[cartesian]: Mitigation for K offset write #1684
Labels
gt4py.cartesian
Issues concerning the current version with support only for cartesian grids.
Comments
This was referenced Oct 9, 2024
This comment has been minimized.
This comment has been minimized.
This was referenced Jan 8, 2025
havogt
pushed a commit
that referenced
this issue
Jan 10, 2025
The `interval` analysis in the unit test `test_K_offset_write_conditional` fails to catch a mistake in the code that leads to a race condition. Work: - Fix the bad interval - Remove not needed restriction on CUDA version Further work to fix the underlying problem and the larger issue of bound check on variable indexing is covered [here](#1684) and [there](#1754)
This was referenced Jan 13, 2025
edopao
added a commit
that referenced
this issue
Jan 15, 2025
Found other test failures related to #1684.
|
Note: this is missing the gt4py.cartesian label. |
FlorianDeconinck
added
the
gt4py.cartesian
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
The new write offset in K means we can go out of bounds without the system being able to warn about it since extent calculation won't be able to cover those access.
Think about the degenerate case of
This was an iron rule for GT4Py that we now allow to be broken so we can match patterns seen regularly in physics.
To mitigate the issue, we should introduce on
DEBUGa bound check, drop before every K offset write which should error out cleanly.DEBUGgenerate a check for bounds before K offset writeThe text was updated successfully, but these errors were encountered: