UbuntuでCUDAを有効にしたOpenCVをビルド #Python3

使用した環境その1

OS:Ubuntu18.04.3 LTS
CPU:Intel Corei7-7700HQ
GPU:NVIDIA Quadro M1200

使用した環境その2(2020/8/15追記)

OS:Ubuntu20.04.1 LTS
CPU:Intel Corei7-7700HQ
GPU:NVIDIA Quadro M1200

必要なものをインストール

GPUのドライバ関係(18.04/20.04共通)

sudo apt update && sudo apt upgrade
sudo ubuntu-drivers autoinstall
sudo reboot

sudo apt-key adv --fetch-keys http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/7fa2af80.pub

wget http://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64/cuda-repo-ubuntu1804_10.1.168-1_amd64.deb
sudo dpkg -i cuda-repo-ubuntu1804_10.1.168-1_amd64.deb
sudo apt update 

sudo apt install cuda cuda-drivers

.bashrcに以下を追記

export PATH="/usr/local/cuda/bin:$PATH"
export LD_LIBRARY_PATH="/usr/local/cuda/lib64:$LD_LIBRARY_PATH"

CUDNNのインストール(18.04/20.04共通)

この記事を参考にインストールからインストール

再起動する(18.04/20.04共通)

sudo reboot

パッケージ(18.04)

sudo apt install python3-dev python3-pip gcc-6 g++-6 cmake build-essential git pkg-config ccache \
unzip ecm extra-cmake-modules mesa-utils libgtk-3-dev ffmpeg \
libavcodec-dev libavformat-dev libswscale-dev libtbb-dev libogre-1.9-dev \
libjpeg-dev libpng-dev libtiff5-dev x264 libxvidcore-dev yasm \
libxine2-dev libv4l-dev libfaac-dev libmp3lame-dev libopencore-amrnb-dev \
libopencore-amrwb-dev libtheora-dev libvorbis-dev libxvidcore-dev \
x264 v4l-utils ffmpeg libdc1394-22 libdc1394-22-dev libtiff5-dev \
qt5-default libeigen3-dev libeigen3-doc tesseract-ocr tesseract-ocr-jpn \
vtk7 tcl-vtk7 python3-vtk7 libgflags-dev autoconf automake libtool \
autoconf-archive libleptonica-dev libtesseract-dev gphoto2 liblapacke-dev

pip3 install numpy

パッケージ(20.04)

sudo apt install python3-dev python3-pip gcc-7 g++-7 cmake build-essential git pkg-config ccache \
unzip extra-cmake-modules mesa-utils libgtk-3-dev ffmpeg \
libavcodec-dev libavformat-dev libswscale-dev libtbb-dev libogre-1.9-dev \
libjpeg-dev libpng-dev libtiff5-dev x264 libxvidcore-dev yasm \
libxine2-dev libv4l-dev libfaac-dev libmp3lame-dev libopencore-amrnb-dev \
libopencore-amrwb-dev libtheora-dev libvorbis-dev libxvidcore-dev \
x264 v4l-utils ffmpeg libdc1394-22 libdc1394-22-dev libtiff5-dev \
qt5-default libeigen3-dev libeigen3-doc tesseract-ocr tesseract-ocr-jpn \
vtk7 tcl-vtk7 python3-vtk7 libgflags-dev autoconf automake libtool \
autoconf-archive libleptonica-dev libtesseract-dev gphoto2 liblapacke-dev

pip3 install numpy

メイン

OpenCVのソースをダウンロード(18.04/20.04共通)

cd ~/
git clone --depth=1 https://github.com/opencv/opencv.git
git clone --depth=1 https://github.com/opencv/opencv_contrib.git
cd opencv && mkdir build && cd build

Cmakeする(18.04)

CC=gcc-6 CXX=g++-6 cmake -D CMAKE_BUILD_TYPE=RELEASE -D OPENCV_EXTRA_MODULES_PATH=../../opencv_contrib/modules \
 -D OPENCV_GENERATE_PKGCONFIG=ON -D BUILD_opencv_apps=ON -D BUILD_opencv_calib3d=ON -D BUILD_opencv_core=ON \
 -D BUILD_opencv_cudaarithm=ON -D BUILD_opencv_cudabgsegm=ON -D BUILD_opencv_cudacodec=ON \
 -D BUILD_opencv_cudafeatures2d=ON -D BUILD_opencv_cudafilters=ON -D BUILD_opencv_cudaimgproc=ON \
 -D BUILD_opencv_cudalegacy=ON -D BUILD_opencv_cudaobjdetect=ON -D BUILD_opencv_cudaoptflow=ON \
 -D BUILD_opencv_cudastereo=ON -D BUILD_opencv_cudawarping=ON -D BUILD_opencv_cudev=ON \
 -D BUILD_opencv_features2d=ON -D BUILD_opencv_flann=ON -D BUILD_opencv_highgui=ON -D BUILD_opencv_imgcodecs=ON \
 -D BUILD_opencv_imgproc=ON -D BUILD_opencv_ml=ON -D BUILD_opencv_objdetect=ON \
 -D BUILD_opencv_photo=ON -D BUILD_opencv_stitching=ON -D BUILD_opencv_superres=ON \
 -D BUILD_opencv_ts=ON -D BUILD_opencv_video=ON -D BUILD_opencv_videoio=ON -D BUILD_opencv_videostab=ON \
 -D WITH_1394=ON -D WITH_CUBLAS=ON -D WITH_CUDA=ON -D WITH_CUFFT=ON -D WITH_EIGEN=ON -D WITH_FFMPEG=ON \
 -D WITH_GDAL=OFF -D WITH_GPHOTO2=ON -D WITH_GIGEAPI=ON -D WITH_GSTREAMER=ON -D WITH_GTK=ON \
 -D WITH_INTELPERC=OFF -D WITH_IPP=ON -D WITH_IPP_A=OFF -D WITH_JASPER=ON -D WITH_JPEG=ON -D WITH_LIBV4L=ON \
 -D WITH_OPENCL=ON -D WITH_OPENCLAMDBLAS=OFF -D WITH_OPENCLAMDFFT=OFF -D WITH_OPENCL_SVM=OFF \
 -D WITH_OPENEXR=ON -D WITH_OPENGL=ON -D WITH_OPENMP=OFF -D WITH_OPENNI=OFF -D WITH_PNG=ON \
 -D WITH_PTHREADS_PF=OFF -D WITH_PVAPI=ON -D WITH_QT=ON -D WITH_TBB=ON -D WITH_TIFF=ON \
 -D WITH_UNICAP=OFF -D WITH_V4L=ON -D WITH_VTK=ON -D WITH_WEBP=ON -D WITH_XIMEA=OFF -D WITH_XINE=OFF \
 -D CUDA_NVCC_FLAGS=--expt-relaxed-constexpr -D CUDA_FAST_MATH=ON -D CUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda \
 -D CUDA_HOST_COMPILER=/usr/bin/gcc-6 -D PYTHON_DEFAULT_EXECUTABLE=python3  ..

Cmakeする(20.04)

CC=gcc-7 CXX=g++-7 cmake -D CMAKE_BUILD_TYPE=RELEASE -D OPENCV_EXTRA_MODULES_PATH=../../opencv_contrib/modules \
 -D OPENCV_GENERATE_PKGCONFIG=ON -D BUILD_opencv_apps=ON -D BUILD_opencv_calib3d=ON -D BUILD_opencv_core=ON \
 -D BUILD_opencv_cudaarithm=ON -D BUILD_opencv_cudabgsegm=ON -D BUILD_opencv_cudacodec=ON \
 -D BUILD_opencv_cudafeatures2d=ON -D BUILD_opencv_cudafilters=ON -D BUILD_opencv_cudaimgproc=ON \
 -D BUILD_opencv_cudalegacy=ON -D BUILD_opencv_cudaobjdetect=ON -D BUILD_opencv_cudaoptflow=ON \
 -D BUILD_opencv_cudastereo=ON -D BUILD_opencv_cudawarping=ON -D BUILD_opencv_cudev=ON \
 -D BUILD_opencv_features2d=ON -D BUILD_opencv_flann=ON -D BUILD_opencv_highgui=ON -D BUILD_opencv_imgcodecs=ON \
 -D BUILD_opencv_imgproc=ON -D BUILD_opencv_ml=ON -D BUILD_opencv_objdetect=ON \
 -D BUILD_opencv_photo=ON -D BUILD_opencv_stitching=ON -D BUILD_opencv_superres=ON \
 -D BUILD_opencv_ts=ON -D BUILD_opencv_video=ON -D BUILD_opencv_videoio=ON -D BUILD_opencv_videostab=ON \
 -D WITH_1394=ON -D WITH_CUBLAS=ON -D WITH_CUDA=ON -D WITH_CUFFT=ON -D WITH_EIGEN=ON -D WITH_FFMPEG=ON \
 -D WITH_GDAL=OFF -D WITH_GPHOTO2=ON -D WITH_GIGEAPI=ON -D WITH_GSTREAMER=ON -D WITH_GTK=ON \
 -D WITH_INTELPERC=OFF -D WITH_IPP=ON -D WITH_IPP_A=OFF -D WITH_JASPER=ON -D WITH_JPEG=ON -D WITH_LIBV4L=ON \
 -D WITH_OPENCL=ON -D WITH_OPENCLAMDBLAS=OFF -D WITH_OPENCLAMDFFT=OFF -D WITH_OPENCL_SVM=OFF \
 -D WITH_OPENEXR=ON -D WITH_OPENGL=ON -D WITH_OPENMP=OFF -D WITH_OPENNI=OFF -D WITH_PNG=ON \
 -D WITH_PTHREADS_PF=OFF -D WITH_PVAPI=ON -D WITH_QT=ON -D WITH_TBB=ON -D WITH_TIFF=ON \
 -D WITH_UNICAP=OFF -D WITH_V4L=ON -D WITH_VTK=ON -D WITH_WEBP=ON -D WITH_XIMEA=OFF -D WITH_XINE=OFF \
 -D CUDA_NVCC_FLAGS=--expt-relaxed-constexpr -D CUDA_FAST_MATH=ON -D CUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda-11.0 \
 -D CUDA_HOST_COMPILER=/usr/bin/gcc-7 ..

makeする(CPUに依存するけどi7-7700HQで1時間くらいかかった)(18.04/20.04共通)

make -j$(nproc)

インストール(18.04/20.04共通)

sudo make install

インストールが成功しているか確認

以下を実行し0より大きい値が返ってくれば成功

python3

import cv2
print(cv2.cuda.getCudaEnabledDeviceCount())

もし失敗してたら(18.04/20.04共通)

cd ~/opencv/
rm -rf build

でbuildフォルダを消してからCmakeからやり直してください

参考

https://qiita.com/shinohe/items/d41f77edac27be9cef57
https://qiita.com/gusugusu1018/items/2115343a8adce378ff4a
https://qiita.com/nakamura21/items/2b2ff6524710d167d618
https://qiita.com/yukoba/items/4733e8602fa4acabcc35