This project evaluated the performance portability of SYCL, a new parallel programming model developed and maintained by the Khronos consortium. SYCL is built on top of the low-level abstraction layer of OpenCL and based on pure C++, enabling developers to write single source code for heterogeneous processors while being able to use standard ISO C++11 features, such as inheritance, templating and namespaces.
The research hypothesis was that, as a high-level heterogeneous model, SYCL can achieve comparable performance to OpenCL, whilst also offering the convenience of using C++ abstractions and a less verbose single source style, therefore lowering the overall complexity of heterogeneous programming in comparison to other similar models. The conclusion was that — in most cases — SYCL is capable of attaining similar performance to OpenCL. In the case of the neutral mini-app, the performance of the SYCL implementation was comparable with the OpenMP version and often surpassed the performance of Kokkos.
The complete dissertation can be read here. The source code for the lattice-boltzmann implementations is available in this repository. Some of the results obtained on GPUs are shown below as a reference comparison between SYCL and OpenCL.
