Skip to content

backlog.md

Latest commit

 

History

History
49 lines (38 loc) · 2.26 KB

backlog.md

File metadata and controls

49 lines (38 loc) · 2.26 KB
Raw

Status and remaining work

Available now

  • The generated gpu package exposes the portable WebGPU surface and xgpu's native extensions through one HashLink API.
  • Native releases use wgpu on Metal, D3D12, Vulkan, and Android OpenGL ES.
  • Promise-returning operations use ash.Future<T>.
  • Buffers, textures, pipelines, passes, query sets, render bundles, surfaces, error scopes, pipeline caches, mesh shaders, and ray-tracing descriptors are represented by the generated API.
  • HXSL render and compute shaders compile to WGSL at Haxe compile time.
  • Releases contain desktop hdlls, mobile static archives, committed Haxe externs, and an Ash-compatible browser side-module bundle.

Feature coverage of the shared API is tracked in xgpu. hlwgpu should not maintain a second operation or enum count.

How it is tested

crates/hlwgpu/test/browser/GpuTest.hx checks what the GPU hands back: compute and rendering by exact readback, copies, queries, and the errors and rejections a program observes through ash.Future. The gpu workflow runs it

  • on stock HashLink with ash_future.hdll (scripts/native_test.sh), on Linux with lavapipe for Vulkan and on Windows with WARP for D3D12;
  • on Ash for Metal, since stock HashLink's JIT does not run on arm64 macOS;
  • in headless Chrome through Ash's page on SwiftShader (scripts/browser_test.sh).

A runner that exposes no adapter reports SKIP rather than failing.

crates/hlwgpu/test/host holds xgpu.wasm to IMPORTS.md with a host that shares no code with Ash (scripts/host_test.sh): its own dylink loader, HashLink carriers, Future hooks and agent worker.

Programs on upstream HashLink take ash.Future from Ash's ash-future haxelib, whose release ZIP carries ash_future.hdll; see using hlwgpu.

Release maintenance

scripts/release_check.py fails CI and a release when xgpu-core, xgpu-backend and xgpu-bindgen are not pinned to one xgpu revision, or when hlwgpu tracks WebGPU IDL of its own; xgpu's vendored IDL is the API's input. CI fails when a build changes the committed haxe/gpu, so a new pin has to be regenerated and committed. A versioned release takes its Haxelib version from its tag, and its notes name the pinned xgpu and the IDL snapshot it carries.