I'm trying to implement the animation of a custom property on a CALayer. I've set up the property as dynamic and it mostly works. In my app, I might have a dozen or more instances of this layer, all of which should animate at the same time. However, sometimes when the animation is triggered, a few of the layers don't perform the animation. They simple jump directly to the final value.
In trying to debug the problem, it looks like when the animations start, the layers with the problem get a redraw on their model layer with the final value of the animated property. The rest of the layers only get redraws on their presentation layers with the transient values of the animated property. It's usually the last of the layers that has the problem, but sometimes it's random layers.
It's so sporadic, I'm not really sure where to look next.
Here's the snippet of code that triggers the animation:
#dynamic aspectModeFraction;
- (id<CAAction>)actionForKey:(NSString *)key
{
if ([key isEqualToString:#"aspectModeFraction"])
{
CABasicAnimation *animation = [CABasicAnimation animationWithKeyPath:key];
animation.timingFunction = [CAMediaTimingFunction functionWithName:kCAMediaTimingFunctionEaseOut];
animation.fromValue = #([[self presentationLayer] aspectModeFraction]);
animation.duration = 2.0f;
animation.delegate = self;
return animation;
}
return [super actionForKey:key];
}
+ (BOOL)needsDisplayForKey:(NSString *)key {
if ([#"fullRect" isEqualToString:key]) { return YES; }
if ([#"localRect" isEqualToString:key]) { return YES; }
if ([#"screens" isEqualToString:key]) { return YES; }
if ([#"aspectModeFraction" isEqualToString:key]) { return YES; }
return [super needsDisplayForKey:key];
}
It looks like you have most of the parts. There are lots of little parts in custom property animation, and this talk may still be helpful, it it kind of looks like you probably have the major things in place.
I'd first suspect the calling code. I'd make sure you're not suppressing animations for instance, or modifying any UIKit pieces off the main thread.
It's worth noting that the trick of using #dynamic to get your layer properties copied may not be officially supported. (It's one of those things that "seems to work" rather than coming out of any doc I know; and when I've spoken with some Apple engineers about it, they'd never heard of doing it that way.) If you're seeing trouble only on 10.9+, you might try replacing your use of #dynamic with an standard property that you copy over in initWithLayer:.
Thanks for the suggestions. I eventually tracked down the problem. I had optimized the drawing code to only do some calculations when this parameter changed. I was marking the calculations as dirty when the parameter was updated so they would be recalculated during the next draw loop. The dirty flag was set via a custom setter on the parameter. When I switched to #dynamic (which seems to be required for CAAnimations?) the custom setter had to go away, so I was updating the dirty flag by having the object do KVO on itself. It turns out that observeValueForKeyPath:... wasn't being called during the animation. It got one update for the final value, but no intermediates.
Anyway, dropping the optimization and doing the full calculations each time resolved the issue. KVO on the animated value not working as expected seemed to be the culprit.
Related
If I create a blank Mac XCode project and layout 500 simple NSView objects side by side in the main window it loads pretty damn fast. If I set wantsLayer=YES on each subview, performance dramatically drops, by several seconds. Why is this the case conceptually? It seems that layers would be faster not slower than regular old NSViews.
You're giving the system more work to do by layer-backing so many views. Layer-backing allows graphic acceleration (for drawing) but it adds a bit of overhead to things like layout, not to mention just creating them and putting them on screen. If used properly, it's not really much of a problem.
Typically, if you had so many "things" to manage on screen at once, you'd have one layer-backed hosting view that manages its own tree of sublayers. "But what about view-based table views?" you ask. Trickery, trickery, I say! Table views don't actually keep all the cell views they manage around; they efficiently reuse them, keeping around only enough to represent what's on screen and/or animating around.
So I'd say this isn't really a problem since it's not a particularly good approach to throw 500+ layer-backed views up for layout and drawing to begin with. :-)
as of 2021, Joshua Nozzi's answer is just the half of the truth.
When you want to utilise that much layers, you should make use of the power of CALayers with Sublayers instead of using NSViews over and over again, each with its own NSCell and possible CALayer backed as you forced it to with -wantsLayer:.
There is nothing wrong with 500 sublayers. Sublayers allow you to be fast when properly structured.
If you want to speed up even more, turn off autoanimation by force on each Layer that does not need it. Yes a lot examples out there show that you can turn off Autoanimation each time you call some property to be changed, slowing down your drawing even more, because it involves even more object messaging.
Keep in mind the following example turns off Auto-animated Keys by their Key name and kicks out the whole action and does not apply onto all Sublayers on purpose. If your structure of Sublayers of Sublayers need it you could call this on according ParentLayer to wipe out the CPU eating Autoanimation. Its much better to let the autoanimation do its job on the layers you actually want it to. Which turns around Apples default paradigm to animate everything. If most action keys are turned off be aware that your layers do show up kind of like a state machine. In the example here #"frame" animations are not wiped out..
void turnOffAutoAnimationsForLayerAndSublayers(CALayer* layer) {
NSNull *no = [NSNull null]; //(id)kCFNull
NSDictionary *dict = [NSDictionary dictionaryWithObjects:#[no,no,no,no,no,no,no,no]
forKeys:#[#"bounds",#"position",#"contents",#"hidden",#"backgroundColor",#"foregroundColor",#"borderWidth",#"borderColor"]];
for (CALayer *layertochange in layer.sublayers ) {
if (layertochange.sublayers.count) {
for (CALayer *sublayertochange in layertochange.sublayers) {
sublayertochange.actions = dict;
}
}
layertochange.actions = dict;
}
}
and use it in your CALayer backed NSView -init like..
CALayer *layer = [CALayer layer];
layer.frame = self.frame;
//...backgroundColor, contents, etc etc..
// here you could even do a for loop adding as much sublayers you want..
for (int y=0; y<100; y++) {
CALayer *sub = [CALayer layer];
//...frame, backgroundColor, contents, of sublayers etc etc..
layer.frame = CGRectMake(0, 0, 10, y*10);
[layer addSublayer:sub];
}
// structure is ready, add it to the base layer
self.layer = layer;
// now tell NSView you really want to make use this layer structure
self.wantsLayer = YES;
// turn off AutoAnimations
turnOffAutoAnimationsForLayerAndSublayers(self.layer);
For layers that are basically all the same there is even a specialised Subclass called CAReplicatorLayer. Allowing you to add even more with less internal draw calls.
You should also know that layers that are hidden are not calculated at all (meaning drawing here). You can still change properties even on hidden Layers. Unhide them, when needed, not earlier. So in example a custom CATextLayer that will change its string property to lets say #"", aka nothing, will still be drawn. But if you Subclass CATextLayer and change its string implementation to lets say..
#interface YourTextAutoHideLayer : CATextLayer
-(void)setString:(id _Nullable )string;
#end
#implementation YourTextAutoHideLayer
-(void)setString:(id _Nullable)string {
self.hidden = [string isEqual:#""];
super.string = string;
}
#end
you speed up drawing even more. A) because less object messaging to hide the layer that has no text anyway and B) once hidden it is not part of the internal draw calls effective speeding up your main CALayers drawing.
On NSTableViews that are NSCell based you usually never use CALayers. There is no need to, NSTableViews manage the visible Cells on purpose to keep in example scrolling smoothly. And still they (NSTableViewCells) have a reuse-mechanism to speed things up. You will very likely not re-invent a tableview with CALayers anyway. But in case you do reinvent, there is a CAScrollLayer class too.
And if that is not enough speed it is worth thinking about a MetalView utilising the power of GPUs.
Edit: code your CALayers not in -drawRect:. -drawRect: is called anytime something changes, in example frame/position on screen or bounds etc.. so try to avoid coding CALayers in there. You can set [self setNeedsLayout:YES]; & [self setNeedsDisplay:YES]; on purpose for a good reason, the reason is you want to avoid too much drawing for basically no change at all. -drawRect: in example was/is such method that is supposed to handle a backing with its own context to draw in. As CALayers have their very own mechanism you can let the drawRect block blank, doing nothing in there, maybe even erasing the method at all if you dont need it.
CALayers are not part of the auto layout system unless you code in -layout.. so again, keep CALayer drawing outside such and you are good. and pssst a lot CALayers properties can be changed even outside the main thread, some definitely not like layer.string.
Any ideas why a screen saver using just a plain ScreenSaverView subclass with a CAEmitterLayer sublayer would render fine on the primary screen and choppy (as if every 2. frame renders there..) on the secondary screen..?
This is my initialization code:
- (id)initWithFrame:(NSRect)frame isPreview:(BOOL)isPreview
{
self = [super initWithFrame:frame isPreview:isPreview];
if (self)
{
CAEmitterLayer* emitterLayer = [MyEmitterFactory emitterLayer:self];
[self setWantsLayer:YES];
[self.layer addSublayer:emitterLayer];
[self setAnimationTimeInterval:1/2.0];
}
return self;
}
Everything else in this subclass is default (as provided by the Xcode template).
Funny enough, backingStoreType does sound like a good candidate to tweak in a ScreenSaverView subclass using CoreAnimation, alas all other modes except the default one are not to be used as per the docs..
(As the animation is powered by Core Animation it doesn't really matter what I put in setAnimationTimeInterval - or remove the call completely, as experiments have shown)
According to the documentation of NSView setWantsLayer:
To create a layer-hosting view, you must call setLayer: and supply your layer object before you call the setWantsLayer: method; the order of these method calls is crucial.
Furthermore: Which OS version is it? Does the choppiness also come up when the two displays are mirrored (or vice versa)?
I have a Cocoa app with ARC enabled. I am using OS 10.8 and Xcode 4.4. The app sporadically crashes upon calls to CGLayerRelease, dumping messages like the following to console:
error for object 0x10a2d2000: entry for pointer being freed from death-row vanished
If I comment out the calls to CGLayerRelease the crashing stops. Do I need to call CGLayerRelease in ARC enabled projects? The documentation isn't very clear about this.
It is hard to post code for this because it is spread throughout a large file for a class that inherits from NSView. The class has a data member CGLayerRef layer;. Then inside the drawRect function:
if (layer == nil) {
layer = CGLayerCreateWithContext(context, self.frame.size, NULL);
// draw some stuff into CGLayerGetContext(layer)
}
CGContextDrawLayerInRect(context, self.bounds, layer);
The class is also a delegate of its window and has the following function
- (void)windowDidEndLiveResize:(NSNotification *)notification {
if (layer) {
//CGLayerRelease(layer);
layer = nil;
}
[self setNeedsDisplay:YES];
}
Also inside a property setter
- (void)setPitch:(NSArray *)aPitch {
pitch = aPitch;
if (layer) {
//CGLayerRelease(layer);
layer = nil;
}
}
Now if I uncomment the calls to CGLayerRelease then I sporadically get the crash mentioned above. I stress "sporadically" because it does not always happen.
Yes, you need to call CGLayerRelease() in ARC. The above assignments look correct, but I suspect you may have others. You're directly accessing your ivars, which is the #1 cause of these kinds of crashes. Do not access your own ivars except in init and dealloc. Create an accessor for setLayer: that correctly releases the old one and creates a new one and always use that.
It's very possible that your CGLayer code is fine, and that it's something else like a CGPath that you're adding to the CGLayer. Move all your ivars to accessors. Otherwise you will be chasing these kinds of bugs forever.
OK, so Apple brought ARC to us, which is great. After refactoring my Application to ARC almost everything works fine and it is a lot easier now to develop and maintain.
There is just one problem I still can't figure out.
My job management program shows different detail information of proposals, orders and so on in their own windows. So I have a special class where WindowControllers gets allocated and initiated with initWithWindowNibName and then the window is shown with showWindow:
DetailWindowController *proposalWindowController = [[DetailWindowController alloc] initWithWindowNibName:#"ThePorposalWindow"];
[proposalWindowController showWindow:nil];
Before ARC the Instance of the WindowController did the release like shown in the documentation:
- (void)windowWillClose:(NSNotification *)notification
{
[self autorelease];
}
But now with ARC this is not possible anymore and what makes it even worse, in my special class where the WindowController is allocated and initiated, the same windowController is released by ARC because there is no pointer to the windowController.
My idea was to copy the windowController into an mutuable array:
[proposalWindowArray addObject:proposalWindowController];
[[proposalWindowArray lastObject] showWindow:nil];
And in the windowControllers delegate method windowWillClose I post a notification to my special class:
- (void)windowWillClose:(NSNotification *)notification
{
[[NSNotificationCenter defaultCenter] postNotificationName:#"ProposalWindowWillClose" object:[[self window] windowController] userInfo:nil];
}
In my special class I listen to the notification and remove the object from the array:
- (void) proposalWindowWasClosed: (NSNotification *) notification
{
[proposalWindowArray removeObjectIdenticalTo:[notification object]];
}
It works, but I still do not believe that this is the correct way.
Does anybody has the same problem or a tip to make it better?
I'd probably use a delegate approach rather than notifications. Generally it is better to have an external object that keeps track of the open windows. Self-retaining objects, like your old system, break the basic points of object ownership and make it hard to find things (such as "give me a list of open windows"). Non-Singletons that are just "floating" out there often come back to bite you in your architecture (I've had to fix this often enough).
That said, sometimes self-ownership is at least convenient, and at worst not-the-end-of-the-world. So self-own. The only difference is that you need to do it explicitly rather than matching a leak and an over-release (which is what your old code was doing).
Create a private strong property. Assign self to it. That will create a retain loop that will keep you around until you set the property to nil.
I think your alternative approach should be correct, but I don't think you need the second notification. You should be able to do:
- (void)windowWillClose:(NSNotification *)notification
{
[proposalWindowArray removeObjectIdenticalTo:self];
}
Assuming the "proposalWindowArray" is a static NSMutableArray.
Without hacks, there is no elegant way to keep an object retained other than having a strong reference to it in some other object. For example, you could keep a static NSMutableArray/NSMutableSet, add your controller there, and remove it in windowsWillClose:. This will be shorter than posting a notification. To make this reusable, create a WindowControllerRegistry singleton with an array, where you add controllers like this one, and which will automatically listen to NSWindowWillCloseNotification and remove them from its array thus releasing ownership.
As a quick workaround, you could perform retain/autorelease calls from non-ARC file:
my_retain(self);
my_autorelease(self);
// ArcDisabled.mm
void my_retain(id obj) { [obj retain]; }
void my_autorelease(id obj) { [obj autorelease]; }
I had this same issue when I switched to ARC. Your solution works, but you're making it too complicated. You can essentially do what you were doing before by having the window release itself when it closes, but in an ARC compatible manner.
The solution is to simply create a property of your class within the class itself. For your example, in DetailWindowController, you would add the following property:
#property (strong) DetailWindowController *theWindowController;
Then when you create the window with your code above, add one line like so:
DetailWindowController *proposalWindowController = [[DetailWindowController alloc] initWithWindowNibName:#"ThePorposalWindow"];
[preferenceController setTheWindowController:proposalWindowController];
[proposalWindowController showWindow:nil];
Then finally, to have ARC release the window when it is closed like you did with the autorelease pre-ARC, in the DetailWindowController class, simply do:
- (void)windowWillClose:(NSNotification *)notification
{
// Let ARC tear this down and clean it up
[self setTheWindowController:nil];
}
I have a UI where the content of an NSCollectionViewItem's View is drawn programmatically through CALayers. I am using a CAConstraintLayoutManager to keep the layot of the sublayers consistent when resizing, but I am getting very poor performance when doing so. It seems that resizing the window, which causes the resize of two CATextLayers so that they fit the root layer's width, and the repositioning of one CATextLayer so that it stays right-aligned, is causing the application to spend most of its time executing the CGSScanConvolveAndIntegrateRGB function (I have used the Time Profiler instrument).
The most "expensive" layer (the one that causes the most stuttering even if it's the only one displayed) is a wrapped multiline CATextLayer. I have absolutely no idea how to get better performance (I have tried not using a CAConstraintLayoutManager and going with layer alignments but I'm getting the same thing). Has anyone had this problem? Is there a way around it?
PS: I have subclassed the layout manager and disabled all the animations during the execution of - (void)layoutSublayersOfLayer:(CALayer *)layer by setting YES to kCATransactionDisableActions in the CATransaction but it doesn't seems to help.
Edit: I have disabled Font Smoothing for the Text Layers and performance has increased a little bit (very little), but it spends an awful amount of time in _ZL9view_drawP7_CAViewdPK11CVTimeStampb (which is something that gets called by a thread of the ATI Radeon driver, I suppose).
I solved it. Kind of. It still seems like a dirty hack to me, but I couldn't find out how to make setNeedsDisplayInRect work so I ended up doing it like this:
In the NSWindow delegate:
-(void)windowWillStartLiveResize:(NSNotification *)notification
{
[[NSNotificationCenter defaultCenter] postNotificationName:#"beginResize" object:nil];
}
-(void)windowDidEndLiveResize:(NSNotification *)notification
{
[[NSNotificationCenter defaultCenter] postNotificationName:#"endResize" object:nil];
}
In my Custom View those two notifications call, respectively, the -(void)beginResize and -(void)endResize selectors. The first one sets a BOOL inLiveResize variable to YES, while the second one sets it to NO and calls setFrameSize again with the new frame size.
I overrode (overridden? Not native english speaker, sorry) the -(void)setFrameSize:(NSSize)newSize method like this:
-(void)setFrameSize:(NSSize)newSize
{
if (inLiveResize) {
NSRect scrollFrame = [[[self superview] enclosingScrollView] documentVisibleRect];
BOOL condition1 = (self.frame.origin.y > (scrollFrame.origin.y - self.frame.size.height));
BOOL condition2 = (self.frame.origin.y < (scrollFrame.origin.y + scrollFrame.size.height + self.frame.size.height));
if (condition1 && condition2)
[super setFrameSize:newSize];
}
else {
[super setFrameSize:newSize]; }}
That's it. This way, only the visible views resize live with the window, while the others get redrawn at the end of the operation. It works, but I don't like how 'dirty' it is, I'm sure there is a more elegant, built-in(ish) way to do this by using the setNeedsDisplayInRect method. I will research more.