NeuroSchedule2.0: A Novel GNN-Based Scheduling Method With RL-Based Preprocessing Optimization for High-Level Synthesis
Abstract
During high-level synthesis (HLS), scheduling is a critical step that determines the execution order of operations and directly affects circuit performance. However, existing scheduling methods suffer from either poor solution quality or limited scalability for complex designs containing hundreds of operations. This article introduces NeuroSchedule2.0, an efficient and effective GNN-based scheduling method that delivers both fast runtime and improved solution quality. NeuroSchedule2.0 is composed of a GNN-based list scheduling (LS) algorithm and a reinforcement learning (RL)-based preprocessing algorithm. Major features are as follows: 1) the learning problem for HLS scheduling is formulated for the first time, and a new machine learning framework is proposed; 2) pretraining models are adopted to further enhance the scalability for various scheduling problems with different settings; and 3) an RL-based preprocessing algorithm is proposed to optimize the topology of data-flow graphs for enhanced solution quality. Experimental results show that NeuroSchedule2.0 reduces the number of scheduled cycles by up to 8.41% on large real-life benchmarks compared with the state-of-the-art neural network-based scheduling method.
