The Deep Learning Compiler: A Comprehensive Survey

The difficulty of deploying various deep learning (DL) models on diverse DL hardware has boosted the research and development compilers in community. Several have been proposed from both industry academia such as Tensorflow XLA TVM. Similarly, take described different frameworks input, then generate optimized codes for output. However, none existing survey analyzed unique design architecture comprehensively. In this paper, we perform a comprehensive by dissecting commonly adopted details, with emphasis oriented multi-level IRs, frontend/backend optimizations. Specifically, provide comparison among aspects. addition, present detailed analysis IRs illustrate optimization techniques. Finally, several insights are highlighted potential directions compiler. This is first paper focusing compilers, which hope can pave road future towards