I'm not sure what you mean. All CUDA code needs to be aware of the programming model the GPU imposes on them, splitting their code manually into threads and warps and blocks and kernels to match. This isn't really transparent at all.