Increase reductions for multicore usage.

[Vizvezdenec]

passed STC
https://tests.stockfishchess.org/tests/view/5e874f5a4411759d9d098696
LLR: 2.94 (-2.94,2.94) {-0.50,1.50}
Total: 9162 W: 1558 L: 1385 D: 6219
Ptnml(0-2): 90, 958, 2343, 1069, 121
passed LTC
https://tests.stockfishchess.org/tests/view/5e8762804411759d9d09869f
LLR: 2.94 (-2.94,2.94) {0.25,1.75}
Total: 79364 W: 9541 L: 9159 D: 60664
Ptnml(0-2): 462, 6880, 24661, 7172, 507
This patch strenghtens effect of reduction increase with core count twice.
Most logical way will be to try to strenghten it even more.
bench 4742598

[vondele]

Thanks!

[xoto10]

Is it possible the increased reduction will go too far on very powerful like at TCEC and decrease strength?
Not sure how we could test this, but just raising the possibility.

[vondele]

I did a test today on 176 threads, 30+0.3TC, and it looks OK:

Score of master vs masterprev: 7 - 5 - 132  [0.507] 144
Elo difference: 4.8 +/- 16.4, LOS: 71.8 %, DrawRatio: 91.7 %

[xoto10]

ok. Obviously some error bounds there, but it doesn't break completely - good! :)

[Vizvezdenec]

Well it wouldn't break anything completely since it's std::log.
So for 8 cores it increases from 25.8 to 26.8 and for 30 cores it's from 26,5 to 28,2 - it's not that massive.
Also original patch was showing better results for 31 core than for 8 cores IIRC.

[vondele]

yes, I would expect this to be more beneficial at higher thread count.

[xoto10]

Sure. I did a spreadsheet of the numbers a day or two ago. I was thinking that bigger numbers can't always be better, and high depth and high thread count is presumably where the problems (if there are any) show up first.
Anyway, as Viz says, it's a log, so the increase gets less and less.
Nice one :)