OpenCL Devices and FPGAs

In the context of CPU and GPU hardware, the attributes of an OpenCL device are fixed and the programmer has very little influence on what the device looks like. An advantage of this characteristic of CPU/GPU systems makes it relatively easy to obtain and use off-the-shelf hardware. This advantage is also a major limitation when compared to FPGA based OpenCL devices. CPU and GPU based systems typically have fixed data paths, memory systems, and I/O architectures. It is not possible, for example, to directly attach high-speed I/O to an OpenCL compute kernel. Similarly, efficient data movement is only performed using bulk memory based transfers.

An OpenCL device for an FPGA is not limited by the constraints of a CPU/GPU device. By taking advantage of the fact that the FPGA starts off as a blank computational canvas, the user can decide the level of device customization that is appropriate to support a single application or a class of applications. In determining the level of customization in a device, the programmer can take advantage of the fact that kernel compute units are not placed in isolation within the FPGA fabric.

FPGA devices capable of supporting OpenCL programs can include, but are not limited to, the following components:

  • DMA engines
  • I/O peripherals such as PCIe and Ethernet
  • Memory controllers
  • Custom interconnects
  • OpenCL compute units
  • RTL-based accelerators
The block diagram below shows an example SDAccel base platform design that includes these components.

Figure: Xilinx FPGA

The creation of Xilinx FPGA based OpenCL devices requires FPGA design expertise and is beyond the scope of SDAccel itself. Devices for SDAccel are created using the Xilinx Vivado® design suite for FPGA designers. SDAccel provides pre-defined devices as well as allows users to augment the tool with third party created devices. A methodology guide describing how to create a device for SDAccel is available upon request from Xilinx.

The devices available in SDAccel are for Virtex®-7, Kintex®-7, and Kintex-UltraScale® devices from Xilinx. These devices are available in a PCIe form factor. The PCIe form factor assumes that the host processor is an x86 based processor and that the FPGA is used for the implementation of compute units.