CUDA 2.0 Programming Guide states that “Atomic functions operating on shared memory and atomic functions operating on 64-bit words are only available for devices of compute capability 1.2 and above.”
On sm_13, I tried to use 64-bit version of atomicAdd() on shared memory, but got compile-time error:
/usr/local/cuda/bin/…/include/sm_12_atomic_functions.h(76): Error: Operand to 64-bit atomic instruction points at shared memory
Does that mean, that sm_13 atomic functions can be either 64-bit or operate on shared memory, but not together?
Can anyone clarify this?
update: my fault, must have read the guide more attentively.