[CUDA] Refactor device initialization

[hdelan]

Some logic for device initialization was split across platform init and device init. This duplicated some calls to get info for max block dim in Y and Z dimension. This rectifies that and makes sure the max block dims are only queried once, which is called from the device constructor.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 12.43%. Comparing base (78ef1ca) to head (2968cc1).
Report is 188 commits behind head on main.

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1363      +/-   ##
==========================================
- Coverage   14.82%   12.43%   -2.40%     
==========================================
  Files         250      241       -9     
  Lines       36220    36242      +22     
  Branches     4094     4111      +17     
==========================================
- Hits         5369     4506     -863     
- Misses      30800    31732     +932     
+ Partials       51        4      -47     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[hdelan]

Friendly ping @konradkusiak97

[konradkusiak97]

Should there be an intel/llvm CI for this?

[hdelan]

Oops forgot to link it. Here it is intel/llvm#12762

[konradkusiak97]

Aha that's strange how it was done before, it could've been partially my fault for not seeing this duplication with my caching PRs. Nice refactor, LGTM!

Just a side note, it seems some of those changes could also be applied to the HIP adapter, do you think it would be feasible to squeeze it together here?

[hdelan]

Just a side note, it seems some of those changes could also be applied to the HIP adapter, do you think it would be feasible to squeeze it together here?

The HIP adapter is only checking this stuff once. It does it in a different manner, ie it doesnt use the ur helper, but I think there isn't anything wrong with the current approach of the HIP adapter so I would be inclined to leave it as it is. Let me know if you agree!

[konradkusiak97]

Just a side note, it seems some of those changes could also be applied to the HIP adapter, do you think it would be feasible to squeeze it together here?

The HIP adapter is only checking this stuff once. It does it in a different manner, ie it doesnt use the ur helper, but I think there isn't anything wrong with the current approach of the HIP adapter so I would be inclined to leave it as it is. Let me know if you agree!

Sounds good!