Many times, I find myself trying to install OpenCV with CUDA support and in every post, I find a different answer, so I figured why not try to add one from scratch for the people that like to see what might be going wrong under the hood instead of using precompiled packages.
Here are the specs and versions we will be using over this blog:
| Component | Version |
|---|---|
| OS | Ubuntu 24.04 |
| CUDA | 12.5 |
| OpenCV | 5.x |
| GCC | 13.2 |
Installing CUDA
As we want to have OpenCV with CUDA from scratch, what better way to start than installing CUDA?
To install CUDA, the instructions provided by Nvidia are too clear, which means that these are too detailed, and you might get lost (as I did). The important section is the following:
https://docs.nvidia.com/cuda/cuda-installation-guide-linux/#download-the-nvidia-cuda-toolkit
In there you can reach a point in which you can select which OS you have, the way you want to install it, and more. The issue that was presented to me was that no Ubuntu24 package is there for CUDA 12.5, I just solved it by using the local runfile (.run file), which installes CUDA toolkit manually. Just be aware, that using this runfile will ask you if you also want to install an Nvidia Driver, which you may want if you haven’t already.
export CUDA_HOME=/usr/local/cuda export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64 export PATH=$PATH:$CUDA_HOME/bin
After this, you can reload a new terminal or source ~/.bashrc (however you want it), then check that CUDA is installed:
$ nvcc --version nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2024 NVIDIA Corporation Built on Thu_Jun__6_02:18:23_PDT_2024 Cuda compilation tools, release 12.5, V12.5.82 Build cuda_12.5.r12.5/compiler.34385749_0
I strongly recommend to check if nvidia-smi is working too:
$ nvidia-smi
Fri Aug 23 14:18:19 2024
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 555.42.06 Driver Version: 555.42.06 CUDA Version: 12.5 |
|-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA GeForce RTX 3070 Ti Off | 00000000:05:00.0 On | N/A |
| 56% 50C P3 42W / 310W | 589MiB / 8192MiB | 7% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=========================================================================================|
| 0 N/A N/A 2718 G /usr/lib/xorg/Xorg 272MiB |
| 0 N/A N/A 3055 G /usr/bin/gnome-shell 61MiB |
| 0 N/A N/A 3712 G ...seed-version=20240821-050100.700000 54MiB |
| 0 N/A N/A 4579 G ...erProcess --variations-seed-version 176MiB |
+-----------------------------------------------------------------------------------------+
Installing OpenCV with CUDA support
Now, what I believe is the real deal here, installing OpenCV with CUDA support… composed of the following steps:
Install dependencies
The following are some core dependencies, not all are 100% needed, just depending on what you need OpenCV for, also others might need to be added.
sudo apt install \
build-essential \
cmake \
cmake-data \
git \
gcc-13 \
g++-13 \
libavcodec-dev \
libavformat-dev \
libgtk2.0-dev \
libjpeg-dev \
libpng-dev \
pkg-config \
python3-numpy \
python3-pip
Clone and set OpenCV repos
We want to build full OpenCV 5.x, even with the contrib modules.
git clone --branch 5.x --depth 1 --single-branch https://github.com/opencv/opencv.git cd opencv git clone --branch 5.x --depth 1 --single-branch https://github.com/opencv/opencv_contrib.git export SRC_DIR=$(pwd)
Build and install OpenCV
First thing is to configure the OpenCV to build:
mkdir build && cd build
cmake \
-D WITH_CUDA=ON \
-D WITH_CUDNN=ON \
-D OPENCV_DNN_CUDA=OFF \
-D CUDA_ARCH_BIN=$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader) \
-D CMAKE_C_COMPILER=gcc-13 \
-D CMAKE_CXX_COMPILER=g++-13 \
-D OPENCV_EXTRA_MODULES_PATH=${SRC_DIR}/opencv_contrib/modules/ \
-D PYTHON3_INCLUDE_DIR=$(python3 -c "from distutils.sysconfig import get_python_inc; print(get_python_inc())") \
-D PYTHON3_PACKAGES_PATH=$(python3 -c "from distutils.sysconfig import get_python_lib; print(get_python_lib())") \
-D PYTHON3_EXECUTABLE="/usr/bin/python3" \
${SRC_DIR}
In this case you can see that you don’t need to modify anything, all the paths for the python3 bindings and the compute capability of the GPU for CUDA is already obtained from a command. We use GCC13, which is supported by both this CUDA and OpenCV versions. But be aware! There is a catch, we didn’t want to use DNN, so we disabled it too, but you can just set it to ON after installing DNN package.
After configure, execute build in that same directory:
make
After that, you can install the library and the Python bindings by using a simple make install:
make install
You can test the installation with a basic command on Python:
python3 -c "import cv2"
If that worked, we have completed the installation of OpenCV with CUDA support, if you run into any trouble, please don’t hesitate to contact us at https://proventusnova.com/contact-us/
