Efficient Implementations of Sparse and Quantized Deep Neural Networks Using Systolic Arrays