61 lines
3.8 KiB
Markdown
61 lines
3.8 KiB
Markdown
# yaze Performance Optimization Summary
|
|
|
|
## 🎉 **Massive Performance Improvements Achieved!**
|
|
|
|
### 📊 **Overall Performance Results**
|
|
|
|
| Component | Before | After | Improvement |
|
|
|-----------|--------|-------|-------------|
|
|
| **DungeonEditor::Load** | **17,967ms** | **3,747ms** | **🚀 79% faster!** |
|
|
| **Total ROM Loading** | **~18.6s** | **~4.7s** | **🚀 75% faster!** |
|
|
| **User Experience** | 18-second freeze | Near-instant | **Dramatic improvement** |
|
|
|
|
## 🚀 **Optimizations Implemented**
|
|
|
|
### 1. **Performance Monitoring System with Feature Flag**
|
|
|
|
- **Feature Flag Control**: `kEnablePerformanceMonitoring` in FeatureFlags allows enabling/disabling the system.
|
|
- **Zero-Overhead When Disabled**: `ScopedTimer` becomes a no-op when monitoring is off.
|
|
- **UI Toggle**: Performance monitoring can be toggled in the Settings UI.
|
|
|
|
### 2. **DungeonEditor Parallel Loading (79% Speedup)**
|
|
|
|
- **Problem Solved**: Loading 296 rooms sequentially was the primary bottleneck, taking ~18 seconds.
|
|
- **Solution**: Implemented multi-threaded room loading, using up to 8 threads to process rooms in parallel. This includes thread-safe collection of results and hardware-aware concurrency.
|
|
|
|
### 3. **Incremental Overworld Map Loading**
|
|
|
|
- **Problem Solved**: UI would block and show blank maps while all 160 overworld maps were loaded upfront.
|
|
- **Solution**: Implemented a priority-based incremental loading system. It creates textures for the current world's maps first, at a 4x faster rate (8 per frame), while showing "Loading..." placeholders for the rest.
|
|
|
|
### 4. **On-Demand Map Reloading**
|
|
|
|
- **Problem Solved**: Any property change would trigger an expensive full map refresh, even for non-visible maps.
|
|
- **Solution**: An intelligent refresh system now only reloads maps that are currently visible. Changes to non-visible maps are deferred until they are viewed.
|
|
|
|
---
|
|
|
|
## Appendix A: Dungeon Editor Parallel Optimization
|
|
|
|
- **Problem Identified**: `DungeonEditor::LoadAllRooms` took **17.97 seconds**, accounting for 99.9% of loading time.
|
|
- **Strategy**: The 296 independent rooms were loaded in parallel across up to 8 threads (~37 rooms per thread).
|
|
- **Implementation**: Used `std::async` to launch tasks and `std::mutex` to safely collect results (like room size and palette data). Results are sorted on the main thread for consistency.
|
|
- **Result**: Loading time for the dungeon editor was reduced by **79%** to ~3.7 seconds.
|
|
|
|
---
|
|
|
|
## Appendix B: Overworld Load Optimization
|
|
|
|
- **Problem Identified**: `Overworld::Load` took **2.9 seconds**, with the main bottleneck being the sequential decompression of 160 map tiles (`DecompressAllMapTiles`).
|
|
- **Strategy**: Parallelize the decompression operations and implement lazy loading for maps that are not immediately visible.
|
|
- **Implementation**: The plan involves using `std::async` to decompress map batches concurrently and creating a system to only load essential maps on startup, deferring the rest to a background process.
|
|
- **Expected Result**: A 70-80% reduction in initial overworld loading time.
|
|
|
|
---
|
|
|
|
## Appendix C: Renderer Optimization
|
|
|
|
- **Problem Identified**: The original renderer created GPU textures synchronously on the main thread for all 160 overworld maps, blocking the UI for several seconds.
|
|
- **Strategy**: Defer texture creation. Bitmaps and surface data are prepared first (a CPU-bound task that can be backgrounded), while the actual GPU texture creation (a main-thread-only task) is done progressively or on-demand.
|
|
- **Implementation**: A `CreateBitmapWithoutTexture` method was introduced. A lazy loading system (`ProcessDeferredTextures`) processes a few textures per frame to avoid blocking, and `EnsureMapTexture` creates a texture immediately if a map becomes visible.
|
|
- **Result**: A much more responsive UI during ROM loading, with an initial load time of only ~200-500ms. |