An Investigation into the Performance and Portability SYCL Compiler Implementations
W. R. Shilpage, S. A. Wright
sellOpen SYCL, DPC++, ComputeCpp, CUDA
In this study, we evaluate the performance portability of the Open SYCL, DPC++, and ComputeCpp compilers with a focus on mini-applications of interest to the plasma physics community. Our evaluation is motivated by Project NEPTUNE (NEutrals & Plasma TUrbulance Numerics for the Exascale), a UK project to develop a new simulation code to aid in the design of a future nuclear fusion power plant.
The results in this study have been collected using Isambard, at the University of Bristol, and the Intel DevCloud. Our evaluation includes the Intel HD Graphics P630 GPU, which is a mid-range integrated GPU provided on some Intel Xeon Coffee Lake and Kaby Lake CPUs. It is included in our evaluation to demonstrate the portability to Intel Xe-HPC GPUs, but we do not expect its performance to be competitive with discrete GPUs.
For each of our evaluations on Isambard, we use version 11.0 of the Clang/LLVM compiler environment. We use a custom-build of the compiler infrastructure, to ensure all required features are available (e.g. OpenMP target offload directives, CUDA, HIP). All of our results are collected with -O3 and other performance relevant compiler flags. We use OpenMPI version 4.1, except on the ThunderX2 platform, where we use version 3.1. We use version 11.2 of the CUDA Toolkit, specifying the correct architecture each time. For Kokkos, we use the OpenMP backend for CPU platforms and the CUDA and HIP backends for GPU platforms. The results presented in this paper are the best runtime achieved on each platform, regardless of maximum parallelism achievable, using the best discovered combination of runtime parameters.
For Open SYCL, we build version 0.9.4 of the compiler from source, enabling it to target CPUs and GPUs through OpenMP and CUDA/HIP, respectively.
We use Intel’s proprietary DPC++ compiler for the CLX, KNL and HD P630 platforms, and we build DPC++ version 16.0 from source for the AMD and NVIDIA platforms. For ThunderX2 and A64FX, we were able to compile bench- marks using DPC++ but we encountered linking errors that we were unable to resolve and so we omit results from these platforms.
We use version 2.10.0 of the ComputeCpp compiler, which is distributed as a pre-built executable. It is only compliant up to the SYCL 1.2.1 standard, and therefore is dependent on an OpenCL driver for each architecture; because of this, our platform set is limited to CLX, KNL, Rome and Milan CPUs.
DATE: 2023-03
DOI: 10.1007/978-3-031-40843-4_45