OpEvo: An Evolutionary Method for Tensor Operator Optimization

Authors: Xiaotian Gao, Wei Cui, Lintao Zhang, Mao Yang12320-12327

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our comprehensive experiment results show that compared with state-of-the-art (SOTA) methods Op Evo can find the best configuration with the lowest variance and least efforts in the number of trials and wall-clock time.
Researcher Affiliation Industry Xiaotian Gao, Wei Cui, Lintao Zhang, Mao Yang Microsoft Research xiaotian.gao, weicu, lintaoz, maoyang@microsoft.com
Pseudocode Yes The Op Evo algorithm is summarized in Algorithm 1.
Open Source Code Yes All code of this work is available online.
Open Datasets No Three different Mat Mul operators are chosen from BERT to evaluate proposed method. ... A classic CNN architecture, Alex Net, is used to evaluate the end-to-end performance of the proposed method...
Dataset Splits No The paper describes experimental runs and evaluation metrics but does not specify traditional train/validation/test dataset splits as it focuses on optimizing tensor operators rather than training a machine learning model on a fixed dataset.
Hardware Specification Yes on both Nvidia (GTX 1080Ti) and AMD (MI50 GPU) platforms.
Software Dependencies Yes compiled and run with CUDA 10.0 or RCOM 2.9.
Experiment Setup Yes Op Evo has two important hyperparameters, the mutation rate q which trade-offs the exploration and exploitation and the parent size λ which governs the diversity in the evolutionary process. In this part, we evaluate Op Evo with different q and λ for better understanding of each introduced technique and the hyperparameter sensitivity.