使用方法
- 复制安装命令，在自己的智能体环境安装前先检查源码包。
- 从原始文档核对所需工具、凭据和支持的宿主。
- 说明输入材料与期望交付物，并对照来源检查输出。
npx skills add https://github.com/scenario-labs/skills --skill scenario-unity-performance开始前需了解
- 安装不包含模型订阅、API 额度或访问文件的权限。
- 仓库星数和安装量反映活跃情况，不等于独立质量或安全测试。
按源仓库内容呈现，保留标题、案例、代码、表格、链接以及原文引用的演示图片。
Unity performance (performance engineer)
Expert performance work is a measurement discipline: a budget in milliseconds, the bounding thread named before any change, the biggest measured cost fixed first, structure before tuning, and every fix proven with before/after numbers from a development player. An agent does all of it headless: AgentProfile in Editor Play mode to iterate, PerfProbe in a batch-mode player for verdicts, PerfCapture to attribute costs from a Profiler capture, metal_trace for GPU time, PerfBuild for size. Target: Unity 6000.3.21f1, URP 17.3, macOS Apple Silicon. If a sibling skill named here is missing from your available skills, ask the user to install it (npx skills add scenario-labs/skills --skill <name>); unattended, proceed from tool schemas and flag the gap.
REQUIRED BACKGROUND: scenario-unity-expert (channels, review loop, 6.3 traps, ut_* toolkit).
Stance (the expert delta)
- Name the bound before touching anything. "If you speed up the GPU when your profile looks
like this it's not going to run any faster" (Peter Hall, Unity Profiler team, _cV1B2hqXGI [00:05:50]). Observed: removing 1.36 MB of garbage per frame did not move a frame bound by 5,946 SetPass calls. GPU bound needs GPU timing; the Frame Debugger has none (Hall [frame 00:09:18]).
- Players decide, the Editor iterates. The authority is a development build on the lowest
target device (profiling e-book p. 18). GRD: 4.4x in a 2M-instance stress test, 2.07x in the Editor but 1.45x standalone in a real game (Oc6T4hh5gaI [frame 00:12:41]); here the Editor swung 1.5x to 3.0x, players 1.9x to 2.8x on a stress fixture. Forecast about 1.5x.
- GC.Alloc sample time is not the cost. Unity records only timestamp and size (e-book p. 27):
54,021 allocations shown as 1.01 ms caused 19.97 ms of GC.Collect (Hall [frame 00:21:52]). Observed: 27,000 allocations per frame displayed 0.8 to 1.1 ms; the allocating code cost 3.8 to 5.4 ms. Count bytes and calls, measure collections, target 0 B per gameplay frame, and keep the probe itself allocation-free (git-amend ham_w48aRJ4 [00:11:30]).
- 6.3 URP draw calls = SRP Batcher + GPU Resident Drawer; the 2024 advice is Built-in-only.
Material "Enable GPU Instancing" off, static batching off, no MaterialPropertyBlock (6.3 Manual table). GRD silently does nothing unless BatchRendererGroup Variants = Keep All and every renderer of the URP asset is Forward+ or Deferred+ (URP 17.3 source, observed). Then prove coverage: GRD draws are named "Hybrid Batch Group" in the Frame Debugger (GUI editor only), each other draw is a renderer that fell out (MPB, LPPV, realtime GI, non-DOTS shader, skinned; 6.3 Manual); headless, check eligibility per renderer after Start.
- Data layout first; spreading work over frames hides cost. 200,000 objects: MonoBehaviour
Update 27 ms to 0.13 ms through manager, structs, Burst, SoA; "it's just spreading the work around" (Jason Booth, NAVbI1HIzCE [00:06:38]-[00:18:17]).
- Attribute every hitch before fixing it, with markers kept in code. Streaming and spawn
systems carry their own ProfilerMarkers and counters; Deep Profile only finds what fills a gap (Hall [00:18:39]-[00:19:12]). Compiles show as Shader.CreateGPUProgram (warm a traced GraphicsStateCollection during loading, Nicolas Borromeo CmD8MVGkDxQ [00:11:14]-[00:14:35]); here warm-up removed all 18 first-use compiles, yet the first frame stayed about 250 ms on Mono.JIT.
- Resident memory kills, untracked usually does not. 0.75 GB untracked was 140 MB resident
(Borromeo [00:32:22]); Editor memory is about 5x the player's (Uuzd39AjFWQ [00:10:03]). On a laptop, paging and VRAM overflow are stutter causes of their own.
- Counts are exact, times drift [added, observed]. On a shared machine (load average 50 to 210
here), interleave A/B runs and require a consistent sign.
Establish first
Target platforms and min-spec device per tier; frame budget (16.67 ms desktop, mobile x 0.65); resolution and quality tier; what may change (code only, or art, design, feel); a repeatable benchmark, time-based when judging streaming (a fixed step per frame reaches boundaries sooner after a speed-up) [added]; scripting backend and build type; the machine actually measured (discrete GPU, on power; probe JSON device) [added]. Defaults: PC-tier URP asset, 60 fps, 1920x1080, development Mono player for iteration, IL2CPP release for size.
Workflow
- Budget and benchmark. GATE: a baseline CSV.
- Iterate in the Editor.
ut_perf.profile(render=False isolates scripts),ab_profilefor
time deltas. GATE: numbers labeled "Editor, iteration".
- Static baseline.
PerfRendering.Audit(batching, GRD eligibility per renderer, graphics jobs
per platform, Build Profile, stack traces, upload buffer, culling), lint_scripts, Project Auditor (project_auditor, Major first, triage_auditor keeps per-frame lines), PerfAudit.AudioImport, bundle_duplicates. GATE: findings triaged by call frequency.
- Measure the player.
build_player(development=True),
run_player(..., shot=png, markers=[...]). GATE: draw_calls > 0, PNG opened, budget_check.
- Classify the bound. Capture wait markers (
capture_player,analyze_capture), longest vs
median frame. GPU: metal_trace + metal_gpu_frames (GPU busy and per-pass ms; Rendering Debugger Display Stats in the GUI), then the pixel-bound test (4K vs 1080p, render scale). GATE: bound named with a number.
- Attribute.
attribute_spikes(your markers, GC.Collect, compiles, MeshCollider cooking,
synchronous reads, physics steps), physics_catchup, memory_budget, density_check, allocations per block (AllocScope, Is.Not.AllocatingGCMemory()), GRD coverage (PerfFrameDebug.Capture or the window). GATE: top three costs named, no unexplained spike.
- Fix one thing at a time, structure first. Scripts: allocations, managers, then Burst. Render
CPU: Material Variants, SRP Batcher shaders, EnableGpuResidentDrawer, one camera, Split graphics jobs on DX12/Vulkan. GPU: dynamic resolution or STP, depth priming and less transparency, LOD where triangles are pixel-sized, post-processing on a fixed slice; shadows with scenario-unity-rendering-lighting. Hitches: async loads, prewarmed pools, PSO warm-up in loading, baked MeshColliders, audio load types, no duplicated bundle dependencies. GATE: compare_line shows the target down, nothing else up, image parity.
- Ship config. IL2CPP release, stripping and code generation chosen, variants stripped against
the used set (P15), flags through BuildPlayerOptions (platform profiles share settings). GATE: build succeeded, size explained, smoke run clean.
- Guard. Allocation tests and Performance Testing sample groups per commit, lint on changed
scripts. GATE: tests pass.
Numbers
| Value | Relative to | Source |
|---|---|---|
| 16.67 / 33.33 ms; x 0.65 mobile | frame budget at 60 / 30 fps; thermal headroom | profiling e-book p. 9, 36 |
| 0 B | managed allocation per gameplay frame | 6.3 Manual; e-book p. 66 |
| up to about 1 ms | per-frame cost of incremental GC write barriers | e-book p. 69 |
| 4.4x / 2.07x / 1.45x | GRD gain: 2M-instance stress / game in Editor / same game standalone | Oc6T4hh5gaI [frame 00:12:41] |
| 217 / 199 / 176 / 161 fps | DX12 Split / Legacy / plain DX12 / Native graphics jobs | Oc6T4hh5gaI [frame 00:03:52] |
| 256 vertices | below it GPU instancing is inefficient: combine instead | 6.3 Manual |
| about 250 ms | one PSO creation stall | Oc6T4hh5gaI [frame 00:04:48] |
| 4 pixel invocations | per one-pixel triangle (2x2 quads); gate 1 triangle per rendered pixel [added] | CmD8MVGkDxQ [00:27:26] |
| 48.55 of 80 ms | per-frame Debug.Log in a development frame | xjsqv8nj0cw [00:10:23] |
| < 200 / > 400 KB | audio: Decompress On Load / Streaming | console e-book p. 104-105 |
| up to 1 ms | CPU per extra active camera on mobile | console e-book p. 89 |
| up to 10 s | startup added by Autoconnect Profiler | e-book p. 47 |
| 50% / 70% / 80% of RAM | memory budget: 2 GB phones / general / dedicated devices | Hall; Unity tutorial; e-book |
| 5,946 to 33 SetPass; 58-60 to 5.2-6.2 ms | 3,000 spheres, MPB tint to 8 Material Variants, dev player | observed |
| 18 to 0 | first-use Shader.CreateGPUProgram after warming 10 PSOs (3.6 ms at load) | observed |
| 177.9 / 110.5 / 116.3 MB | macOS app: IL2CPP Minimal + speed / IL2CPP High + size / Mono | observed |
| 14.75 M to 0.26 M triangles (7.1 to 0.12 per pixel) | 400 dense meshes after Mesh LOD, draw calls unchanged | observed |
57.1 ms frame = Spawn.Burst 26.9 ms + 4 compiles; next frame 3 physics steps | burst of 3,000 Instantiate read through its own marker | observed |
| 48-80 ms vs 0 ms | MeshCollider cooking at spawn vs Physics.BakeMesh in a job at load (40 meshes) | observed |
| 33 us vs 2.6-3.9 us | per Debug.Log, stack traces ScriptOnly vs None (dev player) | observed |
| 2.5 vs 5.8 ms; 1.2x at 4x pixels | GPU busy vs CPU main (Metal System Trace); 4K vs 1080p: not pixel bound | observed |
| 1,944 to 1,002 KB | two bundles sharing a texture: implicit copies vs a shared bundle | observed |
Quality gates
- Measurable: player
budget_checkpass (p95 under budget, no gameplay hitch);gc_checkpass
in the player; attribute_spikes unexplained = 0; physics_steps at most 1 outside a forced spike; memory_budget pass; density_check pass; PerfRendering.Audit counts.error == 0 and every GRD fallout explained; 0 Shader.CreateGPUProgram after loading; no new Major Project Auditor issue in your code; IL2CPP build within budget.
- Visual: the player's last frame (
-perfShot) and Editor captures from the same bookmarks
before and after, ut_review.compare, contact sheet opened; no blank, uniform or magenta frame.
Common mistakes
| Mistake | What it looks like | Fix |
|---|---|---|
| Optimizing the side that is not bound | fewer allocations, same frame time | classify first (wait markers, GPU busy, SetPass) |
| Trusting Editor numbers or one run | "3x faster" gone in the player | player CSV; ab_profile with consistent sign |
| Benchmarking a batch-mode player that renders nothing | 0.3 ms frames, 0 draw calls | PerfProbe submit + -perfShot |
| GRD on, nothing changes | draw calls unchanged | Keep All, Forward+/Deferred+, static batching off, fallouts fixed |
| MaterialPropertyBlock for per-object color | SetPass = draw calls; StdRender.ApplyShader dominates | Material Variants; unity_RendererUserValue with GRD |
| Reading GC.Alloc ms as the cost | "allocations cost 0.5 ms" | count and bytes, GC.Collect ms, A/B |
| A monitor that formats strings each frame | GC column never 0 | preallocated buffers, write once (PerfProbe) |
| Closure in a rarely taken branch | allocations with the feature off | move captured locals into the branch |
| Deep Profile to find a spike | inflated small functions | markers and counters kept in code, -perfMarkers |
| GPU tuned from the Frame Debugger | guessed timings | metal_trace (PIX, Nsight on Windows) |
| Fixed render scale on a laptop | fine cold, hitches when hot or loaded | dynamic resolution; STP at a fixed scale only for a fixed pixel bound |
| Warm-up assumed to fix a first-frame hitch | still 250 ms | capture: JIT means IL2CPP, loading means async |
| Frame-spreading as the fix | lower peak, same total | measure unamortized first, fix structure |
| MeshCollider added at runtime | spawn spike, Physics.BakePhysXCollisionMeshData | Physics.BakeMesh in a job, Prebake Collision Meshes |
Development flag set through EditorUserBuildSettings | every platform profile changed | BuildPlayerOptions or a build profile asset |
Sizing from BuildReport totalSize | 1.9 GB for a 178 MB app | shipped folder on disk |
InstantiateAsync as a hitch fix | 17 to 83 ms frame later | pool and prewarm behind a transition |
| Benchmarking the first launch of a new build | 212 vs 22 ms | discard the first launch |
What not to change without sign-off
Game feel (fixed timestep, Maximum Allowed Timestep, animation timing, input, AI ticks); art direction (texture sizes, lights, LOD thresholds, render scale) without before/after captures; the pipeline, engine version or an ECS rewrite; incremental GC off unless gameplay allocates 0 B in the player and an A/B shows the gain; asyncUploadBufferSize raised blindly (never returned, console e-book p. 31); VSync off in release; the benchmark path between before and after.
Handoffs
- Receives: the brief in numbers from scenario-unity-expert; systems code from
scenario-unity-architecture and scenario-unity-gameplay; pipeline assets from scenario-unity-rendering-lighting; shaders from scenario-unity-shaders; effects from scenario-unity-vfx; UI from scenario-unity-ui; bundles and Addressables from scenario-unity-pipeline-automation.
- Delivers: the bound, attributed costs, before/after tables, CSVs, traces, PNGs: variant
stripping to scenario-unity-shaders, lighting, shadow and post cuts to scenario-unity-rendering-lighting, overdraw budgets to scenario-unity-vfx and scenario-unity-ui, physics settings to scenario-unity-gameplay, duplicate fixes, CI perf tests and build profiles to scenario-unity-pipeline-automation, device and thermal validation to scenario-unity-mobile, Web limits to scenario-unity-web.
Unity 6.3 notes
- SRP Batcher + GRD + GPU occlusion (the 17.3 URP asset holds the occlusion flag); dynamic batching
deprecated in 6.5, obsolete in 6.6.
- Split graphics jobs is the default on DX12/Vulkan/consoles (Oc6T4hh5gaI [00:03:04]); the 6.3 URP
template reads Windows on/Split, Android off/Native, Metal off/Native; per-platform getters are internal (observed).
GraphicsStateCollectionis experimental, traces only in development players; WebGPU tracing in
6.4. FrameDebuggerUtility is internal.
- Profile Analyzer 1.4.0, Memory Profiler 1.1.12, Project Auditor 1.0.2 (built in from 6.4),
Performance Testing 3.5.0 (core in 6.6) and Profiling Core 1.0.3 (needed for ProfilerCounterValue) are preinstalled packages added by name.
- Mesh LOD is 6.2+.
Physics.BakeMesh(int, bool)is obsolete: passmesh.GetEntityId(). 6.3 runs
Boehm (non-generational); CoreCLR is future work.
References
- `references/procedures.md`: eighteen procedures with full calls, each
with its live test and recorded result.
- `references/expert-notes.md`: principles by expert with timestamps,
disagreements and deciding conditions.
- `references/critique.md`: the rubric for judging a performance pass.
- `references/gui-paths.md`: Profiler, Frame Debugger, Rendering
Debugger, Project Auditor and settings paths.
- `references/sources.md`: every source with credentials, URLs, best
timestamps and revision history.
- `scripts/ut_perf.py` (runner side),
`scripts/AgentKit/Performance/*.cs` (Editor jobs), `scripts/Runtime/Performance/*.cs` (PerfProbe, AllocScope).
