Abstract
The difficulty of deploying various deep learning (DL) models on diverse DLhardwares has boosted the research and development of DL compilers in thecommunity. Several DL compilers have been proposed from both industry andacademia such as Tensorflow XLA and TVM. Similarly, the DL compilers take theDL models described in different DL frameworks as input, and then generateoptimized codes for diverse DL hardwares as output. However, none of theexisting survey has analyzed the unique design of the DL compilerscomprehensively. In this paper, we perform a comprehensive survey of existingDL compilers by dissecting the commonly adopted design in details, withemphasis on the DL oriented multi-level IRs, and frontend/backendoptimizations. Specifically, we provide a comprehensive comparison amongexisting DL compilers from various aspects. In addition, we present detailedanalysis of the multi-level IR design and compiler optimization techniques.Finally, several insights are highlighted as the potential research directionsof DL compiler. This is the first survey paper focusing on the unique design ofDL compiler, which we hope can pave the road for future research towards the DLcompiler.