perf: avoid iterator allocations in Frustum.setFromMat4

[willeastcott]

Summary

Frustum.setFromMat4 was destructuring a Float32Array with const [m00, ..., m33] = matrix.data. JavaScript array destructuring uses the iterator protocol, so on every call V8 allocates one { value, done } result object per element (16 per call) plus the iterator object itself.

This function fires per-camera per-layer per-frame, so the iterator-result objects accumulated into measurable GC pressure on a representative scene workload — and even pulled additional V8 HeapNumber boxing into the Plane.normalize subtree downstream.

Replacing the destructuring with explicit indexed reads is behaviour-identical and eliminates the iterator allocations entirely.

Technical details

// before
const [
    m00, m01, m02, m03,
    m10, m11, m12, m13,
    m20, m21, m22, m23,
    m30, m31, m32, m33
] = matrix.data;

// after
const d = matrix.data;
const m00 = d[0], m01 = d[1], m02 = d[2], m03 = d[3];
const m10 = d[4], m11 = d[5], m12 = d[6], m13 = d[7];
const m20 = d[8], m21 = d[9], m22 = d[10], m23 = d[11];
const m30 = d[12], m31 = d[13], m32 = d[14], m33 = d[15];

Measured impact

Captured via V8 sampling heap profiler (node:inspector HeapProfiler.startSampling) at 128 B interval, 1800 frames of a representative scene rendered against NullGraphicsDevice:

Metric	Before	After	Δ
Frustum.setFromMat4 subtree (selfSize incl. children)	98,160 B	36,216 B	−63%
Iterator .next() attribution under this site	~22 KB	0 B	−100%

A targeted microbenchmark of Frustum.setFromMat4 × 20,000 calls drops from ~56 B/iter to ~45 B/iter. The residual is V8 HeapNumber boxing inside the downstream Plane.normalize arithmetic — unrelated to this change.

Public API changes

None.

Test plan

• Existing unit tests pass (npm test)
• Existing examples that exercise area-light culling and shadow frustums render identically on both WebGL2 and WebGPU

[Copilot]

Pull request overview

This PR optimizes a hot path in the math/culling layer by removing Float32Array destructuring from Frustum.setFromMat4, avoiding per-call iterator-related allocations and reducing GC pressure during per-frame frustum updates.

Changes:

• Replaced const [m00, ..., m33] = matrix.data destructuring with explicit indexed reads from matrix.data.
• Introduced a local d alias for matrix.data to keep the indexed reads concise.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[AlexAPPi]

Down with syntactic sugar that harms performance!!!!