Coverage-Guided Tensor Compiler Fuzzing with Joint IR-Pass Mutation

Explore a cutting-edge approach to tensor compiler testing in this 14-minute conference talk from ACM SIGPLAN's OOPSLA. Dive into the world of Deep Learning (DL) systems and the challenges of ensuring their correctness. Learn about Tzer, a novel fuzzing technique designed for the TVM tensor compiler that outperforms existing methods. Discover how Tzer mutates low-level Intermediate Representation (IR) and compiler passes to achieve higher coverage and generate more valuable tests. Gain insights into the experimental results, including the detection of 49 previously unknown bugs in TVM. Understand the importance of tensor compiler reliability and its impact on the efficiency, portability, and scalability of DL systems across various hardware platforms.