Model-driven optimisation of memory hierarchy and multithreading on GPUs