SHF: Medium: Automating High Level Synthesis via Graph-Centric Deep Learning

University Of California-Los Angeles, Los Angeles CA

Investigators

Abstract

Domain-specific accelerators (DSAs), such as those developed in recent years to accelerate deep learning applications, have been shown to offer significant performance and energy efficiency over general-purpose CPUs to meet the ever-increasing performance needs. However, DSAs are hard to design and require deep hardware and circuit-design knowledge to achieve high performance, which are lacking by most software programmers. Although the recent advances in high-level synthesis (HLS) tools made it possible to compile high-level software programs to circuit designs, one still needs to have extensive experience to perform microarchitecture optimizations by restructuring or augmenting the programs, which presents a significant barrier to a typical application-domain expert or software developer to design a DSA. The project aims to leverage machine learning and AI techniques to automate microarchitecture optimization and enable a typical software programmer to be able to design highly efficient hardware DSAs, with the quality comparable to those designed by well-trained circuit designers. As a result, it will enable wider and more rapid adoption of customized computing using DSAs to achieve significant improvement in computing efficiency. This project also plans to integrate the research with education to expose students to exciting opportunities in applying AI and ML techniques to electronic design automation, and broaden the participation in computing via high-school summer programs and partnership with the Center for Excellence in Engineering and Diversity (CEED) and Women in Engineering at UCLA. The project addresses two challenges in automating program transformation for HLS microarchitecture optimization: (1) the evaluation of each HLS design is time-consuming; and (2) the HLS design space is extremely large for brute-force search. The project develops a fully automated framework, named DeepAccel, for evaluating and optimizing the microarchitecture of a DSA design without the invocation of the time-consuming HLS tools. It represents the input C/C++ program as one or a set of graphs with the proper data-flow and control-flow information, including auto-inserted optimization directives (pragmas), and then makes use of the latest advances in graph-based machine learning (ML) and ML-driven optimizations to quickly evaluate each solution candidate and guide the optimization process. The approach is transformative, including the following research components: (1) the project tackles the fundamental representation problem on how to represent programs and associated transformations via graph-representation learning so one can apply the latest advances in deep learning, such as graph neural networks, knowledge distillation, meta-learning, and casual inferencing, to HLS design optimization; (2) the project designs trustworthy and adaptive deep-learning models for HLS performance prediction based on biased and sparsely labeled dataset; and (3) the project uses reinforcement learning and other scalable search algorithms to effectively cope with the combinatoric explosion of the search space. Based on these capabilities, DeepAccel is expected to automate the DSA design process for most performance-oriented software programmers. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.