-
Notifications
You must be signed in to change notification settings - Fork 37
2020.04.22 Meeting Notes
Phil: Originally there was concern that the requirements of LANL was a bit divergence from other projects ("interface" package, for example).
Phil: Reminds us that he would like to demonstrate AMR on GPU by end-of FY. @gshipman big proponent of this.
How do we get from where we are today to AMR GPU by September.
Josh Dolence really would like Sparse Variables to be in for the milestone.
We must have independent milestone requirements by Tuesday, 5/5/2020. Will collaborate offline on generating this document.
We will prefix every shared input block name with "Parthenon". #121
Easy to mess up dimensions - wants to make it harder to incorrectly up-rank an array.
Phil will document the correct way to solve this problem.
Andrew will talk to Karen Tsai offline.
- Forrest invested time testing performance of view of views and believes it's serviceable for now, but the door should be left open to future optimizations
- Josh developed an attempt to make a view of views from a container. Looks in-line with desired properties:
- Ability to flatten higher-d variables into view of 3d views (already done)
- Desire to control index ordering of variables returned (Josh to implement now)
- Desire to return a contiguous block of memory if possible (Jonah to implement later)
- Desire to stash/cache view of views when possible to reduce allocations (Jonah to implement later)
- Josh to open up WIP MR so iterator can be workshopped
- Jonah to produce some preliminary performance numbers and a bug fix
- Forrest to performance test it using his framework
Josh Brown is investigating regression tests. He's been refactoring the python scripts from Athena++ and integrating them with CTest. Hoping to have that done by the end of the week or by early next week.
Jonas Lippuner has been reviewing PRs and been investigating the GPU iterators pull request. Trying to help forrest and Phil come up with the right solution for this.
Andrew Gaspar has been working on Darwin Power9 support (#96, #114) and improving the "public API" for parthenon. (#110)
AthenaArray is gone, replaced with ParArrayND. Refactored stages and containers, inverting the relationship.
Changed it so that input decks can specify which variables are output to graphics.
Pull Request (#108) to better capture application requirements. Allow application developers to specify the required fields.
Started down the path of Kokkosization of Parthenon. Discovered an issue with ParArrayND leading to developer mistakes. Discovered an issue with typedefs that he introduced. Has a fix. #116
Running into CMake issues.
Forrest has been experimenting with Container Iterators. Found surprising performance results - View<View<_>> seems plausible. 20% overhead in low-compute intense-ivity loops, 5% in high compute intense-ivity loops.