Towards Fast and Energy-Efficient Binarized Neural Network Inference on FPGA