64-bit integer atomic instruction to shared memory

Hi,

I found that the PTX 2.0 documentation says, in page 136, that Fermi (CC2.0) supports 64-bit integer atomic instruction to shared memory (e.g., atom.shared.add.u64), but the CUDA 3.0 Programming Guide says otherwise in section B.10.1.1. I tried this new operation on GTX480 and it did appear to work fine. Anyone knows if this is just a typo in the programming guide or something done intentionally?

Thanks,

David