Welcome to the NeuroSLP Repo!

Section 1 - About NeuroSLP

NeuroSLP extends the work done in NeuroVectorizer and enhance llvm's parameter selection for Superword Level Parallelism (SLP).

NeruoSLP uses deep reinforcement learning to estimate the optimal LLVM slp-max-reg-size and slp-threshold parameters source code.

From our evaluation of 1,000 programs NeuroSLP is capable of improving runtime performance by over 6.5% and up to 54% in select applications.

Section 1.1 - Download Links

Our project is available to download with the following links:

Section 1.2 - The Team


Kyle Astroth	kastroth@umich.edu	LinkedIn	GitHub
Sam Gonzalez	samgonza@umich.edu	LinkedIn	GitHub	Website
Carson Hoffman	hoffcar@umich.edu	LinkedIn	GitHub
Xiangyu Qin	qinx@umich.edu	LinkedIn	GitHub

Section 2 - Getting Started

In this section we'll go through everything you'll need to know about setting up, training, and testing NeuroSLP.

Section 2.1 - Dependencies

NeuroSLP has the following dependencies:

# TF2
pip install tensorflow

# Ray
pip install ray==0.8.4

# RLlib
pip install ray[rllib]==0.8.4

# LLVM. Currently tested with clang 14.0.0.1
sudo apt-get install clang-14

# Clang for python
pip install clang

You may also need to install the Anaconda depending on your development environment. For more detailed instructions please refer to NeuroVectorizer's extended documentation here.

Section 2.2 - Configuration

Configuration is mostly achieved by modifying NeuroSlp/preprocess/configure.sh. Some important variables you may need to modify are:

Varible Description

CLANG_PATH Path to libclang

CLANG_BIN_PATH Path to clang binary

SOURCE_DIR Path to files used to generate for code2vec histograms

TEST_SHELL_COMMAND_TIMEOUT Timeout used while compiling programs

Varible	Description
`CLANG_PATH`	Path to libclang
`CLANG_BIN_PATH`	Path to clang binary
`SOURCE_DIR`	Path to files used to generate for code2vec histograms
`TEST_SHELL_COMMAND_TIMEOUT`	Timeout used while compiling programs

Section 2.3 - Training

Training of NeuroSLP is done via the NeuroSlp/slpTrain.sh script. By default NeuroSLP is configured to use the NeuroSlp/slpDataset/large dataset for training and outputs training results to NeuroSlp/slpTrainingResults.

To configure the training dataset and other hyperparameters you will need to modify NeuroSlp/slpAutovec.py.

For more information on available hyperparameters visit the documentation of Ray here. And the documentation for PPO here.

Section 2.4 - Testing

Testing of NeuroSLP is done via the NeuroSlp/slpTest.sh script. By default NeuroSLP is configured to use the NeuroSlp/slpDataset/testLarge1000 dataset for testing and outputs a summary of training results to NeuroSlp/slpTestResults_trainLarge_testLarge1000.out.

To change the test dataset and output path you'll have to modify the NeuroSlp/slpTest.sh script. To modify other hyperparameters you'll have to modify NeuroSlp/slpTempRollout.py instead.

In addition NeuroSlp/slpTempRollout.py can be used to continue training the NeuroSLP from a given checkpoint. For more information on how to use slpTempRollout.py run python slpTempRollout.py --help and visit the documentation provided by NeuroVectorizer here.

Please refer to the documentation linked to in Section 2.3 for more information on the available hyperpameters in Ray.

2.5 - Important Notes

Below are some important notes to keep in mind about the project in no particular order are:

The code2vec embeddings of programs are generated only first invokation of NeuroSlp/slpTrain.sh. If you want to regenerate them delete NeuroSlp/code2vec/data
If you would like to add additional training/test programs for the
NeuroSlp/slpBinCache/O3_runtimes.pkl caches the current runtimes of programs compiled with LLVM's default parameters across multiple runs of training and testing. If you want to recompute the runtimes be sure to delete the file.
NeuroSlp comes with a pretrained version of code2vec embedding using the files in NeuroSlp/slpDataset/full. This embedding is stored in NeuroSlp/slpDataset/*/obs_encodings.pkl. If you would like to produce a new embedding/add to the dataset delete this file and run NuroSlp/slpTrain.sh. This will produce a new obs_encoding.pkl in NeuroSlp/new_garbage_*.
If you want to use another model in the embedding generator (other than code2vec), you need to modify get_obs function in NeuroSlp/envs/neuroslp.py.

NeuroSLP also comes with a variety of helper programs that may be of use to you:

Program Description

NeuroSlp/decompileDir.sh Takes in a directory argument and decompiles every .o files in it into assembly

NeuroSlp/diffRuntimeAsm.py Takes in a directory argument and diffs each .s file in it ignoring code not part of the example function.

NeuroSlp/slpBruteForceRuntimes.py Evaluates the runtime of every combination of slp-max-reg-size and slp-threshold

NeuroSlp/unpackPickle.py Takes in a pkl path argument and outputs a human readable json version of it

NeuroSlp/slpDataset/preprocessLoops.py Takes in input and output directory arguments. Is used to converts all NeuroVectorizer formatted .c test programs in the input directory into NeuroSLP compatible programs.

NeuroSlp/slpDataset/makeDataset.py Takes in a size, input, and output directory argument. Is used to creates random test/training datasets

kyleastroth/NeuroSLP