Search for: All records

Spatial accelerators provide high performance, energy efficiency, and flexibility. Recent design frameworks enable these architectures to be quickly designed and customized to a domain. However, constructing a compiler for this immense design space is challenging, first because accelerators express programs with high-level idioms that are difficult to recognize. Second, it is unpredictable whether certain transformations are beneficial or will lead to infeasible hardware mappings. Our work develops a general spatial-accelerator compiler with two key ideas. First, we propose an approach to recognize and represent useful dataflow idioms, along with a novel idiomatic memory representation. Second, we propose the principle of modular compilation, which combines hardware-aware transformation selection and an iterative approach to handle uncertainty. Our compiler achieves 2.3× speedup, and 98.7× area-normalized speedup over high-end server CPU. 

As Machine Learning (ML) applications become pervasive and computer architects further integrate hardware support, the need to rapidly explore trade-offs between algorithms and hardware becomes pressing. While prior work on hardware accelerators has led to tremendous performance and energy improvements, it can be difficult to generalize these approaches without resorting to special-purpose tools or even languages. Through object-oriented design principles, we describe a general and reusable approach for generating parameterized neural network hardware. Specifically, we describe our experiences with high-level hardware design objects for building neural network hardware based on the open-source Python HDL, PyRTL. By thinking at a higher level of abstraction than simple “hardware modules,”, we open the door to a process by which hardware can be developed with software engineering principles. This creates new opportunities for a tight feedback loop between machine learning algorithm innovation and hardware design reality. Future works considering hardware development for ML applications can benefit from our work analyzing the costs and benefits of abstraction.