As I mentioned this before, it is possible it doesn't work without modifying sgminer to work with it. Strange though, a user claimed it compiled the kernel for him.
This opencl code has 11 kernels instead of just one as in regular sgminer. The speed improvement comes from better use of SIMD by executing shorter kernels instead of a single long one. There's some memory overhead because of it, but it's largelly compensated by the speed improvement.