high performance computing on graphics processing units: hgpu.org

Today, many big data applications require massively parallel tasks to compute complicated mathematical operations. To perform parallel tasks, platforms like CUDA (Compute Unified Device Architecture) and OpenCL (Open Computing Language) are widely used and developed to enhance the throughput of massively parallel tasks. There is also a need for high-level abstractions and platform-independence over those massively parallel computing platforms. Recently, Khronos group announced SYCL (C++ Single-source Heterogeneous Programming for OpenCL), a new cross-platform abstraction layer, to provide an efficient way for single-source heterogeneous computing, with C++-template-level abstractions. However, since there has been no official implementation of SYCL, we currently have several different implementations from various vendors. In this paper, we analyse the characteristics of those SYCL implementations. We also show performance measures of those SYCL implementations, especially for well-known massively parallel tasks. We show that each implementation has its own strength in computing different types of mathematical operations, along with different sizes of data. Our analysis is available for fundamental measurements of the abstract-level cost-effective use of massively parallel computations, especially for big-data applications.