Reformulating the direct convolution for high-performance deep learning inference on ARM processors | Publicación