Crate sgpu_compute

SGPU-Compute - Simple GPU-Compute using WebGPU

This crate aims to provide a simple and easy-to-use interface to run compute shaders with WGPU and WGSL. It is designed to be as simple as possible to use, while still providing a lot of flexibility for performance reason.

Shader buffer

Three buffers are declared by default: - uniform for the uniform buffer - in for the input buffer - out for the output buffer Their types are inferred from the run method. The scratchpad buffer is also available, but it is not required.

Example

use sgpu_compute::prelude::*;

let my_shader = "
   @group(0) @binding(0) var<uniform> coefficient: u32;
   @group(0) @binding(1) var<storage, read> in: array<u32>;
   @group(0) @binding(2) var<storage, read_write> out: array<u32>;

   @compute
   @workgroup_size(8, 1, 1)
   fn main(@builtin(global_invocation_id) global_id: vec3<u32>) {
       out[global_id.x] = coefficient * in[global_id.x];
   }
";

const N_ELEMENT: usize = 64;
const WORKGROUP_SIZE: usize = 8;
const N_WORKGROUP: u32 = (N_ELEMENT / WORKGROUP_SIZE) as u32;

let gpu = GpuCompute::new();
let mut pipeline = gpu.gen_pipeline(
       None,
       [StageDesc {
           name: Some("norm"),
           shader: my_shader,
           entrypoint: "main",
       }],
   );

const COEFFICIENT: u32 = 42;

let input: [u32; N_ELEMENT] = std::array::from_fn(|i| i as u32);
pipeline.write_uniform(&COEFFICIENT);
let result_gpu = pipeline.run(&input, [(N_WORKGROUP, 1, 1)], |vals: &[u32; N_ELEMENT]| *vals);
let result_cpu = input.map(|v| v * COEFFICIENT);
assert_eq!(result_gpu, result_cpu);

Modules

blocking

prelude

Common re-exports for the crate.

Structs

GpuComputeAsync

This is the main struct of the library. It is used to create pipelines and run them. It requires an async runtime to work. If you want a blocking version, you can use the GpuCompute struct. If you don’t use the blocking version disable default features.

PipelineAsync

This struct represents a pipeline. It is used to run async compute shaders. To build it use the gen_pipeline method of the GpuComputeAsync struct.

StageDesc