A Pre-routing Net Wirelength Prediction Method Using an Optimized Convolutional Neural Network

The total wirelength of a circuit implementation is an important metric to evaluate the quality of an FPGA design flow. The wirelengths of all nets of a circuit are determined by routing, but pre-routing stages like placement can use a wirelength prediction model to direct the generation of a placement solution with a shorter total wirelength for routing. The conventional VPR employs a wirelength prediction model based on the bounding box size and the number of sinks of a net, which works well for an FPGA of a regular 2D array structure. However, new FPGA architectures like 3D-FPGA and hierarchical routing cannot use such a simple model. In this work, we propose a method to build an optimized net wirelength prediction model using a convolutional neural network, which can learn routing features from routed nets without manual tunings. The evaluation results show an optimized CNN model also has higher accuracy than the VPR model.