Skip to content

index.md

Latest commit

 

History

History
62 lines (43 loc) · 1.81 KB

index.md

File metadata and controls

62 lines (43 loc) · 1.81 KB
Raw

WebGPU Compute

日本語

WebGpuComputeProvider is the WASM-owned compute provider for browser/WebGPU execution. It implements AIKernel.Abstractions.Compute.IComputeProvider and keeps GPU-specific behavior outside Core.

Capability Surface

The provider exposes:

  • capability: webgpu.compute
  • operations:
    • compute.dispatch
    • compute.vector_add

The capability descriptor records adapter profile, backend name, provider ID, and deterministic metadata for host validation.

Backend Model

The provider separates backend binding from AIKernel compute semantics:

  • IWebGpuBackend defines adapter/device/queue operations.
  • WebGpuNativeBackend is the host-side native binding adapter boundary.
  • WebGpuWasmBackend is the browser/WASM backend and can use IWebGpuJsInterop.

The browser bridge owns WebGPU API calls such as buffer transfer, pipeline creation, and dispatch. The provider owns AIKernel contract validation and fallback routing.

CPU Fallback

When WebGPU is unavailable, compute falls back to AIKernel.Providers.Standard.Compute.CpuComputeProvider.

Fallback is deterministic and used by CI tests so the package can be verified without a physical GPU. Real browser GPU validation remains an E2E/browser test concern.

Vector Add Example

The built-in sample WGSL kernel supports vector addition:

@group(0) @binding(0) var<storage, read> a: array<f32>;
@group(0) @binding(1) var<storage, read> b: array<f32>;
@group(0) @binding(2) var<storage, read_write> out: array<f32>;

@compute @workgroup_size(64)
fn main(@builtin(global_invocation_id) gid: vec3<u32>) {
  let i = gid.x;
  out[i] = a[i] + b[i];
}

Events

Kernel execution can publish GpuKernelExecuted through IEventBus. The event contains deterministic metadata and does not expose browser implementation objects.