Batched Matrix Computations on Hardware Accelerators