UC Santa Barbara
Hierarchical Transactions for Hardware/Software Cosynthesis
- Author(s): Arya, Kunal Arun
- Advisor(s): Brewer, Forrest
- et al.
Modern heterogeneous devices provide of a variety of computationally diverse components holding tremendous performance and power capability. Hardware-software cosynthesis offers system-level synthesis and optimization opportunities to realize the potential of these evolving architectures. Efficiently coordinating high-throughput data to make use of available computational resources requires a myriad of distributed local memories, caching structures, and data motion resources. In fact, storage, caching, and data transfer components comprise the majority of silicon real estate. Conventional automated approaches, unfortunately, do not effectively represent applications in a way that captures data motion and state management which dictate dominant system costs. Consequently, existing cosynthesis methods suffer from poor utility of computational resources. Automated cosynthesis tailored towards memory-centric optimizations can address the challenge, adapting partitioning, scheduling, mapping, and binding techniques to maximize overall system utility.
This research presents a novel hierarchical transaction model that formalizes state and control management through an abstract data/control encapsulation semantic. It is designed from the ground-up to enable efficient synthesis across heterogeneous system components, with an emphasis on memory capacity constraints. It intrinsically encourages a high degree of concurrency and latency tolerance, and provides verification tools to ensure correctness. A unique data/execution hierarchical encapsulation framework guarantees scalable analysis, supporting a novel concept of state and control mobility. A front-end language allows concise expression of designer intent, and is structured with synthesis in mind. Designers express families of valid executions in a minimal format through high-level dependencies, type systems, and computational relationships, allowing synthesis tools to manage lower-level details. This dissertation introduces and exercises the model, discussing language construction, demonstrating control and data-dominated applications, and presenting a synthesis path that exhibits near-linear scalability with problem size.