Saved in:
Bibliographic Details
Main Authors: Wu, Mengdi, Cheng, Xinhao, Liu, Shengyu, Shi, Chunan, Ji, Jianan, Ao, Kit, Velliengiri, Praveen, Miao, Xupeng, Padon, Oded, Jia, Zhihao
Format: Preprint
Published: 2024
Subjects:
Online Access:https://arxiv.org/abs/2405.05751
Tags: Add Tag
No Tags, Be the first to tag this record!
Table of Contents:
  • We introduce Mirage, the first multi-level superoptimizer for tensor programs. A key idea in Mirage is $μ$Graphs, a uniform representation of tensor programs at the kernel, thread block, and thread levels of the GPU compute hierarchy. $μ$Graphs enable Mirage to discover novel optimizations that combine algebraic transformations, schedule transformations, and generation of new custom kernels. To navigate the large search space, Mirage introduces a pruning technique based on abstraction that significantly reduces the search space and provides a certain optimality guarantee. To ensure that the optimized $μ$Graph is equivalent to the input program, Mirage introduces a probabilistic equivalence verification procedure with strong theoretical guarantees. Our evaluation shows that Mirage outperforms existing approaches by up to 3.3$\times$ even for DNNs that are widely used and heavily optimized. Mirage is publicly available at https://github.com/mirage-project/mirage.