Ballistic simulations may be created too frequently

[xster]

Make a normal fling gesture.
~50% of the time, many goBallistic calls are triggered when I would have expected just one.

In this example, it took 78ms cumulative recreating the simulations and causes frame drops.

Call stack

#0      ScrollPositionWithSingleContext.goBallistic (package:flutter/src/widgets/scroll_position_with_single_context.dart:150:16)
#1      BallisticScrollActivity.applyNewDimensions (package:flutter/src/widgets/scroll_activity.dart:399:14)
#2      ScrollPosition.applyNewDimensions (package:flutter/src/widgets/scroll_position.dart:441:14)
#3      ScrollPositionWithSingleContext.applyNewDimensions (package:flutter/src/widgets/scroll_position_with_single_context.dart:110:29)
#4      ScrollPosition.applyContentDimensions (package:flutter/src/widgets/scroll_position.dart:411:7)
#5      RenderViewport.performLayout (package:flutter/src/rendering/viewport.dart:827:20)
#6      RenderObject._layoutWithoutResize (package:flutter/src/rendering/object.dart:1767:7)
#7      PipelineOwner.flushLayout (package:flutter/src/rendering/object.dart:1056:18)
#8      BindingBase&SchedulerBinding&GestureBinding&ServicesBinding&RendererBinding.drawFrame (package:flutter/src/rendering/binding.dart:251:19)
#9      BindingBase&SchedulerBinding&GestureBinding&ServicesBinding&RendererBinding&WidgetsBinding.drawFrame (package:flutter/src/widgets/binding.dart:504:22)
#10     BindingBase&SchedulerBinding&GestureBinding&ServicesBinding&RendererBinding._handlePersistentFrameCallback (package:flutter/src/rendering/binding.dart:189:5)
#11     BindingBase&SchedulerBinding._invokeFrameCallback (package:flutter/src/scheduler/binding.dart:695:15)
#12     BindingBase&SchedulerBinding.handleDrawFrame (package:flutter/src/scheduler/binding.dart:643:9)
#13     _invoke (file:///Users/xster/development/flutter/bin/cache/pkg/sky_engine/lib/ui/hooks.dart:86)
#14     _drawFrame (file:///Users/xster/development/flutter/bin/cache/pkg/sky_engine/lib/ui/hooks.dart:75)

[eseidelGoogle]

Are these used on both iOS and Android or would this cause performance degradations specific to one or the other?

[Hixie]

This happens if something within the list is changing dimensions (because we have to recompute the fling in case it would hit the edge). The question is, what is changing dimensions? And can we report it somehow (e.g. on the timeline) to make it obvious...

This will depend on exactly what you're flinging.

[Hixie]

The other thing is, goBallistic shouldn't be that expensive, yikes.

[Hixie]

Looks from the timeline like maybe this is a long list with many items of various sizes so every time we render a new item, we change our estimate of the total list length, which would trigger this also.

[Hixie]

Not sure what to do with this bug. Assuming I'm right about the cause, it's working as intended (modulo taking so long). Should we keep this open specifically to track down why it's so expensive?

[xster]

Seems like it's mostly coming from recreating new AnimationControllers.

Do you think it makes sense to let the position min/max extents be observables passed into the simulation instead of having them immutably baked in on construction?

Or perhaps create new simulations but keep the old BallisticScrollActivity and AnimationController and just update the simulation? All the math in the simulation's constructor turned out to not be that expensive it seems.

[Hixie]

We need to more precisely understand the cost. We could try doing a microbenchmark of just calling goBallistic in a tight loop and see what appears on the profile.

[knopp]

@xster, @Hixie, this is not just a performance problem. Calling goBallistic during scrolling means that the scrolling stops way sooner than it normally would. Also the momentum from multiple flings is not carried over.

Look at example from #24715 or even Licenses screen from flutter gallery. It is pretty much impossible to scroll through all licenses on iPhone (cupertino mechanics seems to be much more affected by restarting simulation than mountain view). No matter the fling speed or number of flings, scrolling stops on next page.

It seems wrong to me to recreate ballistic simulation on viewport size change then the extent is far from the final position of current simulation.