Using Fermi architecture knowledge to speed up CUDA and OpenCL programs