<@UR2B5DTTL> For log I get comparable results: Aut...

# mathematics

[altavir]

@Aleksei Dievskii For log I get comparable results:
Automatic field log completed in 7721 millis
Raw Viktor log completed in 7067 millis

[Aleksei Dievskii]

both with Graal and OpenJDK?

[altavir]

A moment. This is graal result. Also there is a cross-inline call inside.

[altavir]

Yeah, it seems that on regulat AdoptOpenJDK ln is muuuch slower:

Automatic field log completed in 39923 millis

Raw Viktor log completed in 6480 millis

[altavir]

The ratio of results is more or less the same for GraalVM on Intel i5. Regular inlining is more than 2 times faster than Viktor.

[altavir]

Removing cross-inline did not help a lot, so yeah, it is all about GraalVM magic. More reasons for me to integrate Viktor as a possible back-end.

[Aleksei Dievskii]

well, it stands to reason that a JVM that is internally actively SIMDizing everything will do it better than our JNI hacks.

[altavir]

OpenJDK has SIMD capability. I think that GraalVM just does it better due to more advanced compiler. I am really happy that the results are so good on CE version. It means that we can actually use it in sceince. They say thay EE is even better.

[altavir]

There was a report by Oleg Selajev in past KotlinConf and I had a small discussion with him afterwards. I was most concerned about boxing avoidance in GraalVM, which works great by the way, but it seems like vectorization works great as well.