L4t-base with cuDNN

[naveen.crasta]

Hello,
I would like to install OpenCV 4.5.3 with cuDNN on l4t base docker image. I found that l4t does not have cuDNN pre-installed. How can I install cuDNN into l4t base image?

Best regards,
Naveen

[dusty_nv]

Hi @naveen.crasta, cuDNN is automatically mounted into l4t-base image when you run it with --runtime nvidia. If you want to use OpenCV build with cuDNN enabled, then I recommend using one of the recent l4t-ml containers which have that already installed. For example, if you update your TX2’s JetPack to JetPack 4.6.1, you can use nvcr.io/nvidia/l4t-ml:r32.7.1-py3

[naveen.crasta]

Hi @dusty_nv,
Following your recommendation, I tried to build OpenCV with cuDNN on nvcr.io/nvidia/l4t-ml:r32.7.1-py3 image with the following CUDA flags (see the bottom of the flag list below).

RUN cmake \
     -D CPACK_BINARY_DEB=ON \
     -D CMAKE_BUILD_TYPE=RELEASE \
     -D CMAKE_INSTALL_PREFIX=$CMAKE_INSTALL \
     -D OPENCV_ENABLE_NONFREE=ON \
     -D ENABLE_CCACHE=ON \
     -D BUILD_JPEG=OFF \
     -D BUILD_JASPER=OFF \
     -D BUILD_JAVA=OFF \
     -D BUILD_opencv_python2=OFF \
     -D BUILD_opencv_python3=ON \
     -D INSTALL_PYTHON_EXAMPLES=OFF \
     -D INSTALL_C_EXAMPLES=OFF \
     -D OPENCV_EXTRA_MODULES_PATH='/OpenCV/opencv_contrib/modules' \
     -D PYTHON_DEFAULT_EXECUTABLE=/usr/bin/python3 \
     -D PYTHON3_EXECUTABLE=/usr/bin/python3.8 \
     -D PYTHON3_INCLUDE_DIR=/usr/include/python3.8 \
     -D PYTHON3_LIBRARY=/usr/lib/$(uname -i)-linux-gnu/libpython3.8.so \
     -D PYTHON3_PACKAGES_PATH=/usr/local/lib/python3.8/dist-packages \
     -D BUILD_EXAMPLES=OFF \
     -D WITH_VTK=OFF \
     -D ENABLE_FAST_MATH=ON \
     -D WITH_LIBV4L=ON \
     -D WITH_GSTREAMER=OFF \
     -D WITH_GSTREAMER_0_10=OFF \
     -D WITH_TBB=ON \
     -D WITH_CUDA=ON \
     -D CUDA_ARCH_BIN=5.3,6.2,7.2,7.5 \
     -D CUDA_ARCH_PTX="" \
     -D CUDA_FAST_MATH=ON \
     -D WITH_CUBLAS=ON \
     -D CUDA_TOOLKIT_ROOT_DIR=/usr/local/cuda-10.2 \
     -D CUDA_BIN_PATH=/usr/local/cuda-10.2 \
     -D CUDNN_VERSION='10.2' \
     -D WITH_CUDNN=ON \
     -D OPENCV_DNN_CUDA=ON \
     -D CUDNN_INCLUDE_DIR=/usr/include \
     -D CUDNN_LIBRARY=/usr/include \
     ../ && \
     make -j$(nproc) && make install && make package

However, I ended up with a long list of such errors (I have just pasted the first error)!

– Configuring done
CMake Error: The following variables are used in this project, but they are set to NOTFOUND.
Please set them or make sure they are set and tested correctly in the CMake files:
CUDA_cublas_LIBRARY (ADVANCED)
linked by target “opencv_cudev” in directory /OpenCV/opencv_contrib/modules/cudev
linked by target “opencv_test_cudev” in directory /OpenCV/opencv_contrib/modules/cudev/test
linked by target “opencv_test_core” in directory /OpenCV/opencv/modules/core
linked by target “opencv_perf_core” in directory /OpenCV/opencv/modules/core
linked by target “opencv_core” in directory /OpenCV/opencv/modules/core
linked by target “opencv_test_cudaarithm” in directory /OpenCV/opencv_contrib/modules/cudaarithm

What could be the reason for this error? Is it not possible to cross compile?

Correction: I have Xavier board but I am trying to cross-compile on x86-64 machine using QEMU with the following command:
docker build -t opencv-qemu --platform linux/arm64 .

Thanks,
Naveen

[dusty_nv]

@naveen.crasta

What could be the reason for this error? Is it not possible to cross compile?

Correction: I have Xavier board but I am trying to cross-compile on x86-64 machine using QEMU with the following command:
docker build -t opencv-qemu --platform linux/arm64 .

On JetPack 4.x, in order to use CUDA/cuDNN libraries during docker build operations, you need to set the default docker runtime to nvidia: https://github.com/dusty-nv/jetson-containers#docker-default-runtime

Then these CUDA/cuDNN libraries will get mounted during the docker build. However, this won’t work on emulators, because the libraries aren’t there for them to mount.