Melonjs Performance Agent Skill

Performance

Measure first

Register the debug plugin and read the numbers before changing anything:

import { plugin } from "melonjs";
import { DebugPanelPlugin } from "@melonjs/debug-plugin";

await app.init();
plugin.register(DebugPanelPlugin, "debugPanel");

It reports frame time, draw calls, object counts, and can overlay collision

shapes and bounds. Draw calls is usually the number that matters.

Batching: the biggest lever

Both GPU backends batch across textures, not just within one: the quad

batcher assigns each source image a texture unit and embeds the unit in the

vertex data, so several different images still flush as a single draw. The pool

is finite — the WebGL batcher gets renderer.maxTextures units (the

maxTextures application setting, "auto" = the device limit capped at 32), the

WebGPU batcher eight per segment. Overflow the pool and the batch flushes.

So the goal is not "one image" but "few state changes". What actually ends a

batch:

• Overflowing the texture-unit pool — pack sprites into a TextureAtlas and

prefer one large tileset over several small ones, so a screenful of sprites

fits inside the pool with room to spare.

• Changing blendMode — setBlendMode flushes before it touches GL state.

Group by blend mode where you can.

• Switching batcher — interleaving primitives (strokeRect, fillRect,

debug shapes) or meshes between sprites flushes the quad batcher and rebinds.

Draw all the sprites, then all the primitives.

• Binding a custom shader — a ShaderEffect on the quad batcher also turns

multi-texture batching *off* for the duration, dropping you back to one

texture per draw.

The advanced blend modes are the expensive ones: overlay, difference,

hard-light, soft-light, color-dodge, color-burn, darken and lighten

cannot be expressed as fixed-function blending, so each such draw is bracketed

on its own — capture the destination, redraw offscreen, composite through a

shader. Right for accents, wrong for hundreds of objects. normal, additive,

multiply, screen and exclusion are plain fixed-function state and cost

nothing extra.

Pooling

Anything spawned frequently — bullets, particles, pickups, damage numbers —

should be pooled rather than allocated:

import { createPool } from "melonjs";

const bulletPool = createPool((x, y) => {
    const instance = new Bullet(x, y);
    return { instance, reset: (x, y) => instance.onResetEvent(x, y) };
});

const b = bulletPool.get(x, y);   // recycled: reset(x, y)
                                  // fresh:    new Bullet(x, y)
world.removeChild(b);             // returns to bulletPool automatically

The container returns a child to whichever pool built it, so removal is enough;

you only call release() yourself for something that was never added to the

world.

**me.pool is the older, string-keyed version and is deprecated since

18.0.0.** pool.register(name, Class, true) / pool.pull(name) still work and

print a one-off console notice, but a recycled object is no longer returned on

removal, so pooling through it now costs an allocation per spawn. To register a

class so a Tiled map can name it, use registerTiledObjectClass — that is a

separate registry and is not deprecated.

Two rules that cause subtle bugs when missed:

• onResetEvent must restore everything the object's lifetime mutates —

alpha, tint, scale, animation, velocity. Anything you forget carries into the

next use, which looks like a random visual glitch.

• Pooled objects do not fire onDestroyEvent on removal. Removal always

calls onDeactivateEvent, then tries to return the object to the pool that

built it; only when nothing owns it does it fall through to destroy(),

which is what calls onDestroyEvent. So a successfully recycled object never

sees it. Pair event subscriptions with onActivateEvent /

onDeactivateEvent instead, or you leak handlers.

• **removeChild is DEFERRED, which breaks a pool you re-lend in the same

frame.** The removal is queued and runs after the update and draw stack has

unwound, so an object you remove and immediately hand out again is torn out

from under its new owner when the queue drains: it vanishes, and the frame

after that the removal throws Child is not mine. Take it out immediately

instead, keeping the instance alive for reuse:

  world.removeChildNow(bullet, true);   // keepalive: do NOT destroy it
  this.freeList.push(bullet);

removeChildNow(child, keepalive) with keepalive true skips the destroy,

which is what you want for anything you own the lifetime of. This is the

usual cause of a pooled effect that flickers out one frame after it is

recycled.

The engine's own classes are NOT in this pool. They were registered in it

once, which bought nothing: most had recycling off, so pool.pull("Sprite")

was a plain construction behind a string key. The ones worth recycling have

typed pools instead, reached by key and fully typed:

const tween = getPool("tween").get(target);
const label = getPool("text").get(x, y, { font: "Arial", size: 12 });
getPool("text").release(label);          // yours to release, as with particles

A recycled label comes back with its base-class state reset as well as its

own: floating, blendMode, alpha, tint, the transform, the anchor, the

behaviour flags, any mask and any post effect, and depth back to 0. The

settings object names none of those, so set them again after get(). A pooled

HUD number that is not re-marked floating is drawn through the camera, i.e.

somewhere out in the world rather than on the screen, and addChild(label, z)

is what gives it its depth back.

getPool covers vector2d, vector3d, point, matrix2d, matrix3d,

bounds, color, polygon, line, rectangle, roundedRectangle,

ellipse, tween, particle, text, colorLayer and bitmapTextData.

pool.register still aliases every name you give it under an me. prefix,

pointing at the same entry, so pool.pull("me.Bullet") resolves identically —

and the same alias reaches the Tiled object factory, which is why a map

authored against melonJS 1.x still finds its classes. Prefer the unprefixed

name in new code; do not "correct" an me.-prefixed one, it is not broken.

scale() is multiplicative

A pooled sprite arrives carrying its previous life's transform, and calling

scale() each frame compounds. rotate() is the same. For absolute values,

reset first:

sprite.currentTransform.identity();
sprite.currentTransform.scale(s, s, 1);

Order matters, and it is the opposite of what reads naturally.

Matrix2d.scale multiplies the matrix ROWS, so it applies AFTER whatever the

matrix already holds. Scale first and then rotate; rotate first and the scale

lands in the rotated frame and SHEARS anything non-uniform:

// correct: a long thin quad, rotated
t.identity();
t.scale(length / texW, width / texH);
t.rotate(angle);

// wrong: the same quad arrives as a parallelogram
t.identity();
t.rotate(angle);
t.scale(length / texW, width / texH);

The symptom is a sprite that looks right at 0 and 90 degrees and skews in

between, which reads as a bad texture rather than as a transform bug.

Culling and update cost

• Container recomputes inViewport for every child each frame against every

active camera, and skips both its draw() and its update(dt) when it is

off-screen. That gate is inViewport || alwaysUpdate, and it applies to every

container in every game, whichever physics you use.

• On top of that, the built-in physics adapter gates body integration and

narrow-phase collision on the same inViewport || alwaysUpdate — so an

off-screen body stops simulating even if something else calls its update.

The planck and matter adapters have no such gate: they step the whole world

regardless of the camera.

• Set alwaysUpdate = true on anything that must keep moving off-screen, and

accept the cost. On many objects it removes the saving entirely.

• floating = true does not disable culling — it moves the test into screen

space: a floating renderable is checked against the camera's screen rectangle

(0, 0, camera.width, camera.height) instead of the world view. That is why a

HUD stays put as the camera moves. Container.draw then draws floating

children unconditionally, but update() is still gated, so a floating object

parked outside the screen rectangle stops updating.

Bake once, draw many

When you would otherwise issue hundreds of identical draws every frame, render

once into a CanvasRenderTarget and draw the result as a single sprite. This is

the sanctioned idiom before reaching for a custom shader — hundreds of draws

collapse to one.

Instancing in 3D

InstancedMesh draws one mesh many times in a single call — the difference

between a hundred trees and a hundred thousand. glTF scenes using

EXT_mesh_gpu_instancing load as an InstancedMesh automatically.

It trades per-object control for the draw call: the set gets ONE depth sort key

and ONE ground shadow, and removeInstance swaps the last instance into the

hole so held indices go stale. Scenery yes; anything the game removes or

queries one at a time, usually not. See melonjs-3d for the decision table.

Update loops

• update(dt) should return true only when something changed. Returning

true unconditionally forces redraw work every frame.

• Use dt for motion rather than assuming a fixed step — the engine already

paces it, and a fixed assumption breaks on high-refresh displays.

• Use timer.setTimeout / setInterval, which are pause-aware. window timers

keep firing behind a pause menu.

Audio and assets

• stream: true on long music tracks plays through a streaming element instead

of decoding the whole file into memory.

• pool on an audio clip sizes the reuse pool for finished instances — it is

not a concurrency cap.

• loader.unloadAll() between large levels drops every loaded asset and

destroy()s the shader programs the loader owns — that last part is a real

GPU release. Images, by contrast, only lose their loader-side reference; the

renderer's texture cache is not purged by unloading.

The first frame of a scene is not like the others

A freeze the first time a level appears, and never again, is almost always

shader linking, not your scene build. Measure before assuming: procedural

geometry is usually trivial next to it.

Programs are created lazily. The mesh batcher keys a variant on the feature

combination it is asked for (lit, instanced, instance colours, instance data,

fog) and links it on the first draw that needs it — and compileProgram calls

getProgramParameter(LINK_STATUS) right after linkProgram, which blocks until

the driver finishes. That whole cost lands inside one frame.

Since 20.6 the engine's own programs are built during preload() instead —

loader.preload() calls renderer.prewarm() itself, on every call, when the

application's prewarm setting is on, which is the default. So on a current

version this is one cause you can rule out before looking further, **unless the

game does not use the preloader at all**, which is the one way to miss it — see melonjs-loading-assets. Since 20.10 the

pipelines that depend on the renderable (effects, blend modes, render targets)

are warmed when an effect is attached or a renderable is added to the world, as

long as that happens at least a frame before it is drawn: a pool built at stage

start is covered, an object created and drawn in the same frame is not. Since

20.10.1 a state change also holds the new stage's first frame until what its

onResetEvent built is warmed, for at most 1.5s, so the opening frame no longer

compiles them itself: a few milliseconds once the browser has compiled them

before, longer on a first visit. None of this covers the rest of what a first

frame pays for, below.

How to tell them apart, all from the page:

// 1. is it the build, or the draw?  time your own setup
const t0 = performance.now(); buildLevel(); console.log(performance.now() - t0);

// 2. how long is the bad frame?  rAF deltas, not averages
let last = performance.now();
const tick = () => { const n = performance.now();
  if (n - last > 40) console.log("long frame", (n - last).toFixed(0));
  last = n; requestAnimationFrame(tick); };

// 3. is it programs?  count them, and when
const gl = WebGL2RenderingContext.prototype, link = gl.linkProgram;
gl.linkProgram = function (p) { console.count("linkProgram"); return link.call(this, p); };

The tell is that the second entry into the same scene is cheap: one measured

case went 852ms of stall (worst frame 479ms) on first entry to 277ms (worst

102ms) on the second.

**Careful with the conclusion, though — "the programs were cached" is the

tempting reading and it is wrong.** Prewarming the engine's own programs ahead

of that first entry moves the worst frame by about 3ms, measured. So the 575ms

is not program linking: it is everything else the first *submit* pays for —

texture residency, buffer uploads, and the driver specializing a pipeline

against real bindings — all of which the second entry finds already done.

That distinction decides which lever to reach for. Linking is what prewarm

covers and what shrinking your shader count helps. The rest only goes away if

the geometry is actually drawn once beforehand.

Three things follow.

• Cut the number of programs. Identical ShaderEffects each link their own,

so preload one as a "shader" asset and share it.

• Declare effects as assets, not inline, so the loader can prepare them

before a scene draws — an inline effect does not exist until the stage that

builds it runs.

• Pay what remains behind the loading screen by drawing the scene once there

and discarding it (see melonjs-loading-assets). Linking is only part of a

first frame: texture residency, buffer uploads and the driver specializing

against real bindings all land on the first *submit*, and only an actual draw

front-loads those. A measured case on a fast desktop GPU had shader prewarming

move the worst frame by ~3ms while a warm-up draw moved it by hundreds.

Do not profile this on a software rasterizer — headless Chromium falls back to

SwiftShader, where the absolute numbers are wildly pessimistic and the JS

profile shows the time as (program), outside JS entirely. Ratios transfer;

milliseconds do not.

Render targets, and what hdr costs

A render target is a canvas-sized GPU surface. The engine allocates them

lazily and only where a multi-pass chain actually needs one, which is worth

knowing before assuming an effect is expensive:

• a renderable with no post effects: none
• a renderable with exactly one post effect: still none. It takes the

customShader fast path and draws straight into whatever target is already

bound

• a camera with effects: a pair (one capture, one ping-pong intermediate)
• a sprite with two or more effects: its own pair

So the first lever is the count, not the cost of any one effect: going from

one effect to two on a sprite is what allocates, not going from a cheap shader

to an expensive one.

hdr: true changes the format of that camera pair from 8 bits per channel to

half-float, which is 8 bytes per pixel instead of 4. At 1920×1080 that is

about 8 MB per target instead of 4 MB, so roughly +8 MB for the pair. With

antiAlias on, the capture half is also multisampled and its multisampled

surface takes the same format, so that part doubles too.

Nothing else moves. Sprite chains and the advanced-blend target stay 8-bit

whatever hdr says, for a correctness reason rather than a thrifty one: their

final blit composites with ONE_MINUS_SRC_ALPHA, and additive blending

accumulates alpha past 1 in a float target, which would turn that factor

negative and subtract the backdrop.

With hdr off, nothing changes at all — every target keeps the 8-bit

attachment it has always had, and a game that never adds a camera post-effect

allocates no render targets in the first place. The setting costs nothing

until something is actually using it.

// what you actually got, which is not always what you asked for
if (app.renderer.supportsHDR === false) { /* grade for a chain that clamps */ }

Where the costs actually are

In rough order for a typical 2D game:

• Draw calls — fix with atlases and fewer tilesets
• Overdraw from large blended sprites
• Per-frame allocation — fix with pooling
• alwaysUpdate on objects that do not need it
• The engine itself, which is rarely the problem

Symptom → cause

| symptom | cause |

|---|---|

| high draw-call count | the texture-unit pool overflows, or primitives/meshes are interleaved with sprites |

| draw calls jump when one effect is enabled | a custom shader on the quad batcher disables multi-texture batching |

| frame time spikes and GC pauses | allocating per frame instead of pooling |

| a recycled object looks wrong | onResetEvent does not restore every mutated property |

| sprite grows every frame | scale() is multiplicative — reset the transform |

| off-screen enemies stop moving | update() and built-in physics both gate on inViewport; set alwaysUpdate |

| a HUD element vanishes | floating is culled against the screen rect, and it sits outside it |

| everything updates even when idle | update() returning true unconditionally |

| memory grows with long music | missing stream: true |

| slow only on some machines | Canvas fallback — no GPU tilemap path |

| a freeze the first time a level shows, never again | shader variants linking on first draw — warm them behind the loading screen |

| a freeze the first time an effect appears mid-play (a shield, a hit flash) | the effect attached in the same frame it is drawn — attach it when the pool is built, so the warm-up has a frame to finish |

| the same scene stutters once per entry | a derived texture rebuilt per stage: the texture cache keys on the image object, so each rebuild is a new upload |

| steady movement stutters while the frame rate reads 60 | not a cost: frame-time jitter gives a frame no update and the next two; set syncToDisplay: true (see melonjs-display-and-scaling) |

Related skills

• melonjs-sprites-and-animation — atlases and pooling in detail
• melonjs-renderer-backends — why performance differs per machine
• melonjs-3d — InstancedMesh