Additional Installation to Use the GPU


  • gfortran
  • cmake
  • boost-python
  • CUDA >= 6.0
  • PyCUDA >= 2013.1.1
  • CUDA SciKit >= 0.5.0
  • Mako
  • CULA >= R12 (optional)


  • CUDA:

    • Get the CUDA installers from the CUDA download site and install it.

      sudo dpkg -i cuda-repo-ubuntu1204_6.5-14_amd64.deb
      sudo apt-get update
    • Then you can install the CUDA Toolkit using apt-get.

      sudo apt-get install cuda
    • You should reboot the system afterwards and verify the driver installation with the nvidia-settings utility.

    • Set the environment variable CUDA_HOME to point to the CUDA home directory. Also, add the CUDA binary and library directory to your PATH and LD_LIBRARY_PATH.

      export CUDA_HOME=/usr/local/cuda
      export PATH=${CUDA_HOME}/bin:${PATH}
  • Install PyCUDA with pip. Make sure that PATH is defined as root.

    sudo PATH=$PATH pip install pycuda
  • Install CUDA SciKit with pip.

    sudo pip install pycuda scikits.cuda>=0.5.0a1 Mako
  • CULA (optional):

    • Linear systems can optionally be solved on the GPU using the CULA Dense toolkit.

    • Download and install the full edition of CULA. The full edition is required since the free edition only has single precision functions. The full edition is free for academic use, but requires registration.

    • As recommended by the installation, set the environment variables CULA_ROOT and CULA_INC_PATH to point to the CULA root and include directories. Also, add the CULA library directory to your LD_LIBRARY_PATH.

      export CULA_ROOT=/usr/local/cula
      export CULA_INC_PATH=$CULA_ROOT/include
  • Build the lfd sources with cmake as you would normally do.

    mkdir build_lfd
    cd build_lfd
    cmake /path/to/lfd
    make -j

To use the compiled libraries from python, add the following path to your PYTHONPATH:


For more information, check out the README from the tpsopt module.