[Mlir-commits] [mlir] [mlir][gpu] Add pass for emulating unsupported types. (PR #138087)

[Md Abdullah Shahneous Bari]

mshahneo wrote:

> High-level question: why is this in GPU?
> 
> Shouldn't the translation from bf16 to i16 either be a pass over on Arith or part of the SPIR-V lowering? See also, when going to LLVM, we replace all the 8-bit float types with i8

Hi Krzysztof(@krzysz00),

The reason I wanted keep it in GPU instead of a part of SPIR-V lowering is that, hopefully it can used by other targets as well, not just SPIR-V.

As for doing it as part of AithToSPIRV, it would not work, since, it requires modification to both host and device code. 

That's why to me, gpu seemed like the right place.

https://github.com/llvm/llvm-project/pull/138087