I have been used GPUImage to do some chroma key video merge(I extended GPUImageThreeInputFilter to support four videos blending). But when I try to use GPUImageMovieWriter to write out blended videos, I found some problems - typically those are very flaky to reproduce and it just happens from time to time, which causes lots of headache.
Typical problems/limits you might run into when use GPUImageMovieWriter are:
Output video got cut off Prematurely, i.e, last one second lost #1403 , #1635
Only one audio track got written out #934
Frame dropping, frame time out of sync, some video frame freezes #1501 , #744
Video writing ends before audio’s, which last one or two seconds video freeze #830
Black frames at the beginning #52 , #652
Poor stability, random crashes #773 , #1634
I already have elaborated on solutions to those problems in my blog GPUImage Video Merging: Timestamp Synchronization and GPUImage Movie Writer: Merging All Audio Tracks From Multiple Movies .
But I need a way to incorporate those changes to library nicely without completely writing from scratch so that I could facilitate existing functions of GPUImage.
What we will do is create three classes: THImageMovie, THImageMovieWriter and THImageMovieManager.
To start, let’s duplicate GPUImageMovie and GPUImageMovieWriter and name it to THImageMovie and THImageMovieWriter.
THImageMovie
In THImageMovie, we’re not gonna do many changes:
- ( void ) processAsset
{
reader = [ self createAssetReader ];
AVAssetReaderOutput * readerAudioTrackOutput = nil ;
audioEncodingIsFinished = YES ;
for ( AVAssetReaderOutput * output in reader . outputs ) {
if ( [ output . mediaType isEqualToString : AVMediaTypeAudio ] ) {
audioEncodingIsFinished = NO ;
readerAudioTrackOutput = output ;
}
else if ( [ output . mediaType isEqualToString : AVMediaTypeVideo ] ) {
readerVideoTrackOutput = output ;
}
}
if ([ reader startReading ] == NO )
{
NSLog ( @"Error reading from file at URL: %@" , self . url );
return ;
}
//============================We delete all the code left, just signify that I'm done with asset warming up, ready to be processed
dispatch_group_leave ([ THImageMovieManager shared ]. readingAllReadyDispatchGroup );
}
- ( BOOL ) renderNextFrame {
__unsafe_unretained THImageMovie * weakSelf = self ;
if ( reader . status == AVAssetReaderStatusReading && ( ! _shouldRepeat || keepLooping ))
{
return [ weakSelf readNextVideoFrameFromOutput : readerVideoTrackOutput ];
}
if ( reader . status == AVAssetWriterStatusCompleted ) {
NSLog ( @"movie: %@ reading is done" , self . url . lastPathComponent );
[ reader cancelReading ];
[ weakSelf endProcessing ];
}
return NO ;
}
Above is the most important changes, we basically remove readNextVideoFrameFromOutput in processAsset method, and add a new method - (BOOL)renderNextFrame .
What we do is renderNextFrame is just read next video frame and return YES/NO if reading is completed. One important thing here we dont’ call endProcessing in readNextVideoFrameFromOutput anymore, you could check in detail commit on github.
This is pretty much the changes for movie class.
THImageMovieWriter
In header class THImageMovieWriter.h:
//add one more parameter movies to init method
- ( id ) initWithMovieURL :( NSURL * ) newMovieURL size :( CGSize ) newSize movies :( NSArray * ) movies ;
//remove the instance variable assetWriterPixelBufferInput and move it as property to prevent memory leak when used in block
@property ( nonatomic , retain ) AVAssetWriterInputPixelBufferAdaptor * assetWriterPixelBufferInput ;
in THImageMovieWriter.m, First, let’s add multiple audio tracks reader using AVAssetReaderAudioMixOutput and writer:
#pragma mark setupAssetWriter
- ( void ) setupAudioAssetReader {
NSMutableArray * audioTracks = [ NSMutableArray array ];
for ( THImageMovie * movie in self . movies ){
AVAsset * asset = movie . asset ;
if ( asset ){
NSArray * _audioTracks = [ asset tracksWithMediaType : AVMediaTypeAudio ];
if ( _audioTracks . count > 0 ){
[ audioTracks addObject : _audioTracks . firstObject ];
}
}
}
AVMutableComposition * mixComposition = [ AVMutableComposition composition ];
for ( AVAssetTrack * track in audioTracks ){
if ( ! [ track isKindOfClass :[ NSNull class ]]){
AVMutableCompositionTrack * compositionCommentaryTrack = [ mixComposition addMutableTrackWithMediaType : AVMediaTypeAudio
preferredTrackID: kCMPersistentTrackID_Invalid ];
[ compositionCommentaryTrack insertTimeRange : CMTimeRangeMake ( kCMTimeZero , track . asset . duration )
ofTrack: track
atTime: kCMTimeZero error : nil ];
}
}
self . assetAudioReader = [ AVAssetReader assetReaderWithAsset : mixComposition error : nil ];
self . assetAudioReaderTrackOutput =
[[ AVAssetReaderAudioMixOutput alloc ] initWithAudioTracks :[ mixComposition tracksWithMediaType : AVMediaTypeAudio ]
audioSettings: nil ];
[ self . assetAudioReader addOutput : self . assetAudioReaderTrackOutput ];
}
- ( void ) setupAudioAssetWriter {
double preferredHardwareSampleRate = [[ AVAudioSession sharedInstance ] currentHardwareSampleRate ];
AudioChannelLayout acl ;
bzero ( & acl , sizeof ( acl ));
acl . mChannelLayoutTag = kAudioChannelLayoutTag_Mono ;
NSDictionary * audioOutputSettings = [ NSDictionary dictionaryWithObjectsAndKeys :
[ NSNumber numberWithInt : kAudioFormatMPEG4AAC ], AVFormatIDKey ,
[ NSNumber numberWithInt : 1 ], AVNumberOfChannelsKey ,
[ NSNumber numberWithFloat : preferredHardwareSampleRate ], AVSampleRateKey ,
[ NSData dataWithBytes : & acl length : sizeof ( acl ) ], AVChannelLayoutKey ,
//[ NSNumber numberWithInt:AVAudioQualityLow], AVEncoderAudioQualityKey,
[ NSNumber numberWithInt : 64000 ], AVEncoderBitRateKey ,
nil ];
assetWriterAudioInput = [ AVAssetWriterInput assetWriterInputWithMediaType : AVMediaTypeAudio outputSettings : audioOutputSettings ];
[ assetWriter addInput : assetWriterAudioInput ];
assetWriterAudioInput . expectsMediaDataInRealTime = _encodingLiveVideo ;
}
Next, change startRecording to following:
- ( void ) startRecording ;
{
dispatch_group_notify ([ THImageMovieManager shared ]. readingAllReadyDispatchGroup , dispatch_get_main_queue (), ^ {
NSLog ( @"all set, readers and writer both are ready" );
[ self setupAudioAssetReader ];
[ self setupAudioAssetWriter ];
alreadyFinishedRecording = NO ;
isRecording = YES ;
BOOL aduioReaderStartSuccess = [ self . assetAudioReader startReading ];
if ( ! aduioReaderStartSuccess ){
NSLog ( @"asset audio reader start reading failed: %@" , self . assetAudioReader . error );
return ;
}
startTime = kCMTimeInvalid ;
[ self . assetWriter startWriting ];
[ self . assetWriter startSessionAtSourceTime : kCMTimeZero ];
NSLog ( @"asset write is good to write..." );
[ self kickoffRecording ];
});
}
- ( void ) kickoffRecording {
// If the asset reader and writer both started successfully, create the dispatch group where the reencoding will take place and start a sample-writing session.
self . recordingDispatchGroup = dispatch_group_create ();
self . audioFinished = NO ;
self . videoFinished = NO ;
[ self kickOffAudioWriting ];
[ self kickOffVideoWriting ];
__unsafe_unretained typeof ( self ) weakSelf = self ;
// Set up the notification that the dispatch group will send when the audio and video work have both finished.
dispatch_group_notify ( self . recordingDispatchGroup , [ THImageMovieManager shared ]. mainSerializationQueue , ^ {
weakSelf . videoFinished = NO ;
weakSelf . audioFinished = NO ;
[ weakSelf . assetWriter finishWritingWithCompletionHandler : ^ {
if ( weakSelf . completionBlock ){
weakSelf . completionBlock ();
}
}];
});
}
Here we only start process when all movies are ready and use dispatch_group to ensure audio and video both finished.
- ( void ) kickOffAudioWriting {
dispatch_group_enter ( self . recordingDispatchGroup );
__unsafe_unretained typeof ( self ) weakSelf = self ;
CMTime shortestDuration = kCMTimeInvalid ;
for ( THImageMovie * movie in self . movies ) {
AVAsset * asset = movie . asset ;
if ( CMTIME_IS_INVALID ( shortestDuration )){
shortestDuration = asset . duration ;
} else {
if ( CMTimeCompare ( asset . duration , shortestDuration ) == - 1 ){
shortestDuration = asset . duration ;
}
}
}
[ assetWriterAudioInput requestMediaDataWhenReadyOnQueue :[ THImageMovieManager shared ]. rwAudioSerializationQueue usingBlock : ^ {
// Because the block is called asynchronously, check to see whether its task is complete.
if ( self . audioFinished )
return ;
BOOL completedOrFailed = NO ;
// If the task isn't complete yet, make sure that the input is actually ready for more media data.
while ([ assetWriterAudioInput isReadyForMoreMediaData ] && ! completedOrFailed ) {
// Get the next audio sample buffer, and append it to the output file.
CMSampleBufferRef sampleBuffer = [ self . assetAudioReaderTrackOutput copyNextSampleBuffer ];
if ( sampleBuffer != NULL ) {
CMTime currentSampleTime = CMSampleBufferGetOutputPresentationTimeStamp ( sampleBuffer );
BOOL isDone = CMTimeCompare ( shortestDuration , currentSampleTime ) == - 1 ;
BOOL success = [ assetWriterAudioInput appendSampleBuffer : sampleBuffer ];
weakSelf . audioWroteDuration = CMTimeGetSeconds ( currentSampleTime );
if ( success ) {
//NSLog(@"append audio buffer success");
} else {
NSLog ( @"append audio buffer failed" );
}
CFRelease ( sampleBuffer );
sampleBuffer = NULL ;
completedOrFailed = ! success ;
if ( isDone ){
completedOrFailed = YES ;
}
}
else {
completedOrFailed = YES ;
}
} //end of loop
if ( completedOrFailed ) {
NSLog ( @"kickOffAudioWriting wrint done" );
// Mark the input as finished, but only if we haven't already done so, and then leave the dispatch group (since the audio work has finished).
BOOL oldFinished = self . audioFinished ;
self . audioFinished = YES ;
if ( ! oldFinished ) {
[ assetWriterAudioInput markAsFinished ];
dispatch_group_leave ( self . recordingDispatchGroup );
};
}
}];
}
in audio writng part, we got shortest duration of video and prevent final audio duration in output is longer than video.
- ( void ) kickOffVideoWriting {
dispatch_group_enter ( self . recordingDispatchGroup );
self . isFrameRecieved = NO ;
__unsafe_unretained typeof ( self ) weakSelf = self ;
self . firstVideoFrameTime = - 1 ;
self . onFramePixelBufferReceived = ^ ( CMTime frameTime , CVPixelBufferRef pixel_buffer ){
[ weakSelf . assetWriterPixelBufferInput appendPixelBuffer : pixel_buffer withPresentationTime : frameTime ];
if ( weakSelf . firstVideoFrameTime == - 1 ){
weakSelf . firstVideoFrameTime = CMTimeGetSeconds ( frameTime );
}
CVPixelBufferUnlockBaseAddress ( pixel_buffer , 0 );
weakSelf . videoWroteDuration = CMTimeGetSeconds ( frameTime );
if ( ! [ GPUImageContext supportsFastTextureUpload ])
{
CVPixelBufferRelease ( pixel_buffer );
}
weakSelf . isFrameRecieved = NO ;
};
[ assetWriterVideoInput requestMediaDataWhenReadyOnQueue :[ THImageMovieManager shared ]. rwVideoSerializationQueue usingBlock : ^ {
if ( self . videoFinished )
return ;
BOOL completedOrFailed = NO ;
// If the task isn't complete yet, make sure that the input is actually ready for more media data.
while ([ assetWriterVideoInput isReadyForMoreMediaData ] && ! completedOrFailed ) {
if ( ! self . isFrameRecieved ){
self . isFrameRecieved = YES ;
for ( THImageMovie * movie in self . movies ){
BOOL hasMoreFrame = [ movie renderNextFrame ];
//NSLog(@"--movie: %@, has more frames: %d", movie.url.lastPathComponent, hasMoreFrame);
if ( ! hasMoreFrame ){
completedOrFailed = YES ;
break ;
}
}
}
}
if ( completedOrFailed ){
NSLog ( @"kickOffVideoWriting mark as finish" );
// Mark the input as finished, but only if we haven't already done so, and then leave the dispatch group (since the video work has finished).
BOOL oldFinished = self . videoFinished ;
self . videoFinished = YES ;
if ( ! oldFinished ) {
for ( THImageMovie * movie in self . movies ){
[ movie cancelProcessing ];
}
[ assetWriterVideoInput markAsFinished ];
dispatch_group_leave ( self . recordingDispatchGroup );
}
}
}];
}
Here in video writing, we defined onFramePixelBufferReceived which should be called when final rendered result in GPU is ready in newFrameReadyAtTime . Only if we have recieved merged pixels then we kick off next video reading round. So in this way, we could ensure the frame-perfect video writing and fully synced timestamps.
- ( void ) newFrameReadyAtTime :( CMTime ) frameTime atIndex :( NSInteger ) textureIndex ;
{
if ( ! isRecording )
{
[ firstInputFramebuffer unlock ];
return ;
}
// // Drop frames forced by images and other things with no time constants
// // Also, if two consecutive times with the same value are added to the movie, it aborts recording, so I bail on that case
// if ( (CMTIME_IS_INVALID(frameTime)) || (CMTIME_COMPARE_INLINE(frameTime, ==, previousFrameTime)) || (CMTIME_IS_INDEFINITE(frameTime)) )
// {
// [firstInputFramebuffer unlock];
// return;
// }
//
// if (CMTIME_IS_INVALID(startTime))
// {
// runSynchronouslyOnContextQueue(_movieWriterContext, ^{
// if ((videoInputReadyCallback == NULL) && (assetWriter.status != AVAssetWriterStatusWriting))
// {
// [assetWriter startWriting];
// }
//
// [assetWriter startSessionAtSourceTime:frameTime];
// startTime = frameTime;
// });
// }
GPUImageFramebuffer * inputFramebufferForBlock = firstInputFramebuffer ;
glFinish ();
[ _movieWriterContext useAsCurrentContext ];
[ self renderAtInternalSizeUsingFramebuffer : inputFramebufferForBlock ];
CVPixelBufferRef pixel_buffer = NULL ;
if ([ GPUImageContext supportsFastTextureUpload ])
{
pixel_buffer = renderTarget ;
CVPixelBufferLockBaseAddress ( pixel_buffer , 0 );
}
else
{
CVReturn status = CVPixelBufferPoolCreatePixelBuffer ( NULL , [ self . assetWriterPixelBufferInput pixelBufferPool ], & pixel_buffer );
if (( pixel_buffer == NULL ) || ( status != kCVReturnSuccess ))
{
CVPixelBufferRelease ( pixel_buffer );
return ;
}
else
{
CVPixelBufferLockBaseAddress ( pixel_buffer , 0 );
GLubyte * pixelBufferData = ( GLubyte * ) CVPixelBufferGetBaseAddress ( pixel_buffer );
glReadPixels ( 0 , 0 , videoSize . width , videoSize . height , GL_RGBA , GL_UNSIGNED_BYTE , pixelBufferData );
}
}
runAsynchronouslyOnContextQueue ( _movieWriterContext , ^ {
if ( ! assetWriterVideoInput . readyForMoreMediaData && _encodingLiveVideo )
{
[ inputFramebufferForBlock unlock ];
NSLog ( @"1: Had to drop a video frame: %@" , CFBridgingRelease ( CMTimeCopyDescription ( kCFAllocatorDefault , frameTime )));
return ;
}
// Render the frame with swizzled colors, so that they can be uploaded quickly as BGRA frames
[ _movieWriterContext useAsCurrentContext ];
[ self renderAtInternalSizeUsingFramebuffer : inputFramebufferForBlock ];
CVPixelBufferRef pixel_buffer = NULL ;
if ([ GPUImageContext supportsFastTextureUpload ])
{
pixel_buffer = renderTarget ;
CVPixelBufferLockBaseAddress ( pixel_buffer , 0 );
}
else
{
CVReturn status = CVPixelBufferPoolCreatePixelBuffer ( NULL , [ self . assetWriterPixelBufferInput pixelBufferPool ], & pixel_buffer );
if (( pixel_buffer == NULL ) || ( status != kCVReturnSuccess ))
{
CVPixelBufferRelease ( pixel_buffer );
return ;
}
else
{
CVPixelBufferLockBaseAddress ( pixel_buffer , 0 );
GLubyte * pixelBufferData = ( GLubyte * ) CVPixelBufferGetBaseAddress ( pixel_buffer );
glReadPixels ( 0 , 0 , videoSize . width , videoSize . height , GL_RGBA , GL_UNSIGNED_BYTE , pixelBufferData );
}
}
if ( self . onFramePixelBufferReceived ){
self . onFramePixelBufferReceived ( frameTime , pixel_buffer );
}
[ inputFramebufferForBlock unlock ];
});
}
In newFrameReadyAtTime , we remove the asset write append pixel buffer part and call onFramePixelBufferReceived to notify asset writer to continue next round frame processing.
For THImageMovieManager , we just create some serial queue for video and audio writing, nothing special there.
We’re all good but how to give it run.
Demo Run(Chroma Key)
self . thMovieA = [[ THImageMovie alloc ] initWithURL : rawVideoURL ];
self . thMovieFX = [[ THImageMovie alloc ] initWithURL : fxURL ];
self . filter = [[ GPUImageChromaKeyBlendFilter alloc ] init ];
[ self . thMovieFX addTarget : self . filter ];
[ self . thMovieA addTarget : self . filter ];
NSArray * thMovies = @[ self . thMovieA , self . thMovieFX ];
//setup writer
NSString * pathToMovie = [ NSHomeDirectory () stringByAppendingPathComponent : @"Documents/th_output.mov" ];
unlink ([ pathToMovie UTF8String ]); // If a file already exists, AVAssetWriter won't let you record new frames, so delete the old movie
NSURL * outputURL = [ NSURL fileURLWithPath : pathToMovie ];
self . thMovieWriter = [[ THImageMovieWriter alloc ] initWithMovieURL : outputURL size : CGSizeMake ( 640 . 0 , 640 . 0 ) movies : thMovies ];
[ self . filter addTarget : self . thMovieWriter ];
//---comment those audio one's
//self.movieWriter.shouldPassthroughAudio = YES;
//self.gpuMovieA.audioEncodingTarget = self.movieWriter;
//[self.gpuMovieA enableSynchronizedEncodingUsingMovieWriter:self.movieWriter];
[ self . thMovieA startProcessing ];
[ self . thMovieFX startProcessing ];
[ self . thMovieWriter startRecording ];
__weak typeof ( self ) weakSelf = self ;
[ self . thMovieWriter setCompletionBlock : ^ {
[ weakSelf . thMovieFX endProcessing ];
[ weakSelf . thMovieA endProcessing ];
//no need to call finishRecording
NSLog ( @"movie writing done" );
}];
There you go.
I found this change has resolved all problems above and it runs quite stably(not even with a single crash spotted)
The commit is available on github: Fix movie writer exporting and you could run the sample project called ChromaKeyVideoMerge