Stanford University logo SLAC National Accelerator Laboratory logo Los Alamos National Laboratory logo NVIDIA logo Winner of the R&D 100 Award

Legion

A Data-Centric Parallel Programming System

Github

Pygion

Overview

Pygion is task-based framework for Python based on the Legion programming system. As with Legion, Pygion provides sequential semantics. Tasks may execute asynchronously and in parallel, but only when they preserve the behavior of sequential execution. Pygion analyzes tasks (via Legion) to determine what tasks are safe to execute in parallel and when data movement or synchronization is required for safe execution.

For an overview of Pygion, the original PAW-ATM paper.

from pygion import task

@task
def main():
    print("Hello, Pygion!")

Getting Started

Pygion lives in the main Legion repo under bindings/python. To get started with Pygion, build with Make in that directory:

git clone https://github.com/StanfordLegion/legion.git
cd legion/bindings/python
make LG_RT_DIR=$PWD/../../runtime -j4

Customize the -j flag parameter to control the number of threads used to build, depending on your machine.

Building with CMake

If you prefer to use CMake, then build Legion at the top level -DLegion_BUILD_BINDINGS=ON -DLegion_USE_Python=ON:

git clone https://github.com/StanfordLegion/legion.git
cd legion
mkdir build
cd build
cmake .. -DLegion_BUILD_BINDINGS=ON -DLegion_USE_Python=ON -DCMAKE_INSTALL_PREFIX=$PWD/../install
make install -j4

Examples

A wide variety of examples, with detailed comments, can be found in the examples directory. Most of the important features of Pygion are demonstrated in these examples.

A dedicated tutorial series for Pygion is coming soon.