RFR: 8277617: Adjust AVX3Threshold for copy/fill stubs

[David Holmes]

On 1/12/2021 5:11 pm, Jie Fu wrote:
> On Wed, 1 Dec 2021 03:38:00 GMT, Jie Fu <jiefu at openjdk.org> wrote:
> 
>>>> @dholmes-ora We see about 25% gain on a micro on our latest platform. There is no cpuid bit for this, so the closest was to check for the new serialize ISA supported on this platform.
>>>
>>> It would be better to add a jmh test for this opt.
>>> Thanks.
>>
>>> @DamonFool There are jmh tests for Arraycopy in test/micro/org/openjdk/bench/java/lang/Arraycopy.java.
>>
>> So how about posting the detailed perf data before and after this patch?
>> Thanks.
> 
>>>> @DamonFool There are jmh tests for Arraycopy in test/micro/org/openjdk/bench/java/lang/Arraycopy.java.
>>>
>>>
>>> So how about posting the detailed perf data before and after this patch? Thanks.
>>
>> Before: Benchmark Mode Cnt Score Error Units ArrayCopy.arrayCopyObject avgt 5 19.538 ± 0.073 ns/op ArrayCopy.arrayCopyObjectNonConst avgt 5 20.513 ± 0.104 ns/op ArrayCopy.arrayCopyObjectSameArraysBackward avgt 5 15.919 ± 0.652 ns/op ArrayCopy.arrayCopyObjectSameArraysForward avgt 5 15.669 ± 0.359 ns/op
>>
>> After: Benchmark Mode Cnt Score Error Units ArrayCopy.arrayCopyObject avgt 5 16.957 ± 0.584 ns/op ArrayCopy.arrayCopyObjectNonConst avgt 5 17.221 ± 0.036 ns/op ArrayCopy.arrayCopyObjectSameArraysBackward avgt 5 12.952 ± 0.068 ns/op ArrayCopy.arrayCopyObjectSameArraysForward avgt 5 13.562 ± 0.124 ns/op
> 
> Thanks @sviswa7 for your sharing.
> So the performance number looks good on Intel's latest AVX512 platform.
> 
> We don't use the 64-byte instructions as default on Intel's old AVX512 platforms, right?
> If so, is it possible a performance regression for the old platforms after this patch?
> Thanks.

The old platforms, for which serialize() is not true, will just use 
AVX3Threshold as they do today.

David
----

> -------------
> 
> PR: https://git.openjdk.java.net/jdk/pull/6512
>