Skip to content

Conversation

@tbensonatl
Copy link
Collaborator

For some CUB reduction types, the batches field in the CUB params struct is uninitialized. The field is not needed for the reduction, but it is included when performing cache lookups, so the uninitialized value could cause false negatives (cache misses) and additional plan creations. This did not cause correctness issues but could impact performance. Thus, zero-initialize the batches field to prevent false negatives on cache lookup. The CUB cache is currently disabled by default but can be enabled at build time by the user.

For some CUB reduction types, the batches field in the CUB params struct
is uninitialized. The field is not needed for the reduction, but it is
included when performing cache lookups, so the uninitialized value could
cause false negatives (cache misses) and additional plan creations. This
did not cause correctness issues, but could impact performance. Thus, zero
initialize the batches field to prevent this. The CUB cache is currently
disabled by default, but can be enabled at build time by the user.
@tbensonatl
Copy link
Collaborator Author

/blossom-ci

@tbensonatl tbensonatl merged commit fb6de89 into main Dec 8, 2023
cliffburdick pushed a commit that referenced this pull request Dec 8, 2023
For some CUB reduction types, the batches field in the CUB params struct
is uninitialized. The field is not needed for the reduction, but it is
included when performing cache lookups, so the uninitialized value could
cause false negatives (cache misses) and additional plan creations. This
did not cause correctness issues, but could impact performance. Thus, zero
initialize the batches field to prevent this. The CUB cache is currently
disabled by default, but can be enabled at build time by the user.
@tbensonatl tbensonatl deleted the bugfix/initialize-cub-params-batches branch December 19, 2023 00:37
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants