vvvv-custom-nodes（vvvv 自定义节点）

Writing Custom Nodes for vvvv gamma

ProcessNode Pattern — The Core Pattern

Every stateful C# node in vvvv gamma uses [ProcessNode]:

[ProcessNode]
public class MyTransform : IDisposable
{
    private float _lastInput;
    private float _cachedResult;

    /// <summary>
    /// Transforms input values with caching.
    /// </summary>
    public void Update(
        out float result,       // OUT parameters FIRST
        out string error,       // More out params
        float input = 0f,       // Value inputs with defaults AFTER
        bool reset = false)
    {
        error = null;

        if (input != _lastInput || reset)
        {
            _cachedResult = ExpensiveComputation(input);
            _lastInput = input;
        }

        result = _cachedResult; // ALWAYS output cached data
    }

    public void Dispose() { /* cleanup */ }
}

Prerequisite: [ProcessNode] only works if [assembly: ImportAsIs] is set on the assembly. Projects created by vvvv include this automatically. For library-level ImportAsIs with Namespace/Category parameters, see vvvv-node-libraries.

Non-Negotiable Rules

1. [ProcessNode] attribute on every stateful node class
2. No "Node" in the vvvv-visible name — everything in vvvv is already a node, so "Node" suffix is redundant
3. out parameters FIRST, value inputs with defaults AFTER
4. XML comments on class and Update method (shown as tooltip in vvvv)
5. ZERO allocations in Update — no new, no LINQ, cache everything
6. Change detection — only recompute when inputs actually change
7. Always output latest data — even when no work is done, output cached result
8. No unnecessary public members — data flows through Update in/out params only
9. IDisposable for any node holding native/unmanaged resources

Live Reload Behavior

When the .vl document references a .csproj source project, vvvv compiles C# via Roslyn at runtime. On .cs file save, vvvv recompiles and restarts affected nodes:

1. Dispose() is called on the current instance (if IDisposable)
2. The new constructor runs with a fresh NodeContext
3. Update() resumes on the next frame

Implications for node authors:

• Instance fields reset — any state accumulated during runtime (caches, counters, connections) is lost on code change. This is expected.
• Static fields also reset — the entire in-memory assembly is replaced. Do not rely on static state to survive edits.
• Dispose must be thorough — native handles, network connections, and GPU resources must be released. Leaks accumulate across reloads during development.
• Constructor must be fast — it runs each time you save. Defer heavy initialization to the first Update() call using a _needsInit flag.
• The rules about caching and change detection above exist partly because of this: your code runs in a program that never stops. Allocations in Update() cause GC pressure in a 60 FPS loop that may run for hours.

Class Naming vs Node Name

The rule is: users must never see "Node" in vvvv's node browser. How you achieve this:

// Simple: class name IS the node name — no suffix needed
[ProcessNode]
public class Wander { }              // vvvv shows: "Wander"

// Class has "Node" suffix + Name property strips it — also fine
[ProcessNode(Name = "Scan")]
public class ScanNode { }            // vvvv shows: "Scan"

// Completely different internal name — fine when class manages another type
[ProcessNode(Name = "MeshRenderer", Category = "Stride.Rendering.Custom")]
public class CustomMeshRenderer { }  // vvvv shows: "MeshRenderer"

HasStateOutput — Exposing Node Instance

[ProcessNode(HasStateOutput = true)]
public class ParticleSystem
{
    // The node itself becomes an output pin,
    // allowing downstream nodes to call methods on it
    public void Update(out int particleCount, ...) { ... }
}

Alternative: return this from a method to expose the instance.

Pin Visibility

public void Update(
    out Spread<float> result,
    [Pin(Visibility = PinVisibility.OnlyInspector)] out string error,
    float input = 0f,
    [Pin(Visibility = PinVisibility.Optional)] bool advanced = false)
{
    // PinVisibility values:
    // Visible       — always shown (default)
    // Optional      — user can show/hide
    // Hidden        — not visible, only via inspector
    // OnlyInspector — only in inspector panel (use for debug/error outputs)
}

Pin Groups (Collection Inputs)

For Spread inputs with add/remove buttons in vvvv:

public void Update(
    out float result,
    [Pin(Name = "Input", PinGroupKind = PinGroupKind.Collection, PinGroupDefaultCount = 2)]
    Spread<IRenderer?> input)
{ }

DefaultValue for Complex Types

For defaults that cannot be C# literal expressions:

public void Update(
    [DefaultValue(typeof(Color4), "0.1, 0.1, 0.15, 1.0")] Color4 clearColor,
    [DefaultValue(typeof(Int2), "1920, 1080")] Int2 size,
    bool clear = true)
{ }

Constructor Patterns

Simple node (no special context):

public MyNode() { }

Node needing NodeContext (for AppHost, services, Fuse shader graphs):

public MyNode(NodeContext nodeContext)
{
    _nodeContext = nodeContext;
    // Access: nodeContext.AppHost.IsExported, nodeContext.AppHost.Services, etc.
}

For IFrameClock injection, Stride Game access, logging, and service consumption patterns, see services.md.

Change Detection Patterns

Simple — Direct Field Comparison

private float _lastParam;
private Result _cached;

public void Update(out Result result, float param = 0f)
{
    if (param != _lastParam)
    {
        _cached = Compute(param);
        _lastParam = param;
    }
    result = _cached;
}

Multi-Input — Hash Check

private int _lastHash;
private Config _cached;

public void Update(out Config config, float a = 0f, int b = 0, string c = "")
{
    int hash = HashCode.Combine(a, b, c);
    if (hash != _lastHash)
    {
        _cached = new Config(a, b, c);
        _lastHash = hash;
    }
    config = _cached;
}

Reference Types — Identity Check

if (!ReferenceEquals(newBuffer, _lastBuffer))
{
    ProcessBuffer(newBuffer);
    _lastBuffer = newBuffer;
}

Rebuild Flag — For Pipeline/Effect State

When multiple setters can invalidate state:

private bool _needsRebuild = true;

public void SetShader(ShaderStage vs) { _shader = vs; _needsRebuild = true; }

public void Update(out Effect effect)
{
    if (_needsRebuild)
    {
        RebuildPipeline();
        _needsRebuild = false;
    }
    effect = _cachedEffect;
}

Quick Reference

Return-Based Output (Single Output)

When a node has a single primary output, you can return it directly instead of using out:

[ProcessNode]
public class NoiseSteering
{
    private SteeringConfig? _cached;

    public ISteering Update(
        float strength = 2.0f,
        float noiseFrequency = 0.05f,
        int priority = 0)
    {
        if (_cached is null || _cached.Strength != strength ||
            _cached.NoiseFrequency != noiseFrequency || _cached.Priority != priority)
        {
            _cached = new SteeringConfig(strength, noiseFrequency, priority);
        }
        return _cached;
    }
}

Mix return + out when you have one primary output plus additional outputs:

public ReadOnlySpan<ParticleState> Update(
    SimulationConfig config,
    float deltaTime,
    out TimingStats stats)  // Secondary output via out
{
    // Returns primary output, secondary via out
}

Rising Edge Detection (Bang/Trigger)

For boolean inputs that should trigger once (not every frame they're true):

private bool _lastTrigger;

public void Update(out bool triggered, bool trigger = false)
{
    triggered = trigger && !_lastTrigger; // Rising edge only
    _lastTrigger = trigger;
}

Operation Nodes

vvvv auto-generates nodes from all public C# methods — no attribute needed. Don't create [ProcessNode] wrappers for simple methods that just forward calls. Struct Split() methods also become nodes automatically.

Static methods auto-generate nodes — no [ProcessNode] needed:

public static class MathOps
{
    public static float Remap(float value, float inMin = 0f, float inMax = 1f,
                              float outMin = 0f, float outMax = 1f)
    {
        float t = (value - inMin) / (inMax - inMin);
        return outMin + t * (outMax - outMin);
    }
}

When to Use Which

Memory & Performance in Update Loop

• No new keyword in hot paths
• No LINQ (.Where, .Select, .ToList) — hidden allocations via enumerators
• Cache collections — pre-allocate, reuse arrays/lists
• No string concatenation — use StringBuilder if needed, but avoid in hot path
• Vector types: System.Numerics internally for SIMD, Stride.Core.Mathematics at API boundaries
• Zero-cost conversion: Unsafe.As<StrideVector3, NumericsVector3>(ref val)

For custom regions (delegate-based, ICustomRegion, IRegion\<TInlay\>), see regions.md. For advanced patterns (FragmentSelection, Smell, Dynamic Enums, Settings/Split, Pin Name Derivation), see advanced.md. For service consumption (IFrameClock, Game, Logging), see services.md. For working with public channels from C# nodes, see vvvv-channels. For code examples, see examples.md. For starter templates, see templates/.