Reinforcement learning environments for compiler research
Project description
CompilerGym is a toolkit for exposing compiler optimization problems for reinforcement learning. It allows machine learning researchers to experiment with program optimization techniques without requiring any experience in compilers, and provides a framework for compiler developers to expose new optimization problems for AI.
Table of Contents
Features
With CompilerGym, building ML models for compiler research problems is as easy as building ML models to play video games. Here are some highlights of key features:
-
API: uses the popular Gym interface from OpenAI — use Python to write your agent.
-
Datasets: wraps real world programs (C++ programs, TensorFlow programs, programs from Github, etc.) and a mainstream compiler (LLVM), providing millions of programs for training.
-
Tasks and Actions: interfaces the LLVM compiler for one compiler research problem: phase ordering (more to come). It has a large discrete action space.
-
Representations: provides raw representations of programs, as well as multiple kinds of pre-computed features: you can focus on end-to-end deep learning or features + boosted trees, all the way up to graph models.
-
Rewards: provides appropriate reward functions and loss functions out of the box.
-
Testing: provides a validation process for correctness of results.
-
Baselines: provides some baselines and reports their performance.
-
Competition: provides leaderboards for you to submit your results.
For a glimpse of what's to come, check out our roadmap.
Getting Started
Starting with CompilerGym is simple. If you not already familiar with the gym interface, refer to the getting started guide for an overview of the key concepts.
Installation
Install the latest CompilerGym release using:
pip install -U compiler_gym
The binary works on macOS and Linux (on Ubuntu 18.04, Fedora 28, Debian 10 or newer equivalents).
Building from Source
If you prefer, you may build from source. This requires a modern C++ toolchain and bazel.
macOS
On macOS the required dependencies can be installed using homebrew:
brew install bazelisk zlib
export LDFLAGS="-L/usr/local/opt/zlib/lib"
export CPPFLAGS="-I/usr/local/opt/zlib/include"
export PKG_CONFIG_PATH="/usr/local/opt/zlib/lib/pkgconfig"
Now proceed to All platforms below.
Linux
On debian-based linux systems, install the required toolchain using:
sudo apt install clang-9 libtinfo5 libjpeg-dev patchelf
wget https://github.com/bazelbuild/bazelisk/releases/download/v1.7.5/bazelisk-linux-amd64 -O bazel
chmod +x bazel && mkdir -p ~/.local/bin && mv -v bazel ~/.local/bin
export PATH="$HOME/.local/bin:$PATH"
export CC=clang
export CXX=clang++
All platforms
We recommend using conda to manage the remaining build dependencies. First create a conda environment with the required dependencies:
conda create -n compiler_gym python=3.9 cmake pandoc
conda activate compiler_gym
Then clone the CompilerGym source code using:
git clone https://github.com/facebookresearch/CompilerGym.git
cd CompilerGym
There are two primary git branches: stable
tracks the latest release;
development
is for bleeding edge features that may not yet be mature. Checkout
your preferred branch and install the python development dependencies using:
git checkout stable
make init
The make init
target only needs to be run once on initial setup, or when
pulling remote changes to the CompilerGym repository.
Run the test suite to confirm that everything is working:
make test
To build and install the compiler_gym
python package, run:
make install
NOTE: To use the compiler_gym
package that is installed by make install
you must leave the root directory of this repository. Attempting to import
compiler_gym
while in the root of this repository will cause import errors.
When you are finished, you can deactivate and delete the conda environment using:
conda deactivate
conda env remove -n compiler_gym
Trying it out
In Python, import compiler_gym
to use the environments:
>>> import gym
>>> import compiler_gym # imports the CompilerGym environments
>>> env = gym.make("llvm-autophase-ic-v0") # starts a new environment
>>> env.benchmark = "benchmark://cbench-v1/qsort" # select a program to compile
>>> env.reset() # starts a new compilation session
>>> env.render() # prints the IR of the program
>>> env.step(env.action_space.sample()) # applies a random optimization, updates state/reward/actions
See the documentation website for tutorials, further details, and API reference. Our roadmap of planned features is public, and the changelog summarizes shipped features.
Leaderboards
These leaderboards track the performance of user-submitted algorithms for CompilerGym tasks. To submit a result please see this document.
LLVM Instruction Count
LLVM is a popular open source compiler used widely in industry and research. The
llvm-ic-v0
environment exposes LLVM's optimizing passes as a set of actions
that can be applied to a particular program. The goal of the agent is to select
the sequence of optimizations that lead to the greatest reduction in instruction
count in the program being compiled. Reward is the reduction in instruction
count achieved scaled to the reduction achieved by LLVM's builtin -Oz
pipeline.
This leaderboard tracks the results achieved by algorithms on the llvm-ic-v0
environment on the 23 benchmarks in the cbench-v1
dataset.
Author | Algorithm | Links | Date | Walltime (mean) | Codesize Reduction (geomean) |
---|---|---|---|---|---|
Random search (t=10800) | write-up, results | 2021-03 | 10,512.356s | 1.062× | |
Random search (t=3600) | write-up, results | 2021-03 | 3,630.821s | 1.061× | |
Greedy search | write-up, results | 2021-03 | 169.237s | 1.055× | |
Random search (t=60) | write-up, results | 2021-03 | 91.215s | 1.045× | |
e-Greedy search (e=0.1) | write-up, results | 2021-03 | 152.579s | 1.041× | |
Jiadong Guo | Tabular Q (N=5000, H=10) | write-up, results | 2021-04 | 2534.305 | 1.036× |
Random search (t=10) | write-up, results | 2021-03 | 42.939s | 1.031× | |
Jiadong Guo | Tabular Q (N=2000, H=5) | write-up, results | 2021-04 | 694.105 | 0.988× |
Contributing
We welcome contributions to CompilerGym. If you are interested in contributing please see this document.
Citation
If you use CompilerGym in any of your work, please cite:
@Misc{CompilerGym,
author = {Cummins, Chris and Leather, Hugh and Steiner, Benoit and He, Horace and Chintala, Soumith},
title = {{CompilerGym}: A Reinforcement Learning Toolkit for Compilers},
howpublished = {\url{https://github.com/facebookresearch/CompilerGym/}},
year = {2020}
}
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distributions
Built Distributions
Hashes for compiler_gym-0.1.8-py3-none-manylinux2014_x86_64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | e3e13daa32f79fcc4adc49a6b342fc4f362b0c306c7bcc0599ebb2c7a426ba8a |
|
MD5 | 5f28642c2f35eb358bc10768d9d1f365 |
|
BLAKE2b-256 | 8fe99ec3983ab35eea9c04a13de34498d6baea878d7f0c8aa14df9f47cec075a |
Hashes for compiler_gym-0.1.8-py3-none-macosx_10_9_x86_64.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | ba521a2c1189189dc4def3983158fbd5615ed90d9a2af9d3c2768de8a66a3094 |
|
MD5 | 2f9e58c1b54195c9cef9f613eb56b72c |
|
BLAKE2b-256 | 3111e1b3a0e444f4696d6e0f1633e6e118a417afabcf762b0e6cecb51745a199 |