stablehlo VS gpuweb

I am going to paste a cousin comment:

StableHLO[1] is an interesting project that might help AMD here:

> Our goal is to simplify and accelerate ML development by creating more interoperability between various ML frameworks (such as TensorFlow, JAX and PyTorch) and ML compilers (such as XLA and IREE).

From there, their goal would most likely be to work with XLA/OpenXLA teams on XLA[3] and IREE[2] to make RoCM a better backend.

[1] https://github.com/openxla/stablehlo

[2] https://github.com/openxla/iree

[3] https://www.tensorflow.org/xla

Also see the recently introduced StableHLO and its serialization format: https://github.com/openxla/stablehlo/blob/main/docs/bytecode...

If you mean StableHLO, then it has an MLIR dialect: https://github.com/openxla/stablehlo/blob/main/stablehlo/dia....

In the StableHLO spec, we are talking about this in more abstract terms - "StableHLO opset" - to be able to unambiguously reason about the semantics of StableHLO programs. However, in practice the StableHLO dialect is the primary implementation of the opset at the moment.

I wrote "primary implementation" because e.g. there is also ongoing work on adding StableHLO support to the TFLite flatbuffer schema: https://github.com/tensorflow/tensorflow/blob/master/tensorf.... Having an abstract notion of the StableHLO opset enables us to have a source of truth that all the implementations correspond to.

Works for me.

WebGPU support is behind a couple flags on Linux: https://github.com/gpuweb/gpuweb/wiki/Implementation-Status

People keep spreading this incredibly misleading statement, and yours is even more misleading (suggesting Apple opposed a 'GPU WASM')

By all accounts, Apple's /only/ stance was that if WebGPU used SPIR-V it would be a non-starter for them, due to ongoing legal issues between Apple and Khronos.

Apple actually proposed WebHLSL in collaboration with Microsoft, to have HLSL be the standard.

Mozilla employee's stance[0] was that SPIRV was too low level, did not fit with the goals of WebGPU portability and security, and expressed concern that Khronos may add functionality to SPIRV they cannot support in WebGPU like raytracing instructions .. 'So we'd always be on the verge of forking SPIR-V in some way.'

It was also noted by many people that even if a bytecode format was used, it would still have to be translated to the target (HLSL/DXIL, MSL, etc.) in almost the same way a text format would.

Nobody proposed a 'GPU WASM equivalent' or an alternative bytecode format.

The hard truth is that shader compilation is a fucking nightmare, people do not realize how bad it is across the different native APIs. SPIR-V is good, but it doesn't solve that - and presents other challenges if you are a web browser API. Vulkan and SPIRV are not the golden goose many make them out to be.

[0] https://github.com/gpuweb/gpuweb/issues/847#issuecomment-642...

Yes it is still a bit new. WebGPU is not finished and is still being worked on: https://webgpu.io/

WebGPU currently doesn't support the "bindless" resource access model (see: https://github.com/gpuweb/gpuweb/issues/380).

The "max number of sampled texture per shader stage" is a runtime device limit, and the minimal value for that seems to be 16. So texture atlasses are still a thing in WebGPU.

There's even an implementation under discussion to have the dp4a instruction added to WebGPU (https://github.com/gpuweb/gpuweb/issues/2677)