Turbo aims to be a constraint solver entirely on GPUs. Its theoretical parallel model is described in this paper AAAI 2022. Turbo is part of a larger project called Lattice Land. I write about Turbo in a technical journal where you can learn about the specifics of this solver.
- Cuda compiler
nvcc(>= 12.0). - libxml2
- CMake (>= 3.27)
- Doxygen
The other dependencies will be pulled and compiled automatically by CMake.
You can first clone this repository, and then git checkout the latest released tag to ensure more stability.
The following command will configure and compile Turbo for GPU for the GPU architecture of your computer (native architecture).
cmake --workflow --preset gpu-release --fresh
./build/gpu-release/turbo -s -v -i -t 20000 benchmarks/example_wordpress7_500.fzn
Other compilation builds are also available:
- GPU Debug version:
cmake --workflow --preset gpu-debug --fresh - CPU Debug version:
cmake --workflow --preset cpu-debug --fresh - CPU Release version:
cmake --workflow --preset cpu-release --fresh
Alternatively, you can use these commands without presets and workflow (it is useful compilation scenario falling outside what is provided by the presets):
cmake -DCMAKE_BUILD_TYPE=Release -DGPU=ON -DREDUCE_PTX_SIZE=ON -DCMAKE_VERBOSE_MAKEFILE=ON -Bbuild/gpu-release
cmake --build build/gpu-release
The file turbo.gpu.release.msc can be copied in your Minizinc configuration directory (on Linux: ~/.minizinc/solvers).
You should edit the file and update the paths of executable and mznlib to match the location of Turbo on your system.
You should be able to run Turbo directly from the MiniZinc IDE and from the command line:
minizinc -s -t 60000 --solver turbo.gpu.release benchmarks/mzn-challenge/2022/wordpress/Wordpress7_Offers500.dzn benchmarks/mzn-challenge/2022/wordpress/wordpress.mzn
Please see lattice-land.