PyTorch Setup (C++17, zapcc, QtCreator, Debian, user-space)

This post documents my PyTorch C++ setup. It is intended as a brief how-to.

Goals

• Works with C++17 code (no pre-C++11 ABI)
• Works with the zapcc compiler (personal favorite)
• Works with QtCreator (currently my favored IDE on linux)
• Works with Debian without sudo rights (work constraint)
• Works with CUDA (only realistic way to train larger networks)

Steps

(All versions were current at the time of writing)

Installing PyTorch C++

• go to https://pytorch.org/
• scroll down to configure download
• select:
  • Build: Stable (1.3)
  • OS: Linux
  • Package: LibTorch
  • Language: C++
  • CUDA: 10.1
• download with cxx11 ABI (Important!)

Example link: https://download.pytorch.org/libtorch/cu101/libtorch-cxx11-abi-shared-with-deps-1.3.0.zip

Installing Cuda Toolkit 10.1

• go to https://developer.nvidia.com/cuda-downloads

• select Linux -> x86_64 -> Ubuntu -> 18.04 -> runfile (local)

  (should give you wget <link>)

• instead of executing, extract it via xyz.run --tar mxvf
• run the cuda-installer in the extracted folder
  • do not install driver, samples, demo suite
  • go to Options -> Toolkit Options
    • change Toolkit Install Path to /local/something/cuda-10.1
    • do not create links
    • do not install manpage
  • go to Options -> Library install path
    • change to /local/something/cuda-10.1

Installing CuDNN

• go to https://developer.nvidia.com/cudnn
• download CuDNN
• extract file
• copy lib64 and include folders to /local/something/cuda-10.1/

CMake

Fix Caffe2Targets.cmake

Caffe2 contains hard-coded CUDA paths that are wrong in our installation. Search Caffe2Targets.cmake for lib64/libcudart.so and replace all absolute CUDA paths by your local installation.

PyTorch

I chose to explicitly provide a hint to the pytorch path.

# pytorch
find_package(Torch REQUIRED
    # path containing bin/lib/include/share
    PATHS "/path/to/libtorch/"
)

NOTE: the cmake variable TORCH_LIBRARY was set to a wrong value anyways. In QtCreator Projects -> Build -> CMake -> TORCH_LIBRARY change the value to /path/to/libtorch/lib/libtorch.so.

NOTE: the cmake variable CUDA_TOOLKIT_ROOT_DIR was missing in my case. In QtCreator Projects -> Build -> CMake -> CUDA_TOOLKIT_ROOT_DIR change the value to /path/to/cuda-10.1/.

OpenMP

When using zapcc, openmp is not found by default. We fix this by providing a clang-7 openmp (which zapcc is based on).

# OpenMP
find_package(OpenMP)
if (OPENMP_CXX_FOUND)
    set(OPENMP_TARGET OpenMP::OpenMP_CXX)
elseif(NOT MSVC)
    add_library(openmp INTERFACE)
    target_link_libraries(openmp INTERFACE /usr/lib/llvm-7/lib/libomp.so)
    target_include_directories(openmp INTERFACE /usr/lib/llvm-7/include/openmp)
    set(OPENMP_TARGET openmp)
else()
    set(OPENMP_TARGET "")
endif()

Our Project

target_link_libraries(OurProjectName PUBLIC
    # ... other dependencies
    ${TORCH_LIBRARIES}
    ${OPENMP_TARGET}
)

if (NOT MSVC)
    target_compile_options(OurProjectName PUBLIC
        # ... other options
        -fopenmp
    )
endif()

Running The Program

If the LD_LIBRARY_PATH is not set to /path/to/cuda-10.1/lib64, some libraries will not be found when running the program.

In the QtCreator, this can be added in Projects -> Run -> Run Environment -> Add.

The example code at https://pytorch.org/cppdocs/frontend.html should work now.