fix: eliminate unnecessary buffer allocations during audio playback

packages/csharp/src/AlphaTab.Windows/NAudioSynthOutput.cs

@@ -26,6 +26,7 @@ public class NAudioSynthOutput : WaveProvider32, ISynthOutput, IDisposable
         private int _bufferCount;
         private int _requestedBufferCount;
         private ISynthOutputDevice? _device;
+        private Float32Array? _readWrapper;
 
         /// <inheritdoc />
         public double SampleRate => PreferredSampleRate;
@@ -139,13 +140,18 @@ private void RequestBuffers()
         /// <inheritdoc />
         public override int Read(float[] buffer, int offset, int count)
         {
-            var read = new Float32Array(count);
-
-            var samplesFromBuffer = (int)_circularBuffer.Read(read, 0,
-                System.Math.Min(read.Length, _circularBuffer.Count));
+            // NAudio reuses the same provider buffer across reads, so cache the
+            // Float32Array wrapper to avoid a per-call allocation that otherwise
+            // builds up GC pressure during steady-state playback.
+            var wrapper = _readWrapper;
+            if (wrapper == null || wrapper.Data.Array != buffer)
+            {
+                wrapper = new Float32Array(buffer);
+                _readWrapper = wrapper;
+            }
 
-            Buffer.BlockCopy(read.Data.Array!, read.Data.Offset, buffer, offset * sizeof(float),
-                samplesFromBuffer * sizeof(float));
+            var samplesFromBuffer = (int)_circularBuffer.Read(wrapper, offset,
+                System.Math.Min(count, _circularBuffer.Count));
 
             ((EventEmitterOfT<double>)SamplesPlayed).Trigger(samplesFromBuffer /
                                                              SynthConstants.AudioChannels);

packages/csharp/src/AlphaTab/Core/EcmaScript/Float32Array.cs

@@ -68,7 +68,7 @@ public void Set(Float32Array subarray, double offset)
         System.Buffer.BlockCopy(subarray.Data.Array!,
             subarray.Data.Offset * sizeof(float),
             Data.Array!,
-            Data.Offset + (int)offset * sizeof(float),
+            (Data.Offset + (int)offset) * sizeof(float),
             subarray.Data.Count * sizeof(float));
     }
 

packages/kotlin/src/android/src/main/java/alphaTab/platform/android/AndroidAudioWorker.kt

@@ -2,7 +2,10 @@ package alphaTab.platform.android
 
 import android.media.*
 import java.util.concurrent.*
+import java.util.concurrent.atomic.AtomicLong
 import kotlin.contracts.ExperimentalContracts
+import kotlin.math.max
+import kotlin.math.min
 
 @ExperimentalContracts
 @ExperimentalUnsignedTypes
@@ -61,9 +64,28 @@ internal class AndroidAudioWorker(
                     val samplesFromBuffer = _output.read(_buffer, 0, _buffer.size)
                     if (_previousPosition == -1) {
                         _previousPosition = _track.playbackHeadPosition
+                        _startPosition = _previousPosition
                         _track.getTimestamp(_timestamp)
                     }
-                    _track.write(_buffer, 0, samplesFromBuffer, AudioTrack.WRITE_BLOCKING)
+                    val silenceFloats = _buffer.size - samplesFromBuffer
+                    if (silenceFloats > 0) {
+                        _buffer.fill(0f, samplesFromBuffer, _buffer.size)
+                    }
+                    // write() may return less than requested (or a negative AudioTrack.ERROR_*
+                    // code) when the track is paused/stopped/disconnected mid-write. Only credit
+                    // counters for what actually landed in the track to keep them in sync with
+                    // playbackHeadPosition.
+                    val floatsWritten = _track.write(
+                        _buffer, 0, _buffer.size, AudioTrack.WRITE_BLOCKING
+                    )
+                    if (floatsWritten > 0) {
+                        val realFloatsWritten = min(floatsWritten, samplesFromBuffer)
+                        val silenceFloatsWritten = floatsWritten - realFloatsWritten
+                        _totalFramesWrittenToTrack.addAndGet((floatsWritten / 2).toLong())
+                        if (silenceFloatsWritten > 0) {
+                            _silenceFramesWrittenToTrack.addAndGet((silenceFloatsWritten / 2).toLong())
+                        }
+                    }
                 } else {
                     _playingSemaphore.acquire() // wait for playing to start
                     _playingSemaphore.release() // release semaphore for others
@@ -87,7 +109,14 @@ internal class AndroidAudioWorker(
 
     fun play() {
         if (_track.playState != AudioTrack.PLAYSTATE_PLAYING) {
-            _previousPosition = _track.playbackHeadPosition
+            _previousPosition = -1
+            _startPosition = -1
+            _totalFramesWrittenToTrack.set(0)
+            _silenceFramesWrittenToTrack.set(0)
+            _silenceFramesAccountedAsPlayed = 0
+            _lastTimestampUpdateNanos = -1L
+            _timestamp.nanoTime = 0
+            _timestamp.framePosition = 0
             _track.play()
             _stopped = false
 
@@ -110,14 +139,23 @@ internal class AndroidAudioWorker(
         }
     }
 
-    private var _previousPosition: Int = -1
+    @Volatile private var _previousPosition: Int = -1
+    @Volatile private var _startPosition: Int = -1
+    private val _totalFramesWrittenToTrack = AtomicLong(0)
+    private val _silenceFramesWrittenToTrack = AtomicLong(0)
+    private var _silenceFramesAccountedAsPlayed: Long = 0
     private val _timestamp = AudioTimestamp()
-    private val _lastTimestampUpdate: Long = -1L
+    private var _lastTimestampUpdateNanos: Long = -1L
 
     private fun onUpdatePlayedSamples() {
-        val sinceUpdateInMillis = (System.nanoTime() - _lastTimestampUpdate) / 10e6
-        if (sinceUpdateInMillis >= 10000) {
-            if (!_track.getTimestamp(_timestamp)) {
+        val now = System.nanoTime()
+        val sinceUpdateMs =
+            if (_lastTimestampUpdateNanos == -1L) Long.MAX_VALUE
+            else (now - _lastTimestampUpdateNanos) / 1_000_000L
+        if (sinceUpdateMs >= 10_000L) {
+            if (_track.getTimestamp(_timestamp)) {
+                _lastTimestampUpdateNanos = now
+            } else {
                 _timestamp.nanoTime = 0
                 _timestamp.framePosition = 0
             }
@@ -133,12 +171,35 @@ internal class AndroidAudioWorker(
             return
         }
 
-        val playedSamples = samplePosition - _previousPosition
-        if (playedSamples < 0) {
+        val rawDelta = samplePosition - _previousPosition
+        if (rawDelta < 0) {
             return
         }
-
         _previousPosition = samplePosition
-        _output.onSamplesPlayed(playedSamples)
+
+        val silenceWritten = _silenceFramesWrittenToTrack.get()
+        if (silenceWritten == 0L) {
+            // Happy path: synth has kept the ring buffer fed the entire session — no silence
+            // has ever been queued. Behavior is bit-identical to the pre-fix logic.
+            if (rawDelta > 0) {
+                _output.onSamplesPlayed(rawDelta)
+            }
+            return
+        }
+
+        // Slow path: writer has silence-padded at least once this session. Compensate for
+        // silence the head has now crossed; mathematically equivalent to capping the
+        // cumulative reported count at the cumulative real frames written.
+        val totalWritten = _totalFramesWrittenToTrack.get()
+        val realWritten = totalWritten - silenceWritten
+        val headFromStart = (samplePosition - _startPosition).toLong()
+        val silencePlayedCum = max(0L, headFromStart - realWritten)
+        val silenceCrossedThisTick = silencePlayedCum - _silenceFramesAccountedAsPlayed
+        _silenceFramesAccountedAsPlayed = silencePlayedCum
+
+        val realDelta = rawDelta.toLong() - silenceCrossedThisTick
+        if (realDelta > 0) {
+            _output.onSamplesPlayed(realDelta.toInt())
+        }
     }
 }

packages/kotlin/src/android/src/main/java/alphaTab/platform/android/AndroidSynthOutput.kt

@@ -32,6 +32,7 @@ internal class AndroidSynthOutput(
 
     private lateinit var _audioContext: AndroidAudioWorker
     private lateinit var _circularBuffer: CircularSampleBuffer
+    private var _readWrapper: Float32Array? = null
 
     override val sampleRate: Double
         get() = PreferredSampleRate.toDouble()
@@ -108,12 +109,18 @@ internal class AndroidSynthOutput(
     }
 
     fun read(buffer: FloatArray, offset: Int, sampleCount: Int): Int {
-        val read = Float32Array(sampleCount.toDouble())
-        val actual = _circularBuffer.read(read, 0.0, min(read.length, _circularBuffer.count))
-
-        read.data.copyInto(buffer, offset, 0, sampleCount)
+        // AndroidAudioWorker has one static buffer which is reused, we can cache the read wrapper
+        var wrapper = _readWrapper
+        if (wrapper == null || wrapper.data !== buffer) {
+            wrapper = Float32Array(buffer)
+            _readWrapper = wrapper
+        }
+        val actual = _circularBuffer.read(
+            wrapper,
+            offset.toDouble(),
+            min(sampleCount.toDouble(), _circularBuffer.count)
+        )
         requestBuffers()
-
         return actual.toInt()
     }
 
@@ -122,7 +129,7 @@ internal class AndroidSynthOutput(
     override val sampleRequest: IEventEmitter = EventEmitter()
 
     override suspend fun enumerateOutputDevices(): List<ISynthOutputDevice> {
-        val audioService = context.getSystemService(Context.AUDIO_SERVICE) as AudioManager?
+        val audioService = context.getSystemService(Context.AUDIO_SERVICE) as? AudioManager?
             ?: return List()
 
         return List(