Crate oxicuda_webgpu 

OxiCUDA WebGPU backend — cross-platform GPU compute via wgpu and WGSL.

Provides a WebGpuBackend that implements the oxicuda_backend::ComputeBackend trait from oxicuda-backend, using wgpu to target Vulkan, Metal, Direct3D 12, and the browser WebGPU API from a single Rust crate.

Quick Start

use oxicuda_webgpu::WebGpuBackend;
use oxicuda_backend::ComputeBackend;

let mut backend = WebGpuBackend::new();
backend.init().expect("WebGPU init failed");

// Allocate 1 KiB on the GPU.
let ptr = backend.alloc(1024).expect("alloc failed");
backend.free(ptr).expect("free failed");

Architecture

┌──────────────────────────────────────────┐
│          WebGpuBackend                   │
│  (implements ComputeBackend)             │
└──────────────┬───────────────────────────┘
               │
      ┌────────▼────────┐
      │  WebGpuDevice   │  ← wgpu Instance + Adapter + Device + Queue
      └────────┬────────┘
               │
   ┌───────────▼────────────┐
   │  WebGpuMemoryManager   │  ← buffer pool (u64 handle → wgpu::Buffer)
   └────────────────────────┘

WGSL Shader Generation

The shader module provides helpers that produce WGSL source strings for common kernels (GEMM, element-wise ops, reductions). The shader_ext module adds transpose, row-wise softmax, Blelloch prefix scan, layer normalisation, warp-style subgroup reductions, and emulated-f64 arithmetic, while the fft module provides a radix-2 Cooley-Tukey FFT plan (WgslFftPlan) plus its butterfly shaders.

Dispatch Planning

The planner module resolves workgroup sizes and dispatch grids from adapter Limits (auto-tuning, 2-D dispatch folding, 256-byte texture row alignment) — pure host-side arithmetic requiring no GPU.

Re-exports

pub use backend::WebGpuBackend;

pub use error::WebGpuError;

pub use error::WebGpuResult;

pub use fft::FftDirection;

pub use fft::WgslFftPlan;

pub use planner::DispatchGrid;

pub use planner::Limits;

pub use planner::Workgroup1D;

pub use planner::Workgroup2D;

pub use shader_ext::ScanKind;

Modules

backend

WebGpuBackend — the main entry point for the oxicuda-webgpu crate.

device

WebGPU device wrapper — owns the wgpu instance, adapter, device, and queue.

error

Error types for the oxicuda-webgpu backend.

fft

WGSL radix-2 Cooley-Tukey FFT (Cooley & Tukey 1965) — host plan + shader.

memory

WebGPU buffer manager — allocates, copies, and frees wgpu::Buffer objects through an opaque u64 handle interface that mirrors the CUDA device-pointer model used by the rest of OxiCUDA.

planner

Host-side dispatch and workgroup planning for WebGPU compute.

shader

WGSL shader source generation for common compute kernels.

shader_ext

Additional WGSL shader-source generators that extend crate::shader.

wasm

WASM target support for browser-based GPU compute via WebGPU.