Dynamic Co-Optimization Compiler: Leveraging Multi-Agent Reinforcement Learning for Enhanced DNN Accelerator Performance

TL;DR

This paper presents DCOC, a multi-agent reinforcement learning-based compiler that significantly improves DNN deployment efficiency on hardware by optimizing both software and hardware configurations, leading to higher throughput and faster optimization.

Contribution

The paper introduces a novel multi-agent reinforcement learning framework for hardware/software co-optimization in DNN compilers, achieving superior performance and reduced search space.

Findings

Up to 37.95% throughput improvement.

Optimization time reduced by up to 42.2%.

Outperforms existing state-of-the-art methods.

Abstract

This paper introduces a novel Dynamic Co-Optimization Compiler (DCOC), which employs an adaptive Multi-Agent Reinforcement Learning (MARL) framework to enhance the efficiency of mapping machine learning (ML) models, particularly Deep Neural Networks (DNNs), onto diverse hardware platforms. DCOC incorporates three specialized actor-critic agents within MARL, each dedicated to different optimization facets: one for hardware and two for software. This cooperative strategy results in an integrated hardware/software co-optimization approach, improving the precision and speed of DNN deployments. By focusing on high-confidence configurations, DCOC effectively reduces the search space, achieving remarkable performance over existing methods. Our results demonstrate that DCOC enhances throughput by up to 37.95% while reducing optimization time by up to 42.2% across various DNN models, outperforming current state-of-the-art frameworks.