LOCAL: Low-Complex Mapping Algorithm for Spatial DNN Accelerators

TL;DR

This paper introduces LOCAL, a low-complexity mapping algorithm for spatial DNN accelerators, significantly improving execution time and energy efficiency at the compiler level.

Contribution

The paper proposes a novel low-complexity mapping algorithm, LOCAL, optimized for compiler use in spatial DNN accelerators, with formal problem definition and simulation validation.

Findings

2x to 38x improvements in execution time

Lower energy consumption compared to previous dataflow mechanisms

Effective at the compiler level for mapping operations

Abstract

Deep neural networks are a promising solution for applications that solve problems based on learning data sets. DNN accelerators solve the processing bottleneck as a domain-specific processor. Like other hardware solutions, there must be exact compatibility between the accelerator and other software components, especially the compiler. This paper presents a LOCAL (Low Complexity mapping Algorithm) that is favorable to use at the compiler level to perform mapping operations in one pass with low computation time and energy consumption. We first introduce a formal definition of the design space in order to define the problem's scope, and then we describe the concept of the LOCAL algorithm. The simulation results show 2x to 38x improvements in execution time with lower energy consumption compared to previous proposed dataflow mechanisms.