Tvm

1 Installation

1.1 Build the Shared Library

1.1.1 Requirements

• A recent c++ compiler supporting C++14 (g++5 or higher)

  # We should install gcc-7 and g++-7 because we use cudatoolkit-10.
  sudo apt install gcc-7 g++-7
    
  # Set current gcc and g++ pointing to gcc-7 and g++-7
  # sudo update-alternatives --install <link> <name> <path> <priority>
  sudo update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-7 7
  sudo update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-7 7
    
  # sudo update-alternatives --config <name>
  # Choose the proper priority.
  sudo update-alternatives --config gcc
  sudo update-alternatives --config g++
  

• Cmake

  sudo apt install cmake
  

• LLVM 4.0 or higher

  sudo apt install llvm-10-dev
  

• Cuda

1.1.2 Build

1. Create a build directory where tvm is installed and copy the cmake/config.cmake to the directory:

   # In the directory where tvm is installed
   mkdir build
   cp cmake/config.cmake build
   

2. Edit config.cmake to customize the compilation options:

   • Change set(USE_CUDA OFF) to set(USE_CUDA ON) to enable CUDA.
   • Set set(USE_GRAPH_RUNTIME ON) and set(USE_GRAPH_RUNTIME_DEBUG ON) to help with debugging.
   • Set set(USE_LLVM ON).

3. Build the shared library.

   # In the directory where tvm is installed
   cd build
   cmake ..
   make -j4
   

1.2 Python Package Installation

# Add follow commands to ~/.zshrc.
export TVM_HOME=/path/to/tvm
export PYTHONPATH=$TVM_HOME/python:$PYTHONPATH

1.3 Dependences

• Use RPC tracker:

  conda install tornado
  

• Test c++ with GTest:

  git clone git@github.com:google/googletest.git
  cd googletest
  mkdir build
  cd build
  cmake ..
  make
  sudo make install