Parameterize SubtileGridSizer to allow variation of the padding on non-interface field#347
Conversation
…on non-interface field. This will lead to an API break down the road, as we know require the backend to be given.
|
@oelbert for visibility. In Pace after the January release, you will be required to give |
romanc
left a comment
There was a problem hiding this comment.
If we can avoid it, I wouldn't change the GridSizer base class (see inline).
There was a problem hiding this comment.
just generally if we're already touching this: Is there a case where we want to turn on or off (ij) and k padding independently - physics vs dynamics? Or do we acutally think padding shoud in the long term go completely and not worry about it?
There was a problem hiding this comment.
If we need to do padding for cache line prefetch, we can do it via the strides rather than element count. So I'd vote no, it should just go away.
🧱🧱 BLOCKED BEHIND GridTools/gt4py#2426 🧱🧱
Description
In order to allow zero-copy memory mapping from Fortran (case of GEOS) we need to parameterize
GridSizerand children to allow variation of the padding on non-interface field. This PR takes care of it by requiring the backend and checking it's Fortran friendliness. Further discussion should question the need for padding altogether.This will lead to an API break down the road, as we know require the backend to be given to the
SubtileGridSizer.Also:
How has this been tested?
New unit test.
Checklist