CG-Kit: Code Generation tool-Kit for algorithmic and performance portability

The development of CG-Kit addresses the shifting landscape of high-performance computing (HPC) platforms and the algorithmic complexities of large-scale scientific applications. The key challenges that CG-Kit aims to solve include handling a large code base in Fortran and/or C/C++, subdivision of code into large numbers of units supporting a wide range of physics and numerical methods, different parallelization techniques for distributed and shared memory systems and accelerator devices, and heterogeneity of computing platforms requiring coexisting variants of parallel algorithms. CG-Kit tackles the challenges by providing users with the ability to express their desired control flow and computation-to-resource map in the form a pseudocode-like recipe. It consists of standalone tools that can be combined into highly specific and effective portability and maintainability toolchains.

Algorithmic variants are different realizations of numerical algorithms that lead to the same solution outcome but differ in the details of algorithm design and/or the implementation of how the solution is obtained. The need for variants arises from differences in hardware architecture, and maintaining all variants explicitly is challenging. CG-Kit introduces a feasible way of handling variants and thereby achieves algorithmic portability, where algorithms are adapted to hardware platforms.

Tools

• cflow: Control Flow Graphs
• ctree: Source Code Trees

Install

Install the latest version of CG-Kit:

python -m pip install code-generation-toolkit

License

Released under the GNU General Public License v3.0.

Citation

Paper on arxiv:

@Article{RudiLeeChadhaEtAl24,
  title={{CG}-{K}it: Code Generation Toolkit for Performant and Maintainable
         Variants of Source Code Applied to {F}lash-{X} Hydrodynamics
         Simulations},
  author={Rudi, Johann and Lee, Youngjun and Chadha, Aidan H and Wahib, Mohamed
          and Weide, Klaus and O'Neal, Jared P and Dubey, Anshu},
  year={2024},
  note={arXiv preprint arXiv:2401.03378}
}