Advances in AVFoundation Playback

Session 503 WWDC 2016

AVFoundation is a powerful framework for media operations, providing capture, editing, playback, and export. Learn about new APIs and methods for media playback. Achieve gapless playback transitions between assets, create seamless loops, simplify your playback logic with "autowait", and see how to deliver an even faster playback startup experience.

[ Music ]

Good morning.

[ Applause ]

Welcome to our session on Advances in AVFoundation Playback.

My name is Sam Bushell.

Today we're going to talk about some new enhancements that we've added to try and smooth over some rough edges that some developers have found challenging.

So AVFoundation provides APIs for a very broad selection of multimedia activities, including playback, capture, export, and many kinds of editing.

I'll be focusing mostly on playback.

AVFoundation supports playback from a very wide selection of media formats from local storage.

And in most cases you can take the same file, and you can put it on a web server and then AVFoundation can play that over the network.

The file format in this case is the same, but the IO is over the network.

We call this progressive download playback.

Once we start downloading that file, even if the network characteristics change, we will continue with the same file.

HTTP Live Streaming is more dynamic.

Generally, the base URL refers to a master playlist which introduces multiple playlists for the same content but varying in bit rate and format and maybe in language.

And each of these playlists references segments containing the actual compressed media.

So let's talk about what we're going to talk about today.

We're going to discuss the playback changes to do with the pre-playback buffering period.

We're going to introduce a new API to simplify looping playback of a single file.

We're going to discuss some playback refinements we've made under the hood.

We're going to discuss getting your application ready for wide color video.

And then we'll spend the rest of our time discussing a popular topic optimization of static time in playback apps.

Let's start by waiting for the network.

Because when we play media playback over the Internet, we're at the mercy of the network.

We don't want to start too soon or playback my stall.

We don't want to start too late or the user may give up on us.

We want to start at that Goldilocks moment and start playback when we have enough data that we'll be able to play consistently and not stall.

Here is the existing API.

AVPlayerItem provides three Boolean properties.

playbackLikelyToKeepUp, playbackBufferFull, and playbackBufferEmpty.

playbackBuffer sorry playbackLikelyToKeepUp is true if AVFoundation's algorithm believes that if you were to stop playing now, you could keep on playing without stalling until you got to the end.

playbackBufferFull is true if the buffer does as much as it's going to.

So if you haven't started playing back yet, you might as well.

playbackBufferEmpty means that you are stalling or you're about to stall.

So for progressive download playback in iOS 9 and earlier, AVFoundation clients must monitor these properties themselves and wait until playbackLikelyToKeepUp is true or playbackBufferFull is true before setting the AVPlayer's rate property to 1.

For HTTP Live Streaming, the rules are simpler.

You can set AVPlayer's rate property to 1 as soon as the user chooses to play, and it will automatically wait to buffer sufficient media before playback begins.

We are streamlining the default API contract in the 2016 iOS releases.

iOS, Mac OS, tvOS.

For apps linked on or after iOS 10, Mac OS Sierra, tvOS 10, the same rules for HLS will also apply to progressive download playback.

When the user clicks play, you can immediately set AVPlayer's rate property to 1 or call the play method, which is the same thing.

And AVFoundation will automatically wait to buffer enough to avoid stalling.

If the network drops out during playback and playback stalls, the rate property will stay set to 1.

And so it will again buffer and automatically resume when sufficiently buffered.

If you're using the AVKit or MediaPlayer framework to present your playback UI, it already supports automatic waiting for buffering, and it will continue to.

If your application uses AVFoundation directly and you build your own playback UI, you may need to make some adjustments.

So what should we call this new API?

Well, the word Autoplay has been used in QTKit and also in HTML 5, but we came to the conclusion that from the perspective of this AVPlayer API, the playback is not the automatic part.

It's the waiting.

So the formal name for this API is automatically WaitsToMinimizeStalling.

But you can call it Autoplay if you like.

The network playback now looks like a state machine with three states.

Paused, waiting, and playing.

We start in the pause state until the user chooses to play.

And then the app calls play, and we move to the waiting state.

When the playback likelyToKeepUp property becomes true, the player progresses to the playing state.

Now, if the buffer should become empty, the player will switch back to the waiting state until we're likely to keep up again.

Should the user pause, we'll return to the pause state.

Now there's one further transition available.

Recall that in iOS 9 and earlier before this change, you could call play before playback was likely to keep up and playback would start immediately even if it might stall.

So we preserved this semantic by providing another method, playImmediately (atRate:) which jumps you straight into the playing state from either the paused or the waiting states.

Be aware that this may lead to a stall that the patient waiting state would avoid.

So be careful.

AVPlayer's rate property might not mean what you thought it meant.

Let's recap so everyone's clear.

The player's rate property is the app's requested playback rate.

Not to be confused with the time-based rate of the player item which is the rate at which playback is actually occurring.

We've added two new properties in this release to give you more detail.

One is the timeControlStatus, which tells you which of these states you're in, paused, waiting or playing.

And if you're in the waiting state, the reasonForWaitingToPlay property tells you why.

For example, you could be in the waiting state, so the AVPlayer's rate property could be 1.

The timebased.rate would be 0 because you're waiting.

The timeControlStatus would again say I'm WaitingToPlayAtSpcifiedRate.

And the reasonForWaitingToPlay could be WaitingToMinimizeStallsReason.

So with that background, I'd like to introduce my friend Moritz Wittenhagen who is much braver than me, as he is going to attempt a network playback demo live on stage.

So everyone cross your fingers and give him a hand.

[ Applause ]

Well, good morning everyone.

I want to start by showing you a little bit of the setup we have on stage here.

And I have my iPad which you can see mirrored on the screen there.

And that iPad is joining a network that is hosted by my Mac.

And what that allows me to do is I can use the network link conditioner to actually limit the network connection that this iPad has available.

Can do that using the network link conditioner preference pane.

Sam will tell you in a minute where to find that.

And I've set up a profile called Slow Server that limits this to a mediocre network connection that's a little slower than the media bitrate that we actually want to play.

It's currently turned off.

And we'll leave it off, and let's look at what the iPad does in a decent network situation.

So what I have here is just a selection, and I can just select one video.

Let me do that.

And what you see is that the video immediately loads, and we see that we're currently not playing.

You see this wonderful engineering UI underneath that gives us all the properties and functionality involved in automatic waiting.

This is really just taken from AVPlayer and AVPlayer items.

So these are the properties that you have available if you need to know what automatic waiting is doing.

So right now we are paused, so the rates are all zero.

Current time is at zero.

But the interesting thing is since we're in a fast network, we've loaded 39 seconds of the video, which is actually the whole video.

And we're currently likely to keep up.

What that means is that when I just hit play now, the video just starts playing without any problem.

Now we wanted to see what happens in a bad network situation.

So let's turn on the network link condition on the Mac.

Here we go.

And now not much changed for this video.

Because as I said, it was already buffered.

It had already buffered the whole video.

So when I go back and load this again, I want you to pay attention to loadedTimeRanges and isPlaybackLIkelyToKeepUp again.

So let's do it.

Relaod the video.

And now what we see is that loadedTimeRange is only slowly increase.

And isPlaybackLIkelyToKeepUp is false.

Eventually it will become true.

And at that moment we're at the same state that we were before where now ready to play and playback will just start.

Now let's try this one more time, and this time I will hit play right after I loaded the video.

So this time we don't have enough data, and we go into this waiting state.

And you see the spinner telling the user that playback is waiting.

Eventually we will become ready to play and playback just starts.

There's one more thing we can do.

And that is immediate playback.

So let's also try this.

I go into the video and immediately click play immediately.

And we see that playback starts but then we quickly run into a stall because we didn't have enough buffer to play to the end.

In that case, we'll go into the waiting state and re-buffer until we have enough to play through.

And with that, it was a short demo of automatic waiting.

Go back to Sam and the slides.

[ Applause ]

Thanks, Moritz.

Let's recap what was happening in the middle there.

So when we set a slower network speed, close to the data rate of the movie, the movie started out paused.

When he hit play, it went into the waiting state.

Because playback was not yet likely to keep up.

Notice that at this time, the player's rate was 1, but the timebase rate was 0.

After a few seconds, AVFoundation determined that playback was likely to keep up and so it set the time control it set the state into playing, and now you see that the player rate and the timebase rate are both 1.

It may have occurred to you that there's a little bit more detail available in the timeControlStatus than in the player's rate property.

Remember the player's rate property tells you the app's desired playback rate.

The timeControlStatus also takes into account what's actually happening.

So that might be something you want to take into account when you build a playback UI.

In case you want to try this at home, you might like to know how to find the network link conditioner.

It's not something we invented in my time at least.

It is part of the hardware IO tools download.

To get it, the easiest way is to follow Xcode's menu to More Developer Tools.

And after you log in, you'll find it something like here.

Okay, so on the 2016 SDKs if you link on or after that, your app will act as though you had set this property automatically WaitsToMinimizeStalling to true.

You can set that property to false if you want to go back to the old behavior.

And there's a few reasons why you might want to do this.

In particular, if you use the setRate time atHostTime call to synchronize playback with external timeline, then you must opt out by setting the automatically WaitsToMinimizeStalling property to false.

Otherwise, you will meet a friendly exception.

Your helpful reminder.

Finally, a reminder never use the player's rate to extrapolate current timeout in the future.

If you want to do that, use the item's timebase rate for that instead.

Or use the other APIs in the timebase object.

That's what they're for.

All right, that's it for buffering.

Let's move along to the topic of looping.

I have a question for you.

What's the best way to loop playback of a single item?

Well, one idea would be to set up a listener for the notification that fires when playback has reached the end.

And when you get called, seek back to the beginning and start again.

Well, this idea is a good start.

But unfortunately, it will lead to a gap between the playbacks for two reasons.

The first reason is that there will be latency due to the time it takes for the notification to reach your program and for your second player requests to get back to the playback system.

The second more significant reason is the time needed for prerolling.

It's not actually possible to start media playback instantaneously without some preparation.

It's necessary to load media data and decode some of it before you can actually start playing it out.

This process of filling up the playback pipelines before playback starts is called preroll.

So what we'd like to be able to do here is to have AVFoundation be in on the plan.

If AVFoundation knows about playback item B but early enough, then it can begin prerolling and decoding before item A has finished playing out.

And so it can optimize the transition from A to B.

If item B is super short, then AVFoundation may even start work on the transition to item C.

AVFoundation's tool for achieving this is AVQueuePlayer.

AVQueuePlayer is a subclass of AVPlayer, which has an array of AVPlayer items called the play queue.

The current item is the one in the first position of the array.

Now you can use AVQueuePlayer to optimize transitions between items that are different, but for the case of looping, you can create multiple AVPlayer items from the same AVAsset.

This is just another optimization, since AVFoundation does not have to load and pause the media file multiple times.

And just a reminder, the play queue is not a playlist.

Please do not load the next 10,000 items that you think you might like to play into the play queue.

That's not going to be efficient.

The purpose of the play queue is to provide information about items to be played in the near future so that AVFoundation can optimize transitions.

The design patent when you want to loop a single media file indefinitely is to make a small number of AVPlayer items and put them in the AVQueuePlayer's queue with the action item end property set to advance.

When playback reaches the end of one item, it will be removed from the play queue as playback advances to the next one.

And when you get the notification that that has happened, you can take that finished item, set its current time back to the start, and put it on the end of the play queue to reuse it.

We call this patent the treadmill.

And you can implement the treadmill patent yourself using AVQueuePlayer.

We have sample code to help.

The slightly tricky detail is that you have to set up key value observing to watch when the item is removed and then seek it back to the start.

And then add it to the end of the play queue again.

As you can see, in this code we are deactivating our KVO observer while we change the play queue to avoid any chance of recursion.

So this is clearly doable.

It's just a little fiddley.

And the feedback that we received was that it would be awful swell if we could make this easier.

So we're introducing AVPlayerLooper, which implements the treadmill patent for you.

You give it an AVQueuePlayer.

[ Applause ]

You give it an AVQueue Player and a template AVPlayerItem, and it constructs a small number of copies of that AVPlayerItem, which it then cycles through the play queue until you tell it to stop.

Adopting AVPlayerLooper, the code for the symbol case is really much simpler.

So I want to give you a demo of this on an iPad I have over here.

So here's a piece of sample code.

Video Looper, I'm going to launch that.

And I have added a media file of my own here and we're going to play it with AVPlayerLooper.

[ Music ]

Don't you feel mellow?

Okay, this is clearly looping, and the code is pretty much what I pointed out.

It's fairly simple.

This would be an appropriate tool to use, for example, if you have a tvOS app and you'd like to loop background video behind a title menu.

All right, let's return to slides.

We've talked a bit about how to loop.

I want to spend a moment on what to loop.

Ideally, if you have both audio and video tracks, they should be precisely the same length.

Why? Well, if the audio track is longer, then that means that near the end there's period of time when audio should be playing but video should not.

We have an empty segment of video, so what should the video do?

Should it go away?

Should you freeze on one frame?

Conversely, if the video track is longer, then there's a period of time when the audio should be silent.

So when you build media assets for looping, take the time to make sure that the track durations match up.

In QuickTime Movie files, the track duration is defined by the edit list.

Now if the media asset to loop is not entirely under your control, another possibility is that you could set the AVPlayerItems forward playback end time to the length of the shortest track.

This will have the effect of trimming back the other tracks to match.

All right, next look at an optimization that we've made in the playback pipeline that may have an impact on your applications.

Suppose that we are currently playing, and the lists of playing tracks changes.

For example, we could change the subtitle language or the audio language.

Audio from English to French.

Here I'll change the subtitle language from English to Spanish.

Or we could remove the AVPlayerLayer that was displaying the video.

Or we could add an AVPlayerLayer and begin displaying video.

Well, in all of these cases in iOS 9, AVFoundation will pause playback, adjust the playback pipelines to match the list of enables tracks and then resume playback.

In some cases, this even causes video to snap back to a key frame.

Well, I will say we have received constructive feedback from users and developers about this.

And so I'm happy to say that in iOS 10 and its other 2016 siblings, these changes will no longer cause playback to pause.

Adding or removing the only AVPlayerLayer on a playing AVPlayer, changing the subtitle language or the audio language on a playing AVPlayer or manually disabling or enabling tracks.

We think that this is an enhancement for users and developers.

However, it's a significant change in API behavior, and so I would ask you please take a look in the seeds and see if it leads to any complications in your apps.

If you find an issue with this that looks like it's a bug on our side, then please provide feedback by filing a bug using the Apple Bug Reporter System.

And as always when filing a bug, please try to give us everything we need in order to reproduce the problem ourselves.

Our industry is undergoing a transition to wider color gamuts for digital photography and digital video.

Many developers on iOS have never had to deal with video that wasn't using the standard recommendation 709 color space.

Since that's the standard for high-definition video and that's what we've been shooting since the iPhone 4.

But wider gamut color spaces are coming.

As you may have seen with the newest iPad Pro when running iOS 10, you can capture and display photographs in the P3 color space.

Some third party products are capturing video in P3 also.

So I wanted to give you pointers to the APIs you can adopt in your apps to be prepared for making your apps wide color video aware.

But I need to give you a little bit of background first.

In media files, color space information is part of the metadata of video tracks.

In QuickTime Movie files, it's stored in sample descriptions.

In several Codecs also store it in Codec specific places.

There are three principle parts to this information.

Color Primaries, which specific what the 100 percent red, 100 percent green, and 100 percent blue colors are and also the white point.

Transfer Characteristics, which you may have heard called gamma curves or transfer function.

These define the mapping from pixel values to light levels and answer the question is that a straight line or is it a curve that gives you more detail in the dark areas where our eyes are more sensitive.

And the YCbCr Matrix, the coordinate transform from their RGB space into the space used for efficient compression.

So up here I have some examples.

Now if you haven't heard of it, Recommendation 709 is like the video equivalent of SIGB.

SIGB is actually based on Rec.

709. Wide color can be achieved by using a different set of color primaries.

The P3 color primaries specify values for 100 percent red, 100 percent green, and 100 percent blue that are more vivid then Recommendation 709s.

One more point I want to make.

In our APIs, we generally represent these choices through the use of enumerated strings, since they're easier to print and display and debug.

But in media files, these are represented by numbers.

And these standard tag numbers are defined in an MPEG specification called coding independent code points.

That sounds like a paradox, doesn't it?

How can you be coding independent code points?

Well, it's less than a paradox if you read it as Codec independent code points.

The job of the spec is to make sure that the assignment of these tag numbers is done in a manner that is harmonious all Codecs and file formats.

So the interpretation of numbers will be the same in QuickTime Movie, MPEG-4, H264 and so forth.

All right, with that background, let's look at a few new APIs.

We have introduced a new media characteristic that tells you that at video track is tagged with wider color primaries, something wider than the Rec.

709 primaries.

If your app finds that there is wide gamut video, it might be appropriate for your app to take steps to preserve it, so it isn't clamped back into the 709 space.

If not, it's actually generally best to stay within Rec.

709 for processing.

So you can specify a working color space when you set up an AVPlayerItemVideoOutput or an AVAssetReaderOutput.

And you will then receive buffers that have been converted into that color space.

You can also specify a target color space when setting up an AVAssetWriterInput, in which case the source image buffers that you provide will be converted into that color space prior to compression.

With AVPlayerItemVideoOutput or AVAssetReaderOutput if you don't want image buffers to be converted into a common color space, then you should set the AVVideoAllowWideColorKey to true and then you'll receive buffers in their original color space.

This is effectively a promise that whatever software receives and processes those buffers, whether it's ours or yours, it will examine and honor their color space tags.

There are analogous properties for configuring video compositions.

First, you can specify a working color space for entire video compositions.

Alternatively, if you have a custom video compositor, you may choose to make it wide color aware.

You can declare that your custom video compositor is wide color aware and that it examines and honors color space tags on every single source frame buffer by implementing the optional supportsWideColorSourceFrames property and returning true.

Running it out with a reminder, if you create picture buffers manually, for example, using a pixel buffer pool in metal, then you should explicitly set the color space tags on every buffer by calling core videos APIs.

Most developers won't need to do this.

In most cases when you're using a color space aware API for source buffers, that'll take care of tagging them for you.

By popular request, I'm going to spend the rest of our time discussing some best practices for optimizing playback startup time.

I'll talk about local file playback first.

And then we'll move on to HTTP Live Streaming.

Now some of these optimization techniques may be counterintuitive at first.

They require you to consider things from the perspective of AVFoundation.

And to think about when it gets the information it needs to do what your app is asking it to do.

For example, here is a straightforward piece of code for setting up playback of a local file.

We start with the URL to the file.

We create an AVURLAsset representing the product depositing that file.

We then create an AVPlayerItem to hold the mutable state for playback, and an AVPlayer item to host playback.

And then we create an AVPlayerLayer to connect video playback into our display hierarchy.

Now this code is correct, but it has a small flaw, which maybe you may not initially see.

As soon as the player item is set as the player's current item, the player starts setting up the playback pipeline.

Now it doesn't know the future.

It doesn't know that you're going to set an AVPlayerLayout later.

So it sets things up for audio only playback.

And then when the AVPlayerLayer is added, now AVFoundation knows that the video needs to be decoded too.

And so now it can reconfigure things for audio and video playback.

Now, as I said earlier, we have made enhancements in this year's OS releases to mean that minor changes to the list of playback to the list of enabled tracks do not necessarily cause an interruption.

But it still ideal to start with the information that AVFoundation needs in order to get things right first time.

So I'm going to change this code a little bit.

I'm going to watch where the AVPlayerItem is connected to the AVPlayer.

So now the player is created with no current item, which means it has no reason to build playback pipelines yet.

And that doesn't change when you add the AVPlayerLayer.

Playback pipelines don't get built until the player item becomes the current item.

And by that point, the player know what it needs to get things right first time.

We can generalize this.

First, create the AVPlayerLayer, so first create the AVPlayer and AVPlayerItem objects.

And set whatever properties you need to on them including connecting the AVPlayer to an AVPlayerLayer or an AVPlayerItem to an AVPlayerItemVideoOutput.

Now this might seem crazy, but if you just want playback to start right away, you can tell the player to play before you give it the item to play.

Why would you do this?

Well, if you do it the other way around, the player initially thinks that you wanted to display the still frame at the start of the video.

And it might waste some time on that before it gets the message that actually you just want playback.

Again, starting with the actual goal may shave off a few milliseconds.

Let's move on to HLS.

The timeframes we're trying to optimize with HLS are longer because they're donated by network IO which is much slower than local file storage.

So the potential benefits of optimizations are much more noticeable.

The network IO breaks down into four pieces.

Retrieving the master playlist that's the URL you passed to AVURLAsset.

If the content is protected with fair play streaming, retrieving content keys, retrieving the selected variant playlists for the appropriate bitrate and format of video and audio, and retrieving some media segments that are referenced in that playlist.

Now the media segments will be the highest amount of actual data transfer but with network IO we need to think about round-trip latency.

Some of these stages are serialized.

You can't download things from playlist until you've received the playlist.

So a thing to think about then is can we do any of these things before the user chooses to play?

For example, maybe in your app you display a title card when content is first selected, and that gets the user to say, is this actually the one I wanted to play?

Or do I want to read some information about it.

So the question is, could we do some small amount of network IO speculatively when the user has identified the content they probably want to play before they make it official?

Well, AVURLAsset is a lazy API.

It doesn't begin loading or pausing any data until someone asks it to.

To trigger it to load data from the master playlist, we need to ask it to load a value that would derive from it like duration or available media characteristics with media selection options.

Duration is easy to type.

You don't have to provide a completion handler here unless you're actually going to do something with that value.

Speaking of playlists, they can press really easily, and we've supported compressing them with gzip for many years.

So make sure you're doing that.

Possibly it's just a matter of configuring your server.

If your content is protected using fair play streaming, then there's round-trip involved in negotiating content keys with your server.

And you can trigger that to happen sooner by setting the preloadsEligibleContentKeys property of the asset.resourceLoader to true.

For this to work, the master playlist must contain SESSION-KEY declarations.

So how are we doing so far?

With these techniques, we can start they can get the master playlist and the content keys downloaded, while we're still on the title card.

Now that's pretty cool.

The variant playlists and the segments of data will still load after we hit play.

So you might be asking yourselves, can we push this technique even further?

Well, there is a new API in 2016 called preferredForwardBufferDuration.

Setting it to something low like five seconds will buffer the minimum amount that AVFoundation thinks you need to get started.

But once playback begins, set the override back to zero to allow normal buffering algorithms to take over again.

Here's a list of video variance that might be in a master playlist.

They vary in dimensions and bitrate.

For an Apple TV on a fast connection with a big TV, the 1080p variant might be ideal.

For an iPhone SE, even with a superfast Wi-Fi connection, the 720p variant is the best choice.

It's already higher resolution than the iPhone SE screen, so going bigger probably won't improve any quality.

On a giant iPad Pro, there are a lot of pixels, so we could go up to a big variant for full screen.

But if we play picture-and-picture, we don't need such a high resolution anymore.

And a lower bitrate variant could reduce the size of our cache and help us make more memory available for other apps.

If the network connection is slow on any device, then that's going to be the limiting factor.

So what this means is that AVFoundation needs to take into account both the display dimensions and the network bitrate when choosing the variant.

AVFoundation uses the AVPlayerLayer size on the screen to evaluate the dimensions.

So set up your AVPlayerLayer at the correct size and connect it to the AVPlayer as early as you can.

It can be hidden behind other UI if you're not ready to show video yet.

On a retina iOS device, it's currently necessary to set contentsScale manually.

As for bitrate, well AVFoundation is in a bit of a chicken and egg situation when it comes to playback first beginning.

It has to choose some variant, but it does not know what bitrate it's going to get.

Once it's begun downloading segments, it can use the statistics from those downloads to adjust the choice of variant.

But for that first variant, it hasn't gathered any statistics yet.

So AVFoundation's base algorithm is to pick the first applicable variant in the master playlist.

If that's a low bitrate option, the user will start out seeing something blurry, but AVFoundation will soon decide what the actual network bitrate is and switch up to the appropriate variant.

Well, the question is, what if you would like to try to improve that initial choice?

Well, remember, there is a tradeoff you have to make between initial quality and startup time.

A higher bitrate first segment takes longer to download.

And that means it will take longer to start.

You might decide that it's best to start with a lower bitrate variant in order to start faster.

Well one way to make the tradeoff is to figure out a minimum acceptable quality level you'd like to see on a particular size of screen and start there.

Then AVFoundation will switch to a higher quality after playback begins as the network allows.

And maybe you know one thing that AVFoundation doesn't.

Maybe your app just played a different piece of content.

And maybe you can use that playback's access log to make a better guess about the bitrate that the next playback station is going to get.

So let's suppose that you come up with a hero stick based on startup quality and recent bitrate statistics.

And you decide on a way to choose which variant you want to start with.

Well, how do we plug that choice into AVFoundation?

There are two techniques that have been used.

Here's the first technique.

On the server, you have to sort your variance from highest to lowest.

Like that.

And then in your app, you need to set the player items preferredPeakBitRate to you bitrate guess.

This will eliminate the higher bitrate variance from initial selection.

Shortly after playback starts, you should reset that control back to zero, which will allow AVFoundation to move up to a higher bitrate variance if the network improves.

The second technique is to dynamically rewrite the master playlist in your app and move your preferred choice to the top of the list.

To do this, use a custom URL scheme for the AVURLAsset and implement the AV asset resource loader delegate protocol in which you can supply that rewritten playlist in response to the load request for that custom URL scheme.

I want to remind you to profile your code too.

Look for any delays before you call AVFoundation.

In particular, you do not need to wait for likelyToKeepUp to become true before you set the player rate.

You don't need to now, and in fact, you never have for HLS.

Make sure that you release AVPlayers and AVPlayerItems from old playback sessions so that they do not waste bandwidth in the background.

You can use the Allocations Instrument in Instruments to check the lifespans of AVPlayer and AVPlayerItem objects.

And if you have an application that does other network activity, consider whether you should suspend it during network playback so that the user can take full advantage of available bandwidth for playback.

All right, in conclusion, automaticallyWaits to minimize stalling, Autoplay, Autowait.

It's set to true by default if your app is linked on or after this year's SDKs.

And it provides uniform buffering rules for progressive download and HLS playback.

We've introduced a new API called AVPlayerLooper to simplify using the treadmill patent to loop playback of a single item.

Changing the set of enable tracks during playback no longer always causes a brief pause.

And we've looked at the AVFoundation APIs that you can use to prepare your app for wide color video.

Finally, we talked about optimizing playback startup for local files and for HLS.

In short, avoid accidentally asking for work you don't need.

And for the work you do need, see if you can do it earlier.

We'll have more information at this URL about this session, including sample code that we've shown.

We have some related sessions that you might like to catch up to see in person or catch up online.

The bottom one is an on-demand only one that you can watch in the app.

Thank you for attention.

It's been a pleasure.

I hope you have a great week.

Apple, Inc. AAPL
1 Infinite Loop Cupertino CA 95014 US