[Codegen] Port AMDGPU device lib implementations to MLIR rewrites

[qedawkins]

Overview

For accuracy and/or performance reasons, on AMDGPU backends we prefer device lib implementations of certain math functions (see #19970), for example:

https://github.com/ROCm/llvm-project/tree/amd-staging/amd/device-libs/ocml/src
Erf: https://github.com/ROCm/llvm-project/blob/amd-staging/amd/device-libs/ocml/src/erfF.cl

The erf implementation linked here is quite simple assuming a fast implementation of exp, which is a feature of AMD GPUs that makes this implementation fast. The same reasoning can apply to other targets assuming they provide a fast exp implementation.

Task

1. Port the above device lib implementations to MLIR patterns lowering math.* ops to scf + arith (+ more primitive math ops if needed).
2. Add a has_fast_exp pass option to MathTransformsPass that controls application of the new rewrites.

[benvanik]

+1 to being able to reason about efficient math in MLIR where we can vectorize and perform other magic!

[Megan0704-1]

Hello, I'm new to contributing to IREE and would like to work on this issue!