Learning Programming Massively Parallel Processors

1. Environment

1.1. Method 1: Use Docker Image

The simplest way is to use my docker image jamesnulliu/deeplearning:latest which contains all the softwares you need to build the project:

docker pull jamesnulliu/deeplearning:latest

Check my blog: Docker Container with Nvidia GPU Support if you need any help.

1.2. Method 2: Setup Environment Manually

Or if you are planing to setup your own environment, here are some tips:

You should install all the softwares with corresponding versions listed bellow:

• Miniconda/Anaconda
• gcc >= 12.0, nvcc >= 12.0
• CMake >= 3.30
• Ninja
• vcpkg, pkg-config
• [managed by conda] python >= 3.10, pytorch >= 2.0
• [managed by vcpkg] cxxopts, fmt, spdlog, proxy, gtest, yaml-cpp

🎯Miniconda

Managing python environments with miniconda is always a good choice. Check the official website for an installation guide.

After installation, if you do not intend to install all the packages in base environment, create a new conda environment named PMPP (or whatever you like) and activate it:

# python version should be larger than 3.10
conda create -n PMPP python=3.12
conda activate PMPP  # Activate this environment
# In my experience, when your system gcc version is larger than 12, it is
# highly possible that you have to update libstd++ in conda for running the
# later compiled targets. All you need to do is to run this command:
conda upgrade libstdcxx-ng -c conda-forge

🎯PyTorch

Install pytorch with pip (not conda) in environment PMPP following the steps on the official website. In my case I installed torch-2.6.0 + cuda 12.6.

📝NOTE
All the python packages you installed can be found under the directory of $CONDA_PREFIX/lib/python3.12/site-packages.

🎯CUDA

To compile cuda code, you need to install cuda toolkit on your system. Usually, even if torch-2.6.0 + cuda 12.6 is installed in your conda environment while cuda 12.1 is installed on the system, you can run torch in python without any mistakes. But in some cases, you still have to install cuda 12.6 to exactly match the torch you chose.

You can find all versions of cuda on the official website.

📝NOTE
Installing and using multiple versions of cuda is possible by managing the PATH and LD_LIBRARY_PATH environment variables on linux, and you can do this manually or refering to my methods in this blog.

2. Quick Start

To build the C++ part only:

bash scripts/build.sh

📝NOTE
See "cmake-parameters.md" for details about setting up the build process.

You will find "./build/lib/libPmppTorchOps.so" which is the operator library and "./build/test/pmpp_test" which is the test executable (with gtest).

Execute the test executable to test the library manually:

./build/test/pmpp_test

Note that the test is already integrated into CMake build system (with ctest); In "scripts/build.sh", the last line shows how to run the test:

# $BUILD_DIR is "./build" by default
# Set `GTEST_COLOR` to yes or no to enable or disable colored output

# If the library has not been build, target `all` before `check` is required
cmake --build $BUILD_DIR -j $(nproc) --target all check
# Or if the library has been build, `check` is enough
cmake --build $BUILD_DIR -j $(nproc) --target check

To build and install the python package pmpp in current activated conda environment (pmpp operator library would be built automatically if it has not been built yet):

pip3 install --no-build-isolation -v .

torch.ops.pmpp.vector_add will be available after installation; See test.py for example.