Refactor the buffer system and add methods for synchronize_communication!
#4939
Conversation
|
and splitting some tests to speed up the distributed tests... |
simone-silvestri
changed the title
simone-silvestri
changed the title
synchronize_communication!
| recv = on_architecture(arch, zeros(eltype(data), H, size_y, size(parent(data), 3)))) | ||
| # Either we pass corners or it is a 1D parallelization in x, note that, even for 2D parallelizations, we need a | ||
| # buffer that includes the corners (i.e. a OneDBuffer) if we are at the edge of the domain in y since we are not | ||
| # passing all the corners in that case |
There was a problem hiding this comment.
now that we have AI to make things pretty easy, I think it makes sense to convert all these comments to docstrings. This opens the possibility to getting a gist of the program by looking at the docs
| on_architecture(arch, zeros(eltype(data), H, size(parent(data), 2), size(parent(data), 3)))) | ||
| else | ||
| return (send = on_architecture(arch, zeros(eltype(data), H, size(grid, 2), size(parent(data), 3))), | ||
| recv = on_architecture(arch, zeros(eltype(data), H, size(grid, 2), size(parent(data), 3)))) |
There was a problem hiding this comment.
this named tuple looks like the same thing as OneDBuffer. What's going on here? Do you use the type secretly for dispatch somehow? Do we need to use an explicit type here rather than a named tuple?
A comment might clear this up.
There was a problem hiding this comment.
right it is only for dispatch, I can add another buffer type: TwoDBuffer
|
I'd love if we can take the opportunity of the refactor to clean up the code and make it readable, eg add comments in mysterious parts, use docstrings, and convert golf'd code to a more evocative style |
Co-authored-by: Gregory L. Wagner <[email protected]>
Co-authored-by: Gregory L. Wagner <[email protected]>
github copilot at the rescue :) |
I am starting to dissect #4811 now that the formulation is complete and the tests pass, so that we do not have to merge 2000 lines and 107 files altogether.
This is a very simple start -> a quality of life refactor of the buffer system for distributed computations.