That looks to be just for converting to and from half-precision floats.