In OpenCL 1.2 only the atomic_xchg operation ist supported for single precision floating-point data type. atomic_add, atomic_sub, ...etc should be supported for single precision floating-point data type too.
Thanks