Scala performance fix: Use Double instead of Float to avoid autoboxing

[witi83]

Your code gets twice as fast when you use Double instead of Float. I just replaced all occurrences via search+replace. In my case:

• Before: Rendering took: 2517.6652006ms (average over 5 repetitions)
• After: Rendering took: 1254.6068284ms (average over 5 repetitions)

Among other things, autoboxing occurs when using tuples with Float. Only Doubles are spezialized: https://github.com/scala/scala/blob/v2.13.1/src/library/scala/Tuple2.scala#L24

Besides, your precision rises from 32 bit (Float) to 64 bit (Double), which results in a nicer output image.

[ForNeVeR]

Fun fact is that we were experiencing 25% speed up in F# when switching from float64 to float32.

All of this is getting really messy now.

[athas]

Thanks! The ways of parallel functional programming are mysterious indeed.

[olafurpg]

It doesn't seem to make a difference whether you use Float or Double when running the benchmarks on GraalVM 20.0.0, both are similarly optimized

# With Float (current master after reverting this PR)
[info] Benchmark        (scene)  Mode  Cnt    Score     Error  Units
[info] Bench.construct   rgbbox  avgt    3    0.200 ±   0.067  ms/op
[info] Bench.construct    irreg  avgt    3    5.558 ±   2.903  ms/op
[info] Bench.render      rgbbox  avgt    3  766.837 ± 171.120  ms/op
[info] Bench.render       irreg  avgt    3  268.327 ±  59.585  ms/op

# With Double (current master, includes this PR)
[info] Benchmark        (scene)  Mode  Cnt    Score     Error  Units
[info] Bench.construct   rgbbox  avgt    3    0.211 ±   0.074  ms/op
[info] Bench.construct    irreg  avgt    3    5.878 ±   5.600  ms/op
[info] Bench.render      rgbbox  avgt    3  805.039 ± 152.805  ms/op
[info] Bench.render       irreg  avgt    3  278.762 ± 156.538  ms/op

[witi83]

@olafurpg I'm not familiar with GraalVM and how it handles Boxing.

But using the default JVM, the question here is not Double or Float, the problem lies in Tuple2+Float. Just take a look at the implementation: https://github.com/scala/scala/blob/v2.13.1/src/library/scala/Tuple2.scala#L24 Since Float is not defined, its gets a massive performance penalty due to autoboxing. More information why Float is not defined: https://scalac.io/specialized-generics-object-instantiation/

But yeah, replacing Float with Double is just a quick fix. Rather than this you should refactor the code accordingly and avoid using Tuple+Float