half precision constants in ptx

Hi all, in ptx a float constant is 0f???

Can I write a half2 constant?

Half precision literals are not integrated into CUDA at the moment, but a ticket has in place.

In the meantime, you can use

static_cast<__half>(2.0)

What about at the ptx level?

I think the answer is still no.

At the ptx level, I believe it is possible to first mov a constant into a register, and then do the hfma instruction. Although not sure what the SASS would be like and if it will use the constant cache in this case.