Docker容器构建项目遇CMake错误：无法找到cudart库

环境信息 #

• 主机：Ubuntu 18.04，NVIDIA Jetson TX2
• 预装组件：Docker、nvidia-docker2、l4t-cuda
• 使用Docker镜像：stereolabs/zed:3.7-devel-jetson-jp4.6

核心编译错误 #

CMake Error at /usr/local/lib/python3.6/dist-packages/cmake/data/share/cmake-3.25/Modules/FindPackageHandleStandardArgs.cmake:230 (message):
  Could NOT find CUDA (missing: CUDA_CUDART_LIBRARY) (found suitable version
  "10.2", minimum required is "10.2")
Call Stack (most recent call first):
  /usr/local/lib/python3.6/dist-packages/cmake/data/share/cmake-3.25/Modules/FindPackageHandleStandardArgs.cmake:600 (_FPHSA_FAILURE_MESSAGE)
  /usr/local/lib/python3.6/dist-packages/cmake/data/share/cmake-3.25/Modules/FindCUDA.cmake:1266 (find_package_handle_standard_args)
  CMakeLists.txt:17 (find_package)

CMakeLists.txt完整内容 #

cmake_minimum_required (VERSION 3.5)

project(vision)
enable_testing()

# Variables scopes follow standard rules
# Variables defined here will carry over to its children, ergo subdirectories

# Setup ZED libs
find_package(ZED 3 REQUIRED)
include_directories(${ZED_INCLUDE_DIRS})
link_directories(${ZED_LIBRARY_DIR})

# Setup CUDA libs for zed and ai modules
find_package(CUDA ${ZED_CUDA_VERSION} REQUIRED)
include_directories(${CUDA_INCLUDE_DIRS})
link_directories(${CUDA_LIBRARY_DIRS})

# Setup OpenCV libs
find_package(OpenCV REQUIRED)
include_directories(${OpenCV_INLCUDE_DIRS})
# Check if OpenMP is installed 
find_package(OpenMP)
checkPackage("OpenMP" "OpenMP not found, please install it to improve performances: 'sudo apt install libomp-dev'")

# TensorRT
set(TENSORRT_ROOT /usr/src/tensorrt/)
find_path(TENSORRT_INCLUDE_DIR NvInfer.h
        HINTS ${TENSORRT_ROOT} PATH_SUFFIXES include/)
message(STATUS "Found TensorRT headers at ${TENSORRT_INCLUDE_DIR}")

set(MODEL_INCLUDE ../code/includes)
set(MODEL_LIB_DIR libs)
set(YAML_INCLUDE ../depends/yaml-cpp/include)
set(YAML_LIB_DIR ../depends/yaml-cpp/libs)

include_directories(${MODEL_INCLUDE} ${YAML_INCLUDE})
link_directories(${MODEL_LIB_DIR} ${YAML_LIB_DIR})

# Setup Darknet libs
#find_library(DARKNET_LIBRARY NAMES dark libdark.so libdarknet.so)
#find_package(dark REQUIRED)

# Setup HTTP libs
find_package(httplib REQUIRED)
find_package(nlohmann_json 3.2.0 REQUIRED)

# System libs
SET(SPECIAL_OS_LIBS "pthread")
link_libraries(stdc++fs)

# Optional definitions
add_definitions(-std=c++17 -g -O3)

# Add sub directories
add_subdirectory(zed_module)
add_subdirectory(ai_module)
add_subdirectory(http_api_module)
add_subdirectory(executable_module)

option(RUN_TESTS "Build the tests" off)

if (RUN_TESTS OR CMAKE_BUILD_TYPE MATCHES Debug)
    add_subdirectory(test)

Dockerfile中失败的构建步骤 #

WORKDIR /opt

RUN git clone https://github.com/Cruiz102/Vision-Module
WORKDIR /opt/Vision-Module 
RUN mkdir build-debug && cd build-debug
RUN pwd
WORKDIR /opt/Vision-Module/build-debug
RUN cmake -DCMAKE_BUILD_TYPE=Release ..

/etc/docker/daemon.json配置内容 #

{
    "runtimes": {
        "nvidia": {
            "path": "/usr/bin/nvidia-container-runtime",
            "runtimeArgs": []
        }
    },
    "default-runtime": "nvidia"
}

已尝试但无效的解决方法 #

• 指定工具包路径
• 编辑daemon.json配置
• 重新安装依赖组件
• 更换Docker镜像
• 在主机重新安装cudart库
• 修改编译参数
• 采用交互模式构建容器

可行解决建议 #

1. 显式指定CUDA根目录：Jetson平台CUDA默认路径为/usr/local/cuda-10.2，在CMake命令中传入该路径：

   cmake -DCMAKE_BUILD_TYPE=Release -DCUDA_ROOT=/usr/local/cuda-10.2 ..
   

2. 检查容器内CUDA库存在性：进入容器交互模式，执行以下命令确认libcudart.so是否存在：

   ls /usr/local/cuda/lib64/libcudart.so
   

   若不存在，需确认使用的Docker镜像是否正确包含CUDA runtime环境。
3. 替换CMake CUDA查找模块：CMake 3.25版本的FindCUDA模块对Jetson兼容性较差，改用FindCUDAToolkit模块，修改CMakeLists.txt中CUDA相关代码：

   # 替换原find_package(CUDA ...)部分
   find_package(CUDAToolkit ${ZED_CUDA_VERSION} REQUIRED)
   include_directories(${CUDAToolkit_INCLUDE_DIRS})
   link_directories(${CUDAToolkit_LIBRARY_DIR})
   

4. 确认nvidia容器运行时生效：构建镜像时添加--runtime=nvidia参数，或重启Docker服务确保daemon.json配置生效：

   sudo systemctl daemon-reload
   sudo systemctl restart docker
   

内容的提问来源于stack exchange，提问作者atlas7