reevaluate on modern CPUs/JDKs

[kno10]

It would be great to reevaluate on modern CPUs and JDKs which of these functions are still beneficial to use.

[kno10]

I did some JMH benchmarks, naive (value range 0 to 1), with -Djafama.fastlog=true -Djafama.fastsqrt=true:

Benchmark              (size)  Mode  Cnt    Score   Error  Units
JaFaMa.atan_commons   1000000  avgt   20   50,885 ± 0,786  ns/op
JaFaMa.atan_jafama    1000000  avgt   20    9,572 ± 0,133  ns/op
JaFaMa.atan_java      1000000  avgt   20   46,685 ± 0,344  ns/op
JaFaMa.atan_strict    1000000  avgt   20   46,682 ± 0,224  ns/op
JaFaMa.ceil_commons   1000000  avgt   20    6,641 ± 0,077  ns/op
JaFaMa.ceil_jafama    1000000  avgt   20    5,523 ± 0,022  ns/op
JaFaMa.ceil_java      1000000  avgt   20    4,384 ± 0,004  ns/op
JaFaMa.ceil_strict    1000000  avgt   20    4,478 ± 0,054  ns/op
JaFaMa.cos_commons    1000000  avgt   20   31,059 ± 0,188  ns/op
JaFaMa.cos_jafama     1000000  avgt   20   11,047 ± 0,217  ns/op
JaFaMa.cos_java       1000000  avgt   20   29,382 ± 0,079  ns/op
JaFaMa.cos_strict     1000000  avgt   20   42,947 ± 0,573  ns/op
JaFaMa.exp_commons    1000000  avgt   20   23,770 ± 0,158  ns/op
JaFaMa.exp_jafama     1000000  avgt   20   12,596 ± 0,042  ns/op
JaFaMa.exp_java       1000000  avgt   20   30,689 ± 0,242  ns/op
JaFaMa.exp_strict     1000000  avgt   20   24,657 ± 0,168  ns/op
JaFaMa.log1p_commons  1000000  avgt   20  142,297 ± 2,009  ns/op
JaFaMa.log1p_jafama   1000000  avgt   20   14,763 ± 0,097  ns/op
JaFaMa.log1p_java     1000000  avgt   20   45,126 ± 0,086  ns/op
JaFaMa.log1p_strict   1000000  avgt   20   45,091 ± 0,179  ns/op
JaFaMa.log_commons    1000000  avgt   20   43,293 ± 0,086  ns/op
JaFaMa.log_jafama     1000000  avgt   20   30,487 ± 0,343  ns/op
JaFaMa.log_java       1000000  avgt   20   31,480 ± 0,215  ns/op
JaFaMa.log_strict     1000000  avgt   20   39,382 ± 0,406  ns/op
JaFaMa.round_commons  1000000  avgt   20  5,042 ± 0,026  ns/op
JaFaMa.round_jafama   1000000  avgt   20  5,711 ± 0,064  ns/op
JaFaMa.round_java     1000000  avgt   20  5,334 ± 0,009  ns/op
JaFaMa.round_strict   1000000  avgt   20  5,347 ± 0,020  ns/op
JaFaMa.sin_commons    1000000  avgt   20   23,829 ± 0,147  ns/op
JaFaMa.sin_jafama     1000000  avgt   20   10,737 ± 0,077  ns/op
JaFaMa.sin_java       1000000  avgt   20   30,734 ± 0,091  ns/op
JaFaMa.sin_strict     1000000  avgt   20   36,735 ± 0,082  ns/op
JaFaMa.sqrt_jafama    1000000  avgt   20    3,304 ± 0,064  ns/op
JaFaMa.sqrt_java      1000000  avgt   20    3,282 ± 0,012  ns/op
JaFaMa.sqrt_strict    1000000  avgt   20    3,306 ± 0,060  ns/op
JaFaMa.tan_commons    1000000  avgt   20   54,395 ± 0,188  ns/op
JaFaMa.tan_jafama     1000000  avgt   20   10,073 ± 0,054  ns/op
JaFaMa.tan_java       1000000  avgt   20   40,770 ± 0,069  ns/op
JaFaMa.tan_strict     1000000  avgt   20   48,009 ± 0,120  ns/op

Seems as if JaFaMa still is beneficial, except for log, ceil, sqrt

[jeffhain]

Hello. Sorry for the late reply, I don't check this place often.

Good to see that it's still beneficial for most methods ;) log and sqrt often delegate to hardware so it's hard to beat that, and ceil/floor are simple enough to be naturally fast in the JDK.

One thing I observed got slower with JDK versions from 5 to 7 is the time it takes to initialize the tables (as I say in the readme), but I didn't check if that changed with newer JDKs.

I don't plan on keeping up-to-date benchmarks, unless I make a new version. in which case I might do some related updates.

I have a few non-committed upgrades for some "quick" methods (better constants), and could add some methods that were added to JDK Math, but I don't feel like that would be worth a new version (not releasing too often makes it easier for projects to harmonize their dependencies).

Though, I've been playing around with BigDecimal-like code for exploring some fractals, but mutable (for pooling, which can help a lot), based on binary rather than decimal (makes multiplications and divisions by powers of two much faster), and using parallelism (or not). I might add a "big" package to jafama with these "fast" "big" classes, plus a related FastBigMath class with various kinds of functions: that would be worth a new version.