Reported in https://discord.com/channels/188630481301012481/1097318920991559880/1221836384038551613.
Example score: https://osu.ppy.sh/scores/1855965185
The cause of the overestimation was an error in taiko's score simulator.
In lazer taiko, swell ticks don't give any score anymore, while they did
in stable.
For all intents and purposes, swell ticks can be considered "bonus"
objects that "don't give any actual bonus score". Which is to say,
during simulation of a legacy score swell ticks hit should be treated
as bonus, because if they aren't, then otherwise they will be treated
essentially as *normal hits*, meaning that they will be included in
the *accuracy* portion of score, which breaks all sorts of follow-up
assumptions:
- The accuracy portion of the best possible total score becomes
overinflated in comparison to reality, while the combo portion of
that maximum score becomes underestimated.
- Because the affected score has low accuracy, the estimated accuracy
portion of the score (as given by maximmum accuracy portion of score
times the actual numerical accuracy of the score) is also low.
- However, the next step is estimating the combo portion, which is done
by taking legacy total score, subtracting the aforementioned
estimation for accuracy total score from that, and then dividing
the result by the maximum achievable combo score on the map. Because
most of actual "combo" score from swell ticks was "moved" into the
accuracy portion due to the aforementioned error, the maximum
achievable combo score becomes so small that the estimated combo
portion exceeds 1.
Instead, this change makes it so that gains from swell ticks are treated
as "bonus", which means that they are excluded from the accuracy portion
of score and instead count into the bonus portion of score, bringing the
scores concerned more in line with expectations - although due to
pessimistic assumptions in the simulation of the swell itself,
the conversion will still overestimate total score for affected scores,
just not by *that* much.
Fixes issue that occurs on *about* 246 beatmaps and was first described
by me on discord:
https://discord.com/channels/188630481301012481/188630652340404224/1154367700378865715
and then rediscovered again during work on
https://github.com/ppy/osu/pull/26405:
https://gist.github.com/bdach/414d5289f65b0399fa8f9732245a4f7c#venenog-on-ultmate-end-by-blacky-overdose-631
It so happens that in stable, due to .NET Framework internals, float
math would be performed using x87 registers and opcodes.
.NET (Core) however uses SSE instructions on 32- and 64-bit words.
x87 registers are _80 bits_ wide. Which is notably wider than _both_
float and double. Therefore, on a significant number of beatmaps,
the rounding would not produce correct values due to insufficient
precision.
See following gist for corroboration of the above:
https://gist.github.com/bdach/dcde58d5a3607b0408faa3aa2b67bf10
Thus, to crudely - but, seemingly accurately, after checking across
all ranked maps - emulate this, use `decimal`, which is slow, but has
bigger precision than `double`. The single known exception beatmap
in whose case this results in an incorrect result is
https://osu.ppy.sh/beatmapsets/1156087#osu/2625853
which is considered an "acceptable casualty" of sorts.
Doing this requires some fooling of the compiler / runtime (see second
inline comment in new method). To corroborate that this is required,
you can try the following code snippet:
Console.WriteLine(string.Join(' ', BitConverter.GetBytes(1.3f).Select(x => x.ToString("X2"))));
Console.WriteLine(string.Join(' ', BitConverter.GetBytes(1.3).Select(x => x.ToString("X2"))));
Console.WriteLine();
decimal d1 = (decimal)1.3f;
decimal d2 = (decimal)1.3;
decimal d3 = (decimal)(double)1.3f;
Console.WriteLine(string.Join(' ', decimal.GetBits(d1).SelectMany(BitConverter.GetBytes).Select(x => x.ToString("X2"))));
Console.WriteLine(string.Join(' ', decimal.GetBits(d2).SelectMany(BitConverter.GetBytes).Select(x => x.ToString("X2"))));
Console.WriteLine(string.Join(' ', decimal.GetBits(d3).SelectMany(BitConverter.GetBytes).Select(x => x.ToString("X2"))));
which will print
66 66 A6 3F
CD CC CC CC CC CC F4 3F
0D 00 00 00 00 00 00 00 00 00 00 00 00 00 01 00
0D 00 00 00 00 00 00 00 00 00 00 00 00 00 01 00
8C 5D 89 FB 3B 76 00 00 00 00 00 00 00 00 0E 00
Note that despite `d1` being converted from a less-precise floating-
-point value than `d2`, it still is represented 100% accurately as
a decimal number.
After applying this change, recomputation of legacy scoring attributes
for *all* rulesets will be required.
Firstly, this is intended to be a float division.
Secondly, dividing integers by 0 results in an exception, but dividing
non-zero floats by 0 results in +/- infinity which will be clamped to
the upper range.
In particular, this occurs when the beatmap has 1 hitobject (0 drain
length).
