AMD OpenVX

AMD OpenVX#

The Khronos OpenVX API offers a set of optimized primitives for low-level image processing, computer vision, and neural net operators. The API provides a simple method to write optimized code that is portable across multiple hardware vendors and platforms.

OpenVX allows for resource and execution abstractions, which enable hardware vendors to optimize their implementation for their platform. Performance portability across CPUs, GPUs, and special-function hardware are one of the design goals of the OpenVX specification.

OpenVX is used to build, verify, and coordinate computer vision and neural network graph executions. The graph abstraction enables OpenVX implementation to optimize execution for the underlying hardware. Using optimized OpenVX conformant implementation, software developers can spend more time on algorithmic innovations without worrying about the performance and portability of their applications.

OpenVX specification also defines the VXU or the immediate function library. VXU operators allow developers to use all the OpenVX operators as a directly callable C function without creating a graph first. Applications built using the VXU library do not benefit from the optimizations enabled by graph execution. The VXU library can be the simplest way to use OpenVX and is the first step in porting existing vision applications.

AMD OpenVX is a highly optimized open-source implementation of the Khronos OpenVX 1.3 computer vision specification. It allows for rapid prototyping as well as fast execution on a wide range of computer hardware, including small embedded x86 CPUs and large workstation discrete GPUs.

The amd_openvx project consists of the following components:

The OpenVX framework provides a mechanism to add new vision functions to OpenVX by 3rd party vendors. Look into amd_openvx_extensions for additional OpenVX modules and utilities.

  • vx_loomsl: Radeon LOOM stitching library for live 360-degree video applications

  • vx_nn: OpenVX neural network module that was built on top of MIOpen

  • vx_opencv: OpenVX module that implemented a mechanism to access OpenCV functionality as OpenVX kernels

  • vx_rpp: OpenVX extension providing an interface to some of the Radeon Performance Primitives (RPP) functions. This extension is used to enable rocAL to perform image augmentation.

  • vx_winml: OpenVX module that implemented a mechanism to access Windows Machine Learning(WinML) functionality as OpenVX kernels

Features#

  • The code is highly optimized for both x86 CPU and OpenCL/HIP for GPU

  • Supported hardware spans the range from low power embedded APUs, laptops, desktops, and workstation graphics

  • Supports Windows , Linux , and macOS

  • Includes a “graph optimizer” that looks at the entire processing pipeline and removes/replaces/merges functions to improve performance and minimize bandwidth at runtime

  • Scripting support with RunVX allows for rapid prototyping, without re-compiling at production performance levels

Pre-requisites#

  • CPU: 64-bit SSE4.2 or above

  • GPU: Radeon Professional Graphics Cards or Vega Family of Products (16GB required for vx_loomsl and vx_nn libraries)

    • Windows: install the latest drivers and OpenCL SDK download

    • Linux: install ROCm

Build Instructions#

Build this project to generate AMD OpenVX library

Build using Visual Studio#

  • Install OpenCV with/without contrib download for RunVX tool to support camera capture and image display (optional)

    • OpenCV_DIR environment variable should point to OpenCV/build folder

  • Use amd_openvx/amd_openvx.sln to build for x64 platform

  • If AMD GPU (or OpenCL) is not available, set build flag ENABLE_OPENCL=0in openvx/openvx.vcxproj and runvx/runvx.vcxproj

Note: AMD GPU HIP backend is not supported on Windows

Build using CMake#

  • Install CMake 3.0 or later

  • Use CMake to configure and generate Makefile

Contents