[llvm] [NVPTX] Stop using 16-bit CAS instructions from PTX (PR #120220)

[Artem Belevich via llvm-commits]

Artem-B wrote:

TBH, I do not see much of an improvement if we use CAS explicitly: https://godbolt.org/z/deWcEcn3h

Can you elaborate on what the benefits are on the SASS level? In theory LLVM could potentially optimize explicit CAS a bit better, but considering that all of this is happening very late in the optimization pipeline, we end up with roughly the same SASS that ptxas generates from `atom.add.bf16`. 


https://github.com/llvm/llvm-project/pull/120220