support for half-precision/float16?

[dangmoody]

I've been thinking about this a bit. Would this even be worth it?

It could be a fair amount of work to even know whether or not this is worth doing.

The generator probably needs better benchmarking support to be able to tell if this would be faster.