SYCL

SYCL is a royalty-free, cross-platform abstraction layer for single-source C++ programming of heterogeneous compute devices, standardized by the Khronos Group. It enables developers to write code that targets CPUs, GPUs, FPGAs, and other accelerators from a single source file and compile for different devices without separate host and device languages.

Programs in SYCL create a queue bound to a device and submit work through command groups. The core concepts include buffers, accessors, and handlers, which manage data movement and synchronization, and kernels written as C++ lambdas or function objects. SYCL uses a single-source model, allowing host and device code to be authored together. A complementary memory model, Unified Shared Memory (USM), provides pointer-based memory that can be shared or accessible across host and device.

The execution model is explicit but high-level: data is prepared by buffers/accessors or USM, and kernels execute in parallel over a specified range via parallel_for constructs. Synchronization is performed through wait operations and events.

Implementations and backends vary: multiple vendors provide SYCL implementations, and the standard defines an API that can map to different device runtimes such as OpenCL, CUDA, Level Zero, or HIP. Notable implementations include Intel's oneAPI DPC++, Codeplay's ComputeCpp, hipSYCL, and TriSYCL. SYCL 1.2.1 established an initial widely-deployed version; SYCL 2020 introduced a revised standard with broader features and improved interoperability.

Use cases include high-performance computing, data analytics, and computer graphics, where developers want portable performance across heterogeneous hardware from a single codebase.