Building OpenCV for Windows with CUDA

1. Install CUDA from https://developer.nvidia.com/cuda-downloads;

2. Download cuDNN from https://developer.nvidia.com/cudnn (this requires registration);

3. Install git for windows; for example, use https://gitforwindows.org/;

4. From the command prompt, type git clone https://github.com/opencv/opencv.git; this will create an opencv directory under the current directory (C:\Users\bookWriter, for example);

5. From the command prompt, type git clone https://github.com/opencv/opencv_contrib; this will create an opencv_contrib directory under the current directory;

6. Add C:\Program Files (x86)\Windows Kits\8.1\bin\x86 to the Path environment variables, if not yet included;

7. Install CMake from cmake.org/download; use the .msi file;

8. From the command line, type

   "C:\Program Files\CMake\bin\cmake.exe"
       -B"C:\Users\bookWriter\opencv\build" -
       H"C:\Users\bookWriter\opencv" -
       DOPENCV_EXTRA_MODULES_PATH="C:\Users\bookWriter\opencv_contrib\
       modules" -G"Visual Studio 14 2015 Win64" -
       DBUILD_opencv_world=ON -DWITH_CUDA=ON -DCUDA_FAST_MATH=ON -
       DWITH_CUBLAS=ON -DINSTALL_TESTS=ON -DINSTALL_C_EXAMPLES=ON -
       DBUILD_EXAMPLES=ON -DCMAKE_BUILD_TYPE=Release -
       DBUILD_opencv_gapi=OFF -DWITH_NVCUVID=OFF -DWITH_OPENGL=ON -
       DOPENCV_ENABLE_NONFREE=ON -
       DCUDA_ARCH_PTX=5.2 
   

   C:\Users\bookWriter is the directory within which opencv and opencv_contrib have been downloaded and created; OpenCV is built using Visual Studio 2015 (-G"Visual Studio 14 2015 Win64" option), but other versions of Visual Studio are possible; the code is built for a 5.2 architecture (option DCUDA_ARCH_PTX=5.2); this operation may take several minutes and will create, among others, the OpenCV.sln file under C:\Users\bookWriter\opencv\build;

9. Double click on the OpenCV.sln file;

10. Under Solution 'OpenCV' -> CMakeTargets -> INSTALL -> right click -> build;

11. Add C:\Users\bookWriter\opencv\build\install\x64\vc14\bin to the environment variables.

Please, be aware that OpenCV 4.1.1 requires at least CUDA 10.1.

First OpenCV CUDA example

A first OpenCV CUDA example is reported in Listing 1.

#include <opencv2\opencv.hpp>
#include <opencv2\cudaimgproc.hpp>

#include <opencv2\opencv.hpp>
#include <opencv2\cudaimgproc.hpp>
/********/
/* MAIN */
/********/
int main(){
    std::cout << cv::getBuildInformation() << std::endl;
    
    auto idx = cv::cuda::getCudaEnabledDeviceCount();
    
    if (idx > 0) cv::cuda::printCudaDeviceInfo(0);
    else printf("OpenCV CUDA error!");

    return 0;}

Listing 1. First OpenCV CUDA example.

The code in this example checks whether the OpenCV CUDA installation is in order.
To compile such an example under Visual Studio 2015, do not forget to:

1. add Project -> Properties -> Configuration Properties -> VC++ Directories -> Include Directories -> C:\Users\bookWriter\opencv\build\install\include;

2. add Project -> Properties -> Configuration Properties -> Linker -> General -> C:\Users\bookWriter\opencv\build\install\x64\vc14\lib;

3. add Project -> Properties -> Configuration Properties -> Linker -> Input -> Additional Dependencies -> opencv_world411d.lib for OpenCV version 4.1.1;

Second OpenCV CUDA example

A second OpenCV CUDA example is reported in Listing 2.

#include <iostream>
#include "opencv2\opencv.hpp"
#include "opencv2\core.hpp"
#include "opencv2\highgui.hpp"
#include "opencv2\cudaarithm.hpp"
/********/
/* MAIN */
/********/
int main(){
    cv::Mat h_src = cv::imread("C:\\Users\\bookWriter\\Packt\\
        OpenCVImageTest\\file.png", cv::IMREAD_GRAYSCALE);
    cv::cuda::GpuMat d_dst, d_src;
    d_src.upload(h_src);
    cv::cuda::threshold(d_src, d_dst, 128.0, 255.0, cv::THRESH_BINARY);
    cv::Mat h_dst(d_dst);
    cv::imshow("Processing result", h_dst);
    cv::waitKey();
    return 0; }

Listing 2. Second OpenCV CUDA example.

The code loads a .png image to CPU memory, transfers it to GPU, applies a threshold, moves the result back to the CPU and shows it.