Were RNNs All We Needed? A GPU Programming Perspective