First attempt at a GPU runner

Currently failing with an error I don't understand:

```
wgpu error: Validation Error

Caused by:
  In Device::create_shader_module, label = '...\image_binary.spv'

Shader '...\image_binary.spv' parsing error: InvalidTypeWidth(1)
```

.gitea/workflows/ci.yml

@@ -7,8 +7,6 @@ jobs:
   formatting:
     name: cargo fmt
     runs-on: ubuntu-latest
-    env:
-      RUNNER_TOOL_CACHE: /toolcache
     steps:
       - uses: actions/checkout@v6
       - uses: actions-rust-lang/setup-rust-toolchain@v1
@@ -20,21 +18,20 @@ jobs:
   test:
     name: cargo test
     runs-on: ubuntu-latest
-    env:
-      RUNNER_TOOL_CACHE: /toolcache
     steps:
       - uses: actions/checkout@v6
       - uses: actions-rust-lang/setup-rust-toolchain@v1
-      - run: cargo check
+        with:
+          components: clippy
+      - run: cargo check --all-targets
+      - run: cargo clippy --all-targets
       - run: cargo test
 
   test-gpu:
     name: cargo test (GPU)
     runs-on: ubuntu-latest
-    env:
-      RUNNER_TOOL_CACHE: /toolcache
     steps:
       - uses: actions/checkout@v6
       - uses: actions-rust-lang/setup-rust-toolchain@v1
-      - run: cargo install --git https://github.com/rust-gpu/rust-gpu cargo-gpu
-      - run: cargo gpu check -p enkou-shaders
+      - run: cargo install --git https://github.com/rust-gpu/rust-gpu cargo-gpu --rev 67f1ff2
+      - run: cargo gpu check --auto-install-rust-toolchain -p enkou-shaders

Cargo.toml

@@ -1,7 +1,8 @@
 [workspace]
 members = [
     "enkou-shaders",
-    "enkou-shaders-tests",
+    "examples/image-runner",
+    "examples/image-binary",
 ]
 resolver = "3"
 
@@ -13,6 +14,7 @@ license = "MIT"
 repository = ""
 
 [workspace.lints.rust]
+missing_docs = { level = "warn" }
 unexpected_cfgs = { level = "allow", check-cfg = ['cfg(target_arch, values("spirv"))'] }
 
 [workspace.dependencies]
@@ -21,6 +23,14 @@ spirv-std = { git = "https://github.com/Rust-GPU/rust-gpu.git", rev = "67f1ff2"
 
 anyhow = "1.0.102"
 bytemuck = { version = "1.25.0", features = ["derive"] }
-glam = { version = "0.33.1", default-features = false, features = ["libm"] }
+futures = "0.3.32"
+glam = { version = "0.33.1", default-features = false, features = ["bytemuck", "scalar-math"] }
+image = { version = "0.25.10", default-features = false, features = ["default-formats"]}
+libm = "0.2.16"
+rand = { version = "0.10.1", default-features = false }
+rand_xoshiro = "0.8.1"
 rspirv = "0.13.0"
-
+tempfile = "3.27.0"
+thiserror = "2.0.19"
+wgpu = { version = "30.0.0", features = ["spirv"] }
+xflags = "0.3.2"

enkou-shaders-tests/src/lib.rs

@@ -1,67 +0,0 @@
-#[cfg(test)]
-mod test {
-    use rspirv::binary::parse_bytes;
-    use rspirv::dr::{Module, Operand};
-    use rspirv::spirv::ExecutionModel;
-    use std::sync::OnceLock;
-
-    static SHADER_MODULE: OnceLock<Module> = OnceLock::new();
-
-    fn shader() -> &'static Module {
-        SHADER_MODULE.get_or_init(|| {
-            let shader_bytes = include_bytes!(env!("SHADER_SPV_PATH"));
-
-            let mut loader = rspirv::dr::Loader::new();
-            parse_bytes(shader_bytes, &mut loader).expect("Unable to parse shader");
-
-            loader.module()
-        })
-    }
-
-    fn has_entry_point(execution_model: ExecutionModel, name: &str) -> bool {
-        for ref entry_point in shader().entry_points.iter() {
-            let operands: Vec<Operand> = entry_point
-                .operands
-                .iter()
-                .filter(|op| match op {
-                    Operand::ExecutionModel(_) | Operand::LiteralString(_) => true,
-                    _ => false,
-                })
-                .map(|op| op.clone())
-                .collect();
-
-            assert_eq!(operands.len(), 2);
-            match &operands[0] {
-                Operand::ExecutionModel(actual) => {
-                    if execution_model != *actual {
-                        continue;
-                    }
-                }
-                op => panic!("Unexpected operand; {}", op),
-            }
-
-            match &operands[1] {
-                Operand::LiteralString(actual) => {
-                    if name != actual {
-                        continue;
-                    }
-                }
-                op => panic!("Unexpected operand; {}", op),
-            }
-
-            return true;
-        }
-
-        false
-    }
-
-    #[test]
-    pub fn has_entry_main_fs() {
-        assert!(has_entry_point(ExecutionModel::Fragment, "main_fs"))
-    }
-
-    #[test]
-    pub fn has_entry_main_vs() {
-        assert!(has_entry_point(ExecutionModel::Vertex, "main_vs"))
-    }
-}

enkou-shaders/Cargo.toml

@@ -10,6 +10,9 @@ repository.workspace = true
 workspace = true
 
 [dependencies]
-spirv-std.workspace = true
-glam.workspace = true
 bytemuck.workspace = true
+glam.workspace = true
+libm.workspace = true
+rand.workspace = true
+rand_xoshiro.workspace = true
+spirv-std.workspace = true

enkou-shaders/src/camera.rs

@@ -0,0 +1,201 @@
+//! # Camera
+//!
+//! Map points from the IFS coordinate system to pixel coordinates. This is a lossy transformation.
+use bytemuck::{Pod, Zeroable};
+use glam::{Affine2, IVec2, UVec2, Vec2, vec2};
+use libm::powf;
+
+/// Settings used to map IFS coordinates to pixel coordinates.
+///
+/// The camera is itself an affine transformation, capable of zoom, rotation, and translation
+/// of the IFS coordinates before rendering to the final image.
+#[derive(Copy, Clone, Pod, Zeroable)]
+#[repr(C)]
+pub struct Camera {
+    dimensions: UVec2,
+    transform: Affine2,
+}
+
+impl Camera {
+    /// Construct a new camera for translating IFS coordinates to pixel coordinates.
+    ///
+    /// While the camera is implemented as a single affine transformation, it's helpful
+    /// to express the transform steps individually.
+    ///
+    /// # Arguments
+    ///
+    /// * `dimensions` - Width and height of the output image (in pixels).
+    /// * `center` - Location of the origin in IFS coordinates. Positive `x` shifts the image
+    ///   left, and positive `y` position shifts the image up.
+    /// * `rotate` - Rotation angle (in radians) of IFS coordinates. Rotation is applied after the
+    ///   `center` translation, so it is about the new origin.
+    /// * `zoom` - Zoom factor applied to IFS coordinates. IFS coordinates are scaled by
+    ///   `pow(2, zoom)`, so a zoom factor of 0 is the identity.
+    /// * `scale` - Pixels per unit of IFS coordinates. This parameter is usually chosen such
+    ///   that the largest dimension will cover the range `[-2, 2]`, but values higher or lower
+    ///   can be used as a secondary zoom.
+    pub fn new(dimensions: UVec2, center: Vec2, rotate: f32, zoom: Vec2, scale: Vec2) -> Camera {
+        let ifs_center_transform = Affine2::from_translation(-center);
+        let zoom_transform = Affine2::from_scale(vec2(powf(2.0, zoom.x), powf(2.0, zoom.y)));
+        let scale_transform = Affine2::from_scale(scale);
+        let rotate_transform = Affine2::from_angle(rotate);
+        let image_center_transform = Affine2::from_translation((dimensions / 2).as_vec2());
+
+        let transform = image_center_transform
+            * rotate_transform
+            * scale_transform
+            * zoom_transform
+            * ifs_center_transform;
+
+        Camera {
+            dimensions,
+            transform,
+        }
+    }
+
+    /// Map a point from IFS coordinates to pixel coordinates.
+    ///
+    /// ```
+    /// # use glam::{vec2, ivec2, uvec2, Vec2};
+    /// # use crate::enkou_shaders::camera::Camera;
+    /// // Output image is 600x600 pixels, centered at the origin, no rotation, no zoom,
+    /// // and scaled such that it covers the range [-2, 2].
+    /// // Use the origin as the IFS coordinate, so the pixel coordinate is the center of the image
+    /// let camera = Camera::new(
+    ///     uvec2(600, 600),
+    ///     Vec2::ZERO,
+    ///     0.0,
+    ///     Vec2::ZERO,
+    ///     vec2(150.0, 150.0)
+    /// );
+    /// assert_eq!(camera.transform_point(vec2(0.0, 0.0)), ivec2(300, 300));
+    /// ```
+    pub fn transform_point(&self, point: Vec2) -> IVec2 {
+        self.transform.transform_point2(point).as_ivec2()
+    }
+
+    /// Map a point from IFS coordinates to pixel coordinates (like [`transform_point`](Camera::transform_point)),
+    /// and check that the result is within the provided image dimensions.
+    pub fn transform_point_to_image(&self, point: Vec2) -> Option<UVec2> {
+        let pixel_coordinates = self.transform_point(point);
+        if pixel_coordinates.x < 0
+            || pixel_coordinates.y < 0
+            || (pixel_coordinates.x as u32) >= self.dimensions.x
+            || (pixel_coordinates.y as u32) >= self.dimensions.y
+        {
+            None
+        } else {
+            Some(pixel_coordinates.as_uvec2())
+        }
+    }
+}
+
+/// Shader entry point for running the camera transformation over a list of IFS coordinates
+pub mod entry {
+    use crate::camera::Camera;
+    use spirv_std::glam::{IVec2, Vec2};
+    use spirv_std::spirv;
+
+    /// Transform IFS coordinates to pixel coordinates
+    #[spirv(compute(entry_point_name = "main_camera", threads(1)))]
+    pub fn main_camera(
+        #[spirv(storage_buffer, descriptor_set = 0, binding = 0)] camera: &Camera,
+        #[spirv(storage_buffer, descriptor_set = 0, binding = 1)] coordinates_ifs: &[Vec2],
+        #[spirv(storage_buffer, descriptor_set = 1, binding = 0)] coordinates_pixel: &mut [IVec2],
+    ) {
+        for i in 0..coordinates_ifs.len() {
+            coordinates_pixel[i] = camera.transform_point(coordinates_ifs[i])
+        }
+    }
+}
+
+#[cfg(test)]
+mod test {
+    use crate::camera::Camera;
+    use glam::{Affine2, Vec2, ivec2, uvec2, vec2};
+    use libm::powf;
+
+    #[test]
+    pub fn manual_camera() {
+        let starting_point = vec2(1.0, 1.0);
+
+        // Move the origin; points move right and up by one unit, giving us (2.0, 2.0)
+        let center = vec2(-1.0, -1.0);
+        let point = starting_point - center;
+
+        // Rotate about the new origin; points move counter-clockwise, giving us (-2.0, 2.0)
+        let rotate = 90.0f32.to_radians();
+        let point = Affine2::from_angle(rotate).transform_point2(point);
+
+        // Zoom in by a factor of 1; points will be twice as far from the origin,
+        // giving us (-4.0, 4.0)
+        let zoom = vec2(1.0, 1.0);
+        let point = point * vec2(powf(2.0, zoom.x), powf(2.0, zoom.y));
+
+        // Apply scaling; scale 100 in a 1000 x 1000 image is an effective range
+        // of [-5, 5] in IFS coordinates.
+        // After scaling, the point is (-400.0, 400.0)
+        let scale = vec2(100.0, 100.0);
+        let point = point * scale;
+
+        // Move the origin from (0, 0) to image center,
+        // giving us (100.0, 900.0)
+        let dimensions = uvec2(1000, 1000);
+        let point = point.as_ivec2() + dimensions.as_ivec2() / 2;
+
+        // Check that the camera implementation ends up at the same point
+        let camera = Camera::new(dimensions, center, rotate, zoom, scale);
+
+        // The camera is implemented by composing affine transforms,
+        // which ends up with a slightly different result because of rounding.
+        let error = camera.transform_point(starting_point) - point;
+        assert!(error.x.abs() <= 1);
+        assert!(error.y.abs() <= 1);
+    }
+
+    #[test]
+    pub fn point_outside_camera() {
+        // Scale 250 for an image 1000 x 1000 gives an effective range of [-2, 2]
+        let camera = Camera::new(
+            uvec2(1000, 1000),
+            Vec2::ZERO,
+            0.0,
+            Vec2::ZERO,
+            vec2(250.0, 250.0),
+        );
+
+        // Converting a point outside the effective range is legal, but outside the image bounds
+        assert_eq!(camera.transform_point(vec2(3.0, 3.0)), ivec2(1250, 1250));
+    }
+
+    #[test]
+    pub fn point_outside_camera_negative() {
+        // Scale 250 for an image 1000 x 1000 gives an effective range of [-2, 2]
+        let camera = Camera::new(
+            uvec2(1000, 1000),
+            Vec2::ZERO,
+            0.0,
+            Vec2::ZERO,
+            vec2(250.0, 250.0),
+        );
+
+        // Converting a point outside the effective range is legal, but outside the image bounds
+        assert_eq!(camera.transform_point(vec2(-3.0, -3.0)), ivec2(-250, -250));
+    }
+
+    #[test]
+    pub fn aspect_ratio() {
+        // Scale 100 for an image 1600 x 900 gives an effective X range of [-8, 8],
+        // and effective Y range of [-4.5, 4.5]
+        let camera = Camera::new(
+            uvec2(1600, 900),
+            Vec2::ZERO,
+            0.0,
+            Vec2::ZERO,
+            vec2(100.0, 100.0),
+        );
+
+        // This point is inside the image width, but outside its height
+        assert_eq!(camera.transform_point(vec2(6.0, 6.0)), ivec2(1400, 1050));
+    }
+}

enkou-shaders/src/chaos_game.rs

@@ -0,0 +1,108 @@
+//! # Chaos Game
+//!
+//! Fractal flames are a class of
+//! [iterated function systems](https://en.wikipedia.org/wiki/Iterated_function_system)
+//! that generate images following a simple algorithm:
+//!
+//! - Pick a starting point `(x, y)`
+//! - Iterate:
+//!     - Pick a [`Transform`] from the set of available transforms
+//!     - Apply the current point to the chosen transform, generating a new point `(x, y)`
+//!     - Plot the new point `(x, y)`
+//!
+//! This algorithm is also known as the ["chaos game"](https://en.wikipedia.org/wiki/Chaos_game),
+//! and it forms the basic system for producing images.
+
+use crate::transform::Transform;
+use crate::variation::Variation;
+use rand::distr::{Distribution, StandardUniform};
+use rand::{Rng, RngExt};
+use spirv_std::glam::{Vec2, vec2};
+
+struct BiUnit;
+impl Distribution<f32> for BiUnit {
+    fn sample<R: Rng + ?Sized>(&self, rng: &mut R) -> f32 {
+        rng.sample::<f32, _>(StandardUniform) * 2.0 - 1.0
+    }
+}
+
+/// Iterate one step in the chaos game; choose the next transform, apply it,
+/// and return the resulting point. Also returns the transform index so that
+/// path-dependent weights (the "Xaos" table in Apophysis) can be chosen
+/// for the next iteration step.
+///
+/// # Arguments
+///
+/// * `weights` - Weights are assumed to be normalized; adding all elements together should return the value 1
+pub fn step_chaos_game<R: Rng>(
+    point: Vec2,
+    rng: &mut R,
+    transforms: &[Transform],
+    weights: &[f32],
+    variations: &[Variation],
+) -> (Vec2, u32) {
+    let mut choice_weight = rng.sample::<f32, _>(StandardUniform);
+    let mut transform_index: u32 = 0;
+
+    for i in 0..weights.len() {
+        choice_weight -= weights[i];
+        if choice_weight <= 0.0 {
+            break;
+        }
+
+        transform_index += 1;
+    }
+
+    (
+        transforms[transform_index as usize].transform_point(rng, variations, point),
+        transform_index,
+    )
+}
+
+/// Iterator for chaos game state. Holds the current point and references to all other data
+/// necessary to generate fractal flame images.
+///
+/// New points in the chaos game are produced by iterating on the chaos game.
+pub struct ChaosGame<'a, R: Rng> {
+    current_point: Vec2,
+    rng: &'a mut R,
+    transforms: &'a [Transform],
+    weights: &'a [f32],
+    variations: &'a [Variation],
+}
+
+impl<'a, R: Rng> ChaosGame<'a, R> {
+    /// Create a new chaos game iterator
+    pub fn new(
+        rng: &'a mut R,
+        transforms: &'a [Transform],
+        weights: &'a [f32],
+        variations: &'a [Variation],
+    ) -> Self {
+        let current_point = vec2(rng.sample(BiUnit), rng.sample(BiUnit));
+        ChaosGame {
+            current_point,
+            rng,
+            transforms,
+            weights,
+            variations,
+        }
+    }
+}
+
+impl<'a, R: Rng> Iterator for ChaosGame<'a, R> {
+    type Item = Vec2;
+
+    fn next(&mut self) -> Option<Self::Item> {
+        let (next_point, _) = step_chaos_game(
+            self.current_point,
+            self.rng,
+            self.transforms,
+            self.weights,
+            self.variations,
+        );
+        self.current_point = next_point;
+
+        Some(next_point)
+    }
+}

enkou-shaders/src/lib.rs

@@ -1,36 +1,111 @@
-#![no_std]
+//! # Enkou
+#![cfg_attr(target_arch = "spirv", no_std)]
 
-use bytemuck::{Pod, Zeroable};
-use core::f32::consts::PI;
-use glam::{Vec3, Vec4, vec2, vec3};
-#[cfg(target_arch = "spirv")]
-use spirv_std::num_traits::Float;
-use spirv_std::spirv;
+// SPIR-V backend is unable to compile iteration over items
+#![allow(clippy::needless_range_loop)]
 
-#[derive(Copy, Clone, Pod, Zeroable)]
-#[repr(C)]
-pub struct ShaderConstants {
-    pub width: u32,
-    pub height: u32,
-    pub time: f32,
+pub mod camera;
+pub mod chaos_game;
+pub mod rng;
+pub mod transform;
+pub mod variation;
+
+use glam::Affine2;
+
+/// Utility trait to convert between `flam3` notation and [`glam`].
+#[allow(missing_docs)]
+pub trait Coefficients2 {
+    /// Convert affine transformation coefficients to the [`glam`] representation.
+    /// Parameters use the following form:
+    ///
+    /// ```text
+    /// (a * x + b * y + c, d * x + e * y + f)
+    /// ```
+    ///
+    /// ```
+    /// # use glam::{Affine2, vec2};
+    /// # use crate::enkou_shaders::Coefficients2;
+    /// let coefs = Affine2::from_coefficients(1.0, 2.0, 3.0, 4.0, 5.0, 6.0);
+    /// let (x, y) = (7.0, 8.0);
+    /// assert_eq!(
+    ///     coefs.transform_point2(vec2(x, y)),
+    ///     vec2(
+    ///         coefs.a() * x + coefs.b() * y + coefs.c(),
+    ///         coefs.d() * x + coefs.e() * y + coefs.f()
+    ///     )
+    /// );
+    /// ```
+    fn from_coefficients(a: f32, b: f32, c: f32, d: f32, e: f32, f: f32) -> Affine2;
+
+    /// Convert affine transformation coefficients to the [`glam`] representation.
+    /// Parameters use the following form:
+    ///
+    /// ```text
+    /// (a * x + b * y + c, d * x + e * y + f)
+    /// ```
+    ///
+    /// ```
+    /// # use glam::{Affine2, vec2};
+    /// # use crate::enkou_shaders::Coefficients2;
+    /// let coefs = Affine2::from_coefficients_arr([1.0, 2.0, 3.0, 4.0, 5.0, 6.0]);
+    /// let (x, y) = (7.0, 8.0);
+    /// assert_eq!(
+    ///     coefs.transform_point2(vec2(x, y)),
+    ///     vec2(
+    ///         coefs.a() * x + coefs.b() * y + coefs.c(),
+    ///         coefs.d() * x + coefs.e() * y + coefs.f()
+    ///     )
+    /// );
+    /// ```
+    fn from_coefficients_arr(coefficients: [f32; 6]) -> Affine2;
+
+    fn a(&self) -> f32;
+    fn b(&self) -> f32;
+    fn c(&self) -> f32;
+    fn d(&self) -> f32;
+    fn e(&self) -> f32;
+    fn f(&self) -> f32;
 }
 
-#[spirv(fragment)]
-pub fn main_fs(vtx_color: Vec3, output: &mut Vec4) {
-    *output = Vec4::from((vtx_color, 1.));
-}
+impl Coefficients2 for Affine2 {
+    #[inline]
+    fn from_coefficients(a: f32, b: f32, c: f32, d: f32, e: f32, f: f32) -> Affine2 {
+        Affine2::from_cols_array(&[a, d, b, e, c, f])
+    }
 
-#[spirv(vertex)]
-pub fn main_vs(
-    #[spirv(vertex_index)] vert_id: i32,
-    #[spirv(descriptor_set = 0, binding = 0, storage_buffer)] constants: &ShaderConstants,
-    #[spirv(position)] vtx_pos: &mut Vec4,
-    vtx_color: &mut Vec3,
-) {
-    let speed = 0.4;
-    let time = constants.time * speed + vert_id as f32 * (2. * PI * 120. / 360.);
-    let position = vec2(f32::sin(time), f32::cos(time));
-    *vtx_pos = Vec4::from((position, 0.0, 1.0));
+    #[inline]
+    fn from_coefficients_arr(coefficients: [f32; 6]) -> Affine2 {
+        Affine2::from_coefficients(
+            coefficients[0],
+            coefficients[1],
+            coefficients[2],
+            coefficients[3],
+            coefficients[4],
+            coefficients[5],
+        )
+    }
 
-    *vtx_color = [vec3(1., 0., 0.), vec3(0., 1., 0.), vec3(0., 0., 1.)][vert_id as usize % 3];
+    fn a(&self) -> f32 {
+        self.matrix2.x_axis.x
+    }
+
+    fn b(&self) -> f32 {
+        self.matrix2.y_axis.x
+    }
+
+    fn c(&self) -> f32 {
+        self.translation.x
+    }
+
+    fn d(&self) -> f32 {
+        self.matrix2.x_axis.y
+    }
+
+    fn e(&self) -> f32 {
+        self.matrix2.y_axis.y
+    }
+
+    fn f(&self) -> f32 {
+        self.translation.y
+    }
 }

enkou-shaders/src/rng.rs

@@ -0,0 +1,58 @@
+//! # RNG
+//!
+//! Random number generation utilities for shaders
+use rand::SeedableRng;
+use rand_xoshiro::Xoshiro256StarStar;
+
+/// Convert an RNG state buffer to an instance of [`Xoshiro256StarStar`].
+///
+/// While [`SeedableRng::from_seed`] is an infallible function,
+/// it relies on some methods that can't be compiled by the SPIR-V
+/// backend (specifically, formatting functions in the core crate).
+///
+/// In practice, the xoshiro RNG state is entirely defined by its seed,
+/// so this function does the work of [`SeedableRng::from_seed`] by
+/// transmuting the seed value to an RNG instance.
+///
+/// This function assumes a properly-initialized state array;
+/// output may silently degenerate if the initial state is all zeros,
+/// so this module is private to the crate.
+pub fn xoshiro256starstar_from_seed(
+    rng_state: <Xoshiro256StarStar as SeedableRng>::Seed,
+) -> Xoshiro256StarStar {
+    let mut rng_state_actual = [0u64; 4];
+
+    // NOTE: Bit shifting is bad, but we don't have great alternatives:
+    //  - `chunks_exact` has issues with pointer casting
+    //  - `u64::from_le_bytes` has issues with `OpBitcast` in SPIR-V validation
+    for i in 0..rng_state_actual.len() {
+        for j in 0..size_of::<u64>() {
+            rng_state_actual[i] |= (rng_state[i * size_of::<u64>() + j] as u64) << (j * 8);
+        }
+    }
+
+    unsafe { core::mem::transmute(rng_state_actual) }
+}
+
+#[cfg(test)]
+mod test {
+    use crate::rng::xoshiro256starstar_from_seed;
+    use core::iter::zip;
+    use rand::{RngExt, SeedableRng};
+    use rand_xoshiro::Xoshiro256StarStar;
+
+    #[test]
+    fn match_seeded() {
+        let mut seed: <Xoshiro256StarStar as SeedableRng>::Seed = [0u8; 32];
+        for i in 0..seed.len() {
+            seed[i] = i as u8;
+        }
+
+        let rng1 = Xoshiro256StarStar::from_seed(seed).random_iter::<u64>();
+        let rng2 = xoshiro256starstar_from_seed(seed).random_iter::<u64>();
+
+        zip(rng1, rng2)
+            .take(100)
+            .for_each(|(rng1_value, rng2_value)| assert_eq!(rng1_value, rng2_value));
+    }
+}

enkou-shaders/src/transform.rs

@@ -0,0 +1,97 @@
+//! # Transform
+//!
+//! Transforms are the "functions" in an iterated function system. They take in a point,
+//! and generate a new point. For fractal flames, transforms are always affine,
+//! but produce more interesting images once we add variations.
+use crate::variation::Variation;
+use bytemuck::{Pod, Zeroable};
+use glam::{Affine2, UVec2, Vec2};
+use rand::Rng;
+
+/// Affine transform for use in the [`chaos_game`](crate::chaos_game).
+#[derive(Copy, Clone, Pod, Zeroable)]
+#[repr(C)]
+pub struct Transform {
+    coefficients: Affine2,
+    coefficients_post: Affine2,
+    variation_range: UVec2,
+}
+
+impl Transform {
+    /// Create a new transform from an affine transformation matrix
+    pub fn new(coefficients: Affine2, coefficients_post: Affine2, variation_range: UVec2) -> Self {
+        Transform {
+            coefficients,
+            coefficients_post,
+            variation_range,
+        }
+    }
+
+    /// Apply this transform to a point in IFS coordinates, producing a new point
+    pub fn transform_point<R: Rng>(
+        &self,
+        rng: &mut R,
+        variations: &[Variation],
+        point: Vec2,
+    ) -> Vec2 {
+        let point = self.coefficients.transform_point2(point);
+
+        let mut point_output = Vec2::ZERO;
+
+        let variation_start = self.variation_range.x;
+        let variation_end = self.variation_range.y;
+        for variation_index in variation_start..variation_end {
+            let variation = &variations[variation_index as usize];
+            point_output += variation.transform_point(point, rng, &self.coefficients)
+        }
+
+        self.coefficients_post.transform_point2(point)
+    }
+}
+
+#[cfg(test)]
+mod test {
+    use crate::rng::xoshiro256starstar_from_seed;
+    use crate::transform::Transform;
+    use crate::variation::{Variation, VariationKind};
+    use glam::{Affine2, uvec2, vec2};
+
+    #[test]
+    fn transform_scaling() {
+        let scale_coefficients = vec2(2.0, 0.5);
+        let transform = Transform::new(
+            Affine2::from_scale(scale_coefficients),
+            Affine2::IDENTITY,
+            uvec2(0, 1),
+        );
+
+        let mut rng = xoshiro256starstar_from_seed([0; 32]);
+        let variations = [Variation::IDENTITY];
+        let point = vec2(1.0, 1.0);
+
+        assert_eq!(
+            transform.transform_point(&mut rng, &variations, point),
+            scale_coefficients
+        );
+    }
+
+    #[test]
+    fn transform_scaling_post() {
+        let scale_coefficients = vec2(2.0, 0.5);
+        let transform_pdj = Transform::new(Affine2::IDENTITY, Affine2::IDENTITY, uvec2(0, 1));
+        let transform_pdj_post = Transform::new(
+            Affine2::IDENTITY,
+            Affine2::from_scale(scale_coefficients),
+            uvec2(0, 1),
+        );
+
+        let mut rng = xoshiro256starstar_from_seed([0; 32]);
+        let variations = [Variation::new(VariationKind::Pdj, 1.0, [0.0f32; 4].into())];
+        let point = vec2(1.0, 1.0);
+
+        let point_pdj = transform_pdj.transform_point(&mut rng, &variations, point);
+        let point_pdj_post = transform_pdj_post.transform_point(&mut rng, &variations, point);
+
+        assert_eq!(point_pdj * scale_coefficients, point_pdj_post);
+    }
+}

enkou-shaders/src/variation.rs

@@ -0,0 +1,132 @@
+//! # Variation
+//!
+//! Variations extend the fractal flame iterated function system
+//! with non-linear transforms (as opposed to [`Transform`]s,
+//! which are strictly affine transformations).
+use crate::Coefficients2;
+use bytemuck::{Pod, Zeroable};
+use core::f32::consts::PI;
+use glam::{Affine2, Vec2, vec2};
+use libm::{atan2f, cosf, powf, sinf, sqrtf, tanf};
+use rand::distr::StandardUniform;
+use rand::{Rng, RngExt};
+
+/// Generic variation parameters
+///
+/// Not all variations will use these parameters, but passing them
+/// as an array per variation allows shaders to use a consistent struct size
+/// no matter what the variation actually needs.
+#[derive(Copy, Clone, Pod, Zeroable)]
+#[repr(C)]
+pub struct VariationParams([f32; 4]);
+
+impl From<[f32; 4]> for VariationParams {
+    fn from(v: [f32; 4]) -> Self {
+        VariationParams(v)
+    }
+}
+
+/// Enum for all supported variation types
+///
+/// ID numbers are chosen to match the variation identifier also used by `flam3`
+#[derive(Copy, Clone)]
+#[repr(u32)]
+#[allow(missing_docs)]
+pub enum VariationKind {
+    /// Identity variation, returns the point as-is
+    Linear = 0,
+
+    Julia = 13,
+    Popcorn = 17,
+    Pdj = 24,
+}
+
+// UNSAFE: Sound because enum has guaranteed layout (u32) and defined zero-value
+unsafe impl bytemuck::Zeroable for VariationKind {}
+// UNSAFE: Sound because enum has guaranteed layout (u32) and defined zero-value
+unsafe impl bytemuck::Pod for VariationKind {}
+
+/// Parameters required for shaders to run the variation function.
+///
+/// Not all variations use the [`VariationParams`], but using the struct
+/// makes it easy to provide parameters to the shader.
+#[derive(Copy, Clone, Pod, Zeroable)]
+#[repr(C)]
+pub struct Variation {
+    kind: VariationKind,
+    weight: f32,
+    params: VariationParams,
+}
+
+impl Variation {
+    /// Identity variation; calling [`transform_point`] will yield
+    /// the same point as the input.
+    pub const IDENTITY: Variation = Variation {
+        kind: VariationKind::Linear,
+        weight: 1.0,
+        params: VariationParams([0f32; 4]),
+    };
+
+    /// Create a new variation by providing the variation kind, weight, and parameters.
+    pub const fn new(kind: VariationKind, weight: f32, params: VariationParams) -> Variation {
+        Variation {
+            kind,
+            weight,
+            params,
+        }
+    }
+
+    /// Transform a point by applying this variation.
+    ///
+    /// Output points are scaled by this variation's weight.
+    pub fn transform_point<R: Rng>(
+        &self,
+        point: Vec2,
+        rng: &mut R,
+        coefficients: &Affine2,
+    ) -> Vec2 {
+        (match self.kind {
+            VariationKind::Linear => transform_point_linear(point),
+            VariationKind::Julia => transform_point_julia(point, rng),
+            VariationKind::Popcorn => transform_point_popcorn(point, coefficients),
+            VariationKind::Pdj => transform_point_pdj(point, &self.params),
+        }) * self.weight
+    }
+}
+
+fn transform_point_linear(point: Vec2) -> Vec2 {
+    point
+}
+
+fn transform_point_julia<R: Rng>(point: Vec2, rng: &mut R) -> Vec2 {
+    let x2 = powf(point.x, 2.0);
+    let y2 = powf(point.y, 2.0);
+    let r = sqrtf(x2 + y2);
+
+    let theta = atan2f(point.x, point.y);
+    let omega = if rng.sample::<f32, _>(StandardUniform) > 0.5 {
+        PI
+    } else {
+        0.0
+    };
+
+    let sqrt_r = sqrtf(r);
+    let theta_val = theta / 2.0 + omega;
+
+    vec2(sqrt_r * cosf(theta_val), sqrt_r * sinf(theta_val))
+}
+
+fn transform_point_popcorn(point: Vec2, coefficients: &Affine2) -> Vec2 {
+    vec2(
+        point.x * coefficients.c() * sinf(tanf(3.0 * point.y)),
+        point.y + coefficients.f() * sinf(tanf(3.0 * point.x)),
+    )
+}
+
+fn transform_point_pdj(point: Vec2, params: &VariationParams) -> Vec2 {
+    let (pdj_a, pdj_b, pdj_c, pdj_d) = (params.0[0], params.0[1], params.0[2], params.0[3]);
+    vec2(
+        sinf(pdj_a * point.y) - cosf(pdj_b * point.x),
+        sinf(pdj_c * point.x) - cosf(pdj_d * point.y),
+    )
+}

examples/image-binary/Cargo.toml

@@ -1,18 +1,16 @@
 [package]
-name = "enkou-shaders-tests"
-publish = false
+name = "image-binary"
 version.workspace = true
 authors.workspace = true
 edition.workspace = true
 license.workspace = true
 repository.workspace = true
 
+[dependencies]
+enkou-shaders = { path = "../../enkou-shaders" }
+glam.workspace = true
+spirv-std.workspace = true
+wgpu = { workspace = true, optional = true }
+
 [lints]
 workspace = true
-
-[dependencies]
-rspirv.workspace = true
-
-[build-dependencies]
-anyhow.workspace = true
-cargo-gpu-install.workspace = true

examples/image-binary/src/lib.rs

@@ -0,0 +1,82 @@
+//! # Binary image
+#![cfg_attr(target_arch = "spirv", no_std)]
+use enkou_shaders::camera::Camera;
+use enkou_shaders::chaos_game::ChaosGame;
+use enkou_shaders::rng::xoshiro256starstar_from_seed;
+use enkou_shaders::transform::Transform;
+use enkou_shaders::variation::Variation;
+use glam::{UVec2, UVec4};
+use spirv_std::spirv;
+
+#[cfg(feature = "wgpu")]
+pub use wgpu::*;
+
+const IMAGE_QUALITY: f32 = 1.0;
+const ITERATIONS_FUSE: u32 = 20;
+
+/// Sierpinski Gasket
+#[spirv(compute(entry_point_name = "main_image_binary", threads(1)))]
+pub fn main_image_binary(
+    #[spirv(storage_buffer, descriptor_set = 0, binding = 0)] image_dimensions: &UVec2,
+    #[spirv(storage_buffer, descriptor_set = 0, binding = 1)] transforms: &[Transform],
+    #[spirv(storage_buffer, descriptor_set = 0, binding = 2)] weights: &[f32],
+    #[spirv(storage_buffer, descriptor_set = 0, binding = 3)] variations: &[Variation],
+    #[spirv(storage_buffer, descriptor_set = 0, binding = 4)] camera: &Camera,
+    #[spirv(storage_buffer, descriptor_set = 0, binding = 5)] image_buffer: &mut [UVec4],
+) {
+    // Initialize RNG and run the chaos game
+    let mut rng = xoshiro256starstar_from_seed([4; 32]);
+    let mut chaos_game = ChaosGame::new(&mut rng, transforms, weights, variations);
+
+    // Discard the first few iterations
+    for _ in 0..ITERATIONS_FUSE {
+        chaos_game.next().unwrap();
+    }
+
+    // Plot the remaining points generated by the chaos game
+    let iterations = (image_dimensions.as_vec2().element_product() * IMAGE_QUALITY) as u32;
+    for _ in 0..iterations {
+        let ifs_point = chaos_game.next().unwrap();
+        let pixel_point = camera.transform_point_to_image(ifs_point);
+
+        if let Some(pixel_point) = pixel_point {
+            let pixel_index = pixel_point.y * image_dimensions.x + pixel_point.x;
+            image_buffer[pixel_index as usize] = UVec4::splat(255);
+        }
+    }
+}
+
+#[cfg(feature = "wgpu")]
+pub mod wgpu {
+    const fn bgle(binding: u32, read_only: bool) -> wgpu::BindGroupLayoutEntry {
+        wgpu::BindGroupLayoutEntry {
+            binding,
+            visibility: wgpu::ShaderStages::COMPUTE,
+            ty: wgpu::BindingType::Buffer {
+                ty: wgpu::BufferBindingType::Storage { read_only },
+                has_dynamic_offset: false,
+                min_binding_size: None,
+            },
+            count: None,
+        }
+    }
+
+    pub const BGLE_IMAGE_DIMENSIONS: wgpu::BindGroupLayoutEntry = bgle(0, true);
+    pub const BGLE_TRANSFORMS: wgpu::BindGroupLayoutEntry = bgle(1, true);
+    pub const BGLE_WEIGHTS: wgpu::BindGroupLayoutEntry = bgle(2, true);
+    pub const BGLE_VARIATIONS: wgpu::BindGroupLayoutEntry = bgle(3, true);
+    pub const BGLE_CAMERA: wgpu::BindGroupLayoutEntry = bgle(4, true);
+    pub const BGLE_IMAGE_BUFFER: wgpu::BindGroupLayoutEntry = bgle(5, false);
+
+    pub const BIND_GROUP_IMAGE_BINARY: wgpu::BindGroupLayoutDescriptor = wgpu::BindGroupLayoutDescriptor {
+        label: Some("main_image_binary"),
+        entries: &[
+            BGLE_IMAGE_DIMENSIONS,
+            BGLE_TRANSFORMS,
+            BGLE_WEIGHTS,
+            BGLE_VARIATIONS,
+            BGLE_CAMERA,
+            BGLE_IMAGE_BUFFER,
+        ],
+    };
+}

examples/image-runner/Cargo.toml

@@ -0,0 +1,27 @@
+[package]
+name = "image-runner"
+version.workspace = true
+authors.workspace = true
+edition.workspace = true
+license.workspace = true
+repository.workspace = true
+
+[dependencies]
+enkou-shaders = { path = "../../enkou-shaders" }
+image-binary = { path = "../image-binary", features = ["wgpu"] }
+
+anyhow.workspace = true
+bytemuck.workspace = true
+futures.workspace = true
+glam = { workspace = true, features = ["u8"] }
+image.workspace = true
+tempfile.workspace = true
+wgpu.workspace = true
+xflags.workspace = true
+
+[build-dependencies]
+anyhow.workspace = true
+cargo-gpu-install.workspace = true
+
+[lints]
+workspace = true

examples/image-runner/build.rs

@@ -1,26 +1,28 @@
 use cargo_gpu_install::install::Install;
-use cargo_gpu_install::spirv_builder::{ShaderPanicStrategy, SpirvMetadata};
+use cargo_gpu_install::spirv_builder::{Capability, ShaderPanicStrategy, SpirvMetadata};
 use std::path::PathBuf;
 
 pub fn main() -> anyhow::Result<()> {
     let manifest_dir = env!("CARGO_MANIFEST_DIR");
-    let crate_path = [manifest_dir, "..", "enkou-shaders"]
+    let crate_path = [manifest_dir, "..", "image-binary"]
         .iter()
         .copied()
         .collect::<PathBuf>();
 
-    let mut install = Install::from_shader_crate(crate_path.clone());
-    install.build_script = true;
-    install.auto_install_rust_toolchain = true;
-    let install = install.run()?;
-
+    let install = Install::from_shader_crate(crate_path.clone())
+        .within_build_script()
+        .run()?;
     let mut builder = install.to_spirv_builder(crate_path, "spirv-unknown-vulkan1.3");
     builder.build_script.defaults = true;
     builder.shader_panic_strategy = ShaderPanicStrategy::SilentExit;
     builder.spirv_metadata = SpirvMetadata::Full;
+    builder.capabilities = vec![Capability::Int8, Capability::Int16, Capability::Int64];
 
     let compile_result = builder.build()?;
     let spv_path = compile_result.module.unwrap_single();
-    println!("cargo::rustc-env=SHADER_SPV_PATH={}", spv_path.display());
+    println!(
+        "cargo::rustc-env=SHADER_SPV_PATH_IMAGE_BINARY={}",
+        spv_path.display()
+    );
     Ok(())
 }

examples/image-runner/src/image_binary.rs

@@ -0,0 +1,210 @@
+use enkou_shaders::Coefficients2;
+use enkou_shaders::transform::Transform;
+use enkou_shaders::variation::Variation;
+use futures::channel::oneshot;
+use futures::executor::block_on;
+use glam::{uvec2, Affine2, UVec2, UVec4, Vec2};
+use image::{Rgba, RgbaImage};
+use image_binary::{main_image_binary, BIND_GROUP_IMAGE_BINARY, BGLE_IMAGE_DIMENSIONS, BGLE_TRANSFORMS, BGLE_WEIGHTS, BGLE_CAMERA, BGLE_IMAGE_BUFFER, BGLE_VARIATIONS};
+use std::path::Path;
+use wgpu::util::DeviceExt;
+use enkou_shaders::camera::Camera;
+
+fn transforms() -> [Transform; 3] {
+    [
+        {
+            // F_0: (x / 2, y / 2)
+            let coefficients = Affine2::from_coefficients(0.5, 0.0, 0.0, 0.0, 0.5, 0.0);
+            Transform::new(coefficients, Affine2::IDENTITY, uvec2(0, 1))
+        },
+        {
+            // F_1: ((x + 1) / 2, y / 2)
+            let coefficients = Affine2::from_coefficients(0.5, 0.0, 0.5, 0.0, 0.5, 0.0);
+            Transform::new(coefficients, Affine2::IDENTITY, uvec2(0, 1))
+        },
+        {
+            // F_2: (x / 2, (y + 1) / 2)
+            let coefficients = Affine2::from_coefficients(0.5, 0.0, 0.0, 0.0, 0.5, 0.5);
+            Transform::new(coefficients, Affine2::IDENTITY, uvec2(0, 1))
+        },
+    ]
+}
+
+fn weights() -> [f32; 3] {
+    [1.0 / 3.0; 3]
+}
+
+fn variations() -> [Variation; 1] {
+    [Variation::IDENTITY]
+}
+
+fn camera(image_dimensions: UVec2) -> Camera {
+    Camera::new(
+        image_dimensions,
+        Vec2::ONE * 0.5,
+        0.0,
+        Vec2::ZERO,
+        Vec2::splat(image_dimensions.min_element() as f32),
+    )
+}
+
+pub(crate) fn main_cpu(image_dimensions: UVec2, output_path: &Path) -> Result<(), anyhow::Error> {
+    let mut image_buffer = Vec::<UVec4>::new();
+    image_buffer.resize(image_dimensions.element_product() as usize, UVec4::ZERO);
+
+    main_image_binary(
+        &image_dimensions,
+        &transforms(),
+        &weights(),
+        &variations(),
+        &camera(image_dimensions),
+        &mut image_buffer,
+    );
+
+    let mut image = RgbaImage::new(image_dimensions.x, image_dimensions.y);
+    for (i, color) in image_buffer.into_iter().enumerate() {
+        let image_x = i as u32 % image_dimensions.x;
+        let image_y = i as u32 / image_dimensions.x;
+
+        image.put_pixel(image_x, image_y, color.as_u8vec4().to_array().into());
+    }
+
+    image.save(output_path)?;
+    Ok(())
+}
+
+const SHADER_MODULE: wgpu::ShaderModuleDescriptor = wgpu::include_spirv!(env!("SHADER_SPV_PATH_IMAGE_BINARY"));
+
+fn bge<'a>(entry: &'a wgpu::BindGroupLayoutEntry, buffer: &'a wgpu::Buffer) -> wgpu::BindGroupEntry<'a> {
+    wgpu::BindGroupEntry {
+        binding: entry.binding,
+        resource: buffer.as_entire_binding(),
+    }
+}
+
+pub(crate) fn main_gpu(device: &wgpu::Device, queue: &wgpu::Queue, image_dimensions: UVec2, output_path: &Path) -> Result<(), anyhow::Error> {
+
+    let bind_group_layout = device.create_bind_group_layout(&BIND_GROUP_IMAGE_BINARY);
+
+    let image_dimensions_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
+        label: Some("image_dimensions"),
+        contents: bytemuck::bytes_of(&image_dimensions),
+        usage: wgpu::BufferUsages::COPY_DST | wgpu::BufferUsages::STORAGE,
+    });
+
+    let transforms_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
+        label: Some("transforms"),
+        contents: bytemuck::cast_slice(&transforms()),
+        usage: wgpu::BufferUsages::COPY_DST | wgpu::BufferUsages::STORAGE,
+    });
+
+    let weights_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
+        label: Some("weights"),
+        contents: bytemuck::cast_slice(&weights()),
+        usage: wgpu::BufferUsages::COPY_DST | wgpu::BufferUsages::STORAGE,
+    });
+
+    let variations_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
+        label: Some("variations"),
+        contents: bytemuck::cast_slice(&variations()),
+        usage: wgpu::BufferUsages::COPY_DST | wgpu::BufferUsages::STORAGE,
+    });
+
+    let camera_buffer = device.create_buffer_init(&wgpu::util::BufferInitDescriptor {
+        label: Some("camera"),
+        contents: bytemuck::bytes_of(&camera(image_dimensions)),
+        usage: wgpu::BufferUsages::COPY_DST | wgpu::BufferUsages::STORAGE,
+    });
+
+    let image_buffer_elements = image_dimensions.element_product() as u64;
+    let image_buffer_size = image_buffer_elements * size_of::<UVec4>() as u64;
+    let image_buffer = device.create_buffer(&wgpu::BufferDescriptor {
+        label: Some("image_buffer"),
+        size: image_buffer_size,
+        usage: wgpu::BufferUsages::COPY_SRC | wgpu::BufferUsages::STORAGE,
+        mapped_at_creation: false,
+    });
+
+    let image_staging = device.create_buffer(&wgpu::BufferDescriptor {
+        label: Some("image_buffer_staging"),
+        size: image_buffer_size,
+        usage: wgpu::BufferUsages::MAP_READ | wgpu::BufferUsages::COPY_DST,
+        mapped_at_creation: false,
+    });
+
+    let bind_group = device.create_bind_group(&wgpu::BindGroupDescriptor {
+        label: Some("main_image_binary"),
+        layout: &bind_group_layout,
+        entries: &[
+            bge(&BGLE_IMAGE_DIMENSIONS, &image_dimensions_buffer),
+            bge(&BGLE_TRANSFORMS, &transforms_buffer),
+            bge(&BGLE_WEIGHTS, &weights_buffer),
+            bge(&BGLE_VARIATIONS, &variations_buffer),
+            bge(&BGLE_CAMERA, &camera_buffer),
+            bge(&BGLE_IMAGE_BUFFER, &image_buffer),
+        ],
+    });
+
+    let layout = device.create_pipeline_layout(&wgpu::PipelineLayoutDescriptor {
+        label: Some("main_image_binary"),
+        bind_group_layouts: &[Some(&bind_group_layout)],
+        immediate_size: 0,
+    });
+
+    let module = device.create_shader_module(SHADER_MODULE);
+
+    let compute_pipeline = device.create_compute_pipeline(&wgpu::ComputePipelineDescriptor {
+        label: Some("main_image_binary"),
+        layout: Some(&layout),
+        module: &module,
+        entry_point: Some("main_image_binary"),
+        compilation_options: Default::default(),
+        cache: None,
+    });
+
+    let mut encoder = device.create_command_encoder(&wgpu::CommandEncoderDescriptor {
+        label: Some("main_image_binary"),
+    });
+
+    {
+        let mut compute_pass = encoder.begin_compute_pass(&wgpu::ComputePassDescriptor {
+            label: Some("main_image_binary"),
+            timestamp_writes: None,
+        });
+
+        compute_pass.set_pipeline(&compute_pipeline);
+        compute_pass.set_bind_group(0, &bind_group, &[]);
+    }
+
+    encoder.copy_buffer_to_buffer(&image_buffer, 0, &image_staging, 0, Some(image_buffer_size));
+
+    let (sender, receiver) = oneshot::channel();
+    let image_staging_capturable = image_staging.clone();
+    encoder.map_buffer_on_submit(&image_buffer, wgpu::MapMode::Read, .., move |result| {
+        result.expect("unable to map buffer");
+
+        let staging_buffer_view = image_staging_capturable.get_mapped_range(..).expect("Unable to map staging buffer");
+
+        let mut image = RgbaImage::new(image_dimensions.x, image_dimensions.y);
+        let image_buffer_elements = bytemuck::cast_slice::<u8, UVec4>(staging_buffer_view.as_ref());
+        for (i, element) in image_buffer_elements.iter().enumerate() {
+            let image_x = i as u32 % image_dimensions.x;
+            let image_y = i as u32 / image_dimensions.x;
+
+            let pixel_colors = element.as_u8vec4();
+            image.put_pixel(image_x, image_y, Rgba(*pixel_colors.as_ref()))
+        }
+
+        sender.send(image).expect("Unable to send image");
+    });
+
+    queue.submit(Some(encoder.finish()));
+    device.poll(wgpu::PollType::wait_indefinitely())?;
+    let image = block_on(receiver)?;
+
+    image_staging.unmap();
+
+    image.save(output_path)?;
+
+    Ok(())
+}

examples/image-runner/src/main.rs

@@ -0,0 +1,77 @@
+use glam::{uvec2};
+use std::mem;
+use std::process::Command;
+use std::path::PathBuf;
+use futures::executor::block_on;
+use tempfile::NamedTempFile;
+
+mod image_binary;
+
+fn main() -> Result<(), anyhow::Error> {
+    let instance_future = wgpu::util::new_instance_with_webgpu_detection(wgpu::InstanceDescriptor {
+        backends: Default::default(),
+        flags: Default::default(),
+        memory_budget_thresholds: Default::default(),
+        backend_options: Default::default(),
+        display: None,
+    });
+    let instance = block_on(instance_future);
+
+    let adapter_future = instance.request_adapter(&wgpu::RequestAdapterOptions {
+        power_preference: Default::default(),
+        force_fallback_adapter: false,
+        compatible_surface: None,
+        apply_limit_buckets: false,
+    });
+    let adapter = block_on(adapter_future)?;
+
+    let device_future = adapter.request_device(&wgpu::DeviceDescriptor {
+        label: Some("image-runner"),
+        required_features: Default::default(),
+        required_limits: Default::default(),
+        experimental_features: Default::default(),
+        memory_hints: Default::default(),
+        trace: Default::default(),
+    });
+    let (device, queue) = block_on(device_future)?;
+
+    let flags = xflags::parse_or_exit! {
+        /// Image dimensions to output, as `width,height`
+        optional -d, --dimensions dimensions: String
+
+        /// Output pathname to use
+        optional -o, --output output: PathBuf
+
+        /// Image type to generate
+        required image: String
+    };
+
+    let dimensions = if let Some(dimensions) = flags.dimensions {
+        let (width_str, height_str) = dimensions.split_once(",").ok_or(anyhow::anyhow!("Invalid format for image dimensions"))?;
+        uvec2(width_str.parse()?, height_str.parse()?)
+    } else {
+        uvec2(1600, 900)
+    };
+
+    let output = if let Some(output) = flags.output { output } else {
+        let path = NamedTempFile::with_suffix(".png")?;
+        let pathbuf: PathBuf = path.path().into();
+        mem::forget(path);
+        pathbuf
+    };
+
+    match flags.image.as_ref() {
+        "binary_cpu" => image_binary::main_cpu(dimensions, output.as_ref()),
+        "binary_gpu" => image_binary::main_gpu(&device, &queue, dimensions, output.as_ref()),
+        _ => Err(anyhow::anyhow!("Unrecognized image type"))
+    }?;
+
+    let mut command = cfg_select! {
+        unix => Command::new("xdg-open").arg(temp.path()).spawn(),
+        windows => Command::new("PowerShell").arg("-Command").arg(format!("start {}", output.display())).spawn(),
+        _ => Err(anyhow::anyhow!("No available program to open images"))?
+    }?;
+    command.wait()?;
+
+    Ok(())
+}