I don't really agree with the last part. Since Perim is much much more likely to be non-zero preloading a 0 radius is probably a waste of work. I don't know if the other statement has a similar probablility but it is likely it does. Besides, this is all academic talk, most of what is described should (and probably will) be done by the compiler anyway. The only thing that can cause a real difference is the use of "const" with const values and const methods. The compiler could optimize some of it even to an immediate value then and skip memory reads completely.

But all these optimizations are nothing compared to implementing your calculations in OpenCL. Then it doesn't matter if your cpu is 32 or 64 bit.