Correct #20054-D errors found when compiling on ARM #42
Conversation
The nvcc compiler on ARM is more stringent and detects incorrect shared memory usage in kernels. The static shared memory approach only works with default constructor types. So we update `PathElement` to properly follow default construction rules.
|
CI style checks are failing, possibly because they're pointing to an outdated branch. https://github.com/rapidsai/gputreeshap/actions/runs/7185932141/job/19570236478?pr=42#step:6:1 |
|
I opened #43 to update the CI workflows. If that works, we can merge in the upstream and retry this PR. This is needed to unblock CUDA 12.2 support in rapidsai/cuml#5711. |
|
Just a note, in Bradley's PR that was updating the CI workflows branches, we have learned there are more changes likely needed. Think this would be better taken on by someone familiar with CI on this repo At this point, it would be good to decide whether we are ok with this PR as it stands (assuming the CI issues are left unresolved) As part of figuring this out, we are testing these changes with cuML (using CUDA 12.2) in PR ( rapidsai/cuml#5719 ). Hopefully that gives us a better indication of whether the changes are working the way we expect in cuML CI |
Pick up rapidsai/gputreeshap#42 to fix build on ARM.
The nvcc compiler on ARM is more stringent and detects incorrect shared memory usage in kernels. The static shared memory approach only works with default constructor types.
So we update
PathElementto properly follow default construction rules.