I believe numpy introduced ndarrays with user-specified strides, that made this a O(1) operation. Even if not, it certainly predates pytorch in this regard. Pytorch (and other tensor libraries) just used the ndarray concept from numpy.