Intel HD Graphics 4000 (OpenCL issues)

It seems Intel's GPU cannot effectively handle such constructions:
A = (short16)value.

AMD can optimize it very effectively, Intel CPU OpenCL also produce good work, but HD4000 works very slow with such instructions in kernel.

Although if we will use data stored in __constant memory (like this short16 A={32,32,32...32}), HD4000 will work significantly faster, but all other OpenCL devices will be slower.