An Investigation into the Performance and Portability of SYCL Compiler Implementations

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.