itsjavi Optimize a measured game workload
Read shared context, ready preferences and existing performance evidence. Identify the actual symptom and reproduce it before changing quality. Slow movement may come from speed tuning, input response, collision or animation cadence rather than rendering performance.
Measure a baseline
Use a real graphical run in the selected renderer with a representative camera, actor count, effects and UI. Record device, engine/version, resolution, render scale, vsync or frame cap, warmup and sampling duration. Measure frame-time distribution and spikes as well as average fps. Separate CPU, GPU, memory, loading and input latency where the tools allow it.
Check movement distance over elapsed time and compare it with animation cadence. Headless or fixed-fps tests are useful for simulation invariants but not rendering throughput. Avoid concurrent rendering/encoding jobs during a benchmark. Short desktop samples are diagnostic; sustained thermal and memory checks require the target device and an appropriate run length.
Change the actual bottleneck
- 3D: inspect draw calls/material splits, visible triangle count, skinning, lights/shadows, transparency, textures and post-processing. Preserve silhouette and deformation while simplifying. Reuse or instance compatible geometry and materials.
- 2D: inspect atlas batching, excessive texture switches, transparent overdraw, shader passes, animation updates and oversized textures. Fix native-size filtering and pixel scaling without degrading art unintentionally.
- Tile worlds: profile visible/chunk updates, collision/navigation rebuilds and off-screen work. Test seams and traversal after chunking changes.
- Hybrid scenes: inspect transparent sorting, billboard updates, sprite shadows and mixed-resolution assets.
- Gameplay: profile expensive update loops and unnecessary per-frame work. Retain deterministic timing and input responsiveness.
Set asset budgets from measurements and target requirements, not from a universal polygon number. Reuse source assets and rebuild only affected exports. Record quality tradeoffs, and compare the same scene/camera before and after each meaningful change.
Verify the symptom is improved without introducing visual, movement or collision regressions. Report the measurement conditions and boundary of the conclusion. Stop broadening the optimization when the requested problem and agreed budget are satisfied.