From cdf0b20bc7be20b068b917d4c2b6c3df62e11916 Mon Sep 17 00:00:00 2001
From: Alex P <dtk077@gmail.com>
Date: Wed, 27 Aug 2025 22:57:07 +0000
Subject: [PATCH] perf(audio): optimize validation and buffer pool with caching

- Cache max frame size in validation for hot path performance
- Add lock-free per-goroutine buffer cache to reduce contention
---
 internal/audio/buffer_pool.go | 78 ++++++++++++++++++++++++++++++++++-
 internal/audio/validation.go  | 32 +++++++++++++-
 2 files changed, 106 insertions(+), 4 deletions(-)

diff --git a/internal/audio/buffer_pool.go b/internal/audio/buffer_pool.go
index a015487..a6a09d3 100644
--- a/internal/audio/buffer_pool.go
+++ b/internal/audio/buffer_pool.go
@@ -1,13 +1,43 @@
 package audio
 
 import (
+	"runtime"
 	"sync"
 	"sync/atomic"
 	"time"
+	"unsafe"
 
 	"github.com/jetkvm/kvm/internal/logging"
 )
 
+// Lock-free buffer cache for per-goroutine optimization
+type lockFreeBufferCache struct {
+	buffers [4]*[]byte // Small fixed-size array for lock-free access
+}
+
+// Per-goroutine buffer cache using goroutine-local storage
+var goroutineBufferCache = make(map[int64]*lockFreeBufferCache)
+var goroutineCacheMutex sync.RWMutex
+
+// getGoroutineID extracts goroutine ID from runtime stack for cache key
+func getGoroutineID() int64 {
+	b := make([]byte, 64)
+	b = b[:runtime.Stack(b, false)]
+	// Parse "goroutine 123 [running]:" format
+	for i := 10; i < len(b); i++ {
+		if b[i] == ' ' {
+			id := int64(0)
+			for j := 10; j < i; j++ {
+				if b[j] >= '0' && b[j] <= '9' {
+					id = id*10 + int64(b[j]-'0')
+				}
+			}
+			return id
+		}
+	}
+	return 0
+}
+
 type AudioBufferPool struct {
 	// Atomic fields MUST be first for ARM32 alignment (int64 fields need 8-byte alignment)
 	currentSize int64 // Current pool size (atomic)
@@ -81,7 +111,27 @@ func (p *AudioBufferPool) Get() []byte {
 		}
 	}()
 
-	// First try to get from pre-allocated pool for fastest access
+	// Fast path: Try lock-free per-goroutine cache first
+	gid := getGoroutineID()
+	goroutineCacheMutex.RLock()
+	cache, exists := goroutineBufferCache[gid]
+	goroutineCacheMutex.RUnlock()
+
+	if exists && cache != nil {
+		// Try to get buffer from lock-free cache
+		for i := 0; i < len(cache.buffers); i++ {
+			bufPtr := (*unsafe.Pointer)(unsafe.Pointer(&cache.buffers[i]))
+			buf := (*[]byte)(atomic.LoadPointer(bufPtr))
+			if buf != nil && atomic.CompareAndSwapPointer(bufPtr, unsafe.Pointer(buf), nil) {
+				atomic.AddInt64(&p.hitCount, 1)
+				wasHit = true
+				*buf = (*buf)[:0]
+				return *buf
+			}
+		}
+	}
+
+	// Fallback: Try pre-allocated pool with mutex
 	p.mutex.Lock()
 	if len(p.preallocated) > 0 {
 		lastIdx := len(p.preallocated) - 1
@@ -141,7 +191,31 @@ func (p *AudioBufferPool) Put(buf []byte) {
 	// Reset buffer for reuse - clear any sensitive data
 	resetBuf := buf[:0]
 
-	// First try to return to pre-allocated pool for fastest reuse
+	// Fast path: Try to put in lock-free per-goroutine cache
+	gid := getGoroutineID()
+	goroutineCacheMutex.RLock()
+	cache, exists := goroutineBufferCache[gid]
+	goroutineCacheMutex.RUnlock()
+
+	if !exists {
+		// Create new cache for this goroutine
+		cache = &lockFreeBufferCache{}
+		goroutineCacheMutex.Lock()
+		goroutineBufferCache[gid] = cache
+		goroutineCacheMutex.Unlock()
+	}
+
+	if cache != nil {
+		// Try to store in lock-free cache
+		for i := 0; i < len(cache.buffers); i++ {
+			bufPtr := (*unsafe.Pointer)(unsafe.Pointer(&cache.buffers[i]))
+			if atomic.CompareAndSwapPointer(bufPtr, nil, unsafe.Pointer(&buf)) {
+				return // Successfully cached
+			}
+		}
+	}
+
+	// Fallback: Try to return to pre-allocated pool for fastest reuse
 	p.mutex.Lock()
 	if len(p.preallocated) < p.preallocSize {
 		p.preallocated = append(p.preallocated, &resetBuf)
diff --git a/internal/audio/validation.go b/internal/audio/validation.go
index ec9daf5..b2d682c 100644
--- a/internal/audio/validation.go
+++ b/internal/audio/validation.go
@@ -312,12 +312,40 @@ func ValidateAudioFrameFast(data []byte) error {
 	return nil
 }
 
+// Cached constants for ultra-fast validation (initialized once at startup)
+var (
+	maxFrameSizeCache = 8192 // Will be updated from config during init
+)
+
+// init initializes cached validation constants for optimal performance
+//
+//nolint:gochecknoinits // Required for performance-critical config caching
+func init() {
+	// Cache the maximum frame size to avoid function calls in hot paths
+	if config := GetConfig(); config != nil {
+		maxFrameSizeCache = config.MaxAudioFrameSize
+	}
+	// Fallback to safe default if config unavailable
+	if maxFrameSizeCache <= 0 {
+		maxFrameSizeCache = 8192
+	}
+}
+
 // ValidateAudioFrameUltraFast provides zero-overhead validation for ultra-critical paths
 // This function only checks for nil/empty data and maximum size to prevent buffer overruns
 // Use this in hot audio processing loops where every microsecond matters
+//
+// Performance optimizations:
+// - Uses cached max frame size to avoid config function calls
+// - Single branch condition for optimal CPU pipeline efficiency
+// - Inlined length checks for minimal overhead
+//
+//go:inline
 func ValidateAudioFrameUltraFast(data []byte) error {
-	// Only check for catastrophic failures that could crash the system
-	if len(data) == 0 || len(data) > 8192 { // Hard-coded 8KB safety limit
+	// Single optimized check: empty data OR exceeds cached maximum
+	// This branch prediction friendly pattern minimizes CPU pipeline stalls
+	dataLen := len(data)
+	if dataLen == 0 || dataLen > maxFrameSizeCache {
 		return ErrInvalidFrameData
 	}
 	return nil