[Mlir-commits] [mlir] [mlir][AMDGPU] Add support for AMD f16 math library calls (PR #108809)

[Daniel Hernandez-Juarez]

dhernandez0 wrote:

> Regarding the two f16 calls that you are adding where f32 calls are not present, I am fine with whichever choice makes sense for the hardware / AMDGPU backend but the following needs to be updated accordingly, https://github.com/llvm/llvm-project/blob/main/mlir/lib/Conversion/GPUToROCDL/LowerGpuOpsToROCDLOps.cpp#L338 It controls if for a given type we will make a llvm op or the library call. Otherwise which one happens may vary based on which pattern gets applied first or in the worst case we end up with the op not getting lowered at all and causing a crash.

Thanks for your review! I've updated LowerGpuOpsToROCDLOps.cpp as suggested.

https://github.com/llvm/llvm-project/pull/108809