Rich-T-kid commented on PR #10441: URL: https://github.com/apache/arrow-rs/pull/10441#issuecomment-5396969816
I created this [issue](https://github.com/apache/arrow-rs/issues/10815) to track adding the remaining types to the take kernel benchmarks split into two PR's - https://github.com/apache/arrow-rs/pull/10816 - https://github.com/apache/arrow-rs/pull/10819 I've also split up the take optimizations into 4 separate PRs, some blocked by the benchmark PR's above - optimize `take(fzl)` (this PR) - [optimize](https://github.com/apache/arrow-rs/pull/10812) `take(list<T>)` - [optimize](https://github.com/apache/arrow-rs/pull/10813) `take(Boolean)` - optimize `take(map)` (this PR) - skip assertion in `take_bits()` : this is blocked on the PR's above this since it requires threading some logic throughout the other types/call sites TODO: - optimize `take(union[dense/sparse])` - optimize `take(REE)`. nothing obvious here but maybe we could pre-compute a bitmap of which indices differ from index - 1. this should allow for a two pass approach thats faster than the repeated boolean checks in this loop -> https://github.com/apache/arrow-rs/blob/e1e71ea89b211758e20254296008f96e744ed138/arrow-select/src/take.rs#L966 for the other kernels I didn't find anything to optimize - `take_native` : we already use unsafe accessor methods + pre-allocate. i don't think we could do better for random access reads similar for - - `take_bytes()` + `take_view()` - - `take_dict()` goes through `take_primitive()` -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
