godot-optimization skill by jame581

Godot Optimization

This skill covers performance optimization for Godot 4.3+ projects in both GDScript and C#. It covers the built-in profiler, draw call reduction, physics tuning, GDScript performance patterns, memory management, object pooling, and a reference table of common bottlenecks.

Related skills: godot-debugging for systematic debugging and profiling, godot-code-review for performance review checklist, export-pipeline for release build optimization, physics-system for collision shapes, layers, and physics body types, 2d-essentials for 2D mesh optimization, particle performance, and draw order tuning, multithreading for moving work off the main thread, mobile-development for mobile performance budgets.

1. Using the Profiler

Frame Time Budget

At 60 fps, the entire frame (update, physics, rendering) must complete in 16.6 ms. At 30 fps the budget is 33.3 ms. Any single system that consumes the majority of that budget is a bottleneck.

Target FPS	Frame budget
120	8.3 ms
60	16.6 ms
30	33.3 ms

Reading Profiler Output

Open Debugger > Profiler, click Start, play through the scenario you want to measure, then click Stop.

Frame Time — total wall-clock time for that frame in milliseconds.
Self — time spent inside that function excluding callees. This is the primary hotspot indicator. A function with a high Self time is doing expensive work directly.
Total — time including all callees. Useful for identifying expensive subtrees.
Calls — call count per frame. A function called thousands of times per frame (even if each call is cheap) can dominate the frame.
Click any function name to jump to its source in the script editor.

# Manual micro-benchmark for a specific block
var start := Time.get_ticks_usec()
_run_expensive_operation()
var elapsed := Time

C#:

Monitors Tab

Debugger > Monitors shows real-time engine metrics while the game is running. Click a monitor name to open a live graph. Key monitors to watch:

Monitor	What to watch for
`Time > FPS`	Below target — frame budget overrun
`Time > Process`	High — `_process()` callbacks are expensive
`Time > Physics Process`	High — `_physics_process()` or physics sim is expensive
`Render > Total Draw Calls`	Above ~500 (mobile) or ~2 000 (desktop) — needs batching
`Render > Video RAM`	Steadily growing — unfreed textures or meshes (memory leak)
`Object > Object Count`	Growing across scene reloads — nodes are not being freed
`Physics 3D > Active Bodies`	Large count in simple scenes — bodies not sleeping

C#:

2. Draw Call Optimization

Every distinct mesh, sprite, or canvas item that cannot be batched with its neighbours costs one draw call. Reducing draw calls is one of the highest-leverage optimisations, especially on mobile — wrap 2D groups sharing a texture in CanvasGroup, keep unique-material count low, atlas sprites, and cull off-screen work.

See references/draw-calls.md for the full recipes (CanvasGroup batching constraints, shared shader-parameter materials, texture atlases, VisibleOnScreenNotifier2D/3D culling, and 3D LOD swapping).

3. Physics Optimization

Physics tuning hinges on minimising broadphase work and avoiding mesh colliders on moving bodies. Trim collision masks to only the layers each body actually needs, replace ConcavePolygonShape3D with primitives on anything that moves, and prefer Area2D/3D over per-frame raycasts for overlap detection.

See references/physics-tuning.md for the full recipes (layer/mask bit examples, collision-shape cost table, Engine.physics_ticks_per_second tuning, Area-vs-raycast patterns).

4. GDScript Performance

Hot-path GDScript wins come from eliminating per-frame allocations, comparing StringName instead of String, using typed arrays / PackedArrays, and preload-ing resources at class scope. The same allocation discipline applies to C# (with List<T> in place of typed Array[T] and static readonly StringName fields).

See references/cpu-bottlenecks.md for the full recipes (cached group queries, reused vector locals, &"..." literals, PackedVector2Array, static typing, preload vs load, plus C# parity blocks).

5. Memory Management

Watch Performance.MEMORY_STATIC and OBJECT_COUNT across scene reloads — steady growth means leaked references. Resources loaded by path are cached and shared; call .duplicate() when you need per-instance mutation. Always prefer queue_free() for nodes; free() inside a self-emitted signal will crash.

See references/memory-management.md for the full recipes (Performance singleton queries, ResourceLoader cache semantics, queue_free vs free table, plus the full GDScript and C# object pool implementations for bullets/effects/particles).

6. Common Bottlenecks

Problem	Diagnosis tool	Fix
Too many draw calls	Debugger > Monitors `Render > Total Draw Calls`; Viewport > Debug > Draw Calls overlay	Use `CanvasGroup` for 2D batching; merge meshes for 3D; use texture atlases; reduce unique materials
Heavy GDScript in `_process`	Profiler > Self column shows script functions at top	Move logic to `_physics_process` (runs less often), cache queries, avoid per-frame allocations, consider C# for tight loops
Excessive signal connections	Profiler shows signal dispatch overhead; manually audit `get_signal_connection_list()`	Remove redundant connections; prefer polling over per-frame signals for high-frequency data; use `CONNECT_ONE_SHOT` for fire-and-forget
Unoptimised TileMap	Profiler shows `TileMap._process` or high draw call count	Split into fewer layers; use a single atlas texture per layer; disable `use_parent_material` if not needed; use `TileMapLayer` (Godot 4.3+) instead of legacy TileMap
Large uncompressed textures	Monitors `Render > Video RAM` is high; check Import dock for texture settings	Enable texture compression (VRAM Compressed) in the Import dock; use mipmaps; halve resolution of assets not viewed up-close
Too many active physics bodies	Monitors `Physics 3D > Active Bodies` is high; slow `_physics_process` in Profiler	Enable sleeping on `RigidBody3D` (`can_sleep = true`); lower physics tick rate; replace distant bodies with fake animations; use layers/masks to narrow collision checks
String operations in hot paths	Profiler shows `String` allocation functions; high GC pressure	Replace `String` comparisons with `StringName` (`&"..."`); avoid `String` formatting in `_process`; build strings once and cache
`instantiate()` in hot paths	Profiler shows `PackedScene.instantiate` with high Self time	Implement object pooling (see `references/memory-management.md`); preload scenes at startup; spawn during loading screens rather than during gameplay

7. Top Anti-Patterns

Allocating in _process — new Arrays, Dictionaries, Strings, or Vector constructors per frame. Cache the container, mutate fields in place. See references/cpu-bottlenecks.md.
Mesh colliders on moving bodies — ConcavePolygonShape3D on a CharacterBody3D or RigidBody3D. Use a capsule, box, or convex hull instead. See references/physics-tuning.md.
Unique materials per instance — material_override = SomeMaterial.new() in _ready() breaks batching. Share one material; vary via shader parameters. See references/draw-calls.md.
load() in hot paths — calling load("res://...") from _process or _physics_process. Use const X := preload(...) at class scope. See references/cpu-bottlenecks.md.
instantiate() + queue_free() for short-lived objects — bullets, hit effects, particles. Pool them. See references/memory-management.md.

8. Checklist

Work through this list before shipping or when investigating a performance complaint.

Profiler

Run the Profiler during the most demanding gameplay scenario.
Confirm no single function's Self time exceeds 30% of the frame budget.
Confirm total frame time stays under budget (16.6 ms at 60 fps).

Draw Calls

Draw call count is within target (≤500 mobile, ≤2 000 desktop).
2D sprite groups that share a texture are wrapped in CanvasGroup.
Textures are atlas-packed where possible; duplicate materials are eliminated.
Off-screen nodes use VisibleOnScreenNotifier2D/3D to pause processing.
3D meshes have LOD enabled via import settings or manual swap logic.

Physics

Collision layers and masks are minimal — no body checks layers it never needs.
No moving body uses ConcavePolygonShape — replaced with capsule, box, or convex.
Physics tick rate is appropriate for the game type (30 Hz may be fine for turn-based or top-down).
Area2D/3D is used for overlap detection instead of per-frame raycasts.
RigidBody3D nodes have can_sleep = true where applicable.

GDScript

No Array, Dictionary, or String is allocated inside _process or _physics_process.
All hot-path string comparisons use StringName (&"...").
All arrays in hot paths are typed (Array[T] or PackedArray).
All function parameters and return types in hot paths are statically typed.
All scene and resource references use preload at class scope, not load per frame.

Memory

Performance.get_monitor(Performance.MEMORY_STATIC) is stable between scene reloads.
Resources that require per-instance mutation are .duplicate()d.
All node removals use queue_free() unless synchronous teardown is explicitly required.

Object Pooling

Bullets, hit effects, particles, and other frequently spawned objects use a pool.
Pool initial size is large enough to avoid runtime growth during normal gameplay.
Pooled objects reset all state on reactivation (position, velocity, signals).