Feature request. Currently converting from float to half requires either sending the data to the GPU, or twiddling the bits yourself. Neither are rather savory. Why not include the convert_* functions on the host side? Theoretically they could be implemented as nice SIMD vector operations on the CPU too.
You should be able to get a high-performance version of just this by writing a kernel to run on the CPU that does the conversion and uses CL_USE_HOST_PTR for the input memory object to avoid an extra copy.