claude-dev-suite Unity Performance
Tools
- Profiler (Window > Analysis > Profiler): CPU / GPU / Memory / Rendering / Audio / Physics.
- Frame Debugger (Window > Analysis > Frame Debugger): step through draw calls, see batching breaks.
- Memory Profiler package: snapshots, diff, find leaks.
- Profile Analyzer package: aggregate frames across captures.
- Use Deep Profile sparingly — adds overhead but shows method-level hotspots.
Build for profiling
Make a Development Build with Autoconnect Profiler + Deep Profiling Support. Don't profile editor — numbers are misleading.
GC allocation hunting
Common culprits:
| Cause | Fix |
|---|---|
string concatenation in Update |
StringBuilder reused, or only build when value changes |
List<T>.ToArray() per frame |
Cache list, reuse |
foreach over some collections (boxing) |
for loop over IList<T> |
GetComponent in Update |
Cache in Awake |
LINQ (.Where, .Select) per frame |
Manual loops |
| Lambda capturing locals | Convert to method or pre-allocate delegate |
Read GC Alloc column in Profiler hierarchy — anything > 0 in Update is suspect.
Object pooling
private readonly IObjectPool<Bullet> _pool = new ObjectPool<Bullet>(
createFunc: () => Object.Instantiate(prefab),
actionOnGet: b => b.gameObject.SetActive(true),
actionOnRelease: b => b.gameObject.SetActive(false),
actionOnDestroy: Object.Destroy,
defaultCapacity: 32, maxSize: 256);
public void Fire() {
var b = _pool.Get();
b.transform.position = muzzle.position;
b.OnExpired = () => _pool.Release(b);
}
UnityEngine.Pool.ObjectPool<T> since Unity 2021.1. For SO-driven pools or ECS, use NativeArray-backed structures.
Burst + Jobs
[BurstCompile]
public struct ApplyGravityJob : IJobParallelFor {
public NativeArray<float3> Velocities;
public float DeltaTime;
public void Execute(int i) {
Velocities[i] += new float3(0, -9.81f * DeltaTime, 0);
}
}
void Update() {
var job = new ApplyGravityJob { Velocities = _velocities, DeltaTime = Time.deltaTime };
var handle = job.Schedule(_velocities.Length, 64);
handle.Complete();
}
Burst compiles to highly optimized native; works only on struct jobs with NativeArray/blittable types.
IL2CPP build settings
- IL2CPP Code Generation: "Faster runtime" for shipping; "Faster (smaller) builds" for iteration.
- Managed Stripping Level: Low/Medium for safety; High only with thorough QA +
link.xmlfor reflection-only types. - Engine Code Stripping: enable to drop unused subsystems.
Batching & instancing
| Technique | When |
|---|---|
| Static Batching | Mark static + same material → bake at build |
| Dynamic Batching | Auto for very small meshes; off in URP by default |
| GPU Instancing | Same mesh + same shader, slight per-instance variation via MaterialPropertyBlock |
| SRP Batcher | URP/HDRP default — same shader variant batched into one SetPass |
| GPU Resident Drawer (Unity 6) | Auto-instancing for static objects with shared mesh+mat |
LOD
LOD 0 = full mesh (close)
LOD 1 = half tris (medium)
LOD 2 = quarter tris (far)
LOD 3 = imposter or culled
Use LODGroup with screen-relative transition heights. Combine with Occlusion Culling for indoors.
Anti-patterns
| Anti-pattern | Fix |
|---|---|
| Profiling editor numbers | Profile Development Build only |
Camera.main per frame |
Cache once |
new Vector3(...) in Update with constants |
Make static readonly |
transform.GetChild(i).gameObject repeatedly |
Cache child references |
Per-frame string.Format for HUD text |
Update only on value change |
| Profiler with deep profile shipped to QA | Only enable for short investigations |
Production checklist
- Frame budget hit on target devices (e.g. 16ms PC, 33ms mobile)
- GC alloc per frame ≤ ~0 in steady state
- Object pools for bullets/VFX/UI fly-outs
- Burst-compiled jobs for hot math
- LODGroup + Occlusion Culling tuned
- Shader variants pre-warmed
- Memory snapshot baseline + post-soak captured & diffed