Sample Lab Scenes

DeviceFit includes three specialized, ready-to-run benchmark scenes located in Assets/Decnet/DeviceFit/Samples/.

These scenes allow you to test and verify the entire 11-step tuning loop in under 10 minutes without risking or modifying your production game assets.

Sample Scenes:
├── Lab_Courtyard_GPUShadows.unity   <── GPU Shading & Shadow Cascade Stress
├── Lab_Interaction_CPUSpikes.unity  <── CPU Raycast & GC Memory Spikes
└── Lab_Memory_TextureWall.unity     <── VRAM & Platform Texture Importer Downscaling

---

1. Lab_Courtyard_GPUShadows.unity

Scene Target: Stress GPU rasterization, high-resolution shadow maps, and distance culling.
Active Route: Courtyard_Tour_Route.asset
Expected Bottleneck: GPU_SHADOWS_AND_RASTER

Scene Architecture

• High-density procedural architecture with intricate pillar geometry.
• Real-time directional sunlight configured with 4 shadow cascades and a 150-meter shadow distance.
• Pre-bound camera route (Courtyard_Tour_Route) with 4 waypoints, including visual comparison checkpoints.

What to Test

1. Open the scene and click Tools > Decnet > DeviceFit > Experiment Desk.
2. Run 1. Capture Baseline to establish reference frame times.
3. Observe the bottleneck diagnosis: GPU_SHADOWS_AND_RASTER (Confidence: HIGH).
4. Run candidate trials to evaluate shadow distance reduction (150m $\to$ 80m) and cascade count reduction (4 $\to$ 2).
5. Compare visual shadow fidelity in the Captures & Comparisons window.

---

2. Lab_Interaction_CPUSpikes.unity

Scene Target: Isolate CPU main thread overhead, batching limits, and GC allocation spikes.
Active Component: CpuSpikeFixtureAdapter
Expected Bottleneck: CPU_MAIN_THREAD / MEMORY_GC_PRESSURE

Scene Architecture

• A dynamic interaction arena featuring the CpuSpikeFixtureAdapter component.
• The fixture generates controlled CPU workloads:
  • Multi-threaded and single-threaded physics Raycast sweeps against dense obstacle colliders.
  • Periodic GC allocation bursts (simulating unpooled particle or string allocation churn).

What to Test

1. In the Experiment Desk, observe that frame times spike intermittently above the $p95$ budget.
2. Review the bottleneck verdict: CPU_MAIN_THREAD or MEMORY_GC_PRESSURE.
3. Test LOD bias adjustments (1.0 $\to$ 0.75) to relieve the CPU render thread.

---

3. Lab_Memory_TextureWall.unity

Scene Target: Verify reversible platform texture downscaling and VRAM reduction.
Procedural Assets: TextureWall_Panel_A.png through Panel_D.png (2048x2048)
Expected Bottleneck: Texture Memory Bandwidth / VRAM Footprint

Scene Architecture

• A gallery of quad surfaces mapped with high-resolution 2048 procedural textures (TextureWall_Panel_A through Panel_D).
• Designed to test TextureTrialBuildService.cs (DF-T05).

What to Test

1. In the Experiment Definition, enable allowTextureLimits = true.
2. Run candidate trials. DeviceFit applies a temporary platform override downscaling the textures to $1024$ and $512$.
3. The asset database re-imports the textures, measures frame times and memory footprint, and then cleanly restores the original 2048 settings with zero manual intervention.