Merge branch 'master' of github.com:limelight-stream/limelight-ios

# By Cameron Gutman
# Via Cameron Gutman
* 'master' of github.com:limelight-stream/limelight-ios:
  Change video background to black
  Don't try to be clever and stall the audio thread while we await samples. Start audio output earlier to allow it some extra warmup time.
  Revert "Use a semaphore to prevent us from having to busy wait on the audio buffer queue"
  Fix the memory leak in the video decoder

Limelight/Connection.m

@@ -11,8 +11,6 @@
 #import <AudioUnit/AudioUnit.h>
 #import <AVFoundation/AVFoundation.h>
 
-#include <dispatch/dispatch.h>
-
 #include "Limelight.h"
 #include "opus.h"
 
@@ -42,7 +40,6 @@ struct AUDIO_BUFFER_QUEUE_ENTRY {
 static short decodedPcmBuffer[512];
 static NSLock *audioLock;
 static struct AUDIO_BUFFER_QUEUE_ENTRY *audioBufferQueue;
-static dispatch_semaphore_t audioQueueSemaphore;
 static int audioBufferQueueLength;
 static AudioComponentInstance audioUnit;
 static VideoDecoderRenderer* renderer;
@@ -65,9 +62,8 @@ void DrSubmitDecodeUnit(PDECODE_UNIT decodeUnit)
             entry = entry->next;
         }
         
+        // This function will take our buffer
         [renderer submitDecodeBuffer:data length:decodeUnit->fullLength];
-        
-        free(data);
     }
 }
 
@@ -95,8 +91,6 @@ void ArInit(void)
     opusDecoder = opus_decoder_create(48000, 2, &err);
     
     audioLock = [[NSLock alloc] init];
-    
-    audioQueueSemaphore = dispatch_semaphore_create(0);
 }
 
 void ArRelease(void)
@@ -112,7 +106,6 @@ void ArRelease(void)
 void ArStart(void)
 {
     printf("Start audio\n");
-    AudioOutputUnitStart(audioUnit);
 }
 
 void ArStop(void)
@@ -141,15 +134,9 @@ void ArDecodeAndPlaySample(char* sampleData, int sampleLength)
                 // Clear all values from the buffer queue
                 struct AUDIO_BUFFER_QUEUE_ENTRY *entry;
                 while (audioBufferQueue != NULL) {
-                    // Unlink the current entry
                     entry = audioBufferQueue;
                     audioBufferQueue = entry->next;
-                    
-                    // Decrease the semaphore count and queue length
                     audioBufferQueueLength--;
-                    dispatch_semaphore_wait(audioQueueSemaphore, DISPATCH_TIME_NOW);
-                    
-                    // Free the entry
                     free(entry);
                 }
             }
@@ -164,13 +151,9 @@ void ArDecodeAndPlaySample(char* sampleData, int sampleLength)
                 }
                 lastEntry->next = newEntry;
             }
-            
-            // Increment the queue size and unlock
             audioBufferQueueLength++;
-            [audioLock unlock];
             
-            // Signal the other thread
+            [audioLock unlock];
-            dispatch_semaphore_signal(audioQueueSemaphore);
         }
     }
 }
@@ -312,6 +295,13 @@ void ClDisplayTransientMessage(char* message)
         NSLog(@"Unable to initialize audioUnit: %d", (int32_t)status);
     }
     
+    // We start here because it seems to need some warmup time
+    // before it starts accepting samples
+    status = AudioOutputUnitStart(audioUnit);
+    if (status) {
+        NSLog(@"Unable to start audioUnit: %d", (int32_t)status);
+    }
+    
     return self;
 }
 
@@ -325,9 +315,15 @@ static OSStatus playbackCallback(void *inRefCon,
     // Fill them up as much as you can. Remember to set the size value in each buffer to match how
     // much data is in the buffer.
     
+    bool ranOutOfData = false;
     for (int i = 0; i < ioData->mNumberBuffers; i++) {
         ioData->mBuffers[i].mNumberChannels = 2;
         
+        if (ranOutOfData) {
+            ioData->mBuffers[i].mDataByteSize = 0;
+            continue;
+        }
+        
         if (ioData->mBuffers[i].mDataByteSize != 0) {
             int thisBufferOffset = 0;
             
@@ -337,14 +333,8 @@ static OSStatus playbackCallback(void *inRefCon,
                 continue;
             }
             
-            // Wait for a buffer to be available
             struct AUDIO_BUFFER_QUEUE_ENTRY *audioEntry = NULL;
-            while (audioEntry == NULL)
-            {
-                // Wait for an entry to be present in the queue
-                dispatch_semaphore_wait(audioQueueSemaphore, DISPATCH_TIME_FOREVER);
             
-                // If there's an entry there, dequeue it
             [audioLock lock];
             if (audioBufferQueue != NULL) {
                 // Dequeue this entry temporarily
@@ -353,6 +343,12 @@ static OSStatus playbackCallback(void *inRefCon,
                 audioBufferQueueLength--;
             }
             [audioLock unlock];
+            
+            if (audioEntry == NULL) {
+                // No data left
+                ranOutOfData = true;
+                ioData->mBuffers[i].mDataByteSize = thisBufferOffset;
+                continue;
             }
             
             // Figure out how much data we can write
@@ -373,7 +369,6 @@ static OSStatus playbackCallback(void *inRefCon,
                 audioBufferQueue = audioEntry;
                 audioBufferQueueLength++;
                 [audioLock unlock];
-                dispatch_semaphore_signal(audioQueueSemaphore);
             }
             else {
                 // This entry is fully depleted so free it

Limelight/VideoDecoderRenderer.m

@@ -23,7 +23,7 @@
     
     displayLayer = [[AVSampleBufferDisplayLayer alloc] init];
     displayLayer.bounds = view.bounds;
-    displayLayer.backgroundColor = [UIColor greenColor].CGColor;
+    displayLayer.backgroundColor = [UIColor blackColor].CGColor;
     displayLayer.position = CGPointMake(CGRectGetMidX(view.bounds), CGRectGetMidY(view.bounds));
     displayLayer.videoGravity = AVLayerVideoGravityResizeAspect;
     [view.layer addSublayer:displayLayer];
@@ -46,6 +46,35 @@
     OSStatus status;
     size_t oldOffset = CMBlockBufferGetDataLength(existingBuffer);
     
+    // If we're at index 1 (first NALU in frame), enqueue this buffer to the memory block
+    // so it can handle freeing it when the block buffer is destroyed
+    if (offset == 1) {
+        int dataLength = nalLength - NALU_START_PREFIX_SIZE;
+        
+        // Pass the real buffer pointer directly (no offset)
+        // This will give it to the block buffer to free when it's released.
+        // All further calls to CMBlockBufferAppendMemoryBlock will do so
+        // at an offset and will not be asking the buffer to be freed.
+        status = CMBlockBufferAppendMemoryBlock(existingBuffer, data,
+                                                nalLength + 1, // Add 1 for the offset we decremented
+                                                kCFAllocatorDefault,
+                                                NULL, 0, nalLength + 1, 0);
+        if (status != noErr) {
+            NSLog(@"CMBlockBufferReplaceDataBytes failed: %d", (int)status);
+            return;
+        }
+        
+        // Write the length prefix to existing buffer
+        const uint8_t lengthBytes[] = {(uint8_t)(dataLength >> 24), (uint8_t)(dataLength >> 16),
+            (uint8_t)(dataLength >> 8), (uint8_t)dataLength};
+        status = CMBlockBufferReplaceDataBytes(lengthBytes, existingBuffer,
+                                               oldOffset, NAL_LENGTH_PREFIX_SIZE);
+        if (status != noErr) {
+            NSLog(@"CMBlockBufferReplaceDataBytes failed: %d", (int)status);
+            return;
+        }
+    }
+    else {
         // Append a 4 byte buffer to this block for the length prefix
         status = CMBlockBufferAppendMemoryBlock(existingBuffer, NULL,
                                                 NAL_LENGTH_PREFIX_SIZE,
@@ -67,7 +96,7 @@
             return;
         }
         
-    // Append the rest of the data by simply attaching a reference to the buffer
+        // Attach the buffer by reference to the block buffer
         status = CMBlockBufferAppendMemoryBlock(existingBuffer, &data[offset+NALU_START_PREFIX_SIZE],
                                                 dataLength,
                                                 kCFAllocatorNull, // Don't deallocate data on free
@@ -77,7 +106,9 @@
             return;
         }
     }
+}
 
+// This function must free data
 - (void)submitDecodeBuffer:(unsigned char *)data length:(int)length
 {
     unsigned char nalType = data[FRAME_START_PREFIX_SIZE] & 0x1F;
@@ -120,21 +151,25 @@
             if (status != noErr) {
                 NSLog(@"Failed to create format description: %d", (int)status);
                 formatDesc = NULL;
-                return;
             }
         }
         
+        // Free the data buffer
+        free(data);
+        
         // No frame data to submit for these NALUs
         return;
     }
     
     if (formatDesc == NULL) {
         // Can't decode if we haven't gotten our parameter sets yet
+        free(data);
         return;
     }
     
     if (nalType != 0x1 && nalType != 0x5) {
         // Don't submit parameter set data
+        free(data);
         return;
     }
     
@@ -144,6 +179,7 @@
     status = CMBlockBufferCreateEmpty(NULL, 0, 0, &blockBuffer);
     if (status != noErr) {
         NSLog(@"CMBlockBufferCreateEmpty failed: %d", (int)status);
+        free(data);
         return;
     }
     
@@ -166,6 +202,8 @@
         [self updateBufferForRange:blockBuffer data:data offset:lastOffset length:length - lastOffset];
     }
     
+    // From now on, CMBlockBuffer owns the data pointer and will free it when it's dereferenced
+    
     CMSampleBufferRef sampleBuffer;
     
     status = CMSampleBufferCreate(kCFAllocatorDefault,
@@ -176,6 +214,7 @@
                                   &sampleBuffer);
     if (status != noErr) {
         NSLog(@"CMSampleBufferCreate failed: %d", (int)status);
+        CFRelease(blockBuffer);
         return;
     }
     
@@ -197,6 +236,10 @@
     }
     
     [displayLayer enqueueSampleBuffer:sampleBuffer];
+    
+    // Dereference the buffers
+    CFRelease(blockBuffer);
+    CFRelease(sampleBuffer);
 }
 
 @end