OpenCV tries to load uninstalled Cuda dlls - opencv

Question

I had installed CUDA 9.1 previously and tested my OpenCV with the code below, all worked fine but later on I had to remove it and install 8.0. Now the below code gives errors since the previous dlls are searched.
Here is my test code:
#include <iostream>
#include "opencv2/opencv.hpp"
#include "opencv2/core.hpp"
#include "opencv2/highgui.hpp"
#include "opencv2/cudaarithm.hpp"
using namespace cv;
int main(int argc, char* argv[])
{
try
{
cv::Mat src_host = cv::imread("hdd.png", cv::IMREAD_GRAYSCALE);
cv::cuda::GpuMat dst, src;
src.upload(src_host);
cv::cuda::threshold(src, dst, 128.0, 255.0, cv::THRESH_BINARY);
cv::Mat result_host(dst);
cv::namedWindow("Result", cv::WINDOW_NORMAL);
cv::imshow("Result", result_host);
cv::waitKey();
}
catch (const cv::Exception& ex)
{
std::cout << "Error: " << ex.what() << std::endl;
}
return 0;
}
Here is the error:
The code execution cannot proceed because cudart64_91.dll was not found. Reinstalling the program may fix this problem.
It asks for 2-3 more dlls when I click on OK button, but I am not gonna write them here as I suspect the problem arises from the same source.
The problem is rather obvious. The dlls which are tried to be loaded, belong to the uninstalled version of CUDA (9.1), whereas I have now 8.0. I do not know why my Visual Studio 2013 tries to load the previous ones still.
Before anyone asks, yes I do have my newer installation (8.0) in my PATH. I am using Windows 10 x64, if it matters.

Answer 1

The real problem behind your error:
OpenCV was built with CUDA 9.1, this will not change even if you change the CUDA installation, thus it will the DLLs from CUDA 9.1 will always be required for every program that is compiled with OpenCV. Maybe a module does not have this linked and you can use it... but I am almost sure the main ones do and you won't be able to use them.
Possible solutions:
Build OpenCV with CUDA 8.0 then it will require is CUDA 8.0 DLLs and not the 9.1 ones.
Install CUDA 9.1. Both CUDA can be installed in the same computer, that is why they have this _80 or _91, this way you can have both paths and the computer decide which one is needed... I think it is not possible to have BOTH in the same program though, so be careful with this option. If it is only used by OpenCV then it will be ok.
I would recommend the first option, it is safer to stick to one library version...