Year: 2024
Type: article
Abstract: With the recent evolution of parallel computing and compilation techniques, deep-learning (DL) models have been increasingly deployed with DL compilers for achieving real-time performance in resource-constraint environments.Because the compiler stack is complicated and early drafted, debugging DL compilers is hard, especially when the time and amount of compiler experts are currently limited.What is worse, the massive and often duplicated (i.e., fuzzing bugs) bug reports can include obscure or even no (i.e., semantic bugs) error messages for a model of thousands of layers, making it extremely challenging to understand and localize the compiler failure.To address this emerging challenge, we propose new techniques for automatically reducing and deduplicating DL compiler bug reports: (i) We define the testcase reduction problem in DL compilers for the general graph-level inputs and propose subgraph minimality as the reduction target; (ii) based on the Delta Debugging framework, we perform validity-preserving graph partition guided by heuristics from error-prone patterns; and (iii) based on graph matching, we deduplicate test-cases with graph isomorphism and further make short-cuts to reduction by matching minimized graphs over upcoming ones.We integrated our techniques into a new tool, NNReduce, the first graph-level reducer with sub-graph minimality guarantee, implemented for ONNX model formats.We applied NNReduce to reduce and deduplicate both fuzzer-generated and real-world bug reports as a new dataset for extensive evaluation.Our results show that NNReduce can reduce testcases by up to 3.14× (fuzzing bugs) and 90.9× (real-world bugs) smaller than the existing half-automated reduction tool (i.e., Polygraphy by NVIDIA).NNReduce also improves error-message-based deduplication by achieving an up to 2.36× smaller distinct set of bug reports.
Author: Liu, Jiawei
Language: English
Subfield: Software
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7220773604
