游戏性能瓶颈审计与优化验证
依据性能采样定位游戏帧耗时、内存与加载瓶颈,输出含优化前后对比和质量回归检查的报告。
完整提示词
I want a deep, measurement-driven audit of the game's performance. Main objective: > Identify the real CPU, GPU, memory, IO, asset, simulation and frame-time bottlenecks on the target devices and prove improvements through profiler evidence without degrading visual/gameplay quality. This is not: - "optimize everything" - counting draw calls without GPU context - an assumption that object pooling always helps - an automatic recommendation of a lower resolution - a focus only on average FPS Priority: **frame-time stability > critical gameplay correctness > memory stability > loading/stutter > thermal/sustained performance > average FPS > optimization polish** ## 1. TARGET HARDWARE ## 2. PERFORMANCE TARGET - frame rate - resolution - memory - load time ## 3. FRAME BUDGET At 60 FPS: ~16.67 ms total. At 30 FPS: ~33.33 ms. Use actual target. ## 4. CPU FRAME ## 5. GPU FRAME ## 6. BOTTLENECK CLASS ## 7. P50 ## 8. P95 ## 9. P99 ## 10. FRAME PACING ## 11. STUTTER ## 12. HITCH ## 13. ONE-PERCENT LOW If useful. ## 14. PROFILER Use engine/platform profiler. ## 15. CPU SAMPLE ## 16. INSTRUMENTATION ## 17. GPU CAPTURE ## 18. RENDERDOC/PLATFORM TOOL If applicable. ## 19. DRAW CALL ## 20. BATCHING ## 21. INSTANCING ## 22. STATE CHANGE ## 23. OVERDRAW ## 24. TRANSPARENCY ## 25. SHADER ## 26. SHADER VARIANT ## 27. COMPILE STUTTER ## 28. PIPELINE CACHE ## 29. TEXTURE ## 30. TEXTURE MEMORY ## 31. MIPMAP ## 32. STREAMING ## 33. MODEL ## 34. LOD ## 35. CULLING ## 36. OCCLUSION ## 37. LIGHT ## 38. SHADOW ## 39. POSTPROCESS ## 40. PARTICLE ## 41. UI ## 42. CANVAS/UI REBUILD Engine-specific. ## 43. PHYSICS ## 44. COLLISION ## 45. RIGIDBODY ## 46. RAYCAST ## 47. AI ## 48. PATHFINDING ## 49. ANIMATION ## 50. SKINNING ## 51. AUDIO ## 52. SCRIPT ## 53. ALLOCATION ## 54. GC ## 55. MEMORY ## 56. LEAK ## 57. FRAGMENTATION ## 58. ASSET LIFETIME ## 59. POOL ## 60. SCENE LOAD ## 61. ASYNC LOAD ## 62. DISK IO ## 63. DECOMPRESSION ## 64. NETWORK Multiplayer. ## 65. SERIALIZATION ## 66. JOB SYSTEM ## 67. THREAD ## 68. CONTENTION ## 69. MAIN THREAD ## 70. WORKER ## 71. GPU/CPU SYNC ## 72. READBACK ## 73. V-SYNC ## 74. FRAME CAP ## 75. DYNAMIC RESOLUTION ## 76. QUALITY LEVEL ## 77. THERMAL Mobile/laptop. ## 78. BATTERY ## 79. SUSTAINED LOAD ## 80. MEMORY PRESSURE ## 81. BACKGROUND ## 82. LOW-END DEVICE ## 83. HIGH-END DEVICE ## 84. DIFFERENT DRIVER ## 85. BUILD Development build can distort profiling. ## 86. RELEASE BUILD ## 87. DEBUG OVERHEAD ## 88. WORST-CASE SCENE ## 89. PEAK ENTITY COUNT ## 90. PEAK EFFECTS ## 91. LONG SESSION ## 92. LOAD/UNLOAD LOOP ## 93. TELEPORT ## 94. RAPID SCENE SWITCH ## 95. BENCHMARK ## 96. REPLAYABLE TEST ## 97. OPTIMIZATION PROOF Before/after. ## 98. VISUAL REGRESSION ## 99. GAMEPLAY REGRESSION ## 100. FALSE POSITIVE RULES Do not report: - high draw calls - GC presence - high polygon count - large texture - object pooling absence without measured frame/memory impact. ## 101. EVIDENCE TIERS ```text A - profiler/capture/benchmark proof B - reproducible bottleneck path C - strong measured correlation D - suspected hotspot E - optimization idea ``` ## 102. STATUS ```text CONFIRMED LIKELY NOT VERIFIED CONTROLLED NOT APPLICABLE HARDENING ``` ## 103. SEVERITY P0: performance failure renders critical game state unusable at scale or causes systemic crash/data corruption P1: target hardware cannot sustain required performance or memory P2: material stutter/load/frame issue P3: localized inefficiency P4: optional optimization ## 104. FINDING FORMAT ```text ID: Severity: Status: Evidence tier: Hardware: Build: Scene: Metric: Target: Measured: Profiler evidence: Root cause: Optimization: Before: After: Visual/gameplay delta: Regression test: ``` ## 105. FRAME MATRIX | Scenario | CPU ms | GPU ms | Frame ms | Memory | Target met | |---|---:|---:|---:|---:|---| ## 106. SECOND PASS Profile: - worst scene - low-end target - long session - rapid spawn/despawn - scene load - shader warmup - peak particles - UI-heavy screen - network match - background/resume ## 107. FINAL QUALITY GATE Confirm: - target hardware - frame budget - CPU - GPU - memory - allocation - IO - loading - shader - physics - AI - UI - thermal - release build - long session - before/after proof ## 108. OUTPUT `GAME_PERFORMANCE_AUDIT.md` # FINAL RULE A performance finding without profiler evidence is a hypothesis. An optimization is finished only when: ```text measured bottleneck ↓ targeted change ↓ measured improvement ↓ no unacceptable quality regression ```
分类与标签
编程开发、游戏性能分析、帧时间稳定性、内存优化、性能回归测试