Build an OpenCV YOLOX pipeline to benchmark KleidiCV preprocessing
Introduction
Understand YOLOX and KleidiCV in the detection pipeline
Build OpenCV with and without KleidiCV
Prepare the YOLOX model and test image
Compile and run YOLOX static-image detection
Evaluate KleidiCV preprocessing and pipeline performance
Next Steps

  Log an issue

  Fork and edit

  Discuss on Discord

Check the arm64 environment

Confirm that the board runs a 64-bit Arm userspace:

    

        
        
uname -m

The expected output is:

    

        
        aarch64

Record the operating system, kernel, CPU topology, and CPU features:

    

        
        
cat /etc/os-release
uname -r
lscpu

Keep this information with your benchmark results. KleidiCV selects supported implementations for the available CPU features.

Note

The reference measurements listed in this Learning Path were collected on a Radxa Orion O6 running Debian 12. The board has eight Cortex-A720 CPUs and four Cortex-A520 CPUs. You can use another arm64 Linux system, but CPU topology, supported extensions, and performance results will differ.

Install the build tools

Install the compiler, CMake, Ninja, Git, and image codec development packages:

    

        
        
sudo apt update
sudo apt install -y \
    build-essential \
    ca-certificates \
    cmake \
    file \
    git \
    libjpeg-dev \
    libpng-dev \
    libtiff-dev \
    ninja-build \
    pkg-config \
    util-linux \
    wget

Record the compiler and CMake versions:

    

        
        
c++ --version
cmake --version

Use GCC 10 or later, or Clang 12 or later. If compilation exhausts memory, reduce the --parallel value in the build commands.

Download OpenCV

Create a workspace and check out OpenCV 4.13.0:

    

        
        
mkdir -p ~/opencv-kleidicv-lab
cd ~/opencv-kleidicv-lab
git clone --branch 4.13.0 --depth 1 https://github.com/opencv/opencv.git

Download KleidiCV 26.06

Clone the KleidiCV 26.06 release:

    

        
        
cd ~/opencv-kleidicv-lab
git clone --branch 26.06 --depth 1 \
    https://gitlab.arm.com/kleidi/kleidicv.git

Confirm the selected release:

    

        
        
git -C kleidicv describe --tags --exact-match

The expected output is:

    

        
        26.06

OpenCV 4.13.0 downloads KleidiCV 0.7.0 when KLEIDICV_SOURCE_PATH isn’t provided. For the Learning Path, KleidiCV 26.06 is explicitly selected. This is because KleidiCV 26.06 supports the three-channel, arbitrary-ratio resize used by the letterbox stage.

Both builds use the same OpenCV source tree, compiler, and common CMake options. The intended difference is enabling KleidiCV with its default dispatch policy.

Note

Leave OpenCV 4.13.0 unpatched in both builds. The KleidiCV 26.06 OpenCV integration patch adds a copyMakeBorder() Hardware Acceleration Layer (HAL) hook that’s absent from this OpenCV release. You’ll compare preprocessing without that patch, so padding remains an OpenCV operation. Applying the patch creates a different test and requires new measurements.

Build the baseline OpenCV installation

Configure a baseline OpenCV installation without KleidiCV:

    

        
        
cd ~/opencv-kleidicv-lab

cmake -S opencv -B build-baseline -G Ninja \
    -DCMAKE_BUILD_TYPE=Release \
    -DCMAKE_INSTALL_PREFIX=$HOME/opencv-kleidicv-lab/install-baseline \
    -DBUILD_LIST=core,imgproc,imgcodecs,dnn \
    -DBUILD_EXAMPLES=OFF \
    -DBUILD_opencv_apps=OFF \
    -DBUILD_PERF_TESTS=OFF \
    -DBUILD_TESTS=OFF \
    -DBUILD_SHARED_LIBS=ON \
    -DWITH_IPP=OFF \
    -DWITH_OPENCL=OFF \
    -DWITH_KLEIDICV=OFF

Build and install the baseline:

    

        
        
cmake --build build-baseline --parallel "$(nproc)"
cmake --install build-baseline

List the installed OpenCV libraries:

    

        
        
find install-baseline/lib -maxdepth 1 -type l -name 'libopencv_*.so' \
    | sort

The expected output is:

    

        
        install-baseline/lib/libopencv_core.so
install-baseline/lib/libopencv_dnn.so
install-baseline/lib/libopencv_imgcodecs.so
install-baseline/lib/libopencv_imgproc.so

Confirm the baseline setting:

    

        
        
grep '^WITH_KLEIDICV' build-baseline/CMakeCache.txt

The expected output is:

    

        
        WITH_KLEIDICV:BOOL=OFF

Build OpenCV with KleidiCV

Configure the second build with KleidiCV enabled. Leave its dispatch options at their defaults:

    

        
        
cd ~/opencv-kleidicv-lab

cmake -S opencv -B build-kleidicv-default -G Ninja \
    -DCMAKE_BUILD_TYPE=Release \
    -DCMAKE_INSTALL_PREFIX=$HOME/opencv-kleidicv-lab/install-kleidicv-default \
    -DBUILD_LIST=core,imgproc,imgcodecs,dnn \
    -DBUILD_EXAMPLES=OFF \
    -DBUILD_opencv_apps=OFF \
    -DBUILD_PERF_TESTS=OFF \
    -DBUILD_TESTS=OFF \
    -DBUILD_SHARED_LIBS=ON \
    -DWITH_IPP=OFF \
    -DWITH_OPENCL=OFF \
    -DWITH_KLEIDICV=ON \
    -DKLEIDICV_SOURCE_PATH=$HOME/opencv-kleidicv-lab/kleidicv

Build and install it:

    

        
        
cmake --build build-kleidicv-default --parallel "$(nproc)"
cmake --install build-kleidicv-default

Confirm that KleidiCV is enabled:

    

        
        
grep '^WITH_KLEIDICV' build-kleidicv-default/CMakeCache.txt

The expected output is:

    

        
        WITH_KLEIDICV:BOOL=ON

Confirm that the KleidiCV libraries were built:

    

        
        
find build-kleidicv-default/3rdparty/lib -maxdepth 1 -type f \
    -name 'libkleidicv*.a' -printf '%f\n' | sort

The expected output is:

    

        
        libkleidicv.a
libkleidicv_hal.a
libkleidicv_thread.a

What you’ve accomplished and what’s next

You’ve created baseline and default KleidiCV 26.06 OpenCV installations from the same unpatched source.

Next, you’ll download the YOLOX model and a static test image.

Back
Next