Figure 2: The diagram shows how the Open CL structure enables a host program to use a standard
compiler while the kernel uses a vendor-specific compiler. (Source: Xilinx)
| 4 | | 22 |
Accelerating with Vitis and OpenCL
Vitis enables users to leverage the OpenCL framework to
implement acceleration kernels within the programmable
logic. These acceleration kernels have been defined using
a higher-level language than a traditional RTL.
OpenCL is an industry-standard framework that supports
parallel computing on heterogeneous systems. One of
the core principles behind OpenCL is to enable cross-
platform functionality without the need for code changes.
This allows the same code to be portable across CPUs,
GPUs, FPGAs, DSPs, etc., with the performance scaling
depending upon the platform's capabilities.
OpenCL uses a host and kernel model
(Figure 2). Each
system will have one host, typically x86-based, and several
kernels that provide the acceleration and are usually
GPU-, DSP-, or FPGA-based. The host application is
often developed in C/C++ and uses OpenCL APIs to
support the OpenCL flow. These OpenCL APIs allow the
Xilinx PRODUCT SPOTLIGHT
USING SYSTEM
GENERATOR FOR DSP
FOR ZYNQ ULTRASCALE+
RFSOC
s
host to manage the entire application lifecycle of loading,
configuring, and executing kernels. While supporting
cross-platform portability, the kernel is developed
using the OpenCL C language, based upon C but has
limitations to support cross-platform portability.
This model allows the host program to be compiled using
a standard compiler such as GCC or G++, while the
kernel compiler is vendor-specific.
When working with Xilinx heterogeneous SoC devices,
the Arm Application Processing Unit is the host, while the
programmable logic instantiates kernels. Vitis provides the
developer with everything required to generate, debug,
and analyze both the host and kernel elements when
targeting Xilinx heterogeneous SoC or acceleration cards.
Vitis Platform
To be able to leverage the Vitis OpenCL capabilities, a
base platform is required. This base platform defines both
