.text = text::Cache::new(group, text);
+                    }
+
+                    previous.images = images;
+
+                    previous
+                } else {
+                    Cache {
+                        meshes: if meshes.is_empty() {
+                            None
+                        } else {
+                            Some(mesh::Cache::new(meshes))
+                        },
+                        images,
+                        text: text::Cache::new(group, text),
+                    }
+                }
+            }
+            Self::Cached(cache) => cache,
+        }
+    }
+}
+
+/// A frame for drawing some geometry.
+pub struct Frame {
+    clip_bounds: Rectangle,
+    buffers: BufferStack,
+    meshes: Vec<Mesh>,
+    images: Vec<Image>,
+    text: Vec<Text>,
+    transforms: Transforms,
+    fill_tessellator: tessellation::FillTessellator,
+    stroke_tessellator: tessellation::StrokeTessellator,
+}
+
+impl Frame {
+    /// Creates a new [`Frame`] with the given clip bounds.
+    pub fn new(bounds: Rectangle) -> Frame {
+        Frame {
+            clip_bounds: bounds,
+            buffers: BufferStack::new(),
+            meshes: Vec::new(),
+            images: Vec::new(),
+            text: Vec::new(),
+            transforms: Transforms {
+                previous: Vec::new(),
+                current: Transform(lyon::math::Transform::identity()),
+            },
+            fill_tessellator: tessellation::FillTessellator::new(),
+            stroke_tessellator: tessellation::StrokeTessellator::new(),
+        }
+    }
+}
+
+impl geometry::frame::Backend for Frame {
+    type Geometry = Geometry;
+
+    #[inline]
+    fn width(&self) -> f32 {
+        self.clip_bounds.width
+    }
+
+    #[inline]
+    fn height(&self) -> f32 {
+        self.clip_bounds.height
+    }
+
+    #[inline]
+    fn size(&self) -> Size {
+        self.clip_bounds.size()
+    }
+
+    #[inline]
+    fn center(&self) -> Point {
+        Point::new(self.clip_bounds.width / 2.0, self.clip_bounds.height / 2.0)
+    }
+
+    fn fill(&mut self, path: &Path, fill: impl Into<Fill>) {
+        let Fill { style, rule } = fill.into();
+
+        let mut buffer = self
+            .buffers
+            .get_fill(&self.transforms.current.transform_style(style));
+
+        let options = tessellation::FillOptions::default()
+            .with_fill_rule(into_fill_rule(rule));
+
+        if self.transforms.current.is_identity() {
+            self.fill_tessellator.tessellate_path(
+                path.raw(),
+                &options,
+                buffer.as_mut(),
+            )
+        } else {
+            let path = path.transform(&self.transforms.current.0);
+
+            self.fill_tessellator.tessellate_path(
+                path.raw(),
+                &options,
+                buffer.as_mut(),
+            )
+        }
+        .expect("Tessellate path.");
+    }
+
+    fn fill_rectangle(
+        &mut self,
+        top_left: Point,
+        size: Size,
+        fill: impl Into<Fill>,
+    ) {
+        let Fill { style, rule } = fill.into();
+
+        let mut buffer = self
+            .buffers
+            .get_fill(&self.transforms.current.transform_style(style));
+
+        let top_left = self
+            .transforms
+            .current
+            .0
+            .transform_point(lyon::math::Point::new(top_left.x, top_left.y));
+
+        let size =
+            self.transforms.current.0.transform_vector(
+                lyon::math::Vector::new(size.width, size.height),
+            );
+
+        let options = tessellation::FillOptions::default()
+            .with_fill_rule(into_fill_rule(rule));
+
+        self.fill_tessellator
+            .tessellate_rectangle(
+                &lyon::math::Box2D::new(top_left, top_left + size),
+                &options,
+                buffer.as_mut(),
+            )
+            .expect("Fill rectangle");
+    }
+
+    fn stroke<'a>(&mut self, path: &Path, stroke: impl Into<Stroke<'a>>) {
+        let stroke = stroke.into();
+
+        let mut buffer = self
+            .buffers
+            .get_stroke(&self.transforms.current.transform_style(stroke.style));
+
+        let mut options = tessellation::StrokeOptions::default();
+        options.line_width = stroke.width;
+        options.start_cap = into_line_cap(stroke.line_cap);
+        options.end_cap = into_line_cap(stroke.line_cap);
+        options.line_join = into_line_join(stroke.line_join);
+
+        let path = if stroke.line_dash.segments.is_empty() {
+            Cow::Borrowed(path)
+        } else {
+            Cow::Owned(dashed(path, stroke.line_dash))
+        };
+
+        if self.transforms.current.is_identity() {
+            self.stroke_tessellator.tessellate_path(
+                path.raw(),
+                &options,
+                buffer.as_mut(),
+            )
+        } else {
+            let path = path.transform(&self.transforms.current.0);
+
+            self.stroke_tessellator.tessellate_path(
+                path.raw(),
+                &options,
+                buffer.as_mut(),
+            )
+        }
+        .expect("Stroke path");
+    }
+
+    fn stroke_rectangle<'a>(
+        &mut self,
+        top_left: Point,
+        size: Size,
+        stroke: impl Into<Stroke<'a>>,
+    ) {
+        let stroke = stroke.into();
+
+        let mut buffer = self
+            .buffers
+            .get_stroke(&self.transforms.current.transform_style(stroke.style));
+
+        let top_left = self
+            .transforms
+            .current
+            .0
+            .transform_point(lyon::math::Point::new(top_left.x, top_left.y));
+
+        let size =
+            self.transforms.current.0.transform_vector(
+                lyon::math::Vector::new(size.width, size.height),
+            );
+
+        let mut options = tessellation::StrokeOptions::default();
+        options.line_width = stroke.width;
+        options.start_cap = into_line_cap(stroke.line_cap);
+        options.end_cap = into_line_cap(stroke.line_cap);
+        options.line_join = into_line_join(stroke.line_join);
+
+        self.stroke_tessellator
+            .tessellate_rectangle(
+                &lyon::math::Box2D::new(top_left, top_left + size),
+                &options,
+                buffer.as_mut(),
+            )
+            .expect("Stroke rectangle");
+    }
+
+    fn stroke_text<'a>(
+        &mut self,
+        text: impl Into<geometry::Text>,
+        stroke: impl Into<Stroke<'a>>,
+    ) {
+        let text = text.into();
+        let stroke = stroke.into();
+
+        text.draw_with(|glyph, _color| self.stroke(&glyph, stroke));
+    }
+
+    fn fill_text(&mut self, text: impl Into<geometry::Text>) {
+        let text = text.into();
+
+        let (scale_x, scale_y) = self.transforms.current.scale();
+
+        if self.transforms.current.is_scale_translation()
+            && scale_x == scale_y
+            && scale_x > 0.0
+            && scale_y > 0.0
+        {
+            let (bounds, size, line_height) =
+                if self.transforms.current.is_identity() {
+                    (
+                        Rectangle::new(
+                            text.position,
+                            Size::new(text.max_width, f32::INFINITY),
+                        ),
+                        text.size,
+                        text.line_height,
+                    )
+                } else {
+                    let position =
+                        self.transforms.current.transform_point(text.position);
+
+                    let size = Pixels(text.size.0 * scale_y);
+
+                    let line_height = match text.line_height {
+                        LineHeight::Absolute(size) => {
+                            LineHeight::Absolute(Pixels(size.0 * scale_y))
+                        }
+                        LineHeight::Relative(factor) => {
+                            LineHeight::Relative(factor)
+                        }
+                    };
+
+                    (
+                        Rectangle::new(
+                            position,
+                            Size::new(text.max_width, f32::INFINITY),
+                        ),
+                        size,
+                        line_height,
+                    )
+                };
+
+            self.text.push(Text::Cached {
+                content: text.content,
+                bounds,
+                color: text.color,
+                size,
+                line_height: line_height.to_absolute(size),
+                font: text.font,
+                align_x: text.align_x,
+                align_y: text.align_y,
+                shaping: text.shaping,
+                clip_bounds: self.clip_bounds,
+            });
+        } else {
+            text.draw_with(|path, color| self.fill(&path, color));
+        }
+    }
+
+    #[inline]
+    fn translate(&mut self, translation: Vector) {
+        self.transforms.current.0 =
+            self.transforms
+                .current
+                .0
+                .pre_translate(lyon::math::Vector::new(
+                    translation.x,
+                    translation.y,
+                ));
+    }
+
+    #[inline]
+    fn rotate(&mut self, angle: impl Into<Radians>) {
+        self.transforms.current.0 = self
+            .transforms
+            .current
+            .0
+            .pre_rotate(lyon::math::Angle::radians(angle.into().0));
+    }
+
+    #[inline]
+    fn scale(&mut self, scale: impl Into<f32>) {
+        let scale = scale.into();
+
+        self.scale_nonuniform(Vector { x: scale, y: scale });
+    }
+
+    #[inline]
+    fn scale_nonuniform(&mut self, scale: impl Into<Vector>) {
+        let scale = scale.into();
+
+        self.transforms.current.0 =
+            self.transforms.current.0.pre_scale(scale.x, scale.y);
+    }
+
+    fn push_transform(&mut self) {
+        self.transforms.previous.push(self.transforms.current);
+    }
+
+    fn pop_transform(&mut self) {
+        self.transforms.current = self.transforms.previous.pop().unwrap();
+    }
+
+    fn draft(&mut self, clip_bounds: Rectangle) -> Frame {
+        Frame::new(clip_bounds)
+    }
+
+    fn paste(&mut self, frame: Frame) {
+        self.meshes.extend(frame.meshes);
+        self.meshes
+            .extend(frame.buffers.into_meshes(frame.clip_bounds));
+
+        self.images.extend(frame.images);
+        self.text.extend(frame.text);
+    }
+
+    fn into_geometry(mut self) -> Self::Geometry {
+        self.meshes
+            .extend(self.buffers.into_meshes(self.clip_bounds));
+
+        Geometry::Live {
+            meshes: self.meshes,
+            images: self.images,
+            text: self.text,
+        }
+    }
+
+    fn draw_image(&mut self, bounds: Rectangle, image: impl Into<core::Image>) {
+        let mut image = image.into();
+
+        let (bounds, external_rotation) =
+            self.transforms.current.transform_rectangle(bounds);
+
+        image.rotation += external_rotation;
+        image.border_radius =
+            image.border_radius * self.transforms.current.scale().0;
+
+        self.images.push(Image::Raster {
+            image,
+            bounds,
+            clip_bounds: self.clip_bounds,
+        });
+    }
+
+    fn draw_svg(&mut self, bounds: Rectangle, svg: impl Into<Svg>) {
+        let mut svg = svg.into();
+
+        let (bounds, external_rotation) =
+            self.transforms.current.transform_rectangle(bounds);
+
+        svg.rotation += external_rotation;
+
+        self.images.push(Image::Vector {
+            svg,
+            bounds,
+            clip_bounds: self.clip_bounds,
+        });
+    }
+}
+
+enum Buffer {
+    Solid(tessellation::VertexBuffers<mesh::SolidVertex2D, u32>),
+    Gradient(tessellation::VertexBuffers<mesh::GradientVertex2D, u32>),
+}
+
+struct BufferStack {
+    stack: Vec<Buffer>,
+}
+
+impl BufferStack {
+    fn new() -> Self {
+        Self { stack: Vec::new() }
+    }
+
+    fn get_mut(&mut self, style: &Style) -> &mut Buffer {
+        match style {
+            Style::Solid(_) => match self.stack.last() {
+                Some(Buffer::Solid(_)) => {}
+                _ => {
+                    self.stack.push(Buffer::Solid(
+                        tessellation::VertexBuffers::new(),
+                    ));
+                }
+            },
+            Style::Gradient(_) => match self.stack.last() {
+                Some(Buffer::Gradient(_)) => {}
+                _ => {
+                    self.stack.push(Buffer::Gradient(
+                        tessellation::VertexBuffers::new(),
+                    ));
+                }
+            },
+        }
+
+        self.stack.last_mut().unwrap()
+    }
+
+    fn get_fill<'a>(
+        &'a mut self,
+        style: &Style,
+    ) -> Box<dyn tessellation::FillGeometryBuilder + 'a> {
+        match (style, self.get_mut(style)) {
+            (Style::Solid(color), Buffer::Solid(buffer)) => {
+                Box::new(tessellation::BuffersBuilder::new(
+                    buffer,
+                    TriangleVertex2DBuilder(color::pack(*color)),
+                ))
+            }
+            (Style::Gradient(gradient), Buffer::Gradient(buffer)) => {
+                Box::new(tessellation::BuffersBuilder::new(
+                    buffer,
+                    GradientVertex2DBuilder {
+                        gradient: gradient.pack(),
+                    },
+                ))
+            }
+            _ => unreachable!(),
+        }
+    }
+
+    fn get_stroke<'a>(
+        &'a mut self,
+        style: &Style,
+    ) -> Box<dyn tessellation::StrokeGeometryBuilder + 'a> {
+        match (style, self.get_mut(style)) {
+            (Style::Solid(color), Buffer::Solid(buffer)) => {
+                Box::new(tessellation::BuffersBuilder::new(
+                    buffer,
+                    TriangleVertex2DBuilder(color::pack(*color)),
+                ))
+            }
+            (Style::Gradient(gradient), Buffer::Gradient(buffer)) => {
+                Box::new(tessellation::BuffersBuilder::new(
+                    buffer,
+                    GradientVertex2DBuilder {
+                        gradient: gradient.pack(),
+                    },
+                ))
+            }
+            _ => unreachable!(),
+        }
+    }
+
+    fn into_meshes(self, clip_bounds: Rectangle) -> impl Iterator<Item = Mesh> {
+        self.stack
+            .into_iter()
+            .filter_map(move |buffer| match buffer {
+                Buffer::Solid(buffer) if !buffer.indices.is_empty() => {
+                    Some(Mesh::Solid {
+                        buffers: mesh::Indexed {
+                            vertices: buffer.vertices,
+                            indices: buffer.indices,
+                        },
+                        clip_bounds,
+                        transformation: Transformation::IDENTITY,
+                    })
+                }
+                Buffer::Gradient(buffer) if !buffer.indices.is_empty() => {
+                    Some(Mesh::Gradient {
+                        buffers: mesh::Indexed {
+                            vertices: buffer.vertices,
+                            indices: buffer.indices,
+                        },
+                        clip_bounds,
+                        transformation: Transformation::IDENTITY,
+                    })
+                }
+                _ => None,
+            })
+    }
+}
+
+#[derive(Debug)]
+struct Transforms {
+    previous: Vec<Transform>,
+    current: Transform,
+}
+
+#[derive(Debug, Clone, Copy)]
+struct Transform(lyon::math::Transform);
+
+impl Transform {
+    fn is_identity(&self) -> bool {
+        self.0 == lyon::math::Transform::identity()
+    }
+
+    fn is_scale_translation(&self) -> bool {
+        self.0.m12.abs() < 2.0 * f32::EPSILON
+            && self.0.m21.abs() < 2.0 * f32::EPSILON
+    }
+
+    fn scale(&self) -> (f32, f32) {
+        (self.0.m11, self.0.m22)
+    }
+
+    fn transform_point(&self, point: Point) -> Point {
+        let transformed = self
+            .0
+            .transform_point(euclid::Point2D::new(point.x, point.y));
+
+        Point {
+            x: transformed.x,
+            y: transformed.y,
+        }
+    }
+
+    fn transform_style(&self, style: Style) -> Style {
+        match style {
+            Style::Solid(color) => Style::Solid(color),
+            Style::Gradient(gradient) => {
+                Style::Gradient(self.transform_gradient(gradient))
+            }
+        }
+    }
+
+    fn transform_gradient(&self, mut gradient: Gradient) -> Gradient {
+        match &mut gradient {
+            Gradient::Linear(linear) => {
+                linear.start = self.transform_point(linear.start);
+                linear.end = self.transform_point(linear.end);
+            }
+        }
+
+        gradient
+    }
+
+    fn transform_rectangle(
+        &self,
+        rectangle: Rectangle,
+    ) -> (Rectangle, Radians) {
+        let top_left = self.transform_point(rectangle.position());
+        let top_right = self.transform_point(
+            rectangle.position() + Vector::new(rectangle.width, 0.0),
+        );
+        let bottom_left = self.transform_point(
+            rectangle.position() + Vector::new(0.0, rectangle.height),
+        );
+
+        Rectangle::with_vertices(top_left, top_right, bottom_left)
+    }
+}
+struct GradientVertex2DBuilder {
+    gradient: gradient::Packed,
+}
+
+impl tessellation::FillVertexConstructor<mesh::GradientVertex2D>
+    for GradientVertex2DBuilder
+{
+    fn new_vertex(
+        &mut self,
+        vertex: tessellation::FillVertex<'_>,
+    ) -> mesh::GradientVertex2D {
+        let position = vertex.position();
+
+        mesh::GradientVertex2D {
+            position: [position.x, position.y],
+            gradient: self.gradient,
+        }
+    }
+}
+
+impl tessellation::StrokeVertexConstructor<mesh::GradientVertex2D>
+    for GradientVertex2DBuilder
+{
+    fn new_vertex(
+        &mut self,
+        vertex: tessellation::StrokeVertex<'_, '_>,
+    ) -> mesh::GradientVertex2D {
+        let position = vertex.position();
+
+        mesh::GradientVertex2D {
+            position: [position.x, position.y],
+            gradient: self.gradient,
+        }
+    }
+}
+
+struct TriangleVertex2DBuilder(color::Packed);
+
+impl tessellation::FillVertexConstructor<mesh::SolidVertex2D>
+    for TriangleVertex2DBuilder
+{
+    fn new_vertex(
+        &mut self,
+        vertex: tessellation::FillVertex<'_>,
+    ) -> mesh::SolidVertex2D {
+        let position = vertex.position();
+
+        mesh::SolidVertex2D {
+            position: [position.x, position.y],
+            color: self.0,
+        }
+    }
+}
+
+impl tessellation::StrokeVertexConstructor<mesh::SolidVertex2D>
+    for TriangleVertex2DBuilder
+{
+    fn new_vertex(
+        &mut self,
+        vertex: tessellation::StrokeVertex<'_, '_>,
+    ) -> mesh::SolidVertex2D {
+        let position = vertex.position();
+
+        mesh::SolidVertex2D {
+            position: [position.x, position.y],
+            color: self.0,
+        }
+    }
+}
+
+fn into_line_join(line_join: LineJoin) -> lyon::tessellation::LineJoin {
+    match line_join {
+        LineJoin::Miter => lyon::tessellation::LineJoin::Miter,
+        LineJoin::Round => lyon::tessellation::LineJoin::Round,
+        LineJoin::Bevel => lyon::tessellation::LineJoin::Bevel,
+    }
+}
+
+fn into_line_cap(line_cap: LineCap) -> lyon::tessellation::LineCap {
+    match line_cap {
+        LineCap::Butt => lyon::tessellation::LineCap::Butt,
+        LineCap::Square => lyon::tessellation::LineCap::Square,
+        LineCap::Round => lyon::tessellation::LineCap::Round,
+    }
+}
+
+fn into_fill_rule(rule: fill::Rule) -> lyon::tessellation::FillRule {
+    match rule {
+        fill::Rule::NonZero => lyon::tessellation::FillRule::NonZero,
+        fill::Rule::EvenOdd => lyon::tessellation::FillRule::EvenOdd,
+    }
+}
+
+pub(super) fn dashed(path: &Path, line_dash: LineDash<'_>) -> Path {
+    use lyon::algorithms::walk::{
+        RepeatedPattern, WalkerEvent, walk_along_path,
+    };
+    use lyon::path::iterator::PathIterator;
+
+    Path::new(|builder| {
+        let segments_odd = (line_dash.segments.len() % 2 == 1)
+            .then(|| [line_dash.segments, line_dash.segments].concat());
+
+        let mut draw_line = false;
+
+        walk_along_path(
+            path.raw().iter().flattened(
+                lyon::tessellation::StrokeOptions::DEFAULT_TOLERANCE,
+            ),
+            0.0,
+            lyon::tessellation::StrokeOptions::DEFAULT_TOLERANCE,
+            &mut RepeatedPattern {
+                callback: |event: WalkerEvent<'_>| {
+                    let point = Point {
+                        x: event.position.x,
+                        y: event.position.y,
+                    };
+
+                    if draw_line {
+                        builder.line_to(point);
+                    } else {
+                        builder.move_to(point);
+                    }
+
+                    draw_line = !draw_line;
+
+                    true
+                },
+                index: line_dash.offset,
+                intervals: segments_odd
+                    .as_deref()
+                    .unwrap_or(line_dash.segments),
+            },
+        );
+    })
+}

crates/iced_wgpu/src/image/atlas.rs

@@ -0,0 +1,542 @@
+pub mod entry;
+
+mod allocation;
+mod allocator;
+mod layer;
+
+pub use allocation::Allocation;
+pub use entry::Entry;
+pub use layer::Layer;
+
+use allocator::Allocator;
+
+pub const DEFAULT_SIZE: u32 = 2048;
+pub const MAX_SIZE: u32 = 2048;
+
+use crate::core::Size;
+use crate::graphics::color;
+
+use std::sync::Arc;
+
+#[derive(Debug)]
+pub struct Atlas {
+    size: u32,
+    backend: wgpu::Backend,
+    texture: wgpu::Texture,
+    texture_view: wgpu::TextureView,
+    texture_bind_group: Arc<wgpu::BindGroup>,
+    texture_layout: wgpu::BindGroupLayout,
+    layers: Vec<Layer>,
+}
+
+impl Atlas {
+    pub fn new(
+        device: &wgpu::Device,
+        backend: wgpu::Backend,
+        texture_layout: wgpu::BindGroupLayout,
+    ) -> Self {
+        Self::with_size(device, backend, texture_layout, DEFAULT_SIZE)
+    }
+
+    pub fn with_size(
+        device: &wgpu::Device,
+        backend: wgpu::Backend,
+        texture_layout: wgpu::BindGroupLayout,
+        size: u32,
+    ) -> Self {
+        let size = size.min(MAX_SIZE);
+
+        let layers = match backend {
+            // On the GL backend we start with 2 layers, to help wgpu figure
+            // out that this texture is `GL_TEXTURE_2D_ARRAY` rather than `GL_TEXTURE_2D`
+            // https://github.com/gfx-rs/wgpu/blob/004e3efe84a320d9331371ed31fa50baa2414911/wgpu-hal/src/gles/mod.rs#L371
+            wgpu::Backend::Gl => vec![Layer::Empty, Layer::Empty],
+            _ => vec![Layer::Empty],
+        };
+
+        let extent = wgpu::Extent3d {
+            width: size,
+            height: size,
+            depth_or_array_layers: layers.len() as u32,
+        };
+
+        let texture = device.create_texture(&wgpu::TextureDescriptor {
+            label: Some("iced_wgpu::image texture atlas"),
+            size: extent,
+            mip_level_count: 1,
+            sample_count: 1,
+            dimension: wgpu::TextureDimension::D2,
+            format: if color::GAMMA_CORRECTION {
+                wgpu::TextureFormat::Rgba8UnormSrgb
+            } else {
+                wgpu::TextureFormat::Rgba8Unorm
+            },
+            usage: wgpu::TextureUsages::COPY_DST
+                | wgpu::TextureUsages::COPY_SRC
+                | wgpu::TextureUsages::TEXTURE_BINDING,
+            view_formats: &[],
+        });
+
+        let texture_view = texture.create_view(&wgpu::TextureViewDescriptor {
+            dimension: Some(wgpu::TextureViewDimension::D2Array),
+            ..Default::default()
+        });
+
+        let texture_bind_group =
+            device.create_bind_group(&wgpu::BindGroupDescriptor {
+                label: Some("iced_wgpu::image texture atlas bind group"),
+                layout: &texture_layout,
+                entries: &[wgpu::BindGroupEntry {
+                    binding: 0,
+                    resource: wgpu::BindingResource::TextureView(&texture_view),
+                }],
+            });
+
+        Atlas {
+            size,
+            backend,
+            texture,
+            texture_view,
+            texture_bind_group: Arc::new(texture_bind_group),
+            texture_layout,
+            layers,
+        }
+    }
+
+    pub fn bind_group(&self) -> &Arc<wgpu::BindGroup> {
+        &self.texture_bind_group
+    }
+
+    pub fn upload(
+        &mut self,
+        device: &wgpu::Device,
+        encoder: &mut wgpu::CommandEncoder,
+        belt: &mut wgpu::util::StagingBelt,
+        width: u32,
+        height: u32,
+        pixels: &[u8],
+    ) -> Option<Entry> {
+        let entry = {
+            let current_size = self.layers.len();
+            let entry = self.allocate(width, height)?;
+
+            // We grow the internal texture after allocating if necessary
+            let new_layers = self.layers.len() - current_size;
+            self.grow(new_layers, device, encoder, self.backend);
+
+            entry
+        };
+
+        log::debug!("Allocated atlas entry: {entry:?}");
+
+        match &entry {
+            Entry::Contiguous(allocation) => {
+                self.upload_allocation(
+                    pixels, width, 0, allocation, device, encoder, belt,
+                );
+            }
+            Entry::Fragmented { fragments, .. } => {
+                for fragment in fragments {
+                    let (x, y) = fragment.position;
+                    let offset = 4 * (y * width + x) as usize;
+
+                    self.upload_allocation(
+                        pixels,
+                        width,
+                        offset,
+                        &fragment.allocation,
+                        device,
+                        encoder,
+                        belt,
+                    );
+                }
+            }
+        }
+
+        if log::log_enabled!(log::Level::Debug) {
+            log::debug!(
+                "Atlas layers: {} (busy: {}, allocations: {})",
+                self.layers.len(),
+                self.layers.iter().filter(|layer| !layer.is_empty()).count(),
+                self.layers.iter().map(Layer::allocations).sum::<usize>(),
+            );
+        }
+
+        Some(entry)
+    }
+
+    pub fn remove(&mut self, entry: &Entry) {
+        log::debug!("Removing atlas entry: {entry:?}");
+
+        match entry {
+            Entry::Contiguous(allocation) => {
+                self.deallocate(allocation);
+            }
+            Entry::Fragmented { fragments, .. } => {
+                for fragment in fragments {
+                    self.deallocate(&fragment.allocation);
+                }
+            }
+        }
+    }
+
+    fn allocate(&mut self, width: u32, height: u32) -> Option<Entry> {
+        // Allocate one layer if texture fits perfectly
+        if width == self.size && height == self.size {
+            let mut empty_layers = self
+                .layers
+                .iter_mut()
+                .enumerate()
+                .filter(|(_, layer)| layer.is_empty());
+
+            if let Some((i, layer)) = empty_layers.next() {
+                *layer = Layer::Full;
+
+                return Some(Entry::Contiguous(Allocation::Full {
+                    layer: i,
+                    size: self.size,
+                }));
+            }
+
+            self.layers.push(Layer::Full);
+
+            return Some(Entry::Contiguous(Allocation::Full {
+                layer: self.layers.len() - 1,
+                size: self.size,
+            }));
+        }
+
+        // Split big textures across multiple layers
+        if width > self.size || height > self.size {
+            let mut fragments = Vec::new();
+            let mut y = 0;
+
+            while y < height {
+                let height = std::cmp::min(height - y, self.size);
+                let mut x = 0;
+
+                while x < width {
+                    let width = std::cmp::min(width - x, self.size);
+
+                    let allocation = self.allocate(width, height)?;
+
+                    if let Entry::Contiguous(allocation) = allocation {
+                        fragments.push(entry::Fragment {
+                            position: (x, y),
+                            allocation,
+                        });
+                    }
+
+                    x += width;
+                }
+
+                y += height;
+            }
+
+            return Some(Entry::Fragmented {
+                size: Size::new(width, height),
+                fragments,
+            });
+        }
+
+        // Try allocating on an existing layer
+        for (i, layer) in self.layers.iter_mut().enumerate() {
+            match layer {
+                Layer::Empty => {
+                    let mut allocator = Allocator::new(self.size);
+
+                    if let Some(region) = allocator.allocate(width, height) {
+                        *layer = Layer::Busy(allocator);
+
+                        return Some(Entry::Contiguous(Allocation::Partial {
+                            region,
+                            layer: i,
+                            atlas_size: self.size,
+                        }));
+                    }
+                }
+                Layer::Busy(allocator) => {
+                    if let Some(region) = allocator.allocate(width, height) {
+                        return Some(Entry::Contiguous(Allocation::Partial {
+                            region,
+                            layer: i,
+                            atlas_size: self.size,
+                        }));
+                    }
+                }
+                Layer::Full => {}
+            }
+        }
+
+        // Create new layer with atlas allocator
+        let mut allocator = Allocator::new(self.size);
+
+        if let Some(region) = allocator.allocate(width, height) {
+            self.layers.push(Layer::Busy(allocator));
+
+            return Some(Entry::Contiguous(Allocation::Partial {
+                region,
+                layer: self.layers.len() - 1,
+                atlas_size: self.size,
+            }));
+        }
+
+        // We ran out of memory (?)
+        None
+    }
+
+    fn deallocate(&mut self, allocation: &Allocation) {
+        log::debug!("Deallocating atlas: {allocation:?}");
+
+        match allocation {
+            Allocation::Full { layer, .. } => {
+                self.layers[*layer] = Layer::Empty;
+            }
+            Allocation::Partial { layer, region, .. } => {
+                let layer = &mut self.layers[*layer];
+
+                if let Layer::Busy(allocator) = layer {
+                    allocator.deallocate(region);
+
+                    if allocator.is_empty() {
+                        *layer = Layer::Empty;
+                    }
+                }
+            }
+        }
+    }
+
+    fn upload_allocation(
+        &self,
+        pixels: &[u8],
+        image_width: u32,
+        offset: usize,
+        allocation: &Allocation,
+        device: &wgpu::Device,
+        encoder: &mut wgpu::CommandEncoder,
+        belt: &mut wgpu::util::StagingBelt,
+    ) {
+        let (x, y) = allocation.position();
+        let Size { width, height } = allocation.size();
+        let layer = allocation.layer();
+        let padding = allocation.padding();
+
+        // It is a webgpu requirement that:
+        //   BufferCopyView.layout.bytes_per_row % wgpu::COPY_BYTES_PER_ROW_ALIGNMENT == 0
+        // So we calculate bytes_per_row by rounding width up to the next
+        // multiple of wgpu::COPY_BYTES_PER_ROW_ALIGNMENT.
+        let bytes_per_row = (4 * (width + padding.width * 2))
+            .next_multiple_of(wgpu::COPY_BYTES_PER_ROW_ALIGNMENT)
+            as usize;
+        let total_bytes =
+            bytes_per_row * (height + padding.height * 2) as usize;
+
+        let buffer_slice = belt.allocate(
+            wgpu::BufferSize::new(total_bytes as u64).unwrap(),
+            wgpu::BufferSize::new(8 * 4).unwrap(),
+            device,
+        );
+
+        const PIXEL: usize = 4;
+
+        let mut fragment = buffer_slice.get_mapped_range_mut();
+        let w = width as usize;
+        let h = height as usize;
+        let pad_w = padding.width as usize;
+        let pad_h = padding.height as usize;
+        let stride = PIXEL * w;
+
+        // Copy image rows
+        for row in 0..h {
+            let src = offset + row * PIXEL * image_width as usize;
+            let dst = (row + pad_h) * bytes_per_row;
+
+            fragment[dst + PIXEL * pad_w..dst + PIXEL * pad_w + stride]
+                .copy_from_slice(&pixels[src..src + stride]);
+
+            // Add padding to the sides, if needed
+            for i in 0..pad_w {
+                fragment[dst + PIXEL * i..dst + PIXEL * (i + 1)]
+                    .copy_from_slice(&pixels[src..src + PIXEL]);
+
+                fragment[dst + stride + PIXEL * (pad_w + i)
+                    ..dst + stride + PIXEL * (pad_w + i + 1)]
+                    .copy_from_slice(
+                        &pixels[src + stride - PIXEL..src + stride],
+                    );
+            }
+        }
+
+        // Add padding on top and bottom
+        for row in 0..pad_h {
+            let dst_top = row * bytes_per_row;
+            let dst_bottom = (pad_h + h + row) * bytes_per_row;
+            let src_top = offset;
+            let src_bottom = offset + (h - 1) * PIXEL * image_width as usize;
+
+            // Top
+            fragment[dst_top + PIXEL * pad_w..dst_top + PIXEL * (pad_w + w)]
+                .copy_from_slice(&pixels[src_top..src_top + PIXEL * w]);
+
+            // Bottom
+            fragment
+                [dst_bottom + PIXEL * pad_w..dst_bottom + PIXEL * (pad_w + w)]
+                .copy_from_slice(&pixels[src_bottom..src_bottom + PIXEL * w]);
+
+            // Corners
+            for i in 0..pad_w {
+                // Top left
+                fragment[dst_top + PIXEL * i..dst_top + PIXEL * (i + 1)]
+                    .copy_from_slice(&pixels[offset..offset + PIXEL]);
+
+                // Top right
+                fragment[dst_top + PIXEL * (w + pad_w + i)
+                    ..dst_top + PIXEL * (w + pad_w + i + 1)]
+                    .copy_from_slice(
+                        &pixels[offset + PIXEL * (w - 1)..offset + PIXEL * w],
+                    );
+
+                // Bottom left
+                fragment[dst_bottom + PIXEL * i..dst_bottom + PIXEL * (i + 1)]
+                    .copy_from_slice(&pixels[src_bottom..src_bottom + PIXEL]);
+
+                // Bottom right
+                fragment[dst_bottom + PIXEL * (w + pad_w + i)
+                    ..dst_bottom + PIXEL * (w + pad_w + i + 1)]
+                    .copy_from_slice(
+                        &pixels[src_bottom + PIXEL * (w - 1)
+                            ..src_bottom + PIXEL * w],
+                    );
+            }
+        }
+
+        // Copy actual image
+        encoder.copy_buffer_to_texture(
+            wgpu::TexelCopyBufferInfo {
+                buffer: buffer_slice.buffer(),
+                layout: wgpu::TexelCopyBufferLayout {
+                    offset: buffer_slice.offset(),
+                    bytes_per_row: Some(bytes_per_row as u32),
+                    rows_per_image: Some(height + padding.height * 2),
+                },
+            },
+            wgpu::TexelCopyTextureInfo {
+                texture: &self.texture,
+                mip_level: 0,
+                origin: wgpu::Origin3d {
+                    x: x - padding.width,
+                    y: y - padding.height,
+                    z: layer as u32,
+                },
+                aspect: wgpu::TextureAspect::default(),
+            },
+            wgpu::Extent3d {
+                width: width + padding.width * 2,
+                height: height + padding.height * 2,
+                depth_or_array_layers: 1,
+            },
+        );
+    }
+
+    fn grow(
+        &mut self,
+        amount: usize,
+        device: &wgpu::Device,
+        encoder: &mut wgpu::CommandEncoder,
+        backend: wgpu::Backend,
+    ) {
+        if amount == 0 {
+            return;
+        }
+
+        // On the GL backend if layers.len() is a multiple of 6 we need to help wgpu figure out that this texture
+        // is still a `GL_TEXTURE_2D_ARRAY` rather than `GL_TEXTURE_CUBE_MAP` or `GL_TEXTURE_CUBE_ARRAY`.
+        // This will over-allocate some unused memory on GL, but it's better than not being able to
+        // grow the atlas past multiples of 6!
+        // https://github.com/gfx-rs/wgpu/blob/004e3efe84a320d9331371ed31fa50baa2414911/wgpu-hal/src/gles/mod.rs#L371
+        let depth_or_array_layers = match backend {
+            wgpu::Backend::Gl if self.layers.len().is_multiple_of(6) => {
+                self.layers.len() as u32 + 1
+            }
+            _ => self.layers.len() as u32,
+        };
+
+        let new_texture = device.create_texture(&wgpu::TextureDescriptor {
+            label: Some("iced_wgpu::image texture atlas"),
+            size: wgpu::Extent3d {
+                width: self.size,
+                height: self.size,
+                depth_or_array_layers,
+            },
+            mip_level_count: 1,
+            sample_count: 1,
+            dimension: wgpu::TextureDimension::D2,
+            format: if color::GAMMA_CORRECTION {
+                wgpu::TextureFormat::Rgba8UnormSrgb
+            } else {
+                wgpu::TextureFormat::Rgba8Unorm
+            },
+            usage: wgpu::TextureUsages::COPY_DST
+                | wgpu::TextureUsages::COPY_SRC
+                | wgpu::TextureUsages::TEXTURE_BINDING,
+            view_formats: &[],
+        });
+
+        let amount_to_copy = self.layers.len() - amount;
+
+        for (i, layer) in
+            self.layers.iter_mut().take(amount_to_copy).enumerate()
+        {
+            if layer.is_empty() {
+                continue;
+            }
+
+            encoder.copy_texture_to_texture(
+                wgpu::TexelCopyTextureInfo {
+                    texture: &self.texture,
+                    mip_level: 0,
+                    origin: wgpu::Origin3d {
+                        x: 0,
+                        y: 0,
+                        z: i as u32,
+                    },
+                    aspect: wgpu::TextureAspect::default(),
+                },
+                wgpu::TexelCopyTextureInfo {
+                    texture: &new_texture,
+                    mip_level: 0,
+                    origin: wgpu::Origin3d {
+                        x: 0,
+                        y: 0,
+                        z: i as u32,
+                    },
+                    aspect: wgpu::TextureAspect::default(),
+                },
+                wgpu::Extent3d {
+                    width: self.size,
+                    height: self.size,
+                    depth_or_array_layers: 1,
+                },
+            );
+        }
+
+        self.texture = new_texture;
+        self.texture_view =
+            self.texture.create_view(&wgpu::TextureViewDescriptor {
+                dimension: Some(wgpu::TextureViewDimension::D2Array),
+                ..Default::default()
+            });
+
+        self.texture_bind_group =
+            Arc::new(device.create_bind_group(&wgpu::BindGroupDescriptor {
+                label: Some("iced_wgpu::image texture atlas bind group"),
+                layout: &self.texture_layout,
+                entries: &[wgpu::BindGroupEntry {
+                    binding: 0,
+                    resource: wgpu::BindingResource::TextureView(
+                        &self.texture_view,
+                    ),
+                }],
+            }));
+    }
+}

crates/iced_wgpu/src/image/atlas/allocation.rs

@@ -0,0 +1,52 @@
+use crate::core::Size;
+use crate::image::atlas::allocator;
+
+#[derive(Debug)]
+pub enum Allocation {
+    Partial {
+        layer: usize,
+        region: allocator::Region,
+        atlas_size: u32,
+    },
+    Full {
+        layer: usize,
+        size: u32,
+    },
+}
+
+impl Allocation {
+    pub fn position(&self) -> (u32, u32) {
+        match self {
+            Allocation::Partial { region, .. } => region.position(),
+            Allocation::Full { .. } => (0, 0),
+        }
+    }
+
+    pub fn size(&self) -> Size<u32> {
+        match self {
+            Allocation::Partial { region, .. } => region.size(),
+            Allocation::Full { size, .. } => Size::new(*size, *size),
+        }
+    }
+
+    pub fn padding(&self) -> Size<u32> {
+        match self {
+            Allocation::Partial { region, .. } => region.padding(),
+            Allocation::Full { .. } => Size::new(0, 0),
+        }
+    }
+
+    pub fn layer(&self) -> usize {
+        match self {
+            Allocation::Partial { layer, .. } => *layer,
+            Allocation::Full { layer, .. } => *layer,
+        }
+    }
+
+    pub fn atlas_size(&self) -> u32 {
+        match self {
+            Allocation::Partial { atlas_size, .. } => *atlas_size,
+            Allocation::Full { size, .. } => *size,
+        }
+    }
+}

crates/iced_wgpu/src/image/atlas/allocator.rs

@@ -0,0 +1,108 @@
+use crate::core;
+
+use guillotiere::{AtlasAllocator, Size};
+
+pub struct Allocator {
+    raw: AtlasAllocator,
+    allocations: usize,
+}
+
+impl Allocator {
+    const PADDING: u32 = 1;
+
+    pub fn new(size: u32) -> Allocator {
+        let raw = AtlasAllocator::new(Size::new(size as i32, size as i32));
+
+        Allocator {
+            raw,
+            allocations: 0,
+        }
+    }
+
+    pub fn allocate(&mut self, width: u32, height: u32) -> Option<Region> {
+        let size = self.raw.size();
+
+        let padded_width = width + Self::PADDING * 2;
+        let padded_height = height + Self::PADDING * 2;
+
+        let pad_width = padded_width as i32 <= size.width;
+        let pad_height = padded_height as i32 <= size.height;
+
+        let mut allocation = self.raw.allocate(Size::new(
+            if pad_width { padded_width } else { width } as i32,
+            if pad_height { padded_height } else { height } as i32,
+        ))?;
+
+        if pad_width {
+            allocation.rectangle.min.x += Self::PADDING as i32;
+            allocation.rectangle.max.x -= Self::PADDING as i32;
+        }
+
+        if pad_height {
+            allocation.rectangle.min.y += Self::PADDING as i32;
+            allocation.rectangle.max.y -= Self::PADDING as i32;
+        }
+
+        self.allocations += 1;
+
+        Some(Region {
+            allocation,
+            padding: core::Size::new(
+                if pad_width { Self::PADDING } else { 0 },
+                if pad_height { Self::PADDING } else { 0 },
+            ),
+        })
+    }
+
+    pub fn deallocate(&mut self, region: &Region) {
+        self.raw.deallocate(region.allocation.id);
+
+        self.allocations = self.allocations.saturating_sub(1);
+    }
+
+    pub fn is_empty(&self) -> bool {
+        self.allocations == 0
+    }
+
+    pub fn allocations(&self) -> usize {
+        self.allocations
+    }
+}
+
+pub struct Region {
+    allocation: guillotiere::Allocation,
+    padding: core::Size<u32>,
+}
+
+impl Region {
+    pub fn position(&self) -> (u32, u32) {
+        let rectangle = &self.allocation.rectangle;
+
+        (rectangle.min.x as u32, rectangle.min.y as u32)
+    }
+
+    pub fn size(&self) -> core::Size<u32> {
+        let size = self.allocation.rectangle.size();
+
+        core::Size::new(size.width as u32, size.height as u32)
+    }
+
+    pub fn padding(&self) -> crate::core::Size<u32> {
+        self.padding
+    }
+}
+
+impl std::fmt::Debug for Allocator {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        write!(f, "Allocator")
+    }
+}
+
+impl std::fmt::Debug for Region {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("Region")
+            .field("id", &self.allocation.id)
+            .field("rectangle", &self.allocation.rectangle)
+            .finish()
+    }
+}

crates/iced_wgpu/src/image/atlas/entry.rs

@@ -0,0 +1,27 @@
+use crate::core::Size;
+use crate::image::atlas;
+
+#[derive(Debug)]
+pub enum Entry {
+    Contiguous(atlas::Allocation),
+    Fragmented {
+        size: Size<u32>,
+        fragments: Vec<Fragment>,
+    },
+}
+
+impl Entry {
+    #[cfg(feature = "image")]
+    pub fn size(&self) -> Size<u32> {
+        match self {
+            Entry::Contiguous(allocation) => allocation.size(),
+            Entry::Fragmented { size, .. } => *size,
+        }
+    }
+}
+
+#[derive(Debug)]
+pub struct Fragment {
+    pub position: (u32, u32),
+    pub allocation: atlas::Allocation,
+}

crates/iced_wgpu/src/image/atlas/layer.rs

@@ -0,0 +1,22 @@
+use crate::image::atlas::Allocator;
+
+#[derive(Debug)]
+pub enum Layer {
+    Empty,
+    Busy(Allocator),
+    Full,
+}
+
+impl Layer {
+    pub fn is_empty(&self) -> bool {
+        matches!(self, Layer::Empty)
+    }
+
+    pub fn allocations(&self) -> usize {
+        match self {
+            Layer::Empty => 0,
+            Layer::Busy(allocator) => allocator.allocations(),
+            Layer::Full => 1,
+        }
+    }
+}

crates/iced_wgpu/src/image/cache.rs

@@ -0,0 +1,636 @@
+use crate::core::{self, Size};
+use crate::graphics::Shell;
+use crate::image::atlas::{self, Atlas};
+
+#[cfg(all(feature = "image", not(target_arch = "wasm32")))]
+use worker::Worker;
+
+#[cfg(feature = "image")]
+use std::collections::HashMap;
+
+use std::sync::Arc;
+
+pub struct Cache {
+    atlas: Atlas,
+    #[cfg(feature = "image")]
+    raster: Raster,
+    #[cfg(feature = "svg")]
+    vector: crate::image::vector::Cache,
+    #[cfg(all(feature = "image", not(target_arch = "wasm32")))]
+    worker: Worker,
+}
+
+impl Cache {
+    pub fn new(
+        device: &wgpu::Device,
+        _queue: &wgpu::Queue,
+        backend: wgpu::Backend,
+        layout: wgpu::BindGroupLayout,
+        _shell: &Shell,
+    ) -> Self {
+        #[cfg(all(feature = "image", not(target_arch = "wasm32")))]
+        let worker =
+            Worker::new(device, _queue, backend, layout.clone(), _shell);
+
+        Self {
+            atlas: Atlas::new(device, backend, layout),
+            #[cfg(feature = "image")]
+            raster: Raster {
+                cache: crate::image::raster::Cache::default(),
+                pending: HashMap::new(),
+                belt: wgpu::util::StagingBelt::new(2 * 1024 * 1024),
+            },
+            #[cfg(feature = "svg")]
+            vector: crate::image::vector::Cache::default(),
+            #[cfg(all(feature = "image", not(target_arch = "wasm32")))]
+            worker,
+        }
+    }
+
+    #[cfg(feature = "image")]
+    pub fn allocate_image(
+        &mut self,
+        handle: &core::image::Handle,
+        callback: impl FnOnce(Result<core::image::Allocation, core::image::Error>)
+        + Send
+        + 'static,
+    ) {
+        use crate::image::raster::Memory;
+
+        let callback = Box::new(callback);
+
+        if let Some(callbacks) = self.raster.pending.get_mut(&handle.id()) {
+            callbacks.push(callback);
+            return;
+        }
+
+        if let Some(Memory::Device {
+            allocation, entry, ..
+        }) = self.raster.cache.get_mut(handle)
+        {
+            if let Some(allocation) = allocation
+                .as_ref()
+                .and_then(core::image::Allocation::upgrade)
+            {
+                callback(Ok(allocation));
+                return;
+            }
+
+            #[allow(unsafe_code)]
+            let new = unsafe { core::image::allocate(handle, entry.size()) };
+            *allocation = Some(new.downgrade());
+            callback(Ok(new));
+
+            return;
+        }
+
+        let _ = self.raster.pending.insert(handle.id(), vec![callback]);
+
+        #[cfg(not(target_arch = "wasm32"))]
+        self.worker.load(handle);
+    }
+
+    #[cfg(feature = "image")]
+    pub fn load_image(
+        &mut self,
+        device: &wgpu::Device,
+        queue: &wgpu::Queue,
+        handle: &core::image::Handle,
+    ) -> Result<core::image::Allocation, core::image::Error> {
+        use crate::image::raster::Memory;
+
+        if !self.raster.cache.contains(handle) {
+            self.raster.cache.insert(handle, Memory::load(handle));
+        }
+
+        match self.raster.cache.get_mut(handle).unwrap() {
+            Memory::Host(image) => {
+                let mut encoder = device.create_command_encoder(
+                    &wgpu::CommandEncoderDescriptor {
+                        label: Some("raster image upload"),
+                    },
+                );
+
+                let entry = self.atlas.upload(
+                    device,
+                    &mut encoder,
+                    &mut self.raster.belt,
+                    image.width(),
+                    image.height(),
+                    image,
+                );
+
+                self.raster.belt.finish();
+                let submission = queue.submit([encoder.finish()]);
+                self.raster.belt.recall();
+
+                let Some(entry) = entry else {
+                    return Err(core::image::Error::OutOfMemory);
+                };
+
+                let _ = device.poll(wgpu::PollType::Wait {
+                    submission_index: Some(submission),
+                    timeout: None,
+                });
+
+                #[allow(unsafe_code)]
+                let allocation = unsafe {
+                    core::image::allocate(
+                        handle,
+                        Size::new(image.width(), image.height()),
+                    )
+                };
+
+                self.raster.cache.insert(
+                    handle,
+                    Memory::Device {
+                        entry,
+                        bind_group: None,
+                        allocation: Some(allocation.downgrade()),
+                    },
+                );
+
+                Ok(allocation)
+            }
+            Memory::Device {
+                entry, allocation, ..
+            } => {
+                if let Some(allocation) = allocation
+                    .as_ref()
+                    .and_then(core::image::Allocation::upgrade)
+                {
+                    return Ok(allocation);
+                }
+
+                #[allow(unsafe_code)]
+                let new =
+                    unsafe { core::image::allocate(handle, entry.size()) };
+
+                *allocation = Some(new.downgrade());
+
+                Ok(new)
+            }
+            Memory::Error(error) => Err(error.clone()),
+        }
+    }
+
+    #[cfg(feature = "image")]
+    pub fn measure_image(
+        &mut self,
+        handle: &core::image::Handle,
+    ) -> Option<Size<u32>> {
+        self.receive();
+
+        let image = load_image(
+            &mut self.raster.cache,
+            &mut self.raster.pending,
+            #[cfg(not(target_arch = "wasm32"))]
+            &self.worker,
+            handle,
+            None,
+        )?;
+
+        Some(image.dimensions())
+    }
+
+    #[cfg(feature = "svg")]
+    pub fn measure_svg(&mut self, handle: &core::svg::Handle) -> Size<u32> {
+        // TODO: Concurrency
+        self.vector.load(handle).viewport_dimensions()
+    }
+
+    #[cfg(feature = "image")]
+    pub fn upload_raster(
+        &mut self,
+        device: &wgpu::Device,
+        encoder: &mut wgpu::CommandEncoder,
+        belt: &mut wgpu::util::StagingBelt,
+        handle: &core::image::Handle,
+    ) -> Option<(&atlas::Entry, &Arc<wgpu::BindGroup>)> {
+        use crate::image::raster::Memory;
+
+        self.receive();
+
+        let memory = load_image(
+            &mut self.raster.cache,
+            &mut self.raster.pending,
+            #[cfg(not(target_arch = "wasm32"))]
+            &self.worker,
+            handle,
+            None,
+        )?;
+
+        if let Memory::Device {
+            entry, bind_group, ..
+        } = memory
+        {
+            return Some((
+                entry,
+                bind_group.as_ref().unwrap_or(self.atlas.bind_group()),
+            ));
+        }
+
+        let image = memory.host()?;
+
+        const MAX_SYNC_SIZE: usize = 2 * 1024 * 1024;
+
+        // TODO: Concurrent Wasm support
+        if image.len() < MAX_SYNC_SIZE || cfg!(target_arch = "wasm32") {
+            let entry = self.atlas.upload(
+                device,
+                encoder,
+                belt,
+                image.width(),
+                image.height(),
+                &image,
+            )?;
+
+            *memory = Memory::Device {
+                entry,
+                bind_group: None,
+                allocation: None,
+            };
+
+            if let Memory::Device { entry, .. } = memory {
+                return Some((entry, self.atlas.bind_group()));
+            }
+        }
+
+        if !self.raster.pending.contains_key(&handle.id()) {
+            let _ = self.raster.pending.insert(handle.id(), Vec::new());
+
+            #[cfg(not(target_arch = "wasm32"))]
+            self.worker.upload(handle, image);
+        }
+
+        None
+    }
+
+    #[cfg(feature = "svg")]
+    pub fn upload_vector(
+        &mut self,
+        device: &wgpu::Device,
+        encoder: &mut wgpu::CommandEncoder,
+        belt: &mut wgpu::util::StagingBelt,
+        handle: &core::svg::Handle,
+        color: Option<core::Color>,
+        size: Size,
+        scale: f32,
+    ) -> Option<(&atlas::Entry, &Arc<wgpu::BindGroup>)> {
+        // TODO: Concurrency
+        self.vector
+            .upload(
+                device,
+                encoder,
+                belt,
+                handle,
+                color,
+                size,
+                scale,
+                &mut self.atlas,
+            )
+            .map(|entry| (entry, self.atlas.bind_group()))
+    }
+
+    pub fn trim(&mut self) {
+        #[cfg(feature = "image")]
+        {
+            self.receive();
+            self.raster.cache.trim(&mut self.atlas, |_bind_group| {
+                #[cfg(not(target_arch = "wasm32"))]
+                self.worker.drop(_bind_group);
+            });
+        }
+
+        #[cfg(feature = "svg")]
+        self.vector.trim(&mut self.atlas); // TODO: Concurrency
+    }
+
+    #[cfg(feature = "image")]
+    fn receive(&mut self) {
+        #[cfg(not(target_arch = "wasm32"))]
+        while let Ok(work) = self.worker.try_recv() {
+            use crate::image::raster::Memory;
+
+            match work {
+                worker::Work::Upload {
+                    handle,
+                    entry,
+                    bind_group,
+                } => {
+                    let callbacks = self.raster.pending.remove(&handle.id());
+
+                    let allocation = if let Some(callbacks) = callbacks {
+                        #[allow(unsafe_code)]
+                        let allocation = unsafe {
+                            core::image::allocate(&handle, entry.size())
+                        };
+
+                        let reference = allocation.downgrade();
+
+                        for callback in callbacks {
+                            callback(Ok(allocation.clone()));
+                        }
+
+                        Some(reference)
+                    } else {
+                        None
+                    };
+
+                    self.raster.cache.insert(
+                        &handle,
+                        Memory::Device {
+                            entry,
+                            bind_group: Some(bind_group),
+                            allocation,
+                        },
+                    );
+                }
+                worker::Work::Error { handle, error } => {
+                    let callbacks = self.raster.pending.remove(&handle.id());
+
+                    if let Some(callbacks) = callbacks {
+                        for callback in callbacks {
+                            callback(Err(error.clone()));
+                        }
+                    }
+
+                    self.raster.cache.insert(&handle, Memory::Error(error));
+                }
+            }
+        }
+    }
+}
+
+#[cfg(all(feature = "image", not(target_arch = "wasm32")))]
+impl Drop for Cache {
+    fn drop(&mut self) {
+        self.worker.quit();
+    }
+}
+
+#[cfg(feature = "image")]
+struct Raster {
+    cache: crate::image::raster::Cache,
+    pending: HashMap<core::image::Id, Vec<Callback>>,
+    belt: wgpu::util::StagingBelt,
+}
+
+#[cfg(feature = "image")]
+type Callback =
+    Box<dyn FnOnce(Result<core::image::Allocation, core::image::Error>) + Send>;
+
+#[cfg(feature = "image")]
+fn load_image<'a>(
+    cache: &'a mut crate::image::raster::Cache,
+    pending: &mut HashMap<core::image::Id, Vec<Callback>>,
+    #[cfg(not(target_arch = "wasm32"))] worker: &Worker,
+    handle: &core::image::Handle,
+    callback: Option<Callback>,
+) -> Option<&'a mut crate::image::raster::Memory> {
+    use crate::image::raster::Memory;
+
+    if !cache.contains(handle) {
+        if cfg!(target_arch = "wasm32") {
+            // TODO: Concurrent support for Wasm
+            cache.insert(handle, Memory::load(handle));
+        } else if let core::image::Handle::Rgba { .. } = handle {
+            // Load RGBA handles synchronously, since it's very cheap
+            cache.insert(handle, Memory::load(handle));
+        } else if !pending.contains_key(&handle.id()) {
+            let _ = pending.insert(handle.id(), Vec::from_iter(callback));
+
+            #[cfg(not(target_arch = "wasm32"))]
+            worker.load(handle);
+        }
+    }
+
+    cache.get_mut(handle)
+}
+
+#[cfg(all(feature = "image", not(target_arch = "wasm32")))]
+mod worker {
+    use crate::core::Bytes;
+    use crate::core::image;
+    use crate::graphics::Shell;
+    use crate::image::atlas::{self, Atlas};
+    use crate::image::raster;
+
+    use std::sync::Arc;
+    use std::sync::mpsc;
+    use std::thread;
+
+    pub struct Worker {
+        jobs: mpsc::SyncSender<Job>,
+        quit: mpsc::SyncSender<()>,
+        work: mpsc::Receiver<Work>,
+        handle: Option<std::thread::JoinHandle<()>>,
+    }
+
+    impl Worker {
+        pub fn new(
+            device: &wgpu::Device,
+            queue: &wgpu::Queue,
+            backend: wgpu::Backend,
+            texture_layout: wgpu::BindGroupLayout,
+            shell: &Shell,
+        ) -> Self {
+            let (jobs_sender, jobs_receiver) = mpsc::sync_channel(1_000);
+            let (quit_sender, quit_receiver) = mpsc::sync_channel(1);
+            let (work_sender, work_receiver) = mpsc::sync_channel(1_000);
+
+            let instance = Instance {
+                device: device.clone(),
+                queue: queue.clone(),
+                backend,
+                texture_layout,
+                shell: shell.clone(),
+                belt: wgpu::util::StagingBelt::new(4 * 1024 * 1024),
+                jobs: jobs_receiver,
+                output: work_sender,
+                quit: quit_receiver,
+            };
+
+            let handle = thread::spawn(move || instance.run());
+
+            Self {
+                jobs: jobs_sender,
+                quit: quit_sender,
+                work: work_receiver,
+                handle: Some(handle),
+            }
+        }
+
+        pub fn load(&self, handle: &image::Handle) {
+            let _ = self.jobs.send(Job::Load(handle.clone()));
+        }
+
+        pub fn upload(&self, handle: &image::Handle, image: raster::Image) {
+            let _ = self.jobs.send(Job::Upload {
+                handle: handle.clone(),
+                width: image.width(),
+                height: image.height(),
+                rgba: image.into_raw(),
+            });
+        }
+
+        pub fn drop(&self, bind_group: Arc<wgpu::BindGroup>) {
+            let _ = self.jobs.send(Job::Drop(bind_group));
+        }
+
+        pub fn try_recv(&self) -> Result<Work, mpsc::TryRecvError> {
+            self.work.try_recv()
+        }
+
+        pub fn quit(&mut self) {
+            let _ = self.quit.try_send(());
+            let _ = self.jobs.send(Job::Quit);
+            let _ = self.handle.take().map(thread::JoinHandle::join);
+        }
+    }
+
+    pub struct Instance {
+        device: wgpu::Device,
+        queue: wgpu::Queue,
+        backend: wgpu::Backend,
+        texture_layout: wgpu::BindGroupLayout,
+        shell: Shell,
+        belt: wgpu::util::StagingBelt,
+        jobs: mpsc::Receiver<Job>,
+        output: mpsc::SyncSender<Work>,
+        quit: mpsc::Receiver<()>,
+    }
+
+    #[derive(Debug)]
+    enum Job {
+        Load(image::Handle),
+        Upload {
+            handle: image::Handle,
+            rgba: Bytes,
+            width: u32,
+            height: u32,
+        },
+        Drop(Arc<wgpu::BindGroup>),
+        Quit,
+    }
+
+    pub enum Work {
+        Upload {
+            handle: image::Handle,
+            entry: atlas::Entry,
+            bind_group: Arc<wgpu::BindGroup>,
+        },
+        Error {
+            handle: image::Handle,
+            error: image::Error,
+        },
+    }
+
+    impl Instance {
+        fn run(mut self) {
+            loop {
+                if self.quit.try_recv().is_ok() {
+                    return;
+                }
+
+                let Ok(job) = self.jobs.recv() else {
+                    return;
+                };
+
+                match job {
+                    Job::Load(handle) => {
+                        match crate::graphics::image::load(&handle) {
+                            Ok(image) => self.upload(
+                                handle,
+                                image.width(),
+                                image.height(),
+                                image.into_raw(),
+                                Shell::invalidate_layout,
+                            ),
+                            Err(error) => {
+                                let _ = self
+                                    .output
+                                    .send(Work::Error { handle, error });
+                            }
+                        }
+                    }
+                    Job::Upload {
+                        handle,
+                        rgba,
+                        width,
+                        height,
+                    } => {
+                        self.upload(
+                            handle,
+                            width,
+                            height,
+                            rgba,
+                            Shell::request_redraw,
+                        );
+                    }
+                    Job::Drop(bind_group) => {
+                        drop(bind_group);
+                    }
+                    Job::Quit => return,
+                }
+            }
+        }
+
+        fn upload(
+            &mut self,
+            handle: image::Handle,
+            width: u32,
+            height: u32,
+            rgba: Bytes,
+            callback: fn(&Shell),
+        ) {
+            let mut encoder = self.device.create_command_encoder(
+                &wgpu::CommandEncoderDescriptor {
+                    label: Some("raster image upload"),
+                },
+            );
+
+            let mut atlas = Atlas::with_size(
+                &self.device,
+                self.backend,
+                self.texture_layout.clone(),
+                width.max(height),
+            );
+
+            let Some(entry) = atlas.upload(
+                &self.device,
+                &mut encoder,
+                &mut self.belt,
+                width,
+                height,
+                &rgba,
+            ) else {
+                return;
+            };
+
+            let output = self.output.clone();
+            let shell = self.shell.clone();
+
+            self.belt.finish();
+            let submission = self.queue.submit([encoder.finish()]);
+            self.belt.recall();
+
+            let bind_group = atlas.bind_group().clone();
+
+            self.queue.on_submitted_work_done(move || {
+                let _ = output.send(Work::Upload {
+                    handle,
+                    entry,
+                    bind_group,
+                });
+
+                callback(&shell);
+            });
+
+            let _ = self.device.poll(wgpu::PollType::Wait {
+                submission_index: Some(submission),
+                timeout: None,
+            });
+        }
+    }
+}

crates/iced_wgpu/src/image/mod.rs

@@ -0,0 +1,758 @@
+pub(crate) mod cache;
+pub(crate) use cache::Cache;
+
+mod atlas;
+
+#[cfg(feature = "image")]
+mod raster;
+
+#[cfg(feature = "svg")]
+mod vector;
+
+use crate::Buffer;
+use crate::core::border;
+use crate::core::{Rectangle, Size, Transformation};
+use crate::graphics::Shell;
+
+use bytemuck::{Pod, Zeroable};
+
+use std::mem;
+use std::sync::Arc;
+
+pub use crate::graphics::Image;
+
+pub type Batch = Vec<Image>;
+
+#[derive(Debug, Clone)]
+pub struct Pipeline {
+    raw: wgpu::RenderPipeline,
+    backend: wgpu::Backend,
+    nearest_sampler: wgpu::Sampler,
+    linear_sampler: wgpu::Sampler,
+    texture_layout: wgpu::BindGroupLayout,
+    constant_layout: wgpu::BindGroupLayout,
+}
+
+impl Pipeline {
+    pub fn new(
+        device: &wgpu::Device,
+        format: wgpu::TextureFormat,
+        backend: wgpu::Backend,
+    ) -> Self {
+        let nearest_sampler = device.create_sampler(&wgpu::SamplerDescriptor {
+            address_mode_u: wgpu::AddressMode::ClampToEdge,
+            address_mode_v: wgpu::AddressMode::ClampToEdge,
+            address_mode_w: wgpu::AddressMode::ClampToEdge,
+            min_filter: wgpu::FilterMode::Nearest,
+            mag_filter: wgpu::FilterMode::Nearest,
+            mipmap_filter: wgpu::FilterMode::Nearest,
+            ..Default::default()
+        });
+
+        let linear_sampler = device.create_sampler(&wgpu::SamplerDescriptor {
+            address_mode_u: wgpu::AddressMode::ClampToEdge,
+            address_mode_v: wgpu::AddressMode::ClampToEdge,
+            address_mode_w: wgpu::AddressMode::ClampToEdge,
+            min_filter: wgpu::FilterMode::Linear,
+            mag_filter: wgpu::FilterMode::Linear,
+            mipmap_filter: wgpu::FilterMode::Linear,
+            ..Default::default()
+        });
+
+        let constant_layout =
+            device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
+                label: Some("iced_wgpu::image constants layout"),
+                entries: &[
+                    wgpu::BindGroupLayoutEntry {
+                        binding: 0,
+                        visibility: wgpu::ShaderStages::VERTEX,
+                        ty: wgpu::BindingType::Buffer {
+                            ty: wgpu::BufferBindingType::Uniform,
+                            has_dynamic_offset: false,
+                            min_binding_size: wgpu::BufferSize::new(
+                                mem::size_of::<Uniforms>() as u64,
+                            ),
+                        },
+                        count: None,
+                    },
+                    wgpu::BindGroupLayoutEntry {
+                        binding: 1,
+                        visibility: wgpu::ShaderStages::FRAGMENT,
+                        ty: wgpu::BindingType::Sampler(
+                            wgpu::SamplerBindingType::Filtering,
+                        ),
+                        count: None,
+                    },
+                ],
+            });
+
+        let texture_layout =
+            device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
+                label: Some("iced_wgpu::image texture atlas layout"),
+                entries: &[wgpu::BindGroupLayoutEntry {
+                    binding: 0,
+                    visibility: wgpu::ShaderStages::FRAGMENT,
+                    ty: wgpu::BindingType::Texture {
+                        sample_type: wgpu::TextureSampleType::Float {
+                            filterable: true,
+                        },
+                        view_dimension: wgpu::TextureViewDimension::D2Array,
+                        multisampled: false,
+                    },
+                    count: None,
+                }],
+            });
+
+        let layout =
+            device.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
+                label: Some("iced_wgpu::image pipeline layout"),
+                push_constant_ranges: &[],
+                bind_group_layouts: &[&constant_layout, &texture_layout],
+            });
+
+        let shader =
+            device.create_shader_module(wgpu::ShaderModuleDescriptor {
+                label: Some("iced_wgpu image shader"),
+                source: wgpu::ShaderSource::Wgsl(std::borrow::Cow::Borrowed(
+                    concat!(
+                        include_str!("../shader/vertex.wgsl"),
+                        "\n",
+                        include_str!("../shader/image.wgsl"),
+                    ),
+                )),
+            });
+
+        let pipeline =
+            device.create_render_pipeline(&wgpu::RenderPipelineDescriptor {
+                label: Some("iced_wgpu::image pipeline"),
+                layout: Some(&layout),
+                vertex: wgpu::VertexState {
+                    module: &shader,
+                    entry_point: Some("vs_main"),
+                    buffers: &[wgpu::VertexBufferLayout {
+                        array_stride: mem::size_of::<Instance>() as u64,
+                        step_mode: wgpu::VertexStepMode::Instance,
+                        attributes: &wgpu::vertex_attr_array!(
+                            // Center
+                            0 => Float32x2,
+                            // Clip bounds
+                            1 => Float32x4,
+                            // Border radius
+                            2 => Float32x4,
+                            // Tile
+                            3 => Float32x4,
+                            // Rotation
+                            4 => Float32,
+                            // Opacity
+                            5 => Float32,
+                            // Atlas position
+                            6 => Float32x2,
+                            // Atlas scale
+                            7 => Float32x2,
+                            // Layer
+                            8 => Sint32,
+                            // Snap
+                            9 => Uint32,
+                        ),
+                    }],
+                    compilation_options:
+                        wgpu::PipelineCompilationOptions::default(),
+                },
+                fragment: Some(wgpu::FragmentState {
+                    module: &shader,
+                    entry_point: Some("fs_main"),
+                    targets: &[Some(wgpu::ColorTargetState {
+                        format,
+                        blend: Some(wgpu::BlendState {
+                            color: wgpu::BlendComponent {
+                                src_factor: wgpu::BlendFactor::SrcAlpha,
+                                dst_factor: wgpu::BlendFactor::OneMinusSrcAlpha,
+                                operation: wgpu::BlendOperation::Add,
+                            },
+                            alpha: wgpu::BlendComponent {
+                                src_factor: wgpu::BlendFactor::One,
+                                dst_factor: wgpu::BlendFactor::OneMinusSrcAlpha,
+                                operation: wgpu::BlendOperation::Add,
+                            },
+                        }),
+                        write_mask: wgpu::ColorWrites::ALL,
+                    })],
+                    compilation_options:
+                        wgpu::PipelineCompilationOptions::default(),
+                }),
+                primitive: wgpu::PrimitiveState {
+                    topology: wgpu::PrimitiveTopology::TriangleList,
+                    front_face: wgpu::FrontFace::Cw,
+                    ..Default::default()
+                },
+                depth_stencil: None,
+                multisample: wgpu::MultisampleState {
+                    count: 1,
+                    mask: !0,
+                    alpha_to_coverage_enabled: false,
+                },
+                multiview: None,
+                cache: None,
+            });
+
+        Pipeline {
+            raw: pipeline,
+            backend,
+            nearest_sampler,
+            linear_sampler,
+            texture_layout,
+            constant_layout,
+        }
+    }
+
+    pub fn create_cache(
+        &self,
+        device: &wgpu::Device,
+        queue: &wgpu::Queue,
+        shell: &Shell,
+    ) -> Cache {
+        Cache::new(
+            device,
+            queue,
+            self.backend,
+            self.texture_layout.clone(),
+            shell,
+        )
+    }
+}
+
+#[derive(Default)]
+pub struct State {
+    layers: Vec<Layer>,
+    prepare_layer: usize,
+    nearest_instances: Vec<Instance>,
+    linear_instances: Vec<Instance>,
+}
+
+impl State {
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    pub fn prepare(
+        &mut self,
+        pipeline: &Pipeline,
+        device: &wgpu::Device,
+        belt: &mut wgpu::util::StagingBelt,
+        encoder: &mut wgpu::CommandEncoder,
+        cache: &mut Cache,
+        images: &Batch,
+        transformation: Transformation,
+        scale: f32,
+    ) {
+        if self.layers.len() <= self.prepare_layer {
+            self.layers.push(Layer::new(
+                device,
+                &pipeline.constant_layout,
+                &pipeline.nearest_sampler,
+                &pipeline.linear_sampler,
+            ));
+        }
+
+        let layer = &mut self.layers[self.prepare_layer];
+
+        let mut atlas: Option<Arc<wgpu::BindGroup>> = None;
+
+        for image in images {
+            match &image {
+                #[cfg(feature = "image")]
+                Image::Raster {
+                    image,
+                    bounds,
+                    clip_bounds,
+                } => {
+                    if let Some((atlas_entry, bind_group)) = cache
+                        .upload_raster(device, encoder, belt, &image.handle)
+                    {
+                        match atlas.as_mut() {
+                            None => {
+                                atlas = Some(bind_group.clone());
+                            }
+                            Some(atlas) if atlas != bind_group => {
+                                layer.push(
+                                    atlas,
+                                    &self.nearest_instances,
+                                    &self.linear_instances,
+                                );
+
+                                *atlas = Arc::clone(bind_group);
+                            }
+                            _ => {}
+                        }
+
+                        add_instances(
+                            *bounds,
+                            *clip_bounds,
+                            image.border_radius,
+                            f32::from(image.rotation),
+                            image.opacity,
+                            image.snap,
+                            atlas_entry,
+                            match image.filter_method {
+                                crate::core::image::FilterMethod::Nearest => {
+                                    &mut self.nearest_instances
+                                }
+                                crate::core::image::FilterMethod::Linear => {
+                                    &mut self.linear_instances
+                                }
+                            },
+                        );
+                    }
+                }
+                #[cfg(not(feature = "image"))]
+                Image::Raster { .. } => continue,
+
+                #[cfg(feature = "svg")]
+                Image::Vector {
+                    svg,
+                    bounds,
+                    clip_bounds,
+                } => {
+                    if let Some((atlas_entry, bind_group)) = cache
+                        .upload_vector(
+                            device,
+                            encoder,
+                            belt,
+                            &svg.handle,
+                            svg.color,
+                            bounds.size(),
+                            scale,
+                        )
+                    {
+                        match atlas.as_mut() {
+                            None => {
+                                atlas = Some(bind_group.clone());
+                            }
+                            Some(atlas) if atlas != bind_group => {
+                                layer.push(
+                                    atlas,
+                                    &self.nearest_instances,
+                                    &self.linear_instances,
+                                );
+
+                                *atlas = bind_group.clone();
+                            }
+                            _ => {}
+                        }
+
+                        add_instances(
+                            *bounds,
+                            *clip_bounds,
+                            border::radius(0),
+                            f32::from(svg.rotation),
+                            svg.opacity,
+                            true,
+                            atlas_entry,
+                            &mut self.nearest_instances,
+                        );
+                    }
+                }
+                #[cfg(not(feature = "svg"))]
+                Image::Vector { .. } => continue,
+            }
+        }
+
+        if let Some(atlas) = &atlas {
+            layer.push(atlas, &self.nearest_instances, &self.linear_instances);
+        }
+
+        layer.prepare(
+            device,
+            encoder,
+            belt,
+            transformation,
+            scale,
+            &self.nearest_instances,
+            &self.linear_instances,
+        );
+
+        self.prepare_layer += 1;
+        self.nearest_instances.clear();
+        self.linear_instances.clear();
+    }
+
+    pub fn render<'a>(
+        &'a self,
+        pipeline: &'a Pipeline,
+        layer: usize,
+        bounds: Rectangle<u32>,
+        render_pass: &mut wgpu::RenderPass<'a>,
+    ) {
+        if let Some(layer) = self.layers.get(layer) {
+            render_pass.set_pipeline(&pipeline.raw);
+
+            render_pass.set_scissor_rect(
+                bounds.x,
+                bounds.y,
+                bounds.width,
+                bounds.height,
+            );
+
+            layer.render(render_pass);
+        }
+    }
+
+    pub fn trim(&mut self) {
+        for layer in &mut self.layers[..self.prepare_layer] {
+            layer.clear();
+        }
+
+        self.prepare_layer = 0;
+    }
+}
+
+#[derive(Debug)]
+struct Layer {
+    uniforms: wgpu::Buffer,
+    instances: Buffer<Instance>,
+    nearest: Vec<Group>,
+    nearest_layout: wgpu::BindGroup,
+    nearest_total: usize,
+    linear: Vec<Group>,
+    linear_layout: wgpu::BindGroup,
+    linear_total: usize,
+}
+
+#[derive(Debug)]
+struct Group {
+    atlas: Arc<wgpu::BindGroup>,
+    instance_count: usize,
+}
+
+impl Layer {
+    fn new(
+        device: &wgpu::Device,
+        constant_layout: &wgpu::BindGroupLayout,
+        nearest_sampler: &wgpu::Sampler,
+        linear_sampler: &wgpu::Sampler,
+    ) -> Self {
+        let uniforms = device.create_buffer(&wgpu::BufferDescriptor {
+            label: Some("iced_wgpu::image uniforms buffer"),
+            size: mem::size_of::<Uniforms>() as u64,
+            usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
+            mapped_at_creation: false,
+        });
+
+        let instances = Buffer::new(
+            device,
+            "iced_wgpu::image instance buffer",
+            Instance::INITIAL,
+            wgpu::BufferUsages::VERTEX | wgpu::BufferUsages::COPY_DST,
+        );
+
+        let nearest_layout =
+            device.create_bind_group(&wgpu::BindGroupDescriptor {
+                label: Some("iced_wgpu::image constants bind group"),
+                layout: constant_layout,
+                entries: &[
+                    wgpu::BindGroupEntry {
+                        binding: 0,
+                        resource: wgpu::BindingResource::Buffer(
+                            wgpu::BufferBinding {
+                                buffer: &uniforms,
+                                offset: 0,
+                                size: None,
+                            },
+                        ),
+                    },
+                    wgpu::BindGroupEntry {
+                        binding: 1,
+                        resource: wgpu::BindingResource::Sampler(
+                            nearest_sampler,
+                        ),
+                    },
+                ],
+            });
+
+        let linear_layout =
+            device.create_bind_group(&wgpu::BindGroupDescriptor {
+                label: Some("iced_wgpu::image constants bind group"),
+                layout: constant_layout,
+                entries: &[
+                    wgpu::BindGroupEntry {
+                        binding: 0,
+                        resource: wgpu::BindingResource::Buffer(
+                            wgpu::BufferBinding {
+                                buffer: &uniforms,
+                                offset: 0,
+                                size: None,
+                            },
+                        ),
+                    },
+                    wgpu::BindGroupEntry {
+                        binding: 1,
+                        resource: wgpu::BindingResource::Sampler(
+                            linear_sampler,
+                        ),
+                    },
+                ],
+            });
+
+        Self {
+            uniforms,
+            instances,
+            nearest: Vec::new(),
+            nearest_layout,
+            nearest_total: 0,
+            linear: Vec::new(),
+            linear_layout,
+            linear_total: 0,
+        }
+    }
+
+    fn prepare(
+        &mut self,
+        device: &wgpu::Device,
+        encoder: &mut wgpu::CommandEncoder,
+        belt: &mut wgpu::util::StagingBelt,
+        transformation: Transformation,
+        scale_factor: f32,
+        nearest: &[Instance],
+        linear: &[Instance],
+    ) {
+        let uniforms = Uniforms {
+            transform: transformation.into(),
+            scale_factor,
+            _padding: [0.0; 3],
+        };
+
+        let bytes = bytemuck::bytes_of(&uniforms);
+
+        belt.write_buffer(
+            encoder,
+            &self.uniforms,
+            0,
+            (bytes.len() as u64).try_into().expect("Sized uniforms"),
+            device,
+        )
+        .copy_from_slice(bytes);
+
+        let _ = self
+            .instances
+            .resize(device, self.nearest_total + self.linear_total);
+
+        let mut offset = 0;
+
+        if !nearest.is_empty() {
+            offset += self.instances.write(device, encoder, belt, 0, nearest);
+        }
+
+        if !linear.is_empty() {
+            let _ = self.instances.write(device, encoder, belt, offset, linear);
+        }
+    }
+
+    fn push(
+        &mut self,
+        atlas: &Arc<wgpu::BindGroup>,
+        nearest: &[Instance],
+        linear: &[Instance],
+    ) {
+        let new_nearest = nearest.len() - self.nearest_total;
+
+        if new_nearest > 0 {
+            self.nearest.push(Group {
+                atlas: atlas.clone(),
+                instance_count: new_nearest,
+            });
+
+            self.nearest_total = nearest.len();
+        }
+
+        let new_linear = linear.len() - self.linear_total;
+
+        if new_linear > 0 {
+            self.linear.push(Group {
+                atlas: atlas.clone(),
+                instance_count: new_linear,
+            });
+
+            self.linear_total = linear.len();
+        }
+    }
+
+    fn render<'a>(&'a self, render_pass: &mut wgpu::RenderPass<'a>) {
+        render_pass.set_vertex_buffer(0, self.instances.slice(..));
+
+        let mut offset = 0;
+
+        if !self.nearest.is_empty() {
+            render_pass.set_bind_group(0, &self.nearest_layout, &[]);
+
+            for group in &self.nearest {
+                render_pass.set_bind_group(1, group.atlas.as_ref(), &[]);
+                render_pass
+                    .draw(0..6, offset..offset + group.instance_count as u32);
+
+                offset += group.instance_count as u32;
+            }
+        }
+
+        if !self.linear.is_empty() {
+            render_pass.set_bind_group(0, &self.linear_layout, &[]);
+
+            for group in &self.linear {
+                render_pass.set_bind_group(1, group.atlas.as_ref(), &[]);
+                render_pass
+                    .draw(0..6, offset..offset + group.instance_count as u32);
+
+                offset += group.instance_count as u32;
+            }
+        }
+    }
+
+    fn clear(&mut self) {
+        self.nearest.clear();
+        self.nearest_total = 0;
+
+        self.linear.clear();
+        self.linear_total = 0;
+    }
+}
+
+#[repr(C)]
+#[derive(Debug, Clone, Copy, Zeroable, Pod)]
+struct Instance {
+    _center: [f32; 2],
+    _clip_bounds: [f32; 4],
+    _border_radius: [f32; 4],
+    _tile: [f32; 4],
+    _rotation: f32,
+    _opacity: f32,
+    _position_in_atlas: [f32; 2],
+    _size_in_atlas: [f32; 2],
+    _layer: u32,
+    _snap: u32,
+}
+
+impl Instance {
+    pub const INITIAL: usize = 20;
+}
+
+#[repr(C)]
+#[derive(Debug, Clone, Copy, Zeroable, Pod)]
+struct Uniforms {
+    transform: [f32; 16],
+    scale_factor: f32,
+    // Uniforms must be aligned to their largest member,
+    // this uses a mat4x4<f32> which aligns to 16, so align to that
+    _padding: [f32; 3],
+}
+
+fn add_instances(
+    bounds: Rectangle,
+    clip_bounds: Rectangle,
+    border_radius: border::Radius,
+    rotation: f32,
+    opacity: f32,
+    snap: bool,
+    entry: &atlas::Entry,
+    instances: &mut Vec<Instance>,
+) {
+    let center = [
+        bounds.x + bounds.width / 2.0,
+        bounds.y + bounds.height / 2.0,
+    ];
+
+    let clip_bounds = [
+        clip_bounds.x,
+        clip_bounds.y,
+        clip_bounds.width,
+        clip_bounds.height,
+    ];
+
+    let border_radius = border_radius.into();
+
+    match entry {
+        atlas::Entry::Contiguous(allocation) => {
+            add_instance(
+                center,
+                clip_bounds,
+                border_radius,
+                [bounds.x, bounds.y, bounds.width, bounds.height],
+                rotation,
+                opacity,
+                snap,
+                allocation,
+                instances,
+            );
+        }
+        atlas::Entry::Fragmented { fragments, size } => {
+            let scaling_x = bounds.width / size.width as f32;
+            let scaling_y = bounds.height / size.height as f32;
+
+            for fragment in fragments {
+                let allocation = &fragment.allocation;
+                let (fragment_x, fragment_y) = fragment.position;
+
+                let Size {
+                    width: fragment_width,
+                    height: fragment_height,
+                } = allocation.size();
+
+                let tile = [
+                    bounds.x + fragment_x as f32 * scaling_x,
+                    bounds.y + fragment_y as f32 * scaling_y,
+                    fragment_width as f32 * scaling_x,
+                    fragment_height as f32 * scaling_y,
+                ];
+
+                add_instance(
+                    center,
+                    clip_bounds,
+                    border_radius,
+                    tile,
+                    rotation,
+                    opacity,
+                    snap,
+                    allocation,
+                    instances,
+                );
+            }
+        }
+    }
+}
+
+#[inline]
+fn add_instance(
+    center: [f32; 2],
+    clip_bounds: [f32; 4],
+    border_radius: [f32; 4],
+    tile: [f32; 4],
+    rotation: f32,
+    opacity: f32,
+    snap: bool,
+    allocation: &atlas::Allocation,
+    instances: &mut Vec<Instance>,
+) {
+    let (x, y) = allocation.position();
+    let Size { width, height } = allocation.size();
+    let layer = allocation.layer();
+    let atlas_size = allocation.atlas_size();
+
+    let instance = Instance {
+        _center: center,
+        _clip_bounds: clip_bounds,
+        _border_radius: border_radius,
+        _tile: tile,
+        _rotation: rotation,
+        _opacity: opacity,
+        _position_in_atlas: [
+            x as f32 / atlas_size as f32,
+            y as f32 / atlas_size as f32,
+        ],
+        _size_in_atlas: [
+            width as f32 / atlas_size as f32,
+            height as f32 / atlas_size as f32,
+        ],
+        _layer: layer as u32,
+        _snap: snap as u32,
+    };
+
+    instances.push(instance);
+}

crates/iced_wgpu/src/image/null.rs

@@ -0,0 +1,16 @@
+pub use crate::graphics::Image;
+
+#[derive(Debug, Default)]
+pub struct Batch;
+
+impl Batch {
+    pub fn push(&mut self, _image: Image) {}
+
+    pub fn clear(&mut self) {}
+
+    pub fn is_empty(&self) -> bool {
+        true
+    }
+
+    pub fn append(&mut self, _batch: &mut Self) {}
+}

crates/iced_wgpu/src/image/raster.rs

@@ -0,0 +1,123 @@
+use crate::core::Size;
+use crate::core::image;
+use crate::graphics;
+use crate::image::atlas::{self, Atlas};
+
+use rustc_hash::{FxHashMap, FxHashSet};
+use std::sync::{Arc, Weak};
+
+pub type Image = graphics::image::Buffer;
+
+/// Entry in cache corresponding to an image handle
+#[derive(Debug)]
+pub enum Memory {
+    /// Image data on host
+    Host(Image),
+    /// Storage entry
+    Device {
+        entry: atlas::Entry,
+        bind_group: Option<Arc<wgpu::BindGroup>>,
+        allocation: Option<Weak<image::Memory>>,
+    },
+    Error(image::Error),
+}
+
+impl Memory {
+    pub fn load(handle: &image::Handle) -> Self {
+        match graphics::image::load(handle) {
+            Ok(image) => Self::Host(image),
+            Err(error) => Self::Error(error),
+        }
+    }
+
+    pub fn dimensions(&self) -> Size<u32> {
+        match self {
+            Memory::Host(image) => {
+                let (width, height) = image.dimensions();
+
+                Size::new(width, height)
+            }
+            Memory::Device { entry, .. } => entry.size(),
+            Memory::Error(_) => Size::new(1, 1),
+        }
+    }
+
+    pub fn host(&self) -> Option<Image> {
+        match self {
+            Memory::Host(image) => Some(image.clone()),
+            Memory::Device { .. } | Memory::Error(_) => None,
+        }
+    }
+}
+
+#[derive(Debug, Default)]
+pub struct Cache {
+    map: FxHashMap<image::Id, Memory>,
+    hits: FxHashSet<image::Id>,
+    should_trim: bool,
+}
+
+impl Cache {
+    pub fn get_mut(&mut self, handle: &image::Handle) -> Option<&mut Memory> {
+        let _ = self.hits.insert(handle.id());
+
+        self.map.get_mut(&handle.id())
+    }
+
+    pub fn insert(&mut self, handle: &image::Handle, memory: Memory) {
+        let _ = self.map.insert(handle.id(), memory);
+        let _ = self.hits.insert(handle.id());
+
+        self.should_trim = true;
+    }
+
+    pub fn contains(&self, handle: &image::Handle) -> bool {
+        self.map.contains_key(&handle.id())
+    }
+
+    pub fn trim(
+        &mut self,
+        atlas: &mut Atlas,
+        on_drop: impl Fn(Arc<wgpu::BindGroup>),
+    ) {
+        // Only trim if new entries have landed in the `Cache`
+        if !self.should_trim {
+            return;
+        }
+
+        let hits = &self.hits;
+
+        self.map.retain(|id, memory| {
+            // Retain active allocations
+            if let Memory::Device { allocation, .. } = memory
+                && allocation
+                    .as_ref()
+                    .is_some_and(|allocation| allocation.strong_count() > 0)
+            {
+                return true;
+            }
+
+            let retain = hits.contains(id);
+
+            if !retain {
+                log::debug!("Dropping image allocation: {id:?}");
+
+                if let Memory::Device {
+                    entry, bind_group, ..
+                } = memory
+                {
+                    if let Some(bind_group) = bind_group.take() {
+                        on_drop(bind_group);
+                    } else {
+                        atlas.remove(entry);
+                    }
+                }
+            }
+
+            retain
+        });
+
+        self.hits.clear();
+        self.should_trim = false;
+    }
+}

crates/iced_wgpu/src/image/vector.rs

@@ -0,0 +1,231 @@
+use crate::core::svg;
+use crate::core::{Color, Size};
+use crate::image::atlas::{self, Atlas};
+
+use resvg::tiny_skia;
+use resvg::usvg;
+use rustc_hash::{FxHashMap, FxHashSet};
+use std::fs;
+use std::panic;
+use std::sync::Arc;
+
+/// Entry in cache corresponding to an svg handle
+pub enum Svg {
+    /// Parsed svg
+    Loaded(usvg::Tree),
+    /// Svg not found or failed to parse
+    NotFound,
+}
+
+impl Svg {
+    /// Viewport width and height
+    pub fn viewport_dimensions(&self) -> Size<u32> {
+        match self {
+            Svg::Loaded(tree) => {
+                let size = tree.size();
+
+                Size::new(size.width() as u32, size.height() as u32)
+            }
+            Svg::NotFound => Size::new(1, 1),
+        }
+    }
+}
+
+/// Caches svg vector and raster data
+#[derive(Debug, Default)]
+pub struct Cache {
+    svgs: FxHashMap<u64, Svg>,
+    rasterized: FxHashMap<(u64, u32, u32, ColorFilter), atlas::Entry>,
+    svg_hits: FxHashSet<u64>,
+    rasterized_hits: FxHashSet<(u64, u32, u32, ColorFilter)>,
+    should_trim: bool,
+    fontdb: Option<Arc<usvg::fontdb::Database>>,
+}
+
+type ColorFilter = Option<[u8; 4]>;
+
+impl Cache {
+    /// Load svg
+    pub fn load(&mut self, handle: &svg::Handle) -> &Svg {
+        if self.svgs.contains_key(&handle.id()) {
+            return self.svgs.get(&handle.id()).unwrap();
+        }
+
+        // TODO: Reuse `cosmic-text` font database
+        if self.fontdb.is_none() {
+            let mut fontdb = usvg::fontdb::Database::new();
+            fontdb.load_system_fonts();
+
+            self.fontdb = Some(Arc::new(fontdb));
+        }
+
+        let options = usvg::Options {
+            fontdb: self
+                .fontdb
+                .as_ref()
+                .expect("fontdb must be initialized")
+                .clone(),
+            ..usvg::Options::default()
+        };
+
+        let svg = match handle.data() {
+            svg::Data::Path(path) => fs::read_to_string(path)
+                .ok()
+                .and_then(|contents| {
+                    usvg::Tree::from_str(&contents, &options).ok()
+                })
+                .map(Svg::Loaded)
+                .unwrap_or(Svg::NotFound),
+            svg::Data::Bytes(bytes) => {
+                match usvg::Tree::from_data(bytes, &options) {
+                    Ok(tree) => Svg::Loaded(tree),
+                    Err(_) => Svg::NotFound,
+                }
+            }
+        };
+
+        self.should_trim = true;
+
+        let _ = self.svgs.insert(handle.id(), svg);
+        self.svgs.get(&handle.id()).unwrap()
+    }
+
+    /// Load svg and upload raster data
+    pub fn upload(
+        &mut self,
+        device: &wgpu::Device,
+        encoder: &mut wgpu::CommandEncoder,
+        belt: &mut wgpu::util::StagingBelt,
+        handle: &svg::Handle,
+        color: Option<Color>,
+        size: Size,
+        scale: f32,
+        atlas: &mut Atlas,
+    ) -> Option<&atlas::Entry> {
+        let id = handle.id();
+
+        let (width, height) = (
+            (scale * size.width).ceil() as u32,
+            (scale * size.height).ceil() as u32,
+        );
+
+        let color = color.map(Color::into_rgba8);
+        let key = (id, width, height, color);
+
+        // TODO: Optimize!
+        // We currently rerasterize the SVG when its size changes. This is slow
+        // as heck. A GPU rasterizer like `pathfinder` may perform better.
+        // It would be cool to be able to smooth resize the `svg` example.
+        if self.rasterized.contains_key(&key) {
+            let _ = self.svg_hits.insert(id);
+            let _ = self.rasterized_hits.insert(key);
+
+            return self.rasterized.get(&key);
+        }
+
+        match self.load(handle) {
+            Svg::Loaded(tree) => {
+                if width == 0 || height == 0 {
+                    return None;
+                }
+
+                // TODO: Optimize!
+                // We currently rerasterize the SVG when its size changes. This is slow
+                // as heck. A GPU rasterizer like `pathfinder` may perform better.
+                // It would be cool to be able to smooth resize the `svg` example.
+                let mut img = tiny_skia::Pixmap::new(width, height)?;
+
+                let tree_size = tree.size().to_int_size();
+
+                let target_size = if width > height {
+                    tree_size.scale_to_width(width)
+                } else {
+                    tree_size.scale_to_height(height)
+                };
+
+                let transform = if let Some(target_size) = target_size {
+                    let tree_size = tree_size.to_size();
+                    let target_size = target_size.to_size();
+
+                    tiny_skia::Transform::from_scale(
+                        target_size.width() / tree_size.width(),
+                        target_size.height() / tree_size.height(),
+                    )
+                } else {
+                    tiny_skia::Transform::default()
+                };
+
+                // SVG rendering can panic on malformed or complex vectors.
+                // We catch panics to prevent crashes and continue gracefully.
+                let render =
+                    panic::catch_unwind(panic::AssertUnwindSafe(|| {
+                        resvg::render(tree, transform, &mut img.as_mut());
+                    }));
+
+                if let Err(error) = render {
+                    log::warn!(
+                        "SVG rendering for {handle:?} panicked: {error:?}"
+                    );
+                }
+
+                let mut rgba = img.take();
+
+                if let Some(color) = color {
+                    rgba.chunks_exact_mut(4).for_each(|rgba| {
+                        if rgba[3] > 0 {
+                            rgba[0] = color[0];
+                            rgba[1] = color[1];
+                            rgba[2] = color[2];
+                        }
+                    });
+                }
+
+                let allocation = atlas
+                    .upload(device, encoder, belt, width, height, &rgba)?;
+
+                log::debug!("allocating {id} {width}x{height}");
+
+                let _ = self.svg_hits.insert(id);
+                let _ = self.rasterized_hits.insert(key);
+                let _ = self.rasterized.insert(key, allocation);
+                self.should_trim = true;
+
+                self.rasterized.get(&key)
+            }
+            Svg::NotFound => None,
+        }
+    }
+
+    /// Load svg and upload raster data
+    pub fn trim(&mut self, atlas: &mut Atlas) {
+        if !self.should_trim {
+            return;
+        }
+
+        let svg_hits = &self.svg_hits;
+        let rasterized_hits = &self.rasterized_hits;
+
+        self.svgs.retain(|k, _| svg_hits.contains(k));
+        self.rasterized.retain(|k, entry| {
+            let retain = rasterized_hits.contains(k);
+
+            if !retain {
+                atlas.remove(entry);
+            }
+
+            retain
+        });
+        self.svg_hits.clear();
+        self.rasterized_hits.clear();
+        self.should_trim = false;
+    }
+}
+
+impl std::fmt::Debug for Svg {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        match self {
+            Svg::Loaded(_) => write!(f, "Svg::Loaded"),
+            Svg::NotFound => write!(f, "Svg::NotFound"),
+        }
+    }
+}

crates/iced_wgpu/src/layer.rs

@@ -0,0 +1,405 @@
+use crate::core::{
+    self, Background, Color, Point, Rectangle, Svg, Transformation, renderer,
+};
+use crate::graphics;
+use crate::graphics::Mesh;
+use crate::graphics::color;
+use crate::graphics::layer;
+use crate::graphics::mesh;
+use crate::graphics::text::{Editor, Paragraph};
+use crate::image::{self, Image};
+use crate::primitive::{self, Primitive};
+use crate::quad::{self, Quad};
+use crate::text::{self, Text};
+use crate::triangle;
+
+pub type Stack = layer::Stack<Layer>;
+
+#[derive(Debug)]
+pub struct Layer {
+    pub bounds: Rectangle,
+    pub quads: quad::Batch,
+    pub triangles: triangle::Batch,
+    pub primitives: primitive::Batch,
+    pub images: image::Batch,
+    pub text: text::Batch,
+    pending_meshes: Vec<Mesh>,
+    pending_text: Vec<Text>,
+}
+
+impl Layer {
+    pub fn is_empty(&self) -> bool {
+        self.quads.is_empty()
+            && self.triangles.is_empty()
+            && self.primitives.is_empty()
+            && self.images.is_empty()
+            && self.text.is_empty()
+            && self.pending_meshes.is_empty()
+            && self.pending_text.is_empty()
+    }
+
+    pub fn draw_quad(
+        &mut self,
+        quad: renderer::Quad,
+        background: Background,
+        transformation: Transformation,
+    ) {
+        let bounds = quad.bounds * transformation;
+
+        let quad = Quad {
+            position: [bounds.x, bounds.y],
+            size: [bounds.width, bounds.height],
+            border_color: color::pack(quad.border.color),
+            border_radius: (quad.border.radius * transformation.scale_factor())
+                .into(),
+            border_width: quad.border.width * transformation.scale_factor(),
+            shadow_color: color::pack(quad.shadow.color),
+            shadow_offset: (quad.shadow.offset * transformation.scale_factor())
+                .into(),
+            shadow_blur_radius: quad.shadow.blur_radius
+                * transformation.scale_factor(),
+            snap: quad.snap as u32,
+        };
+
+        self.quads.add(quad, &background);
+    }
+
+    pub fn draw_paragraph(
+        &mut self,
+        paragraph: &Paragraph,
+        position: Point,
+        color: Color,
+        clip_bounds: Rectangle,
+        transformation: Transformation,
+    ) {
+        let paragraph = Text::Paragraph {
+            paragraph: paragraph.downgrade(),
+            position,
+            color,
+            clip_bounds,
+            transformation,
+        };
+
+        self.pending_text.push(paragraph);
+    }
+
+    pub fn draw_editor(
+        &mut self,
+        editor: &Editor,
+        position: Point,
+        color: Color,
+        clip_bounds: Rectangle,
+        transformation: Transformation,
+    ) {
+        let editor = Text::Editor {
+            editor: editor.downgrade(),
+            position,
+            color,
+            clip_bounds,
+            transformation,
+        };
+
+        self.pending_text.push(editor);
+    }
+
+    pub fn draw_text(
+        &mut self,
+        text: crate::core::Text,
+        position: Point,
+        color: Color,
+        clip_bounds: Rectangle,
+        transformation: Transformation,
+    ) {
+        let text = Text::Cached {
+            content: text.content,
+            bounds: Rectangle::new(position, text.bounds) * transformation,
+            color,
+            size: text.size * transformation.scale_factor(),
+            line_height: text.line_height.to_absolute(text.size)
+                * transformation.scale_factor(),
+            font: text.font,
+            align_x: text.align_x,
+            align_y: text.align_y,
+            shaping: text.shaping,
+            clip_bounds: clip_bounds * transformation,
+        };
+
+        self.pending_text.push(text);
+    }
+
+    pub fn draw_text_raw(
+        &mut self,
+        raw: graphics::text::Raw,
+        transformation: Transformation,
+    ) {
+        let raw = Text::Raw {
+            raw,
+            transformation,
+        };
+
+        self.pending_text.push(raw);
+    }
+
+    pub fn draw_image(&mut self, image: Image, transformation: Transformation) {
+        match image {
+            Image::Raster {
+                image,
+                bounds,
+                clip_bounds,
+            } => {
+                self.draw_raster(image, bounds, clip_bounds, transformation);
+            }
+            Image::Vector {
+                svg,
+                bounds,
+                clip_bounds,
+            } => {
+                self.draw_svg(svg, bounds, clip_bounds, transformation);
+            }
+        }
+    }
+
+    pub fn draw_raster(
+        &mut self,
+        image: core::Image,
+        bounds: Rectangle,
+        clip_bounds: Rectangle,
+        transformation: Transformation,
+    ) {
+        let image = Image::Raster {
+            image: core::Image {
+                border_radius: image.border_radius
+                    * transformation.scale_factor(),
+                ..image
+            },
+            bounds: bounds * transformation,
+            clip_bounds: clip_bounds * transformation,
+        };
+
+        self.images.push(image);
+    }
+
+    pub fn draw_svg(
+        &mut self,
+        svg: Svg,
+        bounds: Rectangle,
+        clip_bounds: Rectangle,
+        transformation: Transformation,
+    ) {
+        let svg = Image::Vector {
+            svg,
+            bounds: bounds * transformation,
+            clip_bounds: clip_bounds * transformation,
+        };
+
+        self.images.push(svg);
+    }
+
+    pub fn draw_mesh(
+        &mut self,
+        mut mesh: Mesh,
+        transformation: Transformation,
+    ) {
+        match &mut mesh {
+            Mesh::Solid {
+                transformation: local_transformation,
+                ..
+            }
+            | Mesh::Gradient {
+                transformation: local_transformation,
+                ..
+            } => {
+                *local_transformation = *local_transformation * transformation;
+            }
+        }
+
+        self.pending_meshes.push(mesh);
+    }
+
+    pub fn draw_mesh_group(
+        &mut self,
+        meshes: Vec<Mesh>,
+        transformation: Transformation,
+    ) {
+        self.flush_meshes();
+
+        self.triangles.push(triangle::Item::Group {
+            meshes,
+            transformation,
+        });
+    }
+
+    pub fn draw_mesh_cache(
+        &mut self,
+        cache: mesh::Cache,
+        transformation: Transformation,
+    ) {
+        self.flush_meshes();
+
+        self.triangles.push(triangle::Item::Cached {
+            cache,
+            transformation,
+        });
+    }
+
+    pub fn draw_text_group(
+        &mut self,
+        text: Vec<Text>,
+        transformation: Transformation,
+    ) {
+        self.flush_text();
+
+        self.text.push(text::Item::Group {
+            text,
+            transformation,
+        });
+    }
+
+    pub fn draw_text_cache(
+        &mut self,
+        cache: text::Cache,
+        transformation: Transformation,
+    ) {
+        self.flush_text();
+
+        self.text.push(text::Item::Cached {
+            cache,
+            transformation,
+        });
+    }
+
+    pub fn draw_primitive(
+        &mut self,
+        bounds: Rectangle,
+        primitive: impl Primitive,
+        transformation: Transformation,
+    ) {
+        let bounds = bounds * transformation;
+
+        self.primitives
+            .push(primitive::Instance::new(bounds, primitive));
+    }
+
+    fn flush_meshes(&mut self) {
+        if !self.pending_meshes.is_empty() {
+            self.triangles.push(triangle::Item::Group {
+                transformation: Transformation::IDENTITY,
+                meshes: self.pending_meshes.drain(..).collect(),
+            });
+        }
+    }
+
+    fn flush_text(&mut self) {
+        if !self.pending_text.is_empty() {
+            self.text.push(text::Item::Group {
+                transformation: Transformation::IDENTITY,
+                text: self.pending_text.drain(..).collect(),
+            });
+        }
+    }
+}
+
+impl graphics::Layer for Layer {
+    fn with_bounds(bounds: Rectangle) -> Self {
+        Self {
+            bounds,
+            ..Self::default()
+        }
+    }
+
+    fn bounds(&self) -> Rectangle {
+        self.bounds
+    }
+
+    fn flush(&mut self) {
+        self.flush_meshes();
+        self.flush_text();
+    }
+
+    fn resize(&mut self, bounds: Rectangle) {
+        self.bounds = bounds;
+    }
+
+    fn reset(&mut self) {
+        self.bounds = Rectangle::INFINITE;
+
+        self.quads.clear();
+        self.triangles.clear();
+        self.primitives.clear();
+        self.text.clear();
+        self.images.clear();
+        self.pending_meshes.clear();
+        self.pending_text.clear();
+    }
+
+    fn start(&self) -> usize {
+        if !self.quads.is_empty() {
+            return 1;
+        }
+
+        if !self.triangles.is_empty() {
+            return 2;
+        }
+
+        if !self.primitives.is_empty() {
+            return 3;
+        }
+
+        if !self.images.is_empty() {
+            return 4;
+        }
+
+        if !self.text.is_empty() {
+            return 5;
+        }
+
+        usize::MAX
+    }
+
+    fn end(&self) -> usize {
+        if !self.text.is_empty() {
+            return 5;
+        }
+
+        if !self.images.is_empty() {
+            return 4;
+        }
+
+        if !self.primitives.is_empty() {
+            return 3;
+        }
+
+        if !self.triangles.is_empty() {
+            return 2;
+        }
+
+        if !self.quads.is_empty() {
+            return 1;
+        }
+
+        0
+    }
+
+    fn merge(&mut self, layer: &mut Self) {
+        self.quads.append(&mut layer.quads);
+        self.triangles.append(&mut layer.triangles);
+        self.primitives.append(&mut layer.primitives);
+        self.images.append(&mut layer.images);
+        self.text.append(&mut layer.text);
+    }
+}
+
+impl Default for Layer {
+    fn default() -> Self {
+        Self {
+            bounds: Rectangle::INFINITE,
+            quads: quad::Batch::default(),
+            triangles: triangle::Batch::default(),
+            primitives: primitive::Batch::default(),
+            text: text::Batch::default(),
+            images: image::Batch::default(),
+            pending_meshes: Vec::new(),
+            pending_text: Vec::new(),
+        }
+    }
+}

crates/iced_wgpu/src/lib.rs

@@ -0,0 +1,984 @@
+//! A [`wgpu`] renderer for [Iced].
+//!
+//! ![The native path of the Iced ecosystem](https://github.com/iced-rs/iced/blob/0525d76ff94e828b7b21634fa94a747022001c83/docs/graphs/native.png?raw=true)
+//!
+//! [`wgpu`] supports most modern graphics backends: Vulkan, Metal, DX11, and
+//! DX12 (OpenGL and WebGL are still WIP). Additionally, it will support the
+//! incoming [WebGPU API].
+//!
+//! Currently, `iced_wgpu` supports the following primitives:
+//! - Text, which is rendered using [`glyphon`].
+//! - Quads or rectangles, with rounded borders and a solid background color.
+//! - Clip areas, useful to implement scrollables or hide overflowing content.
+//! - Images and SVG, loaded from memory or the file system.
+//! - Meshes of triangles, useful to draw geometry freely.
+//!
+//! [Iced]: https://github.com/iced-rs/iced
+//! [`wgpu`]: https://github.com/gfx-rs/wgpu-rs
+//! [WebGPU API]: https://gpuweb.github.io/gpuweb/
+//! [`glyphon`]: https://github.com/grovesNL/glyphon
+#![doc(
+    html_logo_url = "https://raw.githubusercontent.com/iced-rs/iced/9ab6923e943f784985e9ef9ca28b10278297225d/docs/logo.svg"
+)]
+#![cfg_attr(docsrs, feature(doc_cfg))]
+#![allow(missing_docs)]
+pub mod layer;
+pub mod primitive;
+pub mod settings;
+pub mod window;
+
+#[cfg(feature = "geometry")]
+pub mod geometry;
+
+mod buffer;
+mod color;
+mod engine;
+mod quad;
+mod text;
+mod triangle;
+
+#[cfg(any(feature = "image", feature = "svg"))]
+#[path = "image/mod.rs"]
+mod image;
+
+#[cfg(not(any(feature = "image", feature = "svg")))]
+#[path = "image/null.rs"]
+mod image;
+
+use buffer::Buffer;
+
+use iced_debug as debug;
+pub use iced_graphics as graphics;
+pub use iced_graphics::core;
+
+pub use wgpu;
+
+pub use engine::Engine;
+pub use layer::Layer;
+pub use primitive::Primitive;
+pub use settings::Settings;
+
+#[cfg(feature = "geometry")]
+pub use geometry::Geometry;
+
+use crate::core::renderer;
+use crate::core::{
+    Background, Color, Font, Pixels, Point, Rectangle, Size, Transformation,
+};
+use crate::graphics::mesh;
+use crate::graphics::text::{Editor, Paragraph};
+use crate::graphics::{Shell, Viewport};
+
+/// A [`wgpu`] graphics renderer for [`iced`].
+///
+/// [`wgpu`]: https://github.com/gfx-rs/wgpu-rs
+/// [`iced`]: https://github.com/iced-rs/iced
+pub struct Renderer {
+    engine: Engine,
+
+    default_font: Font,
+    default_text_size: Pixels,
+    layers: layer::Stack,
+
+    quad: quad::State,
+    triangle: triangle::State,
+    text: text::State,
+    text_viewport: text::Viewport,
+
+    #[cfg(any(feature = "svg", feature = "image"))]
+    image: image::State,
+
+    // TODO: Centralize all the image feature handling
+    #[cfg(any(feature = "svg", feature = "image"))]
+    image_cache: std::cell::RefCell<image::Cache>,
+
+    staging_belt: wgpu::util::StagingBelt,
+}
+
+impl Renderer {
+    pub fn new(
+        engine: Engine,
+        default_font: Font,
+        default_text_size: Pixels,
+    ) -> Self {
+        Self {
+            default_font,
+            default_text_size,
+            layers: layer::Stack::new(),
+
+            quad: quad::State::new(),
+            triangle: triangle::State::new(
+                &engine.device,
+                &engine.triangle_pipeline,
+            ),
+            text: text::State::new(),
+            text_viewport: engine.text_pipeline.create_viewport(&engine.device),
+
+            #[cfg(any(feature = "svg", feature = "image"))]
+            image: image::State::new(),
+
+            #[cfg(any(feature = "svg", feature = "image"))]
+            image_cache: std::cell::RefCell::new(engine.create_image_cache()),
+
+            // TODO: Resize belt smartly (?)
+            // It would be great if the `StagingBelt` API exposed methods
+            // for introspection to detect when a resize may be worth it.
+            staging_belt: wgpu::util::StagingBelt::new(
+                buffer::MAX_WRITE_SIZE as u64,
+            ),
+
+            engine,
+        }
+    }
+
+    fn draw(
+        &mut self,
+        clear_color: Option<Color>,
+        target: &wgpu::TextureView,
+        viewport: &Viewport,
+    ) -> wgpu::CommandEncoder {
+        let mut encoder = self.engine.device.create_command_encoder(
+            &wgpu::CommandEncoderDescriptor {
+                label: Some("iced_wgpu encoder"),
+            },
+        );
+
+        self.prepare(&mut encoder, viewport);
+        self.render(&mut encoder, target, clear_color, viewport);
+
+        self.quad.trim();
+        self.triangle.trim();
+        self.text.trim();
+
+        // TODO: Provide window id (?)
+        self.engine.trim();
+
+        #[cfg(any(feature = "svg", feature = "image"))]
+        {
+            self.image.trim();
+            self.image_cache.borrow_mut().trim();
+        }
+
+        encoder
+    }
+
+    pub fn present(
+        &mut self,
+        clear_color: Option<Color>,
+        _format: wgpu::TextureFormat,
+        frame: &wgpu::TextureView,
+        viewport: &Viewport,
+    ) -> wgpu::SubmissionIndex {
+        let encoder = self.draw(clear_color, frame, viewport);
+
+        self.staging_belt.finish();
+        let submission = self.engine.queue.submit([encoder.finish()]);
+        self.staging_belt.recall();
+        submission
+    }
+
+    /// Renders the current surface to an offscreen buffer.
+    ///
+    /// Returns RGBA bytes of the texture data.
+    pub fn screenshot(
+        &mut self,
+        viewport: &Viewport,
+        background_color: Color,
+    ) -> Vec<u8> {
+        #[derive(Clone, Copy, Debug)]
+        struct BufferDimensions {
+            width: u32,
+            height: u32,
+            unpadded_bytes_per_row: usize,
+            padded_bytes_per_row: usize,
+        }
+
+        impl BufferDimensions {
+            fn new(size: Size<u32>) -> Self {
+                let unpadded_bytes_per_row = size.width as usize * 4; //slice of buffer per row; always RGBA
+                let alignment = wgpu::COPY_BYTES_PER_ROW_ALIGNMENT as usize; //256
+                let padded_bytes_per_row_padding = (alignment
+                    - unpadded_bytes_per_row % alignment)
+                    % alignment;
+                let padded_bytes_per_row =
+                    unpadded_bytes_per_row + padded_bytes_per_row_padding;
+
+                Self {
+                    width: size.width,
+                    height: size.height,
+                    unpadded_bytes_per_row,
+                    padded_bytes_per_row,
+                }
+            }
+        }
+
+        let dimensions = BufferDimensions::new(viewport.physical_size());
+
+        let texture_extent = wgpu::Extent3d {
+            width: dimensions.width,
+            height: dimensions.height,
+            depth_or_array_layers: 1,
+        };
+
+        let texture =
+            self.engine.device.create_texture(&wgpu::TextureDescriptor {
+                label: Some("iced_wgpu.offscreen.source_texture"),
+                size: texture_extent,
+                mip_level_count: 1,
+                sample_count: 1,
+                dimension: wgpu::TextureDimension::D2,
+                format: self.engine.format,
+                usage: wgpu::TextureUsages::RENDER_ATTACHMENT
+                    | wgpu::TextureUsages::COPY_SRC
+                    | wgpu::TextureUsages::TEXTURE_BINDING,
+                view_formats: &[],
+            });
+
+        let view = texture.create_view(&wgpu::TextureViewDescriptor::default());
+
+        let mut encoder = self.draw(Some(background_color), &view, viewport);
+
+        let texture = crate::color::convert(
+            &self.engine.device,
+            &mut encoder,
+            texture,
+            if graphics::color::GAMMA_CORRECTION {
+                wgpu::TextureFormat::Rgba8UnormSrgb
+            } else {
+                wgpu::TextureFormat::Rgba8Unorm
+            },
+        );
+
+        let output_buffer =
+            self.engine.device.create_buffer(&wgpu::BufferDescriptor {
+                label: Some("iced_wgpu.offscreen.output_texture_buffer"),
+                size: (dimensions.padded_bytes_per_row
+                    * dimensions.height as usize) as u64,
+                usage: wgpu::BufferUsages::MAP_READ
+                    | wgpu::BufferUsages::COPY_DST,
+                mapped_at_creation: false,
+            });
+
+        encoder.copy_texture_to_buffer(
+            texture.as_image_copy(),
+            wgpu::TexelCopyBufferInfo {
+                buffer: &output_buffer,
+                layout: wgpu::TexelCopyBufferLayout {
+                    offset: 0,
+                    bytes_per_row: Some(dimensions.padded_bytes_per_row as u32),
+                    rows_per_image: None,
+                },
+            },
+            texture_extent,
+        );
+
+        self.staging_belt.finish();
+        let index = self.engine.queue.submit([encoder.finish()]);
+        self.staging_belt.recall();
+
+        let slice = output_buffer.slice(..);
+        slice.map_async(wgpu::MapMode::Read, |_| {});
+
+        let _ = self.engine.device.poll(wgpu::PollType::Wait {
+            submission_index: Some(index),
+            timeout: None,
+        });
+
+        let mapped_buffer = slice.get_mapped_range();
+
+        mapped_buffer.chunks(dimensions.padded_bytes_per_row).fold(
+            vec![],
+            |mut acc, row| {
+                acc.extend(&row[..dimensions.unpadded_bytes_per_row]);
+                acc
+            },
+        )
+    }
+
+    fn prepare(
+        &mut self,
+        encoder: &mut wgpu::CommandEncoder,
+        viewport: &Viewport,
+    ) {
+        let scale_factor = viewport.scale_factor();
+
+        self.text_viewport
+            .update(&self.engine.queue, viewport.physical_size());
+
+        let physical_bounds = Rectangle::<f32>::from(Rectangle::with_size(
+            viewport.physical_size(),
+        ));
+
+        self.layers.merge();
+
+        for layer in self.layers.iter() {
+            let clip_bounds = layer.bounds * scale_factor;
+
+            if physical_bounds
+                .intersection(&clip_bounds)
+                .and_then(Rectangle::snap)
+                .is_none()
+            {
+                continue;
+            }
+
+            if !layer.quads.is_empty() {
+                let prepare_span = debug::prepare(debug::Primitive::Quad);
+
+                self.quad.prepare(
+                    &self.engine.quad_pipeline,
+                    &self.engine.device,
+                    &mut self.staging_belt,
+                    encoder,
+                    &layer.quads,
+                    viewport.projection(),
+                    scale_factor,
+                );
+
+                prepare_span.finish();
+            }
+
+            if !layer.triangles.is_empty() {
+                let prepare_span = debug::prepare(debug::Primitive::Triangle);
+
+                self.triangle.prepare(
+                    &self.engine.triangle_pipeline,
+                    &self.engine.device,
+                    &mut self.staging_belt,
+                    encoder,
+                    &layer.triangles,
+                    Transformation::scale(scale_factor),
+                    viewport.physical_size(),
+                );
+
+                prepare_span.finish();
+            }
+
+            if !layer.primitives.is_empty() {
+                let prepare_span = debug::prepare(debug::Primitive::Shader);
+
+                let mut primitive_storage = self
+                    .engine
+                    .primitive_storage
+                    .write()
+                    .expect("Write primitive storage");
+
+                for instance in &layer.primitives {
+                    instance.primitive.prepare(
+                        &mut primitive_storage,
+                        &self.engine.device,
+                        &self.engine.queue,
+                        self.engine.format,
+                        &instance.bounds,
+                        viewport,
+                    );
+                }
+
+                prepare_span.finish();
+            }
+
+            #[cfg(any(feature = "svg", feature = "image"))]
+            if !layer.images.is_empty() {
+                let prepare_span = debug::prepare(debug::Primitive::Image);
+
+                self.image.prepare(
+                    &self.engine.image_pipeline,
+                    &self.engine.device,
+                    &mut self.staging_belt,
+                    encoder,
+                    &mut self.image_cache.borrow_mut(),
+                    &layer.images,
+                    viewport.projection(),
+                    scale_factor,
+                );
+
+                prepare_span.finish();
+            }
+
+            if !layer.text.is_empty() {
+                let prepare_span = debug::prepare(debug::Primitive::Text);
+
+                self.text.prepare(
+                    &self.engine.text_pipeline,
+                    &self.engine.device,
+                    &self.engine.queue,
+                    &self.text_viewport,
+                    encoder,
+                    &layer.text,
+                    layer.bounds,
+                    Transformation::scale(scale_factor),
+                );
+
+                prepare_span.finish();
+            }
+        }
+    }
+
+    fn render(
+        &mut self,
+        encoder: &mut wgpu::CommandEncoder,
+        frame: &wgpu::TextureView,
+        clear_color: Option<Color>,
+        viewport: &Viewport,
+    ) {
+        use std::mem::ManuallyDrop;
+
+        let mut render_pass = ManuallyDrop::new(encoder.begin_render_pass(
+            &wgpu::RenderPassDescriptor {
+                label: Some("iced_wgpu render pass"),
+                color_attachments: &[Some(wgpu::RenderPassColorAttachment {
+                    view: frame,
+                    depth_slice: None,
+                    resolve_target: None,
+                    ops: wgpu::Operations {
+                        load: match clear_color {
+                            Some(background_color) => wgpu::LoadOp::Clear({
+                                let [r, g, b, a] =
+                                    graphics::color::pack(background_color)
+                                        .components();
+
+                                wgpu::Color {
+                                    r: f64::from(r),
+                                    g: f64::from(g),
+                                    b: f64::from(b),
+                                    a: f64::from(a),
+                                }
+                            }),
+                            None => wgpu::LoadOp::Load,
+                        },
+                        store: wgpu::StoreOp::Store,
+                    },
+                })],
+                depth_stencil_attachment: None,
+                timestamp_writes: None,
+                occlusion_query_set: None,
+            },
+        ));
+
+        let mut quad_layer = 0;
+        let mut mesh_layer = 0;
+        let mut text_layer = 0;
+
+        #[cfg(any(feature = "svg", feature = "image"))]
+        let mut image_layer = 0;
+
+        let scale_factor = viewport.scale_factor();
+        let physical_bounds = Rectangle::<f32>::from(Rectangle::with_size(
+            viewport.physical_size(),
+        ));
+
+        let scale = Transformation::scale(scale_factor);
+
+        for layer in self.layers.iter() {
+            let Some(physical_bounds) =
+                physical_bounds.intersection(&(layer.bounds * scale_factor))
+            else {
+                continue;
+            };
+
+            let Some(scissor_rect) = physical_bounds.snap() else {
+                continue;
+            };
+
+            if !layer.quads.is_empty() {
+                let render_span = debug::render(debug::Primitive::Quad);
+                self.quad.render(
+                    &self.engine.quad_pipeline,
+                    quad_layer,
+                    scissor_rect,
+                    &layer.quads,
+                    &mut render_pass,
+                );
+                render_span.finish();
+
+                quad_layer += 1;
+            }
+
+            if !layer.triangles.is_empty() {
+                let _ = ManuallyDrop::into_inner(render_pass);
+
+                let render_span = debug::render(debug::Primitive::Triangle);
+                mesh_layer += self.triangle.render(
+                    &self.engine.triangle_pipeline,
+                    encoder,
+                    frame,
+                    mesh_layer,
+                    &layer.triangles,
+                    physical_bounds,
+                    scale,
+                );
+                render_span.finish();
+
+                render_pass = ManuallyDrop::new(encoder.begin_render_pass(
+                    &wgpu::RenderPassDescriptor {
+                        label: Some("iced_wgpu render pass"),
+                        color_attachments: &[Some(
+                            wgpu::RenderPassColorAttachment {
+                                view: frame,
+                                depth_slice: None,
+                                resolve_target: None,
+                                ops: wgpu::Operations {
+                                    load: wgpu::LoadOp::Load,
+                                    store: wgpu::StoreOp::Store,
+                                },
+                            },
+                        )],
+                        depth_stencil_attachment: None,
+                        timestamp_writes: None,
+                        occlusion_query_set: None,
+                    },
+                ));
+            }
+
+            if !layer.primitives.is_empty() {
+                let render_span = debug::render(debug::Primitive::Shader);
+
+                let primitive_storage = self
+                    .engine
+                    .primitive_storage
+                    .read()
+                    .expect("Read primitive storage");
+
+                let mut need_render = Vec::new();
+
+                for instance in &layer.primitives {
+                    let bounds = instance.bounds * scale;
+
+                    if let Some(clip_bounds) = (instance.bounds * scale)
+                        .intersection(&physical_bounds)
+                        .and_then(Rectangle::snap)
+                    {
+                        render_pass.set_viewport(
+                            bounds.x,
+                            bounds.y,
+                            bounds.width,
+                            bounds.height,
+                            0.0,
+                            1.0,
+                        );
+
+                        render_pass.set_scissor_rect(
+                            clip_bounds.x,
+                            clip_bounds.y,
+                            clip_bounds.width,
+                            clip_bounds.height,
+                        );
+
+                        let drawn = instance
+                            .primitive
+                            .draw(&primitive_storage, &mut render_pass);
+
+                        if !drawn {
+                            need_render.push((instance, clip_bounds));
+                        }
+                    }
+                }
+
+                render_pass.set_viewport(
+                    0.0,
+                    0.0,
+                    viewport.physical_width() as f32,
+                    viewport.physical_height() as f32,
+                    0.0,
+                    1.0,
+                );
+
+                render_pass.set_scissor_rect(
+                    0,
+                    0,
+                    viewport.physical_width(),
+                    viewport.physical_height(),
+                );
+
+                if !need_render.is_empty() {
+                    let _ = ManuallyDrop::into_inner(render_pass);
+
+                    for (instance, clip_bounds) in need_render {
+                        instance.primitive.render(
+                            &primitive_storage,
+                            encoder,
+                            frame,
+                            &clip_bounds,
+                        );
+                    }
+
+                    render_pass = ManuallyDrop::new(encoder.begin_render_pass(
+                        &wgpu::RenderPassDescriptor {
+                            label: Some("iced_wgpu render pass"),
+                            color_attachments: &[Some(
+                                wgpu::RenderPassColorAttachment {
+                                    view: frame,
+                                    depth_slice: None,
+                                    resolve_target: None,
+                                    ops: wgpu::Operations {
+                                        load: wgpu::LoadOp::Load,
+                                        store: wgpu::StoreOp::Store,
+                                    },
+                                },
+                            )],
+                            depth_stencil_attachment: None,
+                            timestamp_writes: None,
+                            occlusion_query_set: None,
+                        },
+                    ));
+                }
+
+                render_span.finish();
+            }
+
+            #[cfg(any(feature = "svg", feature = "image"))]
+            if !layer.images.is_empty() {
+                let render_span = debug::render(debug::Primitive::Image);
+                self.image.render(
+                    &self.engine.image_pipeline,
+                    image_layer,
+                    scissor_rect,
+                    &mut render_pass,
+                );
+                render_span.finish();
+
+                image_layer += 1;
+            }
+
+            if !layer.text.is_empty() {
+                let render_span = debug::render(debug::Primitive::Text);
+                text_layer += self.text.render(
+                    &self.engine.text_pipeline,
+                    &self.text_viewport,
+                    text_layer,
+                    &layer.text,
+                    scissor_rect,
+                    &mut render_pass,
+                );
+                render_span.finish();
+            }
+        }
+
+        let _ = ManuallyDrop::into_inner(render_pass);
+
+        debug::layers_rendered(|| {
+            self.layers
+                .iter()
+                .filter(|layer| {
+                    !layer.is_empty()
+                        && physical_bounds
+                            .intersection(&(layer.bounds * scale_factor))
+                            .is_some_and(|viewport| viewport.snap().is_some())
+                })
+                .count()
+        });
+    }
+}
+
+impl core::Renderer for Renderer {
+    fn start_layer(&mut self, bounds: Rectangle) {
+        self.layers.push_clip(bounds);
+    }
+
+    fn end_layer(&mut self) {
+        self.layers.pop_clip();
+    }
+
+    fn start_transformation(&mut self, transformation: Transformation) {
+        self.layers.push_transformation(transformation);
+    }
+
+    fn end_transformation(&mut self) {
+        self.layers.pop_transformation();
+    }
+
+    fn fill_quad(
+        &mut self,
+        quad: core::renderer::Quad,
+        background: impl Into<Background>,
+    ) {
+        let (layer, transformation) = self.layers.current_mut();
+        layer.draw_quad(quad, background.into(), transformation);
+    }
+
+    fn reset(&mut self, new_bounds: Rectangle) {
+        self.layers.reset(new_bounds);
+    }
+
+    fn allocate_image(
+        &mut self,
+        _handle: &core::image::Handle,
+        _callback: impl FnOnce(Result<core::image::Allocation, core::image::Error>)
+        + Send
+        + 'static,
+    ) {
+        #[cfg(feature = "image")]
+        self.image_cache
+            .get_mut()
+            .allocate_image(_handle, _callback);
+    }
+}
+
+impl core::text::Renderer for Renderer {
+    type Font = Font;
+    type Paragraph = Paragraph;
+    type Editor = Editor;
+
+    const ICON_FONT: Font = Font::with_name("Iced-Icons");
+    const CHECKMARK_ICON: char = '\u{f00c}';
+    const ARROW_DOWN_ICON: char = '\u{e800}';
+    const ICED_LOGO: char = '\u{e801}';
+    const SCROLL_UP_ICON: char = '\u{e802}';
+    const SCROLL_DOWN_ICON: char = '\u{e803}';
+    const SCROLL_LEFT_ICON: char = '\u{e804}';
+    const SCROLL_RIGHT_ICON: char = '\u{e805}';
+
+    fn default_font(&self) -> Self::Font {
+        self.default_font
+    }
+
+    fn default_size(&self) -> Pixels {
+        self.default_text_size
+    }
+
+    fn fill_paragraph(
+        &mut self,
+        text: &Self::Paragraph,
+        position: Point,
+        color: Color,
+        clip_bounds: Rectangle,
+    ) {
+        let (layer, transformation) = self.layers.current_mut();
+
+        layer.draw_paragraph(
+            text,
+            position,
+            color,
+            clip_bounds,
+            transformation,
+        );
+    }
+
+    fn fill_editor(
+        &mut self,
+        editor: &Self::Editor,
+        position: Point,
+        color: Color,
+        clip_bounds: Rectangle,
+    ) {
+        let (layer, transformation) = self.layers.current_mut();
+        layer.draw_editor(editor, position, color, clip_bounds, transformation);
+    }
+
+    fn fill_text(
+        &mut self,
+        text: core::Text,
+        position: Point,
+        color: Color,
+        clip_bounds: Rectangle,
+    ) {
+        let (layer, transformation) = self.layers.current_mut();
+        layer.draw_text(text, position, color, clip_bounds, transformation);
+    }
+}
+
+impl graphics::text::Renderer for Renderer {
+    fn fill_raw(&mut self, raw: graphics::text::Raw) {
+        let (layer, transformation) = self.layers.current_mut();
+        layer.draw_text_raw(raw, transformation);
+    }
+}
+
+#[cfg(feature = "image")]
+impl core::image::Renderer for Renderer {
+    type Handle = core::image::Handle;
+
+    fn load_image(
+        &self,
+        handle: &Self::Handle,
+    ) -> Result<core::image::Allocation, core::image::Error> {
+        self.image_cache.borrow_mut().load_image(
+            &self.engine.device,
+            &self.engine.queue,
+            handle,
+        )
+    }
+
+    fn measure_image(&self, handle: &Self::Handle) -> Option<core::Size<u32>> {
+        self.image_cache.borrow_mut().measure_image(handle)
+    }
+
+    fn draw_image(
+        &mut self,
+        image: core::Image,
+        bounds: Rectangle,
+        clip_bounds: Rectangle,
+    ) {
+        let (layer, transformation) = self.layers.current_mut();
+        layer.draw_raster(image, bounds, clip_bounds, transformation);
+    }
+}
+
+#[cfg(feature = "svg")]
+impl core::svg::Renderer for Renderer {
+    fn measure_svg(&self, handle: &core::svg::Handle) -> core::Size<u32> {
+        self.image_cache.borrow_mut().measure_svg(handle)
+    }
+
+    fn draw_svg(
+        &mut self,
+        svg: core::Svg,
+        bounds: Rectangle,
+        clip_bounds: Rectangle,
+    ) {
+        let (layer, transformation) = self.layers.current_mut();
+        layer.draw_svg(svg, bounds, clip_bounds, transformation);
+    }
+}
+
+impl graphics::mesh::Renderer for Renderer {
+    fn draw_mesh(&mut self, mesh: graphics::Mesh) {
+        debug_assert!(
+            !mesh.indices().is_empty(),
+            "Mesh must not have empty indices"
+        );
+
+        debug_assert!(
+            mesh.indices().len().is_multiple_of(3),
+            "Mesh indices length must be a multiple of 3"
+        );
+
+        let (layer, transformation) = self.layers.current_mut();
+        layer.draw_mesh(mesh, transformation);
+    }
+
+    fn draw_mesh_cache(&mut self, cache: mesh::Cache) {
+        let (layer, transformation) = self.layers.current_mut();
+        layer.draw_mesh_cache(cache, transformation);
+    }
+}
+
+#[cfg(feature = "geometry")]
+impl graphics::geometry::Renderer for Renderer {
+    type Geometry = Geometry;
+    type Frame = geometry::Frame;
+
+    fn new_frame(&self, bounds: Rectangle) -> Self::Frame {
+        geometry::Frame::new(bounds)
+    }
+
+    fn draw_geometry(&mut self, geometry: Self::Geometry) {
+        let (layer, transformation) = self.layers.current_mut();
+
+        match geometry {
+            Geometry::Live {
+                meshes,
+                images,
+                text,
+            } => {
+                layer.draw_mesh_group(meshes, transformation);
+
+                for image in images {
+                    layer.draw_image(image, transformation);
+                }
+
+                layer.draw_text_group(text, transformation);
+            }
+            Geometry::Cached(cache) => {
+                if let Some(meshes) = cache.meshes {
+                    layer.draw_mesh_cache(meshes, transformation);
+                }
+
+                if let Some(images) = cache.images {
+                    for image in images.iter().cloned() {
+                        layer.draw_image(image, transformation);
+                    }
+                }
+
+                if let Some(text) = cache.text {
+                    layer.draw_text_cache(text, transformation);
+                }
+            }
+        }
+    }
+}
+
+impl primitive::Renderer for Renderer {
+    fn draw_primitive(&mut self, bounds: Rectangle, primitive: impl Primitive) {
+        let (layer, transformation) = self.layers.current_mut();
+        layer.draw_primitive(bounds, primitive, transformation);
+    }
+}
+
+impl graphics::compositor::Default for crate::Renderer {
+    type Compositor = window::Compositor;
+}
+
+impl renderer::Headless for Renderer {
+    async fn new(
+        default_font: Font,
+        default_text_size: Pixels,
+        backend: Option<&str>,
+    ) -> Option<Self> {
+        if backend.is_some_and(|backend| backend != "wgpu") {
+            return None;
+        }
+
+        let instance = wgpu::Instance::new(&wgpu::InstanceDescriptor {
+            backends: wgpu::Backends::from_env()
+                .unwrap_or(wgpu::Backends::PRIMARY),
+            flags: wgpu::InstanceFlags::empty(),
+            ..wgpu::InstanceDescriptor::default()
+        });
+
+        let adapter = instance
+            .request_adapter(&wgpu::RequestAdapterOptions {
+                power_preference: wgpu::PowerPreference::HighPerformance,
+                force_fallback_adapter: false,
+                compatible_surface: None,
+            })
+            .await
+            .ok()?;
+
+        let (device, queue) = adapter
+            .request_device(&wgpu::DeviceDescriptor {
+                label: Some("iced_wgpu [headless]"),
+                required_features: wgpu::Features::empty(),
+                required_limits: wgpu::Limits {
+                    max_bind_groups: 2,
+                    ..wgpu::Limits::default()
+                },
+                memory_hints: wgpu::MemoryHints::MemoryUsage,
+                trace: wgpu::Trace::Off,
+                experimental_features: wgpu::ExperimentalFeatures::disabled(),
+            })
+            .await
+            .ok()?;
+
+        let engine = Engine::new(
+            &adapter,
+            device,
+            queue,
+            if graphics::color::GAMMA_CORRECTION {
+                wgpu::TextureFormat::Rgba8UnormSrgb
+            } else {
+                wgpu::TextureFormat::Rgba8Unorm
+            },
+            Some(graphics::Antialiasing::MSAAx4),
+            Shell::headless(),
+        );
+
+        Some(Self::new(engine, default_font, default_text_size))
+    }
+
+    fn name(&self) -> String {
+        "wgpu".to_owned()
+    }
+
+    fn screenshot(
+        &mut self,
+        size: Size<u32>,
+        scale_factor: f32,
+        background_color: Color,
+    ) -> Vec<u8> {
+        self.screenshot(
+            &Viewport::with_physical_size(size, scale_factor),
+            background_color,
+        )
+    }
+}

crates/iced_wgpu/src/primitive.rs

@@ -0,0 +1,242 @@
+//! Draw custom primitives.
+use crate::core::{self, Rectangle};
+use crate::graphics::Viewport;
+use crate::graphics::futures::{MaybeSend, MaybeSync};
+
+use rustc_hash::FxHashMap;
+use std::any::{Any, TypeId};
+use std::fmt::Debug;
+
+/// A batch of primitives.
+pub type Batch = Vec<Instance>;
+
+/// A set of methods which allows a [`Primitive`] to be rendered.
+pub trait Primitive: Debug + MaybeSend + MaybeSync + 'static {
+    /// The shared renderer of this [`Primitive`].
+    ///
+    /// Normally, this will contain a bunch of [`wgpu`] state; like
+    /// a rendering pipeline, buffers, and textures.
+    ///
+    /// All instances of this [`Primitive`] type will share the same
+    /// [`Renderer`].
+    type Pipeline: Pipeline + MaybeSend + MaybeSync;
+
+    /// Processes the [`Primitive`], allowing for GPU buffer allocation.
+    fn prepare(
+        &self,
+        pipeline: &mut Self::Pipeline,
+        device: &wgpu::Device,
+        queue: &wgpu::Queue,
+        bounds: &Rectangle,
+        viewport: &Viewport,
+    );
+
+    /// Draws the [`Primitive`] in the given [`wgpu::RenderPass`].
+    ///
+    /// When possible, this should be implemented over [`render`](Self::render)
+    /// since reusing the existing render pass should be considerably more
+    /// efficient than issuing a new one.
+    ///
+    /// The viewport and scissor rect of the render pass provided is set
+    /// to the bounds and clip bounds of the [`Primitive`], respectively.
+    ///
+    /// If you have complex composition needs, then you can leverage
+    /// [`render`](Self::render) by returning `false` here.
+    ///
+    /// By default, it does nothing and returns `false`.
+    fn draw(
+        &self,
+        _pipeline: &Self::Pipeline,
+        _render_pass: &mut wgpu::RenderPass<'_>,
+    ) -> bool {
+        false
+    }
+
+    /// Renders the [`Primitive`], using the given [`wgpu::CommandEncoder`].
+    ///
+    /// This will only be called if [`draw`](Self::draw) returns `false`.
+    ///
+    /// By default, it does nothing.
+    fn render(
+        &self,
+        _pipeline: &Self::Pipeline,
+        _encoder: &mut wgpu::CommandEncoder,
+        _target: &wgpu::TextureView,
+        _clip_bounds: &Rectangle<u32>,
+    ) {
+    }
+}
+
+/// The pipeline of a graphics [`Primitive`].
+pub trait Pipeline: Any + MaybeSend + MaybeSync {
+    /// Creates the [`Pipeline`] of a [`Primitive`].
+    ///
+    /// This will only be called once, when the first [`Primitive`] with this kind
+    /// of [`Pipeline`] is encountered.
+    fn new(
+        device: &wgpu::Device,
+        queue: &wgpu::Queue,
+        format: wgpu::TextureFormat,
+    ) -> Self
+    where
+        Self: Sized;
+
+    /// Trims any cached data in the [`Pipeline`].
+    ///
+    /// This will normally be called at the end of a frame.
+    fn trim(&mut self) {}
+}
+
+pub(crate) trait Stored:
+    Debug + MaybeSend + MaybeSync + 'static
+{
+    fn prepare(
+        &self,
+        storage: &mut Storage,
+        device: &wgpu::Device,
+        queue: &wgpu::Queue,
+        format: wgpu::TextureFormat,
+        bounds: &Rectangle,
+        viewport: &Viewport,
+    );
+
+    fn draw(
+        &self,
+        storage: &Storage,
+        render_pass: &mut wgpu::RenderPass<'_>,
+    ) -> bool;
+
+    fn render(
+        &self,
+        storage: &Storage,
+        encoder: &mut wgpu::CommandEncoder,
+        target: &wgpu::TextureView,
+        clip_bounds: &Rectangle<u32>,
+    );
+}
+
+#[derive(Debug)]
+struct BlackBox<P: Primitive> {
+    primitive: P,
+}
+
+impl<P: Primitive> Stored for BlackBox<P> {
+    fn prepare(
+        &self,
+        storage: &mut Storage,
+        device: &wgpu::Device,
+        queue: &wgpu::Queue,
+        format: wgpu::TextureFormat,
+        bounds: &Rectangle,
+        viewport: &Viewport,
+    ) {
+        if !storage.has::<P>() {
+            storage.store::<P, _>(P::Pipeline::new(device, queue, format));
+        }
+
+        let renderer = storage
+            .get_mut::<P>()
+            .expect("renderer should be initialized")
+            .downcast_mut::<P::Pipeline>()
+            .expect("renderer should have the proper type");
+
+        self.primitive
+            .prepare(renderer, device, queue, bounds, viewport);
+    }
+
+    fn draw(
+        &self,
+        storage: &Storage,
+        render_pass: &mut wgpu::RenderPass<'_>,
+    ) -> bool {
+        let renderer = storage
+            .get::<P>()
+            .expect("renderer should be initialized")
+            .downcast_ref::<P::Pipeline>()
+            .expect("renderer should have the proper type");
+
+        self.primitive.draw(renderer, render_pass)
+    }
+
+    fn render(
+        &self,
+        storage: &Storage,
+        encoder: &mut wgpu::CommandEncoder,
+        target: &wgpu::TextureView,
+        clip_bounds: &Rectangle<u32>,
+    ) {
+        let renderer = storage
+            .get::<P>()
+            .expect("renderer should be initialized")
+            .downcast_ref::<P::Pipeline>()
+            .expect("renderer should have the proper type");
+
+        self.primitive
+            .render(renderer, encoder, target, clip_bounds);
+    }
+}
+
+#[derive(Debug)]
+/// An instance of a specific [`Primitive`].
+pub struct Instance {
+    /// The bounds of the [`Instance`].
+    pub(crate) bounds: Rectangle,
+
+    /// The [`Primitive`] to render.
+    pub(crate) primitive: Box<dyn Stored>,
+}
+
+impl Instance {
+    /// Creates a new [`Instance`] with the given [`Primitive`].
+    pub fn new(bounds: Rectangle, primitive: impl Primitive) -> Self {
+        Instance {
+            bounds,
+            primitive: Box::new(BlackBox { primitive }),
+        }
+    }
+}
+
+/// A renderer than can draw custom primitives.
+pub trait Renderer: core::Renderer {
+    /// Draws a custom primitive.
+    fn draw_primitive(&mut self, bounds: Rectangle, primitive: impl Primitive);
+}
+
+/// Stores custom, user-provided types.
+#[derive(Default)]
+pub struct Storage {
+    pipelines: FxHashMap<TypeId, Box<dyn Pipeline>>,
+}
+
+impl Storage {
+    /// Returns `true` if `Storage` contains a type `T`.
+    pub fn has<T: 'static>(&self) -> bool {
+        self.pipelines.contains_key(&TypeId::of::<T>())
+    }
+
+    /// Inserts the data `T` in to [`Storage`].
+    pub fn store<T: 'static, P: Pipeline>(&mut self, pipeline: P) {
+        let _ = self.pipelines.insert(TypeId::of::<T>(), Box::new(pipeline));
+    }
+
+    /// Returns a reference to the data with type `T` if it exists in [`Storage`].
+    pub fn get<T: 'static>(&self) -> Option<&dyn Any> {
+        self.pipelines
+            .get(&TypeId::of::<T>())
+            .map(|pipeline| pipeline.as_ref() as &dyn Any)
+    }
+
+    /// Returns a mutable reference to the data with type `T` if it exists in [`Storage`].
+    pub fn get_mut<T: 'static>(&mut self) -> Option<&mut dyn Any> {
+        self.pipelines
+            .get_mut(&TypeId::of::<T>())
+            .map(|pipeline| pipeline.as_mut() as &mut dyn Any)
+    }
+
+    /// Trims the cache of all the pipelines in the [`Storage`].
+    pub fn trim(&mut self) {
+        for pipeline in self.pipelines.values_mut() {
+            pipeline.trim();
+        }
+    }
+}

crates/iced_wgpu/src/quad.rs

@@ -0,0 +1,362 @@
+mod gradient;
+mod solid;
+
+use gradient::Gradient;
+use solid::Solid;
+
+use crate::core::{Background, Rectangle, Transformation};
+use crate::graphics;
+use crate::graphics::color;
+
+use bytemuck::{Pod, Zeroable};
+
+use std::mem;
+
+const INITIAL_INSTANCES: usize = 2_000;
+
+/// The properties of a quad.
+#[derive(Clone, Copy, Debug, Pod, Zeroable)]
+#[repr(C)]
+pub struct Quad {
+    /// The position of the [`Quad`].
+    pub position: [f32; 2],
+
+    /// The size of the [`Quad`].
+    pub size: [f32; 2],
+
+    /// The border color of the [`Quad`], in __linear RGB__.
+    pub border_color: color::Packed,
+
+    /// The border radii of the [`Quad`].
+    pub border_radius: [f32; 4],
+
+    /// The border width of the [`Quad`].
+    pub border_width: f32,
+
+    /// The shadow color of the [`Quad`].
+    pub shadow_color: color::Packed,
+
+    /// The shadow offset of the [`Quad`].
+    pub shadow_offset: [f32; 2],
+
+    /// The shadow blur radius of the [`Quad`].
+    pub shadow_blur_radius: f32,
+
+    /// Whether the [`Quad`] should be snapped to the pixel grid.
+    pub snap: u32,
+}
+
+#[derive(Debug, Clone)]
+pub struct Pipeline {
+    solid: solid::Pipeline,
+    gradient: gradient::Pipeline,
+    constant_layout: wgpu::BindGroupLayout,
+}
+
+#[derive(Default)]
+pub struct State {
+    layers: Vec<Layer>,
+    prepare_layer: usize,
+}
+
+impl State {
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    pub fn prepare(
+        &mut self,
+        pipeline: &Pipeline,
+        device: &wgpu::Device,
+        belt: &mut wgpu::util::StagingBelt,
+        encoder: &mut wgpu::CommandEncoder,
+        quads: &Batch,
+        transformation: Transformation,
+        scale: f32,
+    ) {
+        if self.layers.len() <= self.prepare_layer {
+            self.layers
+                .push(Layer::new(device, &pipeline.constant_layout));
+        }
+
+        let layer = &mut self.layers[self.prepare_layer];
+        layer.prepare(device, encoder, belt, quads, transformation, scale);
+
+        self.prepare_layer += 1;
+    }
+
+    pub fn render<'a>(
+        &'a self,
+        pipeline: &'a Pipeline,
+        layer: usize,
+        bounds: Rectangle<u32>,
+        quads: &Batch,
+        render_pass: &mut wgpu::RenderPass<'a>,
+    ) {
+        if let Some(layer) = self.layers.get(layer) {
+            render_pass.set_scissor_rect(
+                bounds.x,
+                bounds.y,
+                bounds.width,
+                bounds.height,
+            );
+
+            let mut solid_offset = 0;
+            let mut gradient_offset = 0;
+
+            for (kind, count) in &quads.order {
+                match kind {
+                    Kind::Solid => {
+                        pipeline.solid.render(
+                            render_pass,
+                            &layer.constants,
+                            &layer.solid,
+                            solid_offset..(solid_offset + count),
+                        );
+
+                        solid_offset += count;
+                    }
+                    Kind::Gradient => {
+                        pipeline.gradient.render(
+                            render_pass,
+                            &layer.constants,
+                            &layer.gradient,
+                            gradient_offset..(gradient_offset + count),
+                        );
+
+                        gradient_offset += count;
+                    }
+                }
+            }
+        }
+    }
+
+    pub fn trim(&mut self) {
+        self.prepare_layer = 0;
+    }
+}
+
+impl Pipeline {
+    pub fn new(device: &wgpu::Device, format: wgpu::TextureFormat) -> Pipeline {
+        let constant_layout =
+            device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
+                label: Some("iced_wgpu::quad uniforms layout"),
+                entries: &[wgpu::BindGroupLayoutEntry {
+                    binding: 0,
+                    visibility: wgpu::ShaderStages::VERTEX,
+                    ty: wgpu::BindingType::Buffer {
+                        ty: wgpu::BufferBindingType::Uniform,
+                        has_dynamic_offset: false,
+                        min_binding_size: wgpu::BufferSize::new(
+                            mem::size_of::<Uniforms>() as wgpu::BufferAddress,
+                        ),
+                    },
+                    count: None,
+                }],
+            });
+
+        Self {
+            solid: solid::Pipeline::new(device, format, &constant_layout),
+            gradient: gradient::Pipeline::new(device, format, &constant_layout),
+            constant_layout,
+        }
+    }
+}
+
+#[derive(Debug)]
+pub struct Layer {
+    constants: wgpu::BindGroup,
+    constants_buffer: wgpu::Buffer,
+    solid: solid::Layer,
+    gradient: gradient::Layer,
+}
+
+impl Layer {
+    pub fn new(
+        device: &wgpu::Device,
+        constant_layout: &wgpu::BindGroupLayout,
+    ) -> Self {
+        let constants_buffer = device.create_buffer(&wgpu::BufferDescriptor {
+            label: Some("iced_wgpu::quad uniforms buffer"),
+            size: mem::size_of::<Uniforms>() as wgpu::BufferAddress,
+            usage: wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
+            mapped_at_creation: false,
+        });
+
+        let constants = device.create_bind_group(&wgpu::BindGroupDescriptor {
+            label: Some("iced_wgpu::quad uniforms bind group"),
+            layout: constant_layout,
+            entries: &[wgpu::BindGroupEntry {
+                binding: 0,
+                resource: constants_buffer.as_entire_binding(),
+            }],
+        });
+
+        Self {
+            constants,
+            constants_buffer,
+            solid: solid::Layer::new(device),
+            gradient: gradient::Layer::new(device),
+        }
+    }
+
+    pub fn prepare(
+        &mut self,
+        device: &wgpu::Device,
+        encoder: &mut wgpu::CommandEncoder,
+        belt: &mut wgpu::util::StagingBelt,
+        quads: &Batch,
+        transformation: Transformation,
+        scale: f32,
+    ) {
+        self.update(device, encoder, belt, transformation, scale);
+
+        if !quads.solids.is_empty() {
+            self.solid.prepare(device, encoder, belt, &quads.solids);
+        }
+
+        if !quads.gradients.is_empty() {
+            self.gradient
+                .prepare(device, encoder, belt, &quads.gradients);
+        }
+    }
+
+    pub fn update(
+        &mut self,
+        device: &wgpu::Device,
+        encoder: &mut wgpu::CommandEncoder,
+        belt: &mut wgpu::util::StagingBelt,
+        transformation: Transformation,
+        scale: f32,
+    ) {
+        let uniforms = Uniforms::new(transformation, scale);
+        let bytes = bytemuck::bytes_of(&uniforms);
+
+        belt.write_buffer(
+            encoder,
+            &self.constants_buffer,
+            0,
+            (bytes.len() as u64).try_into().expect("Sized uniforms"),
+            device,
+        )
+        .copy_from_slice(bytes);
+    }
+}
+
+/// A group of [`Quad`]s rendered together.
+#[derive(Default, Debug)]
+pub struct Batch {
+    /// The solid quads of the [`Layer`].
+    solids: Vec<Solid>,
+
+    /// The gradient quads of the [`Layer`].
+    gradients: Vec<Gradient>,
+
+    /// The quad order of the [`Layer`].
+    order: Order,
+}
+
+/// The quad order of a [`Layer`]; stored as a tuple of the quad type & its count.
+type Order = Vec<(Kind, usize)>;
+
+impl Batch {
+    /// Returns true if there are no quads of any type in [`Quads`].
+    pub fn is_empty(&self) -> bool {
+        self.solids.is_empty() && self.gradients.is_empty()
+    }
+
+    /// Adds a [`Quad`] with the provided `Background` type to the quad [`Layer`].
+    pub fn add(&mut self, quad: Quad, background: &Background) {
+        let kind = match background {
+            Background::Color(color) => {
+                self.solids.push(Solid {
+                    color: color::pack(*color),
+                    quad,
+                });
+
+                Kind::Solid
+            }
+            Background::Gradient(gradient) => {
+                self.gradients.push(Gradient {
+                    gradient: graphics::gradient::pack(
+                        gradient,
+                        Rectangle::new(quad.position.into(), quad.size.into()),
+                    ),
+                    quad,
+                });
+
+                Kind::Gradient
+            }
+        };
+
+        match self.order.last_mut() {
+            Some((last_kind, count)) if kind == *last_kind => {
+                *count += 1;
+            }
+            _ => {
+                self.order.push((kind, 1));
+            }
+        }
+    }
+
+    pub fn clear(&mut self) {
+        self.solids.clear();
+        self.gradients.clear();
+        self.order.clear();
+    }
+
+    pub fn append(&mut self, batch: &mut Batch) {
+        self.solids.append(&mut batch.solids);
+        self.gradients.append(&mut batch.gradients);
+        self.order.append(&mut batch.order);
+    }
+}
+
+#[derive(Debug, Copy, Clone, PartialEq, Eq)]
+/// The kind of a quad.
+enum Kind {
+    /// A solid quad
+    Solid,
+    /// A gradient quad
+    Gradient,
+}
+
+fn color_target_state(
+    format: wgpu::TextureFormat,
+) -> [Option<wgpu::ColorTargetState>; 1] {
+    [Some(wgpu::ColorTargetState {
+        format,
+        blend: Some(wgpu::BlendState::PREMULTIPLIED_ALPHA_BLENDING),
+        write_mask: wgpu::ColorWrites::ALL,
+    })]
+}
+
+#[repr(C)]
+#[derive(Debug, Clone, Copy, bytemuck::Zeroable, bytemuck::Pod)]
+struct Uniforms {
+    transform: [f32; 16],
+    scale: f32,
+    // Uniforms must be aligned to their largest member,
+    // this uses a mat4x4<f32> which aligns to 16, so align to that
+    _padding: [f32; 3],
+}
+
+impl Uniforms {
+    fn new(transformation: Transformation, scale: f32) -> Uniforms {
+        Self {
+            transform: *transformation.as_ref(),
+            scale,
+            _padding: [0.0; 3],
+        }
+    }
+}
+
+impl Default for Uniforms {
+    fn default() -> Self {
+        Self {
+            transform: *Transformation::IDENTITY.as_ref(),
+            scale: 1.0,
+            _padding: [0.0; 3],
+        }
+    }
+}

crates/iced_wgpu/src/quad/gradient.rs

@@ -0,0 +1,187 @@
+use crate::Buffer;
+use crate::graphics::gradient;
+use crate::quad::{self, Quad};
+
+use bytemuck::{Pod, Zeroable};
+use std::ops::Range;
+
+/// A quad filled with interpolated colors.
+#[derive(Clone, Copy, Debug)]
+#[repr(C)]
+pub struct Gradient {
+    /// The background gradient data of the quad.
+    pub gradient: gradient::Packed,
+
+    /// The [`Quad`] data of the [`Gradient`].
+    pub quad: Quad,
+}
+
+#[allow(unsafe_code)]
+unsafe impl Pod for Gradient {}
+
+#[allow(unsafe_code)]
+unsafe impl Zeroable for Gradient {}
+
+#[derive(Debug)]
+pub struct Layer {
+    instances: Buffer<Gradient>,
+    instance_count: usize,
+}
+
+impl Layer {
+    pub fn new(device: &wgpu::Device) -> Self {
+        let instances = Buffer::new(
+            device,
+            "iced_wgpu.quad.gradient.buffer",
+            quad::INITIAL_INSTANCES,
+            wgpu::BufferUsages::VERTEX | wgpu::BufferUsages::COPY_DST,
+        );
+
+        Self {
+            instances,
+            instance_count: 0,
+        }
+    }
+
+    pub fn prepare(
+        &mut self,
+        device: &wgpu::Device,
+        encoder: &mut wgpu::CommandEncoder,
+        belt: &mut wgpu::util::StagingBelt,
+        instances: &[Gradient],
+    ) {
+        let _ = self.instances.resize(device, instances.len());
+        let _ = self.instances.write(device, encoder, belt, 0, instances);
+
+        self.instance_count = instances.len();
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct Pipeline {
+    #[cfg(not(target_arch = "wasm32"))]
+    pipeline: wgpu::RenderPipeline,
+}
+
+impl Pipeline {
+    #[allow(unused_variables)]
+    pub fn new(
+        device: &wgpu::Device,
+        format: wgpu::TextureFormat,
+        constants_layout: &wgpu::BindGroupLayout,
+    ) -> Self {
+        #[cfg(not(target_arch = "wasm32"))]
+        {
+            let layout = device.create_pipeline_layout(
+                &wgpu::PipelineLayoutDescriptor {
+                    label: Some("iced_wgpu.quad.gradient.pipeline"),
+                    push_constant_ranges: &[],
+                    bind_group_layouts: &[constants_layout],
+                },
+            );
+
+            let shader =
+                device.create_shader_module(wgpu::ShaderModuleDescriptor {
+                    label: Some("iced_wgpu.quad.gradient.shader"),
+                    source: wgpu::ShaderSource::Wgsl(
+                        std::borrow::Cow::Borrowed(concat!(
+                            include_str!("../shader/quad.wgsl"),
+                            "\n",
+                            include_str!("../shader/vertex.wgsl"),
+                            "\n",
+                            include_str!("../shader/quad/gradient.wgsl"),
+                            "\n",
+                            include_str!("../shader/color.wgsl"),
+                            "\n",
+                            include_str!("../shader/color/linear_rgb.wgsl")
+                        )),
+                    ),
+                });
+
+            let pipeline = device.create_render_pipeline(
+                &wgpu::RenderPipelineDescriptor {
+                    label: Some("iced_wgpu.quad.gradient.pipeline"),
+                    layout: Some(&layout),
+                    vertex: wgpu::VertexState {
+                        module: &shader,
+                        entry_point: Some("gradient_vs_main"),
+                        buffers: &[wgpu::VertexBufferLayout {
+                            array_stride: std::mem::size_of::<Gradient>()
+                                as u64,
+                            step_mode: wgpu::VertexStepMode::Instance,
+                            attributes: &wgpu::vertex_attr_array!(
+                                // Colors 1-2
+                                0 => Uint32x4,
+                                // Colors 3-4
+                                1 => Uint32x4,
+                                // Colors 5-6
+                                2 => Uint32x4,
+                                // Colors 7-8
+                                3 => Uint32x4,
+                                // Offsets 1-8
+                                4 => Uint32x4,
+                                // Direction
+                                5 => Float32x4,
+                                // Position & Scale
+                                6 => Float32x4,
+                                // Border color
+                                7 => Float32x4,
+                                // Border radius
+                                8 => Float32x4,
+                                // Border width
+                                9 => Float32,
+                                // Snap
+                                10 => Uint32,
+                            ),
+                        }],
+                        compilation_options:
+                            wgpu::PipelineCompilationOptions::default(),
+                    },
+                    fragment: Some(wgpu::FragmentState {
+                        module: &shader,
+                        entry_point: Some("gradient_fs_main"),
+                        targets: &quad::color_target_state(format),
+                        compilation_options:
+                            wgpu::PipelineCompilationOptions::default(),
+                    }),
+                    primitive: wgpu::PrimitiveState {
+                        topology: wgpu::PrimitiveTopology::TriangleList,
+                        front_face: wgpu::FrontFace::Cw,
+                        ..Default::default()
+                    },
+                    depth_stencil: None,
+                    multisample: wgpu::MultisampleState {
+                        count: 1,
+                        mask: !0,
+                        alpha_to_coverage_enabled: false,
+                    },
+                    multiview: None,
+                    cache: None,
+                },
+            );
+
+            Self { pipeline }
+        }
+
+        #[cfg(target_arch = "wasm32")]
+        Self {}
+    }
+
+    #[allow(unused_variables)]
+    pub fn render<'a>(
+        &'a self,
+        render_pass: &mut wgpu::RenderPass<'a>,
+        constants: &'a wgpu::BindGroup,
+        layer: &'a Layer,
+        range: Range<usize>,
+    ) {
+        #[cfg(not(target_arch = "wasm32"))]
+        {
+            render_pass.set_pipeline(&self.pipeline);
+            render_pass.set_bind_group(0, constants, &[]);
+            render_pass.set_vertex_buffer(0, layer.instances.slice(..));
+
+            render_pass.draw(0..6, range.start as u32..range.end as u32);
+        }
+    }
+}

crates/iced_wgpu/src/quad/solid.rs

@@ -0,0 +1,162 @@
+use crate::Buffer;
+use crate::graphics::color;
+use crate::quad::{self, Quad};
+
+use bytemuck::{Pod, Zeroable};
+use std::ops::Range;
+
+/// A quad filled with a solid color.
+#[derive(Clone, Copy, Debug, Pod, Zeroable)]
+#[repr(C)]
+pub struct Solid {
+    /// The background color data of the quad.
+    pub color: color::Packed,
+
+    /// The [`Quad`] data of the [`Solid`].
+    pub quad: Quad,
+}
+
+#[derive(Debug)]
+pub struct Layer {
+    instances: Buffer<Solid>,
+    instance_count: usize,
+}
+
+impl Layer {
+    pub fn new(device: &wgpu::Device) -> Self {
+        let instances = Buffer::new(
+            device,
+            "iced_wgpu.quad.solid.buffer",
+            quad::INITIAL_INSTANCES,
+            wgpu::BufferUsages::VERTEX | wgpu::BufferUsages::COPY_DST,
+        );
+
+        Self {
+            instances,
+            instance_count: 0,
+        }
+    }
+
+    pub fn prepare(
+        &mut self,
+        device: &wgpu::Device,
+        encoder: &mut wgpu::CommandEncoder,
+        belt: &mut wgpu::util::StagingBelt,
+        instances: &[Solid],
+    ) {
+        let _ = self.instances.resize(device, instances.len());
+        let _ = self.instances.write(device, encoder, belt, 0, instances);
+
+        self.instance_count = instances.len();
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct Pipeline {
+    pipeline: wgpu::RenderPipeline,
+}
+
+impl Pipeline {
+    pub fn new(
+        device: &wgpu::Device,
+        format: wgpu::TextureFormat,
+        constants_layout: &wgpu::BindGroupLayout,
+    ) -> Self {
+        let layout =
+            device.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
+                label: Some("iced_wgpu.quad.solid.pipeline"),
+                push_constant_ranges: &[],
+                bind_group_layouts: &[constants_layout],
+            });
+
+        let shader =
+            device.create_shader_module(wgpu::ShaderModuleDescriptor {
+                label: Some("iced_wgpu.quad.solid.shader"),
+                source: wgpu::ShaderSource::Wgsl(std::borrow::Cow::Borrowed(
+                    concat!(
+                        include_str!("../shader/color.wgsl"),
+                        "\n",
+                        include_str!("../shader/quad.wgsl"),
+                        "\n",
+                        include_str!("../shader/vertex.wgsl"),
+                        "\n",
+                        include_str!("../shader/quad/solid.wgsl"),
+                    ),
+                )),
+            });
+
+        let pipeline =
+            device.create_render_pipeline(&wgpu::RenderPipelineDescriptor {
+                label: Some("iced_wgpu.quad.solid.pipeline"),
+                layout: Some(&layout),
+                vertex: wgpu::VertexState {
+                    module: &shader,
+                    entry_point: Some("solid_vs_main"),
+                    buffers: &[wgpu::VertexBufferLayout {
+                        array_stride: std::mem::size_of::<Solid>() as u64,
+                        step_mode: wgpu::VertexStepMode::Instance,
+                        attributes: &wgpu::vertex_attr_array!(
+                            // Color
+                            0 => Float32x4,
+                            // Position
+                            1 => Float32x2,
+                            // Size
+                            2 => Float32x2,
+                            // Border color
+                            3 => Float32x4,
+                            // Border radius
+                            4 => Float32x4,
+                            // Border width
+                            5 => Float32,
+                            // Shadow color
+                            6 => Float32x4,
+                            // Shadow offset
+                            7 => Float32x2,
+                            // Shadow blur radius
+                            8 => Float32,
+                            // Snap
+                            9 => Uint32,
+                        ),
+                    }],
+                    compilation_options:
+                        wgpu::PipelineCompilationOptions::default(),
+                },
+                fragment: Some(wgpu::FragmentState {
+                    module: &shader,
+                    entry_point: Some("solid_fs_main"),
+                    targets: &quad::color_target_state(format),
+                    compilation_options:
+                        wgpu::PipelineCompilationOptions::default(),
+                }),
+                primitive: wgpu::PrimitiveState {
+                    topology: wgpu::PrimitiveTopology::TriangleList,
+                    front_face: wgpu::FrontFace::Cw,
+                    ..Default::default()
+                },
+                depth_stencil: None,
+                multisample: wgpu::MultisampleState {
+                    count: 1,
+                    mask: !0,
+                    alpha_to_coverage_enabled: false,
+                },
+                multiview: None,
+                cache: None,
+            });
+
+        Self { pipeline }
+    }
+
+    pub fn render<'a>(
+        &'a self,
+        render_pass: &mut wgpu::RenderPass<'a>,
+        constants: &'a wgpu::BindGroup,
+        layer: &'a Layer,
+        range: Range<usize>,
+    ) {
+        render_pass.set_pipeline(&self.pipeline);
+        render_pass.set_bind_group(0, constants, &[]);
+        render_pass.set_vertex_buffer(0, layer.instances.slice(..));
+
+        render_pass.draw(0..6, range.start as u32..range.end as u32);
+    }
+}

crates/iced_wgpu/src/settings.rs

@@ -0,0 +1,84 @@
+//! Configure a renderer.
+use crate::core::{Font, Pixels};
+use crate::graphics::{self, Antialiasing};
+
+/// The settings of a [`Renderer`].
+///
+/// [`Renderer`]: crate::Renderer
+#[derive(Debug, Clone, Copy, PartialEq)]
+pub struct Settings {
+    /// The present mode of the [`Renderer`].
+    ///
+    /// [`Renderer`]: crate::Renderer
+    pub present_mode: wgpu::PresentMode,
+
+    /// The graphics backends to use.
+    pub backends: wgpu::Backends,
+
+    /// The default [`Font`] to use.
+    pub default_font: Font,
+
+    /// The default size of text.
+    ///
+    /// By default, it will be set to `16.0`.
+    pub default_text_size: Pixels,
+
+    /// The antialiasing strategy that will be used for triangle primitives.
+    ///
+    /// By default, it is `None`.
+    pub antialiasing: Option<Antialiasing>,
+}
+
+impl Default for Settings {
+    fn default() -> Settings {
+        Settings {
+            present_mode: wgpu::PresentMode::AutoVsync,
+            backends: wgpu::Backends::all(),
+            default_font: Font::default(),
+            default_text_size: Pixels(16.0),
+            antialiasing: None,
+        }
+    }
+}
+
+impl From<graphics::Settings> for Settings {
+    fn from(settings: graphics::Settings) -> Self {
+        Self {
+            present_mode: if settings.vsync {
+                wgpu::PresentMode::AutoVsync
+            } else {
+                wgpu::PresentMode::AutoNoVsync
+            },
+            default_font: settings.default_font,
+            default_text_size: settings.default_text_size,
+            antialiasing: settings.antialiasing,
+            ..Settings::default()
+        }
+    }
+}
+
+/// Obtains a [`wgpu::PresentMode`] from the current environment
+/// configuration, if set.
+///
+/// The value returned by this function can be changed by setting
+/// the `ICED_PRESENT_MODE` env variable. The possible values are:
+///
+/// - `vsync` → [`wgpu::PresentMode::AutoVsync`]
+/// - `no_vsync` → [`wgpu::PresentMode::AutoNoVsync`]
+/// - `immediate` → [`wgpu::PresentMode::Immediate`]
+/// - `fifo` → [`wgpu::PresentMode::Fifo`]
+/// - `fifo_relaxed` → [`wgpu::PresentMode::FifoRelaxed`]
+/// - `mailbox` → [`wgpu::PresentMode::Mailbox`]
+pub fn present_mode_from_env() -> Option<wgpu::PresentMode> {
+    let present_mode = std::env::var("ICED_PRESENT_MODE").ok()?;
+
+    match present_mode.to_lowercase().as_str() {
+        "vsync" => Some(wgpu::PresentMode::AutoVsync),
+        "no_vsync" => Some(wgpu::PresentMode::AutoNoVsync),
+        "immediate" => Some(wgpu::PresentMode::Immediate),
+        "fifo" => Some(wgpu::PresentMode::Fifo),
+        "fifo_relaxed" => Some(wgpu::PresentMode::FifoRelaxed),
+        "mailbox" => Some(wgpu::PresentMode::Mailbox),
+        _ => None,
+    }
+}

crates/iced_wgpu/src/shader/blit.wgsl

@@ -0,0 +1,37 @@
+var<private> uvs: array<vec2<f32>, 6> = array<vec2<f32>, 6>(
+    vec2<f32>(0.0, 0.0),
+    vec2<f32>(1.0, 0.0),
+    vec2<f32>(1.0, 1.0),
+    vec2<f32>(0.0, 0.0),
+    vec2<f32>(0.0, 1.0),
+    vec2<f32>(1.0, 1.0)
+);
+
+@group(0) @binding(0) var u_sampler: sampler;
+@group(0) @binding(1) var<uniform> u_ratio: vec4<f32>;
+@group(1) @binding(0) var u_texture: texture_2d<f32>;
+
+struct VertexInput {
+    @builtin(vertex_index) vertex_index: u32,
+}
+
+struct VertexOutput {
+    @builtin(position) position: vec4<f32>,
+    @location(0) uv: vec2<f32>,
+}
+
+@vertex
+fn vs_main(input: VertexInput) -> VertexOutput {
+    let uv = uvs[input.vertex_index];
+
+    var out: VertexOutput;
+    out.uv = uv * u_ratio.xy;
+    out.position = vec4<f32>(uv * vec2(2.0, -2.0) + vec2(-1.0, 1.0), 0.0, 1.0);
+
+    return out;
+}
+
+@fragment
+fn fs_main(input: VertexOutput) -> @location(0) vec4<f32> {
+    return textureSample(u_texture, u_sampler, input.uv);
+}

crates/iced_wgpu/src/shader/color.wgsl

@@ -0,0 +1,14 @@
+fn premultiply(color: vec4<f32>) -> vec4<f32> {
+    return vec4(color.xyz * color.a, color.a);
+}
+
+fn unpack_color(data: vec2<u32>) -> vec4<f32> {
+    return premultiply(unpack_u32(data));
+}
+
+fn unpack_u32(data: vec2<u32>) -> vec4<f32> {
+    let rg: vec2<f32> = unpack2x16float(data.x);
+    let ba: vec2<f32> = unpack2x16float(data.y);
+
+    return vec4<f32>(rg.y, rg.x, ba.y, ba.x);
+}

crates/iced_wgpu/src/shader/color/linear_rgb.wgsl

@@ -0,0 +1,3 @@
+fn interpolate_color(from_: vec4<f32>, to_: vec4<f32>, factor: f32) -> vec4<f32> {
+    return mix(from_, to_, factor);
+}

crates/iced_wgpu/src/shader/image.wgsl

@@ -0,0 +1,129 @@
+struct Globals {
+    transform: mat4x4<f32>,
+    scale_factor: f32,
+}
+
+@group(0) @binding(0) var<uniform> globals: Globals;
+@group(0) @binding(1) var u_sampler: sampler;
+@group(1) @binding(0) var u_texture: texture_2d_array<f32>;
+
+struct VertexInput {
+    @builtin(vertex_index) vertex_index: u32,
+    @location(0) center: vec2<f32>,
+    @location(1) clip_bounds: vec4<f32>,
+    @location(2) border_radius: vec4<f32>,
+    @location(3) tile: vec4<f32>,
+    @location(4) rotation: f32,
+    @location(5) opacity: f32,
+    @location(6) atlas_pos: vec2<f32>,
+    @location(7) atlas_scale: vec2<f32>,
+    @location(8) layer: i32,
+    @location(9) snap: u32,
+}
+
+struct VertexOutput {
+    @builtin(position) position: vec4<f32>,
+    @location(0) @interpolate(flat) clip_bounds: vec4<f32>,
+    @location(1) @interpolate(flat) border_radius: vec4<f32>,
+    @location(2) @interpolate(flat) atlas: vec4<f32>,
+    @location(3) @interpolate(flat) layer: i32,
+    @location(4) @interpolate(flat) opacity: f32,
+    @location(5) uv: vec2<f32>,
+}
+
+@vertex
+fn vs_main(input: VertexInput) -> VertexOutput {
+    var out: VertexOutput;
+
+    // Generate a vertex position in the range [0, 1] from the vertex index
+    let corner = vertex_position(input.vertex_index);
+
+    let tile = input.tile;
+    let center = input.center;
+
+    // List the unrotated tile corners
+    let corners = array<vec2<f32>, 4>(
+        tile.xy,                              // Top left
+        tile.xy + vec2<f32>(tile.z, 0.0),     // Top right
+        tile.xy + vec2<f32>(0.0, tile.w),     // Bottom left
+        tile.xy + tile.zw                     // Bottom right
+    );
+
+    // Rotate tile corners around center
+    let cos_r = cos(-input.rotation); // Clockwise
+    let sin_r = sin(-input.rotation);
+    var rotated = array<vec2<f32>, 4>();
+
+    for (var i = 0u; i < 4u; i++) {
+        let c = corners[i] - input.center;
+        rotated[i] = vec2<f32>(c.x * cos_r - c.y * sin_r, c.x * sin_r + c.y * cos_r) + input.center;
+    }
+
+    // Find bounding box of rotated tile
+    var min_xy = rotated[0];
+    var max_xy = rotated[0];
+    for (var i = 1u; i < 4u; i++) {
+        min_xy = min(min_xy, rotated[i]);
+        max_xy = max(max_xy, rotated[i]);
+    }
+    let rotated_bounds = vec4<f32>(min_xy, max_xy - min_xy);
+
+    // Intersect with clip bounds
+    let clip_min = max(rotated_bounds.xy, input.clip_bounds.xy);
+    let clip_max = min(rotated_bounds.xy + rotated_bounds.zw, input.clip_bounds.xy + input.clip_bounds.zw);
+    let clipped_tile = vec4<f32>(clip_min, max(vec2<f32>(0.0), clip_max - clip_min));
+
+    // Calculate the vertex position
+    let v_pos = clipped_tile.xy + corner * clipped_tile.zw;
+    out.position = vec4(vec2(globals.scale_factor), 1.0, 1.0) * vec4<f32>(v_pos, 0.0, 1.0);
+    out.clip_bounds = globals.scale_factor * input.clip_bounds;
+
+    // Calculate rotated UV
+    let uv = input.atlas_pos + (v_pos - tile.xy) / tile.zw * input.atlas_scale;
+    let uv_center = input.atlas_pos + input.atlas_scale / 2.0;
+
+    let d = uv - uv_center;
+    out.uv = vec2<f32>(d.x * cos_r - d.y * sin_r, d.x * sin_r + d.y * cos_r) + uv_center;
+
+    // Snap position to the pixel grid
+    if bool(input.snap) {
+        out.position = round(out.position);
+        out.clip_bounds = vec4(
+            round(out.clip_bounds.xy),
+            round(out.clip_bounds.xy + out.clip_bounds.zw) - out.clip_bounds.xy,
+        );
+    }
+
+    out.position = globals.transform * out.position;
+    out.border_radius = globals.scale_factor * min(input.border_radius, vec4(min(input.clip_bounds.z, input.clip_bounds.w) / 2.0));
+    out.atlas = vec4(input.atlas_pos, input.atlas_pos + input.atlas_scale);
+    out.layer = input.layer;
+    out.opacity = input.opacity;
+
+    return out;
+}
+
+@fragment
+fn fs_main(input: VertexOutput) -> @location(0) vec4<f32> {
+    let fragment = input.position.xy;
+    let position = input.clip_bounds.xy;
+    let scale = input.clip_bounds.zw;
+
+    let d = rounded_box_sdf(
+        2.0 * (fragment - position - scale / 2.0),
+        scale,
+        input.border_radius * 2.0,
+    ) / 2.0;
+
+    let antialias: f32 = clamp(1.0 - d, 0.0, 1.0);
+    let inside = all(input.uv >= input.atlas.xy) && all(input.uv <= input.atlas.zw);
+
+    return textureSample(u_texture, u_sampler, input.uv, input.layer) * vec4<f32>(1.0, 1.0, 1.0, antialias * input.opacity * f32(inside));
+}
+
+fn rounded_box_sdf(p: vec2<f32>, size: vec2<f32>, corners: vec4<f32>) -> f32 {
+    var box_half = select(corners.yz, corners.xw, p.x > 0.0);
+    var corner = select(box_half.y, box_half.x, p.y > 0.0);
+    var q = abs(p) - size + corner;
+    return min(max(q.x, q.y), 0.0) + length(max(q, vec2(0.0))) - corner;
+}

crates/iced_wgpu/src/shader/quad.wgsl

@@ -0,0 +1,13 @@
+struct Globals {
+    transform: mat4x4<f32>,
+    scale: f32,
+}
+
+@group(0) @binding(0) var<uniform> globals: Globals;
+
+fn rounded_box_sdf(p: vec2<f32>, size: vec2<f32>, corners: vec4<f32>) -> f32 {
+    var box_half = select(corners.yz, corners.xw, p.x > 0.0);
+    var corner = select(box_half.y, box_half.x, p.y > 0.0);
+    var q = abs(p) - size + corner;
+    return min(max(q.x, q.y), 0.0) + length(max(q, vec2(0.0))) - corner;
+}

crates/iced_wgpu/src/shader/quad/gradient.wgsl

@@ -0,0 +1,184 @@
+struct GradientVertexInput {
+    @builtin(vertex_index) vertex_index: u32,
+    @location(0) @interpolate(flat) colors_1: vec4<u32>,
+    @location(1) @interpolate(flat) colors_2: vec4<u32>,
+    @location(2) @interpolate(flat) colors_3: vec4<u32>,
+    @location(3) @interpolate(flat) colors_4: vec4<u32>,
+    @location(4) @interpolate(flat) offsets: vec4<u32>,
+    @location(5) direction: vec4<f32>,
+    @location(6) position_and_scale: vec4<f32>,
+    @location(7) border_color: vec4<f32>,
+    @location(8) border_radius: vec4<f32>,
+    @location(9) border_width: f32,
+    @location(10) snap: u32,
+}
+
+struct GradientVertexOutput {
+    @builtin(position) position: vec4<f32>,
+    @location(1) @interpolate(flat) colors_1: vec4<u32>,
+    @location(2) @interpolate(flat) colors_2: vec4<u32>,
+    @location(3) @interpolate(flat) colors_3: vec4<u32>,
+    @location(4) @interpolate(flat) colors_4: vec4<u32>,
+    @location(5) @interpolate(flat) offsets: vec4<u32>,
+    @location(6) direction: vec4<f32>,
+    @location(7) position_and_scale: vec4<f32>,
+    @location(8) border_color: vec4<f32>,
+    @location(9) border_radius: vec4<f32>,
+    @location(10) border_width: f32,
+}
+
+@vertex
+fn gradient_vs_main(input: GradientVertexInput) -> GradientVertexOutput {
+    var out: GradientVertexOutput;
+
+    var pos: vec2<f32> = input.position_and_scale.xy * globals.scale;
+    var scale: vec2<f32> = input.position_and_scale.zw * globals.scale;
+
+    var pos_snap = vec2<f32>(0.0, 0.0);
+    var scale_snap = vec2<f32>(0.0, 0.0);
+
+    if bool(input.snap) {
+        pos_snap = round(pos + vec2(0.001, 0.001)) - pos;
+        scale_snap = round(pos + scale + vec2(0.001, 0.001)) - pos - pos_snap - scale;
+    }
+
+    var min_border_radius = min(input.position_and_scale.z, input.position_and_scale.w) * 0.5;
+    var border_radius: vec4<f32> = vec4<f32>(
+        min(input.border_radius.x, min_border_radius),
+        min(input.border_radius.y, min_border_radius),
+        min(input.border_radius.z, min_border_radius),
+        min(input.border_radius.w, min_border_radius)
+    );
+
+    var transform: mat4x4<f32> = mat4x4<f32>(
+        vec4<f32>(scale.x + scale_snap.x + 1.0, 0.0, 0.0, 0.0),
+        vec4<f32>(0.0, scale.y + scale_snap.y + 1.0, 0.0, 0.0),
+        vec4<f32>(0.0, 0.0, 1.0, 0.0),
+        vec4<f32>(pos + pos_snap - vec2<f32>(0.5, 0.5), 0.0, 1.0)
+    );
+
+    out.position = globals.transform * transform * vec4<f32>(vertex_position(input.vertex_index), 0.0, 1.0);
+    out.colors_1 = input.colors_1;
+    out.colors_2 = input.colors_2;
+    out.colors_3 = input.colors_3;
+    out.colors_4 = input.colors_4;
+    out.offsets = input.offsets;
+    out.direction = input.direction * globals.scale;
+    out.position_and_scale = vec4<f32>(pos + pos_snap, scale + scale_snap);
+    out.border_color = premultiply(input.border_color);
+    out.border_radius = border_radius * globals.scale;
+    out.border_width = input.border_width * globals.scale;
+
+    return out;
+}
+
+fn random(coords: vec2<f32>) -> f32 {
+    return fract(sin(dot(coords, vec2(12.9898,78.233))) * 43758.5453);
+}
+
+/// Returns the current interpolated color with a max 8-stop gradient
+fn gradient(
+    raw_position: vec2<f32>,
+    direction: vec4<f32>,
+    colors: array<vec4<f32>, 8>,
+    offsets: array<f32, 8>,
+    last_index: i32
+) -> vec4<f32> {
+    let start = direction.xy;
+    let end = direction.zw;
+
+    let v1 = end - start;
+    let v2 = raw_position - start;
+    let unit = normalize(v1);
+    let coord_offset = dot(unit, v2) / length(v1);
+
+    //need to store these as a var to use dynamic indexing in a loop
+    //this is already added to wgsl spec but not in wgpu yet
+    var colors_arr = colors;
+    var offsets_arr = offsets;
+
+    var color: vec4<f32>;
+
+    let noise_granularity: f32 = 0.3/255.0;
+
+    for (var i: i32 = 0; i < last_index; i++) {
+        let curr_offset = offsets_arr[i];
+        let next_offset = offsets_arr[i+1];
+
+        if (coord_offset <= offsets_arr[0]) {
+            color = colors_arr[0];
+        }
+
+        if (curr_offset <= coord_offset && coord_offset <= next_offset) {
+            let from_ = colors_arr[i];
+            let to_ = colors_arr[i+1];
+            let factor = smoothstep(curr_offset, next_offset, coord_offset);
+
+            color = interpolate_color(from_, to_, factor);
+        }
+
+        if (coord_offset >= offsets_arr[last_index]) {
+            color = colors_arr[last_index];
+        }
+    }
+
+    return color + mix(-noise_granularity, noise_granularity, random(raw_position));
+}
+
+@fragment
+fn gradient_fs_main(input: GradientVertexOutput) -> @location(0) vec4<f32> {
+    let colors = array<vec4<f32>, 8>(
+        unpack_color(input.colors_1.xy),
+        unpack_color(input.colors_1.zw),
+        unpack_color(input.colors_2.xy),
+        unpack_color(input.colors_2.zw),
+        unpack_color(input.colors_3.xy),
+        unpack_color(input.colors_3.zw),
+        unpack_color(input.colors_4.xy),
+        unpack_color(input.colors_4.zw),
+    );
+
+    let offsets_1: vec4<f32> = unpack_u32(input.offsets.xy);
+    let offsets_2: vec4<f32> = unpack_u32(input.offsets.zw);
+
+    var offsets = array<f32, 8>(
+        offsets_1.x,
+        offsets_1.y,
+        offsets_1.z,
+        offsets_1.w,
+        offsets_2.x,
+        offsets_2.y,
+        offsets_2.z,
+        offsets_2.w,
+    );
+
+    //TODO could just pass this in to the shader but is probably more performant to just check it here
+    var last_index = 7;
+    for (var i: i32 = 0; i <= 7; i++) {
+        if (offsets[i] > 1.0) {
+            last_index = i - 1;
+            break;
+        }
+    }
+
+    var mixed_color: vec4<f32> = gradient(input.position.xy, input.direction, colors, offsets, last_index);
+
+    let pos = input.position_and_scale.xy;
+    let scale = input.position_and_scale.zw;
+
+    var dist: f32 = rounded_box_sdf(
+        -(input.position.xy - pos - scale / 2.0) * 2.0,
+        scale,
+        input.border_radius * 2.0
+    ) / 2.0;
+
+    if (input.border_width > 0.0) {
+        mixed_color = mix(
+            mixed_color,
+            input.border_color,
+            clamp(0.5 + dist + input.border_width, 0.0, 1.0)
+        );
+    }
+
+    return mixed_color * clamp(0.5-dist, 0.0, 1.0);
+}

crates/iced_wgpu/src/shader/quad/solid.wgsl

@@ -0,0 +1,102 @@
+struct SolidVertexInput {
+    @builtin(vertex_index) vertex_index: u32,
+    @location(0) color: vec4<f32>,
+    @location(1) pos: vec2<f32>,
+    @location(2) scale: vec2<f32>,
+    @location(3) border_color: vec4<f32>,
+    @location(4) border_radius: vec4<f32>,
+    @location(5) border_width: f32,
+    @location(6) shadow_color: vec4<f32>,
+    @location(7) shadow_offset: vec2<f32>,
+    @location(8) shadow_blur_radius: f32,
+    @location(9) snap: u32,
+}
+
+struct SolidVertexOutput {
+    @builtin(position) position: vec4<f32>,
+    @location(0) color: vec4<f32>,
+    @location(1) border_color: vec4<f32>,
+    @location(2) pos: vec2<f32>,
+    @location(3) scale: vec2<f32>,
+    @location(4) border_radius: vec4<f32>,
+    @location(5) border_width: f32,
+    @location(6) shadow_color: vec4<f32>,
+    @location(7) shadow_offset: vec2<f32>,
+    @location(8) shadow_blur_radius: f32,
+}
+
+@vertex
+fn solid_vs_main(input: SolidVertexInput) -> SolidVertexOutput {
+    var out: SolidVertexOutput;
+
+    var pos: vec2<f32> = (input.pos + min(input.shadow_offset, vec2<f32>(0.0, 0.0)) - input.shadow_blur_radius) * globals.scale;
+    var scale: vec2<f32> = (input.scale + vec2<f32>(abs(input.shadow_offset.x), abs(input.shadow_offset.y)) + input.shadow_blur_radius * 2.0) * globals.scale;
+
+    var pos_snap = vec2<f32>(0.0, 0.0);
+    var scale_snap = vec2<f32>(0.0, 0.0);
+
+    if bool(input.snap) {
+        pos_snap = round(pos + vec2(0.001, 0.001)) - pos;
+        scale_snap = round(pos + scale + vec2(0.001, 0.001)) - pos - pos_snap - scale;
+    }
+
+    let border_radius = min(input.border_radius, vec4(min(input.scale.x, input.scale.y) / 2.0));
+
+    var transform: mat4x4<f32> = mat4x4<f32>(
+        vec4<f32>(scale.x + scale_snap.x + 1.0, 0.0, 0.0, 0.0),
+        vec4<f32>(0.0, scale.y + scale_snap.y + 1.0, 0.0, 0.0),
+        vec4<f32>(0.0, 0.0, 1.0, 0.0),
+        vec4<f32>(pos + pos_snap - vec2<f32>(0.5, 0.5), 0.0, 1.0)
+    );
+
+    out.position = globals.transform * transform * vec4<f32>(vertex_position(input.vertex_index), 0.0, 1.0);
+    out.color = premultiply(input.color);
+    out.border_color = premultiply(input.border_color);
+    out.pos = input.pos * globals.scale + pos_snap;
+    out.scale = input.scale * globals.scale + scale_snap;
+    out.border_radius = border_radius * globals.scale;
+    out.border_width = input.border_width * globals.scale;
+    out.shadow_color = premultiply(input.shadow_color);
+    out.shadow_offset = input.shadow_offset * globals.scale;
+    out.shadow_blur_radius = input.shadow_blur_radius * globals.scale;
+
+    return out;
+}
+
+@fragment
+fn solid_fs_main(
+    input: SolidVertexOutput
+) -> @location(0) vec4<f32> {
+    var mixed_color: vec4<f32> = input.color;
+
+    var dist = rounded_box_sdf(
+        -(input.position.xy - input.pos - input.scale * 0.5) * 2.0,
+        input.scale,
+        input.border_radius * 2.0
+    ) / 2.0;
+
+    if (input.border_width > 0.0) {
+        mixed_color = mix(
+            input.color,
+            input.border_color,
+            clamp(0.5 + dist + input.border_width, 0.0, 1.0)
+        );
+    }
+
+    var quad_alpha: f32 = clamp(0.5-dist, 0.0, 1.0);
+
+    let quad_color = mixed_color * quad_alpha;
+
+    if input.shadow_color.a > 0.0 {
+        var shadow_dist: f32 = rounded_box_sdf(
+            -(input.position.xy - input.pos - input.shadow_offset - input.scale/2.0) * 2.0,
+            input.scale,
+            input.border_radius * 2.0
+        ) / 2.0;
+        let shadow_alpha = 1.0 - smoothstep(-input.shadow_blur_radius, input.shadow_blur_radius, max(shadow_dist, 0.0));
+
+        return mix(quad_color, input.shadow_color, (1.0 - quad_alpha) * shadow_alpha);
+    } else {
+        return quad_color;
+    }
+}

crates/iced_wgpu/src/shader/triangle.wgsl

@@ -0,0 +1,5 @@
+struct Globals {
+    transform: mat4x4<f32>,
+}
+
+@group(0) @binding(0) var<uniform> globals: Globals;

crates/iced_wgpu/src/shader/triangle/gradient.wgsl

@@ -0,0 +1,127 @@
+struct GradientVertexInput {
+    @location(0) v_pos: vec2<f32>,
+    @location(1) @interpolate(flat) colors_1: vec4<u32>,
+    @location(2) @interpolate(flat) colors_2: vec4<u32>,
+    @location(3) @interpolate(flat) colors_3: vec4<u32>,
+    @location(4) @interpolate(flat) colors_4: vec4<u32>,
+    @location(5) @interpolate(flat) offsets: vec4<u32>,
+    @location(6) direction: vec4<f32>,
+}
+
+struct GradientVertexOutput {
+    @builtin(position) position: vec4<f32>,
+    @location(0) raw_position: vec2<f32>,
+    @location(1) @interpolate(flat) colors_1: vec4<u32>,
+    @location(2) @interpolate(flat) colors_2: vec4<u32>,
+    @location(3) @interpolate(flat) colors_3: vec4<u32>,
+    @location(4) @interpolate(flat) colors_4: vec4<u32>,
+    @location(5) @interpolate(flat) offsets: vec4<u32>,
+    @location(6) direction: vec4<f32>,
+}
+
+@vertex
+fn gradient_vs_main(input: GradientVertexInput) -> GradientVertexOutput {
+    var output: GradientVertexOutput;
+
+    output.position = globals.transform * vec4<f32>(input.v_pos, 0.0, 1.0);
+    output.raw_position = input.v_pos;
+    output.colors_1 = input.colors_1;
+    output.colors_2 = input.colors_2;
+    output.colors_3 = input.colors_3;
+    output.colors_4 = input.colors_4;
+    output.offsets = input.offsets;
+    output.direction = input.direction;
+
+    return output;
+}
+
+/// Returns the current interpolated color with a max 8-stop gradient
+fn gradient(
+    raw_position: vec2<f32>,
+    direction: vec4<f32>,
+    colors: array<vec4<f32>, 8>,
+    offsets: array<f32, 8>,
+    last_index: i32
+) -> vec4<f32> {
+    let start = direction.xy;
+    let end = direction.zw;
+
+    let v1 = end - start;
+    let v2 = raw_position - start;
+    let unit = normalize(v1);
+    let coord_offset = dot(unit, v2) / length(v1);
+
+    //need to store these as a var to use dynamic indexing in a loop
+    //this is already added to wgsl spec but not in wgpu yet
+    var colors_arr = colors;
+    var offsets_arr = offsets;
+
+    var color: vec4<f32>;
+
+    let noise_granularity: f32 = 0.3/255.0;
+
+    for (var i: i32 = 0; i < last_index; i++) {
+        let curr_offset = offsets_arr[i];
+        let next_offset = offsets_arr[i+1];
+
+        if (coord_offset <= offsets_arr[0]) {
+            color = colors_arr[0];
+        }
+
+        if (curr_offset <= coord_offset && coord_offset <= next_offset) {
+            let from_ = colors_arr[i];
+            let to_ = colors_arr[i+1];
+            let factor = smoothstep(curr_offset, next_offset, coord_offset);
+
+            color = interpolate_color(from_, to_, factor);
+        }
+
+        if (coord_offset >= offsets_arr[last_index]) {
+            color = colors_arr[last_index];
+        }
+    }
+
+    return color + mix(-noise_granularity, noise_granularity, random(raw_position));
+}
+
+@fragment
+fn gradient_fs_main(input: GradientVertexOutput) -> @location(0) vec4<f32> {
+    let colors = array<vec4<f32>, 8>(
+        unpack_color(input.colors_1.xy),
+        unpack_color(input.colors_1.zw),
+        unpack_color(input.colors_2.xy),
+        unpack_color(input.colors_2.zw),
+        unpack_color(input.colors_3.xy),
+        unpack_color(input.colors_3.zw),
+        unpack_color(input.colors_4.xy),
+        unpack_color(input.colors_4.zw),
+    );
+
+    let offsets_1: vec4<f32> = unpack_u32(input.offsets.xy);
+    let offsets_2: vec4<f32> = unpack_u32(input.offsets.zw);
+
+    var offsets = array<f32, 8>(
+        offsets_1.x,
+        offsets_1.y,
+        offsets_1.z,
+        offsets_1.w,
+        offsets_2.x,
+        offsets_2.y,
+        offsets_2.z,
+        offsets_2.w,
+    );
+
+    var last_index = 7;
+    for (var i: i32 = 0; i <= 7; i++) {
+        if (offsets[i] >= 1.0) {
+            last_index = i;
+            break;
+        }
+    }
+
+    return gradient(input.raw_position, input.direction, colors, offsets, last_index);
+}
+
+fn random(coords: vec2<f32>) -> f32 {
+    return fract(sin(dot(coords, vec2(12.9898,78.233))) * 43758.5453);
+}

crates/iced_wgpu/src/shader/triangle/solid.wgsl

@@ -0,0 +1,24 @@
+struct SolidVertexInput {
+    @location(0) position: vec2<f32>,
+    @location(1) color: vec4<f32>,
+}
+
+struct SolidVertexOutput {
+    @builtin(position) position: vec4<f32>,
+    @location(0) color: vec4<f32>,
+}
+
+@vertex
+fn solid_vs_main(input: SolidVertexInput) -> SolidVertexOutput {
+    var out: SolidVertexOutput;
+
+    out.color = premultiply(input.color);
+    out.position = globals.transform * vec4<f32>(input.position, 0.0, 1.0);
+
+    return out;
+}
+
+@fragment
+fn solid_fs_main(input: SolidVertexOutput) -> @location(0) vec4<f32> {
+    return input.color;
+}

crates/iced_wgpu/src/shader/vertex.wgsl

@@ -0,0 +1,7 @@
+// Compute the normalized quad coordinates based on the vertex index.
+fn vertex_position(vertex_index: u32) -> vec2<f32> {
+    // #: 0 1 2 3 4 5
+    // x: 1 1 0 0 0 1
+    // y: 1 0 0 0 1 1
+    return vec2<f32>((vec2(1u, 2u) + vertex_index) % vec2(6u) < vec2(3u));
+}

crates/iced_wgpu/src/text.rs

@@ -0,0 +1,648 @@
+use crate::core::alignment;
+use crate::core::text::Alignment;
+use crate::core::{Rectangle, Size, Transformation};
+use crate::graphics::cache;
+use crate::graphics::color;
+use crate::graphics::text::cache::{self as text_cache, Cache as BufferCache};
+use crate::graphics::text::{Editor, Paragraph, font_system, to_color};
+
+use rustc_hash::FxHashMap;
+use std::collections::hash_map;
+use std::sync::atomic::{self, AtomicU64};
+use std::sync::{self, Arc, RwLock};
+
+pub use crate::graphics::Text;
+
+const COLOR_MODE: cryoglyph::ColorMode = if color::GAMMA_CORRECTION {
+    cryoglyph::ColorMode::Accurate
+} else {
+    cryoglyph::ColorMode::Web
+};
+
+pub type Batch = Vec<Item>;
+
+#[derive(Debug)]
+pub enum Item {
+    Group {
+        transformation: Transformation,
+        text: Vec<Text>,
+    },
+    Cached {
+        transformation: Transformation,
+        cache: Cache,
+    },
+}
+
+#[derive(Debug, Clone)]
+pub struct Cache {
+    id: Id,
+    group: cache::Group,
+    text: Arc<[Text]>,
+    version: usize,
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, Eq, PartialOrd, Ord, Hash)]
+pub struct Id(u64);
+
+impl Cache {
+    pub fn new(group: cache::Group, text: Vec<Text>) -> Option<Self> {
+        static NEXT_ID: AtomicU64 = AtomicU64::new(0);
+
+        if text.is_empty() {
+            return None;
+        }
+
+        Some(Self {
+            id: Id(NEXT_ID.fetch_add(1, atomic::Ordering::Relaxed)),
+            group,
+            text: Arc::from(text),
+            version: 0,
+        })
+    }
+
+    pub fn update(&mut self, text: Vec<Text>) {
+        if self.text.is_empty() && text.is_empty() {
+            return;
+        }
+
+        self.text = Arc::from(text);
+        self.version += 1;
+    }
+}
+
+struct Upload {
+    renderer: cryoglyph::TextRenderer,
+    buffer_cache: BufferCache,
+    transformation: Transformation,
+    version: usize,
+    group_version: usize,
+    text: sync::Weak<[Text]>,
+    _atlas: sync::Weak<()>,
+}
+
+#[derive(Default)]
+pub struct Storage {
+    groups: FxHashMap<cache::Group, Group>,
+    uploads: FxHashMap<Id, Upload>,
+}
+
+struct Group {
+    atlas: cryoglyph::TextAtlas,
+    version: usize,
+    should_trim: bool,
+    handle: Arc<()>, // Keeps track of active uploads
+}
+
+impl Storage {
+    fn get(&self, cache: &Cache) -> Option<(&cryoglyph::TextAtlas, &Upload)> {
+        if cache.text.is_empty() {
+            return None;
+        }
+
+        self.groups
+            .get(&cache.group)
+            .map(|group| &group.atlas)
+            .zip(self.uploads.get(&cache.id))
+    }
+
+    fn prepare(
+        &mut self,
+        device: &wgpu::Device,
+        queue: &wgpu::Queue,
+        viewport: &cryoglyph::Viewport,
+        encoder: &mut wgpu::CommandEncoder,
+        format: wgpu::TextureFormat,
+        state: &cryoglyph::Cache,
+        cache: &Cache,
+        new_transformation: Transformation,
+        bounds: Rectangle,
+    ) {
+        let group_count = self.groups.len();
+
+        let group = self.groups.entry(cache.group).or_insert_with(|| {
+            log::debug!(
+                "New text atlas: {:?} (total: {})",
+                cache.group,
+                group_count + 1
+            );
+
+            Group {
+                atlas: cryoglyph::TextAtlas::with_color_mode(
+                    device, queue, state, format, COLOR_MODE,
+                ),
+                version: 0,
+                should_trim: false,
+                handle: Arc::new(()),
+            }
+        });
+
+        match self.uploads.entry(cache.id) {
+            hash_map::Entry::Occupied(entry) => {
+                let upload = entry.into_mut();
+
+                if upload.version != cache.version
+                    || upload.group_version != group.version
+                    || upload.transformation != new_transformation
+                {
+                    if !cache.text.is_empty() {
+                        let _ = prepare(
+                            device,
+                            queue,
+                            viewport,
+                            encoder,
+                            &mut upload.renderer,
+                            &mut group.atlas,
+                            &mut upload.buffer_cache,
+                            &cache.text,
+                            bounds,
+                            new_transformation,
+                        );
+                    }
+
+                    // Only trim if glyphs have changed
+                    group.should_trim =
+                        group.should_trim || upload.version != cache.version;
+
+                    upload.text = Arc::downgrade(&cache.text);
+                    upload.version = cache.version;
+                    upload.group_version = group.version;
+                    upload.transformation = new_transformation;
+
+                    upload.buffer_cache.trim();
+                }
+            }
+            hash_map::Entry::Vacant(entry) => {
+                let mut renderer = cryoglyph::TextRenderer::new(
+                    &mut group.atlas,
+                    device,
+                    wgpu::MultisampleState::default(),
+                    None,
+                );
+
+                let mut buffer_cache = BufferCache::new();
+
+                if !cache.text.is_empty() {
+                    let _ = prepare(
+                        device,
+                        queue,
+                        viewport,
+                        encoder,
+                        &mut renderer,
+                        &mut group.atlas,
+                        &mut buffer_cache,
+                        &cache.text,
+                        bounds,
+                        new_transformation,
+                    );
+                }
+
+                let _ = entry.insert(Upload {
+                    renderer,
+                    buffer_cache,
+                    transformation: new_transformation,
+                    version: 0,
+                    group_version: group.version,
+                    text: Arc::downgrade(&cache.text),
+                    _atlas: Arc::downgrade(&group.handle),
+                });
+
+                group.should_trim = cache.group.is_singleton();
+
+                log::debug!(
+                    "New text upload: {} (total: {})",
+                    cache.id.0,
+                    self.uploads.len()
+                );
+            }
+        }
+    }
+
+    pub fn trim(&mut self) {
+        self.uploads
+            .retain(|_id, upload| upload.text.strong_count() > 0);
+
+        self.groups.retain(|id, group| {
+            let active_uploads = Arc::weak_count(&group.handle);
+
+            if active_uploads == 0 {
+                log::debug!("Dropping text atlas: {id:?}");
+
+                return false;
+            }
+
+            if group.should_trim {
+                log::trace!("Trimming text atlas: {id:?}");
+
+                group.atlas.trim();
+                group.should_trim = false;
+
+                // We only need to worry about glyph fighting
+                // when the atlas may be shared by multiple
+                // uploads.
+                if !id.is_singleton() {
+                    log::debug!(
+                        "Invalidating text atlas: {id:?} \
+                        (uploads: {active_uploads})"
+                    );
+
+                    group.version += 1;
+                }
+            }
+
+            true
+        });
+    }
+}
+
+pub struct Viewport(cryoglyph::Viewport);
+
+impl Viewport {
+    pub fn update(&mut self, queue: &wgpu::Queue, resolution: Size<u32>) {
+        self.0.update(
+            queue,
+            cryoglyph::Resolution {
+                width: resolution.width,
+                height: resolution.height,
+            },
+        );
+    }
+}
+
+#[derive(Clone)]
+pub struct Pipeline {
+    format: wgpu::TextureFormat,
+    cache: cryoglyph::Cache,
+    atlas: Arc<RwLock<cryoglyph::TextAtlas>>,
+}
+
+impl Pipeline {
+    pub fn new(
+        device: &wgpu::Device,
+        queue: &wgpu::Queue,
+        format: wgpu::TextureFormat,
+    ) -> Self {
+        let cache = cryoglyph::Cache::new(device);
+        let atlas = cryoglyph::TextAtlas::with_color_mode(
+            device, queue, &cache, format, COLOR_MODE,
+        );
+
+        Pipeline {
+            format,
+            cache,
+            atlas: Arc::new(RwLock::new(atlas)),
+        }
+    }
+
+    pub fn create_viewport(&self, device: &wgpu::Device) -> Viewport {
+        Viewport(cryoglyph::Viewport::new(device, &self.cache))
+    }
+
+    pub fn trim(&self) {
+        self.atlas.write().expect("Write text atlas").trim();
+    }
+}
+
+#[derive(Default)]
+pub struct State {
+    renderers: Vec<cryoglyph::TextRenderer>,
+    prepare_layer: usize,
+    cache: BufferCache,
+    storage: Storage,
+}
+
+impl State {
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    pub fn prepare(
+        &mut self,
+        pipeline: &Pipeline,
+        device: &wgpu::Device,
+        queue: &wgpu::Queue,
+        viewport: &Viewport,
+        encoder: &mut wgpu::CommandEncoder,
+        batch: &Batch,
+        layer_bounds: Rectangle,
+        layer_transformation: Transformation,
+    ) {
+        let mut atlas = pipeline.atlas.write().expect("Write to text atlas");
+
+        for item in batch {
+            match item {
+                Item::Group {
+                    transformation,
+                    text,
+                } => {
+                    if self.renderers.len() <= self.prepare_layer {
+                        self.renderers.push(cryoglyph::TextRenderer::new(
+                            &mut atlas,
+                            device,
+                            wgpu::MultisampleState::default(),
+                            None,
+                        ));
+                    }
+
+                    let renderer = &mut self.renderers[self.prepare_layer];
+                    let result = prepare(
+                        device,
+                        queue,
+                        &viewport.0,
+                        encoder,
+                        renderer,
+                        &mut atlas,
+                        &mut self.cache,
+                        text,
+                        layer_bounds * layer_transformation,
+                        layer_transformation * *transformation,
+                    );
+
+                    match result {
+                        Ok(()) => {
+                            self.prepare_layer += 1;
+                        }
+                        Err(cryoglyph::PrepareError::AtlasFull) => {
+                            // If the atlas cannot grow, then all bets are off.
+                            // Instead of panicking, we will just pray that the result
+                            // will be somewhat readable...
+                        }
+                    }
+                }
+                Item::Cached {
+                    transformation,
+                    cache,
+                } => {
+                    self.storage.prepare(
+                        device,
+                        queue,
+                        &viewport.0,
+                        encoder,
+                        pipeline.format,
+                        &pipeline.cache,
+                        cache,
+                        layer_transformation * *transformation,
+                        layer_bounds * layer_transformation,
+                    );
+                }
+            }
+        }
+    }
+
+    pub fn render<'a>(
+        &'a self,
+        pipeline: &'a Pipeline,
+        viewport: &'a Viewport,
+        start: usize,
+        batch: &'a Batch,
+        bounds: Rectangle<u32>,
+        render_pass: &mut wgpu::RenderPass<'a>,
+    ) -> usize {
+        let atlas = pipeline.atlas.read().expect("Read text atlas");
+        let mut layer_count = 0;
+
+        render_pass.set_scissor_rect(
+            bounds.x,
+            bounds.y,
+            bounds.width,
+            bounds.height,
+        );
+
+        for item in batch {
+            match item {
+                Item::Group { .. } => {
+                    let renderer = &self.renderers[start + layer_count];
+
+                    renderer
+                        .render(&atlas, &viewport.0, render_pass)
+                        .expect("Render text");
+
+                    layer_count += 1;
+                }
+                Item::Cached { cache, .. } => {
+                    if let Some((atlas, upload)) = self.storage.get(cache) {
+                        upload
+                            .renderer
+                            .render(atlas, &viewport.0, render_pass)
+                            .expect("Render cached text");
+                    }
+                }
+            }
+        }
+
+        layer_count
+    }
+
+    pub fn trim(&mut self) {
+        self.cache.trim();
+        self.storage.trim();
+
+        self.prepare_layer = 0;
+    }
+}
+
+fn prepare(
+    device: &wgpu::Device,
+    queue: &wgpu::Queue,
+    viewport: &cryoglyph::Viewport,
+    encoder: &mut wgpu::CommandEncoder,
+    renderer: &mut cryoglyph::TextRenderer,
+    atlas: &mut cryoglyph::TextAtlas,
+    buffer_cache: &mut BufferCache,
+    sections: &[Text],
+    layer_bounds: Rectangle,
+    layer_transformation: Transformation,
+) -> Result<(), cryoglyph::PrepareError> {
+    let mut font_system = font_system().write().expect("Write font system");
+    let font_system = font_system.raw();
+
+    enum Allocation {
+        Paragraph(Paragraph),
+        Editor(Editor),
+        Cache(text_cache::KeyHash),
+        Raw(Arc<cryoglyph::Buffer>),
+    }
+
+    let allocations: Vec<_> = sections
+        .iter()
+        .map(|section| match section {
+            Text::Paragraph { paragraph, .. } => {
+                paragraph.upgrade().map(Allocation::Paragraph)
+            }
+            Text::Editor { editor, .. } => {
+                editor.upgrade().map(Allocation::Editor)
+            }
+            Text::Cached {
+                content,
+                bounds,
+                size,
+                line_height,
+                font,
+                shaping,
+                align_x,
+                ..
+            } => {
+                let (key, _) = buffer_cache.allocate(
+                    font_system,
+                    text_cache::Key {
+                        content,
+                        size: f32::from(*size),
+                        line_height: f32::from(*line_height),
+                        font: *font,
+                        align_x: *align_x,
+                        bounds: Size {
+                            width: bounds.width,
+                            height: bounds.height,
+                        },
+                        shaping: *shaping,
+                    },
+                );
+
+                Some(Allocation::Cache(key))
+            }
+            Text::Raw { raw, .. } => raw.buffer.upgrade().map(Allocation::Raw),
+        })
+        .collect();
+
+    let text_areas = sections.iter().zip(allocations.iter()).filter_map(
+        |(section, allocation)| {
+            let (buffer, position, color, clip_bounds, transformation) =
+                match section {
+                    Text::Paragraph {
+                        position,
+                        color,
+                        clip_bounds,
+                        transformation,
+                        ..
+                    } => {
+                        let Some(Allocation::Paragraph(paragraph)) = allocation
+                        else {
+                            return None;
+                        };
+
+                        (
+                            paragraph.buffer(),
+                            *position,
+                            *color,
+                            *clip_bounds,
+                            *transformation,
+                        )
+                    }
+                    Text::Editor {
+                        position,
+                        color,
+                        clip_bounds,
+                        transformation,
+                        ..
+                    } => {
+                        let Some(Allocation::Editor(editor)) = allocation
+                        else {
+                            return None;
+                        };
+
+                        (
+                            editor.buffer(),
+                            *position,
+                            *color,
+                            *clip_bounds,
+                            *transformation,
+                        )
+                    }
+                    Text::Cached {
+                        bounds,
+                        align_x,
+                        align_y,
+                        color,
+                        clip_bounds,
+                        ..
+                    } => {
+                        let Some(Allocation::Cache(key)) = allocation else {
+                            return None;
+                        };
+
+                        let entry =
+                            buffer_cache.get(key).expect("Get cached buffer");
+
+                        let mut position = bounds.position();
+
+                        position.x = match align_x {
+                            Alignment::Default
+                            | Alignment::Left
+                            | Alignment::Justified => position.x,
+                            Alignment::Center => {
+                                position.x - entry.min_bounds.width / 2.0
+                            }
+                            Alignment::Right => {
+                                position.x - entry.min_bounds.width
+                            }
+                        };
+
+                        position.y = match align_y {
+                            alignment::Vertical::Top => position.y,
+                            alignment::Vertical::Center => {
+                                position.y - entry.min_bounds.height / 2.0
+                            }
+                            alignment::Vertical::Bottom => {
+                                position.y - entry.min_bounds.height
+                            }
+                        };
+
+                        (
+                            &entry.buffer,
+                            position,
+                            *color,
+                            *clip_bounds,
+                            Transformation::IDENTITY,
+                        )
+                    }
+                    Text::Raw {
+                        raw,
+                        transformation,
+                    } => {
+                        let Some(Allocation::Raw(buffer)) = allocation else {
+                            return None;
+                        };
+
+                        (
+                            buffer.as_ref(),
+                            raw.position,
+                            raw.color,
+                            raw.clip_bounds,
+                            *transformation,
+                        )
+                    }
+                };
+
+            let position = position * transformation * layer_transformation;
+
+            let clip_bounds = layer_bounds.intersection(
+                &(clip_bounds * transformation * layer_transformation),
+            )?;
+
+            Some(cryoglyph::TextArea {
+                buffer,
+                left: position.x,
+                top: position.y,
+                scale: transformation.scale_factor()
+                    * layer_transformation.scale_factor(),
+                bounds: cryoglyph::TextBounds {
+                    left: clip_bounds.x.round() as i32,
+                    top: clip_bounds.y.round() as i32,
+                    right: (clip_bounds.x + clip_bounds.width).round() as i32,
+                    bottom: (clip_bounds.y + clip_bounds.height).round() as i32,
+                },
+                default_color: to_color(color),
+            })
+        },
+    );
+
+    renderer.prepare(
+        device,
+        queue,
+        encoder,
+        font_system,
+        atlas,
+        viewport,
+        text_areas,
+        &mut cryoglyph::SwashCache::new(),
+    )
+}

crates/iced_wgpu/src/triangle.rs

@@ -0,0 +1,963 @@
+//! Draw meshes of triangles.
+mod msaa;
+
+use crate::Buffer;
+use crate::core::{Point, Rectangle, Size, Transformation, Vector};
+use crate::graphics::Antialiasing;
+use crate::graphics::mesh::{self, Mesh};
+
+use rustc_hash::FxHashMap;
+use std::collections::hash_map;
+use std::sync::Weak;
+
+const INITIAL_INDEX_COUNT: usize = 1_000;
+const INITIAL_VERTEX_COUNT: usize = 1_000;
+
+pub type Batch = Vec<Item>;
+
+#[derive(Debug)]
+pub enum Item {
+    Group {
+        transformation: Transformation,
+        meshes: Vec<Mesh>,
+    },
+    Cached {
+        transformation: Transformation,
+        cache: mesh::Cache,
+    },
+}
+
+#[derive(Debug)]
+struct Upload {
+    layer: Layer,
+    transformation: Transformation,
+    version: usize,
+    batch: Weak<[Mesh]>,
+}
+
+#[derive(Debug, Default)]
+pub struct Storage {
+    uploads: FxHashMap<mesh::Id, Upload>,
+}
+
+impl Storage {
+    pub fn new() -> Self {
+        Self::default()
+    }
+
+    fn get(&self, cache: &mesh::Cache) -> Option<&Upload> {
+        if cache.is_empty() {
+            return None;
+        }
+
+        self.uploads.get(&cache.id())
+    }
+
+    fn prepare(
+        &mut self,
+        device: &wgpu::Device,
+        encoder: &mut wgpu::CommandEncoder,
+        belt: &mut wgpu::util::StagingBelt,
+        solid: &solid::Pipeline,
+        gradient: &gradient::Pipeline,
+        cache: &mesh::Cache,
+        new_transformation: Transformation,
+    ) {
+        match self.uploads.entry(cache.id()) {
+            hash_map::Entry::Occupied(entry) => {
+                let upload = entry.into_mut();
+
+                if !cache.is_empty()
+                    && (upload.version != cache.version()
+                        || upload.transformation != new_transformation)
+                {
+                    upload.layer.prepare(
+                        device,
+                        encoder,
+                        belt,
+                        solid,
+                        gradient,
+                        cache.batch(),
+                        new_transformation,
+                    );
+
+                    upload.batch = cache.downgrade();
+                    upload.version = cache.version();
+                    upload.transformation = new_transformation;
+                }
+            }
+            hash_map::Entry::Vacant(entry) => {
+                let mut layer = Layer::new(device, solid, gradient);
+
+                layer.prepare(
+                    device,
+                    encoder,
+                    belt,
+                    solid,
+                    gradient,
+                    cache.batch(),
+                    new_transformation,
+                );
+
+                let _ = entry.insert(Upload {
+                    layer,
+                    transformation: new_transformation,
+                    version: 0,
+                    batch: cache.downgrade(),
+                });
+
+                log::debug!(
+                    "New mesh upload: {:?} (total: {})",
+                    cache.id(),
+                    self.uploads.len()
+                );
+            }
+        }
+    }
+
+    pub fn trim(&mut self) {
+        self.uploads
+            .retain(|_id, upload| upload.batch.strong_count() > 0);
+    }
+}
+
+#[derive(Debug, Clone)]
+pub struct Pipeline {
+    msaa: Option<msaa::Pipeline>,
+    solid: solid::Pipeline,
+    gradient: gradient::Pipeline,
+}
+
+pub struct State {
+    msaa: Option<msaa::State>,
+    layers: Vec<Layer>,
+    prepare_layer: usize,
+    storage: Storage,
+}
+
+impl State {
+    pub fn new(device: &wgpu::Device, pipeline: &Pipeline) -> Self {
+        Self {
+            msaa: pipeline
+                .msaa
+                .as_ref()
+                .map(|pipeline| msaa::State::new(device, pipeline)),
+            layers: Vec::new(),
+            prepare_layer: 0,
+            storage: Storage::new(),
+        }
+    }
+
+    pub fn prepare(
+        &mut self,
+        pipeline: &Pipeline,
+        device: &wgpu::Device,
+        belt: &mut wgpu::util::StagingBelt,
+        encoder: &mut wgpu::CommandEncoder,
+        items: &[Item],
+        scale: Transformation,
+        target_size: Size<u32>,
+    ) {
+        let projection = if let Some((state, pipeline)) =
+            self.msaa.as_mut().zip(pipeline.msaa.as_ref())
+        {
+            state.prepare(device, encoder, belt, pipeline, target_size) * scale
+        } else {
+            Transformation::orthographic(target_size.width, target_size.height)
+                * scale
+        };
+
+        for item in items {
+            match item {
+                Item::Group {
+                    transformation,
+                    meshes,
+                } => {
+                    if self.layers.len() <= self.prepare_layer {
+                        self.layers.push(Layer::new(
+                            device,
+                            &pipeline.solid,
+                            &pipeline.gradient,
+                        ));
+                    }
+
+                    let layer = &mut self.layers[self.prepare_layer];
+                    layer.prepare(
+                        device,
+                        encoder,
+                        belt,
+                        &pipeline.solid,
+                        &pipeline.gradient,
+                        meshes,
+                        projection * *transformation,
+                    );
+
+                    self.prepare_layer += 1;
+                }
+                Item::Cached {
+                    transformation,
+                    cache,
+                } => {
+                    self.storage.prepare(
+                        device,
+                        encoder,
+                        belt,
+                        &pipeline.solid,
+                        &pipeline.gradient,
+                        cache,
+                        projection * *transformation,
+                    );
+                }
+            }
+        }
+    }
+
+    pub fn render(
+        &mut self,
+        pipeline: &Pipeline,
+        encoder: &mut wgpu::CommandEncoder,
+        target: &wgpu::TextureView,
+        start: usize,
+        batch: &Batch,
+        bounds: Rectangle,
+        screen_transformation: Transformation,
+    ) -> usize {
+        let mut layer_count = 0;
+
+        let items = batch.iter().filter_map(|item| match item {
+            Item::Group {
+                transformation,
+                meshes,
+            } => {
+                let layer = &self.layers[start + layer_count];
+                layer_count += 1;
+
+                Some((
+                    layer,
+                    meshes.as_slice(),
+                    screen_transformation * *transformation,
+                ))
+            }
+            Item::Cached {
+                transformation,
+                cache,
+            } => {
+                let upload = self.storage.get(cache)?;
+
+                Some((
+                    &upload.layer,
+                    cache.batch(),
+                    screen_transformation * *transformation,
+                ))
+            }
+        });
+
+        render(
+            encoder,
+            target,
+            self.msaa.as_ref().zip(pipeline.msaa.as_ref()),
+            &pipeline.solid,
+            &pipeline.gradient,
+            bounds,
+            items,
+        );
+
+        layer_count
+    }
+
+    pub fn trim(&mut self) {
+        self.storage.trim();
+
+        self.prepare_layer = 0;
+    }
+}
+
+impl Pipeline {
+    pub fn new(
+        device: &wgpu::Device,
+        format: wgpu::TextureFormat,
+        antialiasing: Option<Antialiasing>,
+    ) -> Pipeline {
+        Pipeline {
+            msaa: antialiasing.map(|a| msaa::Pipeline::new(device, format, a)),
+            solid: solid::Pipeline::new(device, format, antialiasing),
+            gradient: gradient::Pipeline::new(device, format, antialiasing),
+        }
+    }
+}
+
+fn render<'a>(
+    encoder: &mut wgpu::CommandEncoder,
+    target: &wgpu::TextureView,
+    mut msaa: Option<(&msaa::State, &msaa::Pipeline)>,
+    solid: &solid::Pipeline,
+    gradient: &gradient::Pipeline,
+    bounds: Rectangle,
+    group: impl Iterator<Item = (&'a Layer, &'a [Mesh], Transformation)>,
+) {
+    {
+        let mut render_pass = if let Some((_state, pipeline)) = &mut msaa {
+            pipeline.render_pass(encoder)
+        } else {
+            encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
+                label: Some("iced_wgpu.triangle.render_pass"),
+                color_attachments: &[Some(wgpu::RenderPassColorAttachment {
+                    view: target,
+                    depth_slice: None,
+                    resolve_target: None,
+                    ops: wgpu::Operations {
+                        load: wgpu::LoadOp::Load,
+                        store: wgpu::StoreOp::Store,
+                    },
+                })],
+                depth_stencil_attachment: None,
+                timestamp_writes: None,
+                occlusion_query_set: None,
+            })
+        };
+
+        for (layer, meshes, transformation) in group {
+            layer.render(
+                solid,
+                gradient,
+                meshes,
+                bounds,
+                transformation,
+                &mut render_pass,
+            );
+        }
+    }
+
+    if let Some((state, pipeline)) = msaa {
+        state.render(pipeline, encoder, target);
+    }
+}
+
+#[derive(Debug)]
+pub struct Layer {
+    index_buffer: Buffer<u32>,
+    solid: solid::Layer,
+    gradient: gradient::Layer,
+}
+
+impl Layer {
+    fn new(
+        device: &wgpu::Device,
+        solid: &solid::Pipeline,
+        gradient: &gradient::Pipeline,
+    ) -> Self {
+        Self {
+            index_buffer: Buffer::new(
+                device,
+                "iced_wgpu.triangle.index_buffer",
+                INITIAL_INDEX_COUNT,
+                wgpu::BufferUsages::INDEX | wgpu::BufferUsages::COPY_DST,
+            ),
+            solid: solid::Layer::new(device, &solid.constants_layout),
+            gradient: gradient::Layer::new(device, &gradient.constants_layout),
+        }
+    }
+
+    fn prepare(
+        &mut self,
+        device: &wgpu::Device,
+        encoder: &mut wgpu::CommandEncoder,
+        belt: &mut wgpu::util::StagingBelt,
+        solid: &solid::Pipeline,
+        gradient: &gradient::Pipeline,
+        meshes: &[Mesh],
+        transformation: Transformation,
+    ) {
+        // Count the total amount of vertices & indices we need to handle
+        let count = mesh::attribute_count_of(meshes);
+
+        // Then we ensure the current attribute buffers are big enough, resizing if necessary.
+        // We are not currently using the return value of these functions as we have no system in
+        // place to calculate mesh diff, or to know whether or not that would be more performant for
+        // the majority of use cases. Therefore we will write GPU data every frame (for now).
+        let _ = self.index_buffer.resize(device, count.indices);
+        let _ = self.solid.vertices.resize(device, count.solid_vertices);
+        let _ = self
+            .gradient
+            .vertices
+            .resize(device, count.gradient_vertices);
+
+        if self.solid.uniforms.resize(device, count.solids) {
+            self.solid.constants = solid::Layer::bind_group(
+                device,
+                &self.solid.uniforms.raw,
+                &solid.constants_layout,
+            );
+        }
+
+        if self.gradient.uniforms.resize(device, count.gradients) {
+            self.gradient.constants = gradient::Layer::bind_group(
+                device,
+                &self.gradient.uniforms.raw,
+                &gradient.constants_layout,
+            );
+        }
+
+        let mut solid_vertex_offset = 0;
+        let mut solid_uniform_offset = 0;
+        let mut gradient_vertex_offset = 0;
+        let mut gradient_uniform_offset = 0;
+        let mut index_offset = 0;
+
+        for mesh in meshes {
+            let clip_bounds = mesh.clip_bounds() * transformation;
+            let snap_distance = clip_bounds
+                .snap()
+                .map(|snapped_bounds| {
+                    Point::new(snapped_bounds.x as f32, snapped_bounds.y as f32)
+                        - clip_bounds.position()
+                })
+                .unwrap_or(Vector::ZERO);
+
+            let uniforms = Uniforms::new(
+                transformation
+                    * mesh.transformation()
+                    * Transformation::translate(
+                        snap_distance.x,
+                        snap_distance.y,
+                    ),
+            );
+
+            let indices = mesh.indices();
+
+            index_offset += self.index_buffer.write(
+                device,
+                encoder,
+                belt,
+                index_offset,
+                indices,
+            );
+
+            match mesh {
+                Mesh::Solid { buffers, .. } => {
+                    solid_vertex_offset += self.solid.vertices.write(
+                        device,
+                        encoder,
+                        belt,
+                        solid_vertex_offset,
+                        &buffers.vertices,
+                    );
+
+                    solid_uniform_offset += self.solid.uniforms.write(
+                        device,
+                        encoder,
+                        belt,
+                        solid_uniform_offset,
+                        &[uniforms],
+                    );
+                }
+                Mesh::Gradient { buffers, .. } => {
+                    gradient_vertex_offset += self.gradient.vertices.write(
+                        device,
+                        encoder,
+                        belt,
+                        gradient_vertex_offset,
+                        &buffers.vertices,
+                    );
+
+                    gradient_uniform_offset += self.gradient.uniforms.write(
+                        device,
+                        encoder,
+                        belt,
+                        gradient_uniform_offset,
+                        &[uniforms],
+                    );
+                }
+            }
+        }
+    }
+
+    fn render<'a>(
+        &'a self,
+        solid: &'a solid::Pipeline,
+        gradient: &'a gradient::Pipeline,
+        meshes: &[Mesh],
+        bounds: Rectangle,
+        transformation: Transformation,
+        render_pass: &mut wgpu::RenderPass<'a>,
+    ) {
+        let mut num_solids = 0;
+        let mut num_gradients = 0;
+        let mut solid_offset = 0;
+        let mut gradient_offset = 0;
+        let mut index_offset = 0;
+        let mut last_is_solid = None;
+
+        for mesh in meshes {
+            let Some(clip_bounds) = bounds
+                .intersection(&(mesh.clip_bounds() * transformation))
+                .and_then(Rectangle::snap)
+            else {
+                match mesh {
+                    Mesh::Solid { buffers, .. } => {
+                        solid_offset += buffers.vertices.len();
+                        num_solids += 1;
+                    }
+                    Mesh::Gradient { buffers, .. } => {
+                        gradient_offset += buffers.vertices.len();
+                        num_gradients += 1;
+                    }
+                }
+                continue;
+            };
+
+            render_pass.set_scissor_rect(
+                clip_bounds.x,
+                clip_bounds.y,
+                clip_bounds.width,
+                clip_bounds.height,
+            );
+
+            match mesh {
+                Mesh::Solid { buffers, .. } => {
+                    if !last_is_solid.unwrap_or(false) {
+                        render_pass.set_pipeline(&solid.pipeline);
+
+                        last_is_solid = Some(true);
+                    }
+
+                    render_pass.set_bind_group(
+                        0,
+                        &self.solid.constants,
+                        &[(num_solids * std::mem::size_of::<Uniforms>())
+                            as u32],
+                    );
+
+                    render_pass.set_vertex_buffer(
+                        0,
+                        self.solid.vertices.range(
+                            solid_offset,
+                            solid_offset + buffers.vertices.len(),
+                        ),
+                    );
+
+                    num_solids += 1;
+                    solid_offset += buffers.vertices.len();
+                }
+                Mesh::Gradient { buffers, .. } => {
+                    if last_is_solid.unwrap_or(true) {
+                        render_pass.set_pipeline(&gradient.pipeline);
+
+                        last_is_solid = Some(false);
+                    }
+
+                    render_pass.set_bind_group(
+                        0,
+                        &self.gradient.constants,
+                        &[(num_gradients * std::mem::size_of::<Uniforms>())
+                            as u32],
+                    );
+
+                    render_pass.set_vertex_buffer(
+                        0,
+                        self.gradient.vertices.range(
+                            gradient_offset,
+                            gradient_offset + buffers.vertices.len(),
+                        ),
+                    );
+
+                    num_gradients += 1;
+                    gradient_offset += buffers.vertices.len();
+                }
+            };
+
+            render_pass.set_index_buffer(
+                self.index_buffer
+                    .range(index_offset, index_offset + mesh.indices().len()),
+                wgpu::IndexFormat::Uint32,
+            );
+
+            render_pass.draw_indexed(0..mesh.indices().len() as u32, 0, 0..1);
+
+            index_offset += mesh.indices().len();
+        }
+    }
+}
+
+fn fragment_target(
+    texture_format: wgpu::TextureFormat,
+) -> wgpu::ColorTargetState {
+    wgpu::ColorTargetState {
+        format: texture_format,
+        blend: Some(wgpu::BlendState::PREMULTIPLIED_ALPHA_BLENDING),
+        write_mask: wgpu::ColorWrites::ALL,
+    }
+}
+
+fn primitive_state() -> wgpu::PrimitiveState {
+    wgpu::PrimitiveState {
+        topology: wgpu::PrimitiveTopology::TriangleList,
+        front_face: wgpu::FrontFace::Cw,
+        ..Default::default()
+    }
+}
+
+fn multisample_state(
+    antialiasing: Option<Antialiasing>,
+) -> wgpu::MultisampleState {
+    wgpu::MultisampleState {
+        count: antialiasing.map(Antialiasing::sample_count).unwrap_or(1),
+        mask: !0,
+        alpha_to_coverage_enabled: false,
+    }
+}
+
+#[derive(Debug, Clone, Copy, bytemuck::Pod, bytemuck::Zeroable)]
+#[repr(C)]
+pub struct Uniforms {
+    transform: [f32; 16],
+    /// Uniform values must be 256-aligned;
+    /// see: [`wgpu::Limits`] `min_uniform_buffer_offset_alignment`.
+    _padding: [f32; 48],
+}
+
+impl Uniforms {
+    pub fn new(transform: Transformation) -> Self {
+        Self {
+            transform: transform.into(),
+            _padding: [0.0; 48],
+        }
+    }
+
+    pub fn entry() -> wgpu::BindGroupLayoutEntry {
+        wgpu::BindGroupLayoutEntry {
+            binding: 0,
+            visibility: wgpu::ShaderStages::VERTEX_FRAGMENT,
+            ty: wgpu::BindingType::Buffer {
+                ty: wgpu::BufferBindingType::Uniform,
+                has_dynamic_offset: true,
+                min_binding_size: wgpu::BufferSize::new(
+                    std::mem::size_of::<Self>() as u64,
+                ),
+            },
+            count: None,
+        }
+    }
+
+    pub fn min_size() -> Option<wgpu::BufferSize> {
+        wgpu::BufferSize::new(std::mem::size_of::<Self>() as u64)
+    }
+}
+
+mod solid {
+    use crate::Buffer;
+    use crate::graphics::Antialiasing;
+    use crate::graphics::mesh;
+    use crate::triangle;
+
+    #[derive(Debug, Clone)]
+    pub struct Pipeline {
+        pub pipeline: wgpu::RenderPipeline,
+        pub constants_layout: wgpu::BindGroupLayout,
+    }
+
+    #[derive(Debug)]
+    pub struct Layer {
+        pub vertices: Buffer<mesh::SolidVertex2D>,
+        pub uniforms: Buffer<triangle::Uniforms>,
+        pub constants: wgpu::BindGroup,
+    }
+
+    impl Layer {
+        pub fn new(
+            device: &wgpu::Device,
+            constants_layout: &wgpu::BindGroupLayout,
+        ) -> Self {
+            let vertices = Buffer::new(
+                device,
+                "iced_wgpu.triangle.solid.vertex_buffer",
+                triangle::INITIAL_VERTEX_COUNT,
+                wgpu::BufferUsages::VERTEX | wgpu::BufferUsages::COPY_DST,
+            );
+
+            let uniforms = Buffer::new(
+                device,
+                "iced_wgpu.triangle.solid.uniforms",
+                1,
+                wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
+            );
+
+            let constants =
+                Self::bind_group(device, &uniforms.raw, constants_layout);
+
+            Self {
+                vertices,
+                uniforms,
+                constants,
+            }
+        }
+
+        pub fn bind_group(
+            device: &wgpu::Device,
+            buffer: &wgpu::Buffer,
+            layout: &wgpu::BindGroupLayout,
+        ) -> wgpu::BindGroup {
+            device.create_bind_group(&wgpu::BindGroupDescriptor {
+                label: Some("iced_wgpu.triangle.solid.bind_group"),
+                layout,
+                entries: &[wgpu::BindGroupEntry {
+                    binding: 0,
+                    resource: wgpu::BindingResource::Buffer(
+                        wgpu::BufferBinding {
+                            buffer,
+                            offset: 0,
+                            size: triangle::Uniforms::min_size(),
+                        },
+                    ),
+                }],
+            })
+        }
+    }
+
+    impl Pipeline {
+        pub fn new(
+            device: &wgpu::Device,
+            format: wgpu::TextureFormat,
+            antialiasing: Option<Antialiasing>,
+        ) -> Self {
+            let constants_layout = device.create_bind_group_layout(
+                &wgpu::BindGroupLayoutDescriptor {
+                    label: Some("iced_wgpu.triangle.solid.bind_group_layout"),
+                    entries: &[triangle::Uniforms::entry()],
+                },
+            );
+
+            let layout = device.create_pipeline_layout(
+                &wgpu::PipelineLayoutDescriptor {
+                    label: Some("iced_wgpu.triangle.solid.pipeline_layout"),
+                    bind_group_layouts: &[&constants_layout],
+                    push_constant_ranges: &[],
+                },
+            );
+
+            let shader =
+                device.create_shader_module(wgpu::ShaderModuleDescriptor {
+                    label: Some("iced_wgpu.triangle.solid.shader"),
+                    source: wgpu::ShaderSource::Wgsl(
+                        std::borrow::Cow::Borrowed(concat!(
+                            include_str!("shader/triangle.wgsl"),
+                            "\n",
+                            include_str!("shader/triangle/solid.wgsl"),
+                            "\n",
+                            include_str!("shader/color.wgsl"),
+                        )),
+                    ),
+                });
+
+            let pipeline =
+                device.create_render_pipeline(
+                    &wgpu::RenderPipelineDescriptor {
+                        label: Some("iced_wgpu::triangle::solid pipeline"),
+                        layout: Some(&layout),
+                        vertex: wgpu::VertexState {
+                            module: &shader,
+                            entry_point: Some("solid_vs_main"),
+                            buffers: &[wgpu::VertexBufferLayout {
+                                array_stride: std::mem::size_of::<
+                                    mesh::SolidVertex2D,
+                                >(
+                                )
+                                    as u64,
+                                step_mode: wgpu::VertexStepMode::Vertex,
+                                attributes: &wgpu::vertex_attr_array!(
+                                    // Position
+                                    0 => Float32x2,
+                                    // Color
+                                    1 => Float32x4,
+                                ),
+                            }],
+                            compilation_options:
+                                wgpu::PipelineCompilationOptions::default(),
+                        },
+                        fragment: Some(wgpu::FragmentState {
+                            module: &shader,
+                            entry_point: Some("solid_fs_main"),
+                            targets: &[Some(triangle::fragment_target(format))],
+                            compilation_options:
+                                wgpu::PipelineCompilationOptions::default(),
+                        }),
+                        primitive: triangle::primitive_state(),
+                        depth_stencil: None,
+                        multisample: triangle::multisample_state(antialiasing),
+                        multiview: None,
+                        cache: None,
+                    },
+                );
+
+            Self {
+                pipeline,
+                constants_layout,
+            }
+        }
+    }
+}
+
+mod gradient {
+    use crate::Buffer;
+    use crate::graphics::Antialiasing;
+    use crate::graphics::mesh;
+    use crate::triangle;
+
+    #[derive(Debug, Clone)]
+    pub struct Pipeline {
+        pub pipeline: wgpu::RenderPipeline,
+        pub constants_layout: wgpu::BindGroupLayout,
+    }
+
+    #[derive(Debug)]
+    pub struct Layer {
+        pub vertices: Buffer<mesh::GradientVertex2D>,
+        pub uniforms: Buffer<triangle::Uniforms>,
+        pub constants: wgpu::BindGroup,
+    }
+
+    impl Layer {
+        pub fn new(
+            device: &wgpu::Device,
+            constants_layout: &wgpu::BindGroupLayout,
+        ) -> Self {
+            let vertices = Buffer::new(
+                device,
+                "iced_wgpu.triangle.gradient.vertex_buffer",
+                triangle::INITIAL_VERTEX_COUNT,
+                wgpu::BufferUsages::VERTEX | wgpu::BufferUsages::COPY_DST,
+            );
+
+            let uniforms = Buffer::new(
+                device,
+                "iced_wgpu.triangle.gradient.uniforms",
+                1,
+                wgpu::BufferUsages::UNIFORM | wgpu::BufferUsages::COPY_DST,
+            );
+
+            let constants =
+                Self::bind_group(device, &uniforms.raw, constants_layout);
+
+            Self {
+                vertices,
+                uniforms,
+                constants,
+            }
+        }
+
+        pub fn bind_group(
+            device: &wgpu::Device,
+            uniform_buffer: &wgpu::Buffer,
+            layout: &wgpu::BindGroupLayout,
+        ) -> wgpu::BindGroup {
+            device.create_bind_group(&wgpu::BindGroupDescriptor {
+                label: Some("iced_wgpu.triangle.gradient.bind_group"),
+                layout,
+                entries: &[wgpu::BindGroupEntry {
+                    binding: 0,
+                    resource: wgpu::BindingResource::Buffer(
+                        wgpu::BufferBinding {
+                            buffer: uniform_buffer,
+                            offset: 0,
+                            size: triangle::Uniforms::min_size(),
+                        },
+                    ),
+                }],
+            })
+        }
+    }
+
+    impl Pipeline {
+        pub fn new(
+            device: &wgpu::Device,
+            format: wgpu::TextureFormat,
+            antialiasing: Option<Antialiasing>,
+        ) -> Self {
+            let constants_layout = device.create_bind_group_layout(
+                &wgpu::BindGroupLayoutDescriptor {
+                    label: Some(
+                        "iced_wgpu.triangle.gradient.bind_group_layout",
+                    ),
+                    entries: &[triangle::Uniforms::entry()],
+                },
+            );
+
+            let layout = device.create_pipeline_layout(
+                &wgpu::PipelineLayoutDescriptor {
+                    label: Some("iced_wgpu.triangle.gradient.pipeline_layout"),
+                    bind_group_layouts: &[&constants_layout],
+                    push_constant_ranges: &[],
+                },
+            );
+
+            let shader =
+                device.create_shader_module(wgpu::ShaderModuleDescriptor {
+                    label: Some("iced_wgpu.triangle.gradient.shader"),
+                    source: wgpu::ShaderSource::Wgsl(
+                        std::borrow::Cow::Borrowed(concat!(
+                            include_str!("shader/triangle.wgsl"),
+                            "\n",
+                            include_str!("shader/triangle/gradient.wgsl"),
+                            "\n",
+                            include_str!("shader/color.wgsl"),
+                            "\n",
+                            include_str!("shader/color/linear_rgb.wgsl")
+                        )),
+                    ),
+                });
+
+            let pipeline = device.create_render_pipeline(
+                &wgpu::RenderPipelineDescriptor {
+                    label: Some("iced_wgpu.triangle.gradient.pipeline"),
+                    layout: Some(&layout),
+                    vertex: wgpu::VertexState {
+                        module: &shader,
+                        entry_point: Some("gradient_vs_main"),
+                        buffers: &[wgpu::VertexBufferLayout {
+                            array_stride: std::mem::size_of::<
+                                mesh::GradientVertex2D,
+                            >()
+                                as u64,
+                            step_mode: wgpu::VertexStepMode::Vertex,
+                            attributes: &wgpu::vertex_attr_array!(
+                                // Position
+                                0 => Float32x2,
+                                // Colors 1-2
+                                1 => Uint32x4,
+                                // Colors 3-4
+                                2 => Uint32x4,
+                                // Colors 5-6
+                                3 => Uint32x4,
+                                // Colors 7-8
+                                4 => Uint32x4,
+                                // Offsets
+                                5 => Uint32x4,
+                                // Direction
+                                6 => Float32x4
+                            ),
+                        }],
+                        compilation_options:
+                            wgpu::PipelineCompilationOptions::default(),
+                    },
+                    fragment: Some(wgpu::FragmentState {
+                        module: &shader,
+                        entry_point: Some("gradient_fs_main"),
+                        targets: &[Some(triangle::fragment_target(format))],
+                        compilation_options:
+                            wgpu::PipelineCompilationOptions::default(),
+                    }),
+                    primitive: triangle::primitive_state(),
+                    depth_stencil: None,
+                    multisample: triangle::multisample_state(antialiasing),
+                    multiview: None,
+                    cache: None,
+                },
+            );
+
+            Self {
+                pipeline,
+                constants_layout,
+            }
+        }
+    }
+}

crates/iced_wgpu/src/triangle/msaa.rs

@@ -0,0 +1,371 @@
+use crate::core::{Size, Transformation};
+use crate::graphics;
+
+use std::num::NonZeroU64;
+use std::sync::{Arc, RwLock};
+
+#[derive(Debug, Clone)]
+pub struct Pipeline {
+    format: wgpu::TextureFormat,
+    sampler: wgpu::Sampler,
+    raw: wgpu::RenderPipeline,
+    constant_layout: wgpu::BindGroupLayout,
+    texture_layout: wgpu::BindGroupLayout,
+    sample_count: u32,
+    targets: Arc<RwLock<Option<Targets>>>,
+}
+
+impl Pipeline {
+    pub fn new(
+        device: &wgpu::Device,
+        format: wgpu::TextureFormat,
+        antialiasing: graphics::Antialiasing,
+    ) -> Pipeline {
+        let sampler =
+            device.create_sampler(&wgpu::SamplerDescriptor::default());
+
+        let constant_layout =
+            device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
+                label: Some("iced_wgpu::triangle:msaa uniforms layout"),
+                entries: &[
+                    wgpu::BindGroupLayoutEntry {
+                        binding: 0,
+                        visibility: wgpu::ShaderStages::FRAGMENT,
+                        ty: wgpu::BindingType::Sampler(
+                            wgpu::SamplerBindingType::NonFiltering,
+                        ),
+                        count: None,
+                    },
+                    wgpu::BindGroupLayoutEntry {
+                        binding: 1,
+                        visibility: wgpu::ShaderStages::VERTEX,
+                        ty: wgpu::BindingType::Buffer {
+                            ty: wgpu::BufferBindingType::Uniform,
+                            has_dynamic_offset: false,
+                            min_binding_size: None,
+                        },
+                        count: None,
+                    },
+                ],
+            });
+
+        let texture_layout =
+            device.create_bind_group_layout(&wgpu::BindGroupLayoutDescriptor {
+                label: Some("iced_wgpu::triangle::msaa texture layout"),
+                entries: &[wgpu::BindGroupLayoutEntry {
+                    binding: 0,
+                    visibility: wgpu::ShaderStages::FRAGMENT,
+                    ty: wgpu::BindingType::Texture {
+                        sample_type: wgpu::TextureSampleType::Float {
+                            filterable: false,
+                        },
+                        view_dimension: wgpu::TextureViewDimension::D2,
+                        multisampled: false,
+                    },
+                    count: None,
+                }],
+            });
+
+        let layout =
+            device.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
+                label: Some("iced_wgpu::triangle::msaa pipeline layout"),
+                push_constant_ranges: &[],
+                bind_group_layouts: &[&constant_layout, &texture_layout],
+            });
+
+        let shader =
+            device.create_shader_module(wgpu::ShaderModuleDescriptor {
+                label: Some("iced_wgpu triangle blit_shader"),
+                source: wgpu::ShaderSource::Wgsl(std::borrow::Cow::Borrowed(
+                    include_str!("../shader/blit.wgsl"),
+                )),
+            });
+
+        let pipeline =
+            device.create_render_pipeline(&wgpu::RenderPipelineDescriptor {
+                label: Some("iced_wgpu::triangle::msaa pipeline"),
+                layout: Some(&layout),
+                vertex: wgpu::VertexState {
+                    module: &shader,
+                    entry_point: Some("vs_main"),
+                    buffers: &[],
+                    compilation_options:
+                        wgpu::PipelineCompilationOptions::default(),
+                },
+                fragment: Some(wgpu::FragmentState {
+                    module: &shader,
+                    entry_point: Some("fs_main"),
+                    targets: &[Some(wgpu::ColorTargetState {
+                        format,
+                        blend: Some(
+                            wgpu::BlendState::PREMULTIPLIED_ALPHA_BLENDING,
+                        ),
+                        write_mask: wgpu::ColorWrites::ALL,
+                    })],
+                    compilation_options:
+                        wgpu::PipelineCompilationOptions::default(),
+                }),
+                primitive: wgpu::PrimitiveState {
+                    topology: wgpu::PrimitiveTopology::TriangleList,
+                    front_face: wgpu::FrontFace::Cw,
+                    ..Default::default()
+                },
+                depth_stencil: None,
+                multisample: wgpu::MultisampleState {
+                    count: 1,
+                    mask: !0,
+                    alpha_to_coverage_enabled: false,
+                },
+                multiview: None,
+                cache: None,
+            });
+
+        Self {
+            format,
+            sampler,
+            raw: pipeline,
+            constant_layout,
+            texture_layout,
+            sample_count: antialiasing.sample_count(),
+            targets: Arc::new(RwLock::new(None)),
+        }
+    }
+
+    fn targets(
+        &self,
+        device: &wgpu::Device,
+        region_size: Size<u32>,
+    ) -> Targets {
+        let mut targets = self.targets.write().expect("Write MSAA targets");
+
+        match targets.as_mut() {
+            Some(targets)
+                if region_size.width <= targets.size.width
+                    && region_size.height <= targets.size.height => {}
+            _ => {
+                *targets = Some(Targets::new(
+                    device,
+                    self.format,
+                    &self.texture_layout,
+                    self.sample_count,
+                    region_size,
+                ));
+            }
+        }
+
+        targets.as_ref().unwrap().clone()
+    }
+
+    pub fn render_pass<'a>(
+        &self,
+        encoder: &'a mut wgpu::CommandEncoder,
+    ) -> wgpu::RenderPass<'a> {
+        let targets = self.targets.read().expect("Read MSAA targets");
+        let targets = targets.as_ref().unwrap();
+
+        encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
+            label: Some("iced_wgpu.triangle.render_pass"),
+            color_attachments: &[Some(wgpu::RenderPassColorAttachment {
+                view: &targets.attachment,
+                depth_slice: None,
+                resolve_target: Some(&targets.resolve),
+                ops: wgpu::Operations {
+                    load: wgpu::LoadOp::Clear(wgpu::Color::TRANSPARENT),
+                    store: wgpu::StoreOp::Store,
+                },
+            })],
+            depth_stencil_attachment: None,
+            timestamp_writes: None,
+            occlusion_query_set: None,
+        })
+    }
+}
+
+#[derive(Debug, Clone)]
+struct Targets {
+    attachment: wgpu::TextureView,
+    resolve: wgpu::TextureView,
+    bind_group: wgpu::BindGroup,
+    size: Size<u32>,
+}
+
+impl Targets {
+    pub fn new(
+        device: &wgpu::Device,
+        format: wgpu::TextureFormat,
+        texture_layout: &wgpu::BindGroupLayout,
+        sample_count: u32,
+        size: Size<u32>,
+    ) -> Targets {
+        let extent = wgpu::Extent3d {
+            width: size.width,
+            height: size.height,
+            depth_or_array_layers: 1,
+        };
+
+        let attachment = device.create_texture(&wgpu::TextureDescriptor {
+            label: Some("iced_wgpu::triangle::msaa attachment"),
+            size: extent,
+            mip_level_count: 1,
+            sample_count,
+            dimension: wgpu::TextureDimension::D2,
+            format,
+            usage: wgpu::TextureUsages::RENDER_ATTACHMENT,
+            view_formats: &[],
+        });
+
+        let resolve = device.create_texture(&wgpu::TextureDescriptor {
+            label: Some("iced_wgpu::triangle::msaa resolve target"),
+            size: extent,
+            mip_level_count: 1,
+            sample_count: 1,
+            dimension: wgpu::TextureDimension::D2,
+            format,
+            usage: wgpu::TextureUsages::RENDER_ATTACHMENT
+                | wgpu::TextureUsages::TEXTURE_BINDING,
+            view_formats: &[],
+        });
+
+        let attachment =
+            attachment.create_view(&wgpu::TextureViewDescriptor::default());
+
+        let resolve =
+            resolve.create_view(&wgpu::TextureViewDescriptor::default());
+
+        let bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
+            label: Some("iced_wgpu::triangle::msaa texture bind group"),
+            layout: texture_layout,
+            entries: &[wgpu::BindGroupEntry {
+                binding: 0,
+                resource: wgpu::BindingResource::TextureView(&resolve),
+            }],
+        });
+
+        Targets {
+            attachment,
+            resolve,
+            bind_group,
+            size,
+        }
+    }
+}
+
+#[derive(Debug, Clone, Copy, PartialEq, bytemuck::Pod, bytemuck::Zeroable)]
+#[repr(C)]
+struct Ratio {
+    u: f32,
+    v: f32,
+    // Padding field for 16-byte alignment.
+    // See https://docs.rs/wgpu/latest/wgpu/struct.DownlevelFlags.html#associatedconstant.BUFFER_BINDINGS_NOT_16_BYTE_ALIGNED
+    _padding: [f32; 2],
+}
+
+pub struct State {
+    ratio: wgpu::Buffer,
+    constants: wgpu::BindGroup,
+    last_ratio: Option<Ratio>,
+}
+
+impl State {
+    pub fn new(device: &wgpu::Device, pipeline: &Pipeline) -> Self {
+        let ratio = device.create_buffer(&wgpu::BufferDescriptor {
+            label: Some("iced_wgpu::triangle::msaa ratio"),
+            size: std::mem::size_of::<Ratio>() as u64,
+            usage: wgpu::BufferUsages::COPY_DST | wgpu::BufferUsages::UNIFORM,
+            mapped_at_creation: false,
+        });
+
+        let constants = device.create_bind_group(&wgpu::BindGroupDescriptor {
+            label: Some("iced_wgpu::triangle::msaa uniforms bind group"),
+            layout: &pipeline.constant_layout,
+            entries: &[
+                wgpu::BindGroupEntry {
+                    binding: 0,
+                    resource: wgpu::BindingResource::Sampler(&pipeline.sampler),
+                },
+                wgpu::BindGroupEntry {
+                    binding: 1,
+                    resource: ratio.as_entire_binding(),
+                },
+            ],
+        });
+
+        Self {
+            ratio,
+            constants,
+            last_ratio: None,
+        }
+    }
+
+    pub fn prepare(
+        &mut self,
+        device: &wgpu::Device,
+        encoder: &mut wgpu::CommandEncoder,
+        belt: &mut wgpu::util::StagingBelt,
+        pipeline: &Pipeline,
+        region_size: Size<u32>,
+    ) -> Transformation {
+        let targets = pipeline.targets(device, region_size);
+
+        let ratio = Ratio {
+            u: region_size.width as f32 / targets.size.width as f32,
+            v: region_size.height as f32 / targets.size.height as f32,
+            _padding: [0.0; 2],
+        };
+
+        if Some(ratio) != self.last_ratio {
+            belt.write_buffer(
+                encoder,
+                &self.ratio,
+                0,
+                NonZeroU64::new(std::mem::size_of::<Ratio>() as u64)
+                    .expect("non-empty ratio"),
+                device,
+            )
+            .copy_from_slice(bytemuck::bytes_of(&ratio));
+
+            self.last_ratio = Some(ratio);
+        }
+
+        Transformation::orthographic(targets.size.width, targets.size.height)
+    }
+
+    pub fn render(
+        &self,
+        pipeline: &Pipeline,
+        encoder: &mut wgpu::CommandEncoder,
+        target: &wgpu::TextureView,
+    ) {
+        let mut render_pass =
+            encoder.begin_render_pass(&wgpu::RenderPassDescriptor {
+                label: Some("iced_wgpu::triangle::msaa render pass"),
+                color_attachments: &[Some(wgpu::RenderPassColorAttachment {
+                    view: target,
+                    depth_slice: None,
+                    resolve_target: None,
+                    ops: wgpu::Operations {
+                        load: wgpu::LoadOp::Load,
+                        store: wgpu::StoreOp::Store,
+                    },
+                })],
+                depth_stencil_attachment: None,
+                timestamp_writes: None,
+                occlusion_query_set: None,
+            });
+
+        render_pass.set_pipeline(&pipeline.raw);
+        render_pass.set_bind_group(0, &self.constants, &[]);
+        render_pass.set_bind_group(
+            1,
+            &pipeline
+                .targets
+                .read()
+                .expect("Read MSAA targets")
+                .as_ref()
+                .unwrap()
+                .bind_group,
+            &[],
+        );
+        render_pass.draw(0..6, 0..1);
+    }
+}

crates/iced_wgpu/src/window.rs

@@ -0,0 +1,5 @@
+//! Display rendering results on windows.
+pub mod compositor;
+
+pub use compositor::Compositor;
+pub use wgpu::Surface;

crates/iced_wgpu/src/window/compositor.rs

@@ -0,0 +1,358 @@
+//! Connect a window with a renderer.
+use crate::core::Color;
+use crate::graphics::color;
+use crate::graphics::compositor;
+use crate::graphics::error;
+use crate::graphics::{self, Shell, Viewport};
+use crate::settings::{self, Settings};
+use crate::{Engine, Renderer};
+
+/// A window graphics backend for iced powered by `wgpu`.
+pub struct Compositor {
+    instance: wgpu::Instance,
+    adapter: wgpu::Adapter,
+    format: wgpu::TextureFormat,
+    alpha_mode: wgpu::CompositeAlphaMode,
+    engine: Engine,
+    settings: Settings,
+}
+
+/// A compositor error.
+#[derive(Debug, Clone, thiserror::Error)]
+pub enum Error {
+    /// The surface creation failed.
+    #[error("the surface creation failed: {0}")]
+    SurfaceCreationFailed(#[from] wgpu::CreateSurfaceError),
+    /// The surface is not compatible.
+    #[error("the surface is not compatible")]
+    IncompatibleSurface,
+    /// No adapter was found for the options requested.
+    #[error("no adapter was found for the options requested: {0:?}")]
+    NoAdapterFound(String),
+    /// No device request succeeded.
+    #[error("no device request succeeded: {0:?}")]
+    RequestDeviceFailed(Vec<(wgpu::Limits, wgpu::RequestDeviceError)>),
+}
+
+impl From<Error> for graphics::Error {
+    fn from(error: Error) -> Self {
+        Self::GraphicsAdapterNotFound {
+            backend: "wgpu",
+            reason: error::Reason::RequestFailed(error.to_string()),
+        }
+    }
+}
+
+impl Compositor {
+    /// Requests a new [`Compositor`] with the given [`Settings`].
+    ///
+    /// Returns `None` if no compatible graphics adapter could be found.
+    pub async fn request<W: compositor::Window>(
+        settings: Settings,
+        compatible_window: Option<W>,
+        shell: Shell,
+    ) -> Result<Self, Error> {
+        let instance = wgpu::util::new_instance_with_webgpu_detection(&wgpu::InstanceDescriptor {
+            backends: settings.backends,
+            flags: if cfg!(feature = "strict-assertions") {
+                wgpu::InstanceFlags::debugging()
+            } else {
+                wgpu::InstanceFlags::empty()
+            },
+            ..Default::default()
+        })
+        .await;
+
+        log::info!("{settings:#?}");
+
+        #[cfg(not(target_arch = "wasm32"))]
+        if log::max_level() >= log::LevelFilter::Info {
+            let available_adapters: Vec<_> = instance
+                .enumerate_adapters(settings.backends)
+                .iter()
+                .map(wgpu::Adapter::get_info)
+                .collect();
+            log::info!("Available adapters: {available_adapters:#?}");
+        }
+
+        #[allow(unsafe_code)]
+        let compatible_surface =
+            compatible_window.and_then(|window| instance.create_surface(window).ok());
+
+        let adapter_options = wgpu::RequestAdapterOptions {
+            power_preference: wgpu::PowerPreference::from_env()
+                .unwrap_or(wgpu::PowerPreference::HighPerformance),
+            compatible_surface: compatible_surface.as_ref(),
+            force_fallback_adapter: false,
+        };
+
+        let adapter = instance
+            .request_adapter(&adapter_options)
+            .await
+            .map_err(|_error| Error::NoAdapterFound(format!("{adapter_options:?}")))?;
+
+        log::info!("Selected: {:#?}", adapter.get_info());
+
+        let (format, alpha_mode) = compatible_surface
+            .as_ref()
+            .and_then(|surface| {
+                let capabilities = surface.get_capabilities(&adapter);
+
+                let formats = capabilities.formats.iter().copied();
+
+                log::info!("Available formats: {formats:#?}");
+
+                let mut formats =
+                    formats.filter(|format| format.required_features() == wgpu::Features::empty());
+
+                let format = if color::GAMMA_CORRECTION {
+                    formats.find(wgpu::TextureFormat::is_srgb)
+                } else {
+                    formats.find(|format| !wgpu::TextureFormat::is_srgb(format))
+                };
+
+                let format = format.or_else(|| {
+                    log::warn!("No format found!");
+
+                    capabilities.formats.first().copied()
+                });
+
+                let alpha_modes = capabilities.alpha_modes;
+
+                log::info!("Available alpha modes: {alpha_modes:#?}");
+
+                let preferred_alpha =
+                    if alpha_modes.contains(&wgpu::CompositeAlphaMode::PostMultiplied) {
+                        wgpu::CompositeAlphaMode::PostMultiplied
+                    } else if alpha_modes.contains(&wgpu::CompositeAlphaMode::PreMultiplied) {
+                        wgpu::CompositeAlphaMode::PreMultiplied
+                    } else {
+                        wgpu::CompositeAlphaMode::Auto
+                    };
+
+                format.zip(Some(preferred_alpha))
+            })
+            .ok_or(Error::IncompatibleSurface)?;
+
+        log::info!("Selected format: {format:?} with alpha mode: {alpha_mode:?}");
+
+        #[cfg(target_arch = "wasm32")]
+        let limits = if adapter.get_info().backend == wgpu::Backend::BrowserWebGpu {
+            vec![
+                wgpu::Limits::default().using_resolution(adapter.limits()),
+                wgpu::Limits::downlevel_webgl2_defaults().using_resolution(adapter.limits()),
+            ]
+        } else {
+            vec![wgpu::Limits::downlevel_webgl2_defaults().using_resolution(adapter.limits())]
+        };
+
+        #[cfg(not(target_arch = "wasm32"))]
+        let limits = vec![wgpu::Limits::default(), wgpu::Limits::downlevel_defaults()];
+
+        let limits = limits.into_iter().map(|limits| wgpu::Limits {
+            max_bind_groups: 2,
+            max_non_sampler_bindings: 2048,
+            ..limits
+        });
+
+        let mut errors = Vec::new();
+
+        for required_limits in limits {
+            let result = adapter
+                .request_device(&wgpu::DeviceDescriptor {
+                    label: Some("iced_wgpu::window::compositor device descriptor"),
+                    required_features: wgpu::Features::empty(),
+                    required_limits: required_limits.clone(),
+                    memory_hints: wgpu::MemoryHints::MemoryUsage,
+                    trace: wgpu::Trace::Off,
+                    experimental_features: wgpu::ExperimentalFeatures::disabled(),
+                })
+                .await;
+
+            match result {
+                Ok((device, queue)) => {
+                    let engine = Engine::new(
+                        &adapter,
+                        device,
+                        queue,
+                        format,
+                        settings.antialiasing,
+                        shell,
+                    );
+
+                    return Ok(Compositor {
+                        instance,
+                        adapter,
+                        format,
+                        alpha_mode,
+                        engine,
+                        settings,
+                    });
+                }
+                Err(error) => {
+                    errors.push((required_limits, error));
+                }
+            }
+        }
+
+        Err(Error::RequestDeviceFailed(errors))
+    }
+}
+
+/// Creates a [`Compositor`] with the given [`Settings`] and window.
+pub async fn new<W: compositor::Window>(
+    settings: Settings,
+    compatible_window: W,
+    shell: Shell,
+) -> Result<Compositor, Error> {
+    Compositor::request(settings, Some(compatible_window), shell).await
+}
+
+/// Presents the given primitives with the given [`Compositor`].
+pub fn present(
+    renderer: &mut Renderer,
+    surface: &mut wgpu::Surface<'static>,
+    viewport: &Viewport,
+    background_color: Color,
+    on_pre_present: impl FnOnce(),
+) -> Result<(), compositor::SurfaceError> {
+    match surface.get_current_texture() {
+        Ok(frame) => {
+            let view = &frame
+                .texture
+                .create_view(&wgpu::TextureViewDescriptor::default());
+
+            let _submission = renderer.present(
+                Some(background_color),
+                frame.texture.format(),
+                view,
+                viewport,
+            );
+
+            // Present the frame
+            on_pre_present();
+            frame.present();
+
+            Ok(())
+        }
+        Err(error) => match error {
+            wgpu::SurfaceError::Timeout => Err(compositor::SurfaceError::Timeout),
+            wgpu::SurfaceError::Outdated => Err(compositor::SurfaceError::Outdated),
+            wgpu::SurfaceError::Lost => Err(compositor::SurfaceError::Lost),
+            wgpu::SurfaceError::OutOfMemory => Err(compositor::SurfaceError::OutOfMemory),
+            wgpu::SurfaceError::Other => Err(compositor::SurfaceError::Other),
+        },
+    }
+}
+
+impl graphics::Compositor for Compositor {
+    type Renderer = Renderer;
+    type Surface = wgpu::Surface<'static>;
+
+    async fn with_backend(
+        settings: graphics::Settings,
+        _display: impl compositor::Display,
+        compatible_window: impl compositor::Window,
+        shell: Shell,
+        backend: Option<&str>,
+    ) -> Result<Self, graphics::Error> {
+        match backend {
+            None | Some("wgpu") => {
+                let mut settings = Settings::from(settings);
+
+                if let Some(backends) = wgpu::Backends::from_env() {
+                    settings.backends = backends;
+                }
+
+                if let Some(present_mode) = settings::present_mode_from_env() {
+                    settings.present_mode = present_mode;
+                }
+
+                Ok(new(settings, compatible_window, shell).await?)
+            }
+            Some(backend) => Err(graphics::Error::GraphicsAdapterNotFound {
+                backend: "wgpu",
+                reason: error::Reason::DidNotMatch {
+                    preferred_backend: backend.to_owned(),
+                },
+            }),
+        }
+    }
+
+    fn create_renderer(&self) -> Self::Renderer {
+        Renderer::new(
+            self.engine.clone(),
+            self.settings.default_font,
+            self.settings.default_text_size,
+        )
+    }
+
+    fn create_surface<W: compositor::Window>(
+        &mut self,
+        window: W,
+        width: u32,
+        height: u32,
+    ) -> Self::Surface {
+        let mut surface = self
+            .instance
+            .create_surface(window)
+            .expect("Create surface");
+
+        if width > 0 && height > 0 {
+            self.configure_surface(&mut surface, width, height);
+        }
+
+        surface
+    }
+
+    fn configure_surface(&mut self, surface: &mut Self::Surface, width: u32, height: u32) {
+        surface.configure(
+            &self.engine.device,
+            &wgpu::SurfaceConfiguration {
+                usage: wgpu::TextureUsages::RENDER_ATTACHMENT,
+                format: self.format,
+                present_mode: self.settings.present_mode,
+                width,
+                height,
+                alpha_mode: self.alpha_mode,
+                view_formats: vec![],
+                desired_maximum_frame_latency: 1,
+            },
+        );
+    }
+
+    fn information(&self) -> compositor::Information {
+        let information = self.adapter.get_info();
+
+        compositor::Information {
+            adapter: information.name,
+            backend: format!("{:?}", information.backend),
+        }
+    }
+
+    fn present(
+        &mut self,
+        renderer: &mut Self::Renderer,
+        surface: &mut Self::Surface,
+        viewport: &Viewport,
+        background_color: Color,
+        on_pre_present: impl FnOnce(),
+    ) -> Result<(), compositor::SurfaceError> {
+        present(
+            renderer,
+            surface,
+            viewport,
+            background_color,
+            on_pre_present,
+        )
+    }
+
+    fn screenshot(
+        &mut self,
+        renderer: &mut Self::Renderer,
+        viewport: &Viewport,
+        background_color: Color,
+    ) -> Vec<u8> {
+        renderer.screenshot(viewport, background_color)
+    }
+}

crates/ocs_web_worker/Cargo.toml

@@ -0,0 +1,15 @@
+[package]
+name = "ocs_web_worker"
+version = "0.1.0"
+edition = "2021"
+publish = false
+
+[lib]
+crate-type = ["cdylib"]
+
+[dependencies]
+acadrust = { version = "0.4", features = ["serde"] }
+bincode = "1.3"
+getrandom = { version = "0.3", features = ["wasm_js"] }
+js-sys = "0.3"
+wasm-bindgen = "0.2"

crates/ocs_web_worker/src/lib.rs

@@ -0,0 +1,32 @@
+use std::io::Cursor;
+
+use acadrust::io::dwg::DwgReader;
+use acadrust::DxfReader;
+use js_sys::Uint8Array;
+use wasm_bindgen::prelude::*;
+
+/// Parse DWG/DXF on a dedicated browser worker and return a compact serialized
+/// document. The main wasm instance only deserializes and installs it, so the
+/// expensive bit/handle/object decode never occupies the browser UI thread.
+#[wasm_bindgen]
+pub fn parse_document(name: String, bytes: Uint8Array) -> Result<Uint8Array, JsValue> {
+    let bytes = bytes.to_vec();
+    let ext = name.rsplit('.').next().unwrap_or_default().to_lowercase();
+    let document = match ext.as_str() {
+        "dwg" => DwgReader::from_stream(Cursor::new(bytes))
+            .read()
+            .map_err(|error| JsValue::from_str(&error.to_string()))?,
+        "dxf" => DxfReader::from_reader(Cursor::new(bytes))
+            .map_err(|error| JsValue::from_str(&error.to_string()))?
+            .read()
+            .map_err(|error| JsValue::from_str(&error.to_string()))?,
+        _ => {
+            return Err(JsValue::from_str(&format!(
+                "Unsupported file format: .{ext}"
+            )))
+        }
+    };
+    let encoded =
+        bincode::serialize(&document).map_err(|error| JsValue::from_str(&error.to_string()))?;
+    Ok(Uint8Array::from(encoded.as_slice()))
+}

index.html

@@ -45,6 +45,8 @@
     <!-- Per-script Noto subsets, fetched lazily at runtime (one alphabet per
          file) so CAD text renders non-Latin scripts on the web. (#141) -->
     <link data-trunk rel="copy-dir" href="web/fonts" />
+    <link data-trunk rel="copy-file" href="web/ocs-parse-worker.js" />
+    <link data-trunk rel="copy-dir" href="web/worker_pkg" />
   </head>
   <body>
     <div id="loading">

scripts/build-web-worker.sh

@@ -0,0 +1,10 @@
+#!/bin/sh
+set -eu
+
+cargo build --release --target wasm32-unknown-unknown --package ocs_web_worker
+mkdir -p web/worker_pkg
+wasm-bindgen \
+  --target web \
+  --out-dir web/worker_pkg \
+  --out-name ocs_web_worker \
+  target/wasm32-unknown-unknown/release/ocs_web_worker.wasm

src/app/command_driver.rs

@@ -1052,9 +1052,10 @@ impl OpenCADStudio {
                 self.tabs[i].active_cmd = None;
                 self.tabs[i].snap_result = None;
                 self.tabs[i].scene.clear_preview_wire();
-                self.push_undo_snapshot(i, "GROUP");
+                let undo = self.begin_group_undo(i, "GROUP");
                 self.tabs[i].scene.create_group(name.clone(), handles);
                 self.tabs[i].dirty = true;
+                self.commit_group_undo(i, undo);
                 self.command_line
                     .push_info(&format!("Group \"{}\" created.", name));
             }
@@ -1062,9 +1063,10 @@ impl OpenCADStudio {
                 self.tabs[i].active_cmd = None;
                 self.tabs[i].snap_result = None;
                 self.tabs[i].scene.clear_preview_wire();
-                self.push_undo_snapshot(i, "UNGROUP");
+                let undo = self.begin_group_undo(i, "UNGROUP");
                 let count = self.tabs[i].scene.delete_groups_containing(&handles);
                 self.tabs[i].dirty = true;
+                self.commit_group_undo(i, undo);
                 if count > 0 {
                     self.command_line
                         .push_info(&format!("{} group(s) dissolved.", count));

src/app/commands/layerprops.rs

@@ -817,10 +817,15 @@ impl OpenCADStudio {
                             self.command_line
                                 .push_error(&format!("DIMSTYLE: '{}' already exists.", name));
                         } else {
+                            let undo = self.begin_dim_style_undo(
+                                i,
+                                "DIMSTYLE NEW",
+                                std::slice::from_ref(&name),
+                            );
                             let style = DimStyle::new(&name);
                             let _ = self.tabs[i].scene.document.dim_styles.add(style);
-                            self.push_undo_snapshot(i, "DIMSTYLE NEW");
                             self.tabs[i].dirty = true;
+                            self.commit_dim_style_undo(i, undo);
                             self.command_line
                                 .push_output(&format!("DIMSTYLE: '{}' created.", name));
                         }
@@ -832,6 +837,11 @@ impl OpenCADStudio {
                         let prop = parts.get(2).map(|s| s.to_lowercase()).unwrap_or_default();
                         let val_str = parts.get(3).map(|s| s.trim()).unwrap_or("");
                         if let Ok(val) = val_str.parse::<f64>() {
+                            let undo = self.begin_dim_style_undo(
+                                i,
+                                "DIMSTYLE SET",
+                                std::slice::from_ref(&style_name),
+                            );
                             if let Some(ds) =
                                 self.tabs[i].scene.document.dim_styles.get_mut(&style_name)
                             {
@@ -882,8 +892,10 @@ impl OpenCADStudio {
                                         return Some(Task::none());
                                     }
                                 }
-                                self.push_undo_snapshot(i, "DIMSTYLE SET");
                                 self.tabs[i].dirty = true;
+                                self.tabs[i].scene
+                                    .invalidate_dim_style_dependencies(&style_name);
+                                self.commit_dim_style_undo(i, undo);
                                 self.command_line.push_output(&format!(
                                     "DIMSTYLE: '{style_name}'.{prop} = {val:.3}"
                                 ));
@@ -1121,10 +1133,15 @@ impl OpenCADStudio {
                             self.command_line
                                 .push_error(&format!("{prefix}: style '{name}' already exists."));
                         } else {
+                            let undo = self.begin_text_style_undo(
+                                i,
+                                "STYLE NEW",
+                                std::slice::from_ref(&name),
+                            );
                             let style = acadrust::tables::TextStyle::new(&name);
                             let _ = self.tabs[i].scene.document.text_styles.add(style);
-                            self.push_undo_snapshot(i, "STYLE NEW");
                             self.tabs[i].dirty = true;
+                            self.commit_text_style_undo(i, undo);
                             self.command_line
                                 .push_output(&format!("{prefix}: style '{name}' created."));
                         }
@@ -1136,18 +1153,28 @@ impl OpenCADStudio {
                         if style_name.is_empty() || font.is_empty() {
                             self.command_line
                                 .push_error(&format!("Usage: {prefix} FONT <style> <font_file>"));
-                        } else if let Some(s) =
+                        } else {
+                            let undo = self.begin_text_style_undo(
+                                i,
+                                "STYLE FONT",
+                                std::slice::from_ref(&style_name),
+                            );
+                            if let Some(style) =
                             self.tabs[i].scene.document.text_styles.get_mut(&style_name)
                         {
-                            s.font_file = font.clone();
-                            self.push_undo_snapshot(i, "STYLE FONT");
+                                style.font_file = font.clone();
                             self.tabs[i].dirty = true;
+                                self.tabs[i]
+                                    .scene
+                                    .invalidate_text_style_dependencies(&style_name);
+                                self.commit_text_style_undo(i, undo);
                             self.command_line.push_output(&format!(
                                 "{prefix}: '{style_name}' font set to '{font}'."
                             ));
                         } else {
                             self.command_line
                                 .push_error(&format!("{prefix}: style '{style_name}' not found."));
+                            }
                         }
                     }
                     "WIDTH" | "W" => {
@@ -1155,12 +1182,20 @@ impl OpenCADStudio {
                         let style_name = parts.get(1).map(|s| s.trim()).unwrap_or("").to_string();
                         let factor_str = parts.get(2).map(|s| s.trim()).unwrap_or("");
                         if let Ok(factor) = factor_str.parse::<f64>() {
-                            if let Some(s) =
+                            let undo = self.begin_text_style_undo(
+                                i,
+                                "STYLE WIDTH",
+                                std::slice::from_ref(&style_name),
+                            );
+                            if let Some(style) =
                                 self.tabs[i].scene.document.text_styles.get_mut(&style_name)
                             {
-                                s.width_factor = factor;
-                                self.push_undo_snapshot(i, "STYLE WIDTH");
+                                style.width_factor = factor;
                                 self.tabs[i].dirty = true;
+                                self.tabs[i]
+                                    .scene
+                                    .invalidate_text_style_dependencies(&style_name);
+                                self.commit_text_style_undo(i, undo);
                                 self.command_line.push_output(&format!(
                                     "{prefix}: '{style_name}' width factor set to {factor:.3}."
                                 ));
@@ -1179,12 +1214,20 @@ impl OpenCADStudio {
                         let style_name = parts.get(1).map(|s| s.trim()).unwrap_or("").to_string();
                         let angle_str = parts.get(2).map(|s| s.trim()).unwrap_or("");
                         if let Ok(deg) = angle_str.parse::<f64>() {
-                            if let Some(s) =
+                            let undo = self.begin_text_style_undo(
+                                i,
+                                "STYLE OBLIQUE",
+                                std::slice::from_ref(&style_name),
+                            );
+                            if let Some(style) =
                                 self.tabs[i].scene.document.text_styles.get_mut(&style_name)
                             {
-                                s.oblique_angle = deg.to_radians();
-                                self.push_undo_snapshot(i, "STYLE OBLIQUE");
+                                style.oblique_angle = deg.to_radians();
                                 self.tabs[i].dirty = true;
+                                self.tabs[i]
+                                    .scene
+                                    .invalidate_text_style_dependencies(&style_name);
+                                self.commit_text_style_undo(i, undo);
                                 self.command_line.push_output(&format!(
                                     "{prefix}: '{style_name}' oblique angle set to {deg:.1}°."
                                 ));

src/app/commands/layers.rs

@@ -23,7 +23,8 @@ impl OpenCADStudio {
                         .into_iter()
                         .map(|(_, e)| e.common().layer.clone())
                         .collect();
-                    self.push_undo_snapshot(i, "LAYOFF");
+                    let names: Vec<String> = layers.iter().cloned().collect();
+                    let undo = self.begin_layer_undo(i, "LAYOFF", &names);
                     for name in &layers {
                         if name == "0" {
                             continue;
@@ -32,8 +33,9 @@ impl OpenCADStudio {
                             dl.turn_off();
                         }
                     }
-                    self.tabs[i].scene.bump_geometry();
+                    self.tabs[i].scene.invalidate_layer_dependencies(&names);
                     self.tabs[i].dirty = true;
+                    self.commit_layer_undo(i, undo);
                     self.refresh_layer_panel();
                     self.command_line.push_info("Layer(s) turned off.");
                 }
@@ -58,7 +60,8 @@ impl OpenCADStudio {
                         .into_iter()
                         .map(|(_, e)| e.common().layer.clone())
                         .collect();
-                    self.push_undo_snapshot(i, "LAYFRZ");
+                    let names: Vec<String> = layers.iter().cloned().collect();
+                    let undo = self.begin_layer_undo(i, "LAYFRZ", &names);
                     for name in &layers {
                         if name == "0" {
                             continue;
@@ -67,8 +70,9 @@ impl OpenCADStudio {
                             dl.freeze();
                         }
                     }
-                    self.tabs[i].scene.bump_geometry();
+                    self.tabs[i].scene.invalidate_layer_dependencies(&names);
                     self.tabs[i].dirty = true;
+                    self.commit_layer_undo(i, undo);
                     self.refresh_layer_panel();
                     self.command_line.push_info("Layer(s) frozen.");
                 }
@@ -266,10 +270,18 @@ impl OpenCADStudio {
                                 "LAYERSTATE: no saved state named \"{arg}\"."
                             ));
                         } else {
-                            self.push_undo_snapshot(i, "LAYERSTATE");
+                            let names: Vec<String> = self.tabs[i]
+                                .scene
+                                .document
+                                .layers
+                                .iter()
+                                .map(|layer| layer.name.clone())
+                                .collect();
+                            let undo = self.begin_layer_undo(i, "LAYERSTATE", &names);
                             let n = self.tabs[i].restore_layer_state(arg).unwrap_or(0);
-                            self.tabs[i].scene.bump_geometry();
+                            self.tabs[i].scene.invalidate_layer_dependencies(&names);
                             self.tabs[i].dirty = true;
+                            self.commit_layer_undo(i, undo);
                             self.refresh_layer_panel();
                             self.command_line.push_output(&format!(
                                 "LAYERSTATE: restored \"{arg}\" ({n} layer(s))."
@@ -312,14 +324,16 @@ impl OpenCADStudio {
                         .into_iter()
                         .map(|(_, e)| e.common().layer.clone())
                         .collect();
-                    self.push_undo_snapshot(i, "LAYLCK");
+                    let names: Vec<String> = layers.iter().cloned().collect();
+                    let undo = self.begin_layer_undo(i, "LAYLCK", &names);
                     for name in &layers {
                         if let Some(dl) = self.tabs[i].scene.document.layers.get_mut(name) {
                             dl.lock();
                         }
                     }
-                    self.tabs[i].scene.bump_geometry();
+                    // Layer locking changes editability only.
                     self.tabs[i].dirty = true;
+                    self.commit_layer_undo(i, undo);
                     self.refresh_layer_panel();
                     self.command_line.push_info("Layer(s) locked.");
                 }
@@ -359,41 +373,43 @@ impl OpenCADStudio {
             }
 
             "LAYON" => {
-                self.push_undo_snapshot(i, "LAYON");
-                for name in self.tabs[i]
+                let names = self.tabs[i]
                     .scene
                     .document
                     .layers
                     .iter()
                     .map(|l| l.name.clone())
-                    .collect::<Vec<_>>()
-                {
+                    .collect::<Vec<_>>();
+                let undo = self.begin_layer_undo(i, "LAYON", &names);
+                for name in &names {
                     if let Some(dl) = self.tabs[i].scene.document.layers.get_mut(&name) {
                         dl.turn_on();
                     }
                 }
-                self.tabs[i].scene.bump_geometry();
+                self.tabs[i].scene.invalidate_layer_dependencies(&names);
                 self.tabs[i].dirty = true;
+                self.commit_layer_undo(i, undo);
                 self.refresh_layer_panel();
                 self.command_line.push_info("All layers turned on.");
             }
 
             "LAYTHW" => {
-                self.push_undo_snapshot(i, "LAYTHW");
-                for name in self.tabs[i]
+                let names = self.tabs[i]
                     .scene
                     .document
                     .layers
                     .iter()
                     .map(|l| l.name.clone())
-                    .collect::<Vec<_>>()
-                {
+                    .collect::<Vec<_>>();
+                let undo = self.begin_layer_undo(i, "LAYTHW", &names);
+                for name in &names {
                     if let Some(dl) = self.tabs[i].scene.document.layers.get_mut(&name) {
                         dl.thaw();
                     }
                 }
-                self.tabs[i].scene.bump_geometry();
+                self.tabs[i].scene.invalidate_layer_dependencies(&names);
                 self.tabs[i].dirty = true;
+                self.commit_layer_undo(i, undo);
                 self.refresh_layer_panel();
                 self.command_line.push_info("All layers thawed.");
             }
@@ -417,14 +433,16 @@ impl OpenCADStudio {
                         .into_iter()
                         .map(|(_, e)| e.common().layer.clone())
                         .collect();
-                    self.push_undo_snapshot(i, "LAYULK");
+                    let names: Vec<String> = layers.iter().cloned().collect();
+                    let undo = self.begin_layer_undo(i, "LAYULK", &names);
                     for name in &layers {
                         if let Some(dl) = self.tabs[i].scene.document.layers.get_mut(name) {
                             dl.unlock();
                         }
                     }
-                    self.tabs[i].scene.bump_geometry();
+                    // Layer unlocking changes editability only.
                     self.tabs[i].dirty = true;
+                    self.commit_layer_undo(i, undo);
                     self.refresh_layer_panel();
                     self.command_line.push_info("Layer(s) unlocked.");
                 }
@@ -442,7 +460,6 @@ impl OpenCADStudio {
                     self.command_line
                         .push_error("LAYISO: select entities on the layers to isolate first.");
                 } else {
-                    self.push_undo_snapshot(i, "LAYISO");
                     let names: Vec<String> = self.tabs[i]
                         .scene
                         .document
@@ -450,15 +467,17 @@ impl OpenCADStudio {
                         .iter()
                         .map(|l| l.name.clone())
                         .collect();
-                    for name in names {
-                        if !sel_layers.contains(&name) {
-                            if let Some(dl) = self.tabs[i].scene.document.layers.get_mut(&name) {
+                    let undo = self.begin_layer_undo(i, "LAYISO", &names);
+                    for name in &names {
+                        if !sel_layers.contains(name) {
+                            if let Some(dl) = self.tabs[i].scene.document.layers.get_mut(name) {
                                 dl.turn_off();
                             }
                         }
                     }
-                    self.tabs[i].scene.bump_geometry();
+                    self.tabs[i].scene.invalidate_layer_dependencies(&names);
                     self.tabs[i].dirty = true;
+                    self.commit_layer_undo(i, undo);
                     self.refresh_layer_panel();
                     self.command_line
                         .push_info(&format!("LAYISO: isolated {} layer(s).", sel_layers.len()));
@@ -511,7 +530,6 @@ impl OpenCADStudio {
 
             // LAYUNISO — restore all layers that were turned off by LAYISO (turn all on)
             "LAYUNISO" => {
-                self.push_undo_snapshot(i, "LAYUNISO");
                 let names: Vec<String> = self.tabs[i]
                     .scene
                     .document
@@ -519,13 +537,15 @@ impl OpenCADStudio {
                     .iter()
                     .map(|l| l.name.clone())
                     .collect();
-                for name in names {
+                let undo = self.begin_layer_undo(i, "LAYUNISO", &names);
+                for name in &names {
                     if let Some(dl) = self.tabs[i].scene.document.layers.get_mut(&name) {
                         dl.turn_on();
                     }
                 }
-                self.tabs[i].scene.bump_geometry();
+                self.tabs[i].scene.invalidate_layer_dependencies(&names);
                 self.tabs[i].dirty = true;
+                self.commit_layer_undo(i, undo);
                 self.refresh_layer_panel();
                 self.command_line
                     .push_info("LAYUNISO: all layers restored.");
@@ -587,9 +607,10 @@ impl OpenCADStudio {
                     self.command_line.push_info(&cmd.prompt());
                     self.tabs[i].active_cmd = Some(Box::new(cmd));
                 } else {
-                    self.push_undo_snapshot(i, "UNGROUP");
+                    let undo = self.begin_group_undo(i, "UNGROUP");
                     let count = self.tabs[i].scene.delete_groups_containing(&handles);
                     self.tabs[i].dirty = true;
+                    self.commit_group_undo(i, undo);
                     if count > 0 {
                         self.command_line
                             .push_info(&format!("{} group(s) dissolved.", count));

src/app/document.rs

@@ -520,7 +520,7 @@ impl HistorySnapshot {
                 .saturating_add(
                     d.structure
                         .as_ref()
-                        .map_or(0, |doc| doc.objects.len().saturating_mul(192)),
+                        .map_or(0, StructureSnapshot::estimated_bytes),
                 )
                 .saturating_add(d.selected_before.len().saturating_mul(16))
                 .saturating_add(d.selected_after.len().saturating_mul(16))
@@ -546,10 +546,67 @@ pub(super) struct DeltaSnapshot {
     pub(super) dirty_after: bool,
     /// Opposite non-entity document state. `apply_delta_state` swaps this with
     /// the live structure, so the same allocation shuttles between undo/redo.
-    pub(super) structure: Option<CadDocument>,
+    pub(super) structure: Option<StructureSnapshot>,
     pub(super) label: String,
 }
 
+#[derive(Clone)]
+pub(super) enum StructureSnapshot {
+    /// Compatibility fallback for genuinely broad structural commands.
+    Full(CadDocument),
+    /// Exact layer-table entries touched by one command.
+    Layers(Vec<TableEntryDelta<acadrust::tables::Layer>>),
+    /// Exact text-style entries touched by one command.
+    TextStyles(Vec<TableEntryDelta<acadrust::tables::TextStyle>>),
+    /// Exact dimension-style entries touched by one command.
+    DimStyles(Vec<TableEntryDelta<acadrust::tables::DimStyle>>),
+    /// Exact object-map entries touched by one command. This supports commands
+    /// such as groups/dictionaries without retaining every unrelated object.
+    Objects(Vec<ObjectEntryDelta>),
+    /// The bounded set of style tables, style objects, current-style pointers,
+    /// and matching ribbon state touched by one Style Manager transaction.
+    Styles {
+        before: super::style_ops::StyleStateSnapshot,
+        after: super::style_ops::StyleStateSnapshot,
+        text_names: Vec<String>,
+        dim_names: Vec<String>,
+        object_handles: Vec<Handle>,
+    },
+}
+
+impl StructureSnapshot {
+    pub(super) fn estimated_bytes(&self) -> usize {
+        match self {
+            Self::Full(doc) => doc.objects.len().saturating_mul(192),
+            Self::Layers(entries) => entries.len().saturating_mul(256),
+            Self::TextStyles(entries) => entries.len().saturating_mul(320),
+            Self::DimStyles(entries) => entries.len().saturating_mul(1024),
+            Self::Objects(entries) => entries.len().saturating_mul(384),
+            Self::Styles { before, after, .. } => before
+                .estimated_bytes()
+                .saturating_add(after.estimated_bytes()),
+        }
+    }
+
+    pub(super) fn is_full(&self) -> bool {
+        matches!(self, Self::Full(_))
+    }
+}
+
+#[derive(Clone)]
+pub(super) struct TableEntryDelta<T> {
+    pub(super) name: String,
+    pub(super) before: Option<T>,
+    pub(super) after: Option<T>,
+}
+
+#[derive(Clone)]
+pub(super) struct ObjectEntryDelta {
+    pub(super) handle: Handle,
+    pub(super) before: Option<acadrust::objects::ObjectType>,
+    pub(super) after: Option<acadrust::objects::ObjectType>,
+}
+
 #[derive(Default)]
 pub(super) struct HistoryState {
     pub(super) undo_stack: Vec<HistorySnapshot>,

src/app/history.rs

@@ -1,5 +1,8 @@
 use super::{
-    document::{DeltaSnapshot, HistorySnapshot, PendingHistorySnapshot},
+    document::{
+        DeltaSnapshot, HistorySnapshot, ObjectEntryDelta, PendingHistorySnapshot,
+        StructureSnapshot, TableEntryDelta,
+    },
     OpenCADStudio,
 };
 use acadrust::{EntityType, Handle};
@@ -71,6 +74,38 @@ pub(super) struct PendingDelta {
     structure_before: Option<acadrust::CadDocument>,
 }
 
+pub(super) struct PendingLayerDelta {
+    label: String,
+    current_layout: String,
+    selected_before: Vec<Handle>,
+    dirty_before: bool,
+    before: Vec<(String, Option<acadrust::tables::Layer>)>,
+}
+
+pub(super) struct PendingTextStyleDelta {
+    label: String,
+    current_layout: String,
+    selected_before: Vec<Handle>,
+    dirty_before: bool,
+    before: Vec<(String, Option<acadrust::tables::TextStyle>)>,
+}
+
+pub(super) struct PendingDimStyleDelta {
+    label: String,
+    current_layout: String,
+    selected_before: Vec<Handle>,
+    dirty_before: bool,
+    before: Vec<(String, Option<acadrust::tables::DimStyle>)>,
+}
+
+pub(super) struct PendingObjectDelta {
+    label: String,
+    current_layout: String,
+    selected_before: Vec<Handle>,
+    dirty_before: bool,
+    before: FxHashMap<Handle, acadrust::objects::ObjectType>,
+}
+
 impl OpenCADStudio {
     pub(super) fn history_label_from_active_cmd(&self, i: usize, fallback: &'static str) -> String {
         self.tabs[i]
@@ -248,7 +283,7 @@ impl OpenCADStudio {
             selected_after,
             dirty_before: pending.dirty_before,
             dirty_after,
-            structure: structure_changed.then_some(pending.structure_before),
+            structure: structure_changed.then_some(StructureSnapshot::Full(pending.structure_before)),
             label: pending.label,
         };
         self.push_undo_entry(i, HistorySnapshot::Delta(delta));
@@ -304,6 +339,263 @@ impl OpenCADStudio {
         })
     }
 
+    pub(super) fn begin_layer_undo(
+        &mut self,
+        i: usize,
+        label: impl Into<String>,
+        names: &[String],
+    ) -> PendingLayerDelta {
+        self.finish_pending_history(i);
+        PendingLayerDelta {
+            label: label.into(),
+            current_layout: self.tabs[i].scene.current_layout.clone(),
+            selected_before: self.tabs[i].scene.selected.iter().copied().collect(),
+            dirty_before: self.tabs[i].dirty,
+            before: names
+                .iter()
+                .map(|name| {
+                    (
+                        name.clone(),
+                        self.tabs[i].scene.document.layers.get(name).cloned(),
+                    )
+                })
+                .collect(),
+        }
+    }
+
+    pub(super) fn commit_layer_undo(&mut self, i: usize, pending: PendingLayerDelta) {
+        let entries: Vec<_> = pending
+            .before
+            .into_iter()
+            .filter_map(|(name, before)| {
+                let after = self.tabs[i].scene.document.layers.get(&name).cloned();
+                (before != after).then_some(TableEntryDelta {
+                    name,
+                    before,
+                    after,
+                })
+            })
+            .collect();
+        if entries.is_empty() {
+            return;
+        }
+        let selected_after = self.tabs[i].scene.selected.iter().copied().collect();
+        let delta = DeltaSnapshot {
+            entities: Vec::new(),
+            current_layout_before: pending.current_layout,
+            current_layout_after: self.tabs[i].scene.current_layout.clone(),
+            selected_before: pending.selected_before,
+            selected_after,
+            dirty_before: pending.dirty_before,
+            dirty_after: self.tabs[i].dirty,
+            structure: Some(StructureSnapshot::Layers(entries)),
+            label: pending.label,
+        };
+        self.push_undo_entry(i, HistorySnapshot::Delta(delta));
+    }
+
+    pub(super) fn begin_text_style_undo(
+        &mut self,
+        i: usize,
+        label: impl Into<String>,
+        names: &[String],
+    ) -> PendingTextStyleDelta {
+        self.finish_pending_history(i);
+        PendingTextStyleDelta {
+            label: label.into(),
+            current_layout: self.tabs[i].scene.current_layout.clone(),
+            selected_before: self.tabs[i].scene.selected.iter().copied().collect(),
+            dirty_before: self.tabs[i].dirty,
+            before: names
+                .iter()
+                .map(|name| {
+                    (
+                        name.clone(),
+                        self.tabs[i].scene.document.text_styles.get(name).cloned(),
+                    )
+                })
+                .collect(),
+        }
+    }
+
+    pub(super) fn commit_text_style_undo(&mut self, i: usize, pending: PendingTextStyleDelta) {
+        let entries: Vec<_> = pending
+            .before
+            .into_iter()
+            .filter_map(|(name, before)| {
+                let after = self.tabs[i].scene.document.text_styles.get(&name).cloned();
+                (before != after).then_some(TableEntryDelta {
+                    name,
+                    before,
+                    after,
+                })
+            })
+            .collect();
+        if entries.is_empty() {
+            return;
+        }
+        let delta = DeltaSnapshot {
+            entities: Vec::new(),
+            current_layout_before: pending.current_layout,
+            current_layout_after: self.tabs[i].scene.current_layout.clone(),
+            selected_before: pending.selected_before,
+            selected_after: self.tabs[i].scene.selected.iter().copied().collect(),
+            dirty_before: pending.dirty_before,
+            dirty_after: self.tabs[i].dirty,
+            structure: Some(StructureSnapshot::TextStyles(entries)),
+            label: pending.label,
+        };
+        self.push_undo_entry(i, HistorySnapshot::Delta(delta));
+    }
+
+    pub(super) fn begin_dim_style_undo(
+        &mut self,
+        i: usize,
+        label: impl Into<String>,
+        names: &[String],
+    ) -> PendingDimStyleDelta {
+        self.finish_pending_history(i);
+        PendingDimStyleDelta {
+            label: label.into(),
+            current_layout: self.tabs[i].scene.current_layout.clone(),
+            selected_before: self.tabs[i].scene.selected.iter().copied().collect(),
+            dirty_before: self.tabs[i].dirty,
+            before: names
+                .iter()
+                .map(|name| {
+                    (
+                        name.clone(),
+                        self.tabs[i].scene.document.dim_styles.get(name).cloned(),
+                    )
+                })
+                .collect(),
+        }
+    }
+
+    pub(super) fn commit_dim_style_undo(&mut self, i: usize, pending: PendingDimStyleDelta) {
+        let entries: Vec<_> = pending
+            .before
+            .into_iter()
+            .filter_map(|(name, before)| {
+                let after = self.tabs[i].scene.document.dim_styles.get(&name).cloned();
+                (before != after).then_some(TableEntryDelta {
+                    name,
+                    before,
+                    after,
+                })
+            })
+            .collect();
+        if entries.is_empty() {
+            return;
+        }
+        let delta = DeltaSnapshot {
+            entities: Vec::new(),
+            current_layout_before: pending.current_layout,
+            current_layout_after: self.tabs[i].scene.current_layout.clone(),
+            selected_before: pending.selected_before,
+            selected_after: self.tabs[i].scene.selected.iter().copied().collect(),
+            dirty_before: pending.dirty_before,
+            dirty_after: self.tabs[i].dirty,
+            structure: Some(StructureSnapshot::DimStyles(entries)),
+            label: pending.label,
+        };
+        self.push_undo_entry(i, HistorySnapshot::Delta(delta));
+    }
+
+    fn group_object_state(&self, i: usize) -> FxHashMap<Handle, acadrust::objects::ObjectType> {
+        use acadrust::objects::ObjectType;
+        let document = &self.tabs[i].scene.document;
+        let dictionary = document.header.acad_group_dict_handle;
+        document
+            .objects
+            .iter()
+            .filter(|(handle, object)| {
+                **handle == dictionary || matches!(object, ObjectType::Group(_))
+            })
+            .map(|(handle, object)| (*handle, object.clone()))
+            .collect()
+    }
+
+    pub(super) fn begin_group_undo(
+        &mut self,
+        i: usize,
+        label: impl Into<String>,
+    ) -> PendingObjectDelta {
+        self.finish_pending_history(i);
+        PendingObjectDelta {
+            label: label.into(),
+            current_layout: self.tabs[i].scene.current_layout.clone(),
+            selected_before: self.tabs[i].scene.selected.iter().copied().collect(),
+            dirty_before: self.tabs[i].dirty,
+            before: self.group_object_state(i),
+        }
+    }
+
+    pub(super) fn commit_group_undo(&mut self, i: usize, pending: PendingObjectDelta) {
+        let after = self.group_object_state(i);
+        let mut handles: HashSet<Handle> = pending.before.keys().copied().collect();
+        handles.extend(after.keys().copied());
+        let entries: Vec<_> = handles
+            .into_iter()
+            .filter_map(|handle| {
+                let before = pending.before.get(&handle).cloned();
+                let after = after.get(&handle).cloned();
+                (before != after).then_some(ObjectEntryDelta {
+                    handle,
+                    before,
+                    after,
+                })
+            })
+            .collect();
+        if entries.is_empty() {
+            return;
+        }
+        let delta = DeltaSnapshot {
+            entities: Vec::new(),
+            current_layout_before: pending.current_layout,
+            current_layout_after: self.tabs[i].scene.current_layout.clone(),
+            selected_before: pending.selected_before,
+            selected_after: self.tabs[i].scene.selected.iter().copied().collect(),
+            dirty_before: pending.dirty_before,
+            dirty_after: self.tabs[i].dirty,
+            structure: Some(StructureSnapshot::Objects(entries)),
+            label: pending.label,
+        };
+        self.push_undo_entry(i, HistorySnapshot::Delta(delta));
+    }
+
+    pub(super) fn commit_style_undo(
+        &mut self,
+        i: usize,
+        before: super::style_ops::StyleStateSnapshot,
+        after: super::style_ops::StyleStateSnapshot,
+        dirty_before: bool,
+    ) {
+        if before == after {
+            return;
+        }
+        self.finish_pending_history(i);
+        let (text_names, dim_names, object_handles) = after.changed_keys(&before);
+        let delta = DeltaSnapshot {
+            entities: Vec::new(),
+            current_layout_before: self.tabs[i].scene.current_layout.clone(),
+            current_layout_after: self.tabs[i].scene.current_layout.clone(),
+            selected_before: self.tabs[i].scene.selected.iter().copied().collect(),
+            selected_after: self.tabs[i].scene.selected.iter().copied().collect(),
+            dirty_before,
+            dirty_after: true,
+            structure: Some(StructureSnapshot::Styles {
+                before,
+                after,
+                text_names,
+                dim_names,
+                object_handles,
+            }),
+            label: "STYLE".to_string(),
+        };
+        self.push_undo_entry(i, HistorySnapshot::Delta(delta));
+    }
+
     /// Copy is delta-safe only when no target is a Dimension and no complete
     /// group is copied: dimensions clone fresh anonymous `*D` block records,
     /// and complete group copies add Group objects / dictionary entries. Both
@@ -397,11 +689,12 @@ impl OpenCADStudio {
             .collect();
         let selected_after = self.tabs[i].scene.selected.iter().copied().collect();
         let dirty_after = self.tabs[i].dirty;
-        let mut structure = pending.structure_before;
+        let mut structure = pending.structure_before.map(StructureSnapshot::Full);
         if let Some(before_structure) = structure.as_mut() {
-            let after_structure = self.tabs[i].scene.document.snapshot_structure();
-            let added_handles: Vec<Handle> = entities
-                .iter()
+            if let StructureSnapshot::Full(before_structure) = before_structure {
+                let after_structure = self.tabs[i].scene.document.snapshot_structure();
+                let added_handles: Vec<Handle> = entities
+                    .iter()
                 .filter_map(|(handle, before, after)| {
                     (before.is_none() && after.is_some()).then_some(*handle)
                 })
@@ -414,6 +707,7 @@ impl OpenCADStudio {
             if *before_structure == after_structure {
                 structure = None;
             }
+            }
         }
         let delta = DeltaSnapshot {
             entities,
@@ -443,9 +737,113 @@ impl OpenCADStudio {
         // Install the chosen side of every entity image. Derived-cache, geometry
         // and UI invalidation are deferred until every requested undo/redo step
         // has been applied.
-        if let Some(structure) = d.structure.take() {
-            let inverse = self.tabs[i].scene.document.swap_structure(structure);
-            d.structure = Some(inverse);
+        if let Some(structure) = d.structure.as_mut() {
+            match structure {
+                StructureSnapshot::Full(stored) => {
+                    let inverse = self.tabs[i].scene.document.swap_structure(
+                        std::mem::replace(stored, acadrust::CadDocument::new()),
+                    );
+                    *stored = inverse;
+                    self.tabs[i].scene.invalidate_dependency_index();
+                }
+                StructureSnapshot::Layers(entries) => {
+                    let names: Vec<String> =
+                        entries.iter().map(|entry| entry.name.clone()).collect();
+                    for entry in entries {
+                        let value = if undo {
+                            entry.before.clone()
+                        } else {
+                            entry.after.clone()
+                        };
+                        if let Some(layer) = value {
+                            self.tabs[i].scene.document.layers.add_or_replace(layer);
+                        } else {
+                            self.tabs[i].scene.document.layers.remove(&entry.name);
+                        }
+                    }
+                    self.tabs[i].scene.invalidate_layer_dependencies(&names);
+                }
+                StructureSnapshot::TextStyles(entries) => {
+                    let names: Vec<String> =
+                        entries.iter().map(|entry| entry.name.clone()).collect();
+                    for entry in entries {
+                        let value = if undo {
+                            entry.before.clone()
+                        } else {
+                            entry.after.clone()
+                        };
+                        if let Some(style) = value {
+                            self.tabs[i]
+                                .scene
+                                .document
+                                .text_styles
+                                .add_or_replace(style);
+                        } else {
+                            self.tabs[i].scene.document.text_styles.remove(&entry.name);
+                        }
+                    }
+                    for name in names {
+                        self.tabs[i].scene.invalidate_text_style_dependencies(&name);
+                    }
+                }
+                StructureSnapshot::DimStyles(entries) => {
+                    let names: Vec<String> =
+                        entries.iter().map(|entry| entry.name.clone()).collect();
+                    for entry in entries {
+                        let value = if undo {
+                            entry.before.clone()
+                        } else {
+                            entry.after.clone()
+                        };
+                        if let Some(style) = value {
+                            self.tabs[i].scene.document.dim_styles.add_or_replace(style);
+                        } else {
+                            self.tabs[i].scene.document.dim_styles.remove(&entry.name);
+                        }
+                    }
+                    for name in names {
+                        self.tabs[i].scene.invalidate_dim_style_dependencies(&name);
+                    }
+                }
+                StructureSnapshot::Objects(entries) => {
+                    for entry in entries {
+                        let value = if undo {
+                            entry.before.clone()
+                        } else {
+                            entry.after.clone()
+                        };
+                        if let Some(object) = value {
+                            self.tabs[i]
+                                .scene
+                                .document
+                                .objects
+                                .insert(entry.handle, object);
+                        } else {
+                            self.tabs[i].scene.document.objects.remove(&entry.handle);
+                        }
+                    }
+                    self.tabs[i].scene.invalidate_dependency_index();
+                }
+                StructureSnapshot::Styles {
+                    before,
+                    after,
+                    text_names,
+                    dim_names,
+                    object_handles,
+                } => {
+                    let snapshot = if undo { before } else { after };
+                    self.restore_style_state(i, snapshot);
+                    self.tabs[i]
+                        .scene
+                        .invalidate_text_style_dependencies_many(text_names);
+                    self.tabs[i]
+                        .scene
+                        .invalidate_dim_style_dependencies_many(dim_names);
+                    self.tabs[i]
+                        .scene
+                        .invalidate_object_style_dependencies(object_handles);
+                }
+            }
         }
         let changes = self.tabs[i].scene.apply_entity_delta(&d.entities, undo);
         let scene = &mut self.tabs[i].scene;
@@ -479,6 +877,7 @@ impl OpenCADStudio {
         &mut self,
         i: usize,
         had_full: bool,
+        structure_changed: bool,
         changes: &[(Handle, crate::scene::ChangeKind)],
     ) {
         self.tabs[i].edit_revision = self.tabs[i].edit_revision.wrapping_add(1);
@@ -524,7 +923,7 @@ impl OpenCADStudio {
         self.tabs[i].active_cmd = None;
         self.tabs[i].snap_result = None;
         self.tabs[i].active_grip = None;
-        if had_full {
+        if structure_changed {
             let doc_layers = self.tabs[i].scene.document.layers.clone();
             let vp_info = self.tabs[i].scene.viewport_list();
             self.tabs[i]
@@ -555,6 +954,7 @@ impl OpenCADStudio {
 
         let mut last_label = String::new();
         let mut had_full = false;
+        let mut structure_changed = false;
         let mut changes = Vec::new();
         for _ in 0..steps {
             let Some(snapshot) = self.tabs[i].history.undo_stack.pop() else {
@@ -566,7 +966,8 @@ impl OpenCADStudio {
                     // Symmetric: undo applies the before side, then the same
                     // delta rides to the redo stack (it still holds the after
                     // side) — no current-state capture needed.
-                    had_full |= d.structure.is_some();
+                    structure_changed |= d.structure.is_some();
+                    had_full |= d.structure.as_ref().is_some_and(StructureSnapshot::is_full);
                     changes.extend(self.apply_delta_state(i, &mut d, true));
                     self.tabs[i]
                         .history
@@ -575,7 +976,7 @@ impl OpenCADStudio {
                 }
             }
         }
-        self.finish_history_apply(i, had_full, &changes);
+        self.finish_history_apply(i, had_full, structure_changed, &changes);
         self.command_line
             .push_output(&format!("Undo: {last_label}"));
     }
@@ -592,6 +993,7 @@ impl OpenCADStudio {
 
         let mut last_label = String::new();
         let mut had_full = false;
+        let mut structure_changed = false;
         let mut changes = Vec::new();
         for _ in 0..steps {
             let Some(snapshot) = self.tabs[i].history.redo_stack.pop() else {
@@ -600,7 +1002,8 @@ impl OpenCADStudio {
             last_label = snapshot.label().to_string();
             match snapshot {
                 HistorySnapshot::Delta(mut d) => {
-                    had_full |= d.structure.is_some();
+                    structure_changed |= d.structure.is_some();
+                    had_full |= d.structure.as_ref().is_some_and(StructureSnapshot::is_full);
                     changes.extend(self.apply_delta_state(i, &mut d, false));
                     self.tabs[i]
                         .history
@@ -609,7 +1012,7 @@ impl OpenCADStudio {
                 }
             }
         }
-        self.finish_history_apply(i, had_full, &changes);
+        self.finish_history_apply(i, had_full, structure_changed, &changes);
         self.command_line
             .push_output(&format!("Redo: {last_label}"));
     }

src/app/mod.rs

@@ -168,7 +168,6 @@ use acadrust::CadDocument;
 use iced::time::Instant;
 use iced::window;
 use iced::{mouse, Point, Task, Theme};
-use std::sync::atomic::AtomicU8;
 use std::sync::Arc;
 
 pub(super) const POLY_START_DELAY_MS: u128 = 150;
@@ -179,14 +178,15 @@ pub(super) const VARIES_LABEL: &str = "*VARIES*";
 // loader thread, read by the UI overlay on every frame.
 pub const OPEN_PHASE_READING: u8 = 0;
 pub const OPEN_PHASE_PARSING: u8 = 1;
-pub const OPEN_PHASE_CACHING: u8 = 2;
-pub const OPEN_PHASE_FINALIZING: u8 = 3;
+pub const OPEN_PHASE_XREF: u8 = 2;
+pub const OPEN_PHASE_CACHING: u8 = 3;
+pub const OPEN_PHASE_FINALIZING: u8 = 4;
 
 #[derive(Debug, Clone)]
 pub struct OpenProgress {
     pub name: String,
     pub size_bytes: u64,
-    pub phase: Arc<AtomicU8>,
+    pub state: Arc<crate::io::OpenProgressState>,
     pub started: Instant,
 }
 
@@ -2098,6 +2098,9 @@ pub enum Message {
     PlotWindowExport,
     /// Callback after the user picks (or cancels) the window-export path.
     PlotWindowExportPath(Option<std::path::PathBuf>),
+    /// Completion of a PDF/preview/print job performed outside the UI thread.
+    /// The boolean restores the Plot dialog after a preview.
+    BackgroundIoFinished(Result<String, String>, bool),
     /// Send current layout to the system printer (via lp / lpr).
     PrintToPrinter,
     /// Callback from the async printer job.
@@ -2299,6 +2302,8 @@ pub enum Message {
     WblockSave(String),
     /// Result of the WBLOCK save path dialog.
     WblockSaveResult(String, Option<std::path::PathBuf>),
+    /// Background extraction/write completion.
+    WblockWriteFinished(String, std::path::PathBuf, Result<(), String>),
     // ── DATAEXTRACTION ────────────────────────────────────────────────────
     /// Save the pre-built CSV string to a file chosen by the user.
     DataExtractionSave(String),
@@ -2309,16 +2314,23 @@ pub enum Message {
     StlExport,
     /// Callback after the user picks (or cancels) the STL save path.
     StlExportPath(Option<std::path::PathBuf>),
+    StlExportFinished(std::path::PathBuf, Result<(), String>),
     // ── STEP export ───────────────────────────────────────────────────────
     /// Trigger STEP AP203 export: show save dialog.
     StepExport,
     /// Callback after the user picks (or cancels) the STEP save path.
     StepExportPath(Option<std::path::PathBuf>),
+    StepExportFinished(std::path::PathBuf, Result<(), String>),
     // ── OBJ import ────────────────────────────────────────────────────────
     /// Trigger OBJ import: show open-file dialog.
     ObjImport,
     /// Callback after the user picks (or cancels) the OBJ file path.
     ObjImportPath(Option<std::path::PathBuf>),
+    ObjImportFinished(
+        u64,
+        std::path::PathBuf,
+        Result<crate::scene::model::mesh_model::MeshModel, String>,
+    ),
 }
 
 impl OpenCADStudio {

src/app/style_ops.rs

@@ -539,8 +539,7 @@ impl OpenCADStudio {
 
     /// Overwrite the live style state with a snapshot (used by commit's undo
     /// dance and by discard).
-    fn restore_style_state(&mut self, snap: &StyleStateSnapshot) {
-        let i = self.active_tab;
+    pub(super) fn restore_style_state(&mut self, i: usize, snap: &StyleStateSnapshot) {
         let doc = &mut self.tabs[i].scene.document;
         doc.text_styles = snap.text_styles.clone();
         doc.dim_styles = snap.dim_styles.clone();
@@ -560,8 +559,10 @@ impl OpenCADStudio {
         doc.header.multiline_style = snap.multiline_style.clone();
         doc.header.current_table_style_name = snap.current_table.clone();
         doc.header.current_mleader_style_name = snap.current_mleader.clone();
-        self.ribbon.active_table_style = snap.active_table.clone();
-        self.ribbon.active_mleader_style = snap.active_mleader.clone();
+        if i == self.active_tab {
+            self.ribbon.active_table_style = snap.active_table.clone();
+            self.ribbon.active_mleader_style = snap.active_mleader.clone();
+        }
         self.tabs[i].active_mleader_style = snap.tab_active_mleader.clone();
     }
 
@@ -587,18 +588,26 @@ impl OpenCADStudio {
             self.sync_ribbon_styles();
             return;
         };
-        // Capture the edited state, rewind to the baseline so the undo entry
-        // restores the pre-edit document, then re-apply the edits on top.
         let edited = self.capture_style_state();
-        self.restore_style_state(&stage.baseline);
-        self.push_undo_snapshot(i, "STYLE");
-        self.restore_style_state(&edited);
-        self.tabs[i].dirty = true;
-        self.tabs[i].scene.bump_geometry();
+        let changed = edited != stage.baseline;
+        if changed {
+            self.tabs[i].dirty = true;
+            let (text_names, dim_names, object_handles) = edited.changed_keys(&stage.baseline);
+            self.tabs[i].scene.invalidate_text_style_dependencies_many(&text_names);
+            self.tabs[i]
+                .scene
+                .invalidate_dim_style_dependencies_many(&dim_names);
+            self.tabs[i]
+                .scene
+                .invalidate_object_style_dependencies(&object_handles);
+            self.commit_style_undo(i, stage.baseline, edited.clone(), stage.dirty_at_open);
+        } else {
+            self.tabs[i].dirty = stage.dirty_at_open;
+        }
         self.sync_ribbon_styles();
         // Re-baseline so further edits in the still-open window stage afresh.
         self.style_stage = Some(StyleStage {
-            dirty_at_open: true,
+            dirty_at_open: self.tabs[i].dirty,
             baseline: edited,
         });
     }
@@ -609,13 +618,14 @@ impl OpenCADStudio {
         let Some(stage) = self.style_stage.take() else {
             return;
         };
-        self.restore_style_state(&stage.baseline);
+        self.restore_style_state(self.active_tab, &stage.baseline);
         self.tabs[self.active_tab].dirty = stage.dirty_at_open;
         self.sync_ribbon_styles();
     }
 }
 
 /// Snapshot of every document field a style manager can touch.
+#[derive(Clone, PartialEq)]
 pub(super) struct StyleStateSnapshot {
     text_styles: acadrust::tables::Table<TextStyle>,
     dim_styles: acadrust::tables::Table<DimStyle>,
@@ -630,6 +640,64 @@ pub(super) struct StyleStateSnapshot {
     tab_active_mleader: String,
 }
 
+impl StyleStateSnapshot {
+    pub(super) fn estimated_bytes(&self) -> usize {
+        self.text_styles
+            .iter()
+            .count()
+            .saturating_mul(320)
+            .saturating_add(self.dim_styles.iter().count().saturating_mul(1024))
+            .saturating_add(self.style_objects.len().saturating_mul(512))
+    }
+
+    pub(super) fn changed_keys(&self, other: &Self) -> (Vec<String>, Vec<String>, Vec<Handle>) {
+        let text_names = self
+            .text_styles
+            .iter()
+            .filter(|style| other.text_styles.get(&style.name) != Some(*style))
+            .map(|style| style.name.clone())
+            .chain(
+                other
+                    .text_styles
+                    .iter()
+                    .filter(|style| self.text_styles.get(&style.name) != Some(*style))
+                    .map(|style| style.name.clone()),
+            )
+            .collect::<rustc_hash::FxHashSet<_>>()
+            .into_iter()
+            .collect();
+        let dim_names = self
+            .dim_styles
+            .iter()
+            .filter(|style| other.dim_styles.get(&style.name) != Some(*style))
+            .map(|style| style.name.clone())
+            .chain(
+                other
+                    .dim_styles
+                    .iter()
+                    .filter(|style| self.dim_styles.get(&style.name) != Some(*style))
+                    .map(|style| style.name.clone()),
+            )
+            .collect::<rustc_hash::FxHashSet<_>>()
+            .into_iter()
+            .collect();
+        let self_objects: rustc_hash::FxHashMap<_, _> =
+            self.style_objects.iter().map(|(h, o)| (*h, o)).collect();
+        let other_objects: rustc_hash::FxHashMap<_, _> =
+            other.style_objects.iter().map(|(h, o)| (*h, o)).collect();
+        let object_handles = self
+            .style_objects
+            .iter()
+            .chain(other.style_objects.iter())
+            .map(|(handle, _)| *handle)
+            .collect::<rustc_hash::FxHashSet<_>>()
+            .into_iter()
+            .filter(|handle| self_objects.get(handle) != other_objects.get(handle))
+            .collect();
+        (text_names, dim_names, object_handles)
+    }
+}
+
 /// An in-progress style-manager transaction.
 pub(super) struct StyleStage {
     dirty_at_open: bool,

src/app/update/file.rs

@@ -20,6 +20,28 @@ use acadrust::{EntityType as AcadEntityType, Handle};
 use iced::time::Instant;
 use iced::{mouse, Point, Task};
 
+pub(super) fn background_task<T, F, M>(work: F, map: M) -> Task<Message>
+where
+    T: Send + 'static,
+    F: FnOnce() -> T + Send + 'static,
+    M: FnOnce(T) -> Message + Send + 'static,
+{
+    #[cfg(not(target_arch = "wasm32"))]
+    {
+        let (tx, rx) = iced::futures::channel::oneshot::channel();
+        std::thread::spawn(move || {
+            let _ = tx.send(work());
+        });
+        Task::perform(
+            async move { rx.await.expect("background export worker dropped") },
+            map,
+        )
+    }
+    #[cfg(target_arch = "wasm32")]
+    {
+        Task::perform(async move { work() }, map)
+    }
+}
 
 impl OpenCADStudio {
     /// Before a save, give every cached truck solid that still has no ACIS
@@ -362,15 +384,16 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
                     // progress, so mark one. The browser picker + parse happen
                     // inside `pick_and_load_web`; the real name is unknown until
                     // then, so show a generic label meanwhile.
+            let state = std::sync::Arc::new(crate::io::OpenProgressState::new(
+                crate::app::OPEN_PHASE_READING,
+            ));
                     self.opening = Some(crate::app::OpenProgress {
                         name: "Opening…".into(),
                         size_bytes: 0,
-                        phase: std::sync::Arc::new(std::sync::atomic::AtomicU8::new(
-                            crate::app::OPEN_PHASE_READING,
-                        )),
+                state: state.clone(),
                         started: Instant::now(),
                     });
-                    Task::perform(crate::io::pick_and_load_web(), Message::FileOpened)
+                    Task::perform(crate::io::pick_and_load_web(state), Message::FileOpened)
                 }
     }
 
@@ -403,13 +426,15 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
         }
     }
 
-    pub(super) fn on_file_opened(&mut self, name: String, path: std::path::PathBuf, doc: acadrust::CadDocument, caches: crate::scene::DerivedCaches) -> Task<Message> {
+    pub(super) fn on_file_opened(&mut self, name: String, path: std::path::PathBuf, doc: acadrust::CadDocument,
+        mut caches: crate::scene::DerivedCaches,
+    ) -> Task<Message> {
                 // If the user clicked Cancel while the parser was running, the
                 // overlay state was cleared and we silently drop the result.
                 if self.opening.is_none() {
                     return Task::none();
                 }
-                let open_started = self.opening.take().map(|p| p.started);
+                let open_started = self.opening.as_ref().map(|p| p.started);
                 let timings = caches.timings;
                 let entity_count = doc.entities().count();
                 self.command_line
@@ -420,6 +445,30 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
                         caches.corrupt_dropped
                     ));
                 }
+        if caches.xref_dropped > 0 {
+            self.command_line.push_error(&format!(
+                "Warning: {} corrupt xref entities dropped",
+                caches.xref_dropped
+            ));
+        }
+        for info in &caches.xrefs {
+            match info.status {
+                crate::io::xref::XrefStatus::Loaded => {
+                    self.command_line
+                        .push_output(&format!("XREF  Loaded \"{}\"", info.name));
+                }
+                crate::io::xref::XrefStatus::NotFound => {
+                    self.command_line.push_error(&format!(
+                        "XREF  Not found: \"{}\" ({})",
+                        info.name, info.path
+                    ));
+                }
+                crate::io::xref::XrefStatus::Unloaded => {
+                    self.command_line
+                        .push_info(&format!("XREF  Unloaded (skipped): \"{}\"", info.name));
+                }
+            }
+        }
                 let thumbs_task = self.push_recent(path.clone());
 
                 let current_is_empty = {
@@ -472,47 +521,6 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
                     1.0
                 };
 
-                // Auto-resolve XREFs relative to the opened file's directory.
-                let mut xref_ms = 0u32;
-                let mut xref_merged = false;
-                if let Some(base_dir) = path.parent() {
-                    // xref content arrives un-purged: parser-garbage entities
-                    // inside the referenced file can trigger infinite loops in
-                    // tessellation. `resolve_xrefs` runs the corrupt-entity
-                    // guard inline as it merges each xref, so no second
-                    // full-document walk is needed here.
-                    let t_xref = Instant::now();
-                    let (xrefs, extra_dropped) =
-                        crate::io::xref::resolve_xrefs(&mut self.tabs[i].scene.document, base_dir);
-                    xref_ms = t_xref.elapsed().as_millis() as u32;
-                    if extra_dropped > 0 {
-                        self.command_line.push_error(&format!(
-                            "Warning: {extra_dropped} corrupt xref entities dropped"
-                        ));
-                    }
-                    for info in &xrefs {
-                        match info.status {
-                            crate::io::xref::XrefStatus::Loaded => {
-                                xref_merged = true;
-                                self.command_line
-                                    .push_output(&format!("XREF  Loaded \"{}\"", info.name));
-                            }
-                            crate::io::xref::XrefStatus::NotFound => {
-                                self.command_line.push_error(&format!(
-                                    "XREF  Not found: \"{}\" ({})",
-                                    info.name, info.path
-                                ));
-                            }
-                            crate::io::xref::XrefStatus::Unloaded => {
-                                self.command_line.push_info(&format!(
-                                    "XREF  Unloaded (skipped): \"{}\"",
-                                    info.name
-                                ));
-                            }
-                        }
-                    }
-                }
-
                 // Open-time breakdown so regressions are visible immediately.
                 // `total` is wall time from the Open click to here (post-xref,
                 // pre-first-frame); the phase figures are the background-thread
@@ -522,7 +530,7 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
                     .unwrap_or(0);
                 self.command_line.push_info(&format!(
                     "  parse {}ms · purge {}ms · caches {}ms · xref {}ms · total {}ms",
-                    timings.parse_ms, timings.purge_ms, timings.caches_ms, xref_ms, total_ms
+                    timings.parse_ms, timings.purge_ms, timings.caches_ms, timings.xref_ms, total_ms
                 ));
 
                 // Caches were built on the background thread inside open_path().
@@ -532,18 +540,11 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
                 self.tabs[i].scene.images = caches.images;
                 self.tabs[i].scene.meshes = caches.meshes;
                 self.tabs[i].scene.block_meshes = caches.block_meshes;
+        let prepared_geometry = caches.prepared_geometry.take();
                 // Invalidate the wire cache so the new document is tessellated.
                 self.tabs[i].scene.bump_geometry();
-                // XREFs are merged into the document AFTER the background worker
-                // built the mesh caches above, so those caches contain none of
-                // the xref'd geometry. The wire pass rebuilds from the document
-                // each frame (bump_geometry covers it), but 3D-solid meshes are
-                // only tessellated by populate — run the incremental variant so
-                // the already-cached host solids are kept and only the newly
-                // merged xref solids (walls, floors, roofs) are tessellated,
-                // avoiding a full re-tessellation of the whole drawing. (#203)
-                if xref_merged {
-                    self.tabs[i].scene.populate_missing_meshes_from_document();
+                if let Some(prepared) = prepared_geometry {
+                    self.tabs[i].scene.install_prepared_open_geometry(prepared);
                 }
                 self.tabs[i].scene.selected = rustc_hash::FxHashSet::default();
                 self.tabs[i].scene.preview_wires = vec![];
@@ -625,42 +626,38 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
                 };
                 #[cfg(target_arch = "wasm32")]
                 let interaction_task = Task::none();
+        if let Some(opening) = &self.opening {
+            opening
+                .state
+                .set(crate::app::OPEN_PHASE_FINALIZING, 10000, 1, 1);
+        }
+        self.opening.take();
                 let pending_open_task = self.drain_pending_open();
                 Task::batch([thumbs_task, pending_open_task, interaction_task])
     }
 
-    pub(super) fn on_wblock_save_result_some(&mut self, block_name: String, path: std::path::PathBuf) -> Task<Message> {
+    pub(super) fn on_wblock_save_result_some(&mut self, block_name: String, path: std::path::PathBuf,
+    ) -> Task<Message> {
                 let i = self.active_tab;
-                let result = if block_name == "*" {
+                let document = self.tabs[i].scene.document.clone();
                     let handles: Vec<_> = self.tabs[i].scene.selected.iter().copied().collect();
+        let worker_name = block_name.clone();
+        let worker_path = path.clone();
+        background_task(
+            move || {
+                let document = if worker_name == "*" {
                     crate::modules::insert::wblock::extract_entities_to_doc(
-                        &self.tabs[i].scene.document,
+                        &document,
                         &handles,
                     )
                 } else {
                     crate::modules::insert::wblock::extract_block_to_doc(
-                        &self.tabs[i].scene.document,
-                        &block_name,
-                    )
-                };
-                match result {
-                    Ok(doc) => match crate::io::save(&doc, &path) {
-                        Ok(()) => {
-                            let fname = path
-                                .file_name()
-                                .map(|n| n.to_string_lossy().into_owned())
-                                .unwrap_or_else(|| path.to_string_lossy().into_owned());
-                            self.command_line.push_output(&format!(
-                                "WBLOCK  Saved \"{block_name}\" → \"{fname}\""
-                            ));
-                        }
-                        Err(e) => self
-                            .command_line
-                            .push_error(&format!("WBLOCK save failed: {e}")),
-                    },
-                    Err(e) => self.command_line.push_error(&format!("WBLOCK: {e}")),
+                        &document,
+                        &worker_name)
                 }
-                Task::none()
+                .map_err(|e| e.to_string())?; crate::io::save(&document, &worker_path).map_err(|e| e.to_string())
+                    },
+            move |result| Message::WblockWriteFinished(block_name, path, result),)
     }
 
     pub(super) fn on_stl_export_path_some(&mut self, path: std::path::PathBuf) -> Task<Message> {
@@ -675,21 +672,15 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
                     .values()
                     .filter_map(|s| s.lods.first().cloned())
                     .collect();
-                let mesh_refs: Vec<&crate::scene::model::mesh_model::MeshModel> = meshes.iter().collect();
-                match crate::io::stl::build_stl(&mesh_refs) {
-                    Some(bytes) => match std::fs::write(&path, bytes) {
-                        Ok(()) => self
-                            .command_line
-                            .push_output(&format!("STLOUT: exported to \"{}\"", path.display())),
-                        Err(e) => self
-                            .command_line
-                            .push_error(&format!("STLOUT: write error: {e}")),
+                let worker_path = path.clone();
+        background_task(
+            move || {
+                let mesh_refs: Vec<_> = meshes.iter().collect();
+                let bytes = crate::io::stl::build_stl(&mesh_refs)
+                    .ok_or_else(|| "no mesh data to export".to_string())?;
+                std::fs::write(&worker_path, bytes).map_err(|e| e.to_string())
                     },
-                    None => self
-                        .command_line
-                        .push_error("STLOUT: no mesh data to export."),
-                }
-                Task::none()
+            move |result| Message::StlExportFinished(path, result),)
     }
 
     pub(super) fn on_step_export_path_some(&mut self, path: std::path::PathBuf) -> Task<Message> {
@@ -701,63 +692,26 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
                     .values()
                     .filter_map(|s| s.lods.first().cloned())
                     .collect();
-                let mesh_refs: Vec<&crate::scene::model::mesh_model::MeshModel> = meshes.iter().collect();
-                match crate::io::step::build_step(&mesh_refs) {
-                    Some(text) => match std::fs::write(&path, text.as_bytes()) {
-                        Ok(()) => self
-                            .command_line
-                            .push_output(&format!("STEPOUT: exported to \"{}\"", path.display())),
-                        Err(e) => self
-                            .command_line
-                            .push_error(&format!("STEPOUT: write error: {e}")),
+                let worker_path = path.clone();
+        background_task(
+            move || {
+                let mesh_refs: Vec<_> = meshes.iter().collect();
+                let text = crate::io::step::build_step(&mesh_refs)
+                    .ok_or_else(|| "no mesh data to export".to_string())?;
+                std::fs::write(&worker_path, text.as_bytes()).map_err(|e| e.to_string())
                     },
-                    None => self
-                        .command_line
-                        .push_error("STEPOUT: no mesh data to export."),
-                }
-                Task::none()
+            move |result| Message::StepExportFinished(path, result),)
     }
 
     pub(super) fn on_obj_import_path_some(&mut self, path: std::path::PathBuf) -> Task<Message> {
-                let src = match std::fs::read_to_string(&path) {
-                    Ok(s) => s,
-                    Err(e) => {
-                        self.command_line
-                            .push_error(&format!("IMPORTOBJ: read error: {e}"));
-                        return Task::none();
-                    }
-                };
-                let color = [0.7f32, 0.7, 0.85, 1.0];
-                match crate::io::obj::parse_obj(&src, color) {
-                    None => {
-                        self.command_line
-                            .push_error("IMPORTOBJ: no usable geometry in file.");
-                    }
-                    Some(mut mesh) => {
-                        let i = self.active_tab;
-                        let file_stem = path
-                            .file_stem()
-                            .map(|s| s.to_string_lossy().into_owned())
-                            .unwrap_or_else(|| "obj_mesh".into());
-                        mesh.name = file_stem.clone();
-                        self.push_undo_snapshot(i, "IMPORTOBJ");
-                        use crate::modules::insert::solid3d_cmds::empty_solid3d;
-                        let entity = empty_solid3d();
-                        let handle = self.tabs[i].scene.add_entity(entity);
-                        if !handle.is_null() {
-                            self.tabs[i]
-                                .scene
-                                .meshes
-                                .insert(handle, crate::scene::MeshLodSet::from_single(mesh));
-                            self.tabs[i].dirty = true;
-                            self.command_line.push_output(&format!(
-                                "IMPORTOBJ: imported \"{}\" as mesh.",
-                                file_stem
-                            ));
-                        }
-                    }
-                }
-                Task::none()
+                let tab_id = self.tabs[self.active_tab].id;
+        let worker_path = path.clone();
+        background_task(
+            move || {
+                let src = std::fs::read_to_string(&worker_path).map_err(|e| e.to_string())?; crate::io::obj::parse_obj(&src, [0.7, 0.7, 0.85, 1.0])
+                            .ok_or_else(|| "no usable geometry in file".to_string())
+            },
+            move |result| Message::ObjImportFinished(tab_id, path, result),)
     }
 
     #[cfg(not(target_arch = "wasm32"))]
@@ -770,6 +724,19 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
         self.sync_truck_solids_to_acis(i);
     }
 
+    #[cfg(target_arch = "wasm32")]
+    fn stamp_thumbnail(&mut self, i: usize, version: acadrust::DxfVersion) {
+        let scene = &self.tabs[i].scene;
+        let preview = crate::io::thumbnail::from_snapshot(
+            &scene.entity_wires(),
+            &scene.camera.borrow(),
+            scene.bg_color,
+            version >= acadrust::DxfVersion::AC1027,
+            self.vp_size,
+        );
+        self.tabs[i].scene.document.preview = preview;
+    }
+
     #[cfg(not(target_arch = "wasm32"))]
     pub(in crate::app) fn queue_native_save(
         &mut self,
@@ -1484,7 +1451,11 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
                     _ => (paper_w, paper_h),
                 };
 
-                match crate::io::pdf_export::export_pdf(
+        let plot_style = self.active_plot_style.clone();
+        let worker_path = path.clone();
+        background_task(
+            move || {
+                crate::io::pdf_export::export_pdf(
                     &wires,
                     hatches.as_slice(),
                     wipeouts.as_slice(),
@@ -1495,23 +1466,20 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
                     rotation_deg,
                     1.0,
                     None,
-                    &path,
-                    self.active_plot_style.as_ref(),
-                ) {
-                    // Full path, not just the file name — when the export was
-                    // driven by EXPORTPDF <path> the user needs to see where
-                    // the file actually landed. (#369)
-                    Ok(()) => self
-                        .command_line
-                        .push_info(&format!("Exported: {}", path.display())),
-                    Err(e) => self.command_line.push_error(&format!("Export failed: {e}")),
-                }
-                Task::none()
+                    &worker_path,
+                    plot_style.as_ref(),
+                )
+                .map(|_| format!("Exported: {}", worker_path.display()))
+                .map_err(|e| format!("Export failed: {e}"))
+            },
+            |result| Message::BackgroundIoFinished(result, false),
+        )
     }
 
     /// Export the pending model-space plot window (set by PLOTWINDOW while on
     /// the Model tab) to PDF, using the chosen paper size/orientation/scale.
-    pub(super) fn on_plot_window_export_path_some(&mut self, path: std::path::PathBuf) -> Task<Message> {
+    pub(super) fn on_plot_window_export_path_some(&mut self, path: std::path::PathBuf,
+    ) -> Task<Message> {
                 use crate::io::paper_sizes::{sheet_mm, window_to_sheet, PlotScale};
                 let i = self.active_tab;
                 if self.tabs[i].scene.current_layout != "Model" {
@@ -1569,7 +1537,12 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
                 // same as the wires) so the final sheet-mm rect lands at (ox, oy).
                 let clip = Some(((ox / scale) as f32, (oy / scale) as f32, win_w as f32, win_h as f32));
 
-                let res = crate::io::pdf_export::export_pdf(
+                let plot_style = self.active_plot_style.clone();
+        let worker_path = path.clone();
+        self.close_active_modal();
+        background_task(
+            move || {
+                crate::io::pdf_export::export_pdf(
                     &wires,
                     hatches.as_slice(),
                     wipeouts.as_slice(),
@@ -1580,20 +1553,18 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
                     0,
                     scale as f32,
                     clip,
-                    &path,
-                    self.active_plot_style.as_ref(),
-                );
-                match res {
-                    Ok(()) => {
-                        self.command_line.push_info(&format!(
+                    &worker_path,
+                    plot_style.as_ref(),
+                )
+                .map(|_| {format!(
                             "Plotted window to {}",
-                            path.file_name().unwrap_or_default().to_string_lossy()
-                        ));
-                        self.close_active_modal();
-                    }
-                    Err(e) => self.command_line.push_error(&format!("Plot failed: {e}")),
-                }
-                Task::none()
+                        worker_path
+                            .file_name().unwrap_or_default().to_string_lossy()
+                        )
+                }).map_err(|e| format!("Plot failed: {e}"))
+            },
+            |result| Message::BackgroundIoFinished(result, false),
+        )
     }
 
     /// Build the render inputs and page geometry for a full-layout plot: wires
@@ -1678,12 +1649,12 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
             self.layout_plot_params();
         let plot_style = self.active_plot_style.clone();
         self.command_line.push_info("Sending to system printer…");
-        Task::perform(
-            async move {
+        background_task(
+            move || {
+                iced::futures::executor::block_on(
                 crate::io::print_to_printer::print_wires(
                     wires, hatches, wipeouts, eff_w, eff_h, ox, oy, rotation_deg, plot_style,
-                )
-                .await
+                ))
             },
             Message::PrintResult,
         )
@@ -2220,16 +2191,16 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
                     return Task::none();
                 };
                 let tmp = std::env::temp_dir().join("open_cad_studio_preview.pdf");
-                let exp = crate::io::pdf_export::export_pdf(
+                return background_task(
+                    move || {
+                        crate::io::pdf_export::export_pdf(
                     &w_wires, &w_hatches, &w_wipeouts, sw, sh, wox, woy, 0, wscale, wclip, &tmp,
                     plot_style.as_ref(),
+                ).and_then(|_| crate::io::print_to_printer::open_in_viewer(&tmp))
+                        .map(|_| "Opened plot preview.".to_string()).map_err(|e| format!("Preview failed: {e}"))
+                    },
+                    |result| Message::BackgroundIoFinished(result, true),
                 );
-                match exp.and_then(|_| crate::io::print_to_printer::open_in_viewer(&tmp)) {
-                    Ok(()) => self.command_line.push_info("Opened plot preview."),
-                    Err(e) => self.command_line.push_error(&format!("Preview failed: {e}")),
-                }
-                self.active_modal = Some(crate::app::ModalKind::Plot);
-                return Task::none();
             }
             if d.to_file {
                 // Tested clipped export (opens a save dialog).
@@ -2244,25 +2215,27 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
                 return Task::none();
             };
             let tmp = std::env::temp_dir().join("open_cad_studio_print.pdf");
-            let exp = crate::io::pdf_export::export_pdf(
+            let opts = self.plot_print_options(&d);
+            return background_task(
+                move || {
+                    crate::io::pdf_export::export_pdf(
                 &w_wires, &w_hatches, &w_wipeouts, sw, sh, wox, woy, 0, wscale, wclip, &tmp,
                 plot_style.as_ref(),
+            ).and_then(|_| crate::io::print_to_printer::print_existing_pdf(&tmp, &opts))
+                    .map(|printer| format!("Sent to printer: {printer}"))
+                    .map_err(|e| format!("Print failed: {e}"))
+                },
+                |result| Message::BackgroundIoFinished(result, false),
             );
-            let opts = self.plot_print_options(&d);
-            match exp.and_then(|_| crate::io::print_to_printer::print_existing_pdf(&tmp, &opts)) {
-                Ok(printer) => self
-                    .command_line
-                    .push_info(&format!("Sent to printer: {printer}")),
-                Err(e) => self.command_line.push_error(&format!("Print failed: {e}")),
-            }
-            return Task::none();
         }
 
         let (wires, hatches, wipeouts, eff_w, eff_h, ox, oy, rot) = self.layout_plot_params();
 
         if preview {
             let tmp = std::env::temp_dir().join("open_cad_studio_preview.pdf");
-            let res = crate::io::pdf_export::export_pdf(
+            return background_task(
+                move || {
+                    crate::io::pdf_export::export_pdf(
                 &wires,
                 &hatches,
                 &wipeouts,
@@ -2275,14 +2248,11 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
                 None,
                 &tmp,
                 plot_style.as_ref(),
+            ).and_then(|_| crate::io::print_to_printer::open_in_viewer(&tmp))
+                    .map(|_| "Opened plot preview.".to_string()).map_err(|e| format!("Preview failed: {e}"))
+                },
+                |result| Message::BackgroundIoFinished(result, true),
             );
-            match res.and_then(|_| crate::io::print_to_printer::open_in_viewer(&tmp)) {
-                Ok(()) => self.command_line.push_info("Opened plot preview."),
-                Err(e) => self.command_line.push_error(&format!("Preview failed: {e}")),
-            }
-            // Preview leaves the dialog open for further tweaks.
-            self.active_modal = Some(crate::app::ModalKind::Plot);
-            return Task::none();
         }
 
         if d.to_file {
@@ -2292,12 +2262,12 @@ pub(super) fn on_open_file(&mut self) -> Task<Message> {
 
         let opts = self.plot_print_options(&d);
         self.command_line.push_info("Sending to system printer…");
-        Task::perform(
-            async move {
+        background_task(
+            move || {
+                iced::futures::executor::block_on(
                 crate::io::print_to_printer::print_wires_with(
                     wires, hatches, wipeouts, eff_w, eff_h, ox, oy, rot, plot_style, opts,
-                )
-                .await
+                ))
             },
             Message::PrintResult,
         )

src/app/update/mod.rs

@@ -449,20 +449,26 @@ impl OpenCADStudio {
                     .file_name()
                     .map(|n| n.to_string_lossy().into_owned())
                     .unwrap_or_else(|| "unknown".into());
-                let phase = std::sync::Arc::new(std::sync::atomic::AtomicU8::new(
+                let progress = std::sync::Arc::new(crate::io::OpenProgressState::new(
                     super::OPEN_PHASE_READING,
                 ));
                 self.opening = Some(super::OpenProgress {
                     name: name.clone(),
                     size_bytes,
-                    phase: phase.clone(),
+                    state: progress.clone(),
                     started: Instant::now(),
                 });
                 let size_label = format_size(size_bytes);
                 self.command_line
                     .push_info(&format!("Opening \"{name}\" ({size_label})…"));
+                let model_bg = self.default_bg_color.unwrap_or([
+                    33.0 / 255.0,
+                    40.0 / 255.0,
+                    48.0 / 255.0,
+                    1.0,
+                ]);
                 Task::perform(
-                    crate::io::open_path_with_phase(path, phase),
+                    crate::io::open_path_with_phase(path, progress, model_bg),
                     Message::FileOpened,
                 )
             }
@@ -579,6 +585,19 @@ impl OpenCADStudio {
 
             Message::WblockSaveResult(_, None) => Task::none(),
 
+            Message::WblockWriteFinished(block_name, path, result) => {
+                match result {
+                    Ok(()) => self.command_line.push_output(&format!(
+                        "WBLOCK  Saved \"{block_name}\" → \"{}\"",
+                        path.display()
+                    )),
+                    Err(error) => self
+                        .command_line
+                        .push_error(&format!("WBLOCK save failed: {error}")),
+                }
+                Task::none()
+            }
+
             Message::DataExtractionSave(csv) => {
                 let csv_clone = csv.clone();
                 Task::perform(
@@ -643,6 +662,16 @@ impl OpenCADStudio {
 
             Message::StlExportPath(None) => Task::none(),
 
+            Message::StlExportFinished(path, result) => {
+                match result {
+                    Ok(()) => self
+                        .command_line
+                        .push_output(&format!("STLOUT: exported to \"{}\"", path.display())),
+                    Err(error) => self.command_line.push_error(&format!("STLOUT: {error}")),
+                }
+                Task::none()
+            }
+
             // ── STEP AP203 export ─────────────────────────────────────────
             Message::StepExport => {
                 let i = self.active_tab;
@@ -670,6 +699,16 @@ impl OpenCADStudio {
 
             Message::StepExportPath(None) => Task::none(),
 
+            Message::StepExportFinished(path, result) => {
+                match result {
+                    Ok(()) => self
+                        .command_line
+                        .push_output(&format!("STEPOUT: exported to \"{}\"", path.display())),
+                    Err(error) => self.command_line.push_error(&format!("STEPOUT: {error}")),
+                }
+                Task::none()
+            }
+
             // ── OBJ import ────────────────────────────────────────────────
             Message::ObjImport => Task::perform(
                 async {
@@ -688,6 +727,38 @@ impl OpenCADStudio {
 
             Message::ObjImportPath(None) => Task::none(),
 
+            Message::ObjImportFinished(tab_id, path, result) => {
+                match result {
+                    Err(error) => self.command_line.push_error(&format!("IMPORTOBJ: {error}")),
+                    Ok(mut mesh) => {
+                        let Some(i) = self.tabs.iter().position(|tab| tab.id == tab_id) else {
+                            self.command_line
+                                .push_info("IMPORTOBJ: target drawing was closed.");
+                            return Task::none();
+                        };
+                        let file_stem = path
+                            .file_stem()
+                            .map(|s| s.to_string_lossy().into_owned())
+                            .unwrap_or_else(|| "obj_mesh".into());
+                        mesh.name = file_stem.clone();
+                        self.push_undo_snapshot(i, "IMPORTOBJ");
+                        let entity = crate::modules::insert::solid3d_cmds::empty_solid3d();
+                        let handle = self.tabs[i].scene.add_entity(entity);
+                        if !handle.is_null() {
+                            self.tabs[i]
+                                .scene
+                                .meshes
+                                .insert(handle, crate::scene::MeshLodSet::from_single(mesh));
+                            self.tabs[i].dirty = true;
+                            self.command_line.push_output(&format!(
+                                "IMPORTOBJ: imported \"{file_stem}\" as mesh."
+                            ));
+                        }
+                    }
+                }
+                Task::none()
+            }
+
             Message::SaveFile => self.on_save_file(),
 
             Message::SaveAs => {
@@ -1132,7 +1203,7 @@ impl OpenCADStudio {
                 let targets = self.layer_row_action_targets(i, idx);
                 if let Some(on) = on {
                     if !targets.is_empty() {
-                        self.push_undo_snapshot(i, "LAYER OFF/ON");
+                        let undo = self.begin_layer_undo(i, "LAYER OFF/ON", &targets);
                         for name in &targets {
                             if let Some(dl) = self.tabs[i].scene.document.layers.get_mut(name) {
                                 dl.flags.off = !on;
@@ -1143,8 +1214,9 @@ impl OpenCADStudio {
                                 pl.visible = on;
                             }
                         }
-                        self.tabs[i].scene.bump_geometry();
+                        self.tabs[i].scene.invalidate_layer_dependencies(&targets);
                         self.tabs[i].dirty = true;
+                        self.commit_layer_undo(i, undo);
                         self.command_line.push_output(&format!(
                             "{} layer(s) turned {}",
                             targets.len(),
@@ -1171,7 +1243,7 @@ impl OpenCADStudio {
                 let targets = self.layer_row_action_targets(i, idx);
                 if let Some(locked) = locked {
                     if !targets.is_empty() {
-                        self.push_undo_snapshot(i, "LAYER LOCK/UNLOCK");
+                        let undo = self.begin_layer_undo(i, "LAYER LOCK/UNLOCK", &targets);
                         for name in &targets {
                             if let Some(dl) = self.tabs[i].scene.document.layers.get_mut(name) {
                                 dl.flags.locked = locked;
@@ -1182,8 +1254,9 @@ impl OpenCADStudio {
                                 pl.locked = locked;
                             }
                         }
-                        self.tabs[i].scene.bump_geometry();
+                        // Lock state affects editability, not rendered geometry.
                         self.tabs[i].dirty = true;
+                        self.commit_layer_undo(i, undo);
                         self.command_line.push_output(&format!(
                             "{} layer(s) {}",
                             targets.len(),
@@ -1201,7 +1274,7 @@ impl OpenCADStudio {
                 let targets = self.layer_row_action_targets(i, idx);
                 if let Some(frozen) = frozen {
                     if !targets.is_empty() {
-                        self.push_undo_snapshot(i, "LAYER FREEZE");
+                        let undo = self.begin_layer_undo(i, "LAYER FREEZE", &targets);
                         for name in &targets {
                             if let Some(dl) = self.tabs[i].scene.document.layers.get_mut(name) {
                                 if frozen {
@@ -1216,8 +1289,9 @@ impl OpenCADStudio {
                                 pl.frozen = frozen;
                             }
                         }
-                        self.tabs[i].scene.bump_geometry();
+                        self.tabs[i].scene.invalidate_layer_dependencies(&targets);
                         self.tabs[i].dirty = true;
+                        self.commit_layer_undo(i, undo);
                         self.command_line.push_output(&format!(
                             "{} layer(s) {}",
                             targets.len(),
@@ -1321,6 +1395,7 @@ impl OpenCADStudio {
                 // Apply to every selected layer (multi-select), not just one.
                 let names = self.selected_layer_names(i);
                 if !names.is_empty() {
+                    let undo = self.begin_layer_undo(i, "LAYER COLOR", &names);
                     use crate::ui::window::layers::iced_color_from_acad;
                     let new_color = iced_color_from_acad(&AcadColor::Index(aci));
                     for name in &names {
@@ -1334,10 +1409,11 @@ impl OpenCADStudio {
                         }
                     }
                     self.tabs[i].dirty = true;
+                    self.commit_layer_undo(i, undo);
                     // ByLayer color is baked into the cached wires at
                     // tessellation time, so bump the geometry epoch to
                     // invalidate the wire cache and repaint with the new color.
-                    self.tabs[i].scene.bump_geometry();
+                    self.tabs[i].scene.invalidate_layer_dependencies(&names);
                     self.tabs[i].layers.color_picker_row = None;
                     self.tabs[i].layers.color_full_palette = false;
                     self.sync_ribbon_layers();
@@ -1349,6 +1425,7 @@ impl OpenCADStudio {
                 let i = self.active_tab;
                 let names = self.selected_layer_names(i);
                 if !names.is_empty() {
+                    let undo = self.begin_layer_undo(i, "LAYER LINETYPE", &names);
                     for name in &names {
                         if let Some(dl) = self.tabs[i].scene.document.layers.get_mut(name) {
                             dl.line_type = lt.clone();
@@ -1360,8 +1437,9 @@ impl OpenCADStudio {
                         }
                     }
                     self.tabs[i].dirty = true;
+                    self.commit_layer_undo(i, undo);
                     // Linetype is baked into the cached wires; repaint.
-                    self.tabs[i].scene.bump_geometry();
+                    self.tabs[i].scene.invalidate_layer_dependencies(&names);
                 }
                 Task::none()
             }
@@ -1370,6 +1448,7 @@ impl OpenCADStudio {
                 let i = self.active_tab;
                 let names = self.selected_layer_names(i);
                 if !names.is_empty() {
+                    let undo = self.begin_layer_undo(i, "LAYER LINEWEIGHT", &names);
                     for name in &names {
                         if let Some(dl) = self.tabs[i].scene.document.layers.get_mut(name) {
                             dl.line_weight = lw;
@@ -1381,8 +1460,9 @@ impl OpenCADStudio {
                         }
                     }
                     self.tabs[i].dirty = true;
+                    self.commit_layer_undo(i, undo);
                     // Lineweight is baked into the cached wires; repaint.
-                    self.tabs[i].scene.bump_geometry();
+                    self.tabs[i].scene.invalidate_layer_dependencies(&names);
                 }
                 Task::none()
             }
@@ -3984,6 +4064,17 @@ impl OpenCADStudio {
             }
             Message::PlotWindowExportPath(Some(path)) => self.on_plot_window_export_path_some(path),
 
+            Message::BackgroundIoFinished(result, reopen_plot) => {
+                match result {
+                    Ok(message) => self.command_line.push_info(&message),
+                    Err(error) => self.command_line.push_error(&error),
+                }
+                if reopen_plot {
+                    self.active_modal = Some(crate::app::ModalKind::Plot);
+                }
+                Task::none()
+            }
+
             // ── Print to system printer ───────────────────────────────────────
             Message::PrintToPrinter => self.on_print_to_printer(),
             Message::PrintResult(Ok(printer)) => {

src/io/mod.rs

@@ -18,20 +18,58 @@ pub mod patterns;
 pub mod update_check;
 pub mod paper_sizes;
 pub mod thumbnail;
+#[cfg(target_arch = "wasm32")]
+mod web_worker;
 
 use crate::scene::DerivedCaches;
 use acadrust::entities::EntityType;
 use acadrust::io::dwg::DwgReader;
 use acadrust::{CadDocument, DwgWriter, DxfReader, DxfWriter};
 use std::path::{Path, PathBuf};
-use std::sync::atomic::{AtomicU8, Ordering};
+use std::sync::atomic::{AtomicU16, AtomicU32, AtomicU8, Ordering};
 use std::sync::Arc;
 
-// Phase tags written into the shared atomic so the UI overlay can display a
-// human-readable label. Kept in sync with the constants in `crate::app`.
-const PHASE_PARSING: u8 = 1;
-const PHASE_CACHING: u8 = 2;
-const PHASE_FINALIZING: u8 = 3;
+/// Thread-safe state shared by the native loader and the open overlay.
+///
+/// `basis_points` is monotonic in 0..=10000. `completed/total` describe the
+/// current sub-stage and let diagnostics distinguish real progress from a
+/// cosmetic timer.
+#[derive(Debug)]
+pub struct OpenProgressState {
+    pub phase: AtomicU8,
+    pub basis_points: AtomicU16,
+    pub completed: AtomicU32,
+    pub total: AtomicU32,
+}
+
+impl OpenProgressState {
+    pub fn new(phase: u8) -> Self {
+        Self {
+            phase: AtomicU8::new(phase),
+            basis_points: AtomicU16::new(0),
+            completed: AtomicU32::new(0),
+            total: AtomicU32::new(1),
+        }
+    }
+
+    pub fn set(&self, phase: u8, basis_points: u16, completed: usize, total: usize) {
+        self.completed
+            .store(completed.min(u32::MAX as usize) as u32, Ordering::Relaxed);
+        self.total.store(
+            total.max(1).min(u32::MAX as usize) as u32,
+            Ordering::Relaxed,
+        );
+        self.basis_points
+            .fetch_max(basis_points.min(10000), Ordering::Relaxed);
+        self.phase.store(phase, Ordering::Release);
+    }
+
+    pub fn set_fraction(&self, phase: u8, base: u16, span: u16, completed: usize, total: usize) {
+        let denominator = total.max(1) as u64;
+        let value = base as u64 + (completed.min(total.max(1)) as u64 * span as u64 / denominator);
+        self.set(phase, value.min(10000) as u16, completed, total);
+    }
+}
 
 // ── Open ──────────────────────────────────────────────────────────────────
 
@@ -61,37 +99,103 @@ pub async fn pick_open_path() -> Option<(PathBuf, u64)> {
 /// thread runs.
 pub async fn open_path_with_phase(
     path: PathBuf,
-    phase: Arc<AtomicU8>,
+    progress: Arc<OpenProgressState>,
+    model_bg: [f32; 4],
 ) -> Result<(String, PathBuf, CadDocument, DerivedCaches), String> {
     let name = path
         .file_name()
         .map(|n| n.to_string_lossy().into_owned())
         .unwrap_or_else(|| "unknown".into());
     let path2 = path.clone();
-    let phase2 = phase.clone();
-    let (doc, caches) = std::thread::spawn(move || -> Result<_, String> {
+    let progress2 = progress.clone();
+    let (sender, receiver) = iced::futures::channel::oneshot::channel();
+    std::thread::Builder::new()
+        .name("ocs-file-open".to_string())
+        .spawn(move || {
+            let result = (|| -> Result<_, String> {
         use iced::time::Instant;
-        phase2.store(PHASE_PARSING, Ordering::Relaxed);
+                progress2.set(crate::app::OPEN_PHASE_PARSING, 200, 0, 1000);
         let t_parse = Instant::now();
-        let mut doc = load_file(&path2)?;
+        let parser_progress = {
+                    let progress = Arc::clone(&progress2);
+                    let callback: Arc<dyn Fn(u16) + Send + Sync> = Arc::new(move |value| {
+                        progress.set_fraction(
+                            crate::app::OPEN_PHASE_PARSING,
+                            200,
+                            5600,
+                            value as usize,
+                            1000,
+                        );
+                    });
+                    callback
+                };
+                let mut doc = load_file_with_progress(&path2, Some(parser_progress))?;
         let parse_ms = t_parse.elapsed().as_millis() as u32;
+                progress2.set(crate::app::OPEN_PHASE_PARSING, 5800, 1000, 1000);
         let t_purge = Instant::now();
         let dropped = purge_corrupt_entities(&mut doc);
         let purge_ms = t_purge.elapsed().as_millis() as u32;
-        phase2.store(PHASE_CACHING, Ordering::Relaxed);
+                progress2.set(crate::app::OPEN_PHASE_XREF, 6000, 0, 1);
+                let t_xref = Instant::now();
+                let (xref_infos, xref_dropped) = if let Some(base_dir) = path2.parent() {
+                    let xref_progress = {
+                        let progress = Arc::clone(&progress2);
+                        let callback: Arc<dyn Fn(usize, usize) + Send + Sync> =
+                            Arc::new(move |completed, total| {
+                                progress.set_fraction(
+                                    crate::app::OPEN_PHASE_XREF,
+                                    6000,
+                                    1400,
+                                    completed,
+                                    total,
+                                );
+                            });
+                        callback
+                    };
+                    crate::io::xref::resolve_xrefs_with_progress(
+                        &mut doc,
+                        base_dir,
+                        Some(xref_progress),
+                    )
+                } else {
+                    (Vec::new(), 0)
+                };
+                let xref_ms = t_xref.elapsed().as_millis() as u32;
+                progress2.set(crate::app::OPEN_PHASE_CACHING, 7400, 0, 10000);
         let t_caches = Instant::now();
-        let mut caches = crate::scene::build_derived_caches(&doc);
+        let cache_progress = |value: u16| {
+                    progress2.set_fraction(
+                        crate::app::OPEN_PHASE_CACHING,
+                        7400,
+                        2200,
+                        value as usize,
+                        10000,
+                    );
+                };
+                let mut caches = crate::scene::build_derived_caches_with_progress(&doc, &cache_progress);
         caches.timings = crate::scene::OpenTimings {
             parse_ms,
             purge_ms,
             caches_ms: t_caches.elapsed().as_millis() as u32,
+                    xref_ms,
         };
         caches.corrupt_dropped = dropped;
-        phase2.store(PHASE_FINALIZING, Ordering::Relaxed);
+                caches.xref_dropped = xref_dropped;
+                caches.xrefs = xref_infos;
+                progress2.set(crate::app::OPEN_PHASE_FINALIZING, 9600, 0, 1);
+                let (prepared_doc, prepared_geometry) =
+                    crate::scene::prepare_open_geometry(doc, &caches, model_bg);
+                doc = prepared_doc;
+                caches.prepared_geometry = Some(prepared_geometry);
+                progress2.set(crate::app::OPEN_PHASE_FINALIZING, 9950, 1, 1);
         Ok((doc, caches))
+    })();
+            let _ = sender.send(result);
     })
-    .join()
-    .map_err(|_| "parser thread panicked".to_string())??;
+        .map_err(|error| format!("failed to start parser thread: {error}"))?;
+    let (doc, caches) = receiver
+        .await
+        .map_err(|_| "parser thread stopped without a result".to_string())??;
     Ok((name, path, doc, caches))
 }
 
@@ -102,6 +206,7 @@ pub async fn open_path_with_phase(
 /// stands in for the document path.
 #[cfg(target_arch = "wasm32")]
 pub async fn pick_and_load_web(
+    progress: Arc<OpenProgressState>,
 ) -> Result<(String, PathBuf, CadDocument, DerivedCaches), String> {
     let handle = crate::sys::file_dialog()
         .set_title("Open CAD file")
@@ -111,11 +216,23 @@ pub async fn pick_and_load_web(
         .await
         .ok_or_else(|| "Cancelled".to_string())?;
     let name = handle.file_name();
+    progress.set(crate::app::OPEN_PHASE_READING, 500, 1, 2);
     let bytes = handle.read().await;
-    let mut doc = load_bytes(&name, bytes)?;
+    progress.set(crate::app::OPEN_PHASE_PARSING, 1000, 0, 1);
+    let mut doc = match web_worker::parse_document(&name, bytes).await {
+        Ok(document) => document,
+        Err(error) => return Err(format!("Web parser worker: {error}")),
+    };
+    if name.to_ascii_lowercase().ends_with(".dxf") {
+        fix_dxf_dimension_rotations(&mut doc);
+    }
+    fix_viewport_status_flags(&mut doc);
+    fix_current_style_names(&mut doc);
+    progress.set(crate::app::OPEN_PHASE_CACHING, 7000, 0, 1);
     let dropped = purge_corrupt_entities(&mut doc);
     let mut caches = crate::scene::build_derived_caches(&doc);
     caches.corrupt_dropped = dropped;
+    progress.set(crate::app::OPEN_PHASE_FINALIZING, 9900, 1, 1);
     let path = PathBuf::from(&name);
     Ok((name, path, doc, caches))
 }
@@ -168,6 +285,13 @@ fn sniff_dwg_or_dxf(path: &Path) -> String {
 }
 
 pub fn load_file(path: &Path) -> Result<CadDocument, String> {
+    load_file_with_progress(path, None)
+}
+
+pub(crate) fn load_file_with_progress(
+    path: &Path,
+    _progress: Option<Arc<dyn Fn(u16) + Send + Sync>>,
+) -> Result<CadDocument, String> {
     let ext = path
         .extension()
         .map(|e| e.to_string_lossy().to_lowercase())
@@ -184,10 +308,15 @@ pub fn load_file(path: &Path) -> Result<CadDocument, String> {
     match effective.as_str() {
         "dwg" => {
             #[cfg(not(target_arch = "wasm32"))]
-            let mut doc = DwgReader::from_mmap(path)
-                .map_err(|e| e.to_string())?
-                .read()
+            let mut doc = {
+                let mut reader = DwgReader::from_mmap(path)
                 .map_err(|e| e.to_string())?;
+                if let Some(progress) = _progress {
+                    reader.set_progress_callback(progress);
+                }
+                reader.read()
+                .map_err(|e| e.to_string())?
+            };
             #[cfg(target_arch = "wasm32")]
             let mut doc = DwgReader::from_file(path)
                 .map_err(|e| e.to_string())?

src/io/web_worker.rs

@@ -0,0 +1,77 @@
+use std::cell::RefCell;
+use std::rc::Rc;
+
+use acadrust::CadDocument;
+use js_sys::{Array, Object, Reflect, Uint8Array};
+use wasm_bindgen::closure::Closure;
+use wasm_bindgen::{JsCast, JsValue};
+use web_sys::{ErrorEvent, MessageEvent, Worker, WorkerOptions, WorkerType};
+
+pub(super) async fn parse_document(name: &str, bytes: Vec<u8>) -> Result<CadDocument, String> {
+    let options = WorkerOptions::new();
+    options.set_type(WorkerType::Module);
+    let worker = Worker::new_with_options("ocs-parse-worker.js", &options).map_err(js_error)?;
+
+    let (sender, receiver) = iced::futures::channel::oneshot::channel();
+    let sender = Rc::new(RefCell::new(Some(sender)));
+    let message_sender = sender.clone();
+    let on_message = Closure::<dyn FnMut(MessageEvent)>::new(move |event: MessageEvent| {
+        let data = event.data();
+        let ok = Reflect::get(&data, &JsValue::from_str("ok"))
+            .ok()
+            .and_then(|value| value.as_bool())
+            .unwrap_or(false);
+        let result = if ok {
+            Reflect::get(&data, &JsValue::from_str("data"))
+                .map_err(js_error)
+                .and_then(|value| {
+                    let bytes = Uint8Array::new(&value).to_vec();
+                    bincode::deserialize(&bytes).map_err(|error| error.to_string())
+                })
+        } else {
+            Err(Reflect::get(&data, &JsValue::from_str("error"))
+                .ok()
+                .and_then(|value| value.as_string())
+                .unwrap_or_else(|| "CAD parser worker failed".to_string()))
+        };
+        if let Some(sender) = message_sender.borrow_mut().take() {
+            let _ = sender.send(result);
+        }
+    });
+    worker.set_onmessage(Some(on_message.as_ref().unchecked_ref()));
+
+    let error_sender = sender;
+    let on_error = Closure::<dyn FnMut(ErrorEvent)>::new(move |event: ErrorEvent| {
+        if let Some(sender) = error_sender.borrow_mut().take() {
+            let _ = sender.send(Err(event.message()));
+        }
+    });
+    worker.set_onerror(Some(on_error.as_ref().unchecked_ref()));
+
+    let payload = Object::new();
+    Reflect::set(
+        &payload,
+        &JsValue::from_str("name"),
+        &JsValue::from_str(name),
+    )
+    .map_err(js_error)?;
+    let input = Uint8Array::from(bytes.as_slice());
+    Reflect::set(&payload, &JsValue::from_str("bytes"), &input.buffer()).map_err(js_error)?;
+    let transfer = Array::new();
+    transfer.push(&input.buffer());
+    worker
+        .post_message_with_transfer(&payload, &transfer)
+        .map_err(js_error)?;
+
+    let result = receiver
+        .await
+        .map_err(|_| "CAD parser worker closed without a result".to_string())?;
+    worker.terminate();
+    result
+}
+
+fn js_error(value: JsValue) -> String {
+    value
+        .as_string()
+        .unwrap_or_else(|| format!("browser worker error: {value:?}"))
+}

src/io/xref.rs

@@ -41,6 +41,19 @@ pub struct XrefInfo {
 /// just like the host doc, so it gets the same corrupt-entity guard. Folding
 /// it in here avoids a second full-document `entities()` walk after resolve.
 pub fn resolve_xrefs(doc: &mut CadDocument, base_dir: &Path) -> (Vec<XrefInfo>, usize) {
+    resolve_xrefs_with_progress(doc, base_dir, None)
+}
+
+/// Resolve XREFs while reporting completed work units.
+///
+/// Each reference contributes 1000 parse units and 1000 merge units. Parsing
+/// progress comes directly from the DWG reader when available, so one large
+/// XREF advances smoothly instead of making the file-open bar appear frozen.
+pub fn resolve_xrefs_with_progress(
+    doc: &mut CadDocument,
+    base_dir: &Path,
+    progress: Option<std::sync::Arc<dyn Fn(usize, usize) + Send + Sync>>,
+) -> (Vec<XrefInfo>, usize) {
     // Auto-resolve every xref — frustum + LOD culling keep GPU cost bounded.
     let xref_entries: Vec<(String, String, Handle)> = doc
         .block_records
@@ -48,6 +61,11 @@ pub fn resolve_xrefs(doc: &mut CadDocument, base_dir: &Path) -> (Vec<XrefInfo>,
         .filter(|br| (br.flags.is_xref || br.flags.is_xref_overlay) && !br.xref_path.is_empty())
         .map(|br| (br.name.clone(), br.xref_path.clone(), br.handle))
         .collect();
+    let xref_count = xref_entries.len();
+    let total_units = xref_count.saturating_mul(2000);
+    if let Some(progress) = &progress {
+        progress(0, total_units);
+    }
 
     // Phase 1 — parse every referenced file in parallel. Each `load_file`
     // reads and decodes an independent DWG/DXF and touches nothing in the host
@@ -55,11 +73,40 @@ pub fn resolve_xrefs(doc: &mut CadDocument, base_dir: &Path) -> (Vec<XrefInfo>,
     // instead of running back-to-back. (The merge in phase 2 mutates `doc`, so
     // it stays serial.) `resolve_path` is pure and `base_dir` is shared &-ref.
     use crate::par::prelude::*;
+    let parse_units: std::sync::Arc<Vec<std::sync::atomic::AtomicU16>> = std::sync::Arc::new(
+        (0..xref_count)
+            .map(|_| std::sync::atomic::AtomicU16::new(0))
+            .collect(),
+    );
     let parsed: Vec<(String, String, Handle, Option<PathBuf>, Option<CadDocument>)> = xref_entries
         .into_par_iter()
-        .map(|(block_name, raw_path, br_handle)| {
+        .enumerate()
+        .map(|(xref_index, (block_name, raw_path, br_handle))| {
             let resolved = resolve_path(&raw_path, base_dir);
-            let xref_doc = resolved.as_ref().and_then(|p| super::load_file(p).ok());
+            let units = std::sync::Arc::clone(&parse_units);
+            let nested_progress = progress.as_ref().map(|progress| {
+                let progress = std::sync::Arc::clone(progress);
+                let callback: std::sync::Arc<dyn Fn(u16) + Send + Sync> =
+                    std::sync::Arc::new(move |value| {
+                        units[xref_index]
+                            .store(value.min(1000), std::sync::atomic::Ordering::Relaxed);
+                        let completed = units
+                            .iter()
+                            .map(|unit| unit.load(std::sync::atomic::Ordering::Relaxed) as usize)
+                            .sum();
+                        progress(completed, total_units);
+                    });
+                callback
+            });
+            let xref_doc = resolved.as_ref().and_then(|p| super::load_file_with_progress(p, nested_progress).ok());
+            parse_units[xref_index].store(1000, std::sync::atomic::Ordering::Relaxed);
+            if let Some(progress) = &progress {
+                let completed = parse_units
+                    .iter()
+                    .map(|unit| unit.load(std::sync::atomic::Ordering::Relaxed) as usize)
+                    .sum();
+                progress(completed, total_units);
+            }
             (block_name, raw_path, br_handle, resolved, xref_doc)
         })
         .collect();
@@ -68,7 +115,8 @@ pub fn resolve_xrefs(doc: &mut CadDocument, base_dir: &Path) -> (Vec<XrefInfo>,
     // block order (par_iter preserves it), so handle allocation is deterministic.
     let mut result = Vec::with_capacity(parsed.len());
     let mut dropped = 0usize;
-    for (block_name, raw_path, br_handle, resolved, xref_doc) in parsed {
+    for (merge_index, (block_name, raw_path, br_handle, resolved, xref_doc)) in parsed.into_iter().enumerate()
+    {
         let status = if let Some(xref_doc) = xref_doc {
             ensure_block_entities(doc, &block_name);
             dropped += merge_xref_into_block(doc, &block_name, br_handle, xref_doc);
@@ -85,6 +133,14 @@ pub fn resolve_xrefs(doc: &mut CadDocument, base_dir: &Path) -> (Vec<XrefInfo>,
                 .unwrap_or(raw_path),
             status,
         });
+        if let Some(progress) = &progress {
+            progress(
+                xref_count
+                    .saturating_mul(1000)
+                    .saturating_add((merge_index + 1).saturating_mul(1000)),
+                total_units,
+            );
+        }
     }
 
     (result, dropped)

src/modules/draw/modify/entity_index.rs

@@ -0,0 +1,99 @@
+//! Shared broad phase for interactive modify commands.
+//!
+//! The scene already narrows the cursor pick. Commands still need fast access
+//! to the picked analytic entity and, for intersection-heavy operations such
+//! as TRIM, to nearby boundary entities. Keeping this compact command-local
+//! index avoids cloning/scanning the complete drawing on every mouse move.
+
+use acadrust::{EntityType, Handle};
+use rustc_hash::{FxHashMap, FxHashSet};
+
+use crate::scene::convert::tess::entity_world_aabb_f64;
+use crate::scene::pick::quadtree::QuadTree;
+
+pub(super) struct ModifyEntityIndex {
+    by_handle: FxHashMap<Handle, usize>,
+    tree: Option<QuadTree>,
+    unbounded: Vec<Handle>,
+}
+
+impl ModifyEntityIndex {
+    pub(super) fn build(entities: &[EntityType]) -> Self {
+        let mut by_handle = FxHashMap::default();
+        let mut bounded = Vec::new();
+        let mut unbounded = Vec::new();
+        let mut world = [
+            f64::INFINITY,
+            f64::INFINITY,
+            f64::NEG_INFINITY,
+            f64::NEG_INFINITY,
+        ];
+
+        for (index, entity) in entities.iter().enumerate() {
+            let handle = entity.common().handle;
+            by_handle.insert(handle, index);
+            if let Some(aabb) = entity_world_aabb_f64(entity) {
+                world[0] = world[0].min(aabb[0]);
+                world[1] = world[1].min(aabb[1]);
+                world[2] = world[2].max(aabb[2]);
+                world[3] = world[3].max(aabb[3]);
+                bounded.push((handle, aabb));
+            } else {
+                unbounded.push(handle);
+            }
+        }
+
+        let tree = if bounded.is_empty() {
+            None
+        } else {
+            let span = (world[2] - world[0]).max(world[3] - world[1]).max(1.0);
+            let pad = span * 1.0e-9 + 1.0e-6;
+            let mut tree = QuadTree::new([
+                world[0] - pad,
+                world[1] - pad,
+                world[2] + pad,
+                world[3] + pad,
+            ]);
+            for (handle, aabb) in bounded {
+                tree.insert(handle, aabb);
+            }
+            Some(tree)
+        };
+
+        Self {
+            by_handle,
+            tree,
+            unbounded,
+        }
+    }
+
+    #[inline]
+    pub(super) fn get<'a>(
+        &self,
+        entities: &'a [EntityType],
+        handle: Handle,
+    ) -> Option<&'a EntityType> {
+        self.by_handle
+            .get(&handle)
+            .and_then(|index| entities.get(*index))
+    }
+
+    pub(super) fn nearby_handles(
+        &self,
+        entities: &[EntityType],
+        handle: Handle,
+    ) -> Option<FxHashSet<Handle>> {
+        let entity = self.get(entities, handle)?;
+        let aabb = entity_world_aabb_f64(entity)?;
+        let span = (aabb[2] - aabb[0]).max(aabb[3] - aabb[1]).max(1.0);
+        let pad = span * 1.0e-10 + 1.0e-7;
+        let query = [aabb[0] - pad, aabb[1] - pad, aabb[2] + pad, aabb[3] + pad];
+        let mut handles: FxHashSet<Handle> = self
+            .tree
+            .as_ref()
+            .map(|tree| tree.query_rect(query).into_iter().collect())
+            .unwrap_or_default();
+        handles.extend(self.unbounded.iter().copied());
+        Some(handles)
+    }
+}

src/modules/draw/modify/fillet.rs

@@ -21,6 +21,8 @@ use crate::modules::draw::defaults;
 use crate::modules::IconKind;
 use crate::scene::model::wire_model::WireModel;
 
+use super::entity_index::ModifyEntityIndex;
+
 // ── Dropdown constants ─────────────────────────────────────────────────────
 
 pub const DROPDOWN_ID: &str = "fillet_chamfer";
@@ -1204,6 +1206,7 @@ pub struct FilletCommand {
     radius: f64,
     step: FilletStep,
     all_entities: Vec<EntityType>,
+    entity_index: ModifyEntityIndex,
     /// First-object pick to restore after a radius entry made mid-selection
     /// (i.e. "R" pressed after the first object was already picked), so the
     /// command resumes at the second pick instead of restarting selection.
@@ -1212,10 +1215,12 @@ pub struct FilletCommand {
 
 impl FilletCommand {
     pub fn new(radius: f64, all_entities: Vec<EntityType>) -> Self {
+        let entity_index = ModifyEntityIndex::build(&all_entities);
         Self {
             radius: radius as f64,
             step: FilletStep::First,
             all_entities,
+            entity_index,
             resume_second: None,
         }
     }
@@ -1348,9 +1353,8 @@ impl CadCommand for FilletCommand {
             FilletStep::WaitingForRadius => return CmdResult::NeedPoint,
             FilletStep::First => {
                 let e1 = self
-                    .all_entities
-                    .iter()
-                    .find(|e| e.common().handle == handle)
+                    .entity_index
+                    .get(&self.all_entities, handle)
                     .and_then(|entity| match entity {
                         EntityType::LwPolyline(p) => {
                             Some(FilletEntity::from_lwpoly(p, handle, click))
@@ -1380,9 +1384,8 @@ impl CadCommand for FilletCommand {
                 }
 
                 let e2 = self
-                    .all_entities
-                    .iter()
-                    .find(|e| e.common().handle == handle)
+                    .entity_index
+                    .get(&self.all_entities, handle)
                     .and_then(|entity| match entity {
                         EntityType::LwPolyline(p) => {
                             Some(FilletEntity::from_lwpoly(p, handle, click))
@@ -1426,9 +1429,8 @@ impl CadCommand for FilletCommand {
             FilletStep::WaitingForRadius => vec![],
             FilletStep::First => {
                 let pts = self
-                    .all_entities
-                    .iter()
-                    .find(|e| e.common().handle == handle)
+                    .entity_index
+                    .get(&self.all_entities, handle)
                     .and_then(|e| match e {
                         EntityType::LwPolyline(p) => Some(lwpoly_seg_hover_pts(p, click)),
                         _ => FilletEntity::from_entity(e).map(|fe| entity_pts(&fe.to_entity_type())),
@@ -1449,9 +1451,8 @@ impl CadCommand for FilletCommand {
                 let e1 = e1.clone();
                 let click1 = *click1;
                 let e2 = self
-                    .all_entities
-                    .iter()
-                    .find(|e| e.common().handle == handle)
+                    .entity_index
+                    .get(&self.all_entities, handle)
                     .and_then(|entity| match entity {
                         EntityType::LwPolyline(p) => {
                             Some(FilletEntity::from_lwpoly(p, handle, click))
@@ -1583,6 +1584,7 @@ pub struct ChamferCommand {
     dist2: f64,
     step: ChamferStep,
     all_entities: Vec<EntityType>,
+    entity_index: ModifyEntityIndex,
     /// First-object pick (line or polyline segment) to restore after a
     /// distance entry made mid-selection, so the command resumes at the
     /// second pick instead of restarting selection.
@@ -1591,11 +1593,13 @@ pub struct ChamferCommand {
 
 impl ChamferCommand {
     pub fn new(dist: f64, all_entities: Vec<EntityType>) -> Self {
+        let entity_index = ModifyEntityIndex::build(&all_entities);
         Self {
             dist1: dist as f64,
             dist2: defaults::get_chamfer_dist2(),
             step: ChamferStep::First,
             all_entities,
+            entity_index,
             resume_pick: None,
         }
     }
@@ -1777,9 +1781,7 @@ impl CadCommand for ChamferCommand {
             }
             ChamferStep::First => {
                 match self
-                    .all_entities
-                    .iter()
-                    .find(|e| e.common().handle == handle)
+                    .entity_index.get(&self.all_entities, handle)
                 {
                     Some(EntityType::Line(l)) => {
                         self.step = ChamferStep::Second {
@@ -1821,9 +1823,8 @@ impl CadCommand for ChamferCommand {
                 }
 
                 let l2 = self
-                    .all_entities
-                    .iter()
-                    .find(|e| e.common().handle == handle)
+                    .entity_index
+                    .get(&self.all_entities, handle)
                     .and_then(|e| {
                         if let EntityType::Line(l) = e {
                             Some(l.clone())
@@ -1857,9 +1858,8 @@ impl CadCommand for ChamferCommand {
             ChamferStep::WaitingForDist1 | ChamferStep::WaitingForDist2 => return vec![],
             ChamferStep::First => {
                 let pts = self
-                    .all_entities
-                    .iter()
-                    .find(|e| e.common().handle == handle)
+                    .entity_index
+                    .get(&self.all_entities, handle)
                     .and_then(|e| match e {
                         EntityType::Line(l) => Some(line_pts(l)),
                         EntityType::LwPolyline(p) => Some(lwpoly_seg_hover_pts(p, click)),
@@ -1880,9 +1880,8 @@ impl CadCommand for ChamferCommand {
                 let l1 = l1.clone();
                 let click1 = *click1;
                 let l2 = self
-                    .all_entities
-                    .iter()
-                    .find(|e| e.common().handle == handle)
+                    .entity_index
+                    .get(&self.all_entities, handle)
                     .and_then(|e| {
                         if let EntityType::Line(l) = e {
                             Some(l.clone())

src/modules/draw/modify/mod.rs

@@ -4,6 +4,7 @@ pub mod attedit;
 pub mod break_cmd;
 pub mod copy;
 pub mod delete;
+mod entity_index;
 pub mod explode;
 pub mod fillet;
 pub mod join;

src/modules/draw/modify/offset.rs

@@ -25,6 +25,8 @@ use crate::modules::draw::defaults;
 use crate::modules::{IconKind, ModuleEvent, ToolDef};
 use crate::scene::model::wire_model::WireModel;
 
+use super::entity_index::ModifyEntityIndex;
+
 // ── Ribbon definition ──────────────────────────────────────────────────────
 
 pub fn tool() -> ToolDef {
@@ -691,6 +693,7 @@ enum Step {
 pub struct OffsetCommand {
     step: Step,
     all_entities: Vec<EntityType>,
+    entity_index: ModifyEntityIndex,
     /// Pre-selected offsettable objects (pick-first, #422); consumed when the
     /// distance step resolves.
     preselected: Vec<EntityType>,
@@ -712,9 +715,11 @@ pub fn is_offsettable(e: &EntityType) -> bool {
 
 impl OffsetCommand {
     pub fn new(all_entities: Vec<EntityType>) -> Self {
+        let entity_index = ModifyEntityIndex::build(&all_entities);
         Self {
             step: Step::Distance,
             all_entities,
+            entity_index,
             preselected: Vec::new(),
         }
     }
@@ -722,9 +727,11 @@ impl OffsetCommand {
     /// Pick-first flow (#422): the distance step still comes first, then the
     /// pre-selected objects go straight to the side step.
     pub fn with_selection(all_entities: Vec<EntityType>, targets: Vec<EntityType>) -> Self {
+        let entity_index = ModifyEntityIndex::build(&all_entities);
         Self {
             step: Step::Distance,
             all_entities,
+            entity_index,
             preselected: targets,
         }
     }
@@ -801,9 +808,7 @@ impl CadCommand for OffsetCommand {
         }
 
         let entity = self
-            .all_entities
-            .iter()
-            .find(|e| e.common().handle == handle)
+            .entity_index.get(&self.all_entities, handle)
             .cloned();
 
         // Accept every type compute_offset can offset — including XLine (#296),
@@ -881,9 +886,7 @@ impl CadCommand for OffsetCommand {
             return vec![];
         }
         if let Some(entity) = self
-            .all_entities
-            .iter()
-            .find(|e| e.common().handle == handle)
+            .entity_index.get(&self.all_entities, handle)
         {
             let pts = entity_wire_pts(entity);
             if !pts.is_empty() {

src/modules/draw/modify/trim.rs

@@ -27,6 +27,8 @@ use crate::modules::draw::modify::spline_ops::{
 use crate::modules::IconKind;
 use crate::scene::model::wire_model::WireModel;
 
+use super::entity_index::ModifyEntityIndex;
+
 // ── Dropdown constants ─────────────────────────────────────────────────────
 
 pub const DROPDOWN_ID: &str = "trim_extend";
@@ -269,6 +271,20 @@ enum Geo {
     },
 }
 
+impl Geo {
+    fn handle(&self) -> Handle {
+        match self {
+            Self::Line { handle, .. }
+            | Self::Arc { handle, .. }
+            | Self::Circle { handle, .. }
+            | Self::Ray { handle, .. }
+            | Self::InfLine { handle, .. }
+            | Self::Ellipse { handle, .. }
+            | Self::Spline { handle, .. } => *handle,
+        }
+    }
+}
+
 fn build_geos(entities: &[EntityType]) -> Vec<Geo> {
     let mut out = Vec::new();
     for e in entities {
@@ -2797,6 +2813,7 @@ fn crossing_preview_wire(p1: [f64; 2], cursor: [f64; 2], name: &str) -> WireMode
 
 pub struct TrimCommand {
     all_entities: Vec<EntityType>,
+    entity_index: ModifyEntityIndex,
     geos: Vec<Geo>,
     mode: TrimMode,
     /// Cutting-edge selection; empty = every object cuts (quick mode).
@@ -2809,9 +2826,11 @@ pub struct TrimCommand {
 
 impl TrimCommand {
     pub fn new(all_entities: Vec<EntityType>) -> Self {
+        let entity_index = ModifyEntityIndex::build(&all_entities);
         let geos = build_geos(&all_entities);
         Self {
             all_entities,
+            entity_index,
             geos,
             mode: TrimMode::Pick,
             edge_set: Vec::new(),
@@ -2823,6 +2842,7 @@ impl TrimCommand {
     /// Boundary geometry from the edge selection (or everything when none),
     /// with the Edge option's implied extrapolation applied on top.
     fn rebuild_geos(&mut self) {
+        self.entity_index = ModifyEntityIndex::build(&self.all_entities);
         self.geos = if self.edge_set.is_empty() {
             build_geos(&self.all_entities)
         } else {
@@ -2839,6 +2859,23 @@ impl TrimCommand {
         }
     }
 
+    /// Exact analytic boundaries whose boxes overlap the picked entity. The
+    /// host already narrowed the cursor pick; this second broad phase keeps
+    /// TRIM preview/intersection work local on dense drawings.
+    fn nearby_geos(&self, handle: Handle) -> Vec<Geo> {
+        if self.implied_edges {
+            return self.geos.clone();
+        }
+        let Some(handles) = self.entity_index.nearby_handles(&self.all_entities, handle) else {
+            return self.geos.clone();
+        };
+        self.geos
+            .iter()
+            .filter(|geo| handles.contains(&geo.handle()))
+            .cloned()
+            .collect()
+    }
+
     fn fence_run(
         &mut self,
         fence: &[[f64; 2]],
@@ -2974,7 +3011,8 @@ impl CadCommand for TrimCommand {
                     pick_extend_at(&self.all_entities, &self.geos, handle, px, py)
                         .map(|e| vec![e])
                 } else {
-                    pick_trim_at(&self.all_entities, &self.geos, handle, px, py)
+                    let geos = self.nearby_geos(handle);
+                    pick_trim_at(&self.all_entities, &geos, handle, px, py)
                 };
                 if let Some(new_entities) = new_entities {
                     // Snapshot is updated in on_entity_replaced once we know
@@ -3108,10 +3146,10 @@ impl CadCommand for TrimCommand {
             return vec![];
         }
 
+        let nearby_geos = self.nearby_geos(handle);
+        let geos = nearby_geos.as_slice();
         let entity = self
-            .all_entities
-            .iter()
-            .find(|e| e.common().handle == handle);
+            .entity_index.get(&self.all_entities, handle);
 
         let mut hover_wires = match entity {
             Some(EntityType::Line(l)) => {
@@ -3119,7 +3157,7 @@ impl CadCommand for TrimCommand {
                 let ay = l.start.y;
                 let bx = l.end.x;
                 let by = l.end.y;
-                let ts = line_seg_ts(ax, ay, bx, by, handle, &self.geos);
+                let ts = line_seg_ts(ax, ay, bx, by, handle, geos);
                 if ts.is_empty() {
                     return vec![];
                 }
@@ -3152,7 +3190,7 @@ impl CadCommand for TrimCommand {
                 let cy = a.center.y;
                 let a0 = a.start_angle;
                 let a1 = a.end_angle;
-                let ts = arc_seg_ts(cx, cy, a.radius, a0, a1, handle, &self.geos);
+                let ts = arc_seg_ts(cx, cy, a.radius, a0, a1, handle, geos);
                 if ts.is_empty() {
                     return vec![];
                 }
@@ -3176,7 +3214,7 @@ impl CadCommand for TrimCommand {
             Some(EntityType::Circle(c)) => {
                 let cx = c.center.x;
                 let cy = c.center.y;
-                let ts = arc_seg_ts(cx, cy, c.radius, 0.0, TAU, handle, &self.geos);
+                let ts = arc_seg_ts(cx, cy, c.radius, 0.0, TAU, handle, geos);
                 if ts.len() < 2 {
                     return vec![];
                 }
@@ -3205,7 +3243,7 @@ impl CadCommand for TrimCommand {
                 let by = r.base_point.y;
                 let ex = bx + r.direction.x * TRIM_EXTENT;
                 let ey = by + r.direction.y * TRIM_EXTENT;
-                let ts = line_seg_ts(bx, by, ex, ey, handle, &self.geos);
+                let ts = line_seg_ts(bx, by, ex, ey, handle, geos);
                 if ts.is_empty() {
                     return vec![];
                 }
@@ -3245,7 +3283,7 @@ impl CadCommand for TrimCommand {
                 let ey_start = by - x.direction.y * TRIM_EXTENT;
                 let ex_end = bx + x.direction.x * TRIM_EXTENT;
                 let ey_end = by + x.direction.y * TRIM_EXTENT;
-                let ts = line_seg_ts(ex_start, ey_start, ex_end, ey_end, handle, &self.geos);
+                let ts = line_seg_ts(ex_start, ey_start, ex_end, ey_end, handle, geos);
                 if ts.is_empty() {
                     return vec![];
                 }
@@ -3294,7 +3332,7 @@ impl CadCommand for TrimCommand {
                     t1 += TAU;
                 }
                 let ts = ellipse_seg_ts(
-                    e.center.x, e.center.y, a, b, nx, ny, t0, t1, handle, &self.geos,
+                    e.center.x, e.center.y, a, b, nx, ny, t0, t1, handle, geos,
                 );
                 if ts.is_empty() {
                     return vec![];
@@ -3321,7 +3359,7 @@ impl CadCommand for TrimCommand {
                 out
             }
             Some(EntityType::Spline(s)) => {
-                let ts = spline_seg_ts(s, handle, &self.geos);
+                let ts = spline_seg_ts(s, handle, geos);
                 if ts.is_empty() {
                     return vec![];
                 }
@@ -3348,7 +3386,7 @@ impl CadCommand for TrimCommand {
                 out
             }
             Some(EntityType::LwPolyline(p)) => {
-                let Some(survivors) = trim_lwpolyline(p, pt.x as f64, pt.y as f64, &self.geos)
+                let Some(survivors) = trim_lwpolyline(p, pt.x as f64, pt.y as f64, geos)
                 else {
                     return vec![];
                 };
@@ -3371,7 +3409,7 @@ impl CadCommand for TrimCommand {
         if self.implied_edges && !hover_wires.is_empty() {
             if let (Some(orig), Some(pieces)) = (
                 entity,
-                pick_trim_at(&self.all_entities, &self.geos, handle, pt.x, pt.y),
+                pick_trim_at(&self.all_entities, geos, handle, pt.x, pt.y),
             ) {
                 let cuts = piece_cut_points(orig, &pieces);
                 hover_wires.extend(implied_cut_guides(&self.all_entities, handle, &cuts));
@@ -3477,6 +3515,7 @@ impl CadCommand for TrimCommand {
 
 pub struct ExtendCommand {
     all_entities: Vec<EntityType>,
+    entity_index: ModifyEntityIndex,
     geos: Vec<Geo>,
     mode: TrimMode,
     /// Boundary-edge selection; empty = every object is a boundary.
@@ -3489,9 +3528,11 @@ pub struct ExtendCommand {
 
 impl ExtendCommand {
     pub fn new(all_entities: Vec<EntityType>) -> Self {
+        let entity_index = ModifyEntityIndex::build(&all_entities);
         let geos = build_geos(&all_entities);
         Self {
             all_entities,
+            entity_index,
             geos,
             mode: TrimMode::Pick,
             edge_set: Vec::new(),
@@ -3501,6 +3542,7 @@ impl ExtendCommand {
     }
 
     fn rebuild_geos(&mut self) {
+        self.entity_index = ModifyEntityIndex::build(&self.all_entities);
         self.geos = if self.edge_set.is_empty() {
             build_geos(&self.all_entities)
         } else {
@@ -3681,18 +3723,14 @@ impl CadCommand for ExtendCommand {
                 let mut out: Vec<WireModel> = self
                     .edge_set
                     .iter()
-                    .filter_map(|h| {
-                        self.all_entities.iter().find(|e| e.common().handle == *h)
-                    })
+                    .filter_map(|h| self.entity_index.get(&self.all_entities, *h))
                     .map(|e| {
                         WireModel::solid("edge_sel".into(), entity_pts(e), OPT_YELLOW, false)
                     })
                     .collect();
                 if !handle.is_null() && !self.edge_set.contains(&handle) {
                     if let Some(e) = self
-                        .all_entities
-                        .iter()
-                        .find(|e| e.common().handle == handle)
+                        .entity_index.get(&self.all_entities, handle)
                     {
                         let mut c = OPT_YELLOW;
                         c[3] = 0.45;
@@ -3732,9 +3770,7 @@ impl CadCommand for ExtendCommand {
         }
 
         let entity = self
-            .all_entities
-            .iter()
-            .find(|e| e.common().handle == handle);
+            .entity_index.get(&self.all_entities, handle);
         match entity {
             Some(EntityType::Line(l)) => {
                 let ax = l.start.x;

src/scene/cache/block_cache.rs

@@ -15,7 +15,7 @@
 // a visited set so a self-referential block produces a marker rather than
 // recursing forever.
 
-use rustc_hash::FxHashMap as HashMap;
+use rustc_hash::{FxHashMap as HashMap, FxHashSet as HashSet};
 use std::sync::Arc;
 
 use acadrust::types::{Color as AcadColor, LineWeight, Transform, Vector3};
@@ -153,6 +153,29 @@ pub struct BlockDefn {
 #[derive(Default, Debug)]
 pub struct BlockCache {
     defns: HashMap<String, Arc<BlockDefn>>,
+    prototype_blocks: HashSet<String>,
+    /// Fully expanded prototype for repeated, non-array inserts. The key omits
+    /// translation but includes the linear transform and every inherited style
+    /// input. A matching insert therefore reuses all nested expansion/style
+    /// work and only applies its translation to the immutable prototype.
+    expansion_prototypes: std::sync::Mutex<
+        HashMap<ExpansionPrototypeKey, Arc<std::sync::Mutex<Option<Arc<CachedExpansion>>>>>,
+    >,
+}
+
+#[derive(Clone, Debug, PartialEq, Eq, Hash)]
+struct ExpansionPrototypeKey {
+    block_name: String,
+    linear: [u64; 9],
+    insert_style: Vec<u32>,
+    selected: bool,
+    is_xref: bool,
+}
+
+#[derive(Debug)]
+struct CachedExpansion {
+    translation: [f64; 3],
+    wires: Arc<Vec<WireModel>>,
 }
 
 impl BlockCache {
@@ -181,6 +204,18 @@ impl BlockCache {
         use crate::par::prelude::*;
         let mut cache = Self::new();
         let referenced = collect_referenced_blocks(doc);
+        let mut reference_counts: HashMap<String, usize> = HashMap::default();
+        for entity in doc.entities() {
+            if let EntityType::Insert(insert) = entity {
+                *reference_counts
+                    .entry(insert.block_name.clone())
+                    .or_default() += insert.instance_count();
+            }
+        }
+        cache.prototype_blocks = reference_counts
+            .into_iter()
+            .filter_map(|(name, count)| (count > 1).then_some(name))
+            .collect();
         // Each defn is built independently: nested INSERTs are stored as
         // by-name references (`LocalSub::Nested`), never expanded here, so a
         // block's build never depends on another block's defn. That makes the
@@ -269,7 +304,6 @@ impl BlockCache {
 /// Walk all entities + all block_record contents collecting every distinct
 /// `block_name` that appears in an Insert (transitively).
 fn collect_referenced_blocks(doc: &CadDocument) -> Vec<String> {
-    use rustc_hash::FxHashSet as HashSet;
     let mut seen: HashSet<String> = HashSet::default();
     let mut queue: Vec<String> = Vec::new();
 
@@ -795,6 +829,61 @@ pub fn expand_insert(
         xform = xform.then(&scale_about_p);
     }
     let name = ins_handle.value().to_string();
+    let prototype_key = if view_aabb.is_none()
+        && world_per_pixel.is_none()
+        && !ins.is_array()
+        && cache.prototype_blocks.contains(&ins.block_name)
+    {
+        Some(expansion_prototype_key(
+            ins,
+            &xform,
+            ins_resolved_color,
+            ins_pat_len,
+            ins_pat,
+            ins_lw_px,
+            ins_layer,
+            selected,
+            pslt_factor,
+            is_xref,
+            bg_color,
+            anno_scale,
+        ))
+    } else {
+        None
+    };
+    let prototype_slot = prototype_key.as_ref().map(|key| {
+        let mut prototypes = cache
+            .expansion_prototypes
+            .lock()
+            .unwrap_or_else(|poisoned| poisoned.into_inner());
+        Arc::clone(
+            prototypes
+                .entry(key.clone())
+                .or_insert_with(|| Arc::new(std::sync::Mutex::new(None))),
+        )
+    });
+    let mut prototype_guard = prototype_slot
+        .as_ref()
+        .map(|slot| slot.lock().unwrap_or_else(|poisoned| poisoned.into_inner()));
+    if let Some(cached) = prototype_guard
+        .as_ref()
+        .and_then(|guard| guard.as_ref())
+        .cloned()
+    {
+        let translation = transform_translation(&xform);
+        let delta = [
+            translation[0] - cached.translation[0],
+            translation[1] - cached.translation[1],
+            translation[2] - cached.translation[2],
+        ];
+        return Some(
+            cached
+                .wires
+                .iter()
+                .map(|wire| translated_prototype_wire(wire, &name, delta))
+                .collect(),
+        );
+    }
     let mut batches = Batches::default();
     let mut visited: Vec<String> = Vec::with_capacity(8);
 
@@ -848,7 +937,135 @@ pub fn expand_insert(
         };
         expand_defn(defn, &base_xform, &ctx, &mut batches, &mut visited, 0, (0.0, 1.0));
     }
-    Some(batches.finalize(&name, selected, bg_color))
+    let wires = batches.finalize(&name, selected, bg_color);
+    if let Some(guard) = prototype_guard.as_mut() {
+        let cached = Arc::new(CachedExpansion {
+            translation: transform_translation(&xform),
+            wires: Arc::new(wires.clone()),
+        });
+        **guard = Some(cached);
+    }
+    Some(wires)
+}
+
+fn transform_translation(transform: &Transform) -> [f64; 3] {
+    [
+        transform.matrix.m[0][3],
+        transform.matrix.m[1][3],
+        transform.matrix.m[2][3],
+    ]
+}
+
+#[allow(clippy::too_many_arguments)]
+fn expansion_prototype_key(
+    ins: &acadrust::entities::Insert,
+    transform: &Transform,
+    ins_color: [f32; 4],
+    ins_pat_len: f32,
+    ins_pat: [f32; 8],
+    ins_lw_px: f32,
+    ins_layer: crate::scene::view::render::InheritStyle,
+    selected: bool,
+    pslt_factor: f32,
+    is_xref: bool,
+    bg_color: [f32; 4],
+    anno_scale: f32,
+) -> ExpansionPrototypeKey {
+    let matrix = &transform.matrix.m;
+    let linear = [
+        matrix[0][0].to_bits(),
+        matrix[0][1].to_bits(),
+        matrix[0][2].to_bits(),
+        matrix[1][0].to_bits(),
+        matrix[1][1].to_bits(),
+        matrix[1][2].to_bits(),
+        matrix[2][0].to_bits(),
+        matrix[2][1].to_bits(),
+        matrix[2][2].to_bits(),
+    ];
+    let mut insert_style = Vec::with_capacity(32);
+    insert_style.extend(ins_color.map(f32::to_bits));
+    insert_style.push(ins_pat_len.to_bits());
+    insert_style.extend(ins_pat.map(f32::to_bits));
+    insert_style.push(ins_lw_px.to_bits());
+    insert_style.extend(ins_layer.color.map(f32::to_bits));
+    insert_style.push(ins_layer.pat_len.to_bits());
+    insert_style.extend(ins_layer.pat.map(f32::to_bits));
+    insert_style.push(ins_layer.lw_px.to_bits());
+    insert_style.push(pslt_factor.to_bits());
+    insert_style.extend(bg_color.map(f32::to_bits));
+    insert_style.push(anno_scale.to_bits());
+    ExpansionPrototypeKey {
+        block_name: ins.block_name.clone(),
+        linear,
+        insert_style,
+        selected,
+        is_xref,
+    }
+}
+
+fn translated_prototype_wire(source: &WireModel, name: &str, delta: [f64; 3]) -> WireModel {
+    let mut wire = source.clone();
+    wire.name = name.to_string();
+    translate_double_single(&mut wire.points, &mut wire.points_low, delta);
+    translate_double_single(&mut wire.fill_tris, &mut wire.fill_tris_low, delta);
+    translate_double_single(&mut wire.pick_tris, &mut wire.pick_tris_low, delta);
+    for (point, _) in &mut wire.snap_pts {
+        point.x += delta[0];
+        point.y += delta[1];
+        point.z += delta[2];
+    }
+    for point in &mut wire.key_vertices {
+        point[0] += delta[0];
+        point[1] += delta[1];
+        point[2] += delta[2];
+    }
+    let delta_f32 = [delta[0] as f32, delta[1] as f32, delta[2] as f32];
+    for tangent in &mut wire.tangent_geoms {
+        match tangent {
+            TangentGeom::Line { p1, p2 } => {
+                for axis in 0..3 {
+                    p1[axis] += delta_f32[axis];
+                    p2[axis] += delta_f32[axis];
+                }
+            }
+            TangentGeom::Circle { center, .. } => {
+                for axis in 0..3 {
+                    center[axis] += delta_f32[axis];
+                }
+            }
+        }
+    }
+    if !wire.text_verts.is_empty() {
+        wire.text_verts =
+            crate::scene::model::wire_model::map_text_verts(&wire.text_verts, |x, y, z| {
+                (x + delta[0], y + delta[1], z + delta[2])
+            });
+    }
+    if wire.aabb != WireModel::UNBOUNDED_AABB {
+        wire.aabb[0] += delta_f32[0];
+        wire.aabb[1] += delta_f32[1];
+        wire.aabb[2] += delta_f32[0];
+        wire.aabb[3] += delta_f32[1];
+    }
+    wire
+}
+
+fn translate_double_single(points: &mut [[f32; 3]], lows: &mut Vec<[f32; 3]>, delta: [f64; 3]) {
+    if points.is_empty() {
+        return;
+    }
+    if lows.len() != points.len() {
+        lows.resize(points.len(), [0.0; 3]);
+    }
+    for (point, low) in points.iter_mut().zip(lows.iter_mut()) {
+        for axis in 0..3 {
+            let value = point[axis] as f64 + low[axis] as f64 + delta[axis];
+            let high = value as f32;
+            point[axis] = high;
+            low[axis] = (value - high as f64) as f32;
+        }
+    }
 }
 
 fn aabb_pixel_size(local_aabb: [f32; 4], world_per_pixel: f32) -> f32 {

src/scene/entity.rs

@@ -199,6 +199,7 @@ impl Scene {
         };
 
         if !handle.is_null() {
+            self.invalidate_dependency_index();
             if let Some(model) = hatch_seed {
                 self.hatches.insert(handle, model);
             }
@@ -275,6 +276,7 @@ impl Scene {
                 }
             }
         }
+        self.invalidate_dependency_index();
         self.bump_geometry();
         true
     }
@@ -353,6 +355,7 @@ impl Scene {
             return false;
         };
         *slot = entity;
+        self.invalidate_dependency_index();
 
         // Drop stale derived caches for this handle, then reseed for the new
         // entity's type (which may differ from the old one).
@@ -2092,6 +2095,7 @@ impl Scene {
     /// Rebuild hatch / image / mesh caches after the document is modified
     /// outside the normal `add_entity` path (e.g. REFCLOSE SAVE).
     pub fn rebuild_derived_caches(&mut self) {
+        self.invalidate_dependency_index();
         self.populate_hatches_from_document_unbumped();
         self.populate_images_from_document_unbumped();
         self.populate_meshes_impl(false, false);

src/scene/mod.rs

@@ -45,6 +45,22 @@ pub(super) struct EntityIndex {
     pub unbounded_handles: Vec<Handle>,
 }
 
+#[derive(Default)]
+struct DependencyTargets {
+    render_handles: HashSet<Handle>,
+    source_handles: HashSet<Handle>,
+    touches_block_definition: bool,
+}
+
+#[derive(Default)]
+struct SceneDependencyIndex {
+    layers: HashMap<String, DependencyTargets>,
+    text_styles: HashMap<String, DependencyTargets>,
+    dim_styles: HashMap<String, DependencyTargets>,
+    object_styles: HashMap<Handle, DependencyTargets>,
+    blocks: HashMap<String, HashSet<Handle>>,
+}
+
 fn hatch_interaction_aabb(hatch: &model::hatch_model::HatchModel) -> Option<[f64; 4]> {
     let mut aabb = [
         f64::INFINITY,
@@ -238,6 +254,21 @@ struct ResidentWireLayout {
     marker_start: usize,
 }
 
+#[derive(Clone)]
+pub struct PreparedOpenGeometry {
+    pub wires: Arc<Vec<WireModel>>,
+    pub interaction_index: Option<Arc<crate::scene::pick::interaction_index::InteractionIndex>>,
+}
+
+impl std::fmt::Debug for PreparedOpenGeometry {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("PreparedOpenGeometry")
+            .field("wires", &self.wires.len())
+            .field("interaction_index", &self.interaction_index.is_some())
+            .finish()
+    }
+}
+
 #[derive(Debug)]
 pub(crate) struct WireGpuPatch {
     pub(crate) changes: Arc<Vec<(Handle, ChangeKind)>>,
@@ -357,6 +388,15 @@ pub struct DerivedCaches {
     /// Reported back to the UI so the user knows when a file had parser-junk
     /// entities silently dropped.
     pub corrupt_dropped: usize,
+    /// Corrupt entities dropped while resolving referenced drawings.
+    pub xref_dropped: usize,
+    /// XREF resolution results produced by the loader worker. Keeping these in
+    /// the open bundle prevents parsing and merging references on the UI thread.
+    pub xrefs: Vec<crate::io::xref::XrefInfo>,
+    /// Model wire set and its spatial interaction index, prepared on the loader
+    /// thread. Installing these prevents the first visible frame from paying a
+    /// whole-drawing tessellation/index build while the progress overlay freezes.
+    pub prepared_geometry: Option<PreparedOpenGeometry>,
     /// Background-thread open-phase timings in milliseconds (parse, purge,
     /// derived-cache build). Filled in by `open_path_with_phase`; surfaced in
     /// the open-complete breakdown log so open-time regressions are visible.
@@ -369,11 +409,30 @@ pub struct OpenTimings {
     pub parse_ms: u32,
     pub purge_ms: u32,
     pub caches_ms: u32,
+    pub xref_ms: u32,
 }
 
 /// Build hatch / image / mesh caches from a document without needing `&mut Scene`.
 /// Intended to run on a background thread during file load.
 pub fn build_derived_caches(doc: &CadDocument) -> DerivedCaches {
+    build_derived_caches_impl(doc, None)
+}
+
+/// Build open-time caches while reporting monotonic progress in 0..=10000.
+///
+/// The callback is UI-agnostic and may run from Rayon workers. Callers should
+/// keep it cheap, normally just updating atomics.
+pub fn build_derived_caches_with_progress(
+    doc: &CadDocument,
+    progress: &(dyn Fn(u16) + Sync),
+) -> DerivedCaches {
+    build_derived_caches_impl(doc, Some(progress))
+}
+
+fn build_derived_caches_impl(
+    doc: &CadDocument,
+    progress: Option<&(dyn Fn(u16) + Sync)>,
+) -> DerivedCaches {
     // A new drawing must not inherit the previous one's resolved images — drop
     // the memoised set so each reference re-reads / re-fetches once here (and
     // stays cached across this document's later cache rebuilds).
@@ -428,7 +487,8 @@ pub fn build_derived_caches(doc: &CadDocument) -> DerivedCaches {
     let mut image_handles: Vec<Handle> = Vec::new();
     let mut mesh_handles: Vec<Handle> = Vec::new();
     let mut centers: Vec<[f64; 3]> = Vec::new();
-    for e in doc.entities() {
+    let entity_total = doc.entity_count().max(1);
+    for (index, e) in doc.entities().enumerate() {
         let h = e.common().handle;
         match e {
             EntityType::Hatch(_) | EntityType::Solid(_) => hatch_handles.push(h),
@@ -444,6 +504,14 @@ pub fn build_derived_caches(doc: &CadDocument) -> DerivedCaches {
         if let Some(c) = offset_centroid(e, model_block, &prep) {
             centers.push(c);
         }
+        if index & 0x1fff == 0 {
+            if let Some(progress) = progress {
+                progress(((index as u64 * 4000) / entity_total as u64) as u16);
+            }
+        }
+    }
+    if let Some(progress) = progress {
+        progress(4000);
     }
     let (local_center, local_extent_max) = cluster_extent_from_centers(centers, &doc.header);
 
@@ -453,6 +521,19 @@ pub fn build_derived_caches(doc: &CadDocument) -> DerivedCaches {
     // the per-layout adaptation kicks in later regardless).
     const LOAD_BG: [f32; 4] = [33.0 / 255.0, 40.0 / 255.0, 48.0 / 255.0, 1.0];
 
+    let detail_total = hatch_handles
+        .len()
+        .saturating_add(image_handles.len())
+        .saturating_add(mesh_handles.len())
+        .max(1);
+    let detail_done = std::sync::atomic::AtomicUsize::new(0);
+    let report_detail = |done: usize| {
+        if let Some(progress) = progress {
+            let value = 4000u64 + done as u64 * 6000 / detail_total as u64;
+            progress(value.min(10000) as u16);
+        }
+    };
+
     // hatches
     let hatches: HashMap<Handle, HatchModel> = hatch_handles
         .par_iter()
@@ -465,15 +546,23 @@ pub fn build_derived_caches(doc: &CadDocument) -> DerivedCaches {
                 EntityType::Solid(solid) => Some(Scene::solid_hatch_model(solid, color)),
                 _ => None,
             };
-            model.map(|m| (handle, m))
+            let result = model.map(|m| (handle, m));
+            let done = detail_done.fetch_add(1, Ordering::Relaxed) + 1;
+            if done & 0xff == 0 || done == detail_total {
+                report_detail(done);
+            }
+            result
         })
         .collect();
 
     // images
     let images: HashMap<Handle, ImageModel> = image_handles
         .par_iter()
-        .filter_map(|&handle| match doc.get_entity(handle)? {
-            EntityType::RasterImage(img) => ImageModel::from_raster_image(img).map(|m| (handle, m)),
+        .filter_map(|&handle| {
+            let result = match doc.get_entity(handle)? {
+                EntityType::RasterImage(img) => {
+                    ImageModel::from_raster_image(img).map(|m| (handle, m))
+                }
             EntityType::Ole2Frame(ole) => ImageModel::from_ole2frame(ole).map(|m| (handle, m)),
             EntityType::Underlay(u) => match doc.objects.get(&u.definition_handle) {
                 Some(acadrust::objects::ObjectType::UnderlayDefinition(def)) => {
@@ -482,6 +571,12 @@ pub fn build_derived_caches(doc: &CadDocument) -> DerivedCaches {
                 _ => None,
             },
             _ => None,
+            };
+            let done = detail_done.fetch_add(1, Ordering::Relaxed) + 1;
+            if done & 0xff == 0 || done == detail_total {
+                report_detail(done);
+            }
+            result
         })
         .collect();
 
@@ -511,10 +606,16 @@ pub fn build_derived_caches(doc: &CadDocument) -> DerivedCaches {
             let (raw, ..) = view::render::render_style_for(doc, e);
             let color = view::render::adapt_to_bg(raw, LOAD_BG);
             let top_level = layout_blocks.contains(&e.common().owner_handle);
-            crate::entities::solid3d::tessellate_volume(e, color, facet_res, isolines).map(|m| {
+            let result = crate::entities::solid3d::tessellate_volume(e, color, facet_res, isolines)
+                .map(|m| {
                 let m = if top_level { offset_mesh_lod_set(m) } else { m };
                 (handle, m, top_level)
-            })
+                });
+            let done = detail_done.fetch_add(1, Ordering::Relaxed) + 1;
+            if done & 0xff == 0 || done == detail_total {
+                report_detail(done);
+            }
+            result
         })
         .collect();
     let mut meshes: HashMap<Handle, MeshLodSet> = HashMap::default();
@@ -527,6 +628,10 @@ pub fn build_derived_caches(doc: &CadDocument) -> DerivedCaches {
         }
     }
 
+    if let Some(progress) = progress {
+        progress(10000);
+    }
+
     DerivedCaches {
         local_extent_max,
         local_center,
@@ -535,10 +640,54 @@ pub fn build_derived_caches(doc: &CadDocument) -> DerivedCaches {
         meshes,
         block_meshes,
         corrupt_dropped: 0,
+        xref_dropped: 0,
+        xrefs: Vec::new(),
+        prepared_geometry: None,
         timings: OpenTimings::default(),
     }
 }
 
+/// Prepare the expensive first Model wire set and spatial interaction index on
+/// the loader thread. The temporary `Scene` never crosses threads (it contains
+/// `Rc`/`RefCell` state); only its Send-safe document and immutable prepared
+/// geometry are returned.
+pub fn prepare_open_geometry(
+    doc: CadDocument,
+    caches: &DerivedCaches,
+    model_bg: [f32; 4],
+) -> (CadDocument, PreparedOpenGeometry) {
+    let mut scene = Scene::new();
+    scene.document = doc;
+    scene.local_extent_max = caches.local_extent_max;
+    scene.local_center = caches.local_center;
+    scene.bg_color = model_bg;
+    let cannoscale_value = scene.document.header.annotation_scale_value;
+    scene.annotation_scale = if cannoscale_value > 1e-9 {
+        (1.0 / cannoscale_value) as f32
+    } else {
+        1.0
+    };
+    scene.current_layout = "Model".to_string();
+    let camera = scene.camera.borrow().clone();
+    let wires = scene.model_tile_wires_arc(0, &camera, 1.0, 1.0);
+    let interaction_index = if scene.interaction_index_worthwhile(&wires) {
+        let index =
+            Arc::new(crate::scene::pick::interaction_index::InteractionIndex::build(&wires));
+        index.prepare_screen();
+        Some(index)
+    } else {
+        None
+    };
+    let doc = std::mem::replace(&mut scene.document, CadDocument::new());
+    (
+        doc,
+        PreparedOpenGeometry {
+            wires,
+            interaction_index,
+        },
+    )
+}
+
 /// Mirrors `cache::block_cache::SANE_EXTENT` — wire coords past this magnitude
 /// are treated as corruption rather than precision-relevant geometry.
 const CLUSTER_SANE_EXTENT: f64 = 1.0e8;
@@ -1165,13 +1314,10 @@ pub struct Scene {
     /// filtered variants. Viewports sharing a frozen set share one entry (like
     /// the resident wire set). Empty for a viewport with no frozen layers (it
     /// reuses the unfiltered `*_arc` sets directly).
-    frozen_hatch_cache:
-        RefCell<HashMap<(String, u64), (u64, u64, Arc<Vec<HatchModel>>)>>,
-    frozen_wipeout_cache:
-        RefCell<HashMap<(String, u64), (u64, Arc<Vec<HatchModel>>)>>,
+    frozen_hatch_cache: RefCell<HashMap<(String, u64), (u64, u64, Arc<Vec<HatchModel>>)>>,
+    frozen_wipeout_cache: RefCell<HashMap<(String, u64), (u64, Arc<Vec<HatchModel>>)>>,
     frozen_image_cache: RefCell<HashMap<u64, (u64, Arc<Vec<ImageModel>>)>>,
-    frozen_mesh_cache:
-        RefCell<HashMap<(String, u64), (u64, Arc<Vec<MeshLodSet>>)>>,
+    frozen_mesh_cache: RefCell<HashMap<(String, u64), (u64, Arc<Vec<MeshLodSet>>)>>,
     /// Cached block-INSERT hatches for hit-testing, keyed by geometry_epoch.
     /// Building this explodes every model-space INSERT, so without the cache a
     /// heavy block-instanced drawing re-explodes thousands of inserts on every
@@ -1258,12 +1404,16 @@ pub struct Scene {
     /// Reverse map: entity_handle → block_record_handle, built from entity_handles lists.
     /// Keyed by geometry_epoch. Eliminates the O(B) fallback scan in belongs_to_visible_block.
     entity_block_map_cache: RefCell<Option<(u64, HashMap<Handle, Handle>)>>,
+    /// Reverse dependencies from layer/style/block definitions to the top-level
+    /// entities whose resident wire runs actually change. Kept independent from
+    /// `geometry_epoch`: a layer colour toggle can reuse the index, invalidate
+    /// only its dependants, and avoid a whole-document scan on every toggle.
+    dependency_index_cache: RefCell<Option<SceneDependencyIndex>>,
     /// Tessellated block definitions in block-local coords, keyed by render
     /// background and block epoch. Model and Paper adapt black/white colours
     /// differently; retaining both variants prevents a full block rebuild on
     /// every layout-tab switch.
-    block_defn_cache:
-        RefCell<HashMap<[u32; 4], (u64, Arc<cache::block_cache::BlockCache>)>>,
+    block_defn_cache: RefCell<HashMap<[u32; 4], (u64, Arc<cache::block_cache::BlockCache>)>>,
     /// Spatial index + always-emit list for top-level entities
     /// (Phase 2.1). Lazily rebuilt by `entity_index()` on
     /// `geometry_epoch` change. See `EntityIndex` for what each side
@@ -1455,6 +1605,7 @@ impl Scene {
             annotation_affects_wires: std::cell::Cell::new(None),
             model_extents_cache: RefCell::new(None),
             entity_block_map_cache: RefCell::new(None),
+            dependency_index_cache: RefCell::new(None),
             block_defn_cache: RefCell::new(HashMap::default()),
             entity_index_cache: RefCell::new(None),
             last_render_aspect: std::cell::Cell::new(16.0 / 9.0),
@@ -1555,6 +1706,30 @@ impl Scene {
         arc
     }
 
+    /// Install loader-thread geometry into this scene's Model resident cache.
+    ///
+    /// The target scene has a different epoch from the temporary loader scene,
+    /// so only immutable geometry is transferred and re-stamped here. The
+    /// optional interaction index is cached against the exact same `Arc`.
+    pub fn install_prepared_open_geometry(&self, prepared: PreparedOpenGeometry) {
+        let block = self.model_space_block_handle();
+        let key = Self::resident_wire_key(block, self.bg_color, None, None);
+        let gen = WIRE_CONTENT_GEN.fetch_add(1, Ordering::Relaxed);
+        self.last_model_wire_gen.set(gen);
+        self.resident_wire_sets.borrow_mut().insert(
+            key,
+            ResidentWireSet {
+                epoch: self.geometry_epoch,
+                gen,
+                wires: Arc::clone(&prepared.wires),
+                layout: None,
+            },
+        );
+        if let Some(index) = prepared.interaction_index {
+            self.cache_interaction_index(self.geometry_epoch, prepared.wires, index);
+        }
+    }
+
     /// Push one delta onto the journal, evicting the oldest past the cap and
     /// raising the floor so a consumer that fell behind falls back to a full
     /// rebuild. Every `geometry_epoch` bump must call this exactly once so the
@@ -1779,6 +1954,10 @@ impl Scene {
             }
         }
 
+        if !changes.is_empty() {
+            self.invalidate_dependency_index();
+        }
+
         changes
     }
 
@@ -3245,30 +3424,7 @@ impl Scene {
         } else {
             self.paper_bg_color
         };
-        let key = {
-            let mut k: u64 = 0xcbf2_9ce4_8422_2325;
-            let mut mix = |x: u64| k = k.rotate_left(17) ^ x.wrapping_mul(0x9E37_79B9_7F4A_7C15);
-            mix(block.value());
-            for c in bg {
-                mix(c.to_bits() as u64);
-            }
-            mix(anno_scale_override
-                .map(|a| a.to_bits() as u64)
-                .unwrap_or(u64::MAX));
-            match frozen_layers {
-                Some(f) => {
-                    // Order-independent fold of the frozen-layer set.
-                    let mut acc: u64 = 0;
-                    for h in f {
-                        acc ^= h.value().wrapping_mul(0x9E37_79B9_7F4A_7C15);
-                    }
-                    mix(acc);
-                    mix(f.len() as u64);
-                }
-                None => mix(u64::MAX - 1),
-            }
-            k
-        };
+        let key = Self::resident_wire_key(block, bg, anno_scale_override, frozen_layers);
         {
             let sets = self.resident_wire_sets.borrow();
             if let Some(set) = sets.get(&key) {
@@ -3331,6 +3487,36 @@ impl Scene {
         arc
     }
 
+    fn resident_wire_key(
+        block: Handle,
+        bg: [f32; 4],
+        anno_scale_override: Option<f32>,
+        frozen_layers: Option<&HashSet<Handle>>,
+    ) -> u64 {
+        let mut key: u64 = 0xcbf2_9ce4_8422_2325;
+        let mut mix =
+            |value: u64| key = key.rotate_left(17) ^ value.wrapping_mul(0x9E37_79B9_7F4A_7C15);
+        mix(block.value());
+        for component in bg {
+            mix(component.to_bits() as u64);
+        }
+        mix(anno_scale_override
+            .map(|scale| scale.to_bits() as u64)
+            .unwrap_or(u64::MAX));
+        match frozen_layers {
+            Some(frozen) => {
+                let mut signature = 0u64;
+                for handle in frozen {
+                    signature ^= handle.value().wrapping_mul(0x9E37_79B9_7F4A_7C15);
+                }
+                mix(signature);
+                mix(frozen.len() as u64);
+            }
+            None => mix(u64::MAX - 1),
+        }
+        key
+    }
+
     /// The GPU wire-arena handoff for a viewport whose content id is `gen`:
     /// `(prev_gen, changed handles)` when the Model set reached `gen` via an
     /// incremental resident patch, else `None`. Read (not consumed) so every
@@ -6375,6 +6561,253 @@ impl Scene {
         })
     }
 
+    fn rebuild_dependency_index(&self) -> SceneDependencyIndex {
+        let layout_blocks: HashSet<Handle> = self
+            .document
+            .objects
+            .values()
+            .filter_map(|object| match object {
+                acadrust::objects::ObjectType::Layout(layout) if !layout.block_record.is_null() => {
+                    Some(layout.block_record)
+                }
+                _ => None,
+            })
+            .collect();
+        let block_names: HashMap<Handle, String> = self
+            .document
+            .block_records
+            .iter()
+            .map(|record| (record.handle, record.name.to_ascii_uppercase()))
+            .collect();
+        let membership: HashMap<Handle, Handle> = self
+            .document
+            .block_records
+            .iter()
+            .flat_map(|record| {
+                record
+                    .entity_handles
+                    .iter()
+                    .copied()
+                    .map(move |handle| (handle, record.handle))
+            })
+            .collect();
+
+        let mut roots: HashMap<String, HashSet<Handle>> = HashMap::default();
+        let mut parents: HashMap<String, HashSet<String>> = HashMap::default();
+        for entity in self.document.entities() {
+            let EntityType::Insert(insert) = entity else {
+                continue;
+            };
+            let target = insert.block_name.to_ascii_uppercase();
+            let common = &insert.common;
+            let owner = if common.owner_handle.is_null() {
+                membership
+                    .get(&common.handle)
+                    .copied()
+                    .unwrap_or(Handle::NULL)
+            } else {
+                common.owner_handle
+            };
+            if owner.is_null() || layout_blocks.contains(&owner) {
+                roots.entry(target).or_default().insert(common.handle);
+            } else if let Some(parent) = block_names.get(&owner) {
+                parents.entry(target).or_default().insert(parent.clone());
+            }
+        }
+        // Propagate top-level INSERT users through nested block references.
+        // Fixed-point form is cycle-safe and block graphs are normally shallow.
+        let mut changed = true;
+        while changed {
+            changed = false;
+            for (child, parent_names) in &parents {
+                let inherited: Vec<Handle> = parent_names
+                    .iter()
+                    .flat_map(|parent| roots.get(parent).into_iter().flatten().copied())
+                    .collect();
+                let entry = roots.entry(child.clone()).or_default();
+                let before = entry.len();
+                entry.extend(inherited);
+                changed |= entry.len() != before;
+            }
+        }
+
+        let mut index = SceneDependencyIndex {
+            blocks: roots.clone(),
+            ..SceneDependencyIndex::default()
+        };
+        for entity in self.document.entities() {
+            let common = entity.common();
+            let owner = if common.owner_handle.is_null() {
+                membership
+                    .get(&common.handle)
+                    .copied()
+                    .unwrap_or(Handle::NULL)
+            } else {
+                common.owner_handle
+            };
+            let inside_block = !owner.is_null() && !layout_blocks.contains(&owner);
+            let render_handles: HashSet<Handle> = if inside_block {
+                block_names
+                    .get(&owner)
+                    .and_then(|name| roots.get(name))
+                    .cloned()
+                    .unwrap_or_default()
+            } else {
+                std::iter::once(common.handle).collect()
+            };
+            let add = |map: &mut HashMap<String, DependencyTargets>, name: &str| {
+                let target = map.entry(name.to_ascii_uppercase()).or_default();
+                target.render_handles.extend(render_handles.iter().copied());
+                target.source_handles.insert(common.handle);
+                target.touches_block_definition |= inside_block;
+            };
+            let add_handle = |map: &mut HashMap<Handle, DependencyTargets>,
+                              handle: Option<Handle>| {
+                let Some(handle) = handle.filter(|handle| !handle.is_null()) else {
+                    return;
+                };
+                let target = map.entry(handle).or_default();
+                target.render_handles.extend(render_handles.iter().copied());
+                target.source_handles.insert(common.handle);
+                target.touches_block_definition |= inside_block;
+            };
+            add(&mut index.layers, &common.layer);
+            match entity {
+                EntityType::Text(text) => add(&mut index.text_styles, &text.style),
+                EntityType::MText(text) => add(&mut index.text_styles, &text.style),
+                EntityType::AttributeDefinition(attribute) => {
+                    add(&mut index.text_styles, &attribute.text_style)
+                }
+                EntityType::AttributeEntity(attribute) => {
+                    add(&mut index.text_styles, &attribute.text_style)
+                }
+                EntityType::Insert(insert) => {
+                    for attribute in &insert.attributes {
+                        add(&mut index.text_styles, &attribute.text_style);
+                    }
+                }
+                EntityType::Dimension(dimension) => {
+                    add(&mut index.dim_styles, &dimension.base().style_name)
+                }
+                EntityType::Table(table) => {
+                    add_handle(&mut index.object_styles, table.table_style_handle)
+                }
+                EntityType::MultiLeader(leader) => {
+                    add_handle(&mut index.object_styles, leader.style_handle)
+                }
+                EntityType::MLine(line) => add_handle(&mut index.object_styles, line.style_handle),
+                _ => {}
+            }
+        }
+        index
+    }
+
+    fn dependency_targets(&self, kind: &str, names: &[String]) -> DependencyTargets {
+        if self.dependency_index_cache.borrow().is_none() {
+            *self.dependency_index_cache.borrow_mut() = Some(self.rebuild_dependency_index());
+        }
+        let cache = self.dependency_index_cache.borrow();
+        let index = cache.as_ref().unwrap();
+        let map = match kind {
+            "layer" => &index.layers,
+            "text" => &index.text_styles,
+            "dim" => &index.dim_styles,
+            _ => unreachable!(),
+        };
+        let mut combined = DependencyTargets::default();
+        for name in names {
+            let Some(target) = map.get(&name.to_ascii_uppercase()) else {
+                continue;
+            };
+            combined
+                .render_handles
+                .extend(target.render_handles.iter().copied());
+            combined
+                .source_handles
+                .extend(target.source_handles.iter().copied());
+            combined.touches_block_definition |= target.touches_block_definition;
+        }
+        combined
+    }
+
+    fn invalidate_dependency_targets(&mut self, targets: DependencyTargets) {
+        if targets.render_handles.is_empty() {
+            return;
+        }
+        if targets.touches_block_definition {
+            self.block_epoch = GEOMETRY_EPOCH.fetch_add(1, Ordering::Relaxed);
+        }
+        let sources: Vec<Handle> = targets.source_handles.iter().copied().collect();
+        self.recolor_meshes_for_handles(&sources);
+        let changes: Vec<(Handle, ChangeKind)> = targets
+            .render_handles
+            .into_iter()
+            .map(|handle| (handle, ChangeKind::Modified))
+            .collect();
+        self.bump_entities(&changes);
+    }
+
+    pub fn invalidate_layer_dependencies(&mut self, names: &[String]) {
+        let targets = self.dependency_targets("layer", names);
+        self.invalidate_dependency_targets(targets);
+    }
+
+    pub fn invalidate_text_style_dependencies(&mut self, name: &str) {
+        self.invalidate_text_style_dependencies_many(&[name.to_string()]);
+    }
+
+    pub fn invalidate_text_style_dependencies_many(&mut self, names: &[String]) {
+        let targets = self.dependency_targets("text", names);
+        self.invalidate_dependency_targets(targets);
+    }
+
+    pub fn invalidate_dim_style_dependencies(&mut self, name: &str) {
+        self.invalidate_dim_style_dependencies_many(&[name.to_string()]);
+    }
+
+    pub fn invalidate_dim_style_dependencies_many(&mut self, names: &[String]) {
+        let targets = self.dependency_targets("dim", names);
+        self.invalidate_dependency_targets(targets);
+    }
+
+    pub fn invalidate_object_style_dependencies(&mut self, handles: &[Handle]) {
+        if self.dependency_index_cache.borrow().is_none() {
+            *self.dependency_index_cache.borrow_mut() = Some(self.rebuild_dependency_index());
+        }
+        let mut combined = DependencyTargets::default();
+        if let Some(index) = self.dependency_index_cache.borrow().as_ref() {
+            for handle in handles {
+                let Some(target) = index.object_styles.get(handle) else {
+                    continue;
+                };
+                combined
+                    .render_handles
+                    .extend(target.render_handles.iter().copied());
+                combined
+                    .source_handles
+                    .extend(target.source_handles.iter().copied());
+                combined.touches_block_definition |= target.touches_block_definition;
+            }
+        }
+        self.invalidate_dependency_targets(combined);
+    }
+
+    pub fn block_dependency_handles(&self, name: &str) -> Vec<Handle> {
+        if self.dependency_index_cache.borrow().is_none() {
+            *self.dependency_index_cache.borrow_mut() = Some(self.rebuild_dependency_index());
+        }
+        self.dependency_index_cache
+            .borrow()
+            .as_ref()
+            .and_then(|index| index.blocks.get(&name.to_ascii_uppercase()))
+            .map(|handles| handles.iter().copied().collect())
+            .unwrap_or_default()
+    }
+
+    pub(crate) fn invalidate_dependency_index(&self) {
+        self.dependency_index_cache.borrow_mut().take();
+    }
+
     /// Spatial index + always-emit list for top-level entities. Lazily
     /// rebuilt on `geometry_epoch` change.
     ///

src/scene/pipeline/mod.rs

@@ -47,7 +47,7 @@ pub struct Pipeline {
     /// Used to draw ghost copies of selected wires through occluding geometry.
     wire_xray_pipeline: wgpu::RenderPipeline,
     /// Layout for the per-wire `WireConst` storage buffer (group 1 of the wire /
-    /// xray pipelines). `Some` only on the fast native path; `None` in packed
+    /// xray pipelines). `Some` on native/WebGPU fast paths; `None` in packed
     /// compatibility mode. Passed to `WireGpu::from_run` / `from_batch`.
     pub(crate) wire_const_bgl: Option<wgpu::BindGroupLayout>,
     wipeout_pipeline: wgpu::RenderPipeline,
@@ -338,8 +338,8 @@ impl Pipeline {
         });
 
         // ── Wire pipeline ──────────────────────────────────────────────────
-        // Select once per device. The fast native path hoists shared constants
-        // into storage; compatibility mode keeps them in 10 packed attributes.
+        // Select once per device. Native and WebGPU hoist shared constants into
+        // storage when limits allow; WebGL2/compat keeps packed attributes.
         let wire_mode = wire_gpu::WirePipelineMode::select(device);
         let renderer_mode_name =
             if wire_mode.uses_storage() { "fast-storage" } else { "packed-compat" };
@@ -357,12 +357,9 @@ impl Pipeline {
             renderer_mode_name,
             device.limits().max_storage_buffers_per_shader_stage
         );
-        #[cfg(not(target_arch = "wasm32"))]
         let wire_const_bgl = wire_mode
             .uses_storage()
             .then(|| wire_gpu::WireConst::bind_group_layout(device));
-        #[cfg(target_arch = "wasm32")]
-        let wire_const_bgl: Option<wgpu::BindGroupLayout> = None;
         let mut wire_bgls: Vec<&wgpu::BindGroupLayout> = vec![&frame_bgl];
         if let Some(bgl) = &wire_const_bgl {
             wire_bgls.push(bgl);
@@ -379,7 +376,6 @@ impl Pipeline {
         let wire_shader = device.create_shader_module(wgpu::ShaderModuleDescriptor {
             label: Some("wire.shader"),
             source: wgpu::ShaderSource::Wgsl(std::borrow::Cow::Borrowed(match wire_mode {
-                #[cfg(not(target_arch = "wasm32"))]
                 wire_gpu::WirePipelineMode::IndexedStorage => {
                     include_str!("../../shaders/wire_indexed.wgsl")
                 }
@@ -3157,6 +3153,14 @@ fn create_msaa_texture(
 pub struct MultiPipeline {
     pub(crate) inners: Vec<Pipeline>,
     format: wgpu::TextureFormat,
+    /// Stable viewport identity → pipeline slot map. Paper viewports used to
+    /// occupy slots by their current list position, so switching layouts or
+    /// scrolling a sheet could assign an existing viewport to a different
+    /// slot and throw away all of its GPU caches. Keep the association across
+    /// tab switches and only recycle genuinely cold slots.
+    pub(crate) slot_by_instance: rustc_hash::FxHashMap<u64, usize>,
+    slot_last_used: Vec<u64>,
+    slot_clock: u64,
     /// The resident wire batches, keyed by `wire_content_id` and shared across
     /// every slot (and every pane — one `MultiPipeline` backs all of them) that
     /// renders the same content. `prepare` builds an entry once on a cache miss
@@ -3189,8 +3193,70 @@ impl MultiPipeline {
         let n = n.max(1);
         while self.inners.len() < n {
             self.inners.push(Pipeline::new(device, queue, self.format));
+            self.slot_last_used.push(0);
         }
     }
+
+    /// Resolve stable slots for the viewport identities in one primitive.
+    /// Thirty-two hot slots cover ordinary tiled/paper drawings. A cold slot
+    /// is recycled only after several other prepare calls, which prevents
+    /// sibling Model panes prepared in the same frame from evicting each
+    /// other. If every slot is still hot, growing is safer than a visible
+    /// rebuild hitch.
+    pub(crate) fn resolve_slots(
+        &mut self,
+        device: &wgpu::Device,
+        queue: &wgpu::Queue,
+        instance_ids: &[u64],
+    ) -> Vec<usize> {
+        const SOFT_LIMIT: usize = 32;
+        const HOT_WINDOW: u64 = 8;
+
+        self.slot_clock = self.slot_clock.wrapping_add(1).max(1);
+        let now = self.slot_clock;
+        let reserved: rustc_hash::FxHashSet<u64> = instance_ids.iter().copied().collect();
+        let mut slots = Vec::with_capacity(instance_ids.len());
+
+        for &instance_id in instance_ids {
+            let slot = if let Some(&slot) = self.slot_by_instance.get(&instance_id) {
+                slot
+            } else {
+                let vacant = self
+                    .inners
+                    .iter()
+                    .position(|inner| inner.slot_id == u64::MAX);
+                let recyclable = vacant.or_else(|| {
+                    (self.inners.len() >= SOFT_LIMIT)
+                        .then(|| {
+                            self.inners
+                                .iter()
+                                .enumerate()
+                                .filter(|(_, inner)| !reserved.contains(&inner.slot_id))
+                                .filter(|(slot, _)| {
+                                    now.saturating_sub(self.slot_last_used[*slot]) > HOT_WINDOW
+                                })
+                                .min_by_key(|(slot, _)| self.slot_last_used[*slot])
+                                .map(|(slot, _)| slot)
+                        })
+                        .flatten()
+                });
+                let slot = recyclable.unwrap_or_else(|| {
+                    let slot = self.inners.len();
+                    self.ensure_len(device, queue, slot + 1);
+                    slot
+                });
+                let old_id = self.inners[slot].slot_id;
+                if old_id != u64::MAX {
+                    self.slot_by_instance.remove(&old_id);
+                }
+                self.slot_by_instance.insert(instance_id, slot);
+                slot
+            };
+            self.slot_last_used[slot] = now;
+            slots.push(slot);
+        }
+        slots
+    }
 }
 
 /// Send wgpu's uncaptured validation errors to stderr instead of the default
@@ -3227,6 +3293,9 @@ impl iced::widget::shader::Pipeline for MultiPipeline {
         Self {
             inners: vec![Pipeline::new(device, queue, format)],
             format,
+            slot_by_instance: rustc_hash::FxHashMap::default(),
+            slot_last_used: vec![0],
+            slot_clock: 0,
             wire_buffer_cache: rustc_hash::FxHashMap::default(),
         }
     }

src/scene/pipeline/wire_gpu.rs

@@ -58,18 +58,17 @@ fn instance_buffer_mapped<T: bytemuck::Pod>(
 
 // ── Instance layout ───────────────────────────────────────────────────────
 
-// ── Native: slim per-segment instance + shared per-wire constants ───────────
+// ── WebGPU/native: slim per-segment instance + shared constants ─────────────
 //
 // Every segment of a wire used to carry the wire's color / line-weight / dash
 // pattern / draw-depth (~44 B) on each instance — re-fetched once per segment
-// even though it's constant along the wire. On native we hoist those into a
+// even though it's constant along the wire. On storage-capable adapters we hoist those into a
 // per-wire `WireConst` storage buffer indexed by `wire_id`, so the instance
 // keeps only the per-segment data (endpoints + arc-length distances). Cuts the
 // instance from 104 B to one 64-byte cache line and removes the redundant
 // per-segment re-fetch of the shared constants. WebGL2 has no vertex-stage
 // storage buffers, so the wasm build below keeps the original self-contained
 // fat instance.
-#[cfg(not(target_arch = "wasm32"))]
 #[repr(C)]
 #[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
 pub struct WireInstance {
@@ -87,8 +86,6 @@ pub struct WireInstance {
     /// exactly one 64-byte cache line.
     pub taper_ratio: [u16; 2],
 }
-
-#[cfg(not(target_arch = "wasm32"))]
 impl WireInstance {
     pub fn layout<'a>() -> wgpu::VertexBufferLayout<'a> {
         // Must match `InstanceIn` in wire_indexed.wgsl.
@@ -110,11 +107,10 @@ impl WireInstance {
     }
 }
 
-/// Per-wire constants shared by every segment of a wire (native only). std430
+/// Per-wire constants shared by every segment of a wire. std430
 /// layout: three vec4 then eight scalars = 80 B, matching `WireConst` in
 /// wire_indexed.wgsl. `align_end` / `align_total` carry the "A"-type endpoint
 /// alignment (see `wire_distances`); 0.0 total = no alignment.
-#[cfg(not(target_arch = "wasm32"))]
 #[repr(C)]
 #[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
 pub struct WireConst {
@@ -134,8 +130,6 @@ pub struct WireConst {
     pub _pad1: f32,
     pub _pad2: f32,
 }
-
-#[cfg(not(target_arch = "wasm32"))]
 impl WireConst {
     /// Bind-group layout for the per-wire storage buffer (group 1 of the wire /
     /// xray pipelines). Read-only storage, visible to the vertex stage.
@@ -158,8 +152,8 @@ impl WireConst {
 
 // ── Packed compatibility instance (no vertex-stage storage) ────────────────
 //
-// Web always uses this layout. Native selects it at runtime for adapters whose
-// storage-buffer limits are insufficient, or when --compat-renderer is set.
+// WebGL2 and limited native adapters use this layout. WebGPU selects the
+// indexed storage path when its reported limits satisfy the same requirement.
 #[repr(C)]
 #[derive(Copy, Clone, bytemuck::Pod, bytemuck::Zeroable)]
 pub struct PackedWireInstance {
@@ -224,21 +218,15 @@ impl PackedWireInstance {
     }
 }
 
-#[cfg(target_arch = "wasm32")]
-pub type WireInstance = PackedWireInstance;
-
 /// Wire and hatch pipelines switch together: the fast path uses storage
 /// buffers; the compatibility path carries wire constants in packed vertex
 /// attributes and hatch data in a texture.
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
 pub enum WirePipelineMode {
-    #[cfg(not(target_arch = "wasm32"))]
     IndexedStorage,
     Packed,
 }
-
-#[cfg(not(target_arch = "wasm32"))]
-fn select_native_pipeline(max_storage_buffers_per_stage: u32, forced: bool) -> WirePipelineMode {
+fn select_pipeline(max_storage_buffers_per_stage: u32, forced: bool) -> WirePipelineMode {
     const REQUIRED_STORAGE_BUFFERS_PER_STAGE: u32 = 5;
     if forced || max_storage_buffers_per_stage < REQUIRED_STORAGE_BUFFERS_PER_STAGE {
         WirePipelineMode::Packed
@@ -249,23 +237,15 @@ fn select_native_pipeline(max_storage_buffers_per_stage: u32, forced: bool) -> W
 
 impl WirePipelineMode {
     pub fn select(device: &wgpu::Device) -> Self {
-        #[cfg(target_arch = "wasm32")]
-        {
-            let _ = device;
-            Self::Packed
-        }
         #[cfg(not(target_arch = "wasm32"))]
-        {
-            select_native_pipeline(
-                device.limits().max_storage_buffers_per_shader_stage,
-                crate::cli::gui_config().compat_renderer,
-            )
-        }
+        let forced = crate::cli::gui_config().compat_renderer;
+        #[cfg(target_arch = "wasm32")]
+        let forced = false;
+        select_pipeline(device.limits().max_storage_buffers_per_shader_stage, forced)
     }
 
     pub fn uses_storage(self) -> bool {
         match self {
-            #[cfg(not(target_arch = "wasm32"))]
             Self::IndexedStorage => true,
             Self::Packed => false,
         }
@@ -273,7 +253,6 @@ impl WirePipelineMode {
 
     pub fn layout<'a>(self) -> wgpu::VertexBufferLayout<'a> {
         match self {
-            #[cfg(not(target_arch = "wasm32"))]
             Self::IndexedStorage => WireInstance::layout(),
             Self::Packed => PackedWireInstance::layout(),
         }
@@ -494,9 +473,8 @@ fn emit_wire_packed(
     instances
 }
 
-/// Native: emit slim per-segment instances (positions + distances + `wire_id`)
+/// Storage path: emit slim per-segment instances (positions + distances + `wire_id`)
 /// plus the one `WireConst` record every segment of this wire shares.
-#[cfg(not(target_arch = "wasm32"))]
 pub(crate) fn emit_wire_native(
     wire: &WireModel,
     wire_id: u32,
@@ -579,11 +557,10 @@ pub(crate) fn wire_draw_depth(
     }
 }
 
-/// Build the shared per-wire `WireConst` storage buffer and its bind group
-/// (native only). All instance-buffer chunks from one build reference the same
+/// Build the shared per-wire `WireConst` storage buffer and its bind group.
+/// All instance-buffer chunks from one build reference the same
 /// buffer via their global `wire_id`, so a single bind group is cloned into
 /// each chunk.
-#[cfg(not(target_arch = "wasm32"))]
 fn build_const_bind_group(
     device: &wgpu::Device,
     bgl: &wgpu::BindGroupLayout,
@@ -677,7 +654,6 @@ impl WireGpu {
         mesh_edge: bool,
         const_bgl: Option<&wgpu::BindGroupLayout>,
     ) -> Vec<Self> {
-        #[cfg(not(target_arch = "wasm32"))]
         if let Some(const_bgl) = const_bgl {
             const MAX_INSTANCES: usize =
                 268_435_456 / std::mem::size_of::<WireInstance>();
@@ -769,7 +745,6 @@ impl WireGpu {
         if total_segs == 0 {
             return vec![];
         }
-        #[cfg(not(target_arch = "wasm32"))]
         if let Some(const_bgl) = const_bgl {
             // GPU max buffer size is 256 MB; chunk to stay within the limit.
             const MAX_INSTANCES: usize =

src/scene/selection.rs

@@ -391,7 +391,10 @@ impl Scene {
         // Report the exact erased handles so derived caches drop just those and
         // the resident set removes only their wires (bump_entities drops them
         // from the tessellation memos too).
-        self.bump_entities(&erased);
+        if !erased.is_empty() {
+            self.invalidate_dependency_index();
+            self.bump_entities(&erased);
+        }
     }
 
     /// Restore erased Arc-backed entities without re-linking their still-present
@@ -414,6 +417,7 @@ impl Scene {
             }
         }
         if !changes.is_empty() {
+            self.invalidate_dependency_index();
             self.bump_entities(&changes);
         }
         restored

src/scene/view/render.rs

@@ -152,13 +152,6 @@ pub struct Primitive {
     pub(in crate::scene) viewports: Vec<ViewportData>,
     /// Background color used to clear each viewport's MSAA buffer.
     pub(in crate::scene) bg_color: [f32; 4],
-    /// First `MultiPipeline` inner slot this primitive owns. Paper space (one
-    /// shader widget, many viewports) uses 0. Per-pane Model widgets each own a
-    /// distinct slot (= their tile index) so several shader widgets can share
-    /// the type-keyed pipeline storage without clobbering one another — all
-    /// `prepare` calls run before all `render` calls, so disjoint slots are
-    /// safe.
-    pub(in crate::scene) base_slot: usize,
     /// One input-to-render sample, carried only when PERF tracing is enabled.
     pub(in crate::scene) nav_perf: Option<NavPerfSample>,
 }
@@ -250,10 +243,11 @@ impl shader::Primitive for Primitive {
         let phys = viewport.physical_size();
         let full_size = Size::new(phys.width, phys.height);
         let scale = viewport.scale_factor() as f32;
-        pipeline.ensure_len(device, queue, self.base_slot + self.viewports.len());
+        let instance_ids: Vec<u64> = self.viewports.iter().map(|vp| vp.instance_id).collect();
+        let slots = pipeline.resolve_slots(device, queue, &instance_ids);
 
         for (i, vp) in self.viewports.iter().enumerate() {
-            let inner = &mut pipeline.inners[self.base_slot + i];
+            let inner = &mut pipeline.inners[slots[i]];
             // Pipeline slots are addressed by list index, but off-canvas
             // viewports are dropped from the list — so a slot can be reused by a
             // DIFFERENT viewport across frames (e.g. the first viewport scrolls
@@ -442,7 +436,10 @@ impl shader::Primitive for Primitive {
                 // the whole wire buffer. Only for the scissor-free, mesh-free
                 // (single-batch) Model set; scissored paper viewports and mixed
                 // 2D/3D sets fall through to the shared batched path below.
+                #[cfg(not(target_arch = "wasm32"))]
                 let mut arena_served = false;
+                #[cfg(target_arch = "wasm32")]
+                let arena_served = false;
                 #[cfg(not(target_arch = "wasm32"))]
                 let _perf = crate::perf::enabled();
                 #[cfg(not(target_arch = "wasm32"))]
@@ -895,9 +892,12 @@ impl shader::Primitive for Primitive {
         let ch = clip.height as f32;
         let clip_right = clip.x + clip.width;
         let clip_bottom = clip.y + clip.height;
-        for (i, vp) in self.viewports.iter().enumerate() {
-            let Some(inner) = pipeline.inners.get(self.base_slot + i) else {
-                break;
+        for vp in &self.viewports {
+            let Some(slot) = pipeline.slot_by_instance.get(&vp.instance_id) else {
+                continue;
+            };
+            let Some(inner) = pipeline.inners.get(*slot) else {
+                continue;
             };
             // Where the viewport would land on the surface in absolute
             // pixels (i32 because either edge may stick off the canvas).
@@ -1433,14 +1433,9 @@ impl Scene {
             sample.build_ms = nav_build_started.elapsed().as_secs_f64() * 1000.0;
             sample
         });
-        // Model panes permanently own slots 0..N. Paper starts after them so
-        // its sheet/content viewports never evict the Model slot's wire arena,
-        // textures, mesh batches and render cache during a layout-tab switch.
-        let base_slot = self.model_tiles.borrow().len();
         Primitive {
             viewports,
             bg_color,
-            base_slot,
             nav_perf: perf_nav,
         }
     }
@@ -1466,7 +1461,6 @@ impl Scene {
             return Primitive {
                 viewports: vec![],
                 bg_color,
-                base_slot: tile_idx,
                 nav_perf: None,
             };
         };
@@ -1514,7 +1508,6 @@ impl Scene {
         Primitive {
             viewports,
             bg_color,
-            base_slot: tile_idx,
             nav_perf: perf_nav,
         }
     }

src/ui/window/open_progress.rs

@@ -1,7 +1,7 @@
 //! Modal overlay shown while a CAD file is being loaded.
 //!
-//! Displays the file name, size, current phase, an indeterminate animated
-//! progress bar, and a Cancel button.
+//! Displays the file name, size, current phase, measured progress, and a
+//! Cancel button.
 
 use iced::time::Instant;
 use iced::widget::{button, column, container, row, stack, text, Space};
@@ -10,20 +10,18 @@ use std::sync::atomic::Ordering;
 
 use crate::app::{
     Message, OpenProgress, OPEN_PHASE_CACHING, OPEN_PHASE_FINALIZING, OPEN_PHASE_PARSING,
-    OPEN_PHASE_READING,
+    OPEN_PHASE_READING, OPEN_PHASE_XREF,
 };
 
 const CARD_WIDTH: f32 = 420.0;
 const BAR_TRACK_WIDTH: f32 = 380.0;
 const BAR_TRACK_HEIGHT: f32 = 6.0;
-const BAR_WINDOW_WIDTH: f32 = 100.0;
-/// Period of one back-and-forth bounce, in milliseconds.
-const BAR_PERIOD_MS: f32 = 1800.0;
 
 fn phase_label(phase: u8) -> &'static str {
     match phase {
         OPEN_PHASE_READING => "Reading file…",
         OPEN_PHASE_PARSING => "Parsing entities…",
+        OPEN_PHASE_XREF => "Loading references…",
         OPEN_PHASE_CACHING => "Building scene caches…",
         OPEN_PHASE_FINALIZING => "Finalizing…",
         _ => "Working…",
@@ -46,29 +44,22 @@ fn format_size(bytes: u64) -> String {
     }
 }
 
-/// Compute the left-offset of the moving highlight inside the track.
-/// Bounces left↔right using a triangle wave so the user sees motion even when
-/// the actual phase atomic stays put for a while.
-fn bar_offset(elapsed_ms: f32) -> f32 {
-    let travel = (BAR_TRACK_WIDTH - BAR_WINDOW_WIDTH).max(0.0);
-    let cycle = (elapsed_ms / BAR_PERIOD_MS).fract();
-    let tri = if cycle < 0.5 {
-        cycle * 2.0
-    } else {
-        (1.0 - cycle) * 2.0
-    };
-    tri * travel
-}
+pub fn view<'a>(progress: &'a OpenProgress, _now: Instant) -> Element<'a, Message> {
+    let phase = progress.state.phase.load(Ordering::Acquire);
+    let basis_points = progress
+        .state
+        .basis_points
+        .load(Ordering::Relaxed)
+        .min(10000);
+    let fraction = basis_points as f32 / 10000.0;
+    let fill_width = BAR_TRACK_WIDTH * fraction;
+    let trailing = (BAR_TRACK_WIDTH - fill_width).max(0.0);
 
-pub fn view<'a>(progress: &'a OpenProgress, now: Instant) -> Element<'a, Message> {
-    let phase = progress.phase.load(Ordering::Relaxed);
-    let elapsed_ms = now.saturating_duration_since(progress.started).as_millis() as f32;
-
-    // ── Animated indeterminate bar ────────────────────────────────────────
-    let offset = bar_offset(elapsed_ms);
-    let trailing = (BAR_TRACK_WIDTH - BAR_WINDOW_WIDTH - offset).max(0.0);
-
-    let bar_window: Element<'_, Message> = container(Space::new().width(Length::Fixed(BAR_WINDOW_WIDTH)).height(Length::Fixed(BAR_TRACK_HEIGHT)))
+    let bar_fill: Element<'_, Message> = container(
+        Space::new()
+            .width(Length::Fixed(fill_width))
+            .height(Length::Fixed(BAR_TRACK_HEIGHT)),
+    )
         .style(|_: &Theme| container::Style {
             background: Some(Background::Color(Color {
                 r: 0.30,
@@ -84,11 +75,8 @@ pub fn view<'a>(progress: &'a OpenProgress, now: Instant) -> Element<'a, Message
         })
         .into();
 
-    let bar_moving: Element<'_, Message> = row![
-        Space::new()
-            .width(Length::Fixed(offset))
-            .height(Length::Fixed(BAR_TRACK_HEIGHT)),
-        bar_window,
+    let bar_value: Element<'_, Message> = row![
+        bar_fill,
         Space::new()
             .width(Length::Fixed(trailing))
             .height(Length::Fixed(BAR_TRACK_HEIGHT)),
@@ -111,7 +99,7 @@ pub fn view<'a>(progress: &'a OpenProgress, now: Instant) -> Element<'a, Message
                     },
                     ..Default::default()
                 }),
-            bar_moving,
+            bar_value,
         ]
         .width(Length::Fixed(BAR_TRACK_WIDTH))
         .height(Length::Fixed(BAR_TRACK_HEIGHT)),
@@ -136,7 +124,11 @@ pub fn view<'a>(progress: &'a OpenProgress, now: Instant) -> Element<'a, Message
         a: 1.0,
     });
 
-    let phase_line = text(phase_label(phase))
+    let phase_line = text(format!(
+        "{}  {:.1}%",
+        phase_label(phase),
+        basis_points as f32 / 100.0
+    ))
         .size(12)
         .color(Color {
             r: 0.70,

web/ocs-parse-worker.js

@@ -0,0 +1,20 @@
+import init, { parse_document } from "./worker_pkg/ocs_web_worker.js";
+
+const ready = init();
+
+self.onmessage = async ({ data }) => {
+  try {
+    await ready;
+    const encoded = parse_document(data.name, new Uint8Array(data.bytes));
+    // wasm-bindgen returns a view into WebAssembly.Memory. Copy to a standalone
+    // ArrayBuffer before transferring it, otherwise the worker's wasm memory
+    // itself would be detached.
+    const transferable = encoded.slice();
+    self.postMessage({ ok: true, data: transferable.buffer }, [transferable.buffer]);
+  } catch (error) {
+    self.postMessage({
+      ok: false,
+      error: error instanceof Error ? error.message : String(error),
+    });
+  }
+};