Some of the trig functions have tolerance issues for fp16. It sounds like we just need to adjust the tolerance for fp16 as it is currently defaulted to .0008, but ...