[BUG] Illegal memory access to collision counter in concurrent_unordered_multimap

[gaohao95]

Describe the bug
In cpp/src/hash/concurrent_unordered_multimap.cuh, member element m_collisions is allocated in host memory, but the host pointer is passed to atomicAdd in device function insert, here, and here.

Steps/Code to reproduce bug
Any code using concurrent_unordered_multimap with count_collisions set to true should observe illegal memory access in the kernel.

[jrhemstad]

Since the concurrent_unordered_multimap is not public facing and the count_collisions functionality is not used internally, this will be fairly low priority to fix. Happy to review any PRs w/ the fix.

[gaohao95]

Since the concurrent_unordered_multimap is not public facing and the count_collisions functionality is not used internally, this will be fairly low priority to fix.

No problem at all. It's not a blocker for me as well. Just want to raise an issue here so that it can get tracked.

[jrhemstad]

No problem at all. It's not a blocker for me as well. Just want to raise an issue here so that it can get tracked.

Thanks. I think this oversight is a holdover from when the concurrent_unordered_multimap object itself was allocated in managed memory.

[github-actions]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[vyasr]

No longer relevant as of #10642.