Towards improved parallelism through order reduction of accessing data in nD matrices.