Skip to content

Latest commit

 

History

History
136 lines (102 loc) · 3.57 KB

performance-optimization.md

File metadata and controls

136 lines (102 loc) · 3.57 KB

Performance Optimization

Overview

Performance optimization system covering rendering, prediction latency, memory management, and monitoring. Target metrics: 60fps rendering, <100ms prediction latency, <150MB RAM.

Key Files

File Class/Function Purpose
src/main/kotlin/tribixbite/cleverkeys/CleverKeysService.kt onCreate(), onDestroy() Lifecycle and cleanup
src/main/kotlin/tribixbite/cleverkeys/Keyboard2View.kt onDraw() Rendering
src/main/kotlin/tribixbite/cleverkeys/OnnxSwipePredictorImpl.kt Neural inference ONNX performance
AndroidManifest.xml hardwareAccelerated GPU rendering

Performance Budget

Operation Target Maximum
Frame render 16ms 33ms (30fps)
Key press latency 30ms 50ms
ONNX inference 50ms 100ms
Layout load 50ms 100ms
Theme switch 100ms 200ms

Implementation Details

Hardware Acceleration

Enabled in AndroidManifest.xml:

<manifest android:hardwareAccelerated="true">
    <application android:hardwareAccelerated="true">
        <!-- All activities inherit true -->
    </application>
</manifest>

Performance Monitoring System

class PerformanceMonitor {
    // Frame timing
    fun startFrameTiming()
    fun recordFrame(durationMs: Long)

    // Operation timing
    fun startOperation(name: String): TimingToken
    fun endOperation(token: TimingToken)

    // Memory tracking
    fun recordMemorySnapshot()

    // Reporting
    fun getPerformanceReport(): PerformanceReport
}

Resource Cleanup (CleverKeysService.onDestroy)

90+ components cleaned up in proper order:

override fun onDestroy() {
    // Cancel all coroutines
    serviceScope.cancel()

    // Stop performance monitoring
    performanceProfiler?.cleanup()

    // Unregister listeners
    prefs.unregisterOnSharedPreferenceChangeListener(this)

    // Clear view references
    inputView = null

    // Release neural resources
    neuralEngine?.release()

    super.onDestroy()
}

Memory Optimization Strategies

Strategy Implementation
Object pooling Reuse Paint, Path, Rect in hot paths
Bitmap caching Cache key background bitmaps
WeakReference Large objects not critical to hold
LRU cache Prediction results
Tensor pooling Pre-allocated buffers for ONNX

Rendering Optimization

  1. Hardware layers for animated views
  2. Cache Paint objects to reduce allocations
  3. Use Canvas.drawBitmap for complex renderings
  4. Avoid overdraw with opaque backgrounds

ONNX Inference Optimization

// OptimizedTensorPool prevents allocation during inference
class OptimizedTensorPool {
    private val featureBuffer = FloatArray(150 * 6)
    private val keyBuffer = LongArray(150)

    fun getFeatureTensor(): OnnxTensor {
        return OnnxTensor.createTensor(env, featureBuffer, shape)
    }
}

Metrics to Track

Category Metrics
Frame Times Min, max, avg, p95, p99
Key Latency Touch to feedback time
Prediction Latency Swipe to suggestions time
Memory Usage Heap size, native memory
GC Events Frequency, duration
ONNX Inference Model load time, inference time

Profiling Tools

Tool Purpose
Android Profiler CPU, memory, network
Systrace System-level analysis
GPU Overdraw Rendering optimization
Layout Inspector View hierarchy
LeakCanary Memory leak detection
Perfetto Advanced tracing