mbutrovich commented on PR #12058: URL: https://github.com/apache/arrow-rs/pull/12058#issuecomment-6865287650
> Could you share benchmark numbers for them, or leave them on `try_op!`? `Int64` `mul` now uses `overflowing_mul` instead of multiplying in `i128` (6035cda2f6). Without and with 10% nulls, `int64_multiply_checked` measured 1.35x and 2.25x on [Neoverse-V2](https://github.com/apache/arrow-rs/pull/12058#issuecomment-6864850084), 2.29x and 2.86x on [Xeon 8581C](https://github.com/apache/arrow-rs/pull/12058#issuecomment-6864870951), and 2.03x and 2.55x with [`x86-64-v3`](https://github.com/apache/arrow-rs/pull/12058#issuecomment-6864872512), where it was 0.65x and 0.84x. In `i128`, the multiply stayed a scalar `imul` while LLVM packed each row's overflow flag into vector registers around it. With `overflowing_mul`, `x86-64` runs a scalar `imul`, `seto` and `or` per row, and `x86-64-v3` handles eight rows per iteration without stack spills. The unchanged `int64_add` measured 0.98x to 1.01x in the same runs. -- 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]
