OpenCL Platform Model

The OpenCL™ platform model defines the logical representation of all hardware capable of executing an OpenCL program. At the most fundamental level all platforms are defined by the grouping of a processor and one or more devices. The host processor, which runs the OS for the system, is also responsible for the general bookkeeping and task launch duties associated with the execution of parallel programs such as OpenCL applications. The device is the element in the system on which the kernels of the application are executed. The device is further divided into a set of compute units. The number of compute units depends on the target hardware for a specific application. A compute unit is defined as the element in the hardware device onto which a work group of a kernel is executed. This device is responsible for executing the operations of the assigned work group to completion. In accordance to the OpenCL standard division of work groups into work items, a compute unit is further subdivided into processing elements. A processing element is the data path in the compute unit, which is responsible for executing the operations of one work item. A conceptual view of this model is shown in the following figure.

An OpenCL platform always starts with a host processor. For the case of platforms created with Xilinx® devices, the host processor is an x86 or Power8 based processor communicating to the devices using a PCIe™ solution. The host processor has the following responsibilities:

Manage the operating system and enable drivers for all devices.
Execute the application host program.
Set up all global memory buffers and manage data transfer between the host and the device.
Monitor the status of all compute units in the system.

In all OpenCL platforms, the host processor tasks are executed using a common set of API functions. The implementation of the OpenCL API functions is provided by the hardware vendor and is referred to as the runtime library.

The OpenCL runtime library, which is provided by the hardware vendor, is the logical layer in a platform that is responsible for translating user commands described by the OpenCL API into hardware specific commands for a given device. For example, when the application programmer allocates a memory buffer using the clCreateBuffer API call, it is the responsibility of the runtime library to keep track of where the allocated buffer physically resides in the system, and of the mechanism required for buffer access. It is important for the application programmer to keep in mind that the OpenCL API is portable across vendors, but the runtime library provided by a vendor is not. Therefore, OpenCL applications have to be linked at compile time with the runtime library that is paired with the target execution device.

The other component of a platform is the device. A device in the context of an OpenCL API is the physical collection of hardware resources onto which the application kernels are executed. A platform must have at least one device available for the execution of kernels. Also, per the OpenCL platform model, all devices in a platform do not have to be of identical type.