Tensor Contractions with Extended BLAS Kernels on CPU and GPU