An Investigation into the Performance and Portability of SYCL Compiler Implementations

TY - GEN

T1 - An Investigation into the Performance and Portability of SYCL Compiler Implementations

AU - Shilpage, Wageesha R.

AU - Wright, Steven A.

N1 - Funding Information: Many of the results in this paper were gathered on the Isambard UK National Tier-2 HPC Service (http://gw4.ac.uk/isambard/) operated by GW4 and the UK Met Office, and funded by EPSRC (EP/P020224/1). Access to the Intel HD Graphics P630 GPU was provided by Intel through the Intel Developer Cloud. The ExCALIBUR programme (https://excalibur.ac.uk/) is supported by the UKRI Strategic Priorities Fund. The programme is co-delivered by the Met Office and EPSRC in partnership with the Public Sector Research Establishment, the UK Atomic Energy Authority (UKAEA) and UKRI research councils, including NERC, MRC and STFC. Publisher Copyright: © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.

PY - 2023/8/25

Y1 - 2023/8/25

N2 - In June 2022, Frontier became the first Supercomputer to “officially” break the ExaFLOP/s barrier on LINPACK, achieving a peak performance of 1.1×10^18 floating-point operations per second using AMD Instinct accelerators. Developing high performance applications for such platforms typically requires the adoption of vendor-specific programming models, which in turn may limit portability. SYCL is a high-level, single-source language based on C++17, developed by the Khronos group to overcome the shortcomings of those vendor-specific HPC programming models. In this paper we present an initial study into the SYCL parallel programming model and its implementing compilers, to understand its performance and portability, and how this compares to other parallel programming models. We use three major SYCL implementations for our evaluation – Open SYCL (previously hipSYCL), DPC++, and ComputeCpp – on a range of CPU and GPU hardware from Intel, AMD, Fujitsu, Marvell, and NVIDIA. Our results show that for a simple finite difference mini-application, SYCL can offer competitive performance to native approaches, while for a more complex finite-element mini-application, significant performance degradation is observed. Our findings suggest that development work is required at the compiler- and application-level to ensure SYCL is competitive with alternative approaches.

AB - In June 2022, Frontier became the first Supercomputer to “officially” break the ExaFLOP/s barrier on LINPACK, achieving a peak performance of 1.1×10^18 floating-point operations per second using AMD Instinct accelerators. Developing high performance applications for such platforms typically requires the adoption of vendor-specific programming models, which in turn may limit portability. SYCL is a high-level, single-source language based on C++17, developed by the Khronos group to overcome the shortcomings of those vendor-specific HPC programming models. In this paper we present an initial study into the SYCL parallel programming model and its implementing compilers, to understand its performance and portability, and how this compares to other parallel programming models. We use three major SYCL implementations for our evaluation – Open SYCL (previously hipSYCL), DPC++, and ComputeCpp – on a range of CPU and GPU hardware from Intel, AMD, Fujitsu, Marvell, and NVIDIA. Our results show that for a simple finite difference mini-application, SYCL can offer competitive performance to native approaches, while for a more complex finite-element mini-application, significant performance degradation is observed. Our findings suggest that development work is required at the compiler- and application-level to ensure SYCL is competitive with alternative approaches.

KW - High-Performance Computing

KW - Performance Portability

KW - SYCL

UR - http://www.scopus.com/inward/record.url?scp=85171337277&partnerID=8YFLogxK

U2 - 10.1007/978-3-031-40843-4_45

DO - 10.1007/978-3-031-40843-4_45

M3 - Conference contribution

AN - SCOPUS:85171337277

SN - 9783031408427

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

SP - 605

EP - 619

BT - High Performance Computing - ISC High Performance 2023 International Workshops, Revised Selected Papers

A2 - Bienz, Amanda

A2 - Weiland, Michèle

A2 - Baboulin, Marc

A2 - Kruse, Carola

PB - Springer Science and Business Media Deutschland GmbH

T2 - 38th International Conference on High Performance Computing, ISC High Performance 2023

Y2 - 21 May 2023 through 25 May 2023

ER -