GEVO

GEVO (Gpu optimization using EVOlutionary computation) is a tool for automatically discovering optimization opportunities and tuning the performance of GPU kernels in the LLVM representation.

Package dependency

Python 3.6.1 or above

Python packages dependency is list in the setup.py and pyproject.toml, though they will be taken care by python pip.

NVIDIA CUDA 11

While GEVO can work with any CUDA version so long LLVM compiler can compile it, certain benchmark I tested requires Nvidia intrinsic only supported by the most recent CUDA. Thus, CUDA 11 is enforced.

❗ Note: Since 2018.8, nvidia driver 418 and beyond has a profiling permission restriction. Please refer to this nvidia document and this nvidia forum thread for how to lift the restriction for nvprof, which is the profiling tool GEVO use to measure the CUDA kernel runtime and power comsumption.

LLVM 11

LLVM 11.1 + CUDA 11.1 has been tested with current GEVO, even though LLVM 11 does not officially support CUDA 11. The oldest LLVM I have used in the past is LLVM 8. It might work, but no guarantee. More information might be found in LLVM document: "Compileing CUDA with clang"

Usually, LLVM in the system package management system is much older than the current development. If you are using Ubuntu/Debian, please refer to "LLVM nightly packages" for how to install the most recent LLVM into the system. Or just follow the following commands referred from above LLVM site for convenience.

wget https://apt.llvm.org/llvm.sh
chmod +x llvm.sh
sudo ./llvm.sh 11

Installation

Simply install GEVO from Python Package Index using pip

pip install gevo --user

Or build and install manually

git clone https://github.com/lioujheyu/gevo
cd gevo
git submodule update --init --recursive
pip install . --user

Evolve a CUDA program using GEVO

Prepare the CUDA program

GEVO interacts with target CUDA program with following requirement.

• profile.json as the profiling description - to specify how to run the target CUDA program with necessary arguments.
• A modified CUDA program with its kernel(s) loaded externally from gevo.ptx file
• A method to compile the kerel code separatedly through LLVM into cuda-device-only-kernel.ll

Please check with profile.json and Makefile/CUDA code of each benchmark (under benchmark/rodinia_3.1/cuda) as examples.

The detail steps of preparing these files are under construction.

Start the evolution

gevo-evolve is the main entry point for code evolution. Execute it under the directory of your chosen CUDA program

gevo-evolve -P <profile.json>

This start the evolution with default parameters. Please use --help for more information including the default values for each argument, like number of individual in the population or mutation rate etc.. The command in evolve.sh under each benchmark is great example for reference as well

Citation

If this project helps your research, please cite the following paper in your publication.

@article{liou2020gevo,
  author = {Liou, Jhe-Yu and Wang, Xiaodong and Forrest, Stephanie and Wu, Carole-Jean},
  title = {GEVO: GPU Code Optimization Using Evolutionary Computation},
  year = {2020},
  journal = {ACM Trans. Archit. Code Optim.}
}