Question

I’m looking at an application that does not require 32bit of precision, 12-16bit will be enough.

Is thrust capable of dealing with float16’s (i.e. packing/unpacking two 16bit precision floats into a 32bit word)?

Should I use fixed-point arithmetic?

Answer 1

CUDA hardware does not include native support for half-precision arithmetic, only conversion to and from float.

Since C does not have a built-in half type, the conversion intrinsics use unsigned short:

unsigned short __float2half_rn( float );
float __half2float( unsigned short );