Efficient Memory Partitioning in Software Defined Hardware

02/02/2022

∙

by   Matthew Feldman, et al.

∙

Stanford University

∙

8

∙

share

As programmers turn to software-defined hardware (SDH) to maintain a high level of productivity while programming hardware to run complex algorithms, heavy-lifting must be done by the compiler to automatically partition on-chip arrays. In this paper, we introduce an automatic memory partitioning system that can quickly compute more efficient partitioning schemes than prior systems. Our system employs a variety of resource-saving optimizations and an ML cost model to select the best partitioning scheme from an array of candidates. We compared our system against various state-of-the-art SDH compilers and FPGAs on a variety of benchmarks and found that our system generates solutions that, on average, consume 40.3 78.3