[llvm] [AMDGPU] Use RTZ for newer fp16 interp instructions (PR #86235)

[Jay Foad via llvm-commits]

jayfoad wrote:

> > Why set the rounding mode when there are V_INTERP_P10_RTZ_F16_F32 and V_INTERP_P2_RTZ_F16_F32 which always use rtz?
> 
> I agree this is a little bit messy. As mentioned in previous discussion (https://reviews.llvm.org/D50633?id=160427#inline-445640), These default fp16 interp instructions need to run under rtz rounding mode to make output value in the expected range. I think this should be fixed. Meanwhile I agree it would be better we expose the rtz version and the frontend stop using the default version.

My preference would be to expose the RTZ versions now _instead_ of this patch.

https://github.com/llvm/llvm-project/pull/86235