RL-Guided Runtime-Constrained Heuristic Exploration for Logic Synthesis

Too Big to Fail? Active Few-Shot Learning Guided Logic Synthesis

“…A blackbox optimizer can cover in a limited time. Hence, although RL-based approaches have shown promise in generating good quality synthesis recipes, they cannot scale up to industrialsized benchmarks [7], [8]. These observations indicate the following key questions computationally-efficient ML-enhanced logic synthesis:…”

“…In this work, we adopt a GCN-based architecture for the predictor (Figure 1); GCNs have been used in RL-based approaches as well [7], [8], but with different architectures. Our predictor architecture is shown in Figure 1a and contains two parallel paths.…”

“…Traditionally, EDA engineers choose from a set handcrafted synthesis recipes, for instance, resyn2 or compress2rs [4] based on intuition and experience. Yet, the space of all synthesis recipes is vast; recent work has shown ML-guided exploration of this vast solution space can yield synthesis recipes that are tailored for each design and outperform handcrafted recipes in terms of quality-of-result (QoR) [5]- [8]. In these papers, finding a good synthesis recipe is formulated as a classification problem [5]-which seeks to distinguish good "angel" recipes from bad "devil" recipes-or as in [6]- [8] as an Markov decision process (MDP) that is solved using reinforcement learning (RL).…”

“…Yet, the space of all synthesis recipes is vast; recent work has shown ML-guided exploration of this vast solution space can yield synthesis recipes that are tailored for each design and outperform handcrafted recipes in terms of quality-of-result (QoR) [5]- [8]. In these papers, finding a good synthesis recipe is formulated as a classification problem [5]-which seeks to distinguish good "angel" recipes from bad "devil" recipes-or as in [6]- [8] as an Markov decision process (MDP) that is solved using reinforcement learning (RL). However, the prior work is deficient on two fronts: (1) scalability to large designs; and (2) generalizability to previously unseen netlists.…”

“…Bulls-Eye produces high-quality synthesis recipes for benchmarks on which SOTA methods fail or time-out. -MAB 30003 [7] AIG+state encoding GCN, RL 2675 [8] AIG+state encoding GCN,RL 32060 Bulls-Eye AIG+recipe encoding GCN,CNN 114771 † DRiLLS claims the model can be reused, without empirical evaluation compared to conventional approaches. Compared to prior SOTA, Bulls-Eye achieves better QoR with an average 1.04x ( vs. [8]) and 11.21x (vs. [7]) run-time speedup.…”