Question: ScanTileState reliance on longlong2 vector type for reading / writing coherently.

[IlyaGrebnov]

ScanTileState (as well as ReduceByKeyScanTileState) currently rely on 128 bit longlong2 vector type for reading / writing descriptor state coherently without using __threadfence or splitting state into aggregate + status word. This is understandable for atomic types, but 128 bit types are not atomic. From "Parallel Thread Execution ISA Version 8.0":

8.2.3. Memory Operations on Vector Data Types The memory consistency model relates operations executed on memory locations with scalar data types, which have a maximum size and alignment of 64 bits. Memory operations with a vector data type are modeled as a set of equivalent memory operations with a scalar data type, executed in an unspecified order on the elements in the vector.

Similarly libcu++ suggests that "atomic_ref does not support types larger than 8 bytes."

So, how this currently works in CUB? Does SASS actually support 128 bit atomics, but it is not guaranteed by PTX and could break in future hardware?