Atomics can work on up to a 64-bit (properly-aligned) quantity at a time. So you cannot do an atomic add on a float4 considering the entire float4 quantity - that is a 128 bit quantity. Furthermore the atomic engine doesn’t know anything about elementwise addition for a vector type.
The only types supported by atomicAdd are those listed in the programming guide. And if you wanted to perform multiple additions across various elements of a vector type, each of those additions would need to be handled by a single atomicAdd instruction.
You could possibly construct your own custom Atomic to perform an atomic addition on two of the quantities at a time. I think that is unlikely to provide better performance than 2 native atomics, however.