如何在CMake中寻找CUDNN？

Question

显然，我已经正确安装了CUDA和CUDNN，但是FindCUDA可以找到CUDA，但FindCUDNN.cmake无法找到CUDNN。

我需要检查什么以调试这个问题？ 操作系统：Windows 10 cmake版本：3.17.2 CUDA版本：10.2 CUDNN版本：cudnn-10.2-windows10-x64-v7.6.5.32

使用安装程序安装了CUDA，路径为C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2 根据指示设置了PATH

将CUDNN文件复制到了相应的bin、include和lib文件夹中。

[cmake] -- Found CUDA: C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v10.2 (found version "10.2") 
[cmake] CMake Error at C:/Program Files/CMake/share/cmake-3.17/Modules/FindPackageHandleStandardArgs.cmake:164 (message):
[cmake]   Could NOT find CUDNN (missing: CUDNN_LIBRARY CUDNN_INCLUDE_DIR)
[cmake] Call Stack (most recent call first):
[cmake]   C:/Program Files/CMake/share/cmake-3.17/Modules/FindPackageHandleStandardArgs.cmake:445 (_FPHSA_FAILURE_MESSAGE)
[cmake]   cmake/FindCUDNN.cmake:90 (find_package_handle_standard_args)
[cmake]   CMakeLists.txt:17 (find_package)

FindCUDNN.cmake: https://gist.github.com/WurmD/26af2940d8dec7cc48d38bc30fd1b3ef 的翻译：

- WurmD

Answer 1

...
set "cudaRoot=C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.3"
...
-DCUDNN_LIBRARY="%cudaRoot%\lib" ^
-DCUDNN_INCLUDE_DIR="%cudaRoot%\include" ^
...

- Lei Chen

Answer 2

我正在使用Visual Studio 2019编译OpenCV；因为我已经解决了这个问题，所以我会假设你也在用 - 因为我很累。

0. 确保在安装CUDNN后重新启动计算机
1. 新克隆opencv和opencv_contrib的GIT存储库
2. 确保在opencv目录中创建BUILD文件夹
3. 设置所需的环境变量；类似于以下内容，但请填写您的路径：

---

    set "CUDA_PATH=............/NVIDIA GPU Computing Toolkit/CUDA/v11.0"
    set "opencvSource=opencv"
    set "opencvExtraModules=opencv_contrib\modules"
    set "opencvBuild=%opencvSource%\build"
    set "compiler=Visual Studio 16 2019"
    set "buildType=Release"
    set "python3_executable=............/AppData/Local/Programs/Python/Python38/python.exe"
    set "python3_include_dir=............/AppData/Local/Programs/Python/Python38/include"
    set "python3_library=............/AppData/Local/Programs/Python/Python38/libs/python38.lib"
    set "python3_numpy_include_dirs=............/AppData/Local/Programs/Python/Python38/Lib/site-packages/numpy/core/include"
    set "python3_packages_path=............/AppData/Local/Programs/Python/Python38/Lib/site-packages"

---

4. 重新运行CMAKE（显然需要根据您的选项进行更改），我的选项是：

---

    cmake ^
    -B"%opencvBuild%/" ^
    -H"%opencvSource%/" ^
    -G"%compiler%" ^
    -Ax64 ^
    -DCMAKE_BUILD_TYPE=%buildType% ^
    -DBUILD_opencv_python3=ON ^
    -DBUILD_opencv_world=ON ^
    -DBUILD_EXAMPLES=OFF ^
    -DBUILD_opencv_python_bindings_generator=ON ^
    -DOPENCV_PYTHON3_VERSION=3.8.3 ^
    -DPYTHON3_EXECUTABLE="%python3_executable%" ^
    -DPYTHON3_INCLUDE_DIR="%python3_include_dir%" ^
    -DPYTHON3_LIBRARY="%python3_library%" ^
    -DPYTHON3_NUMPY_INCLUDE_DIRS="%python3_numpy_include_dirs%" ^
    -DPYTHON3_PACKAGES_PATH="%python3_packages_path%" ^
    -DINSTALL_TESTS=OFF ^
    -DINSTALL_C_EXAMPLES=OFF ^
    -DOPENCV_EXTRA_MODULES_PATH="%opencvExtraModules%/" ^
    -DWITH_CUDA=ON ^
    -DCUDA_ARCH_BIN=6.1^
    -DCUDA_ARCH_PTX=7.5^
    -DCUDA_FAST_MATH=ON ^
    -DBUILD_PROTOBUF=ON 

- Kyle

Answer 3

我也遇到过这个问题。将Win32更改为x64对我有帮助。 我猜在尝试为Win32构建时，它正在查看不同的路径。

- Wim Vanhenden

Answer 4

在安装新版CuDNN时，请确保通过“文件”菜单和“删除缓存”选项来清除CMake的缓存。这将删除所有设置，但可以识别新版CuDNN。

- Chris Fowl