KokkosView

KokkosView refers to the core multi-dimensional array type provided by the Kokkos library, typically implemented as Kokkos::View. It serves as a portable, high-level container for data that participates in Kokkos parallel operations and memory management across diverse hardware backends, such as CPUs and GPUs. A Kokkos::View manages the allocation, access, and lifetime of data and is designed to work efficiently within Kokkos execution spaces and memory spaces.

A View is a templated object parameterized by the data type, the rank (dimension) of the array, a memory layout (for example, LayoutLeft or LayoutRight), and a memory space (for example, HostSpace, CudaSpace, or others). The rank is encoded in the type via the number of pointers in the DataType (for example, double* for 1D, double** for 2D, etc.). A typical instantiation might be Kokkos::View<double**, Kokkos::LayoutLeft, Kokkos::HostSpace> for a 2D array stored in host memory with a specific layout. Views can be created with or without preallocated memory and may be configured as owning or non-owning, depending on how they are constructed.

Views support deep and shallow copy semantics. Copying a View by default creates another View that references the same underlying data (shallow copy), while a separate deep_copy operation can copy data between views, potentially across memory spaces. Views also provide subviews to obtain smaller views of portions of an existing array, enabling flexible data slicing without additional allocations.

Core features include element access via operator(), obtaining extents and sizes, and interoperability with Kokkos kernels such as parallel_for and parallel_reduce. Views enable portability by allowing code to run with the same syntax on different backends and memory spaces, making KokkosView a central building block for performance-portable numerical software.