On Wed, 23 Oct 2024 15:46:31 GMT, Joe Darcy <da...@openjdk.org> wrote:
>> src/jdk.incubator.vector/share/classes/jdk/incubator/vector/Float16.java
>> line 865:
>>
>>> 863: public static Float16 min(Float16 a, Float16 b) {
>>> 864: return
>>> shortBitsToFloat16(floatToFloat16(Math.min(a.floatValue(),
>>> 865:
>>> b.floatValue()) ));
>>
>> I assume we will optimize these min/max implementations in the future.
>> Otherwise, the extra space should be removed.
>
> The general intention is that the back-end intrinsification efforts that use
> native hardware instructions, where present, will provide the main speed
> boost. Therefore, the preference has been to keep the Java sources in Float16
> as straightforward and "obviously correct" as possible.
Yes, C2 optimizations and hardware instruction generation using intrinsics and
other means where needed will follow in one or more subsequent PRs, where the
auto-vectorizer and Vector API can utilize a common backend.
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/21574#discussion_r1813197718
