SYCL-BLAS: Combining Expression Trees and Kernel Fusion on Heterogeneous Systems

The support for heterogenous platforms requires multiple specialised devices collaborate to execute an application. The SYCL standard publishes by Khronos, providing a C++ abstraction layer on top of OpenCL that provides single-source programming for a large number of heterogeneous devices. Single-source programming and task data-flow approach enable SYCL developers to leverage modern programming techniques on heterogeneous platforms. In this paper, we present how SYCL combines expression tree templates and kernel fusion to develop SYCL-BLAS, an efficient BLAS implementation for heterogeneous platforms. The use of templates permits to generate BLAS kernels related to each BLAS routine. whereas kernel fusion describes how to merge the expression trees, enlarging the BLAS kernels. These features prove that SYCL can be used to quickly develop libraries for heterogeneous systems by providing sufficient levels of abstraction. Our experiments compare the performances of clBLAS and SYCL-BLAS on a server equipped with an Intel Core i7-6700K CPU and an AMD R9 GPU.