Exploiting Hyper-Loop Parallelism in Vectorization to Improve Memory Performance on CUDA GPGPU