Details
-
Improvement
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
None
Description
The way that we dispatch to SSE4 hash functions is a remnant from the Impala codebase, which checks CpuInfo on every iteration in debug builds:
https://github.com/apache/arrow/blob/master/cpp/src/arrow/util/hash-util.h#L43
However, the static model_name_ member is causing some non-determinism related to static member lifetime as reported in ARROW-3241.
I'm proposing to refactor CpuInfo into a singleton pattern and handle SIMD vs non-SIMD dispatch at a higher level rather than at the lowest level like it is now. This should hopefully make the issue in ARROW-3241 go away
Attachments
Issue Links
- links to