Scala perf optimizations

[isaacl]

Major improvement for Clock encode, some small speed up for move encoding as well.

• Remove Array.tabulate and .foreach in hot paths as they require Autoboxing.
• Adjustments based on javap source inspection and JMH benchmarks.

                Java    Master   This
Huffman.Encode  8.57ms  8.68ms   8.27ms
Huffman.Decode  7.37ms  5.64ms   5.62ms
Clock.Encode    185ns   312ns    192ns
Clock.Decode    206ns   231ns    220ns

[ornicar]

I'm all for optimizing hot paths, but that's not one. It's invoked only once per game. So I think we should use the idiomatic isEmpty, which btw is supposed to be inlined:

@`inline` def isEmpty: Boolean = xs.length == 0

[isaacl]

Once per ply

[ornicar]

This is not a hot path, it runs once per game. I don't think we should be re-implementing scala's foreach where it can't possibly make a difference.

[isaacl]

Once per ply -- and yes, it does make a difference in the OverallEncodingTest.testEncode bench. (260->190ns/op)

[ornicar]

b733ef4 sbt 'benchmarks/jmh:run -i 5 -wi 3 -f1 -t1 org.lichess.compression.benchmark.*'

[info] Benchmark                       Mode  Cnt     Score    Error  Units
[info] BitOpsTest.testRead             avgt    5   106.057 ±  0.401  ns/op
[info] HuffmanPgnBench.decode          avgt    5  4176.123 ± 40.581  us/op
[info] HuffmanPgnBench.encode          avgt    5  7065.805 ± 28.124  us/op
[info] LinearEstimateTest.testEncode   avgt    5    32.803 ±  0.167  ns/op
[info] LowBitTruncTest.testEncode      avgt    5     3.645 ±  0.011  ns/op
[info] OverallEncodingTest.testDecode  avgt    5   129.727 ±  1.060  ns/op
[info] OverallEncodingTest.testEncode  avgt    5   159.204 ±  3.736  ns/op
[info] VarIntEncodingTest.testDecode   avgt    5    83.967 ±  0.284  ns/op
[info] VarIntEncodingTest.testEncode   avgt    5   125.071 ±  7.903  ns/op

[ornicar]

ffe3e19 sbt 'benchmarks/jmh:run -i 5 -wi 3 -f1 -t1 org.lichess.compression.benchmark.OverallEncodingTest'

[info] Benchmark                       Mode  Cnt    Score   Error  Units
[info] OverallEncodingTest.testDecode  avgt    5  130.638 ± 3.088  ns/op
[info] OverallEncodingTest.testEncode  avgt    5  157.204 ± 2.064  ns/op

as expected within statistical error

[ornicar]

Also I'm rather confused because I just tried benchmarking master 877666f and got this:

877666f sbt 'benchmarks/jmh:run -i 5 -wi 3 -f1 -t1 org.lichess.compression.benchmark.OverallEncodingTest'

[info] Benchmark                       Mode  Cnt    Score   Error  Units
[info] OverallEncodingTest.testDecode  avgt    5  128.926 ± 2.996  ns/op
[info] OverallEncodingTest.testEncode  avgt    5  141.193 ± 1.757  ns/op

Showing similar/better perf than this PR.

Using openjdk 21.0.9 / Linux 6.17.2.

I'm sure this PR should be better, I just don't understand how it can't be measured on my machine.

[isaacl]

That's odd -- I get consistently better benchmarks on my machine, most of the changes here I benchmarked individually, with the exception of a couple foreach array replacements that I did in one go.

I did see a separate odd quirk in jmh, where Huffman.encode sometimes runs with 8.5ms/run and sometimes with 5.6ms/run and restarting the run seems to flip a coin of which one it's going to be.

[ornicar]

since scala3 we have an inline keyword that guarantees inlining https://docs.scala-lang.org/scala3/guides/macros/inline.html#inline-methods

[ornicar]

https://monitor.lichess.ovh/d/nAXOnXrWz/game-pgn-encoder?orgId=1&from=now-7d&to=now&timezone=utc&var-instance=lila@ocean&var-field=mean&refresh=5m&viewPanel=panel-2

[isaacl]

@ornicar will investigate. Definitely not expected.