계산과 데이터를 공동으로 최적화하는 딥 러닝 컴파일 프레임워크

Abstract

In recent years, deep learning algorithms and deep learning processors have been widely adopted in the industry. The challenge of fully leveraging the performance of deep learning processors from a software perspective has become a focal and difficult point in compiler research. Current deep learning compilation frameworks primarily emphasize optimizing the computational parts of programs, offering limited data optimization, which does not tap into the peak performance of deep learning processors. This paper analyzes the characteristics of deep learning algorithms and hardware platforms and introduces a deep learning compilation framework, JOCD, that co-optimizes computation and data. We evaluated JOCD on the FPGA platform. The experimental results show that for typical deep learning applications, the performance of models generated by JOCD can reach 86.5% of the performance of manually optimized models.