Comments / metadata in assembly listings don't make sense for code vectorized using Vector API

[Piotr Tarsa]

Hi Paul,

I've submitted my "mandelbrot" implementation (i.e. the one discussed
in this topic) to the benchmark owner, but sadly he rejected it (due
to usage of non-standard API, i.e. the Vector API):
https://salsa.debian.org/benchmarksgame-team/benchmarksgame/-/issues/424

I've set up public repo on GitHub instead:
https://github.com/tarsa/benchmarksgame-java-fast
I hope it will help someone when experimenting with either Project
Panama or Project Valhalla.

Regards,
Piotr

wt., 2 lut 2021 o 18:54 Paul Sandoz <paul.sandoz at oracle.com> napisał(a):
>
> Hi Piotr,
>
> Thanks for the update.
>
> If the 30 iterations are performed from within Java itself I suspect you will see improvements, since the same compilation would not have to be performed every time. But, I dunno if that is within the benchmark rules.
>
> Paul.
>
> > On Jan 30, 2021, at 5:14 AM, Piotr Tarsa <piotr.tarsa at gmail.com> wrote:
> >
> > Hi,
> >
> > I was busy with other things, so that's why I delayed the reply.
> >
> > wt., 19 sty 2021 o 21:17 Paul Sandoz <paul.sandoz at oracle.com> napisał(a):
> >>
> >> Hi Piotr,
> >>
> >> Thanks for further sharing. I am glad you managed to make progress. I was not aware there were some benchmark rules you needed to adhere to.
> >
> > Rules are here:
> > https://urldefense.com/v3/__https://benchmarksgame-team.pages.debian.net/benchmarksgame/description/mandelbrot.html*mandelbrot__;Iw!!GqivPVa7Brio!LsQqaL3nuiK_2L-DKtjCpVy_JZKb0yX3_AVdzglYqHMkLO4m6m2o83LEz1lwmQOpVg$
> > They require bit perfect output and also the same algorithm. Well, in
> > the end it's a benchmark of programming languages, not of algorithms.
> >
> >> Re: masks, yes there is still work to do for some mask operations.
> >
> > OK, good to know.
> >
> >> Re: execution from the command line. You can run with -XX:-TieredCompilation (Remi, thanks for the correction in the prior email :-) ), and it's also possible reduce the compilation threshold (at the expense of potentially less accurate profiling information) using say -XX:CompileThreshold=1000 (the default is 10000).
> >> It’s always a bit tricky to compare a static (Rust) vs. dynamic system that needs to warm up.
> >>
> >> Paul.
> >
> > I've tested the provided options, but they don't improve performance
> > on the real benchmark:
> >
> > $ time for run in {1..30}; do ~/devel/jdk-16/bin/java
> > -XX:-TieredCompilation -XX:CompileThreshold=1000 --add-modules
> > jdk.incubator.vector -cp
> > target/classes/:/home/piotrek/.m2/repository/org/openjdk/jmh/jmh-core/1.27/jmh-core-1.27.jar
> > pl.tarsa.mandelbrot_simd_1 16000 > /dev/null; done
> > WARNING: Using incubator modules: jdk.incubator.vector
> > WARNING: Using incubator modules: jdk.incubator.vector
> > ... (repeated for each run)
> > WARNING: Using incubator modules: jdk.incubator.vector
> > WARNING: Using incubator modules: jdk.incubator.vector
> > real 0m42,213s
> > user 2m35,740s
> > sys 0m4,094s
> >
> > $ time for run in {1..30}; do ~/devel/jdk-16/bin/java
> > -XX:-TieredCompilation --add-modules jdk.incubator.vector -cp
> > target/classes/:/home/piotrek/.m2/repository/org/openjdk/jmh/jmh-core/1.27/jmh-core-1.27.jar
> > pl.tarsa.mandelbrot_simd_1 16000 > /dev/null; done
> > WARNING: Using incubator modules: jdk.incubator.vector
> > WARNING: Using incubator modules: jdk.incubator.vector
> > ... (repeated for each run)
> > WARNING: Using incubator modules: jdk.incubator.vector
> > WARNING: Using incubator modules: jdk.incubator.vector
> > real 0m40,038s
> > user 2m23,227s
> > sys 0m3,515s
> >
> > $ time for run in {1..30}; do ~/devel/jdk-16/bin/java --add-modules
> > jdk.incubator.vector -cp
> > target/classes/:/home/piotrek/.m2/repository/org/openjdk/jmh/jmh-core/1.27/jmh-core-1.27.jar
> > pl.tarsa.mandelbrot_simd_1 16000 > /dev/null; done
> > WARNING: Using incubator modules: jdk.incubator.vector
> > WARNING: Using incubator modules: jdk.incubator.vector
> > ... (repeated for each run)
> > WARNING: Using incubator modules: jdk.incubator.vector
> > WARNING: Using incubator modules: jdk.incubator.vector
> > real 0m37,743s
> > user 2m16,316s
> > sys 0m3,758s
> >
> > Looks like the default settings yield best performance. That's a
> > positive thing, actually.
> >
> > I'll probably send my version to benchmarks game maintainer when he
> > switches to Java 16 and then leave tuning to others.
> >
> > Thanks for the conversation,
> > Piotr
> >
>