I want to use gocv in Go with CUDA support on windows 11. But when I run my Program, the task is only using the CPU.
This shows me this command.
nvidia-smi
No processes are found on GPU during executing my Go Code.
gocv is installed as described here and works properly. The instruction is not 100% correct, but I got it running.
Anyway, testing it with following command fails:
PS C:\...\cuda>go run .\main.go
# gocv.io/x/gocv/cuda
In file included from arithm.cpp:2:
arithm.h:9:10: fatal error: opencv2/cudaarithm.hpp: No such file or directory
#include <opencv2/cudaarithm.hpp>
^~~~~~~~~~~~~~~~~~~~~~~~
compilation terminated.
So after litte research I found this thread on github. I did the git clone and created my build directory. I changed the CMakeLists.txt
set(OpenCV_DIR "C:/opencv/build")
set(CUDA_TOOLKIT_ROOT_DIR "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.8")
I run cmake .. -G "Visual Studio 17 2022" and I got the cvglue.sln - File in my build directory.
This throws no error.
But when I build project INSTALL I get errors and it creates only "C:\opencv\cvglue_4.7.0\lib\gocv.lib".
During building the project in VS 2022 I got lot of errors like:
2>imgproc.obj : error LNK2001: Nicht aufgelöstes externes Symbol ""public: __cdecl cv::Mat::~Mat(void)" (??1Mat@cv@@QEAA@XZ)".
Schweregrad Code Beschreibung Projekt Pfad Datei Zeile Quelle Unterdrückungszustand Details
Fehler LNK2001 Nicht aufgelöstes externes Symbol ""class cv::debug_build_guard::_InputOutputArray const & __cdecl cv::noArray(void)" (?noArray@cv@@YAAEBV_InputOutputArray@debug_build_guard@1@XZ)". gocv C:\opencv\build\install\gocv\cvglue\build\gocv C:\opencv\build\install\gocv\cvglue\build\gocv\core.obj 1 Build
Is there any way to debug this? Does anybody know how to set include paths correctly?
Or was anybody able to run CUDA under windows in gocv and can give me instructions?
I tried to include the C:\opencv\build\install\include\opencv2 in VS but it did not change anything.
After lot of try and error and research I got it running!
I want to share my procedure and give you detailed instructions:
Optional you can build OpenCV from source if you are able to. Maybe higher versions of NVIDIA tools are also possible, but I did not try it. It could be that OpenCV 4.9.0 needs CUDA 12.3 cuDNN 8.9.7
First I installed Visual Studio 17 2022 and CMake. There is nothing special.
Then I run the NVIDIA Toolkit installer. After the installation, the toolkit should be installed in C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.3
Copy the files from cuDNN\bin, \lib\x64 and \include to CUDA\v12.3\bin, lib\x64 and \include.Filestructure From the Nvidia Video Codec SDK I copied the files from Lib to lib and from Interface to include. Additionally I added the following Paths in my Environment:
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.3\include
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.3\lib\x64
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.3\bin
It should look like this: Environment Variables
Now it is time to add OpenCV. Run the follwing commands in a Command Prompt:
cd C:\
mkdir opencv
cd C:\opencv
mkdir build
Extract the OpenCV 4.9.0 Prebuild opencv_contrib_cuda_4.9.0_win_amd64.7z to the build Directory.
Now let's add gocv.
cd %GOPATH%\pkg\mod\github.com
mkdir garfeng
cd garfeng
git clone https://github.com/garfeng/gocv.git
cd gocv
git checkout cuda_win
cd cvglue/
mkdir build
cd build
Now make changes to the gocv\cvglue\CMakeLists.txt File
Change the following two lines:
set(OpenCV_DIR C:/opencv/4.6.0-cuda/3.5-8.6/)
set(CUDA_TOOLKIT_ROOT_DIR D:/cuda/)
To:
set(OpenCV_DIR C:/opencv/build/x64/vc17/lib)
set(CUDA_TOOLKIT_ROOT_DIR "C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.3")
Then run:
cmake .. -G "Visual Studio 17 2022"
This can throw following error:
Could NOT find CUDAToolkit: Found unsuitable version "12.3.52", but required is exact version "12.3.103" (found C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.3/include)
For workaround I opened
C:\opencv\build\x64\vc17\lib\OpenCVConfig.cmake
I changed
set(OpenCV_CUDA_VERSION "12.3.103")
To:
set(OpenCV_CUDA_VERSION "12.3.52")
And re-run cmake .. -G "Visual Studio 17 2022"
CMake will Generate a Visual studio solution in the build dir. Open the cvglue.sln with VS, then build Release x64, build project INSTALL in the solution explorer.
It will copy all of the dlls to dir C:/opencv/cvglue_4.7.0
It takes some time to build. Rename the Folder only to cvglue
/c/opencv/cvglue
$ tree
.
├── bin
│ ├── gocv.dll
│ ├── gocv_contrib.dll
│ └── gocv_cuda.dll
└── lib
├── gocv.lib
├── gocv_contrib.lib
└── gocv_cuda.lib
2 directories, 6 files
Add the new created lib and bin folder to your environment paths as well.
C:\opencv\cvglue\bin
C:\opencv\cvglue\lib
Now it is time to validate.
cd %GOPATH%\pkg\mod\github.com\garfeng\gocv\cmd\cuda
go run main.go
gocv version: 0.32.0
cuda information:
Device 0: "NVIDIA GeForce RTX 4060 Laptop GPU" 8188Mb, sm_89, Driver/Runtime ver.12.50/12.30
Usage Create your go project. Import gocv.io/x/gocv as usually. Add the following line to your go.mod
replace gocv.io/x/gocv v0.36.1 => github.com/garfeng/gocv v0.31.5
Maybe you have to rename \garfeng\gocv\
To
\garfeng\gocv@v0.31.5\
That's it.
References: https://github.com/garfeng/gocv/blob/cuda_win/UseCudaOnWindows.md