A General Framework for Efficient Logic Synthesis

Logic synthesis synthesize circuit structures to optimize specific targets given reasonable constraints and runtime, utilizing a set of well-defined operators. Most existing synthesis operators are designed heuristically with a general paradigm of traversing circuit cuts or sub-graphs and making local replacements, when certain transformations offer synthesis gains. These transformations are usually obtained via invoking certain synthesis engines and can be time-consuming. However, most of these computations are redundant and lead to no further replacements. The efficiency of these operators is crucial for improving runtime and optimization convergence. Therefore, we propose a general framework for efficient logic synthesis that prunes unnecessary logic transformation and gain measurements, while maintaining synthesis effectiveness. To address the diverse needs of different operators, considering factors such as model inference time, ease of deployment, and interpretability, we propose several methods tailored to specific scenarios and unify them into a general framework. The framework is validated through experiments on both public and industrial circuits. For example, the expertise based scoring method can accelerate drw by about 35% with negligible effectiveness loss, and a GNN-based method achieves up to 3.1x faster runtime of mfs2.