Presentation
DAPO: Design Structure Aware Pass Ordering in High-Level Synthesis with Graph Contrastive and Reinforcement Learning
DescriptionHigh-Level Synthesis (HLS) tools have become increasingly popular for facilitating the design of domain-specific accelerators (DSAs). However, the quality of the final design produced by HLS tools highly depends on the optimization pass sequence during compilation. Finding an optimally tailored pass sequence for each design, commonly known as the pass ordering problem, is NP-hard. Traditional heuristic approaches such as Simulated Annealing and Greedy Search require substantial computational time and effort to yield satisfactory solutions for individual designs. Machine learning methods that employ reinforcement learning (RL) or graph neural networks (GNNs) often struggle with poor generalization due to inefficient representations of program features.
To address the aforementioned challenges, we first propose our pass-order-oriented graph---a heterogeneous graph that effectively captures both semantic and structural features, offering a more informative representation of the input design compared to traditional graphs. Next, we introduce a technique for generating representative program embeddings using contrastive learning, which enhances the GNN's ability to generalize across different designs by learning the distinctions between HLS programs. Since the middle-end of commercial tools are inaccessible to users, hindering pass ordering exploration, we enhance Light-HLS, a lightweight open-source HLS tool that provides accurate and rapid latency results and synthesizes the input HLS design into Verilog code. By integrating these methods within a reinforcement learning flow, we propose the DAPO framework for pass order optimization, which achieves an average performance improvement of 1.8x compared to the SOTA academic tool AutoPhase and a 1.2x improvement w.r.t. the -O3 optimization level in LLVM-18.1.0 while saving 4.36x compilation time. We perform cross-validation on 80 real-world HLS designs showcasing the generalizability of the DAPO's inference model.
To address the aforementioned challenges, we first propose our pass-order-oriented graph---a heterogeneous graph that effectively captures both semantic and structural features, offering a more informative representation of the input design compared to traditional graphs. Next, we introduce a technique for generating representative program embeddings using contrastive learning, which enhances the GNN's ability to generalize across different designs by learning the distinctions between HLS programs. Since the middle-end of commercial tools are inaccessible to users, hindering pass ordering exploration, we enhance Light-HLS, a lightweight open-source HLS tool that provides accurate and rapid latency results and synthesizes the input HLS design into Verilog code. By integrating these methods within a reinforcement learning flow, we propose the DAPO framework for pass order optimization, which achieves an average performance improvement of 1.8x compared to the SOTA academic tool AutoPhase and a 1.2x improvement w.r.t. the -O3 optimization level in LLVM-18.1.0 while saving 4.36x compilation time. We perform cross-validation on 80 real-world HLS designs showcasing the generalizability of the DAPO's inference model.
Event Type
Networking
Work-in-Progress Poster
TimeMonday, June 236:00pm - 7:00pm PDT
LocationLevel 2 Lobby