NDRange

NDRange, short for N-dimensional range, is a term used in OpenCL to describe the size and shape of the index space over which a kernel is executed. It represents an N-dimensional domain of work-items, typically 1D, 2D, or 3D. Each work-item corresponds to a point in this domain and can perform computation for that point. Inside the kernel, the coordinates of a work-item are obtained using built-in functions such as get_global_id(d), where d is the dimension index.

The NDRange is defined when launching a kernel on a command queue. It is specified by the global_work_size parameter, which gives the total number of work-items in each dimension, and by the optional global_work_offset, which can shift the starting coordinates. The work-items are further organized into work-groups defined by local_work_size, with one or more work-items per group. If a local size is provided, the global size should typically be a multiple of the local size in each dimension; if local_work_size is NULL, the device may choose an appropriate size.

Execution proceeds by mapping the NDRange into work-groups. Each work-group has its own local coordinates, accessible via get_local_id(d), while the global coordinates are accessible via get_global_id(d). The total number of work-items equals the product of the global_work_size components, and the total number of work-groups depends on the global and local sizes per dimension. This model supports data-parallel computations over arrays, images, or other structured data.

In OpenCL, the API function clEnqueueNDRangeKernel enqueues kernels with a specified dimension (1, 2, or 3) and arrays for global_work_size, global_work_offset, and local_work_size. The NDRange concept is central to the OpenCL programming model and differs from CUDA terminology, though both describe parallel execution domains.