grimoire / amirstan_plugin

Useful tensorrt plugin. For pytorch and mmdetection model conversion.

Amirstan_plugin

Amirstan plugin contains some useful TensorRT plugins. These plugins are used to support other projects:

• torch2trt_dynamic

• mmdetection-to-tensorrt

Requirement

• TensorRT >= 8.0

Installation

• Install TensorRT: NVIDIA TensorRT

From sources

• clone the repo and create build folder

  git clone --depth=1 https://github.com/grimoire/amirstan_plugin.git
  cd amirstan_plugin
  git submodule update --init --progress --depth=1
  mkdir build
  cd build

• build the project with:

  • plugins only

    cmake -DTENSORRT_DIR=${path_to_tensorrt} ..
    make -j$(nproc)

  • or with DeepStream support

    cmake -DTENSORRT_DIR=${path_to_tensorrt} -DWITH_DEEPSTREAM=true -DDeepStream_DIR=${path_to_deepstream} ..
    make -j$(nproc)

• set the envoirment variable(in ~/.bashrc):

export AMIRSTAN_LIBRARY_PATH=<amirstan_plugin_root>/build/lib

Using Conan

• Install Conan:

  pip install conan

• Register grimoire's Conan remote:

  conan remote add grimoire https://grimoire.jfrog.io/artifactory/api/conan/grimoire-conan

• Add a conanfile.txt file to your project's root with the following content:

  [requires]
  amirstan_plugin/0.5.0
  
  [generators]
  cmake

• Additionaly, you can add a few options under the [options] section to configure your build:

  • tensorrt_dir: path where TensorRT is located. Default ~/SDK/TensorRT.
  • with_deepstream: whether to compile with deepstream support. Default False.
  • deepstream_dir: path where deepstream is located. Default /opt/nvidia/deepstream/deepstream
  • cub_root_dir: Default ./third_party/cub
  • cuda_arch: list of CUDA architectures to compile for. Default 61;62;70;72;75;80;86

  For example, to use a custom TensorRT dir and compile for a specific CUDA architecture:

  [requires]
  amirstan_plugin/0.5.0
  
  [generators]
  cmake
  
  [options]
  amirstan_plugin:tensorrt_dir=/usr/include/x86_64-linux-gnu
  amirstan_plugin:cuda_arch=75

• Add the following lines to your project root's CMakeLists.txt:

  INCLUDE(${CMAKE_BINARY_DIR}/conanbuildinfo.cmake)
  CONAN_BASIC_SETUP()

• Add conan libs to the linking stage:

  target_link_libraries(trt_sample PUBLIC ${CONAN_LIBS} ${CUDA_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT} ${TensorRT_LIBRARIES})

• Compile your project:

  mkdir build
  cd build
  conan install .. -s compiler.libcxx=libstdc++11 --build=missing 
  cmake .. 
  make -j$(nproc)