Use JVM Vector API?

[CommanderTvis]

A possible way to implement the array are intrinsics introduced with JEP 338.

[dievsky]

Last time I looked at it (half a year ago), there was no SIMD mathematics there, and viktor's "area of expertise" is exactly the SIMD mathematics (exp, log etc.).

[CommanderTvis]

https://openjdk.java.net/jeps/338#Goals

[dievsky]

What exactly should I look for under that link?

[CommanderTvis]

Reliable runtime compilation and performance on x64 and AArch64 architectures: The Java runtime, specifically the HotSpot C2 compiler, shall compile, on capable x64 architectures, a sequence of vector operations to a corresponding sequence of vector hardware instructions, such as those supported by Streaming SIMD Extensions (SSE) and Advanced Vector Extensions (AVX) extensions, thereby generating efficient and performant code. The programmer shall have confidence that the vector operations they express will reliably map closely to associated hardware vector instructions. The same shall also apply to capable ARM AArch64 architectures compiling to a sequence of vector hardware instructions supported by Neon.

[dievsky]

I must shamefully admit that it's pretty hard for me to communicate with people speaking in links and quotes.
My point was that (last time I checked) Vector API was missing SIMDized exponent, logarithm, etc, Your counterpoint was a link to the Vector API proposal and a quote of its broadly stated goals.
If you intended to say they have, in fact, implemented SIMDized exponent in the last six months, why not just say so? :))

[CommanderTvis]

As for me, SIMD support means that at least vector arithmetic is optimized.

[CommanderTvis]

I think it would be OK even if they didn't provide other elementary functions: there's nothing wrong in falling back into scalar mathematics for everything except arithmetic.

[dievsky]

Ah, now I get it.

The funny thing is that the array loop arithmetic is already SIMDized, even in JDK 1.8. Every time I've checked, the C2 compiler successfully emitted AVX instructions for arithmetic operations. That's why we stopped delegating the arithmetic to JNI since 1.1.0: the JNI actually performs worse than a simple loop.

So switching to Vector API will currently make the code less readable and portable, while not improving the performance. (Well, okay, maybe the performance will improve by a couple percent because we will skip the warmup phase.)