Integration with hypre - optimise AMG solvers

jonas-schmitt / evostencils

Automating the Design of Multigrid Methods with Evolutionary Program Synthesis

GNU Affero General Public License v3.0

11 stars 4 forks source link

Integration with hypre - optimise AMG solvers #3

Closed dinesh-parthasarathy closed 11 months ago

dinesh-parthasarathy commented 1 year ago

map each solver individual to cmd line input arguments in hypre
evaluate each individual automatically with hypre, by passing run time args.
modify program optimization to evaluate a bunch of individuals in a single program run (to eliminate redundant amg setup phase time)
Store fitness of all individuals in every generation (for visualizing the evolution process as a gif/animation later)
Added scripts / IPython notebooks for analyzing evolved solvers: Weak scaling studies, visualizing solver structure, Error mode analysis using FFT, etc.

jonas-schmitt commented 1 year ago

I'm currently in the process of checking if the ExaStencils code generation still works with the proposed changes.

There seem to be some issues with ExaStencils and newer Java versions (I'm currently testing openjdk-17 under Debian), which means that I probably need to update the installation instructions. This is something I have to fix myself.
Could you provide instructions how to install Hypre and set it up with EvoStencils? For instance, you could provide a notebook that guides the user through all steps necessary to perform the installation, setup and execution of EvoStencils with the Hypre backend. You can use the tutorial.ipynb notebook as a template.

As soon as these changes have been made, I'd do one last integration check with the provided notebook / instructions for the Hypre backend and then I'd be happy to perform the merge.

jonas-schmitt commented 11 months ago

I've now confirmed that your fork still seems to work with the ExaStencils backend at least for openjdk 11. I yet have to check the compatibility with other versions.

However, there is an issue with the evaluation (evaluate_multiple_objectives function), since use_hypre=True is set globally. To resolve this issue, we should introduce a parameter for that option, that either has to be specified by the user or can automatically be determined based on the context the function is called in.

I think in general, the hypre backend is not yet production-ready and probably requires some cleanup and testing. I think it is thus a better idea to merge it to a separate hypre branch for now and perform all subsequent integration efforts there until the code is ready to be merged to the main branch. I will thus perform the merge there and would suggest that you continue working on this new branch.