Proj CDeepFuzz Paper Reading: NNSmith: Generating Diverse and Valid Test Cases for Deep Learning Compilers
Abstract
背景：Deep-learning compiler例如TVM和TensorRT被使用来optimize DNN模型以达到更好的性能要求；在Deep Learning Compiler中的bug可能导致语义改变
本文：NNSmith
Task: fuzz deep-learning compilers
Method:
- using light-weight operator specifications to generate test models
- gradient-search to find model inputs that avoid any floating-point exceptional values to remove false alarm
- differential testing between libraries
实验：
对象：TVM, TensorRT, ONNXRuntime, PyTorch
目前支持：tvm, pt2, torchjit, tensorrt, on runtime, xla, tflite
效果：detect 72 bugs, 58 confirmed, 51 fixed
