What do you mean by “read atomically”?
Atomic operations perform read-modify-write cycles where the difference to ordinary operations is that no writes from other threads are allowed in between.
So I’m not quite sure what operation an “atomic read” by itself would perform.
Just reading a 32-bit word should be fine without atomic operations. As current GPUs don’t support misaligned access, 32-bit words are always read in a single transaction and there is no risk of an intermitting write from another thread.