Optimizing Loop Operation and Dataflow in FPGA Acceleration of Deep Convolutional Neural Networks