RFR: 8309130: x86_64 AVX512 intrinsics for Arrays.sort methods (int, long, float and double arrays) [v3]

[Andrew Haley]

On Tue, 30 May 2023 20:10:09 GMT, Srinivas Vamsi Parasa <duke at openjdk.org> wrote:

>> The goal is to develop faster sort routines for x86_64 CPUs by taking advantage of AVX512 instructions. This enhancement provides an order of magnitude speedup for Arrays.sort() using int, long, float and double arrays.
>> 
>> This PR shows upto ~13x improvement for 32-bit datatypes (int, float) and upto 8x improvement for 64-bit datatypes (long, double) as shown in the performance data below.
>> 
>> **Arrays.sort performance data** 
>> 
>> |	Arrays.sort benchmark	|	Array Size	|	Baseline (us/op)	|	AVX512 Sort (us/op)	|	Speedup	|
>> |	---	|	---	|	---	|	---	|	---	|
>> |	ArraysSort.doubleSort	|	100	|	0.639	|	0.217	|	2.9x	|
>> |	ArraysSort.doubleSort	|	1000	|	8.707	|	3.421	|	2.5x	|
>> |	ArraysSort.doubleSort	|	10000	|	349.267	|	43.56	|	**8.0x**	|
>> |	ArraysSort.doubleSort	|	100000	|	4721.17	|	579.819	|	**8.1x**	|
>> |	ArraysSort.floatSort	|	100	|	0.722	|	0.129	|	5.6x	|
>> |	ArraysSort.floatSort	|	1000	|	9.1	|	2.356	|	3.9x	|
>> |	ArraysSort.floatSort	|	10000	|	336.472	|	26.706	|	**12.6x**	|
>> |	ArraysSort.floatSort	|	100000	|	4804.716	|	427.397	|	**11.2x**	|
>> |	ArraysSort.intSort	|	100	|	0.61	|	0.111	|	5.5x	|
>> |	ArraysSort.intSort	|	1000	|	8.534	|	2.025	|	4.2x	|
>> |	ArraysSort.intSort	|	10000	|	310.97	|	24.082	|	**12.9x**	|
>> |	ArraysSort.intSort	|	100000	|	4484.94	|	381.01	|	**11.8x**	|
>> |	ArraysSort.longSort	|	100	|	0.636	|	0.28	|	2.3x	|
>> |	ArraysSort.longSort	|	1000	|	8.646	|	4.425	|	2.0x	|
>> |	ArraysSort.longSort	|	10000	|	322.116	|	53.094	|	**6.1x**	|
>> |	ArraysSort.longSort	|	100000	|	4448.171	|	696.773	|	**6.4x**	|
>
> Srinivas Vamsi Parasa has updated the pull request with a new target base due to a merge or a rebase. The incremental webrev excludes the unrelated changes brought in by the merge/rebase. The pull request contains three additional commits since the last revision:
> 
>  - Merge branch 'master' of https://git.openjdk.java.net/jdk into avx512sort
>  - remove libstdc++
>  - 8309130: x86_64 AVX512 intrinsics for Arrays.sort methods (int, long, float and double arrays)

What happens to really short arrays? Your patch should include macro benchmarks for e.g. 50 and 10.

-------------

PR Comment: https://git.openjdk.org/jdk/pull/14227#issuecomment-1570447615