간단논문 정리 TVM: An Automated End-to-End Optimizing Compiler for Deep Learning (OSDI 18)

제목

TVM: An Automated End-to-End Optimizing Compiler for Deep Learning

저자

Tianqi Chen and Thierry Moreau, University of Washington; Ziheng Jiang, University of Washington, AWS; Lianmin Zheng, Shanghai Jiao Tong University; Eddie Yan, Haichen Shen, and Meghan Cowan, University of Washington; Leyuan Wang, UC Davis, AWS; Yuwei Hu, Cornell; Luis Ceze, Carlos Guestrin, and Arvind Krishnamurthy

Motivation

 기존 머신러닝을 이용한 compiler optimizaion 방법에서는  human experts를 이용한다양한 hardward back-end(GPU,FPGA,ASIC)이 늘어남에 따라 그 구조에 적합한 complier optimization이 달라 질 수 밖에 없다.  

Contribution

해당논문은 머신러닝 High level Graph 연산을 ML 기반으로 특정 device 적합한 excutable 코드를 만들도록 수행하는 방법을 제시 

Content

Graph level modification & hareware-aware optimization 

• Operator Fusion 
  • Combines many small ops 
• Constant Folding 
  • Pre-computes static graphs  
• Static Memory Planning Pass 
  • Pre-allocates memory for needed tensors 
• Data Layout Transformations 
  • Optimize data storage for each backend  
• cost model에 ML을 이용 
  • Query에서 추출한  feature 를XGBoost 를 이용하여  costs 를 예측 

references

https://arxiv.org/abs/1802.04799

Project Page

https://tvm.apache.org