配置CUDA与cuDNN后编译dlib仍提示未找到cuDNN的问题求助

问题背景 #

环境配置：

• Windows 11
• Visual Studio 17 2022
• cuDNN v9.0
• CUDA v12.3

已配置的环境变量：

$env:CUDA_PATH = "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.3"

系统Path包含：

• C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.3\bin
• C:\Program Files\NVIDIA\CUDNN\v9.0\bin

操作流程：

1. 克隆并初始化dlib仓库：

git clone https://github.com/davisking/dlib.git
cd dlib
git submodule init
git submodule update

2. 创建build目录：

cd dlib
mkdir build
cd build

3. 执行CMake配置：

cmake -G "Visual Studio 17 2022" .. -DDLIB_USE_CUDA=1 -DUSE_AVX_INSTRUCTIONS=1 

出现错误：

This warning is for project developers. Use -Wno-dev to suppress it.

-- Found CUDA: C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.3 (found suitable version "12.3", minimum required is "7.5")
-- Looking for cuDNN install...
-- *** cuDNN V5.0 OR GREATER NOT FOUND. ***
-- *** Dlib requires cuDNN V5.0 OR GREATER. Since cuDNN is not found DLIB WILL NOT USE CUDA. ***
-- *** If you have cuDNN then set CMAKE_PREFIX_PATH to include cuDNNs folder. ***
-- Disabling CUDA support for dlib. DLIB WILL NOT USE CUDA
-- Searching for FFMPEG/LIBAV
-- Could NOT find PkgConfig (missing: PKG_CONFIG_EXECUTABLE)
-- PkgConfig could not be found, FFMPEG wont be available
-- Configuring done (7.6s)
-- Generating done (0.0s)
-- Build files have been written to: E:/pythonProj/dlib/build

尝试添加CMAKE_PREFIX_PATH指定cuDNN根目录后，问题未解决：

cmake -G "Visual Studio 17 2022" .. -DDLIB_USE_CUDA=1 -DUSE_AVX_INSTRUCTIONS=1 -DCMAKE_PREFIX_PATH="C:\Program Files\NVIDIA\CUDNN\v9.0"

解决方法 #

查看dlib的CMakeLists.txt后发现，需要额外指定CMAKE_INCLUDE_PATH和CMAKE_LIBRARY_PATH参数，明确指向cuDNN对应CUDA版本的头文件和库文件路径（cuDNN v9.0按CUDA版本划分了子目录）。

最终可用的CMake命令：

cmake -G "Visual Studio 17 2022" .. `
-DDLIB_USE_CUDA=1 -DUSE_AVX_INSTRUCTIONS=1 `
-DCMAKE_PREFIX_PATH="C:\Program Files\NVIDIA\CUDNN\v9.0;C:\Program Files (x86)\Intel\oneAPI\mkl\latest" `
-DCMAKE_INCLUDE_PATH="C:\Program Files\NVIDIA\CUDNN\v9.0\include\12.3;C:\Program Files (x86)\Intel\oneAPI\mkl\latest\include" `
-DCMAKE_LIBRARY_PATH="C:\Program Files\NVIDIA\CUDNN\v9.0\lib\12.3\x64;C:\Program Files (x86)\Intel\oneAPI\mkl\latest\lib"

关键说明 #

cuDNN v9.0的目录结构与旧版本不同，它会为每个CUDA版本创建独立的子目录（比如对应CUDA 12.3的include/12.3和lib/12.3/x64）。仅指定cuDNN根目录的CMAKE_PREFIX_PATH无法让CMake定位到正确的头文件和库，因此必须通过CMAKE_INCLUDE_PATH和CMAKE_LIBRARY_PATH直接指向这些细分路径。

内容的提问来源于stack exchange，提问作者Trevor