ProSpec: GPU内核的配置文件引导专门化

IF 4.3 2区计算机科学 Q2 COMPUTER SCIENCE, INFORMATION SYSTEMS

Information and Software Technology Pub Date : 2025-09-25 DOI:10.1016/j.infsof.2025.107901

Jiali Liang, Lin Chen, Xiangyu Zhang, Yanhui Li, Yuming Zhou

{"title":"ProSpec: GPU内核的配置文件引导专门化","authors":"Jiali Liang, Lin Chen, Xiangyu Zhang, Yanhui Li, Yuming Zhou","doi":"10.1016/j.infsof.2025.107901","DOIUrl":null,"url":null,"abstract":"<div><div>General-purpose GPUs are widely used for computational acceleration in various fields. Designing high-performance GPU kernels is challenging due to dynamic kernel variables and complex GPU architectures.</div><div>Leveraging runtime profiling to identify value-related inefficiencies is effective for optimizing GPU kernels, but it faces several challenges: (1) high profiling overhead, (2) limited analysis of inter-variable correlations, and (3) lack of automated optimization mechanisms.</div><div>In this paper, we propose a profile-guided optimization technique named ProSpec for GPU Kernel Specialization. It offloads profile collection to CPUs, analyzes inefficiency patterns dependent on multiple hot values, and generates optimization feedback for automatic kernel specialization.</div><div>The prototype of ProSpec, implemented over the LLVM infrastructure, is evaluated on the Rodinia and Polybench benchmarks. It achieves a maximum speedup of 5.619x and an average of 1.417x on optimized applications, maintaining a low profiling overhead of around 1.01x.</div><div>Compared to state-of-the-art methods, ProSpec leads in the number of improved kernels and further optimizes half of those already optimized by other tools.</div></div>","PeriodicalId":54983,"journal":{"name":"Information and Software Technology","volume":"188 ","pages":"Article 107901"},"PeriodicalIF":4.3000,"publicationDate":"2025-09-25","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":"0","resultStr":"{\"title\":\"ProSpec: Profile-guided Specialization for GPU Kernels\",\"authors\":\"Jiali Liang, Lin Chen, Xiangyu Zhang, Yanhui Li, Yuming Zhou\",\"doi\":\"10.1016/j.infsof.2025.107901\",\"DOIUrl\":null,\"url\":null,\"abstract\":\"<div><div>General-purpose GPUs are widely used for computational acceleration in various fields. Designing high-performance GPU kernels is challenging due to dynamic kernel variables and complex GPU architectures.</div><div>Leveraging runtime profiling to identify value-related inefficiencies is effective for optimizing GPU kernels, but it faces several challenges: (1) high profiling overhead, (2) limited analysis of inter-variable correlations, and (3) lack of automated optimization mechanisms.</div><div>In this paper, we propose a profile-guided optimization technique named ProSpec for GPU Kernel Specialization. It offloads profile collection to CPUs, analyzes inefficiency patterns dependent on multiple hot values, and generates optimization feedback for automatic kernel specialization.</div><div>The prototype of ProSpec, implemented over the LLVM infrastructure, is evaluated on the Rodinia and Polybench benchmarks. It achieves a maximum speedup of 5.619x and an average of 1.417x on optimized applications, maintaining a low profiling overhead of around 1.01x.</div><div>Compared to state-of-the-art methods, ProSpec leads in the number of improved kernels and further optimizes half of those already optimized by other tools.</div></div>\",\"PeriodicalId\":54983,\"journal\":{\"name\":\"Information and Software Technology\",\"volume\":\"188 \",\"pages\":\"Article 107901\"},\"PeriodicalIF\":4.3000,\"publicationDate\":\"2025-09-25\",\"publicationTypes\":\"Journal Article\",\"fieldsOfStudy\":null,\"isOpenAccess\":false,\"openAccessPdf\":\"\",\"citationCount\":\"0\",\"resultStr\":null,\"platform\":\"Semanticscholar\",\"paperid\":null,\"PeriodicalName\":\"Information and Software Technology\",\"FirstCategoryId\":\"94\",\"ListUrlMain\":\"https://www.sciencedirect.com/science/article/pii/S095058492500240X\",\"RegionNum\":2,\"RegionCategory\":\"计算机科学\",\"ArticlePicture\":[],\"TitleCN\":null,\"AbstractTextCN\":null,\"PMCID\":null,\"EPubDate\":\"\",\"PubModel\":\"\",\"JCR\":\"Q2\",\"JCRName\":\"COMPUTER SCIENCE, INFORMATION SYSTEMS\",\"Score\":null,\"Total\":0}","platform":"Semanticscholar","paperid":null,"PeriodicalName":"Information and Software Technology","FirstCategoryId":"94","ListUrlMain":"https://www.sciencedirect.com/science/article/pii/S095058492500240X","RegionNum":2,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":null,"EPubDate":"","PubModel":"","JCR":"Q2","JCRName":"COMPUTER SCIENCE, INFORMATION SYSTEMS","Score":null,"Total":0}

引用次数: 0

摘要

通用图形处理器被广泛应用于各个领域的计算加速。由于动态内核变量和复杂的GPU架构，设计高性能GPU内核具有挑战性。利用运行时分析来识别与值相关的低效率对于优化GPU内核是有效的，但它面临着几个挑战：(1)高分析开销，(2)对变量间相关性的有限分析，以及(3)缺乏自动优化机制。本文提出了一种基于配置文件的GPU内核专门化优化技术——ProSpec。它将配置文件收集卸载到cpu，分析依赖于多个热值的低效率模式，并为自动内核专门化生成优化反馈。在LLVM基础架构上实现的ProSpec原型在Rodinia和Polybench基准测试上进行了评估。在优化的应用程序上，它实现了5.619倍的最大加速和1.417倍的平均加速，保持了1.01倍左右的低性能开销。与最先进的方法相比，ProSpec在改进内核的数量上领先，并进一步优化了其他工具已经优化的一半。

本文章由计算机程序翻译，如有差异，请以英文原文为准。

查看原文本刊更多论文

ProSpec: Profile-guided Specialization for GPU Kernels

General-purpose GPUs are widely used for computational acceleration in various fields. Designing high-performance GPU kernels is challenging due to dynamic kernel variables and complex GPU architectures.

Leveraging runtime profiling to identify value-related inefficiencies is effective for optimizing GPU kernels, but it faces several challenges: (1) high profiling overhead, (2) limited analysis of inter-variable correlations, and (3) lack of automated optimization mechanisms.

In this paper, we propose a profile-guided optimization technique named ProSpec for GPU Kernel Specialization. It offloads profile collection to CPUs, analyzes inefficiency patterns dependent on multiple hot values, and generates optimization feedback for automatic kernel specialization.

The prototype of ProSpec, implemented over the LLVM infrastructure, is evaluated on the Rodinia and Polybench benchmarks. It achieves a maximum speedup of 5.619x and an average of 1.417x on optimized applications, maintaining a low profiling overhead of around 1.01x.

Compared to state-of-the-art methods, ProSpec leads in the number of improved kernels and further optimizes half of those already optimized by other tools.

求助全文

通过发布文献求助，成功后即可免费获取论文全文。去求助

来源期刊

Information and Software Technology 工程技术-计算机：软件工程

CiteScore

9.10

自引率

7.70%

发文量

164

审稿时长

9.6 weeks

期刊介绍： Information and Software Technology is the international archival journal focusing on research and experience that contributes to the improvement of software development practices. The journal''s scope includes methods and techniques to better engineer software and manage its development. Articles submitted for review should have a clear component of software engineering or address ways to improve the engineering and management of software development. Areas covered by the journal include: • Software management, quality and metrics, • Software processes, • Software architecture, modelling, specification, design and programming • Functional and non-functional software requirements • Software testing and verification & validation • Empirical studies of all aspects of engineering and managing software development Short Communications is a new section dedicated to short papers addressing new ideas, controversial opinions, "Negative" results and much more. Read the Guide for authors for more information. The journal encourages and welcomes submissions of systematic literature studies (reviews and maps) within the scope of the journal. Information and Software Technology is the premiere outlet for systematic literature studies in software engineering.