Optimizing CUDA Shared Memory Usage