refactor(audio): improve memory management with atomic operations and chunk allocation

- Replace mutex-protected refCount with atomic operations in ZeroCopyFramePool
- Implement chunk-based allocation in AudioBufferPool to reduce allocations
- Add proper reference counting with atomic operations in ZeroCopyAudioFrame
- Optimize buffer pool sizing based on buffer size

internal/audio/util_buffer_pool.go

@@ -354,6 +354,12 @@ type AudioBufferPool struct {
 	// Memory optimization fields
 	preallocated []*[]byte // Pre-allocated buffers for immediate use
 	preallocSize int       // Number of pre-allocated buffers
+	
+	// Chunk-based allocation optimization
+	chunkSize     int      // Size of each memory chunk
+	chunks        [][]byte // Pre-allocated memory chunks
+	chunkOffsets  []int    // Current offset in each chunk
+	chunkMutex    sync.Mutex // Protects chunk allocation
 }
 
 func NewAudioBufferPool(bufferSize int) *AudioBufferPool {
@@ -379,29 +385,74 @@ func NewAudioBufferPool(bufferSize int) *AudioBufferPool {
 		preallocSize = minPrealloc
 	}
 
-	// Pre-allocate with exact capacity to avoid slice growth
+	// Calculate max pool size based on buffer size to prevent memory bloat
-	preallocated := make([]*[]byte, 0, preallocSize)
+	maxPoolSize := 256 // Default
+	if bufferSize > 8192 {
+		maxPoolSize = 64 // Much smaller for very large buffers
+	} else if bufferSize > 4096 {
+		maxPoolSize = 128 // Smaller for large buffers
+	} else if bufferSize > 1024 {
+		maxPoolSize = 192 // Medium for medium buffers
+	}
+
+	// Calculate chunk size - allocate larger chunks to reduce allocation frequency
+	chunkSize := bufferSize * 64 // Each chunk holds 64 buffers worth of memory
+	if chunkSize < 64*1024 {
+		chunkSize = 64 * 1024 // Minimum 64KB chunks
+	}
+
+	p := &AudioBufferPool{
+		bufferSize:   bufferSize,
+		maxPoolSize:  maxPoolSize,
+		preallocated: make([]*[]byte, 0, preallocSize),
+		preallocSize: preallocSize,
+		chunkSize:    chunkSize,
+		chunks:       make([][]byte, 0, 4), // Start with capacity for 4 chunks
+		chunkOffsets: make([]int, 0, 4),
+	}
+
+	// Configure sync.Pool with optimized allocation
+	p.pool.New = func() interface{} {
+		// Use chunk-based allocation instead of individual make()
+		buf := p.allocateFromChunk()
+		return &buf
+	}
 
 	// Pre-allocate buffers with optimized capacity
 	for i := 0; i < preallocSize; i++ {
-		// Use exact buffer size to prevent over-allocation
+		// Use chunk-based allocation to prevent over-allocation
-		buf := make([]byte, 0, bufferSize)
+		buf := p.allocateFromChunk()
-		preallocated = append(preallocated, &buf)
+		p.preallocated = append(p.preallocated, &buf)
 	}
 
-	return &AudioBufferPool{
+	return p
-		bufferSize:   bufferSize,
+}
-		maxPoolSize:  GetConfig().MaxPoolSize * 2, // Double the max pool size for better buffering
+
-		preallocated: preallocated,
+// allocateFromChunk allocates a buffer from pre-allocated memory chunks
-		preallocSize: preallocSize,
+func (p *AudioBufferPool) allocateFromChunk() []byte {
-		pool: sync.Pool{
+	p.chunkMutex.Lock()
-			New: func() interface{} {
+	defer p.chunkMutex.Unlock()
-				// Allocate exact size to minimize memory waste
+	
-				buf := make([]byte, 0, bufferSize)
+	// Try to allocate from existing chunks
-				return &buf
+	for i := 0; i < len(p.chunks); i++ {
-			},
+		if p.chunkOffsets[i]+p.bufferSize <= len(p.chunks[i]) {
-		},
+			// Slice from the chunk
+			start := p.chunkOffsets[i]
+			end := start + p.bufferSize
+			buf := p.chunks[i][start:end:end] // Use 3-index slice to set capacity
+			p.chunkOffsets[i] = end
+			return buf[:0] // Return with zero length but correct capacity
 		}
+	}
+	
+	// Need to allocate a new chunk
+	newChunk := make([]byte, p.chunkSize)
+	p.chunks = append(p.chunks, newChunk)
+	p.chunkOffsets = append(p.chunkOffsets, p.bufferSize)
+	
+	// Return buffer from the new chunk
+	buf := newChunk[0:p.bufferSize:p.bufferSize]
+	return buf[:0] // Return with zero length but correct capacity
 }
 
 func (p *AudioBufferPool) Get() []byte {
@@ -459,10 +510,10 @@ func (p *AudioBufferPool) Get() []byte {
 		// Buffer too small, fall through to allocation
 	}
 
-	// Pool miss - allocate new buffer with exact capacity
+	// Pool miss - allocate new buffer from chunk
 	// Direct miss count update to avoid sampling complexity in critical path
 	atomic.AddInt64(&p.missCount, 1)
-	return make([]byte, 0, p.bufferSize)
+	return p.allocateFromChunk()
 }
 
 func (p *AudioBufferPool) Put(buf []byte) {

internal/audio/zero_copy.go

@@ -147,7 +147,7 @@ func (p *ZeroCopyFramePool) Get() *ZeroCopyAudioFrame {
 		// If we've allocated too many frames, force pool reuse
 		frame := p.pool.Get().(*ZeroCopyAudioFrame)
 		frame.mutex.Lock()
-		frame.refCount = 1
+		atomic.StoreInt32(&frame.refCount, 1)
 		frame.length = 0
 		frame.data = frame.data[:0]
 		frame.mutex.Unlock()
@@ -163,11 +163,12 @@ func (p *ZeroCopyFramePool) Get() *ZeroCopyAudioFrame {
 		p.mutex.Unlock()
 
 		frame.mutex.Lock()
-		frame.refCount = 1
+		atomic.StoreInt32(&frame.refCount, 1)
 		frame.length = 0
 		frame.data = frame.data[:0]
 		frame.mutex.Unlock()
 
+		atomic.AddInt64(&p.hitCount, 1)
 		return frame
 	}
 	p.mutex.Unlock()
@@ -175,7 +176,7 @@ func (p *ZeroCopyFramePool) Get() *ZeroCopyAudioFrame {
 	// Try sync.Pool next and track allocation
 	frame := p.pool.Get().(*ZeroCopyAudioFrame)
 	frame.mutex.Lock()
-	frame.refCount = 1
+	atomic.StoreInt32(&frame.refCount, 1)
 	frame.length = 0
 	frame.data = frame.data[:0]
 	frame.mutex.Unlock()
@@ -191,10 +192,9 @@ func (p *ZeroCopyFramePool) Put(frame *ZeroCopyAudioFrame) {
 		return
 	}
 
+	// Reset frame state for reuse
 	frame.mutex.Lock()
-	frame.refCount--
+	atomic.StoreInt32(&frame.refCount, 0)
-	if frame.refCount <= 0 {
-		frame.refCount = 0
 	frame.length = 0
 	frame.data = frame.data[:0]
 	frame.mutex.Unlock()
@@ -219,15 +219,7 @@ func (p *ZeroCopyFramePool) Put(frame *ZeroCopyAudioFrame) {
 
 	// Return to sync.Pool
 	p.pool.Put(frame)
-		// Metrics collection removed
-		if false {
 	atomic.AddInt64(&p.counter, 1)
-		}
-	} else {
-		frame.mutex.Unlock()
-	}
-
-	// Metrics recording removed - granular metrics collector was unused
 }
 
 // Data returns the frame data as a slice (zero-copy view)
@@ -271,18 +263,28 @@ func (f *ZeroCopyAudioFrame) SetDataDirect(data []byte) {
 	f.pooled = false // Direct assignment means we can't pool this frame
 }
 
-// AddRef increments the reference count for shared access
+// AddRef increments the reference count atomically
 func (f *ZeroCopyAudioFrame) AddRef() {
-	f.mutex.Lock()
+	atomic.AddInt32(&f.refCount, 1)
-	f.refCount++
-	f.mutex.Unlock()
 }
 
-// Release decrements the reference count
+// Release decrements the reference count atomically
-func (f *ZeroCopyAudioFrame) Release() {
+// Returns true if this was the final reference
-	f.mutex.Lock()
+func (f *ZeroCopyAudioFrame) Release() bool {
-	f.refCount--
+	newCount := atomic.AddInt32(&f.refCount, -1)
-	f.mutex.Unlock()
+	if newCount == 0 {
+		// Final reference released, return to pool if pooled
+		if f.pooled {
+			globalZeroCopyPool.Put(f)
+		}
+		return true
+	}
+	return false
+}
+
+// RefCount returns the current reference count atomically
+func (f *ZeroCopyAudioFrame) RefCount() int32 {
+	return atomic.LoadInt32(&f.refCount)
 }
 
 // Length returns the current data length