Towards Hardware-Software Co-Design For Energy-Efficient Deep Learning