Memory leak in Apple Devices (MacOS, iOS)

[saif-ellafi]

hello,

context: I start and dispose threeJs each time I call an animation:

      stateInitializeStarted = true;
      isReinitializing = true;
      disposeThreeJs();
      threeJs = null;
      three.loading.clear();
      isSetup = false;
      hasRenderedOnce = false;
      
      var completer = Completer<void>();

      threeJs = three.ThreeJS(
        size: renderWidth != null && renderHeight != null ? Size(renderWidth!.toDouble(), renderHeight!.toDouble()) : null,
        onSetupComplete: () { 
          if (!completer.isCompleted) {
            completer.complete(); 
          }
          setStateSafe(() { isSetup = true; }); 
        },
        setup: setup,
        settings: three.Settings(
          alpha: true,
          clearAlpha: 0.0,
          clearColor: 0x000000,
          animate: false,
        )
      );

And when finishing:

      isRollSfxCancelled = true;
      await _fadeController.reverse();

      _stopAnimateTicker();

      Dice3DView._cleanupCompleter?.complete();
      Dice3DView._cleanupCompleter = null;
      Dice3DView.cleanupFuture = null;

      Dice3DLoader.stopAllSfx();

      resultText = null;
      clearDice();
      _stopFixedUpdateTimer();

      _disposeThreeJsAfterRoll();

where

  void _disposeThreeJsAfterRoll() {
    disposeThreeJs(); threeJs = null;
    world = null;
    isSetup = false;
    hasRenderedOnce = false;
    stateInitializeStarted = false;
    isReinitializing = false;
    setStateSafe(() {});
  }

each time I roll dice in my app, with version 0.2.0, memory increases really badly:

The same happens in iOS

Does not seem to happen on Windows at all, staying stable over there. Always rolling the same 2d8 dice models:

[saif-ellafi]

Previous version (0.1.7) - seems the same (debugMode)

release mode:

..

[Knightro63]

Hi @saif-ellafi,

All apple products share the same darwin for angle, it is probably a flutter_angle issue. It must not be deleting the IOSurfaceRef. I changed it from unmanaged to managed texture for swift, which did fix the deposing of angles class and the resizing memory leak and crashing, more than likely the garbage collector is not collecting it if you don't fully close and go out of the route.

I will have to see if there is another way to force garbage collection on the disposed texture.

[saif-ellafi]

In my debug tests this is massive as after about 10 dice rolls it crashes everything of course. I wonder if in release mode and other devices things are not so grim...

Alternatively if there is no way to find a solution, I could roll back to leaving three_js open in the background at all times, at the cost of CPU/GPU/Battery Drain on idle

[saif-ellafi]

In general, compared to 0.1.7 I feel that stopping (dipose) and restarting three_js (init) is much slower and heavier now, could I be right or would we need numbers? It somehow continues to feel better in my benchmarks to stop and start the engine on each roll.

[Knightro63]

Hi @saif-ellafi ,

From 0.1.7 to 0.2.0 there is not much difference.

After some testing today on my Mac in debug. I was able to open and close a widget that had threejs in it about 100 times with about 20mb in crease overtime. It does have an initial increase of about 400mb when starting. Then it reduces about 220 when dispose.

E.g.

Start app: 227
Init threejs: 650
Dispose: 420
Init: 650
Dispose:421
Init: 651

And so on. It seems to keep classes that it should have released. Until I get out of that route all together. Then it goes down to about 331 still holding on to something.

[saif-ellafi]

Hey that's not bad, then gives me some food for thought into figuring out whether I an loading textures in the wrong way, or what could I be doing wrong! Will put some hours on it after the first release these days.

[saif-ellafi]

Been playing around the latest version this morning @Knightro63 - MacOS looks good and memory not going up anymore. Any magical ninja releases with the fix? Have not yet tried iOS, but will do in a moment

[Knightro63]

Hi @saif-ellafi,

I removed most of the global parameters that probably were clogging up the memory. I am still hunting for more to see if I can make it better.