jianxind edited a comment on pull request #7029: URL: https://github.com/apache/arrow/pull/7029#issuecomment-618855696
cc @emkornfield The AVX512 path is straightforward as the helper of mask_compress/mask_expand API provide by AVX512. For potential path-finding of SSE/AVX2, as you pointed in the Jira, a solution with fixed lookup table may help, I will work the chance then but it definitely need take more time thus I commit this done part firstly. Below is the benchmark data on Avx512 device before/after the intrinsics: Before: BM_PlainEncodingSpacedFloat/1024 1471 ns 1469 ns 476373 bytes_per_second=2.59603G/s BM_PlainEncodingSpacedDouble/1024 1498 ns 1496 ns 468131 bytes_per_second=5.09834G/s BM_PlainDecodingSpacedFloat/1024 1266 ns 1265 ns 554320 bytes_per_second=3.01623G/s BM_PlainDecodingSpacedDouble/1024 920 ns 919 ns 759151 bytes_per_second=8.30509G/s After: BM_PlainEncodingSpacedFloat/1024 513 ns 512 ns 1374561 bytes_per_second=7.44416G/s BM_PlainEncodingSpacedDouble/1024 635 ns 634 ns 1108739 bytes_per_second=12.0322G/s BM_PlainDecodingSpacedFloat/1024 217 ns 217 ns 3233406 bytes_per_second=17.613G/s BM_PlainDecodingSpacedDouble/1024 309 ns 309 ns 2267740 bytes_per_second=24.7257G/s ---------------------------------------------------------------- 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. For queries about this service, please contact Infrastructure at: [email protected]
