FLIP: An Artificial Neural Network-based Post-routing Incremental Placer

Pre-routing placement algorithms cannot fully exploit opportunities for wirelength reduction because they work with wirelength estimates that can be inaccurate. Hence, we need to employ post-routing incremental placers to refine placement and improve figures of merit. This paper proposes an artificial neural network (ANN) based post-routing incremental placer, FLIP (Fast Learning-based Incremental Placer), to obtain significant reductions in wirelengths in a given runtime. We train an ANN model to predict an optimum location for a cell in a given $10\times 10$ grid using a novel method of synthetically generated dataset in which 50% points are guaranteed to be optimum. Furthermore, we employ this ANN model to perform incremental placement for the entire layout using a sliding window method. We demonstrate the effectiveness of the proposed technique on ICCAD 2020 benchmark designs. For a given runtime, the proposed placer delivers $36\times $ more wirelength reduction than the deterministic incremental placer and $3\times $ more wirelength reduction than the random incremental placer.