Proj CDeepFuzz Paper Reading: A Comprehensive Study of Deep Learning Compiler Bugs

Abstract

背景：深度学习编译器处理的深度学习模型与命令式程序有根本的不同，因为深度学习模型中的程序逻辑是隐式的。（the DL models processed by DL compilers differ fundamentally from imperative programs in that the program logic in DL models is implicit）
本文：TVMFuzz
Github: https://github.com/ShenQingchao/DLCstudy
Task: 1. Review Deep Learning Compiler Bugs, including root causes, symptoms, occurred stages 2. TVMFuzz: Test TVM DL Compiler based on TVM's original test suite
Method:

  1. 分析TVM, Glow和nGraph中的603个错误
  2. Test TVM DL Compiler based on TVM's original test suite

Discussions:

  1. 12项发现，例如很大一部分（近 20%）的 DL 编译器错误与类型有关，尤其是张量类型。

效果：

  1. +8 TVM bugs
posted @ 2023-09-08 04:17  雪溯  阅读(13)  评论(0)  编辑  收藏  举报