Generating Efficient Tensor Contractions for GPUs