Compile a KernelsMem program to imperative code with kernels. This is mostly (but not entirely) the same process no matter if we are targeting OpenCL or CUDA. The important distinctions (the host level code) are introduced later.