Giters

jasaw / tegra210-cuda

CUDA libraries and header files for Tegra210 platform

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

tegra210-cuda

CUDA libraries and header files for Tegra210 platform

nVidia does not provide a convinient way to download and install aarch64 CUDA libraries for the purpose of cross compiling CUDA enabled applications and libraries. This project packages the libraries into a more cross-compile friendly format.

JetPack 4.3 contents:

  • CUDA version 10.0
  • libcudnn version 8.0.0.145

JetPack 4.4 contents:

  • CUDA version 10.2
  • libcudnn version 8.0.0.145

JetPack 4.6 Rev 2 contents:

  • CUDA version 10.2
  • libcudnn version 8.2.1.32

How to extract CUDA libraries from JetPack

These were the steps taken to extract and repackage the CUDA libraries.

Download JetPack 4.6 Rev 2 libraries.

We have 3 options here:

  1. Download from the direct links below:

  2. Download SDK Manager from https://developer.nvidia.com/nvidia-sdk-manager

    • Install the SDK Manager. It is usually installed here: /opt/nvidia/sdkmanager/sdkmanager-gui

  3. Use SDK Manager to generate sdkml3_jetpack_l4t_46_ga_r2.json file, then download straight from the URLs in the file.

Install the downloaded packages

  1. Manually install the downloaded packages. The downloaded packages are usually found in ~/Downloads/nvidia/sdkm_downloads directory.
  2. On x86_64 host, install cuda-repo-cross-aarch64-ubuntu1804-10-2-local_10.2.460-1_all.deb and cuda-repo-ubuntu1804-10-2-local_10.2.460-450.115-1_amd64.deb
    • sudo dpkg --force-all -i cuda-repo-cross-aarch64-ubuntu1804-10-2-local_10.2.460-1_all.deb
    • sudo dpkg --force-all -i cuda-repo-ubuntu1804-10-2-local_10.2.460-450.115-1_amd64.deb
    • sudo apt-key add /var/cuda-repo-10-2-local-10.2*/*.pub'
    • sudo apt-get -y update
    • sudo apt install -y gnupg libgomp1 libfreeimage-dev libopenmpi-dev openmpi-bin
    • sudo apt-get -y --allow-downgrades install cuda-toolkit-10-2 cuda-cross-aarch64-10-2
    • Development files will be installed in /usr/local/cuda-10.2 directory.
  3. On a Jetson Nano running Ubuntu, install cuda-repo-l4t-10-2-local_10.2.460-1_arm64.deb
    • sudo dpkg --force-all -i cuda-repo-l4t-10-2-local_10.2.460-1_arm64.deb
    • sudo apt-key add /var/cuda-repo-10-2-local-10.2*/*.pub'
    • sudo apt-get -y update
    • sudo apt install -y gnupg libgomp1 libfreeimage-dev libopenmpi-dev openmpi-bin
    • sudo apt-get -y --allow-downgrades install cuda-toolkit-10-2 cuda-cross-aarch64-10-2
    • Development files will be installed in /usr/local/cuda-10.2 directory.
  4. On a Jetson Nano running Ubuntu, extract libcudnn8_8.2.1.32-1+cuda10.2_arm64.deb and libcudnn8-dev_8.2.1.32-1+cuda10.2_arm64.deb
    • mkdir -p ~/libcudnn-8.2.1.32/dev
    • mkdir -p ~/libcudnn-8.2.1.32/rt
    • mkdir -p /tmp/libcudnn-8.2.1.32/dev
    • mkdir -p /tmp/libcudnn-8.2.1.32/rt
    • cd /tmp/libcudnn-8.2.1.32/rt
    • ar x libcudnn8_8.2.1.32-1+cuda10.2_arm64.deb && tar xf data.tar.xz
    • mv usr ~/libcudnn-8.2.1.32/rt/
    • cd /tmp/libcudnn-8.2.1.32/dev
    • ar x libcudnn8-dev_8.2.1.32-1+cuda10.2_arm64.deb && tar xf data.tar.xz
    • mv usr ~/libcudnn-8.2.1.32/dev/
  5. Remove docs and samples from /usr/local/cuda-10.2 and ~/libcudnn-8.2.1.32/dev/.

Compress the libraries

  1. On a Jetson Nano running Ubuntu, compress CUDA core libraries
    • mkdir -p ~/cuda-10.2-tegra194/usr/local
    • cp -arf /usr/local/cuda-10.2 ~/cuda-10.2-tegra194/usr/local
    • cd ~
    • XZ_OPT=-9 tar cfJ cuda-10.2-tegra194.tar.xz cuda-10.2-tegra194
  2. On a Jetson Nano running Ubuntu, compress CUDA DNN library
    • cd ~
    • XZ_OPT=-9 tar cfJ libcudnn-8.2.1.32-tegra194.tar.xz libcudnn-8.2.1.32

Alternatively, run the repack script

  1. Place all deb files in the dl directory.
  2. Run ./repack_deb_lib.sh.

Files too big, split and reassemble

To split, run: split -b 1800M -a 1 --numeric-suffixes=0 cuda-10.2-ubuntu1804.tar.xz cuda-10.2-ubuntu1804.tar.xz.part

To reassemble, run: cat cuda-10.2-ubuntu1804.tar.xz.part? > cuda-10.2-ubuntu1804.tar.xz

About

CUDA libraries and header files for Tegra210 platform

Languages

Language:Shell 100.0%