Agenda

• Individual/group updates
• Performance questions
• C++20 modules
• Default parameter handling
• Review non-WIP PRs

Individual/group updates

LR

• Coalesced buffer comm merged. People should test downstream.
• Added additional solver features
  • boundary conditions (within task lists in solvers)

JM

• Small fixes/QoL improvements in various place
  • AMR criteria on Sparse
  • PostInitialze hooks
  • Error check for static mesh refinement
  • Contour1D script
  • Fix support for non-cart coords in yt. PG will ping
• Updated from AD: close to get input compiler open source to that we can use it upstream

AR

• Two new features for type based packs
• WIP PR for virtual fields
  • ready for testing and feedback
• Scratch variable system
  • also ready for testing
  • same memory aliased by various types/shapes at runtime
  • is there a potential issue with tasking (race conditions wrt order)

JD

• adding Python interfaces for pgen in
• potentially also for boundary conditions
• Python interpreter embedded in riot
• big Q: how fast is it (especially when running on GPUs)

PG

• Updated CI
• Fixed VariableExists check in hdf5 restart
• AMD Frontier issues
  • (Re)discovered "GPU memory access faults" on Frontier
  • Can be mitigated by disabling direct intra-node GPU-to-GPU comm (which comes at a significant performance hit, e.g., 40% in small reproducer)
  • Will discuss with OLCF and HPE during office hour
  • JM interested in joining

Discussion on performance

• register usage has potentially significant impact on performance (though in a somewhat nonlinear, hard to predict way)
• (hidden) gradual increase of register usage? -> potentially
• -> reintroduce automated performance monitoring (with real life example)
• get Galen/CS/... involved

C++20 modules

• is this sth we may want to look into?
• eventually yes, right now not too high on the agenda
• might be a good opportunity to cleanup/rethink src/interfaces

Discussion on default parameter handling

• GetOrAdd defaults between Parthenon and downstream codes might cause some runtime errors
• Could introduce some (Parthenon) internal version of GetOrAdd to keep track
• however, potential number of variables currently limited
• keep as is for now as the error messages are clear

Non WIP

• #1323 just code shuffle: PM will review
• #1322 JM PG will review
• #1319 PG will review, otherwise
• #1308 PM will add unit test
• #1320 ready to go

tentative next meeting Oct 9