This is a starter template for an OpenCV with CUDA application.
So by definition the prerequisite for usage is that you have an NVIDIA graphics card with CUDA support.
The backbone of the template is the cmake-init Project Template .
Not much is added to this, just ThirdParty.cmake file that finds installed dependencies (OpenCV being the only
one for the moment) and a source file source/main/main.cpp that compiles to a target
that tests if OpenCV actually uses CUDA.
For maximum control over the CMake configuration options, it is suggested that you download OpenCV from the official repo, check out the tag of your choice at and configure/build/install it following these steps:
cd opencv
mkdir build
cd build
run configure_opencv.sh script :
sudo cmake -DCMAKE_BUILD_TYPE=RELEASE \
-DCMAKE_C_COMPILER=/usr/bin/gcc-9 \
-DCMAKE_CXX_COMPILER=/usr/bin/g++-9 \
-DCMAKE_INSTALL_PREFIX=/usr/local \
-DWITH_CUDA=ON \
-DWITH_CUDDNN=ON \
-DOPENCV_DNN_CUDA=ON \
-DENABLE_FAST_MATH=1 \
-DCUDA_FAST_MATH=1 \
-DCUDA_ARCH_BIN=7.5 \
-DWITH_CUBLAS=1 \
-DCUDA_NVCC_EXECUTABLE=/opt/cuda/bin/nvcc \
-DBUILD_opencv_cudalegacy=OFF \
-DINSTALL_PYTHON_EXAMPLES=OFF \
-DOPENCV_EXTRA_MODULES_PATH=../../opencv_contrib/modules \
-DBUILD_EXAMPLES=OFF \
-DBUILD_TESTS=OFF \
-DBUILD_INFO_SKIP_EXTRA_MODULES=OFF \
-DBUILD_opencv_cudacodec=ON \
Notes on this script:
- You need g++-9 to compile OpenCV with CUDA
- CUDA_ARCH_BIN is a parameter that depends on your graphic card, you will find it on the NVIDIA site.
- CUDA needs to be installed in your system
- cudnn (Nvidia CUDA Deep Neural Network) needs to be installed in your system
- OPENCV_EXTRA_MODULES_PATH is the path to opencv-contrib (you must clone the repo and set the parameter to the local path)
After you have configured the project with cmake, find the number of your CPU cores with
nproc and use the number to build the project with make
make -j24 (for a processor with 24 cores.)
Last, install OpenCV in your system with
sudo make install
With successful installation of the CUDA enabled OpenCV libraries, find_package(OpenCV REQUIRED) in
ThirdParty.cmake file will find and link the OpenCV libraries to your project.
You configure and build the application with CMake as usual.
See ADAPT.md on how to adapt the template to your application.
Important: Your system must have a CUDA compatible Nvidia Driver installed and the version must be compatible CUDA version. See resources for more info.