SimiLie
Loading...
Searching...
No Matches
About

SimiLie

Github repo here. Reference slides here.

SimiLie is a performance-portable (CPU & GPU) C++20 library aiming to extent the capabilities of Kokkos and DDC to offer a complete toolkit able to perform tensor calculus, differential calculus and solving PDE. A key paradigm is the exclusive support of structured meshes to avoid sparse linear algebra and produce mostly-embarrassing parallel code. It should be able to address any multiphysical problem on fixed grid from eventually-relativistic classical field theory (in particular: solid & fluid mechanics, electromagnetism and gravitation).

The physics module contains the available physical models and local material operators.

The multidomains module couples structured patches through sampling policies and global differential stencils. Its potential flow example illustrates mapped geometry, circulation cuts and the ONELAB interface.

Important
Absolute WIP, there is no guarantee for proper working. Only OPENMP and CUDA backends have been tested.

AI-operability

Coding agents are extensively used to develop SimiLie. The AGENTS.md file provides the procedure to operate and develop with SimiLie, with the instruction to automatically detect the presence of a GPU and use the similie_env Dockerfile with OPENMP or CUDA environment/compilation. Ie.:

vibe
run tests

Should detect the presence of a GPU, build the corresponding docker image, compile Similie with g++ or nvcc and run the tests.

Development plan

  1. Short term: simulate electromagnetism on manifold.
  2. Mid term: simulate geometrodynamics.
  3. Long term: generalize to all kinds of classical gauge theories.
  4. Extra-long term (several years): quantization.

References