[Mlir-commits] [mlir] [MLIR][ROCDL] Remove patterns for ops supported as intrinsics in the AMDGPU backend (PR #102971)

[Adrian Kuegel]

akuegel wrote:

This patch now breaks lowering for bf16 ExpOp or LogOp. For bf16, we upcast to F32, and now we have no lowering for F32 anymore.
Moreover I wonder whether it actually makes a difference in the final lowering? If there are intrinsics, shouldn't the other pattern also eventually lower to them?

https://github.com/llvm/llvm-project/pull/102971