MTAudioProcessingTap EXC_BAD_ACCESS , doesnt always fire the finalize callback. how to Release it? - ios

Question

Im trying to implement MTAudioProcessingTap and it works great. The problem is when Im done using the Tap and I reinstaniate my class and create a new Tap.
How Im supposely releasing the tap
1- I retain the tap as a property when created, hoping I can access it and release it later
2- In deinit() method of the class, I set the audiomix to nil and try to do a self.tap?.release()
The thing is.. sometimes it works and calls the FINALIZE callback and everything is great, and sometimes it doesn't and just crashes at the tapProcess Callback line:
let selfMediaInput = Unmanaged<VideoMediaInput>.fromOpaque(MTAudioProcessingTapGetStorage(tap)).takeUnretainedValue()
Here's the full code: https://gist.github.com/omarojo/03d08165a1a7962cb30c17ec01f809a3
import Foundation
import UIKit
import AVFoundation;
import MediaToolbox
protocol VideoMediaInputDelegate: class {
func videoFrameRefresh(sampleBuffer: CMSampleBuffer) //could be audio or video
}
class VideoMediaInput: NSObject {
private let queue = DispatchQueue(label: "com.GenerateMetal.VideoMediaInput")
var videoURL: URL!
weak var delegate: VideoMediaInputDelegate?
private var playerItemObserver: NSKeyValueObservation?
var displayLink: CADisplayLink!
var player = AVPlayer()
var playerItem: AVPlayerItem!
let videoOutput = AVPlayerItemVideoOutput(pixelBufferAttributes: [String(kCVPixelBufferPixelFormatTypeKey): NSNumber(value: kCVPixelFormatType_32BGRA)])
var audioProcessingFormat: AudioStreamBasicDescription?//UnsafePointer<AudioStreamBasicDescription>?
var tap: Unmanaged<MTAudioProcessingTap>?
override init(){
}
convenience init(url: URL){
self.init()
self.videoURL = url
self.playerItem = AVPlayerItem(url: url)
playerItemObserver = playerItem.observe(\.status) { [weak self] item, _ in
guard item.status == .readyToPlay else { return }
self?.playerItemObserver = nil
self?.player.play()
}
setupProcessingTap()
player.replaceCurrentItem(with: playerItem)
player.currentItem!.add(videoOutput)
NotificationCenter.default.removeObserver(self)
NotificationCenter.default.addObserver(forName: NSNotification.Name.AVPlayerItemDidPlayToEndTime, object: nil, queue: nil) {[weak self] notification in
if let weakSelf = self {
/*
Setting actionAtItemEnd to None prevents the movie from getting paused at item end. A very simplistic, and not gapless, looped playback.
*/
weakSelf.player.actionAtItemEnd = .none
weakSelf.player.seek(to: CMTime.zero)
weakSelf.player.play()
}
}
NotificationCenter.default.addObserver(
self,
selector: #selector(applicationDidBecomeActive(_:)),
name: UIApplication.didBecomeActiveNotification,
object: nil)
}
func stopAllProcesses(){
self.queue.sync {
self.player.pause()
self.player.isMuted = true
self.player.currentItem?.audioMix = nil
self.playerItem.audioMix = nil
self.playerItem = nil
self.tap?.release()
}
}
deinit{
print(">> VideoInput deinited !!!! 📌📌")
if let link = self.displayLink {
link.invalidate()
}
NotificationCenter.default.removeObserver(self)
stopAllProcesses()
}
public func playVideo(){
if (player.currentItem != nil) {
print("Starting playback!")
player.play()
}
}
public func pauseVideo(){
if (player.currentItem != nil) {
print("Pausing playback!")
player.pause()
}
}
#objc func applicationDidBecomeActive(_ notification: NSNotification) {
playVideo()
}
//MARK: GET AUDIO BUFFERS
func setupProcessingTap(){
var callbacks = MTAudioProcessingTapCallbacks(
version: kMTAudioProcessingTapCallbacksVersion_0,
clientInfo: UnsafeMutableRawPointer(Unmanaged.passUnretained(self).toOpaque()),
init: tapInit,
finalize: tapFinalize,
prepare: tapPrepare,
unprepare: tapUnprepare,
process: tapProcess)
var tap: Unmanaged<MTAudioProcessingTap>?
let err = MTAudioProcessingTapCreate(kCFAllocatorDefault, &callbacks, kMTAudioProcessingTapCreationFlag_PostEffects, &tap)
self.tap = tap
print("err: \(err)\n")
if err == noErr {
}
print("tracks? \(playerItem.asset.tracks)\n")
let audioTrack = playerItem.asset.tracks(withMediaType: AVMediaType.audio).first!
let inputParams = AVMutableAudioMixInputParameters(track: audioTrack)
inputParams.audioTapProcessor = tap?.takeRetainedValue()//tap?.takeUnretainedValue()
// tap?.release()
// print("inputParms: \(inputParams), \(inputParams.audioTapProcessor)\n")
let audioMix = AVMutableAudioMix()
audioMix.inputParameters = [inputParams]
playerItem.audioMix = audioMix
}
//MARK: TAP CALLBACKS
let tapInit: MTAudioProcessingTapInitCallback = {
(tap, clientInfo, tapStorageOut) in
tapStorageOut.pointee = clientInfo
print("init \(tap, clientInfo, tapStorageOut)\n")
}
let tapFinalize: MTAudioProcessingTapFinalizeCallback = {
(tap) in
print("finalize \(tap)\n")
}
let tapPrepare: MTAudioProcessingTapPrepareCallback = {
(tap, itemCount, basicDescription) in
print("prepare: \(tap, itemCount, basicDescription)\n")
let selfMediaInput = Unmanaged<VideoMediaInput>.fromOpaque(MTAudioProcessingTapGetStorage(tap)).takeUnretainedValue()
selfMediaInput.audioProcessingFormat = AudioStreamBasicDescription(mSampleRate: basicDescription.pointee.mSampleRate,
mFormatID: basicDescription.pointee.mFormatID, mFormatFlags: basicDescription.pointee.mFormatFlags, mBytesPerPacket: basicDescription.pointee.mBytesPerPacket, mFramesPerPacket: basicDescription.pointee.mFramesPerPacket, mBytesPerFrame: basicDescription.pointee.mBytesPerFrame, mChannelsPerFrame: basicDescription.pointee.mChannelsPerFrame, mBitsPerChannel: basicDescription.pointee.mBitsPerChannel, mReserved: basicDescription.pointee.mReserved)
}
let tapUnprepare: MTAudioProcessingTapUnprepareCallback = {
(tap) in
print("unprepare \(tap)\n")
}
let tapProcess: MTAudioProcessingTapProcessCallback = {
(tap, numberFrames, flags, bufferListInOut, numberFramesOut, flagsOut) in
print("callback \(bufferListInOut)\n")
let selfMediaInput = Unmanaged<VideoMediaInput>.fromOpaque(MTAudioProcessingTapGetStorage(tap)).takeUnretainedValue()
let status = MTAudioProcessingTapGetSourceAudio(tap, numberFrames, bufferListInOut, flagsOut, nil, numberFramesOut)
//print("get audio: \(status)\n")
if status != noErr {
print("Error TAPGetSourceAudio :\(String(describing: status.description))")
return
}
selfMediaInput.processAudioData(audioData: bufferListInOut, framesNumber: UInt32(numberFrames))
}
func processAudioData(audioData: UnsafeMutablePointer<AudioBufferList>, framesNumber: UInt32) {
var sbuf: CMSampleBuffer?
var status : OSStatus?
var format: CMFormatDescription?
//FORMAT
// var audioFormat = self.audioProcessingFormat//self.audioProcessingFormat?.pointee
guard var audioFormat = self.audioProcessingFormat else {
return
}
status = CMAudioFormatDescriptionCreate(allocator: kCFAllocatorDefault, asbd: &audioFormat, layoutSize: 0, layout: nil, magicCookieSize: 0, magicCookie: nil, extensions: nil, formatDescriptionOut: &format)
if status != noErr {
print("Error CMAudioFormatDescriptionCreater :\(String(describing: status?.description))")
return
}
print(">> Audio Buffer mSampleRate:\(Int32(audioFormat.mSampleRate))")
var timing = CMSampleTimingInfo(duration: CMTimeMake(value: 1, timescale: Int32(audioFormat.mSampleRate)), presentationTimeStamp: self.player.currentTime(), decodeTimeStamp: CMTime.invalid)
status = CMSampleBufferCreate(allocator: kCFAllocatorDefault,
dataBuffer: nil,
dataReady: Bool(truncating: 0),
makeDataReadyCallback: nil,
refcon: nil,
formatDescription: format,
sampleCount: CMItemCount(framesNumber),
sampleTimingEntryCount: 1,
sampleTimingArray: &timing,
sampleSizeEntryCount: 0, sampleSizeArray: nil,
sampleBufferOut: &sbuf);
if status != noErr {
print("Error CMSampleBufferCreate :\(String(describing: status?.description))")
return
}
status = CMSampleBufferSetDataBufferFromAudioBufferList(sbuf!,
blockBufferAllocator: kCFAllocatorDefault ,
blockBufferMemoryAllocator: kCFAllocatorDefault,
flags: 0,
bufferList: audioData)
if status != noErr {
print("Error cCMSampleBufferSetDataBufferFromAudioBufferList :\(String(describing: status?.description))")
return
}
let currentSampleTime = CMSampleBufferGetOutputPresentationTimeStamp(sbuf!);
print(" audio buffer at time: \(currentSampleTime)")
self.delegate?.videoFrameRefresh(sampleBuffer: sbuf!)
}
}
How I use my class
self.inputVideoMedia = nil
self.inputVideoMedia = VideoMediaInput(url: videoURL)
self.inputVideoMedia!.delegate = self
the second time I do that.. it crashes (but not always). The times it doesnt crash I can see printed in the console the FINALIZE print.

Answer 1

If VideoMediaInput is deallocated before the tap is deallocated (which can happen as there seems to be no way to synchronously stop a tap), then the tap callback can choke on a reference to your deallocated class.
You can fix this by passing (a wrapped, I guess) weak reference to your class. You can do it like this:
First delete your tap instance variable, and any references to it - it's not needed. Then make these changes:
class VideoMediaInput: NSObject {
class TapCookie {
weak var input: VideoMediaInput?
deinit {
print("TapCookie deinit")
}
}
...
func setupProcessingTap(){
let cookie = TapCookie()
cookie.input = self
var callbacks = MTAudioProcessingTapCallbacks(
version: kMTAudioProcessingTapCallbacksVersion_0,
clientInfo: UnsafeMutableRawPointer(Unmanaged.passRetained(cookie).toOpaque()),
init: tapInit,
finalize: tapFinalize,
prepare: tapPrepare,
unprepare: tapUnprepare,
process: tapProcess)
...
let tapFinalize: MTAudioProcessingTapFinalizeCallback = {
(tap) in
print("finalize \(tap)\n")
// release cookie
Unmanaged<TapCookie>.fromOpaque(MTAudioProcessingTapGetStorage(tap)).release()
}
let tapPrepare: MTAudioProcessingTapPrepareCallback = {
(tap, itemCount, basicDescription) in
print("prepare: \(tap, itemCount, basicDescription)\n")
let cookie = Unmanaged<TapCookie>.fromOpaque(MTAudioProcessingTapGetStorage(tap)).takeUnretainedValue()
let selfMediaInput = cookie.input!
...
let tapProcess: MTAudioProcessingTapProcessCallback = {
(tap, numberFrames, flags, bufferListInOut, numberFramesOut, flagsOut) in
print("callback \(bufferListInOut)\n")
let cookie = Unmanaged<TapCookie>.fromOpaque(MTAudioProcessingTapGetStorage(tap)).takeUnretainedValue()
guard let selfMediaInput = cookie.input else {
print("Tap callback: VideoMediaInput was deallocated!")
return
}
...
I'm not sure if the cookie class is necessary, it exists only to wrap the weak reference. Cutting edge Swift experts may know how to mash the weakness through all the teenage mutant ninja raw pointers, but I don't.

Answer 2

The audio context runs in its own real-time thread. So audio processes don't stop synchronously with a stop or cancel function call, but some unknown time later (on the order of the duration of some number of audio samples in some internal audio buffers), after the real-time thread drains.
Thus, audio buffers, objects, and callbacks should not be released (or reallocated) until some (unknown, but less than a couple seconds) time after stopping any real-time audio stream.
Depending on deallocation object messages or instance variable states (including weak references) betweens real-time threads is reported to be currently unsafe in Swift (see WWDC 2018 session on audio). Thus, I recommend using semaphores (outside of a real-time context, such as audio), or posix memory barriers (inside a bridged call to a C function). (...until some future version of Swift figures out a real-time concurrency mechanism.) (...especially on iOS or Apple Silicon (M1) devices which can re-order memory writes).