Cartesian Biweekly 2020 07 21

Action Points

Review of action points of last meeting.

[ ] @Enrique G. P Start collecting ideas for explicit indexing and directional offsets for fields in doc

• @Enrique G. P will help prioritize the items
  [ ] @Johann D Will send around poll to find time for a gt4py mini-workshop
• Purpose of the meeting will be to decide syntax and other details related to the workshop topics
• Focus on 1-2 issues
  [ ] @Johann D Will setup a meeting to review GDP implementation with gt4py team

Action points of this meeting:

[x] Complete Doodle poll:
- https://doodle.com/poll/zn2nhrv46et9hiem
****[ ] Submit items for mini-workshop in this document:
- https://docs.google.com/document/d/1hFQ0TfO7fAiz7Xh-LE3RQucesI_k0tuCGku6pmz9qnY/edit
[x] Review Region GDP in light of figure discussion
[ ] Schedule a meeting to discuss before next instance of this meeting
[ ] Characterize minimum stencil overhead
[ ] Add absolute indexing and offset-center writes to the list of items in the GTScript syntax workshop
[ ] Select examples for off-center writes
[ ] Check whether no-check functionality already in StencilObject
[ ] Determine best way to share meeting notes — prefer a single link
[ ] Convention for GT4Py issue submission — will iterate
[ ] Present results of student projects that @Oli F and @Tobias W are working on — deadline is Aug 7

Discussion Points

Math functions

• @Tobias W is working on this from the SIR side in Dawn and from the GT4Py side

Stencil overhead

• Observed ~5ms overhead
  • Known issue
  • Experiment with how the overhead scales with number of input fields
  • Python code generated in modules will be removed
  • Characterize pybind11 overhead — this will be the minimum overhead
  • What is the minimum acceptable overhead (upper bound)? — within a few percent of the dycore runtime
  • Not currently a high priority for GT4py team
  • @Johann D is planning to submit PR for debug mode
  • Functionality might already be there — need to be check

Regions

• @Enrique G. P and @Johann D had a meeting to discuss the Regions GDP

• Discussion of @Enrique G. P ‘s figure:
  

  • Fields are aligned for the local compute domain via origins
  • Local compute domains need to be mapped to global application domain and its halos
  • Halos can be used to locate corner and edge points
  • Temporaries are generated on the fly — compute temporary origins from those of input fields
  • Need a global frame of reference — each field could store its location on the global grid, maybe only one field needs this, others can compute from origin
  • Information will also be needed for halo updates
  • Halo and origin are separate concepts
  • What does the region object need to know or compute?
  • Should the generated code answer these questions? — will be more efficient than Python

• @Johann D plans to proceed with reference implementation

Vertical treatment

• Absolute indexing
• Off-center writes
  • Always in the vertical
  • Field adapters to shift the fields
  • Shift the output field and write into the shifted field — will appear as centered write
  • Can one pass the same field with different origins?
  • Separate the vertical computations from the stencil
  • Could be done in the GT4Py frontend
  • How will analysis work in Dawn, e.g., field versioning?
• These to be added to the GTScript syntax workshop
• Increasing in priority as more FV3 code is ported