Implementing Efficient, Portable Computations for Machine Learning