The Performance Impact of Formulating Computations in SYCL on CPUs and GPUs

We explore the different ways one can formulate structured mesh stencil computations as well as unstructured mesh computations in SYCL: multidimensional parallel_for computations with or without nd_range, and different ways of avoiding race conflicts in unstructured meshes (atomics and two ways of coloring). These formulations are briefly contrasted to other parallelization approaches (OpenMP, CUDA). We then proceed to evaluate a range of modern parallel architectures from Intel, AMD, and NVIDIA – including CPUs and GPUs – using the DPC++ compiler, and hipSYCL (aka OpenSYCL).

Download Presentation


The oneAPI Specification joins the UXL Foundation. For more info:

Press Release