adriangb commented on PR #18832: URL: https://github.com/apache/datafusion/pull/18832#issuecomment-3556437805
Here's what I'm seeing so far: Function & State | main (us) | specialized (us) | Change -- | -- | -- | -- in_list_utf8(5) (1024, 0) IN (1, 0) | 3.6205 | 4.8594 | Regressed (+34.239%) in_list_utf8(10) (1024, 0) IN (1, 0) | 3.6363 | 5.1001 | Regressed (+40.312%) in_list_utf8(20) (1024, 0) IN (1, 0) | 3.6707 | 4.9045 | Regressed (+33.600%) in_list_f32 (1024, 0) IN (1, 0) | 3.5255 | 3.2288 | Improved (−8.2971%) in_list_i32 (1024, 0) IN (1, 0) | 3.4923 | 3.6403 | Regressed (+3.8772%) in_list_utf8(5) (1024, 0.2) IN (1, 0) | 4.9284 | 6.3583 | Regressed (+29.002%) in_list_utf8(10) (1024, 0.2) IN (1, 0) | 4.9788 | 6.3273 | Regressed (+27.064%) in_list_utf8(20) (1024, 0.2) IN (1, 0) | 5.1145 | 6.4905 | Regressed (+26.907%) in_list_f32 (1024, 0.2) IN (1, 0) | 4.4894 | 4.2540 | Improved (−5.2512%) in_list_i32 (1024, 0.2) IN (1, 0) | 4.4055 | 4.2869 | Improved (−2.7526%) in_list_utf8(5) (1024, 0) IN (3, 0) | 3.5576 | 4.8733 | Regressed (+37.000%) in_list_utf8(10) (1024, 0) IN (3, 0) | 3.4999 | 4.8737 | Regressed (+37.108%) in_list_utf8(20) (1024, 0) IN (3, 0) | 3.4897 | 4.9039 | Regressed (+40.449%) in_list_f32 (1024, 0) IN (3, 0) | 3.5114 | 3.2550 | Improved (−7.3270%) in_list_i32 (1024, 0) IN (3, 0) | 3.5199 | 3.4879 | Noise Threshold (−0.9158%) in_list_utf8(5) (1024, 0.2) IN (3, 0) | 5.0196 | 6.2232 | Regressed (+23.865%) in_list_utf8(10) (1024, 0.2) IN (3, 0) | 5.0453 | 6.5280 | Regressed (+28.004%) in_list_utf8(20) (1024, 0.2) IN (3, 0) | 5.1005 | 11.124 | Regressed (+118.13%) in_list_f32 (1024, 0.2) IN (3, 0) | 4.5049 | 4.3126 | Improved (−4.6367%) in_list_i32 (1024, 0.2) IN (3, 0) | 4.4026 | 4.4424 | Noise Threshold (+1.0449%) in_list_utf8(5) (1024, 0) IN (10, 0) | 3.5275 | 4.9709 | Regressed (+40.969%) in_list_utf8(10) (1024, 0) IN (10, 0) | 3.5613 | 4.9125 | Regressed (+37.386%) in_list_utf8(20) (1024, 0) IN (10, 0) | 3.5589 | 4.8611 | Regressed (+36.578%) in_list_f32 (1024, 0) IN (10, 0) | 3.5596 | 3.2017 | Improved (−10.003%) in_list_i32 (1024, 0) IN (10, 0) | 3.4917 | 3.5190 | Noise Threshold (+1.0216%) in_list_utf8(5) (1024, 0.2) IN (10, 0) | 5.0368 | 6.6162 | Regressed (+31.725%) in_list_utf8(10) (1024, 0.2) IN (10, 0) | 5.0980 | 6.6660 | Regressed (+30.039%) in_list_utf8(20) (1024, 0.2) IN (10, 0) | 5.2543 | 6.5350 | Regressed (+24.335%) in_list_f32 (1024, 0.2) IN (10, 0) | 4.5609 | 4.3248 | Improved (−5.3746%) in_list_i32 (1024, 0.2) IN (10, 0) | 4.4460 | 4.3543 | Improved (−2.7754%) in_list_utf8(5) (1024, 0) IN (100, 0) | 3.6542 | 4.9992 | Regressed (+36.952%) in_list_utf8(10) (1024, 0) IN (100, 0) | 3.6529 | 4.8772 | Regressed (+33.560%) in_list_utf8(20) (1024, 0) IN (100, 0) | 3.6155 | 5.0250 | Regressed (+39.127%) in_list_f32 (1024, 0) IN (100, 0) | 3.6029 | 3.2448 | Improved (−9.7603%) in_list_i32 (1024, 0) IN (100, 0) | 3.6048 | 3.4770 | Improved (−3.5307%) in_list_utf8(5) (1024, 0.2) IN (100, 0) | 5.3988 | 6.6108 | Regressed (+22.866%) in_list_utf8(10) (1024, 0.2) IN (100, 0) | 5.4776 | 6.6567 | Regressed (+21.591%) in_list_utf8(20) (1024, 0.2) IN (100, 0) | 5.3721 | 6.6470 | Regressed (+24.058%) in_list_f32 (1024, 0.2) IN (100, 0) | 4.7693 | 4.3281 | Improved (−9.9814%) in_list_i32 (1024, 0.2) IN (100, 0) | 4.5402 | 4.4076 | Improved (−3.5247%) I think we'd need to add benchmarks for other primitive types. And it's interesting that Utf8 regresses a lot across the board. I guess the benefits of vectorization / computing the entire hashes at once outweighs the dynamic dispatch? -- 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] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
