win10编译opencv4.5.0（gpu版）

正常官方推荐的exe安装&＃xff0c;调用cv::dnn模块运行&＃xff0c;超级慢&＃xff0c;而且打印信息&＃xff1a;

setUpNet DNN module was not built with CUDA backend; switching to CPU

即使加上加速代码&＃xff0c;依然超级慢&＃xff0c;跑512x512图像&＃xff0c;分割网络需要1s&＃xff0c;2080ti 7.5算力&＃xff1a;

this->loc_net.setPreferableBackend(cv::dnn::DNN_BACKEND_CUDA);
this->loc_net.setPreferableTarget(cv::dnn::DNN_TARGET_CUDA);

所以需要编译源码&＃xff0c;开启cuda加速模块。

&＃xff08;1&＃xff09;win10&＃xff0c;visual studio 2019 (试过2015编译失败)&＃xff0c;cuda10.2&＃xff0c;cudnn8.0.3&＃xff0c;opencv4.5.0, opencv-contrib-4.5.0.

&＃xff08;2&＃xff09;或者&＃xff0c;visual studio 2019&＃43;cuda11.4&＃43;cudnn8.2.2&＃xff08;8.0.3&＃xff09;&＃43;opencv4.5.0&＃43;opencv-contrib4.5.0

&＃xff08;3&＃xff09; visual studio 2019&＃43;cuda11.1&＃43;cudnn8.0.3&＃43;opencv4.5.0&＃43;opencv-contrib4.5.0

下载opencv4.5.0

Releases - OpenCVhttps://opencv.org/releases/page/2/下载opencv-contrib4.5.0

https://github.com/opencv/opencv_contrib/archive/refs/tags/4.5.0.ziphttps://github.com/opencv/opencv_contrib/archive/refs/tags/4.5.0.zip下载.cache文件夹

opencv4.5.0对应的.cache文件夹-互联网文档类资源-CSDN下载里面的文件完整&＃xff0c;如下所示&＃xff1a;adedataffmpegippicvnvidia_optica更多下载资源、学习资料请访问CSDN下载频道.https://download.csdn.net/download/jizhidexiaoming/82456363

下载安装cmake 3.23

https://github.com/Kitware/CMake/releases/download/v3.23.0-rc2/cmake-3.23.0-rc2-windows-x86_64.msihttps://github.com/Kitware/CMake/releases/download/v3.23.0-rc2/cmake-3.23.0-rc2-windows-x86_64.msicmake配置源路径和目标路径

点击configure 

等待几分钟。

&＃xff08;1&＃xff09;勾选CUDA&＃xff1b;

&＃xff08;2&＃xff09;设置opencv_contrib路径

 &＃xff08;3&＃xff09;勾选build_opencv_world

&＃xff08;4&＃xff09;去掉python&＃xff0c;test、java加快后面的编译

再次点击configure 

&＃xff08;5&＃xff09;cuda_arch_bin算力设置

算力查询&＃xff1a; 

CUDA GPU | NVIDIA Developerhttps://developer.nvidia.com/zh-cn/cuda-gpus

或者cuda_generation设置为auto

&＃xff08;6&＃xff09;勾选cuda_fast_math

&＃xff08;7&＃xff09; 去掉setupvars

点击configure&＃xff0c;再点击generate

这时候在生成的文件夹下搜索opencv_world450d.lib是没有的&＃xff0c;需要vs继续编译。

 vs2019打开opencv.sln&＃xff0c;进行编译

 打开界面如下。

 任务栏-》生成-》配置管理器&＃xff0c;勾选INSTALL

 点击生成解决方案。等待两小时。。。编译成功。

&＃xff08;1&＃xff09;2080ti的显卡算力是7.5&＃xff0c;在2080ti编译的opencv&＃xff0c;放到3090显卡算力8.6上跑报错&＃xff1a;

OpenCV(4.5.0) Error: No CUDA support (OpenCV was not built to work with the selected device.

备忘下&＃xff0c;原因还未知。。。继续第二步探索

&＃xff08;2&＃xff09;3090显卡上编译opencv&＃xff0c;设置cuda_arch_bin &＃61; 7.5;8.6&＃xff0c;vs2019编译时报错&＃xff1a;

CMake Error at cuda_compile_1_generated_gpu_mat.cu.obj.Debug.cmake:22.

 备忘下&＃xff0c;原因还未知。。。继续第三步探索

&＃xff08;3&＃xff09;3090显卡上编译opencv&＃xff0c;设置cuda_generation&＃61; auto&＃xff0c;点击configure后&＃xff0c;cuda_arch_bin会自动识别到当前值应该为8.6

 继续探索

&＃xff08;4&＃xff09;编译3090编译opencv时报错&＃xff08;此时设置cuda_arch_bin&＃61;8.6&＃xff09;&＃xff1a;

nvcc fatal   : Unsupported gpu architecture &＃39;compute_86&＃39;

LINK : fatal error LNK1104: 无法打开文件“..\..\lib\Debug\opencv_world450d.lib”

根据以下链接建议升高cuda版本&＃xff08;当前是10.2&＃xff0c;建议升到11.4&＃xff09;。

ubuntu - nvcc fatal : Unsupported gpu architecture &＃39;compute_86&＃39; - Stack Overflow

 &＃xff08;5&＃xff09;基于第4步&＃xff0c;编译成功后进行测试&＃xff0c;问题&＃xff1a;

initCUDABackend CUDA backend will fallback to the CPU implementation for

原因未知。待解决。。。

&＃xff08;6&＃xff09;cuda\vc140.pdb, has an obsolete format, delete it and recompile

成功编译完release版本后&＃xff0c;紧接着编译debug报这个错误&＃xff0c;原因可能是有release生成的文件冲突了&＃xff0c;解决办法&＃xff1a;删掉编译release生成的所有文件&＃xff0c;重新编译&＃xff0c;或者更彻底的办法是重新走一遍上面的流程&＃xff0c;即用cmake重新生成opencv.sln&＃xff0c;然后重新编译debug版本。