logo
CanyamAI summaries for academic research
language
MobileMobile
  • Phone FeatureWeChat Mini Program
EnglishEnglish
noticeNotice
BlogBlog
delete
arrow
Return

ProSpec: Profile-guided Specialization for GPU Kernels

delete2025-09-25
delete0
PRE
AI
J
Jiali Liang
L
Lin Chen *
张
张祥宇 (Xiangyu Zhang)
Y
Yanhui Li
Y
Yuming Zhou
DOI：10.1016/j.infsof.2025.107901delete
deleteOriginal
deleteOriginal request for help
deleteShare
deleteSave
Abstract

Abstract

En 中文
• We propose ProSpec, a novel Profile-Guided Optimization (PGO) approach for automating the optimization of GPU kernels by collecting and analyzing runtime parameter information. • We develop a prototype system of ProSpec based on LLVM, implementing its instrumentation, analysis, and optimization components. It optimizes GPU kernels with minimal overhead, and it automatically applies optimizations, boosting GPU performance without manual tuning. • We validate the effectiveness of ProSpec on popular GPU benchmarks, surpassing state-of-the-art methods in several experimental metrics.

Journal

Information and Software Technology cover
Information and Software Technology
IF:
4.3
Papers:
3.7K
Citations:
7.7K

Organization

No organization information available