KokkosExecutionSpace

KokkosExecutionSpace is the execution-space abstraction in the Kokkos library that represents the backend on which parallel work is executed. It models where and how kernels run, independent of the data they operate on. An ExecutionSpace corresponds to a specific backend such as Serial, OpenMP, Threads, CUDA, HIP, or SYCL, enabling performance portability across diverse hardware without changing the core algorithms.

In typical use, kernels are launched via parallel_for, parallel_reduce, or parallel_scan with a policy that specifies an ExecutionSpace. Data containers in Kokkos, such as Views, are associated with a MemorySpace, while the ExecutionSpace handles the scheduling and execution of operations. The default ExecutionSpace can be queried or configured, and an application must initialize the Kokkos runtime (Kokkos::initialize) before use and finalize it on exit.

The design allows the same code to run on different architectures by simply selecting a different ExecutionSpace, or by compiling with support for multiple backends. This enables developers to target CPUs and accelerators (GPUs) from a single source tree, improving portability and maintainability. Backends may differ in capabilities, such as available synchronization primitives or memory models, so some features can have varying behavior across ExecutionSpaces.

Kokkos provides a set of predefined ExecutionSpace types and mechanisms to customize execution and memory configuration. Understanding the ExecutionSpace concept is central to writing portable, high-performance code with Kokkos, as it determines how and where parallel work is dispatched during execution.