TensorRT Open Source Software

This repository contains the Open Source Software (OSS) components of NVIDIA TensorRT. Included are the sources for TensorRT plugins and parsers (Caffe and ONNX), as well as sample applications demonstrating usage and capabilities of the TensorRT platform.

Prerequisites

To build the TensorRT OSS components, ensure you meet the following package requirements:

System Packages

• CUDA

  • Recommended versions:
  • cuda-11.0 + cuDNN-8.0
  • cuda-10.2 + cuDNN-8.0

• GNU Make >= v4.1

• CMake >= v3.13

• Python >= v3.6.5

• PIP >= v19.0

  • PyPI packages
  • numpy
  • onnx 1.6.0
  • onnxruntime >= 1.3.0
  • pytest

• Essential libraries and utilities

  • Git, pkg-config, Wget, Zlib

• Cross compilation for Jetson platforms requires JetPack's host component installation

  • JetPack >= 4.4

• Windows requires Visual Studio 2017 either Community or Enterprise Version

  • Visual Studio

• Cross compilation for QNX requires the qnx developer toolchain

  • QNX

Optional Packages

• Containerized builds

  • Docker >= 19.03
  • NVIDIA Docker >= 2.0 or nvidia-container-toolkit

• Code formatting tools

  • Clang-format
  • Git-clang-format

• Required PyPI packages for Demos

  • Tensorflow-gpu == 1.15.0

TensorRT Release

• TensorRT v7.1

NOTE: Along with the TensorRT OSS components, the following source packages will also be downloaded, and they are not required to be installed on the system.

• ONNX-TensorRT v7.1
• CUB v1.8.0
• Protobuf v3.0.0

Downloading The TensorRT Components

1. Download TensorRT OSS sources.

   Example: Bash

   git clone -b master https://github.com/nvidia/TensorRT TensorRT
   cd TensorRT
   git submodule update --init --recursive
   export TRT_SOURCE=`pwd`

   Example: Powershell

   git clone -b master https://github.com/nvidia/TensorRT TensorRT
   cd TensorRT
   git submodule update --init --recursive
   $Env:TRT_SOURCE = $(Get-Location)

2. Download the TensorRT binary release.

   To build the TensorRT OSS, obtain the corresponding TensorRT 7.1 binary release from NVidia Developer Zone. For a list of key features, known and fixed issues, refer to the TensorRT 7.1 Release Notes.

   Example: Ubuntu 18.04 with cuda-11.0

   Download and extract the latest TensorRT 7.1 GA package for Ubuntu 18.04 and CUDA 11.0

   cd ~/Downloads
   tar -xvzf TensorRT-7.1.3.4.Ubuntu-18.04.x86_64-gnu.cuda-11.0.cudnn8.0.tar.gz
   export TRT_RELEASE=`pwd`/TensorRT-7.1.3.4

   Example: Ubuntu 18.04 with cuda-11.0 on PowerPC

   Download and extract the latest TensorRT 7.1 GA package for Ubuntu 18.04 and CUDA 11.0

   cd ~/Downloads
   # Download TensorRT-7.1.3.4.Ubuntu-18.04.powerpc64le-gnu.cuda-11.0.cudnn8.0.tar.gz
   tar -xvzf TensorRT-7.1.3.4.Ubuntu-18.04.powerpc64le-gnu.cuda-11.0.cudnn8.0.tar.gz
   export TRT_RELEASE=`pwd`/TensorRT-7.1.3.4

   Example: CentOS/RedHat 7 with cuda-11.0

   Download and extract the TensorRT 7.1 GA for CentOS/RedHat 7 and CUDA 11.0 tar package

   cd ~/Downloads
   tar -xvzf TensorRT-7.1.3.4.CentOS-8.0.x86_64-gnu.cuda-11.0.cudnn8.0.tar.gz
   export TRT_RELEASE=`pwd`/TensorRT-7.1.3.4

   Example: Ubuntu 16.04 with cuda-11.0

   Download and extract the TensorRT 7.1 GA for Ubuntu 16.04 and CUDA 11.0 tar package

   cd ~/Downloads
   tar -xvzf TensorRT-7.1.3.4.Ubuntu-16.04.x86_64-gnu.cuda-11.0.cudnn8.0.tar.gz
   export TRT_RELEASE=`pwd`/TensorRT-7.1.3.4

   Example: Ubuntu18.04 cross compile QNX with cuda-10.2

   Download and extract the TensorRT 7.1 GA for QNX and CUDA 10.2 tar package

   cd ~/Downloads
   tar -xvzf TensorRT-7.1.3.4.Ubuntu-18.04.aarch64-qnx.cuda-10.2.cudnn7.6.tar.gz
   export TRT_RELEASE=`pwd`/TensorRT-7.1.3.4
   export QNX_HOST=/path/to/qnx/toolchain/host/linux/x86_64
   export QNX_TARGET=/path/to/qnx/toolchain/target/qnx7

   Example: Windows with cuda-11.0

   Download and extract the TensorRT 7.1 GA for Windows and CUDA 11.0 zip package and add msbuild to PATH

   cd ~\Downloads
   Expand-Archive .\TensorRT-7.1.3.4.Windows10.x86_64.cuda-11.0.cudnn8.0.zip
   $Env:TRT_RELEASE = '$(Get-Location)\TensorRT-7.1.3.4'
   $Env:PATH += 'C:\Program Files (x86)\Microsoft Visual Studio\2017\Professional\MSBuild\15.0\Bin\'

3. Download JetPack toolchain for cross-compilation.[OPTIONAL]

   JetPack example

   Using the SDK manager, download the host components of the PDK version or Jetpack specified in the name of the Dockerfile. To do this:

   1. [SDK Manager Step 01] Log into the SDK manager
   2. [SDK Manager Step 02] Select the correct platform and Target OS System (should be corresponding to the name of the Dockerfile you are building (e.g. Jetson AGX Xavier, Linux Jetpack 4.4), then click Continue
   3. [SDK Manager Step 03] Under Download & Install Options make note of or change the download folder and Select Download now. Install later. then agree to the license terms and click Continue You should now have all expected files to build the container. Move these into the docker/jetpack_files folder.

Setting Up The Build Environment

• Install the System Packages list of components in the Prerequisites section.

• Alternatively, use the build containers as described below:

1. Generate the TensorRT build container.

   The docker container can be built using the included Dockerfiles and build script. The build container is configured with the environment and packages required for building TensorRT OSS.

   Example: Ubuntu 18.04 with cuda-11.0

   ./docker/build.sh --file docker/ubuntu.Dockerfile --tag tensorrt-ubuntu --os 18.04 --cuda 11.0

   Example: Ubuntu 16.04 with cuda-11.0

   ./docker/build.sh --file docker/ubuntu.Dockerfile --tag tensorrt-ubuntu1604 --os 16.04 --cuda 11.0

   Example: CentOS/RedHat 7 with cuda-11.0

   ./docker/build.sh --file docker/centos.Dockerfile --tag tensorrt-centos --os 7 --cuda 11.0

   Example: Cross compile for JetPack 4.4 with cuda-10.2

   ./docker/build.sh --file docker/ubuntu-cross-aarch64.Dockerfile --tag tensorrt-ubuntu-jetpack --os 18.04 --cuda 10.2

   Example: Cross compile for PowerPC with cuda-11.0

   ./docker/build.sh --file docker/ubuntu-cross-ppc64le.Dockerfile --tag tensorrt-ubuntu-ppc --os 18.04 --cuda 11.0

2. Launch the TensorRT build container.

   ./docker/launch.sh --tag tensorrt-ubuntu --gpus all --release $TRT_RELEASE --source $TRT_SOURCE

   NOTE: To run TensorRT/CUDA programs in the build container, install NVIDIA Docker support. Docker versions < 19.03 require nvidia-docker2 and --runtime=nvidia flag for docker run commands. On versions >= 19.03, you need the nvidia-container-toolkit package and --gpus all flag.

Building The TensorRT OSS Components

• Generate Makefiles and build.

  Example: Linux

   cd $TRT_SOURCE
   mkdir -p build && cd build
   cmake .. -DTRT_LIB_DIR=$TRT_RELEASE/lib -DTRT_OUT_DIR=`pwd`/out
   make -j$(nproc)

  Example: Bare-metal build on Jetson (ARM64) with cuda-10.2

  cd $TRT_SOURCE
  mkdir -p build && cd build
  cmake .. -DTRT_LIB_DIR=$TRT_RELEASE/lib -DTRT_OUT_DIR=`pwd`/out -DTRT_PLATFORM_ID=aarch64 -DCUDA_VERSION=10.2
  make -j$(nproc)

  Example: Cross compile for QNX with cuda-10.2

   cd $TRT_SOURCE
   mkdir -p build && cd build
   cmake .. -DTRT_LIB_DIR=$TRT_RELEASE/lib -DTRT_OUT_DIR=`pwd`/out -DCMAKE_TOOLCHAIN_FILE=$TRT_SOURCE/cmake/toolchains/cmake_qnx.toolchain
   make -j$(nproc)

  Example: Powershell

   cd $Env:TRT_SOURCE
   mkdir -p build ; cd build
   cmake .. -DTRT_LIB_DIR=$Env:TRT_RELEASE\lib -DTRT_OUT_DIR='$(Get-Location)\out' -DCMAKE_TOOLCHAIN_FILE=..\cmake\toolchains\cmake_x64_win.toolchain
   msbuild ALL_BUILD.vcxproj

  NOTE:

  1. The default CUDA version used by CMake is 11.0. To override this, for example to 10.2, append -DCUDA_VERSION=10.2 to the cmake command.
  2. Samples may fail to link on CentOS7. To work around this create the following symbolic link: ln -s $TRT_OUT_DIR/libnvinfer_plugin.so $TRT_OUT_DIR/libnvinfer_plugin.so.7

  The required CMake arguments are:

  • TRT_LIB_DIR: Path to the TensorRT installation directory containing libraries.

  • TRT_OUT_DIR: Output directory where generated build artifacts will be copied.

  The following CMake build parameters are optional:

  • CMAKE_BUILD_TYPE: Specify if binaries generated are for release or debug (contain debug symbols). Values consists of [Release] | Debug

  • CUDA_VERISON: The version of CUDA to target, for example [11.0].

  • CUDNN_VERSION: The version of cuDNN to target, for example [8.0].

  • NVCR_SUFFIX: Optional nvcr/cuda image suffix. Set to "-rc" for CUDA11 RC builds until general availability. Blank by default.

  • PROTOBUF_VERSION: The version of Protobuf to use, for example [3.0.0]. Note: Changing this will not configure CMake to use a system version of Protobuf, it will configure CMake to download and try building that version.

  • CMAKE_TOOLCHAIN_FILE: The path to a toolchain file for cross compilation.

  • BUILD_PARSERS: Specify if the parsers should be built, for example [ON] | OFF. If turned OFF, CMake will try to find precompiled versions of the parser libraries to use in compiling samples. First in ${TRT_LIB_DIR}, then on the system. If the build type is Debug, then it will prefer debug builds of the libraries before release versions if available.

  • BUILD_PLUGINS: Specify if the plugins should be built, for example [ON] | OFF. If turned OFF, CMake will try to find a precompiled version of the plugin library to use in compiling samples. First in ${TRT_LIB_DIR}, then on the system. If the build type is Debug, then it will prefer debug builds of the libraries before release versions if available.

  • BUILD_SAMPLES: Specify if the samples should be built, for example [ON] | OFF.

  Other build options with limited applicability:

  • CUB_VERSION: The version of CUB to use, for example [1.8.0].

  • GPU_ARCHS: GPU (SM) architectures to target. By default we generate CUDA code for all major SMs. Specific SM versions can be specified here as a quoted space-separated list to reduce compilation time and binary size. Table of compute capabilities of NVIDIA GPUs can be found here. Examples:

    • NVidia A100: -DGPU_ARCHS="80"
    • Tesla T4, GeForce RTX 2080: -DGPU_ARCHS="75"
    • Titan V, Tesla V100: -DGPU_ARCHS="70"
    • Multiple SMs: -DGPU_ARCHS="80 75"

  • TRT_PLATFORM_ID: Bare-metal build (unlike containerized cross-compilation) on non Linux/x86 platforms must explicitly specify the target platform. Currently supported options: x86_64 (default), aarch64

(Optional - recommended) installing the tensorrt python API bindings

whl files for the TensorRT python API are in the python directory of the TensorRT release

Example install for python3:

pip3 install $TRT_RELEASE/python/tensorrt-7.1.3.4-cp36-none-linux_x86_64.whl

Useful Resources

TensorRT

• TensorRT Homepage
• TensorRT Developer Guide
• TensorRT Sample Support Guide
• TensorRT Discussion Forums

Known Issues

TensorRT 7.1

• See Release Notes.