#wgsl #wgpu #shader

wgsl_bindgen

Generate typesafe Rust bindings for wgsl shaders in wgpu

15 releases

new 0.6.12 Feb 25, 2024
0.6.11 Feb 24, 2024
0.5.3 Feb 11, 2024

#44 in Graphics APIs

Download history 38/week @ 2024-01-28 37/week @ 2024-02-04 75/week @ 2024-02-11 1128/week @ 2024-02-18

1,278 downloads per month

MIT license

235KB
6K SLoC

wgsl-bindgen

Latest Version docs.rs
An experimental library for generating typesafe Rust bindings from WGSL shaders to wgpu.

wgsl_bindgen, powered by naga-oil, is a tool that integrates into your Rust build process. It parses WGSL shaders and generates corresponding Rust modules. These modules contain type definitions and boilerplate code that match your shaders, reducing the risk of runtime errors by catching mismatches at compile time.

The tool facilitates a shader-focused workflow. When you modify your WGSL shaders, the changes are automatically reflected in the Rust code. This immediate feedback helps catch errors early, making it easier to work with shaders.

Features

  • Supports import syntax and many more features from naga oil flavour.
  • BYOT - Bring your own types for wgsl matrix, vectors types. Bindgen automatically also include assertion to test alignment and sizes for your types.
  • Rust structs for vertex, storage, and uniform buffers
  • Either use encase or bytemuck derives, and optionally serde for generated structs.
  • Const validation of WGSL memory layout for provided vector and matrix types and generated structs when using bytemuck
  • More strongly typed bind group and bindings initialization
  • Shader module initialization using either embedded source string, or compose modules for extensibility.
  • Ability to add additional scan directories for shader imports useful for shader unit testing.

Usage

When enabling derives for crates like bytemuck, serde, or encase, these dependencies should also be added to the Cargo.toml with the appropriate derive features. See the provided example project for basic usage.

[dependencies]
bytemuck = "..."

[build-dependencies]
wgsl_bindgen = "..."

Then, in your build.rs:

use wgsl_bindgen::WgslBindgenOptionBuilder;

fn main() {
  let bindgen = WgslBindgenOptionBuilder::default()
    .add_entry_point("src/pbr.wgsl")
    .add_entry_point("src/pfx.wgsl")
    .output_file("src/shader.rs")
    .build()
    .unwrap();

  bindgen.generate().unwrap();
}

This will generate Rust bindings for the WGSL shader at src/pbr.wgsl, src/pfx.wgsl and write them to src/shader.rs. See the example crate for how to use the generated code. Run the example with cargo run.

Wgsl Import Resolution

wgsl_bindgen uses a specific strategy to resolve the import paths in your WGSL source code. This process is handled by the ModulePathResolver::generate_possible_paths function.

Consider the following directory structure:

/my_project
├── src
│   ├── shaders
│   │   ├── main.wgsl
│   │   ├── utils
│   │   │   ├── math.wgsl
│   ├── main.rs
├── Cargo.toml

And the following import statement in main.wgsl:

import utils::math;

Here's how wgsl_bindgen resolves the import path:

  1. The function first checks if the import module name (utils::math) starts with the module prefix. If a module prefix is set and matches, it removes the prefix and treats the rest of the import module name as a relative path from the entry source directory converting the double semicolor :: to forward slash / from the directory of the current source file (src/shaders).
  2. If the import module name does not start with the module prefix, it treats the entire import module name as a relative path from the directory of the current source file. In this case, it will look for utils/math.wgsl in the same directory as main.wgsl.
  3. The function then returns a set of possible import paths. The actual file that the import statement refers to is the first file in this set that exists. In this case, it would successfully find and import src/shaders/utils/math.wgsl.
  4. If not, the second possible path it would have tried would be src/shaders/utils.wgsl treating math as an item within utils.wgsl had it existed.

This strategy allows wgsl_bindgen to handle a variety of import statement formats and directory structures, providing flexibility in how you organize your WGSL source files.

Memory Layout

WGSL structs have different memory layout requirements than Rust structs or standard layout algorithms like repr(C) or repr(packed). Matching the expected layout to share data between the CPU and GPU can be tedious and error prone. wgsl_bindgen offers options to add derives for encase to handle padding and alignment at runtime or bytemuck for enforcing padding and alignment at compile time.

When deriving bytemuck, wgsl_bindgen will use naga's layout calculations to add const assertions to ensure that all fields of host-shareable types (structs for uniform and storage buffers) have the correct offset, size, and alignment expected by WGSL.

Bind Groups

wgpu uses resource bindings organized into bind groups to define global shader resources like textures and buffers. Shaders can have many resource bindings organized into up to 4 bind groups. wgsl_bindgen will generate types and functions for initializing and setting these bind groups in a more typesafe way. Adding, removing, or changing bind groups in the WGSl shader will typically result in a compile error instead of a runtime error when compiling the code without updating the code for creating or using these bind groups.

While bind groups can easily be set all at once using the bind_groups::set_bind_groups function, it's recommended to organize bindings into bindgroups based on their update frequency. Bind group 0 will change the least frequently like per frame resources with bind group 3 changing most frequently like per draw resources. Bind groups can be set individually using their set(render_pass) method. This can provide a small performance improvement for scenes with many draw calls. See descriptor table frequency (DX12) and descriptor set frequency (Vulkan) for details.

Organizing bind groups in this way can also help to better organize rendering resources in application code instead of redundantly storing all resources with each object. The bindgroups::BindGroup0 may only need to be stored once while bindgroups::BindGroup3 may be stored for each mesh in the scene. Note that bind groups store references to their underlying resource bindings, so it is not necessary to recreate a bind group if the only the uniform or storage buffer contents change. Avoid creating new bind groups during rendering if possible for best performance.

Limitations

  • It may be necessary to disable running this function for shaders with unsupported types or features. Please make an issue if any new or existing WGSL syntax is unsupported. The goal is just to generate most of the tedious and error prone boilerplate required to use WGSL shaders with wgpu.
  • Most but not all WGSL types are currently supported.
  • Vertex attributes using floating point types in WGSL like vec2<f32> are assumed to use float inputs instead of normalized attributes like unorm or snorm integers.
  • All textures are assumed to be filterable and all samplers are assumed to be filtering. This may lead to compatibility issues. This can usually be resolved by requesting the native only feature TEXTURE_ADAPTER_SPECIFIC_FORMAT_FEATURES.
  • It's possible to achieve slightly better performance than the generated code in some cases like avoiding redundant bind group bindings or adjusting resource shader stage visibility. This should be addressed by using some handwritten code where appropriate.

Differences from the wgsl_to_wgpu

  • Supports WGSL import syntax and many more features from naga oil flavour.
  • You can only choose either bytemuck or encase for serialization
  • Bytemuck mode supports Runtime-Sized-Array as generic const array in rust.
  • Bytemuck mode correctly adds padding for mat3x3, vec3, whereas original would fail at compile assertions. (The fork was mostly born out of reason to use bytemuck and ensure it works in all cases instead of refusing certain types.)
  • User can provide their own wgsl type mappings using quote library
  • Expect small api surface breaking change.

Publishing Crates

The provided example project outputs the generated bindings to the src/ directory for documentation purposes. This approach is also fine for applications. Published crates should follow the recommendations for build scripts in the Cargo Book.

use miette::{IntoDiagnostic, Result};
use wgsl_bindgen::{WgslTypeSerializeStrategy, WgslBindgenOptionBuilder, GlamWgslTypeMap};

// src/build.rs
fn main() -> Result<()> {
    WgslBindgenOptionBuilder::default()
        .add_entry_point("src/shader/testbed.wgsl")
        .add_entry_point("src/shader/triangle.wgsl")
        .serialization_strategy(WgslTypeSerializeStrategy::Bytemuck)
        .wgsl_type_map(GlamWgslTypeMap)
        .derive_serde(false)
        .output_file("src/shader.rs")
        .build()?
        .generate()
        .into_diagnostic()
}

The generated code will need to be included in one of the normal source files. This includes adding any nested modules as needed.

// src/lib.rs
mod shader;

Dependencies

~11–24MB
~329K SLoC