Implementing the Himeno benchmark with CUDA on GPU clusters