macOS Spatial/Metal 工程师

Architecture for 50K node Metal rendering:

Instanced drawing: All 50K nodes rendered in a single draw call using instanced rendering. Per-instance data (position, color, size) stored in a Metal buffer updated per frame.

Buffer strategy: Triple buffering with MTLResourceStorageModeShared for CPU→GPU data. Layout:

struct NodeInstance {
    var position: SIMD3<Float>  // 12 bytes
    var color: SIMD4<Float>     // 16 bytes
    var scale: Float            // 4 bytes
    // 32 bytes per node × 50,000 = 1.6MB per buffer
}

GPU-based layout: Force-directed layout runs entirely on GPU via compute shader. Each frame: compute forces → update positions → render. CPU only handles interaction events.

Culling: Frustum culling in compute shader discards nodes outside view. With typical camera angles, renders ~15-20K visible nodes per frame.

Performance budget:

• Compute (layout + culling): 3ms
• Vertex + fragment: 2ms
• Total: 5ms per frame → 200fps theoretical, 90fps with stereo + headroom

Edge rendering: Edges as instanced line segments in a separate draw call. Only render edges connected to visible nodes.

Pinch-to-select implementation for spatial graph:

Step 1: Raycast from gaze direction

func handleSpatialTap(_ value: SpatialTapGesture.Value) {
    let ray = value.location3D
    let hitResult = scene.raycast(from: ray.origin, direction: ray.direction)
    if let node = hitResult.first?.entity as? NodeEntity {
        selectNode(node)
    }
}

Step 2: Hit testing against node bounding spheres. Each node has a collision sphere slightly larger than its visual radius (1.5x) for comfortable selection. Test ray-sphere intersection on GPU, return closest hit.

Step 3: Visual feedback. On hover (gaze without pinch): node glows with emission increase. On select (pinch): node scales to 1.2x with spring animation, connected edges highlight.

Step 4: Handle tracking loss. If hand tracking is lost during a pinch gesture, cancel the selection — don't commit a half-gesture. Use gestureState.isActive to detect tracking loss.

Key UX rule: Selection target size must be at least 60pt in angular size at the node's distance. If nodes are too small at current zoom, increase collision radius or require zoom-in first.