[ARROW-8141] [C++] Optimize BM_PlainDecodingBoolean performance using AVX512 Intrinsics API

• Type: Improvement
• Status: Resolved
• Priority: Minor
• Resolution: Fixed
• Fix Version/s: 0.17.0
• Component/s: C++
• Labels: pull-request-available
• External issue URL: https://github.com/apache/arrow/issues/24346
• Assignee: Frank Du
• Reporter: Frank Du
• Created: 18/Mar/20 03:19
• Updated: 11/Jan/23 07:58
• Resolved: 19/Mar/20 02:26

Description

We are running benchmark on the arrow avx512 build, perf show unpack1_32 as the major hotspot for BM_PlainDecodingBoolean indicator.

Implement this func with Intrinsics code show big improvements. See below the results on CLX 8280 cpu which is capable of AVX512.

Indictor	default sse build	avx512 build	avx512 build + Intrinsics	Intrinsics improvements
BM_PlainDecodingBoolean/1024(G/s)	1.55394	3.77701	5.02805	1.331224964
BM_PlainDecodingBoolean/4096(G/s)	1.83472	5.3826	8.3443	1.550235945
BM_PlainDecodingBoolean/32768(G/s)	2.00957	6.1258	10.3793	1.694358288
BM_PlainDecodingBoolean/65536(G/s)	2.02249	6.20035	10.5778	1.706000468

Issue Links

• links to: GitHub Pull Request #6650