Proj CDeepFuzz Paper Reading: Coverage-Guided Tensor Compiler Fuzzing with Joint IR-Pass Mutation

Abstract

背景：

  1. 目前已有工作test underlying operator-level DL Libraries，但测试compier的工作比较少
  2. compiler: compile high-level tensor computation graphs into high-performance binaries
  3. high-level IR has limited mutation space

本文：TZER
task: mutating the low-level IR for TVM tensor compiler
Method: 同时利用general purpose和tensor-compiler-specific mutators mutators guided by coverage feedback, pass mutation
Bug Type: Differential Testing with/without O0 binary(inconsistency, worse performance), Runtime Failure(Crash, Exception)

实验：
Competitor: TVMFuzz, LibFuzzer, LEMON
效果：

  1. +75% coverage, +50% valuable tests than 2nd-based tech
  2. use ablation study to validate different components of TZER
  3. +49 unknown bugs, + 37 bugs confirmed, + 25 fixed

posted @ 2023-08-06 20:33  雪溯  阅读(8)  评论(0)  编辑  收藏  举报