MLGO: a Machine Learning Guided Compiler Optimizations Framework

Trofin, Mircea; Qian, Yundi; Brevdo, Eugene; Lin, Zinan; Choromański, Krzysztof; Li, David

doi:10.48550/arxiv.2101.04808

Cited by 14 publications

(21 citation statements)

References 22 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Of late, there has been considerable interest in applying deep learning techniques to compilers in the areas of phase ordering [17], selection of optimization heuristics [10] and as part of optimization itself in register allocation [13] and inlining [28]. Machine learnt models have been used in optimization heuristics selection such as prediction of unroll factors [26], inlining decisions [25], vectorization [15], [22] etc.…”

Section: Related Workmentioning

confidence: 99%

Learning to Combine Instructions in LLVM Compiler

Mannarswamy¹,

Das²

2022

Preprint

View full text Add to dashboard Cite

Instruction combiner (IC) is a critical compiler optimization pass, which replaces a sequence of instructions with an equivalent and optimized instruction sequence at basic block level. There can be thousands of instruction-combining patterns which need to be frequently updated as new coding idioms/applications and novel hardware evolve over time. This results in frequent updates to the IC optimization pass thereby incurring considerable human effort and high software maintenance costs. To mitigate these challenges associated with the traditional IC, we design and implement a Neural Instruction Combiner (NIC) and demonstrate its feasibility by integrating it into the standard LLVM compiler optimization pipeline.NIC leverages neural sequence-to-sequence (Seq2Seq) models for generating optimized encoded IR sequence from the unoptimized encoded IR sequence. To the best of our knowledge, ours is the first work demonstrating the feasibility of a neural instruction combiner built into a full-fledged compiler pipeline. Given the novelty of this task, we built a new dataset for training our NIC neural model. We show that NIC achieves exact match results percentage of 72% for optimized sequences as compared to traditional IC and neural machine translation metric Bleu precision score of 0.94, demonstrating its feasibility in a production compiler pipeline.

show abstract

Section: Related Workmentioning

confidence: 99%

Learning to Combine Instructions in LLVM Compiler

Mannarswamy¹,

Das²

2022

Preprint

View full text Add to dashboard Cite

show abstract

“…Unlike SL algorithms which train on large static datasets, RL algorithms often dynamically collect data throughout training 2 . As discussed further in Section 7, select previous works approach compiler autotuning using DRL [18,19,20,21]. Here, states are derived from characteristics of program code, actions are applied code optimizations or heuristic settings, and rewards are derived from performance measurements.…”

Section: Reinforcement Learningmentioning

confidence: 99%

“…Previous works have begun to explore the use of reinforcement learning for compiler optimization [18,19,20,21,31,32]. Huang et al [19] optimize compiler HLS phase ordering using cycle count reduction as a reward signal to guide learning using a framework they call AutoPhase.…”

Section: Related Workmentioning

confidence: 99%

“…Haj-Ali et al [18], drive the vectorization pass with pragma directives to inform a LLVM compiler heuristic and build an automated feature extraction framework to represent source code using a framework they call NeuroVectorizer. Trofin et al [21] formulates the inlining-for-size problem as a MDP, using native size reduction as a reward signal to guide an agent to learn the optimal policy for a production LLVM compiler, using a framework they call MLGO. Aside from [31], which uses imitation learning, each of these previous works use an on-policy reinforcement learning strategy.…”

Section: Related Workmentioning

confidence: 99%

See 1 more Smart Citation

Generating GPU Compiler Heuristics using Reinforcement Learning

Colbert¹,

Daly²,

Rubin³

2021

Preprint

View full text Add to dashboard Cite

GPU compilers are complex software programs with many optimizations specific to target hardware. These optimizations are often controlled by heuristics hand-designed by compiler experts using timeand resource-intensive processes. In this paper, we developed a GPU compiler autotuning framework that uses off-policy deep reinforcement learning to generate heuristics that improve the frame rates of graphics applications. Furthermore, we demonstrate the resilience of these learned heuristics to frequent compiler updates by analyzing their stability across a year of code check-ins without retraining. We show that our machine learning-based compiler autotuning framework matches or surpasses the frame rates for 98% of graphics benchmarks with an average uplift of 1.6% up to 15.8%.

show abstract

“…There is a growing body of work that shows how the performance and portability of compiler optimizations can be improved through autotuning [1], machine learning [2], and reinforcement learning [3], [4], [5]. The goal of these approaches is to supplement or replace the optimization decisions made by hand-crafted heuristics with decisions derived from empirical data.…”

Section: Introductionmentioning

confidence: 99%

CompilerGym: Robust, Performant Compiler Optimization Environments for AI Research

Cummins¹,

Wasti²,

Guo³

et al. 2021

Preprint

View full text Add to dashboard Cite

Interest in applying Artificial Intelligence (AI) techniques to compiler optimizations is increasing rapidly, but compiler research has a high entry barrier. Unlike in other domains, compiler and AI researchers do not have access to the datasets and frameworks that enable fast iteration and development of ideas, and getting started requires a significant engineering investment. What is needed is an easy, reusable experimental infrastructure for real world compiler optimization tasks that can serve as a common benchmark for comparing techniques, and as a platform to accelerate progress in the field.We introduce CompilerGym 1 , a set of environments for real world compiler optimization tasks, and a toolkit for exposing new optimization tasks to compiler researchers. CompilerGym enables anyone to experiment on production compiler optimization problems through an easy-to-use package, regardless of their experience with compilers. We build upon the popular OpenAI Gym interface enabling researchers to interact with compilers using Python and a familiar API.We describe the CompilerGym architecture and implementation, characterize the optimization spaces and computational efficiencies of three included compiler environments, and provide extensive empirical evaluations. Compared to prior works, Com-pilerGym offers larger datasets and optimization spaces, is 27× more computationally efficient, is fault-tolerant, and capable of detecting reproducibility bugs in the underlying compilers.In making it easy for anyone to experiment with compilersirrespective of their background -we aim to accelerate progress in the AI and compiler research domains.

show abstract

MLGO: a Machine Learning Guided Compiler Optimizations Framework

Cited by 14 publications

References 22 publications

Learning to Combine Instructions in LLVM Compiler

Learning to Combine Instructions in LLVM Compiler

Generating GPU Compiler Heuristics using Reinforcement Learning

CompilerGym: Robust, Performant Compiler Optimization Environments for AI Research

Contact Info

Product

Resources

About