Coverage-guided tensor compiler fuzzing with joint IR-pass mutation

In the past decade, Deep Learning (DL) systems have been widely deployed in various application domains to facilitate our daily life, e.g., natural language processing, healthcare, activity recognition, and autonomous driving. Meanwhile, it is extremely challenging ensure correctness of DL (e.g., due their intrinsic nondeterminism), bugs can cause serious consequences may even threaten human lives. literature, researchers explored techniques test, analyze, verify models, since quality directly affects corresponding system behaviors. Recently, also proposed novel for testing underlying operator-level libraries (such as TensorFlow PyTorch), which provide general binary implementations each high-level operator are foundation running models on different hardware platforms. However, there still limited work targeting reliability emerging tensor compilers (also known compilers), aim automatically compile computation graphs into high-performance binaries better efficiency, portability, scalability than traditional libraries. Therefore, this paper, we target important problem compiler testing, Tzer, a practical fuzzing technique used TVM compiler. Tzer focuses mutating low-level Intermediate Representation (IR) mutation space IR. More specifically, leverages both general-purpose tensor-compiler-specific mutators guided by coverage feedback diverse evolutionary IR mutation; furthermore, passes (i.e., transformations) optimization, performs pass tandem with more effective fuzzing. Our experimental results show that substantially outperforms existing 75% higher 50% valuable tests 2nd-best technique. Also, components validated via ablation study. To date, has detected 49 previously unknown TVM, 37 confirmed 25 fixed (PR merged).