Performance portability evaluation of blocked stencil computations on GPUs