Michael Kenzel, Bernhard KerblORCID iD, Martin Winter, Markus Steinberger
CUDA and Applications to Task-based Programming
In Eurographics 2021 - Tutorials, pages 1-5. May 2021.

Information

  • Publication Type: Other Reviewed Publication
  • Workgroup(s)/Project(s):
  • Date: 2021
  • Booktitle: Eurographics 2021 - Tutorials
  • DOI: 10.2312/egt.20211037
  • Editor: Carol O'Sullivan and Dieter Schmalstieg
  • Location: Vienna
  • Open Access: yes
  • Publisher: The Eurographics Association
  • Pages: 5
  • Pages: 1 – 5
  • Keywords: Parallel Programming, GPU

Abstract

To provide a profound understanding of how CUDA applications can achieve peak performance, the first two parts of this tutorial outline the modern CUDA architecture. Following a basic introduction, we expose how language features are linked to---and constrained by---the underlying physical hardware components. Furthermore, we describe common applications for massively parallel programming, offer a detailed breakdown of potential issues, and list ways to mitigate performance impacts. An exemplary analysis of PTX and SASS snippets illustrates how code patterns in CUDA are mapped to actual hardware instructions.

In parts 3 and 4, we focus on novel features that were enabled by the arrival of CUDA 10+ toolkits and the Volta+ architectures, such as ITS, tensor cores, and the graph API. In addition to basic use case demonstrations, we outline our own experiences with these capabilities and their potential performance benefits. We also discuss how long-standing best practices are affected by these changes and describe common caveats for dealing with legacy code on recent GPU models. We show how these considerations can be implemented in practice by presenting state-of-the-art research into task-based GPU scheduling, and how the dynamic adjustment of thread roles and group configurations can significantly increase performance.

Additional Files and Images

Additional images and videos

teaser
teaser

Additional files

Weblinks

BibTeX

@inproceedings{kenzel_michael_2021_cuda,
  title =      "CUDA and Applications to Task-based Programming",
  author =     "Michael  Kenzel and Bernhard Kerbl and Martin Winter and
               Markus Steinberger",
  year =       "2021",
  abstract =   "To provide a profound understanding of how CUDA applications
               can achieve peak performance, the first two parts of this
               tutorial outline the modern CUDA architecture. Following a
               basic introduction, we expose how language features are
               linked to---and constrained by---the underlying physical
               hardware components. Furthermore, we describe common
               applications for massively parallel programming, offer a
               detailed breakdown of potential issues, and list ways to
               mitigate performance impacts. An exemplary analysis of PTX
               and SASS snippets illustrates how code patterns in CUDA are
               mapped to actual hardware instructions.  In parts 3 and 4,
               we focus on novel features that were enabled by the arrival
               of CUDA 10+ toolkits and the Volta+ architectures, such as
               ITS, tensor cores, and the graph API. In addition to basic
               use case demonstrations, we outline our own experiences with
               these capabilities and their potential performance benefits.
               We also discuss how long-standing best practices are
               affected by these changes and describe common caveats for
               dealing with legacy code on recent GPU models. We show how
               these considerations can be implemented in practice by
               presenting state-of-the-art research into task-based GPU
               scheduling, and how the dynamic adjustment of thread roles
               and group configurations can significantly increase
               performance.",
  month =      may,
  booktitle =  "Eurographics 2021 - Tutorials",
  doi =        "10.2312/egt.20211037",
  editor =     "Carol O'Sullivan and Dieter Schmalstieg",
  location =   "Vienna",
  publisher =  "The Eurographics Association",
  pages =      "5",
  pages =      "1--5",
  keywords =   "Parallel Programming, GPU",
  URL =        "https://www.cg.tuwien.ac.at/research/publications/2021/kenzel_michael_2021_cuda/",
}