Paper | Slides | Artifact

GenCoG is a DSL-based approach to generating computation graphs for TVM testing. It contains (1) GenCoGL, a domain-specific language for specifying type constraints of operators, and (2) an incremental generation approach with expressivity-directed strategy and concolic constraint solving.

• Implementation of GenCoG;
• Adapted version or reimplementation of baselines: LEMON , Muffin , Luo et al., and NNSmith.
• Bug list, triggering Python scripts and bug reports.

First, make sure TVM is installed. GenCoG works on v0.8 and v0.9, and it may also support later versions. Then, run pip install -r requirements/core.txt to get all dependencies of GenCoG.

If you want to run the experiments in the paper, run pip install -r requirements/exp.txt to install dependencies of the baselines.

Please first create a subdirectory out in the root directory of this project to store all the outputs.

python3 run_test.py

A working directory out/run-%Y%m%d-%H%M%S will be created. Each generated program will be run in a separate process. If the process exits abnormally, the test case will be kept and the error message will also be stored. Otherwise, the case will be deleted.

python3 dedup_case.py -d ${WORK_DIR}

It deduplicates the cases with similar error messages, which indicate that they may share the same root cause.

python3 reduce_case.py -d ${WORK_DIR}

It reduces each test case to a possibly simpler graph with fewer vertices.

Refer to files in gencog/op for how to write constraint specification for an operator and register it in OpRegistry.

Type constraints of operators in different DL compilers are possibly different. Some specifications may need to be rewritten.

A new code generator is also required for generating high-level IR for the new DL compiler, from the in-memory graph representation of GenCoG. gencog/graph/relay.py is the code generator for Relay. You can refer to this file to implement your own generator.

@inproceedings{wang2023gencog,
    author = {Zihan Wang, Pengbo Nie, Xinyuan Miao, Yuting Chen, Chengcheng Wan, Lei Bu, Jianjun Zhao},
    title = {GenCoG: A DSL-Based Approach to Generating Computation Graphs for TVM Testing},
    year = {2023},
    publisher = {ACM},
    address = {New York, NY, USA},
    doi = {10.1145/3597926.3598105},
    booktitle = {Proceedings of the 32nd ACM SIGSOFT International Symposium on Software Testing and 
    Analysis},
    numpages = {13},
    series = {ISSTA ’23}
}