ai2cm / pace

Implementation of the FV3GFS / SHiELD atmospheric model in Python
https://ai2cm.github.io/pace/
Apache License 2.0
37 stars 12 forks source link

FVSubgridZ failure on `dace:gpu` backend #410

Open FlorianDeconinck opened 1 year ago

FlorianDeconinck commented 1 year ago

FVSubgridZ leads to a CUDA 701 when ran under dace:gpu. The underlying cause is an overflow of the registers leading to cuda launch failure.

Temporary workaround as introduced in PR #400 is a skip_test that cancelled our translate test for this particular backend.

Avenue for solutions:

WARNING: this probably also fail when doing orchestration, it remains untested