15 releases (breaking)
new 0.11.0 | Oct 31, 2024 |
---|---|
0.10.0 | Aug 29, 2024 |
0.9.0 | Jul 27, 2024 |
0.6.0 | Feb 23, 2024 |
0.1.0 | Jul 11, 2021 |
#56 in Graphics APIs
287 downloads per month
Used in xc3_wgpu
170KB
4K
SLoC
wgsl_to_wgpu
A library for generating typesafe Rust bindings from WGSL shaders to wgpu.
wgsl_to_wgpu is designed to be incorporated into the compilation process using a build script. The WGSL shaders are parsed using naga to generate a corresponding Rust module. The generated Rust module contains the type definitions and boilerplate code needed to work with the WGSL shader module. Using the generated code can also reduce many instances of invalid API usage.
wgsl_to_wgpu facilitates a shader focused workflow where edits to WGSL code are automatically reflected in the corresponding Rust file. For example, changing the type of a uniform in WGSL will raise a compile error in Rust code using the generated struct to initialize the buffer.
Features
- more strongly typed bind group and bindings initialization
- shader module initialization
- Rust structs for vertex, storage, and uniform buffers
- Rust constants for shader constants and pipeline parameters like workgroup size
- optional derives for encase, bytemuck, and serde
- const validation of the WGSL memory layout for generated structs when using bytemuck
- pipeline-overrideable constants
- push constants
Usage
Add the following lines to the Cargo.toml
and fill in the appropriate versions for wgsl_to_wgpu
.
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.
[build-dependencies]
wgsl_to_wgpu = "..."
See the example crate for how to use the generated code. Run the example with cargo run
.
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_to_wgpu 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_to_wgpu 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. It's strongly recommended to use types like vec4 or mat4 instead of vec3 or mat3 with bytemuck to avoid alignment mismatches. Structs used only as vertex input structs have their layout manually specified using std::mem::offset_of
and do not generate layout validation assertions.
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_to_wgpu 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.
- This library is not a rendering library and will not generate any high level abstractions like a material or scene graph. The goal is just to generate most of the tedious and error prone boilerplate required to use WGSL shaders with wgpu.
- The generated code will not prevent accidentally calling a function from an unrelated generated module. It's recommended to name the shader module with the same name as the shader and use unique shader names to avoid issues. Using generated code from a different shader module may be desirable in some cases such as using the same camera struct definition in multiple WGSL shaders.
- The current implementation assumes all shader stages are part of a single WGSL source file. Shader modules split across files may be supported in a future release.
- Uniform and storage buffers can be initialized using the wrong generated Rust struct. WGPU will still validate the size of the buffer binding at runtime.
- 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. This should be addressed by using some handwritten code where appropriate.
- Pipelines only define a single push constant range shared by all stages for simplicity. Set the data using a single call with offset 0 and the appropriate used stages like
wgpu::ShaderStages::VERTEX_FRAGMENT
.
Publishing Crates
Rust expects build scripts to not modify files outside of OUT_DIR. 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 wgsl_to_wgpu::{create_shader_module_embedded, WriteOptions};
// src/build.rs
fn main() {
println!("cargo:rerun-if-changed=src/model.wgsl");
// Generate the Rust bindings and write to a file.
let text = create_shader_module_embedded(wgsl_source, WriteOptions::default()).unwrap();
let out_dir = std::env::var("OUT_DIR").unwrap();
std::fs::write(format!("{out_dir}/model.rs"), text.as_bytes()).unwrap();
}
The generated code will need to be included in one of the normal source files. This includes adding any nested modules as needed.
// src/shader.rs
pub mod model {
include!(concat!(env!("OUT_DIR"), "/model.rs"));
}
Dependencies
~7–15MB
~201K SLoC