This repository contains the source code for creating large-scale performance datasets, training and evaluating the LLMPerf model. The model is used to predict the performance of OpenCL kernels on a specific device.
The folder structure is as follows:
performance-model
: Source code for training and evaluating the LLMPerf model.cldrive
: Source code for automatically running OpenCL kernels on a specific device and collecting performance data.mem-access-analysis
: Source code for inserting memory access hook into OpenCL kernels, used to collect memory access traces and generate bound-aware datasets.
If you use this codebase, or otherwise find our work valuable, please cite our paper:
@inproceedings{nguyen2024llmperf,
title={LLMPerf: GPU Performance Modeling meets Large Language Models},
author={Nguyen-Nhat, Minh-Khoi and Do, Hoang Duy Nguyen and Le, Huyen Thao and Dao, Thanh Tuan},
booktitle={Proceedings of the International Symposium on the Modeling, Analysis, and Simulation of Computer and Telecommunication Systems},
year={2024},
organization={IEEE}
}