Automatic Compilation of CNN on Reconfigurable Array Processor Based on Scalable Instruction Group Library

Abstract

Although reconfigurable architecture is an inevitable choice for dealing with high-intensive applications, the opacity of hardware structure programming and the fuzziness or imprecision of many dependencies in applications make it difficult to develop applications and take full advantage of reconfigurable architecture. In order to address these challenges, this paper proposes an automatic compilation framework based on a scalable instruction group library, using convolutional neural network (CNN) applications as an example. The proposed framework first identifies functions such as data extraction, data distribution, data summary, and data processing, and divides the reconfigurable processors into multiple logical types. Next, the calculation modes are extracted by analyzing the CNN calculation process. Based on these calculation modes, efficient assembly instruction groups are designed, constituting the scalable instruction group library. Using this library, efficient mapping for CNN applications written in high-level languages can be easily realized on reconfigurable array processors. The experimental results demonstrate that the proposed method reduces the difficulty of programming and achieves better speedup performance compared with OpenMP and LLVM parallel compilation models. Moreover, the scalability of the instruction group library provides helpful guidance for the reconfigurable implementation of domain-oriented applications.

Keywords

• Reconfigurable architecture
• reconfigurable computing
• compilation techniques
• CNN