[Mlir-commits] [mlir] [mlir][ArithToAMDGPU] Only use the scaled conversions for f8E8M0FNU scales (PR #221386)

[Hung-Kuan Tseng]

Tim096 wrote:

@krzysz00 this is the "reject non-f8E8M0 scales in ArithToAMDGPU, then fold away the truncation" you described on #215295, and the AMDGPU side of #217892. @tgymnich (the original lowering) @Muzammiluddin-Syed-ECE @kuhar FYI. One question for you: the non-E8M0 scales are now left for `arith-expand` rather than expanded to a multiply or divide inside this pass; if you would rather have the pass emit the arith ops itself so no pipeline change is needed, I can do that here. Would you mind taking a look when you have a moment? Thanks.

https://github.com/llvm/llvm-project/pull/221386