-
-
Notifications
You must be signed in to change notification settings - Fork 2.2k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Truncate start time of juice stream nested objects #28005
base: master
Are you sure you want to change the base?
Conversation
Intended as at least a partial solution to ppy#26436. In testing, two of the three replays from the issue are FCs with this change, and the third is improved (and not fully fixed). As it turns out, stable truncates the end time of sliders to integers: https://github.com/peppy/osu-stable-reference/blob/79addff0f5d2a328059d2842d6d2968cfb740956/osu!/GameplayElements/HitObjects/Osu/SliderOsu.cs#L1037 and the start time of all juice stream parts is also truncated to integers: https://github.com/peppy/osu-stable-reference/blob/79addff0f5d2a328059d2842d6d2968cfb740956/osu!/GameplayElements/HitObjects/Fruits/SliderFruits.cs#L86-L166 This matters for replay playback when mappers push limits. For instance, let's take one case from the issue. In one of the maps involved, there was a juice-stream-ending fruit at time 7093.9655 according to lazer, which was also a hyperfruit. The broken replay on this map included one frame at time 7093, with the catcher position before the hyperdash, and one frame at time 7097, with the catcher position *after* the hyperdash. Which meant that the replay handler moved *out of the way* of the hyperfruit at time 7093.9655, deciding that it is *after* the replay frame that was supposed to be catching it. (For reference, the relevant replay playback code is here: https://github.com/ppy/osu/blob/3da5831075b187e967ca300dcba0a6761f07d1dd/osu.Game.Rulesets.Catch/Replays/CatchFramedReplayInputHandler.cs#L20-L31 Note the handling of `position`. Any frame with an action is important, which means that as long as any key is held, interpolation will not take place.) On stable this is not a thing, because the fruit's end time was being truncated to `int`, therefore moving it back to time 7093 and restoring temporal integrity. This probably doesn't matter in other rulesets that much because the input tolerances in something like osu! or taiko are much higher. catch is rather knife-edge, what with mappers doing the "edge dash" / "pixel jump" stuff (tl;dr: placing a circle just barely outside of hyperdash range, so that a perfect normal dash is required to catch it). Thus, this is applied locally to catch for now until proven necessary to put it elsewhere too.
!diffcalc |
Hmmmmmm there's not that many maps impacted although a few cases I'd probably want to look at closer (highest nomod SR swing upwards is +11.8%, lowest is -33% - anything outside of ±10% I'd probably want to comb over more closely). Not sure, might be good to see at least some basic acknowledgement that from a conceptual standpoint it's fine to continue trying this change before I commit? |
!diffcalc |
I feel like I've been very acutely aware of this, at least with respect to catch's conversion, but didn't do it because it led to even more incorrect values. I feel like this probably also has an effect here but maybe that was where I noticed the issue to a greater effect: I'd be curious to see how this works combined with #25725 on the test cases in that PR. |
I did try and see if that other PR fixes the broken replays from the issue in the OP but it didn't. Didn't do much else checking beyond that, will do so whenever able. |
The answer seems to be that the two seem to be completely orthogonal / independent of each other.
Which makes sense, given the existing tolerance in conversion tests:
I wouldn't expect this fix to change things too much there, as opposed to replays, wherein very strict equality of object start time begins to matter.
I suppose truncation would have to be applied here yes. I'm not sure it matters in the hyperdashing context which this PR is intended to address primarily (can a tiny droplet be hyper?)
Well if you can remember what that was about exactly then I'd appreciate it because it may save me spending however many hours more down this hole. |
It's more that truncation being applied at somewhat of a post-processing stage led to flat out wrong values because it's expected to be applied progressively during the event calculations.
Unfortunately not. It's going to be a matter of exploration. |
RFC. PRing early, not necessarily ready to be shipped yet.
Intended as at least a partial solution to #26436. In testing, two of the three replays from the issue are FCs with this change, and the third is improved (and not fully fixed).
As it turns out, stable truncates the end time of sliders to integers and the start time of all juice stream parts is also truncated to integers.
This matters for replay playback when mappers push limits. For instance, let's take one case from the issue. In one of the maps involved, there was a juice-stream-ending fruit at time 7093.9655 according to lazer, which was also a hyperfruit.
The broken replay on this map included one frame at time 7093, with the catcher position before the hyperdash, and one frame at time 7097, with the catcher position after the hyperdash. Which meant that the replay handler moved out of the way of the hyperfruit at time 7093.9655, deciding that it is after the replay frame that was supposed to be catching it.
(For reference, the relevant replay playback code is here. Note the handling of
position
. Any frame with an action is important, which means that as long as any key is held, interpolation will not take place.)On stable this is not a thing, because the fruit's end time was being truncated to
int
, therefore moving it back to time 7093 and restoring temporal integrity.This probably doesn't matter in other rulesets that much because the input tolerances in something like osu! are much higher. catch is rather knife-edge, what with mappers doing the "edge dash" / "pixel jump" stuff (tl;dr: placing a circle just barely outside of hyperdash range, so that a perfect normal dash is required to catch it). Thus, this is applied locally to catch for now until proven necessary to put it elsewhere too.