I just started writing some code using the fp16 library, and I was just wondering if this is ever going to happen, because writing kernels that take advantage of the performance is barely better than writing assembly. Multiple thread indexes and comparison functions done in __Half2 are exponentially hard to do than __Half.
Comparison functions might be a bit better if we could choose what the output would be.
Example
__Half2 hlt2(__Half2 a, __Half2 b, __Half a_out, __Half b_out)