[Epic] Add support for WebGPU

[ScottTodd]

Tracking issue for overall WebGPU support in IREE.

Background

WebGPU is a recently released API for running graphics and compute workloads on Graphics Processing Units in web browsers. It is similar to modern graphics/compute APIs like Vulkan, D3D12, and Metal, making it a promising target for IREE.

At a high level, we need:

• IREE's compiler to generate WGSL code
• IREE's runtime to have a WebGPU HAL implementation and general web platform support

The details can get tricky though:

• WGSL is more restrictive than SPIR-V
• WebGPU's compute API is more restrictive than Vulkan
• Runtime support requires compiling from C to JavaScript/WebAssembly (e.g. via Emscripten and some libraries)

Task breakdown

Compiler tasks

• [x] https://github.com/openxla/iree/issues/7840
• [ ] https://github.com/openxla/iree/issues/10906
• [x] https://github.com/openxla/iree/issues/10908
• [x] https://github.com/openxla/iree/issues/10142
• [x] https://github.com/openxla/iree/issues/10145
• [x] https://github.com/openxla/iree/pull/10224
• [ ] https://github.com/openxla/iree/issues/11054
• [x] https://github.com/openxla/iree/issues/11229
• [ ] https://github.com/openxla/iree/issues/11321
• [ ] https://github.com/openxla/iree/issues/12509
• [x] https://github.com/openxla/iree/pull/13965

Runtime tasks

• [x] https://github.com/openxla/iree/pull/13952
• [x] https://github.com/openxla/iree/pull/13669
• [ ] https://github.com/openxla/iree/issues/13809
• [x] https://github.com/openxla/iree/pull/13962
• [ ] Web Workers for threading (optional)
• [ ] Implement semaphores and use for asynchronous queue execution callbacks

Infrastructure / developer experience tasks

• [x] WebGPU/WGSL compiler builds and tests running on presubmit
• [x] General Wasm/Emscripten runtime build running on postsubmit
• [ ] WebGPU runtime build running on postsubmit
• [ ] All web builds running on presubmit (move out of experimental/)
• [ ] Write async unit tests, runnable in some way
• [ ] Integrate tests into CI
• [ ] Document how to use Wasm/WebGPU builds

Other related issues

• [ ] https://github.com/openxla/iree/issues/6499
• [ ] https://github.com/openxla/iree/issues/8327

# Status

last updated 2023-08-03

• All code is merged onto the main branch
• The application at experimental/web/sample_webgpu can programs in web browsers without crashing but with incorrect results
  • multiple_results.mlir: absf(-1.23, -4.56) returns (4.56, 0) at first, then (1.23, 4.56) on future calls with the same loaded program state
  • mobilebert, posenet, mobilessd: no errors, correct number of outputs, incorrect numerical results (way off llvm-cpu/local-task, stable across runs)
• A few compiler options are still needed to work around codegen feature gaps

Previous status reports:

2023-05-26

• The webgpu branch hosting the runtime HAL and a sample webpage (.c, .js, Emscripten) is close to main
  • https://github.com/openxla/iree/pull/13815 gets it ready for merging into experimental/
• The application at experimental/web/sample_webgpu can run simple_abs, fullyconnected, and MobileSSD using WebGPU in web browsers
  • tested Chrome 113 on Windows and Chrome Dev on Linux with Vulkan launch flags
  • Posenet, MobileBERT, and other programs with more than one output hit https://github.com/openxla/iree/issues/13809
• https://github.com/openxla/iree/pull/13820 will async-ify experimental/web/sample_webgpu/ using iree_loop_emscripten
• A few compiler options are required as workarounds, depending on the input program. Currently these are --iree-codegen-gpu-native-math-precision=true and alias-mutable-bindings (no tool flag, edit in source)

2022-12-14

IREE WebGPU Status 2022-12-14

2022-10-04

IREE WebGPU Status 2022-10-04