Look at kernel double-precision floats (cl_khr_fp64 option), use if possible. Look for platforms that DO NOT have this option, and see how that works.