Experience Report: Writing a Portable GPU Runtime with OpenMP 5.1

Shilei Tian; Jon Chesterfield; Johannes Doerfert; Barbara Chapman

doi:10.1007/978-3-030-85262-7_11

Experience Report: Writing a Portable GPU Runtime with OpenMP 5.1

Shilei Tian
, Jon Chesterfield
, Johannes Doerfert
, Barbara Chapman

Computer Science

Stony Brook University
AMD
Argonne National Laboratory

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

13 Scopus citations

Abstract

GPU runtimes are historically implemented in CUDA or other vendor specific languages dedicated to GPU programming. In this work we show that OpenMP 5.1, with minor compiler extensions, is capable of replacing existing solutions without a performance penalty. The result is a performant and portable GPU runtime that can be compiled with LLVM/Clang to Nvidia and AMD GPUs without the need for CUDA or HIP during its development and compilation. While we tried to be OpenMP compliant, we identified the need for compiler extensions to achieve the CUDA performance with our OpenMP runtime. We hope that future versions of OpenMP adopt our extensions to make device programming in OpenMP also portable across compilers, not only across execution platforms. The library we ported to OpenMP is the OpenMP device runtime that provides OpenMP functionality on the GPU. This work opens the door for shipping OpenMP offloading with a Linux distribution’s LLVM package as the package manager would not need a vendor SDK to build the compiler and runtimes. Furthermore, our OpenMP device runtime can support a new GPU target through the use of a few compiler intrinsics rather than requiring a reimplementation of the entire runtime.

Original language	English
Title of host publication	OpenMP
Subtitle of host publication	Enabling Massive Node-Level Parallelism - 17th International Workshop on OpenMP, IWOMP 2021, Proceedings
Editors	Simon McIntosh-Smith, Bronis R. de Supinski, Jannis Klinkenberg
Publisher	Springer Science and Business Media Deutschland GmbH
Pages	159-169
Number of pages	11
ISBN (Print)	9783030852610
DOIs	https://doi.org/10.1007/978-3-030-85262-7_11
State	Published - 2021
Event	17th International Workshop on OpenMP, IWOMP 2021 - Bristol, United Kingdom Duration: Sep 14 2021 → Sep 16 2021

Publication series

Name	Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume	12870 LNCS
ISSN (Print)	0302-9743
ISSN (Electronic)	1611-3349

Conference

Conference	17th International Workshop on OpenMP, IWOMP 2021
Country/Territory	United Kingdom
City	Bristol
Period	09/14/21 → 09/16/21

Keywords

Accelerator
LLVM
OpenMP
Portability
Runtimes
Target offloading

Access to Document

10.1007/978-3-030-85262-7_11

Cite this

Tian, S., Chesterfield, J., Doerfert, J., & Chapman, B. (2021). Experience Report: Writing a Portable GPU Runtime with OpenMP 5.1. In S. McIntosh-Smith, B. R. de Supinski, & J. Klinkenberg (Eds.), OpenMP: Enabling Massive Node-Level Parallelism - 17th International Workshop on OpenMP, IWOMP 2021, Proceedings (pp. 159-169). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 12870 LNCS). Springer Science and Business Media Deutschland GmbH. https://doi.org/10.1007/978-3-030-85262-7_11

Tian, Shilei ; Chesterfield, Jon ; Doerfert, Johannes et al. / Experience Report : Writing a Portable GPU Runtime with OpenMP 5.1. OpenMP: Enabling Massive Node-Level Parallelism - 17th International Workshop on OpenMP, IWOMP 2021, Proceedings. editor / Simon McIntosh-Smith ; Bronis R. de Supinski ; Jannis Klinkenberg. Springer Science and Business Media Deutschland GmbH, 2021. pp. 159-169 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).

@inproceedings{ef16dd42985947ecb136b2f229a80158,

title = "Experience Report: Writing a Portable GPU Runtime with OpenMP 5.1",

abstract = "GPU runtimes are historically implemented in CUDA or other vendor specific languages dedicated to GPU programming. In this work we show that OpenMP 5.1, with minor compiler extensions, is capable of replacing existing solutions without a performance penalty. The result is a performant and portable GPU runtime that can be compiled with LLVM/Clang to Nvidia and AMD GPUs without the need for CUDA or HIP during its development and compilation. While we tried to be OpenMP compliant, we identified the need for compiler extensions to achieve the CUDA performance with our OpenMP runtime. We hope that future versions of OpenMP adopt our extensions to make device programming in OpenMP also portable across compilers, not only across execution platforms. The library we ported to OpenMP is the OpenMP device runtime that provides OpenMP functionality on the GPU. This work opens the door for shipping OpenMP offloading with a Linux distribution{\textquoteright}s LLVM package as the package manager would not need a vendor SDK to build the compiler and runtimes. Furthermore, our OpenMP device runtime can support a new GPU target through the use of a few compiler intrinsics rather than requiring a reimplementation of the entire runtime.",

keywords = "Accelerator, LLVM, OpenMP, Portability, Runtimes, Target offloading",

author = "Shilei Tian and Jon Chesterfield and Johannes Doerfert and Barbara Chapman",

note = "Publisher Copyright: {\textcopyright} 2021, Springer Nature Switzerland AG.; 17th International Workshop on OpenMP, IWOMP 2021 ; Conference date: 14-09-2021 Through 16-09-2021",

year = "2021",

doi = "10.1007/978-3-030-85262-7\_11",

language = "English",

isbn = "9783030852610",

series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",

publisher = "Springer Science and Business Media Deutschland GmbH",

pages = "159--169",

editor = "Simon McIntosh-Smith and \{de Supinski\}, \{Bronis R.\} and Jannis Klinkenberg",

booktitle = "OpenMP",

}

Tian, S, Chesterfield, J, Doerfert, J & Chapman, B 2021, Experience Report: Writing a Portable GPU Runtime with OpenMP 5.1. in S McIntosh-Smith, BR de Supinski & J Klinkenberg (eds), OpenMP: Enabling Massive Node-Level Parallelism - 17th International Workshop on OpenMP, IWOMP 2021, Proceedings. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 12870 LNCS, Springer Science and Business Media Deutschland GmbH, pp. 159-169, 17th International Workshop on OpenMP, IWOMP 2021, Bristol, United Kingdom, 09/14/21. https://doi.org/10.1007/978-3-030-85262-7_11

Experience Report: Writing a Portable GPU Runtime with OpenMP 5.1. / Tian, Shilei; Chesterfield, Jon; Doerfert, Johannes et al.
OpenMP: Enabling Massive Node-Level Parallelism - 17th International Workshop on OpenMP, IWOMP 2021, Proceedings. ed. / Simon McIntosh-Smith; Bronis R. de Supinski; Jannis Klinkenberg. Springer Science and Business Media Deutschland GmbH, 2021. p. 159-169 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 12870 LNCS).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

TY - GEN

T1 - Experience Report

T2 - 17th International Workshop on OpenMP, IWOMP 2021

AU - Tian, Shilei

AU - Chesterfield, Jon

AU - Doerfert, Johannes

AU - Chapman, Barbara

PY - 2021

Y1 - 2021

N2 - GPU runtimes are historically implemented in CUDA or other vendor specific languages dedicated to GPU programming. In this work we show that OpenMP 5.1, with minor compiler extensions, is capable of replacing existing solutions without a performance penalty. The result is a performant and portable GPU runtime that can be compiled with LLVM/Clang to Nvidia and AMD GPUs without the need for CUDA or HIP during its development and compilation. While we tried to be OpenMP compliant, we identified the need for compiler extensions to achieve the CUDA performance with our OpenMP runtime. We hope that future versions of OpenMP adopt our extensions to make device programming in OpenMP also portable across compilers, not only across execution platforms. The library we ported to OpenMP is the OpenMP device runtime that provides OpenMP functionality on the GPU. This work opens the door for shipping OpenMP offloading with a Linux distribution’s LLVM package as the package manager would not need a vendor SDK to build the compiler and runtimes. Furthermore, our OpenMP device runtime can support a new GPU target through the use of a few compiler intrinsics rather than requiring a reimplementation of the entire runtime.

AB - GPU runtimes are historically implemented in CUDA or other vendor specific languages dedicated to GPU programming. In this work we show that OpenMP 5.1, with minor compiler extensions, is capable of replacing existing solutions without a performance penalty. The result is a performant and portable GPU runtime that can be compiled with LLVM/Clang to Nvidia and AMD GPUs without the need for CUDA or HIP during its development and compilation. While we tried to be OpenMP compliant, we identified the need for compiler extensions to achieve the CUDA performance with our OpenMP runtime. We hope that future versions of OpenMP adopt our extensions to make device programming in OpenMP also portable across compilers, not only across execution platforms. The library we ported to OpenMP is the OpenMP device runtime that provides OpenMP functionality on the GPU. This work opens the door for shipping OpenMP offloading with a Linux distribution’s LLVM package as the package manager would not need a vendor SDK to build the compiler and runtimes. Furthermore, our OpenMP device runtime can support a new GPU target through the use of a few compiler intrinsics rather than requiring a reimplementation of the entire runtime.

KW - Accelerator

KW - LLVM

KW - OpenMP

KW - Portability

KW - Runtimes

KW - Target offloading

UR - https://www.scopus.com/pages/publications/85115441799

U2 - 10.1007/978-3-030-85262-7_11

DO - 10.1007/978-3-030-85262-7_11

M3 - Conference contribution

AN - SCOPUS:85115441799

SN - 9783030852610

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 159

EP - 169

BT - OpenMP

A2 - McIntosh-Smith, Simon

A2 - de Supinski, Bronis R.

A2 - Klinkenberg, Jannis

PB - Springer Science and Business Media Deutschland GmbH

Y2 - 14 September 2021 through 16 September 2021

ER -

Tian S, Chesterfield J, Doerfert J, Chapman B. Experience Report: Writing a Portable GPU Runtime with OpenMP 5.1. In McIntosh-Smith S, de Supinski BR, Klinkenberg J, editors, OpenMP: Enabling Massive Node-Level Parallelism - 17th International Workshop on OpenMP, IWOMP 2021, Proceedings. Springer Science and Business Media Deutschland GmbH. 2021. p. 159-169. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). doi: 10.1007/978-3-030-85262-7_11

Experience Report: Writing a Portable GPU Runtime with OpenMP 5.1

Abstract

Publication series

Conference

Keywords

Access to Document

Other files and links

Fingerprint

Cite this