2022.06.09 Meeting Notes

Agenda

• Individual/group updates
• Referee report
• Intel update
• Developer meeting

Individual/group updates

LR

• created a couple of PR ready
  • sparse boundaries/buffer machinery working and ready for review
    • performance should be on par with original implementation
    • also implemented flux correction in one
• working on state for ParArrayNDs (e.g., to include per variable threshold)
  • extending ParArrayND to this end: ParArray inherits from State
  • added new tests
  • looking for feedback on design decisions
  • next step, pull this through packing machinery (to also include state)
  • hopefully this is also fix the allocation bug for sparse
  • does not address the "isAllocated" calls inside kernels (JD is working in a separate branch to only pack allocated variables)

JD

• working on some performance related items
• some of our classes seem prevent vectorization in a simple 3D/4D loops Pulling out raw pointer to pencil for innermost loop seems to help significantly (e.g., for WENO5 recon in Phoebus)

JMM

• looking at interpolation tables (which often span many orders of magnitude and involve log)
• tried to use approximate log and exp, but got problems with functions that are not continuously invertible
• manually wrote/rediscovered function -- should be on the arxiv soon

BR

• particle PRs got merged
• potential next step is to reduce host-device transfer by using parallel scan
• otherwise more focus on downstream dev

FG

• working on face centered fields, looking at various implementation to see what fits best
• will also look at LR PR

JS

• PM is pushing forward with rad hydro torus
• Higher resolution needs more angles (needs more memory)

PG

• PR reviews
• scaling tests

Meeting with Intel

• key issue may be lack of support for rdc/polymophism (required for particles) but way also not want it anyway for performance reasons on CUDA
• otherwise looking forward to additional support getting sycl backend work

Developer meeting

• tentative date: first full week of Oct (3-7)
• JD will find a room on site (target <= 20 people)
• talk about additional people to invite during next meeting

Next meeting Thu 30.06