Work

Dual Avenues: CGRA Mapping Using Zero-Suppressed Binary Decision Diagrams, and Register Optimization for High-Level Synthesis

Year: 2025
Type: dissertation
Abstract: The relentless advancement of modern applications, such as the Internet of Things (IoT), Artificial Intelligence (AI), and big data applications, has created an unprecedented demand for computing power under increasingly stringent time-to-market pressure, challenging traditional paradigms like Moore’s Law and Dennard Scaling. In response, both the academia and the industry have revitalized their focus in two distinct yet parallel trajectories: Coarse-Grained Reconfigurable Arrays (CGRAs) and High-Level Synthesis (HLS). These technologies address the pressing need for scalable computational architectures and efficient design methodologies capable of adapting to evolving technology landscapes without incurring the traditional costs of hardware design. The first part of this thesis focuses on CGRAs, presenting a novel enumeration-based methodology for mapping that leverages the power of Zero-Supressed Binary Decision Diagrams (ZDDs). This approach significantly broadens the search horizon for feasible mappings, transcending traditional greedy search techniques by systematically exploring tens or hundreds of thousands of potential solutions within seconds, delivering a better understanding of the richness of the solution space. In the realm of HLS, this work pioneers a linear time solution for interprocedural register allocation, achieving efficient runtime with near-optimal register efficiency. Additionally, it develops an innovative Integer Linear Programming (ILP) formulation for register-pressure-aware scheduling, optimally solvable in polynomial time, which significantly reduces register count and design area without compromising performance. Together, these contributions tackle critical challenges in HLS optimization, spanning both compilation and hardware efficiency. Overall, the dissertation not only offers solutions to specific challenges in CGRA mapping and HLS, but also proposes new methodologies that can be generalized for broader applications in EDA and beyond for efficient support of the growing complexity and demands of future technological advancements.
Author: Beidas, Rami
Language: English
Open Access status: green
OpenAlex ID: W7133025759
Cites: 174
Cited by:
Related to: 10