Implement JobSystem for parallel task management and integrate with LocalToWorldSystem

- Introduced LabelMetrics and RollingLabelMetrics structs to capture detailed profiling data.
- Implemented a RollingWindow class to maintain a rolling average of metrics.
- Updated Frame class to collect and store label metrics, including memory allocation and garbage collection statistics.
- Enhanced ScopeNode class to support self-time calculations and child duration tracking.
- Improved the Profiler class to manage frame budgets and rolling metrics.
- Refactored the ProfilerVisualizer to support new timeline rendering features, including zoom and pan functionality.
- Added tooltip support for detailed node information in the flame graph.
- Enhanced BoundedQueue to be thread-safe with proper locking mechanisms for concurrent access.

Nerfed.Editor/Systems/EditorProfilerWindow.cs

@@ -1,6 +1,7 @@
 using ImGuiNET;
 using MoonTools.ECS;
 using Nerfed.Runtime;
+using System.Numerics;
 
 namespace Nerfed.Editor.Systems
 {
@@ -12,7 +13,10 @@ namespace Nerfed.Editor.Systems
 
         private int selectedFrame = 0;
         private int previousSelectedFrame = -1;
-        private IOrderedEnumerable<KeyValuePair<string, (double ms, uint calls)>> orderedCombinedData = null;
+        private IOrderedEnumerable<KeyValuePair<string, (double ms, double selfMs, uint calls, double avgMs, double p95Ms, long allocBytes, long selfAllocBytes)>> orderedCombinedData = null;
+        private IOrderedEnumerable<KeyValuePair<int, Profiler.RollingThreadMetrics>> orderedThreadRollingData = null;
+        private readonly ProfilerVisualizer.TimelineState timelineState = new ProfilerVisualizer.TimelineState();
+        private readonly List<Profiler.Frame> frameSnapshot = new List<Profiler.Frame>(256);
 
         public EditorProfilerWindow(World world) : base(world)
         {
@@ -25,49 +29,108 @@ namespace Nerfed.Editor.Systems
                 return;
             }
 
+            if (Profiler.CopyFramesTo(frameSnapshot) <= 0)
+            {
+                return;
+            }
+
+            selectedFrame = Math.Clamp(selectedFrame, 0, frameSnapshot.Count - 1);
+            timelineState.SelectedFrameIndex = Math.Clamp(timelineState.SelectedFrameIndex, -1, frameSnapshot.Count - 1);
+            timelineState.VisibleFrameCount = Math.Clamp(timelineState.VisibleFrameCount, 1, frameSnapshot.Count);
+
             ImGui.Begin("Profiler");
 
-            ImGui.BeginChild("Toolbar", new System.Numerics.Vector2(0, 0), ImGuiChildFlags.AutoResizeY);
+            ImGui.BeginChild("Toolbar", new Vector2(0, 0), ImGuiChildFlags.AutoResizeY);
             if (ImGui.RadioButton("Recording", Profiler.IsRecording))
             {
                 Profiler.SetActive(!Profiler.IsRecording);
             }
 
             ImGui.SameLine();
+            ImGui.Text("Follow");
+            ImGui.SameLine();
+            ImGui.Checkbox("##follow-timeline", ref timelineState.FollowLatest);
+            ImGui.SameLine();
+
+            int visibleFrameCount = timelineState.VisibleFrameCount;
+            ImGui.SetNextItemWidth(130f);
+            if (ImGui.SliderInt("Window", ref visibleFrameCount, 1, frameSnapshot.Count))
+            {
+                timelineState.VisibleFrameCount = visibleFrameCount;
+                timelineState.FollowLatest = false;
+            }
+
+            ImGui.SameLine();
+            if (ImGui.Button("Reset Zoom"))
+            {
+                timelineState.Zoom = 1f;
+                timelineState.PanTicks = 0d;
+                timelineState.FollowLatest = true;
+            }
+
+            ImGui.SameLine();
+            int mode = (int)Profiler.Mode;
+            ImGui.SetNextItemWidth(130f);
+            if (ImGui.Combo("Mode", ref mode, "Instrumented\0Sampled\0"))
+            {
+                Profiler.Mode = (Profiler.CaptureMode)mode;
+            }
+
+            ImGui.SameLine();
+            int stride = Profiler.SamplingStride;
+            ImGui.SetNextItemWidth(90f);
+            if (ImGui.SliderInt("Stride", ref stride, 1, 64))
+            {
+                Profiler.SamplingStride = stride;
+            }
 
             if (Profiler.IsRecording)
             {
                 // Select last frame when recording to see latest frame data.
-                selectedFrame = Profiler.Frames.Count - 1;
+                selectedFrame = frameSnapshot.Count - 1;
-            }
-            if (ImGui.SliderInt(string.Empty, ref selectedFrame, 0, Profiler.Frames.Count - 1))
-            {
-                // Stop recording when browsing frames.
-                Profiler.SetActive(false);
             }
 
-            Profiler.Frame frame = Profiler.Frames.ElementAt(selectedFrame);
+            int sliderFrame = selectedFrame;
+            if (ImGui.SliderInt("Frame", ref sliderFrame, 0, frameSnapshot.Count - 1))
+            {
+                selectedFrame = sliderFrame;
+                timelineState.SelectedFrameIndex = selectedFrame;
+                timelineState.FollowLatest = false;
+            }
+
+            Profiler.Frame frame = frameSnapshot[selectedFrame];
             double ms = frame.ElapsedMilliseconds();
             double s = 1000;
             ImGui.Text($"Frame: {frame.FrameCount} ({ms:0.000} ms | {(s / ms):0} fps)");
+            ImGui.Text($"Budget: {frame.BudgetMilliseconds:0.00} ms ({(frame.OverBudget ? "over" : "within")})");
+            ImGui.Text($"Thread Budget: {Profiler.ThreadBudgetMilliseconds:0.00} ms | Capture: {Profiler.Mode}");
+            ImGui.Text($"Alloc: {frame.AllocatedBytesDelta / 1024d:0.0} KB | GC: G0 {frame.Gen0CollectionsDelta}, G1 {frame.Gen1CollectionsDelta}, G2 {frame.Gen2CollectionsDelta}");
             ImGui.EndChild();
 
-            if (!Profiler.IsRecording) {
+            ProfilerVisualizer.TimelineRenderResult timelineResult = DrawFlameGraph(frameSnapshot, timelineState);
-                if (previousSelectedFrame != selectedFrame)
+            if (timelineResult.SelectionChanged)
-                {
+            {
-                    previousSelectedFrame = selectedFrame;
+                selectedFrame = timelineResult.SelectedFrameIndex;
-                    orderedCombinedData = CalculateCombinedData(frame);
-                }
-
-                DrawFlameGraph(frame);
-
-                DrawHierachy(frame);
-
-                ImGui.SameLine();
-
-                DrawCombined(orderedCombinedData);
             }
 
+            selectedFrame = Math.Clamp(selectedFrame, 0, frameSnapshot.Count - 1);
+            frame = frameSnapshot[selectedFrame];
+
+            if (previousSelectedFrame != selectedFrame)
+            {
+                previousSelectedFrame = selectedFrame;
+                orderedCombinedData = CalculateCombinedData(frame);
+                orderedThreadRollingData = CalculateThreadRollingData();
+            }
+
+            DrawThreadRolling(orderedThreadRollingData);
+
+            DrawHierachy(frame);
+
+            ImGui.SameLine();
+
+            DrawCombined(orderedCombinedData);
+
             ImGui.End();
         }
 
@@ -78,13 +141,18 @@ namespace Nerfed.Editor.Systems
                 return;
             }
 
-            ImGui.BeginChild("Hierachy", new System.Numerics.Vector2(150, 0), ImGuiChildFlags.ResizeX);
+            ImGui.BeginChild("Hierachy", new Vector2(150, 0), ImGuiChildFlags.ResizeX);
 
-            if (ImGui.BeginTable("ProfilerData", 3, tableFlags, new System.Numerics.Vector2(0, 0)))
+            if (ImGui.BeginTable("ProfilerData", 8, tableFlags, new Vector2(0, 0)))
             {
-                ImGui.TableSetupColumn("name", ImGuiTableColumnFlags.WidthStretch, 0.8f, 0);
+                ImGui.TableSetupColumn("name", ImGuiTableColumnFlags.WidthStretch, 0.30f, 0);
-                ImGui.TableSetupColumn("thread", ImGuiTableColumnFlags.WidthStretch, 0.2f, 1);
+                ImGui.TableSetupColumn("category", ImGuiTableColumnFlags.WidthStretch, 0.12f, 1);
-                ImGui.TableSetupColumn("ms", ImGuiTableColumnFlags.WidthStretch, 0.2f, 1);
+                ImGui.TableSetupColumn("tags", ImGuiTableColumnFlags.WidthStretch, 0.08f, 2);
+                ImGui.TableSetupColumn("thread", ImGuiTableColumnFlags.WidthStretch, 0.08f, 3);
+                ImGui.TableSetupColumn("ms", ImGuiTableColumnFlags.WidthStretch, 0.10f, 4);
+                ImGui.TableSetupColumn("self", ImGuiTableColumnFlags.WidthStretch, 0.10f, 5);
+                ImGui.TableSetupColumn("alloc(B)", ImGuiTableColumnFlags.WidthStretch, 0.11f, 6);
+                ImGui.TableSetupColumn("self alloc", ImGuiTableColumnFlags.WidthStretch, 0.11f, 7);
                 ImGui.TableSetupScrollFreeze(0, 1); // Make row always visible
                 ImGui.TableHeadersRow();
 
@@ -115,10 +183,29 @@ namespace Nerfed.Editor.Systems
                 isOpen = ImGui.TreeNodeEx(node.Label, treeNodeFlags);
             }
 
+            ImGui.TableNextColumn();
+            ImGui.Text($"{node.Category}");
+            ImGui.TableNextColumn();
+            ImGui.Text($"0x{node.TagMask:X}");
             ImGui.TableNextColumn();
             ImGui.Text($"{node.ManagedThreadId}");
             ImGui.TableNextColumn();
             ImGui.Text($"{node.ElapsedMilliseconds():0.000}");
+            ImGui.TableNextColumn();
+            ImGui.Text($"{node.SelfMilliseconds():0.000}");
+            ImGui.TableNextColumn();
+            if (node.ProfilerSetupBytes > 0)
+            {
+                ImGui.Text($"{node.AllocatedBytes} !");
+                if (ImGui.IsItemHovered())
+                    ImGui.SetTooltip($"{node.ProfilerSetupBytes} B is profiler warmup overhead");
+            }
+            else
+            {
+                ImGui.Text($"{node.AllocatedBytes}");
+            }
+            ImGui.TableNextColumn();
+            ImGui.Text($"{node.SelfAllocatedBytes()}");
 
             if (isOpen)
             {
@@ -130,24 +217,29 @@ namespace Nerfed.Editor.Systems
             }
         }
 
-        private static void DrawCombined(in IOrderedEnumerable<KeyValuePair<string, (double ms, uint calls)>> orderedCombinedData)
+        private static void DrawCombined(in IOrderedEnumerable<KeyValuePair<string, (double ms, double selfMs, uint calls, double avgMs, double p95Ms, long allocBytes, long selfAllocBytes)>> orderedCombinedData)
         {
             if(orderedCombinedData == null)
             {
                 return;
             }
 
-            ImGui.BeginChild("Combined", new System.Numerics.Vector2(0, 0));
+            ImGui.BeginChild("Combined", new Vector2(0, 0));
 
-            if (ImGui.BeginTable("ProfilerCombinedData", 3, tableFlags, new System.Numerics.Vector2(0, 0)))
+            if (ImGui.BeginTable("ProfilerCombinedData", 8, tableFlags, new Vector2(0, 0)))
             {
-                ImGui.TableSetupColumn("name", ImGuiTableColumnFlags.WidthStretch, 0.6f, 0);
+                ImGui.TableSetupColumn("name", ImGuiTableColumnFlags.WidthStretch, 0.32f, 0);
-                ImGui.TableSetupColumn("ms", ImGuiTableColumnFlags.WidthStretch, 0.2f, 1);
+                ImGui.TableSetupColumn("ms", ImGuiTableColumnFlags.WidthStretch, 0.11f, 1);
-                ImGui.TableSetupColumn("calls", ImGuiTableColumnFlags.WidthStretch, 0.2f, 2);
+                ImGui.TableSetupColumn("self", ImGuiTableColumnFlags.WidthStretch, 0.11f, 2);
+                ImGui.TableSetupColumn("calls", ImGuiTableColumnFlags.WidthStretch, 0.08f, 3);
+                ImGui.TableSetupColumn("avg", ImGuiTableColumnFlags.WidthStretch, 0.09f, 4);
+                ImGui.TableSetupColumn("p95", ImGuiTableColumnFlags.WidthStretch, 0.09f, 5);
+                ImGui.TableSetupColumn("alloc(B)", ImGuiTableColumnFlags.WidthStretch, 0.10f, 6);
+                ImGui.TableSetupColumn("self alloc", ImGuiTableColumnFlags.WidthStretch, 0.10f, 7);
                 ImGui.TableSetupScrollFreeze(0, 1); // Make row always visible
                 ImGui.TableHeadersRow();
 
-                foreach (KeyValuePair<string, (double ms, uint calls)> combinedData in orderedCombinedData)
+                foreach (KeyValuePair<string, (double ms, double selfMs, uint calls, double avgMs, double p95Ms, long allocBytes, long selfAllocBytes)> combinedData in orderedCombinedData)
                 {
                     ImGui.TableNextRow();
                     ImGui.TableNextColumn();
@@ -155,7 +247,17 @@ namespace Nerfed.Editor.Systems
                     ImGui.TableNextColumn();
                     ImGui.Text($"{combinedData.Value.ms:0.000}");
                     ImGui.TableNextColumn();
+                    ImGui.Text($"{combinedData.Value.selfMs:0.000}");
+                    ImGui.TableNextColumn();
                     ImGui.Text($"{combinedData.Value.calls}");
+                    ImGui.TableNextColumn();
+                    ImGui.Text($"{combinedData.Value.avgMs:0.000}");
+                    ImGui.TableNextColumn();
+                    ImGui.Text($"{combinedData.Value.p95Ms:0.000}");
+                    ImGui.TableNextColumn();
+                    ImGui.Text($"{combinedData.Value.allocBytes}");
+                    ImGui.TableNextColumn();
+                    ImGui.Text($"{combinedData.Value.selfAllocBytes}");
                 }
 
                 ImGui.EndTable();
@@ -164,41 +266,78 @@ namespace Nerfed.Editor.Systems
             ImGui.EndChild();
         }
 
-        private static IOrderedEnumerable<KeyValuePair<string, (double ms, uint calls)>> CalculateCombinedData(Profiler.Frame frame)
+        private static IOrderedEnumerable<KeyValuePair<string, (double ms, double selfMs, uint calls, double avgMs, double p95Ms, long allocBytes, long selfAllocBytes)>> CalculateCombinedData(Profiler.Frame frame)
         {
-            Dictionary<string, (double ms, uint calls)> combinedRecordData = new Dictionary<string, (double ms, uint calls)>(128);
+            IReadOnlyDictionary<string, Profiler.RollingLabelMetrics> rollingData = Profiler.GetRollingLabelMetricsSnapshot();
-            foreach (Profiler.ScopeNode node in frame.RootNodes)
+            Dictionary<string, (double ms, double selfMs, uint calls, double avgMs, double p95Ms, long allocBytes, long selfAllocBytes)> combinedRecordData = new Dictionary<string, (double ms, double selfMs, uint calls, double avgMs, double p95Ms, long allocBytes, long selfAllocBytes)>(128);
+            foreach (KeyValuePair<string, Profiler.LabelMetrics> metric in frame.LabelMetrics)
             {
-                CalculateCombinedData(node, in combinedRecordData);
+                Profiler.RollingLabelMetrics rolling = default;
+                if (rollingData.TryGetValue(metric.Key, out Profiler.RollingLabelMetrics found))
+                {
+                    rolling = found;
+                }
+
+                combinedRecordData[metric.Key] = (metric.Value.InclusiveMs, metric.Value.SelfMs, metric.Value.Calls, rolling.AverageMs, rolling.P95Ms, metric.Value.AllocatedBytes, metric.Value.SelfAllocatedBytes);
             }
             return combinedRecordData.OrderByDescending(x => x.Value.ms);
         }
 
-        private static void CalculateCombinedData(Profiler.ScopeNode node, in Dictionary<string, (double ms, uint calls)> combinedRecordData)
+        private static IOrderedEnumerable<KeyValuePair<int, Profiler.RollingThreadMetrics>> CalculateThreadRollingData()
         {
-            if (combinedRecordData.TryGetValue(node.Label, out (double ms, uint calls) combined))
+            IReadOnlyDictionary<int, Profiler.RollingThreadMetrics> rollingData = Profiler.GetRollingThreadMetricsSnapshot();
-            {
+            return rollingData.OrderByDescending(x => x.Value.P95Ms);
-                combinedRecordData[node.Label] = (combined.ms + node.ElapsedMilliseconds(), combined.calls + 1);
-            }
-            else
-            {
-                combinedRecordData.Add(node.Label, (node.ElapsedMilliseconds(), 1));
-            }
-
-            for (int i = 0; i < node.Children.Count; i++)
-            {
-                CalculateCombinedData(node.Children[i], combinedRecordData);
-            }
         }
 
-        private static void DrawFlameGraph(Profiler.Frame frame)
+        private static void DrawThreadRolling(in IOrderedEnumerable<KeyValuePair<int, Profiler.RollingThreadMetrics>> orderedThreadRollingData)
         {
-            if (frame == null)
+            if (orderedThreadRollingData == null)
             {
                 return;
             }
 
-            ProfilerVisualizer.RenderFlameGraph(frame);
+            ImGui.BeginChild("ThreadRolling", new Vector2(0, 140), ImGuiChildFlags.Border);
+            if (ImGui.BeginTable("ProfilerThreadRollingData", 6, tableFlags, new Vector2(0, 0)))
+            {
+                ImGui.TableSetupColumn("thread", ImGuiTableColumnFlags.WidthStretch, 0.15f, 0);
+                ImGui.TableSetupColumn("avg", ImGuiTableColumnFlags.WidthStretch, 0.20f, 1);
+                ImGui.TableSetupColumn("p95", ImGuiTableColumnFlags.WidthStretch, 0.20f, 2);
+                ImGui.TableSetupColumn("max", ImGuiTableColumnFlags.WidthStretch, 0.20f, 3);
+                ImGui.TableSetupColumn("samples", ImGuiTableColumnFlags.WidthStretch, 0.15f, 4);
+                ImGui.TableSetupColumn("misses", ImGuiTableColumnFlags.WidthStretch, 0.15f, 5);
+                ImGui.TableHeadersRow();
+
+                foreach (KeyValuePair<int, Profiler.RollingThreadMetrics> metric in orderedThreadRollingData)
+                {
+                    ImGui.TableNextRow();
+                    ImGui.TableNextColumn();
+                    ImGui.Text($"T{metric.Key}");
+                    ImGui.TableNextColumn();
+                    ImGui.Text($"{metric.Value.AverageMs:0.000}");
+                    ImGui.TableNextColumn();
+                    ImGui.Text($"{metric.Value.P95Ms:0.000}");
+                    ImGui.TableNextColumn();
+                    ImGui.Text($"{metric.Value.MaxMs:0.000}");
+                    ImGui.TableNextColumn();
+                    ImGui.Text($"{metric.Value.Samples}");
+                    ImGui.TableNextColumn();
+                    ImGui.Text($"{metric.Value.BudgetMisses}");
+                }
+
+                ImGui.EndTable();
+            }
+
+            ImGui.EndChild();
+        }
+
+        private static ProfilerVisualizer.TimelineRenderResult DrawFlameGraph(IReadOnlyList<Profiler.Frame> frames, ProfilerVisualizer.TimelineState timelineState)
+        {
+            if (frames == null || frames.Count == 0)
+            {
+                return default;
+            }
+
+            return ProfilerVisualizer.RenderTimeline(frames, timelineState);
         }
     }
 }

Nerfed.Runtime/Engine.cs

@@ -71,6 +71,8 @@ public static class Engine
 
         AudioDevice = new AudioDevice();
 
+        JobSystem.Default.Initialize();
+
         OnInitialize?.Invoke();
 
         while (!quit)
@@ -84,6 +86,7 @@ public static class Engine
         MainWindow.Dispose();
         GraphicsDevice.Dispose();
         AudioDevice.Dispose();
+        JobSystem.Default.Shutdown();
         SDL.SDL_Quit();
     }
 

Nerfed.Runtime/JobSystem.cs

@@ -0,0 +1,109 @@
+using System.Diagnostics;
+
+namespace Nerfed.Runtime;
+
+public sealed class JobSystem : IDisposable
+{
+    public static readonly JobSystem Default = new JobSystem();
+
+    private Thread[] workers;
+    private SemaphoreSlim startSignal;
+    private CountdownEvent completionEvent;
+    private volatile bool running;
+
+    // Shared per-dispatch state written by main thread before workers wake.
+    private volatile Action<int> currentAction;
+    private int workCount;
+    private int nextIndex; // grabbed with Interlocked.Increment for work-stealing
+
+    public int WorkerCount => workers?.Length ?? 0;
+    public bool IsInitialized => workers != null;
+
+    public void Initialize(int threadCount = -1)
+    {
+        if (IsInitialized)
+        {
+            throw new InvalidOperationException("JobSystem is already initialized. Call Shutdown first.");
+        }
+
+        threadCount = threadCount < 0
+            ? Math.Max(1, Environment.ProcessorCount - 2)
+            : Math.Max(1, threadCount);
+
+        running = true;
+        startSignal = new SemaphoreSlim(0);
+        completionEvent = new CountdownEvent(1);
+        workers = new Thread[threadCount];
+
+        for (int i = 0; i < threadCount; i++)
+        {
+            workers[i] = new Thread(WorkerLoop)
+            {
+                IsBackground = true,
+                Name = $"Job-{i}",
+            };
+            workers[i].Start();
+        }
+    }
+
+    public void Dispatch(int count, Action<int> action)
+    {
+        if (count <= 0) return;
+
+        if (!IsInitialized)
+        {
+            // Safe fallback: run inline if Initialize was never called.
+            for (int i = 0; i < count; i++)
+                action(i);
+            return;
+        }
+
+        currentAction = action;
+        workCount = count;
+        Volatile.Write(ref nextIndex, 0);
+
+        completionEvent.Reset(workers.Length);
+        startSignal.Release(workers.Length);
+        completionEvent.Wait();
+
+        currentAction = null;
+    }
+
+    public void Shutdown()
+    {
+        if (!IsInitialized) return;
+
+        running = false;
+        startSignal.Release(workers.Length); // wake all workers so they can see running=false and exit
+
+        foreach (Thread t in workers)
+            t.Join();
+
+        completionEvent.Dispose();
+        startSignal.Dispose();
+        workers = null;
+    }
+
+    public void Dispose() => Shutdown();
+
+    private void WorkerLoop()
+    {
+        while (true)
+        {
+            startSignal.Wait();
+            if (!running) return;
+
+            Action<int> action = currentAction;
+            int total = workCount;
+
+            while (true)
+            {
+                int index = Interlocked.Increment(ref nextIndex) - 1;
+                if (index >= total) break;
+                action(index);
+            }
+
+            completionEvent.Signal();
+        }
+    }
+}

Nerfed.Runtime/Profiler.cs

@@ -10,6 +10,11 @@ public struct ProfilerScope : IDisposable
         Profiler.BeginSample(label);
     }
 
+    public ProfilerScope(string label, string category, ulong tagMask = 0)
+    {
+        Profiler.BeginSample(label, category, tagMask);
+    }
+
     public void Dispose()
     {
         Profiler.EndSample();
@@ -18,18 +23,289 @@ public struct ProfilerScope : IDisposable
 
 public static class Profiler
 {
-    public class Frame(uint frameCount)
+    public enum CaptureMode
     {
-        public uint FrameCount { get; } = frameCount;
+        Instrumented = 0,
-        public long StartTime { get; } = Stopwatch.GetTimestamp();
+        SampledInstrumentation = 1,
+    }
+
+    private sealed class ThreadProfilerState
+    {
+        public readonly Stack<ScopeNode> Scopes = new Stack<ScopeNode>();
+        public readonly Stack<bool> CaptureDecisions = new Stack<bool>();
+        public int SampleCursor;
+        public int ThreadId;
+    }
+
+    public readonly struct LabelMetrics
+    {
+        public LabelMetrics(double inclusiveMs, double selfMs, uint calls, double minInclusiveMs, double maxInclusiveMs, long allocatedBytes, long selfAllocatedBytes)
+        {
+            InclusiveMs = inclusiveMs;
+            SelfMs = selfMs;
+            Calls = calls;
+            MinInclusiveMs = minInclusiveMs;
+            MaxInclusiveMs = maxInclusiveMs;
+            AllocatedBytes = allocatedBytes;
+            SelfAllocatedBytes = selfAllocatedBytes;
+        }
+
+        public double InclusiveMs { get; }
+        public double SelfMs { get; }
+        public uint Calls { get; }
+        public double MinInclusiveMs { get; }
+        public double MaxInclusiveMs { get; }
+        public long AllocatedBytes { get; }
+        public long SelfAllocatedBytes { get; }
+    }
+
+    public readonly struct RollingLabelMetrics
+    {
+        public RollingLabelMetrics(double averageMs, double minMs, double maxMs, double p95Ms, int samples)
+        {
+            AverageMs = averageMs;
+            MinMs = minMs;
+            MaxMs = maxMs;
+            P95Ms = p95Ms;
+            Samples = samples;
+        }
+
+        public double AverageMs { get; }
+        public double MinMs { get; }
+        public double MaxMs { get; }
+        public double P95Ms { get; }
+        public int Samples { get; }
+    }
+
+    public readonly struct ThreadMetrics
+    {
+        public ThreadMetrics(double inclusiveMs, double selfMs, uint calls, bool overBudget)
+        {
+            InclusiveMs = inclusiveMs;
+            SelfMs = selfMs;
+            Calls = calls;
+            OverBudget = overBudget;
+        }
+
+        public double InclusiveMs { get; }
+        public double SelfMs { get; }
+        public uint Calls { get; }
+        public bool OverBudget { get; }
+    }
+
+    public readonly struct RollingThreadMetrics
+    {
+        public RollingThreadMetrics(double averageMs, double p95Ms, double maxMs, int samples, int budgetMisses)
+        {
+            AverageMs = averageMs;
+            P95Ms = p95Ms;
+            MaxMs = maxMs;
+            Samples = samples;
+            BudgetMisses = budgetMisses;
+        }
+
+        public double AverageMs { get; }
+        public double P95Ms { get; }
+        public double MaxMs { get; }
+        public int Samples { get; }
+        public int BudgetMisses { get; }
+    }
+
+    private sealed class RollingWindow
+    {
+        private readonly double[] values;
+        private readonly double[] sortBuffer; // pre-allocated; avoids per-snapshot heap allocation
+        private int index;
+        private int count;
+
+        public RollingWindow(int capacity)
+        {
+            int size = Math.Max(8, capacity);
+            values = new double[size];
+            sortBuffer = new double[size];
+        }
+
+        public void Add(double value)
+        {
+            values[index] = value;
+            index = (index + 1) % values.Length;
+            if (count < values.Length)
+            {
+                count++;
+            }
+        }
+
+        public RollingLabelMetrics Snapshot()
+        {
+            if (count == 0)
+            {
+                return default;
+            }
+
+            double sum = 0;
+            double min = double.MaxValue;
+            double max = double.MinValue;
+
+            int start = (index - count + values.Length) % values.Length;
+            for (int i = 0; i < count; i++)
+            {
+                double value = values[(start + i) % values.Length];
+                sortBuffer[i] = value;
+                sum += value;
+                min = Math.Min(min, value);
+                max = Math.Max(max, value);
+            }
+
+            Array.Sort(sortBuffer, 0, count);
+            int percentileIndex = (int)Math.Ceiling((count - 1) * 0.95d);
+            double p95 = sortBuffer[Math.Clamp(percentileIndex, 0, count - 1)];
+            return new RollingLabelMetrics(sum / count, min, max, p95, count);
+        }
+    }
+
+    private sealed class RollingThreadWindow
+    {
+        private readonly double[] durations;
+        private readonly double[] sortBuffer; // pre-allocated; avoids per-snapshot heap allocation
+        private readonly byte[] misses;
+        private int index;
+        private int count;
+
+        public RollingThreadWindow(int capacity)
+        {
+            int size = Math.Max(8, capacity);
+            durations = new double[size];
+            sortBuffer = new double[size];
+            misses = new byte[size];
+        }
+
+        public void Add(double durationMs, bool budgetMiss)
+        {
+            durations[index] = durationMs;
+            misses[index] = budgetMiss ? (byte)1 : (byte)0;
+            index = (index + 1) % durations.Length;
+            if (count < durations.Length)
+            {
+                count++;
+            }
+        }
+
+        public RollingThreadMetrics Snapshot()
+        {
+            if (count == 0)
+            {
+                return default;
+            }
+
+            double sum = 0;
+            double max = double.MinValue;
+            int budgetMisses = 0;
+
+            int start = (index - count + durations.Length) % durations.Length;
+            for (int i = 0; i < count; i++)
+            {
+                int at = (start + i) % durations.Length;
+                double value = durations[at];
+                sum += value;
+                max = Math.Max(max, value);
+                budgetMisses += misses[at];
+                sortBuffer[i] = value;
+            }
+
+            Array.Sort(sortBuffer, 0, count);
+            int percentileIndex = (int)Math.Ceiling((count - 1) * 0.95d);
+            double p95 = sortBuffer[Math.Clamp(percentileIndex, 0, count - 1)];
+            return new RollingThreadMetrics(sum / count, p95, max, count, budgetMisses);
+        }
+    }
+
+    public class Frame
+    {
+        private readonly List<ScopeNode> rootNodes = new List<ScopeNode>(8);
+        private readonly object rootNodesLock = new object();
+        private readonly Dictionary<string, LabelMetrics> labelMetrics = new Dictionary<string, LabelMetrics>(128, StringComparer.Ordinal);
+        private readonly Dictionary<string, LabelMetrics> categoryMetrics = new Dictionary<string, LabelMetrics>(32, StringComparer.Ordinal);
+        private readonly Dictionary<int, ThreadMetrics> threadMetrics = new Dictionary<int, ThreadMetrics>(16);
+        private readonly List<int> knownThreadIds = new List<int>(16); // avoids Keys.ToArray() in ApplyThreadBudgetFlags
+
+        public uint FrameCount { get; private set; }
+        public long StartTime { get; private set; }
         public long EndTime { get; private set; }
 
-        // Use a concurrent list to collect all thread root nodes per frame.
+        public IReadOnlyList<ScopeNode> RootNodes => rootNodes;
-        public ConcurrentBag<ScopeNode> RootNodes = new ConcurrentBag<ScopeNode>();
+        public IReadOnlyDictionary<string, LabelMetrics> LabelMetrics => labelMetrics;
+        public IReadOnlyDictionary<string, LabelMetrics> CategoryMetrics => categoryMetrics;
+        public IReadOnlyDictionary<int, ThreadMetrics> ThreadMetrics => threadMetrics;
 
-        internal void End()
+        // Return concrete types so callers can use the struct enumerator and avoid boxing.
+        internal Dictionary<string, LabelMetrics> LabelMetricsRaw => labelMetrics;
+        internal Dictionary<int, ThreadMetrics> ThreadMetricsRaw => threadMetrics;
+
+        public long AllocatedBytesStart { get; private set; }
+        public long AllocatedBytesEnd { get; private set; }
+        public long AllocatedBytesDelta { get; private set; }
+        public int Gen0CollectionsStart { get; private set; }
+        public int Gen1CollectionsStart { get; private set; }
+        public int Gen2CollectionsStart { get; private set; }
+        public int Gen0CollectionsEnd { get; private set; }
+        public int Gen1CollectionsEnd { get; private set; }
+        public int Gen2CollectionsEnd { get; private set; }
+        public int Gen0CollectionsDelta { get; private set; }
+        public int Gen1CollectionsDelta { get; private set; }
+        public int Gen2CollectionsDelta { get; private set; }
+        public bool OverBudget { get; private set; }
+        public double BudgetMilliseconds { get; private set; }
+
+        internal void Reset(uint frameCount)
+        {
+            FrameCount = frameCount;
+            StartTime = Stopwatch.GetTimestamp();
+            EndTime = 0;
+            OverBudget = false;
+            BudgetMilliseconds = 0;
+            AllocatedBytesStart = GC.GetTotalAllocatedBytes(false);
+            AllocatedBytesEnd = 0;
+            AllocatedBytesDelta = 0;
+            Gen0CollectionsStart = GC.CollectionCount(0);
+            Gen1CollectionsStart = GC.CollectionCount(1);
+            Gen2CollectionsStart = GC.CollectionCount(2);
+            Gen0CollectionsEnd = 0;
+            Gen1CollectionsEnd = 0;
+            Gen2CollectionsEnd = 0;
+            Gen0CollectionsDelta = 0;
+            Gen1CollectionsDelta = 0;
+            Gen2CollectionsDelta = 0;
+            lock (rootNodesLock)
+            {
+                rootNodes.Clear();
+            }
+        }
+
+        internal void AddRootNode(ScopeNode rootNode)
+        {
+            lock (rootNodesLock)
+            {
+                rootNodes.Add(rootNode);
+            }
+        }
+
+        internal void End(double budgetMilliseconds)
         {
             EndTime = Stopwatch.GetTimestamp();
+            BudgetMilliseconds = budgetMilliseconds;
+            OverBudget = budgetMilliseconds > 0 && ElapsedMilliseconds() > budgetMilliseconds;
+
+            AllocatedBytesEnd = GC.GetTotalAllocatedBytes(false);
+            AllocatedBytesDelta = AllocatedBytesEnd - AllocatedBytesStart;
+
+            Gen0CollectionsEnd = GC.CollectionCount(0);
+            Gen1CollectionsEnd = GC.CollectionCount(1);
+            Gen2CollectionsEnd = GC.CollectionCount(2);
+            Gen0CollectionsDelta = Gen0CollectionsEnd - Gen0CollectionsStart;
+            Gen1CollectionsDelta = Gen1CollectionsEnd - Gen1CollectionsStart;
+            Gen2CollectionsDelta = Gen2CollectionsEnd - Gen2CollectionsStart;
+
+            BuildLabelMetrics();
         }
 
         public double ElapsedMilliseconds()
@@ -37,53 +313,280 @@ public static class Profiler
             long elapsedTicks = EndTime - StartTime;
             return ((double)(elapsedTicks * 1000)) / Stopwatch.Frequency;
         }
+
+        private void BuildLabelMetrics()
+        {
+            labelMetrics.Clear();
+            categoryMetrics.Clear();
+            threadMetrics.Clear();
+            knownThreadIds.Clear();
+            lock (rootNodesLock)
+            {
+                for (int i = 0; i < rootNodes.Count; i++)
+                {
+                    AccumulateLabelMetrics(rootNodes[i]);
+                }
+
+                for (int i = 0; i < rootNodes.Count; i++)
+                {
+                    ScopeNode rootNode = rootNodes[i];
+                    for (int j = 0; j < rootNode.Children.Count; j++)
+                    {
+                        AccumulateThreadMetrics(rootNode.ManagedThreadId, rootNode.Children[j]);
+                    }
+                }
+            }
+
+            ApplyThreadBudgetFlags();
+        }
+
+        private void AccumulateLabelMetrics(ScopeNode node)
+        {
+            double inclusiveMs = node.ElapsedMilliseconds();
+            double selfMs = node.SelfMilliseconds();
+            long allocBytes = node.AllocatedBytes;
+            long selfAllocBytes = node.SelfAllocatedBytes();
+
+            if (labelMetrics.TryGetValue(node.Label, out LabelMetrics current))
+            {
+                labelMetrics[node.Label] = new LabelMetrics(
+                    current.InclusiveMs + inclusiveMs,
+                    current.SelfMs + selfMs,
+                    current.Calls + 1,
+                    Math.Min(current.MinInclusiveMs, inclusiveMs),
+                    Math.Max(current.MaxInclusiveMs, inclusiveMs),
+                    current.AllocatedBytes + allocBytes,
+                    current.SelfAllocatedBytes + selfAllocBytes);
+            }
+            else
+            {
+                labelMetrics[node.Label] = new LabelMetrics(inclusiveMs, selfMs, 1, inclusiveMs, inclusiveMs, allocBytes, selfAllocBytes);
+            }
+
+            if (categoryMetrics.TryGetValue(node.Category, out LabelMetrics categoryCurrent))
+            {
+                categoryMetrics[node.Category] = new LabelMetrics(
+                    categoryCurrent.InclusiveMs + inclusiveMs,
+                    categoryCurrent.SelfMs + selfMs,
+                    categoryCurrent.Calls + 1,
+                    Math.Min(categoryCurrent.MinInclusiveMs, inclusiveMs),
+                    Math.Max(categoryCurrent.MaxInclusiveMs, inclusiveMs),
+                    categoryCurrent.AllocatedBytes + allocBytes,
+                    categoryCurrent.SelfAllocatedBytes + selfAllocBytes);
+            }
+            else
+            {
+                categoryMetrics[node.Category] = new LabelMetrics(inclusiveMs, selfMs, 1, inclusiveMs, inclusiveMs, allocBytes, selfAllocBytes);
+            }
+
+            for (int i = 0; i < node.Children.Count; i++)
+            {
+                AccumulateLabelMetrics(node.Children[i]);
+            }
+        }
+
+        private void AccumulateThreadMetrics(int threadId, ScopeNode node)
+        {
+            double inclusiveMs = node.ElapsedMilliseconds();
+            double selfMs = node.SelfMilliseconds();
+
+            if (threadMetrics.TryGetValue(threadId, out ThreadMetrics current))
+            {
+                threadMetrics[threadId] = new ThreadMetrics(current.InclusiveMs + inclusiveMs, current.SelfMs + selfMs, current.Calls + 1, false);
+            }
+            else
+            {
+                threadMetrics[threadId] = new ThreadMetrics(inclusiveMs, selfMs, 1, false);
+                knownThreadIds.Add(threadId);
+            }
+
+            for (int i = 0; i < node.Children.Count; i++)
+            {
+                AccumulateThreadMetrics(threadId, node.Children[i]);
+            }
+        }
+
+        private void ApplyThreadBudgetFlags()
+        {
+            double perThreadBudget = Math.Max(0d, ThreadBudgetMilliseconds);
+            if (perThreadBudget <= 0d)
+            {
+                return;
+            }
+
+            // knownThreadIds avoids Keys.ToArray() allocation
+            for (int i = 0; i < knownThreadIds.Count; i++)
+            {
+                int key = knownThreadIds[i];
+                ThreadMetrics metric = threadMetrics[key];
+                threadMetrics[key] = new ThreadMetrics(metric.InclusiveMs, metric.SelfMs, metric.Calls, metric.InclusiveMs > perThreadBudget);
+            }
+        }
     }
 
-    public class ScopeNode(string label)
+    public class ScopeNode
     {
-        public string Label { get; } = label;
+        public string Label { get; private set; } = string.Empty;
-        public long StartTime { get; private set; } = Stopwatch.GetTimestamp();  // Start time in ticks
+        public string Category { get; private set; } = DefaultCategory;
+        public ulong TagMask { get; private set; }
+        public long StartTime { get; private set; }
         public long EndTime { get; private set; }
-        public int ManagedThreadId { get; } = Environment.CurrentManagedThreadId;
+        public int ManagedThreadId { get; private set; }
         public List<ScopeNode> Children { get; } = new List<ScopeNode>();
+        public long AllocatedBytes { get; private set; }
+        public long ProfilerSetupBytes { get; private set; }
+        internal ScopeNode Parent { get; private set; }
+        internal long ChildrenDurationTicks { get; private set; }
+        internal long ChildrenAllocatedBytes { get; private set; }
+        private long allocatedBytesAtStart;
+
+        internal void Reset(string label, string category, ulong tagMask, int managedThreadId, ScopeNode parent)
+        {
+            Label = label;
+            Category = string.IsNullOrEmpty(category) ? DefaultCategory : category;
+            TagMask = tagMask;
+            ManagedThreadId = managedThreadId;
+            Parent = parent;
+            StartTime = Stopwatch.GetTimestamp();
+            EndTime = 0;
+            ChildrenDurationTicks = 0;
+            ChildrenAllocatedBytes = 0;
+            AllocatedBytes = 0;
+            ProfilerSetupBytes = 0;
+            Children.Clear();
+            allocatedBytesAtStart = GC.GetAllocatedBytesForCurrentThread();
+        }
 
         internal void End()
         {
-            EndTime = Stopwatch.GetTimestamp();  // End time in ticks
+            if (EndTime != 0)
+            {
+                return;
+            }
+
+            EndTime = Stopwatch.GetTimestamp();
+            if (Parent != null)
+            {
+                // Root nodes are ended from FinalizeCurrentFrame on the main thread, so their
+                // GC counter would be from the wrong thread. Only track alloc on non-root nodes.
+                AllocatedBytes = Math.Max(0, GC.GetAllocatedBytesForCurrentThread() - allocatedBytesAtStart);
+                Parent.ChildrenDurationTicks += Math.Max(0, EndTime - StartTime);
+                Parent.ChildrenAllocatedBytes += AllocatedBytes;
+            }
         }
 
         public double ElapsedMilliseconds()
         {
-            return ((double)(EndTime - StartTime)) * 1000 / Stopwatch.Frequency;  // Convert ticks to ms
+            return ((double)(Math.Max(0, EndTime - StartTime))) * 1000 / Stopwatch.Frequency;
         }
 
-        // Add a child node (used for nested scopes)
+        public double SelfMilliseconds()
-        internal ScopeNode AddChild(string label)
         {
-            ScopeNode child = new ScopeNode(label);
+            long elapsedTicks = Math.Max(0, EndTime - StartTime);
+            long selfTicks = Math.Max(0, elapsedTicks - ChildrenDurationTicks);
+            return ((double)selfTicks) * 1000 / Stopwatch.Frequency;
+        }
+
+        public long SelfAllocatedBytes()
+        {
+            return Math.Max(0, AllocatedBytes - ChildrenAllocatedBytes);
+        }
+
+        // Called after all profiler setup (Children.Add + scopes.Push) to measure overhead within this scope's window.
+        internal void NoteSetupOverhead()
+        {
+            ProfilerSetupBytes = Math.Max(0, GC.GetAllocatedBytesForCurrentThread() - allocatedBytesAtStart);
+        }
+
+        internal ScopeNode AddChild(string label, string category, ulong tagMask)
+        {
+            ScopeNode child = RentNode(label, category, tagMask, ManagedThreadId, this);
             Children.Add(child);
             return child;
         }
     }
 
     private const int maxFrames = 128;
+    private const int rollingWindowSize = 240;
+    private const string DefaultCategory = "General";
 
     public static bool IsRecording { get; private set; } = true;
+    public static double FrameBudgetMilliseconds { get; set; } = 16.667;
+    public static double ThreadBudgetMilliseconds { get; set; } = 8.333;
+    public static CaptureMode Mode { get; set; } = CaptureMode.Instrumented;
+    public static int SamplingStride { get; set; } = 8;
 
-    // Store only the last x amount of frames in memory.
     public static readonly BoundedQueue<Frame> Frames = new(maxFrames);
 
-    // Use ThreadLocal to store a stack of ScopeNodes per thread and enable tracking of thread-local values.
+    // trackAllValues=false; registeredThreadStates avoids ThreadLocal.Values allocating a ReadOnlyCollection each call
-    private static readonly ThreadLocal<Stack<ScopeNode>> threadLocalScopes = new ThreadLocal<Stack<ScopeNode>>(() => new Stack<ScopeNode>(), true);
+    private static readonly ThreadLocal<ThreadProfilerState> threadStates =
+        new ThreadLocal<ThreadProfilerState>(() =>
+        {
+            ThreadProfilerState state = new ThreadProfilerState();
+            lock (registeredThreadStatesLock)
+            {
+                registeredThreadStates.Add(state);
+            }
+
+            return state;
+        });
+
+    private static readonly List<ThreadProfilerState> registeredThreadStates = new List<ThreadProfilerState>(8);
+    private static readonly object registeredThreadStatesLock = new object();
+
+    private static readonly ConcurrentDictionary<int, string> threadRootLabelCache = new ConcurrentDictionary<int, string>();
+    private static readonly ConcurrentBag<ScopeNode> nodePool = new ConcurrentBag<ScopeNode>();
+    private static readonly ConcurrentBag<Frame> framePool = new ConcurrentBag<Frame>(); // pooled; avoids per-frame Frame allocation
+    private static readonly Dictionary<string, RollingWindow> rollingWindows = new Dictionary<string, RollingWindow>(256, StringComparer.Ordinal);
+    private static readonly Dictionary<int, RollingThreadWindow> rollingThreadWindows = new Dictionary<int, RollingThreadWindow>(16);
+    private static readonly object rollingWindowsLock = new object();
 
     private static Frame currentFrame = null;
     private static uint frameCount = 0;
 
     public static void SetActive(bool isRecording)
     {
+        if (IsRecording && !isRecording)
+        {
+            FinalizeCurrentFrame();
+        }
+
         IsRecording = isRecording;
     }
 
+    public static int CopyFramesTo(List<Frame> destination)
+    {
+        return Frames.CopyTo(destination);
+    }
+
+    public static IReadOnlyDictionary<string, RollingLabelMetrics> GetRollingLabelMetricsSnapshot()
+    {
+        lock (rollingWindowsLock)
+        {
+            Dictionary<string, RollingLabelMetrics> snapshot = new Dictionary<string, RollingLabelMetrics>(rollingWindows.Count, StringComparer.Ordinal);
+            foreach (KeyValuePair<string, RollingWindow> pair in rollingWindows)
+            {
+                snapshot[pair.Key] = pair.Value.Snapshot();
+            }
+
+            return snapshot;
+        }
+    }
+
+    public static IReadOnlyDictionary<int, RollingThreadMetrics> GetRollingThreadMetricsSnapshot()
+    {
+        lock (rollingWindowsLock)
+        {
+            Dictionary<int, RollingThreadMetrics> snapshot = new Dictionary<int, RollingThreadMetrics>(rollingThreadWindows.Count);
+            foreach (KeyValuePair<int, RollingThreadWindow> pair in rollingThreadWindows)
+            {
+                snapshot[pair.Key] = pair.Value.Snapshot();
+            }
+
+            return snapshot;
+        }
+    }
+
     [Conditional("PROFILING")]
     public static void BeginFrame()
     {
@@ -92,7 +595,12 @@ public static class Profiler
             return;
         }
 
-        currentFrame = new Frame(frameCount);
+        if (currentFrame != null)
+        {
+            FinalizeCurrentFrame();
+        }
+
+        currentFrame = RentFrame(frameCount);
     }
 
     [Conditional("PROFILING")]
@@ -103,64 +611,202 @@ public static class Profiler
             return;
         }
 
-        foreach (Stack<ScopeNode> scopes in threadLocalScopes.Values)
+        FinalizeCurrentFrame();
-        {
-            if (scopes.Count > 0)
-            {
-                // Pop the left over root nodes.
-                ScopeNode currentScope = scopes.Pop();
-                currentScope.End();
-            }
-
-            // Clean up the thread-local stack to ensure it's empty for the next frame.
-            scopes.Clear();
-        }
-
-        currentFrame.End();
-        Frames.Enqueue(currentFrame);
-        frameCount++;
     }
 
     [Conditional("PROFILING")]
     public static void BeginSample(string label)
     {
-        if (!IsRecording)
+        BeginSample(label, DefaultCategory, 0);
+    }
+
+    [Conditional("PROFILING")]
+    public static void BeginSample(string label, string category, ulong tagMask = 0)
+    {
+        if (!IsRecording || currentFrame == null)
         {
             return;
         }
 
-        Stack<ScopeNode> scopes = threadLocalScopes.Value;  // Get the stack for the current thread
+        ThreadProfilerState state = threadStates.Value;
+        state.ThreadId = Environment.CurrentManagedThreadId;
+
+        bool parentCaptured = state.CaptureDecisions.Count > 0 && state.CaptureDecisions.Peek();
+        bool capture = parentCaptured || Mode == CaptureMode.Instrumented || ShouldSample(state);
+        state.CaptureDecisions.Push(capture);
+
+        if (!capture)
+        {
+            return;
+        }
+
+        Stack<ScopeNode> scopes = state.Scopes;
+        Frame frame = currentFrame;
+        if (frame == null)
+        {
+            state.CaptureDecisions.Pop();
+            return;
+        }
 
         if (scopes.Count == 0)
         {
-            // First scope for this thread (new root for this thread)
+            int threadId = state.ThreadId;
-            ScopeNode rootScopeNode = new ScopeNode($"Thread-{Environment.CurrentManagedThreadId}");
+            ScopeNode rootScopeNode = RentNode(GetThreadRootLabel(threadId), DefaultCategory, 0, threadId, null);
             scopes.Push(rootScopeNode);
-            currentFrame.RootNodes.Add(rootScopeNode);  // Add root node to the frame list
+            frame.AddRootNode(rootScopeNode);
         }
 
-        // Create a new child under the current top of the stack
+        ScopeNode newScope = scopes.Peek().AddChild(label, category, tagMask);
-        ScopeNode newScope = scopes.Peek().AddChild(label);
+        scopes.Push(newScope);
-
+        newScope.NoteSetupOverhead();
-        scopes.Push(newScope);  // Push new scope to the thread's stack
     }
 
     [Conditional("PROFILING")]
     public static void EndSample()
     {
-        if (!IsRecording)
+        if (!IsRecording || currentFrame == null)
         {
             return;
         }
 
-        Stack<ScopeNode> scopes = threadLocalScopes.Value;
+        ThreadProfilerState state = threadStates.Value;
+        if (state.CaptureDecisions.Count == 0)
+        {
+            return;
+        }
 
-        if (scopes.Count > 0)
+        bool captured = state.CaptureDecisions.Pop();
+        if (!captured)
+        {
+            return;
+        }
+
+        Stack<ScopeNode> scopes = state.Scopes;
+
+        if (scopes.Count > 1)
         {
-            // Only pop if this is not the root node.
-            //ScopeNode currentScope = scopes.Count > 1 ? scopes.Pop() : scopes.Peek();
             ScopeNode currentScope = scopes.Pop();
             currentScope.End();
         }
     }
-}
+
+    private static bool ShouldSample(ThreadProfilerState state)
+    {
+        int stride = Math.Max(1, SamplingStride);
+        state.SampleCursor++;
+        return state.SampleCursor % stride == 0;
+    }
+
+    private static string GetThreadRootLabel(int threadId)
+    {
+        return threadRootLabelCache.GetOrAdd(threadId, static id => $"Thread-{id}");
+    }
+
+    private static Frame RentFrame(uint count)
+    {
+        if (!framePool.TryTake(out Frame frame))
+        {
+            frame = new Frame();
+        }
+
+        frame.Reset(count);
+        return frame;
+    }
+
+    private static ScopeNode RentNode(string label, string category, ulong tagMask, int managedThreadId, ScopeNode parent)
+    {
+        if (!nodePool.TryTake(out ScopeNode node))
+        {
+            node = new ScopeNode();
+        }
+
+        node.Reset(label, category, tagMask, managedThreadId, parent);
+        return node;
+    }
+
+    private static void ReturnNodeTree(ScopeNode node)
+    {
+        for (int i = 0; i < node.Children.Count; i++)
+        {
+            ReturnNodeTree(node.Children[i]);
+        }
+
+        node.Reset(string.Empty, DefaultCategory, 0, 0, null);
+        nodePool.Add(node);
+    }
+
+    private static void FinalizeCurrentFrame()
+    {
+        Frame frame = currentFrame;
+        if (frame == null)
+        {
+            return;
+        }
+
+        lock (registeredThreadStatesLock)
+        {
+            for (int i = 0; i < registeredThreadStates.Count; i++)
+            {
+                ThreadProfilerState state = registeredThreadStates[i];
+                Stack<ScopeNode> scopes = state.Scopes;
+                while (scopes.Count > 0)
+                {
+                    scopes.Pop().End();
+                }
+
+                state.CaptureDecisions.Clear();
+            }
+        }
+
+        frame.End(FrameBudgetMilliseconds);
+
+        if (Frames.Enqueue(frame, out Frame evictedFrame))
+        {
+            for (int i = 0; i < evictedFrame.RootNodes.Count; i++)
+            {
+                ReturnNodeTree(evictedFrame.RootNodes[i]);
+            }
+
+            framePool.Add(evictedFrame);
+        }
+
+        UpdateRollingWindows(frame);
+        UpdateRollingThreadWindows(frame);
+        frameCount++;
+        currentFrame = null;
+    }
+
+    private static void UpdateRollingWindows(Frame frame)
+    {
+        lock (rollingWindowsLock)
+        {
+            foreach (KeyValuePair<string, LabelMetrics> pair in frame.LabelMetricsRaw)
+            {
+                if (!rollingWindows.TryGetValue(pair.Key, out RollingWindow window))
+                {
+                    window = new RollingWindow(rollingWindowSize);
+                    rollingWindows.Add(pair.Key, window);
+                }
+
+                window.Add(pair.Value.InclusiveMs);
+            }
+        }
+    }
+
+    private static void UpdateRollingThreadWindows(Frame frame)
+    {
+        lock (rollingWindowsLock)
+        {
+            foreach (KeyValuePair<int, ThreadMetrics> pair in frame.ThreadMetricsRaw)
+            {
+                if (!rollingThreadWindows.TryGetValue(pair.Key, out RollingThreadWindow window))
+                {
+                    window = new RollingThreadWindow(rollingWindowSize);
+                    rollingThreadWindows.Add(pair.Key, window);
+                }
+
+                window.Add(pair.Value.InclusiveMs, pair.Value.OverBudget);
+            }
+        }
+    }
+}

Nerfed.Runtime/ProfilerVisualizer.cs

@@ -5,152 +5,552 @@ namespace Nerfed.Runtime;
 
 public static class ProfilerVisualizer
 {
-    private const float barHeight = 20f;
+    public sealed class TimelineState
-    private const float barPadding = 2f;
-
-    // Render the flame graph across multiple threads
-    public static void RenderFlameGraph(Profiler.Frame frame)
     {
-        if (frame == null) return;
+        public int SelectedFrameIndex = -1;
-        if (frame.RootNodes == null) return;
+        public int WindowStartIndex = 0;
-
+        public int VisibleFrameCount = 64;
-        // Calculate the total timeline duration (max end time across all nodes)
+        public bool FollowLatest = true;
-        double totalDuration = frame.EndTime - frame.StartTime;
+        public float Zoom = 1f;
-        double startTime = frame.StartTime;
+        public double PanTicks = 0;
-
-        // Precompute the maximum depth for each thread's call stack
-        Dictionary<int, int> threadMaxDepths = new Dictionary<int, int>();
-        foreach (IGrouping<int, Profiler.ScopeNode> threadGroup in frame.RootNodes.GroupBy(node => node.ManagedThreadId))
-        {
-            int maxDepth = 0;
-            foreach (Profiler.ScopeNode rootNode in threadGroup)
-            {
-                maxDepth = Math.Max(maxDepth, GetMaxDepth(rootNode, 0));
-            }
-            threadMaxDepths[threadGroup.Key] = maxDepth;
-        }
-
-        // Start a child window to support scrolling
-        ImGui.BeginChild("FlameGraph", new Vector2(0, 64), ImGuiChildFlags.Border | ImGuiChildFlags.ResizeY, ImGuiWindowFlags.HorizontalScrollbar | ImGuiWindowFlags.AlwaysVerticalScrollbar);
-
-        ImDrawListPtr drawList = ImGui.GetWindowDrawList();
-        Vector2 windowPos = ImGui.GetCursorScreenPos();
-
-        // Sort nodes by ThreadID, ensuring main thread (Thread ID 1) is on top
-        IOrderedEnumerable<IGrouping<int, Profiler.ScopeNode>> threadGroups = frame.RootNodes.GroupBy(node => node.ManagedThreadId).OrderBy(g => g.Key);
-
-        // Initial Y position for drawing
-        float baseY = windowPos.Y;
-        bool alternate = false;
-        float contentWidth = ImGui.GetContentRegionAvail().X;
-
-        // Draw each thread's flame graph row by row
-        foreach (IGrouping<int, Profiler.ScopeNode> threadGroup in threadGroups)
-        {
-            int threadId = threadGroup.Key;
-
-            // Compute the base Y position for this thread
-            float threadBaseY = baseY;
-
-            // Calculate the maximum height for this thread's flame graph
-            float threadHeight = (threadMaxDepths[threadId] + 1) * (barHeight + barPadding);
-
-            // Draw the alternating background for each thread row
-            uint backgroundColor = ImGui.ColorConvertFloat4ToU32(alternate ? new Vector4(0.2f, 0.2f, 0.2f, 1f) : new Vector4(0.1f, 0.1f, 0.1f, 1f));
-            drawList.AddRectFilled(new Vector2(windowPos.X, threadBaseY), new Vector2(windowPos.X + contentWidth, threadBaseY + threadHeight), backgroundColor);
-
-            alternate = !alternate;
-
-            // Draw each root node in the group (one per thread)
-            foreach (Profiler.ScopeNode rootNode in threadGroup)
-            {
-                RenderNode(drawList, rootNode, startTime, totalDuration, windowPos.X, threadBaseY, 0, contentWidth, false);
-            }
-
-            // Move to the next thread's row (max depth * height per level)
-            baseY += (threadMaxDepths[threadId] + 1) * (barHeight + barPadding);
-        }
-
-        // Ensure that ImGui knows the size of the content.
-        ImGui.Dummy(new Vector2(contentWidth, baseY));
-
-        ImGui.EndChild();
     }
 
-    private static void RenderNode(ImDrawListPtr drawList, Profiler.ScopeNode node, double startTime, double totalDuration, float startX, float baseY, int depth, float contentWidth, bool alternate)
+    public readonly struct TimelineRenderResult
     {
-        if (node == null) return;
+        public TimelineRenderResult(int selectedFrameIndex, bool selectionChanged, bool userNavigated)
-
-        double nodeStartTime = node.StartTime - startTime;
-        double nodeEndTime = node.EndTime - startTime;
-        double nodeDuration = nodeEndTime - nodeStartTime;
-
-        // Calculate the position and width of the bar based on time
-        float xPos = (float)(startX + (nodeStartTime / totalDuration) * contentWidth);
-        float width = (float)((nodeDuration / totalDuration) * contentWidth);
-
-        // Calculate the Y position based on depth
-        float yPos = baseY + (depth * (barHeight + barPadding)) + (barPadding * 0.5f);
-
-        // Define the rectangle bounds for the node
-        Vector2 min = new Vector2(xPos, yPos);
-        Vector2 max = new Vector2(xPos + width, yPos + barHeight);
-
-        // Define color.
-        Vector4 barColor = alternate ? new Vector4(0.4f, 0.6f, 0.9f, 1f) : new Vector4(0.4f, 0.5f, 0.8f, 1f);
-        Vector4 textColor = new Vector4(1f, 1f, 1f, 1f);
-
-        if (depth != 0)
         {
-            // Draw the bar for the node (colored based on thread depth)
+            SelectedFrameIndex = selectedFrameIndex;
-            drawList.AddRectFilled(min, max, ImGui.ColorConvertFloat4ToU32(barColor));
+            SelectionChanged = selectionChanged;
+            UserNavigated = userNavigated;
+        }
 
-            // Draw the label if it fits inside the bar
+        public int SelectedFrameIndex { get; }
-            string label = $"{node.Label} ({node.ElapsedMilliseconds():0.000} ms)";
+        public bool SelectionChanged { get; }
-            if (width > ImGui.CalcTextSize(label).X)
+        public bool UserNavigated { get; }
+    }
+
+    private readonly struct HoverEntry
+    {
+        public HoverEntry(Profiler.ScopeNode node, Profiler.Frame frame, int frameIndex, int depth, double timelineStartTicks)
+        {
+            Node = node;
+            Frame = frame;
+            FrameIndex = frameIndex;
+            Depth = depth;
+            DurationMs = TicksToMilliseconds(node.EndTime - node.StartTime);
+            SelfMs = node.SelfMilliseconds();
+            StartInFrameMs = TicksToMilliseconds(node.StartTime - frame.StartTime);
+            EndInFrameMs = TicksToMilliseconds(node.EndTime - frame.StartTime);
+            StartInTimelineMs = TicksToMilliseconds(node.StartTime - timelineStartTicks);
+            EndInTimelineMs = TicksToMilliseconds(node.EndTime - timelineStartTicks);
+        }
+
+        public Profiler.ScopeNode Node { get; }
+        public Profiler.Frame Frame { get; }
+        public int FrameIndex { get; }
+        public int Depth { get; }
+        public double DurationMs { get; }
+        public double SelfMs { get; }
+        public double StartInFrameMs { get; }
+        public double EndInFrameMs { get; }
+        public double StartInTimelineMs { get; }
+        public double EndInTimelineMs { get; }
+    }
+
+    private const float BarHeight = 18f;
+    private const float BarPadding = 2f;
+    private const float ThreadGap = 8f;
+    private const float HeaderHeight = 28f;
+    private const float TimelineHeight = 220f;
+    private const float MinTextWidth = 36f;
+    private static readonly double TickToMs = 1000d / System.Diagnostics.Stopwatch.Frequency;
+
+    // Backwards-compatible entry point used by existing call sites.
+    public static void RenderFlameGraph(Profiler.Frame frame)
+    {
+        if (frame == null)
+        {
+            return;
+        }
+
+        List<Profiler.Frame> frames = new List<Profiler.Frame>(1)
+        {
+            frame
+        };
+
+        TimelineState state = new TimelineState
+        {
+            VisibleFrameCount = 1,
+            SelectedFrameIndex = 0,
+            FollowLatest = true
+        };
+
+        RenderTimeline(frames, state);
+    }
+
+    public static TimelineRenderResult RenderTimeline(IReadOnlyList<Profiler.Frame> frames, TimelineState state)
+    {
+        if (frames == null || frames.Count == 0 || state == null)
+        {
+            return default;
+        }
+
+        bool selectionChanged = false;
+        bool userNavigated = false;
+
+        int frameCount = frames.Count;
+        state.VisibleFrameCount = Math.Clamp(state.VisibleFrameCount, 1, frameCount);
+        state.Zoom = Math.Clamp(state.Zoom, 1f, 128f);
+
+        int maxStartIndex = Math.Max(0, frameCount - state.VisibleFrameCount);
+        if (state.FollowLatest)
+        {
+            state.WindowStartIndex = maxStartIndex;
+        }
+        else
+        {
+            state.WindowStartIndex = Math.Clamp(state.WindowStartIndex, 0, maxStartIndex);
+        }
+
+        int visibleStartIndex = state.WindowStartIndex;
+        int visibleEndIndex = visibleStartIndex + state.VisibleFrameCount - 1;
+
+        if (state.SelectedFrameIndex < 0)
+        {
+            state.SelectedFrameIndex = visibleEndIndex;
+            selectionChanged = true;
+        }
+        state.SelectedFrameIndex = Math.Clamp(state.SelectedFrameIndex, visibleStartIndex, visibleEndIndex);
+
+        Profiler.Frame firstFrame = frames[visibleStartIndex];
+        Profiler.Frame lastFrame = frames[visibleEndIndex];
+
+        double timelineStartTicks = firstFrame.StartTime;
+        double timelineEndTicks = Math.Max(lastFrame.EndTime, firstFrame.StartTime + 1);
+        double timelineDurationTicks = Math.Max(1d, timelineEndTicks - timelineStartTicks);
+
+        double visibleDurationTicks = Math.Max(1d, timelineDurationTicks / state.Zoom);
+        double maxPanTicks = Math.Max(0d, timelineDurationTicks - visibleDurationTicks);
+        if (state.FollowLatest)
+        {
+            state.PanTicks = maxPanTicks;
+        }
+        else
+        {
+            state.PanTicks = Math.Clamp(state.PanTicks, 0d, maxPanTicks);
+        }
+
+        double visibleStartTicks = timelineStartTicks + state.PanTicks;
+        double visibleEndTicks = visibleStartTicks + visibleDurationTicks;
+
+        Dictionary<int, int> threadDepths = BuildThreadDepths(frames, visibleStartIndex, visibleEndIndex);
+        List<int> threadOrder = threadDepths.Keys.OrderBy(x => x).ToList();
+
+        Dictionary<int, float> threadBaseY = new Dictionary<int, float>(threadOrder.Count);
+        float yCursor = HeaderHeight;
+        for (int i = 0; i < threadOrder.Count; i++)
+        {
+            int threadId = threadOrder[i];
+            threadBaseY[threadId] = yCursor;
+            yCursor += ((threadDepths[threadId] + 1) * (BarHeight + BarPadding)) + ThreadGap;
+        }
+
+        float contentHeight = Math.Max(TimelineHeight, yCursor + 6f);
+
+        ImGui.BeginChild("ProfilerTimeline", new Vector2(0, TimelineHeight), ImGuiChildFlags.Border | ImGuiChildFlags.ResizeY, ImGuiWindowFlags.AlwaysVerticalScrollbar);
+
+        ImDrawListPtr drawList = ImGui.GetWindowDrawList();
+        Vector2 origin = ImGui.GetCursorScreenPos();
+        Vector2 viewSize = ImGui.GetContentRegionAvail();
+        float canvasWidth = Math.Max(1f, viewSize.X);
+
+        uint frameBgColor = ImGui.GetColorU32(ImGuiCol.FrameBg);
+        uint frameBgHoveredColor = ImGui.GetColorU32(ImGuiCol.FrameBgHovered);
+        uint headerColor = ImGui.GetColorU32(ImGuiCol.Header);
+        uint headerHoveredColor = ImGui.GetColorU32(ImGuiCol.HeaderHovered);
+        uint textColor = ImGui.GetColorU32(ImGuiCol.Text);
+        uint borderColor = ImGui.GetColorU32(ImGuiCol.Border);
+
+        float clipMinX = origin.X;
+        float clipMaxX = origin.X + canvasWidth;
+        float clipMinY = origin.Y;
+        float clipMaxY = origin.Y + Math.Max(1f, ImGui.GetWindowHeight());
+
+        DrawTimelineHeader(drawList, origin, canvasWidth, timelineStartTicks, visibleStartTicks, visibleDurationTicks, textColor, borderColor);
+
+        HoverEntry? hovered = null;
+
+        for (int frameIndex = visibleStartIndex; frameIndex <= visibleEndIndex; frameIndex++)
+        {
+            Profiler.Frame frame = frames[frameIndex];
+            float frameStartX = ToScreenX(frame.StartTime, visibleStartTicks, visibleDurationTicks, origin.X, canvasWidth);
+            float frameEndX = ToScreenX(frame.EndTime, visibleStartTicks, visibleDurationTicks, origin.X, canvasWidth);
+
+            if (frameEndX < clipMinX || frameStartX > clipMaxX)
             {
-                drawList.AddText(new Vector2(xPos + barPadding, yPos + barPadding), ImGui.ColorConvertFloat4ToU32(textColor), label);
+                continue;
             }
 
-            // Add tooltip on hover
+            bool isSelectedFrame = frameIndex == state.SelectedFrameIndex;
-            if (ImGui.IsMouseHoveringRect(min, max))
+            uint frameShadeColor = (frameIndex & 1) == 0 ? frameBgColor : frameBgHoveredColor;
+            if (isSelectedFrame)
             {
-                // Show tooltip when hovering over the node
+                frameShadeColor = LerpColor(headerColor, headerHoveredColor, 0.45f);
-                ImGui.BeginTooltip();
+            }
-                ImGui.Text($"{node.Label}");
+
-                ImGui.Text($"{node.ElapsedMilliseconds():0.000} ms");
+            drawList.AddRectFilled(new Vector2(frameStartX, origin.Y + HeaderHeight), new Vector2(frameEndX, origin.Y + contentHeight), SetAlpha(frameShadeColor, isSelectedFrame ? 0.20f : 0.08f));
-                ImGui.Text($"{node.ManagedThreadId}");
+            drawList.AddLine(new Vector2(frameStartX, origin.Y + HeaderHeight), new Vector2(frameStartX, origin.Y + contentHeight), SetAlpha(borderColor, 0.55f), 1f);
-                ImGui.EndTooltip();
+
+            foreach (Profiler.ScopeNode root in frame.RootNodes)
+            {
+                if (!threadBaseY.TryGetValue(root.ManagedThreadId, out float baseY))
+                {
+                    continue;
+                }
+
+                DrawThreadLabel(drawList, origin.X, origin.Y + baseY, root.ManagedThreadId, textColor);
+                for (int i = 0; i < root.Children.Count; i++)
+                {
+                    RenderNode(
+                        drawList,
+                        root.Children[i],
+                        frame,
+                        frameIndex,
+                        baseY,
+                        0,
+                        visibleStartTicks,
+                        visibleDurationTicks,
+                        origin.X,
+                        origin.Y,
+                        canvasWidth,
+                        clipMinX,
+                        clipMaxX,
+                        clipMaxY,
+                        ref hovered,
+                        textColor,
+                        headerColor,
+                        headerHoveredColor,
+                        frameIndex == state.SelectedFrameIndex);
+                }
+            }
+
+            drawList.AddLine(new Vector2(frameEndX, origin.Y + HeaderHeight), new Vector2(frameEndX, origin.Y + contentHeight), SetAlpha(borderColor, 0.30f), 1f);
+        }
+
+        ImGui.Dummy(new Vector2(canvasWidth, contentHeight));
+
+        bool windowHovered = ImGui.IsWindowHovered(ImGuiHoveredFlags.ChildWindows);
+        if (windowHovered)
+        {
+            HandleZoomAndPan(state, timelineStartTicks, timelineDurationTicks, visibleDurationTicks, origin.X, canvasWidth, ref userNavigated);
+        }
+
+        if (windowHovered && ImGui.IsMouseClicked(ImGuiMouseButton.Left))
+        {
+            int clickedFrame = FindFrameIndexByMouseX(frames, visibleStartIndex, visibleEndIndex, visibleStartTicks, visibleDurationTicks, origin.X, canvasWidth, ImGui.GetMousePos().X);
+            if (clickedFrame >= visibleStartIndex && clickedFrame <= visibleEndIndex && clickedFrame != state.SelectedFrameIndex)
+            {
+                state.SelectedFrameIndex = clickedFrame;
+                selectionChanged = true;
+                state.FollowLatest = false;
+                userNavigated = true;
+            }
+        }
+
+        if (hovered.HasValue)
+        {
+            if (ImGui.IsMouseClicked(ImGuiMouseButton.Left) && hovered.Value.FrameIndex != state.SelectedFrameIndex)
+            {
+                state.SelectedFrameIndex = hovered.Value.FrameIndex;
+                selectionChanged = true;
+                state.FollowLatest = false;
+                userNavigated = true;
+            }
+
+            DrawTooltip(hovered.Value);
+        }
+
+        ImGui.EndChild();
+
+        return new TimelineRenderResult(state.SelectedFrameIndex, selectionChanged, userNavigated);
+    }
+
+    private static void HandleZoomAndPan(TimelineState state, double timelineStartTicks, double timelineDurationTicks, double visibleDurationTicks, float originX, float canvasWidth, ref bool userNavigated)
+    {
+        ImGuiIOPtr io = ImGui.GetIO();
+        if (Math.Abs(io.MouseWheel) < float.Epsilon)
+        {
+            return;
+        }
+
+        if (!io.KeyCtrl && !io.KeyShift)
+        {
+            return; // plain scroll goes to ImGui vertical scrolling
+        }
+
+        float wheel = io.MouseWheel;
+        io.MouseWheel = 0; // consume so the child window doesn't also scroll vertically
+
+        if (io.KeyCtrl)
+        {
+            float previousZoom = state.Zoom;
+            double visibleStartTicksBefore = timelineStartTicks + state.PanTicks;
+            double mouseT = Math.Clamp((ImGui.GetMousePos().X - originX) / Math.Max(1f, canvasWidth), 0f, 1f);
+            double pivotTick = visibleStartTicksBefore + (visibleDurationTicks * mouseT);
+
+            state.Zoom = Math.Clamp(state.Zoom * MathF.Pow(1.12f, wheel), 1f, 128f);
+            if (Math.Abs(previousZoom - state.Zoom) > float.Epsilon)
+            {
+                double newVisibleDurationTicks = Math.Max(1d, timelineDurationTicks / state.Zoom);
+                double newVisibleStartTicks = pivotTick - (newVisibleDurationTicks * mouseT);
+                state.PanTicks = Math.Clamp(newVisibleStartTicks - timelineStartTicks, 0d, Math.Max(0d, timelineDurationTicks - newVisibleDurationTicks));
+                state.FollowLatest = false;
+                userNavigated = true;
             }
         }
         else
         {
-            // Aka root node.
+            // Shift + scroll: horizontal pan
-            string label = $"{node.Label}";
+            state.PanTicks -= wheel * (visibleDurationTicks * 0.10d);
-            drawList.AddText(new Vector2(startX + barPadding, yPos + barPadding), ImGui.ColorConvertFloat4ToU32(textColor), label);
+            state.FollowLatest = false;
-        }
+            userNavigated = true;
-
-        // Draw each child node under this node
-        foreach (Profiler.ScopeNode child in node.Children)
-        {
-            alternate = !alternate;
-            RenderNode(drawList, child, startTime, totalDuration, startX, baseY, depth + 1, contentWidth, alternate);
         }
     }
 
-    // Recursive function to calculate the maximum depth of the node tree
+    private static int FindFrameIndexByMouseX(IReadOnlyList<Profiler.Frame> frames, int visibleStartIndex, int visibleEndIndex, double visibleStartTicks, double visibleDurationTicks, float originX, float width, float mouseX)
-    private static int GetMaxDepth(Profiler.ScopeNode node, int currentDepth)
     {
-        if (node.Children == null || node.Children.Count == 0)
+        double t = Math.Clamp((mouseX - originX) / Math.Max(1f, width), 0f, 1f);
+        double timelineTicks = visibleStartTicks + (visibleDurationTicks * t);
+
+        for (int i = visibleStartIndex; i <= visibleEndIndex; i++)
         {
-            return currentDepth;
+            Profiler.Frame frame = frames[i];
+            if (timelineTicks >= frame.StartTime && timelineTicks <= frame.EndTime)
+            {
+                return i;
+            }
         }
 
-        int maxDepth = currentDepth;
+        return -1;
-        foreach (Profiler.ScopeNode child in node.Children)
+    }
+
+    private static void DrawTimelineHeader(ImDrawListPtr drawList, Vector2 origin, float width, double timelineStartTicks, double visibleStartTicks, double visibleDurationTicks, uint textColor, uint borderColor)
+    {
+        drawList.AddLine(new Vector2(origin.X, origin.Y + HeaderHeight), new Vector2(origin.X + width, origin.Y + HeaderHeight), SetAlpha(borderColor, 0.65f), 1f);
+
+        int tickCount = Math.Clamp((int)(width / 130f), 4, 20);
+        for (int i = 0; i <= tickCount; i++)
         {
-            maxDepth = Math.Max(maxDepth, GetMaxDepth(child, currentDepth + 1));
+            float t = i / (float)tickCount;
+            float x = origin.X + (t * width);
+            drawList.AddLine(new Vector2(x, origin.Y + HeaderHeight - 8f), new Vector2(x, origin.Y + HeaderHeight), SetAlpha(borderColor, 0.8f), 1f);
+
+            double ms = TicksToMilliseconds((visibleStartTicks - timelineStartTicks) + (visibleDurationTicks * t));
+            drawList.AddText(new Vector2(x + 2f, origin.Y + 4f), textColor, $"+{ms:0.0} ms");
+        }
+    }
+
+    private static void DrawThreadLabel(ImDrawListPtr drawList, float x, float y, int threadId, uint textColor)
+    {
+        drawList.AddText(new Vector2(x + 4f, y + 2f), SetAlpha(textColor, 0.85f), $"T{threadId}");
+    }
+
+    private static void DrawTooltip(HoverEntry hover)
+    {
+        ImGui.BeginTooltip();
+        ImGui.Text($"{hover.Node.Label}");
+        ImGui.Separator();
+        ImGui.Text($"Category: {hover.Node.Category}");
+        ImGui.Text($"Tags: 0x{hover.Node.TagMask:X}");
+        ImGui.Text($"Frame: {hover.Frame.FrameCount} (idx {hover.FrameIndex})");
+        ImGui.Text($"Thread: {hover.Node.ManagedThreadId}");
+        ImGui.Text($"Depth: {hover.Depth}");
+        ImGui.Text($"Duration: {hover.DurationMs:0.000} ms");
+        ImGui.Text($"Self: {hover.SelfMs:0.000} ms");
+        ImGui.Text($"Frame Start: {hover.StartInFrameMs:0.000} ms");
+        ImGui.Text($"Frame End: {hover.EndInFrameMs:0.000} ms");
+        ImGui.Text($"Timeline Start: {hover.StartInTimelineMs:0.000} ms");
+        ImGui.Text($"Timeline End: {hover.EndInTimelineMs:0.000} ms");
+        ImGui.Text($"Children: {hover.Node.Children.Count}");
+        if (hover.Node.ProfilerSetupBytes > 0)
+        {
+            ImGui.Separator();
+            ImGui.TextColored(new Vector4(1f, 0.75f, 0f, 1f), $"\u26a0 {hover.Node.ProfilerSetupBytes} B of alloc is profiler warmup overhead");
+        }
+        ImGui.EndTooltip();
+    }
+
+    private static void RenderNode(
+        ImDrawListPtr drawList,
+        Profiler.ScopeNode node,
+        Profiler.Frame frame,
+        int frameIndex,
+        float baseY,
+        int depth,
+        double visibleStartTicks,
+        double visibleDurationTicks,
+        float originX,
+        float originY,
+        float width,
+        float clipMinX,
+        float clipMaxX,
+        float clipMaxY,
+        ref HoverEntry? hovered,
+        uint textColor,
+        uint headerColor,
+        uint headerHoveredColor,
+        bool selectedFrame)
+    {
+        long nodeEndTime = Math.Max(node.EndTime, node.StartTime + 1);
+        if (nodeEndTime < visibleStartTicks || node.StartTime > visibleStartTicks + visibleDurationTicks)
+        {
+            return;
+        }
+
+        float y = originY + baseY + (depth * (BarHeight + BarPadding));
+        if (y > clipMaxY)
+        {
+            return;
+        }
+
+        float minX = ToScreenX(node.StartTime, visibleStartTicks, visibleDurationTicks, originX, width);
+        float maxX = ToScreenX(nodeEndTime, visibleStartTicks, visibleDurationTicks, originX, width);
+        if (maxX < clipMinX || minX > clipMaxX)
+        {
+            return;
+        }
+
+        float barWidth = Math.Max(1f, maxX - minX);
+        Vector2 min = new Vector2(minX, y + (BarPadding * 0.5f));
+        Vector2 max = new Vector2(minX + barWidth, y + (BarPadding * 0.5f) + BarHeight);
+
+        uint barColor = BuildBarColor(node.Label, depth, selectedFrame);
+        uint borderColor = LerpColor(headerColor, headerHoveredColor, 0.45f);
+        drawList.AddRectFilled(min, max, barColor, 3f);
+        drawList.AddRect(min, max, SetAlpha(borderColor, 0.55f), 3f, ImDrawFlags.None, 1f);
+
+        Vector2 mousePos = ImGui.GetMousePos();
+        bool isHovered = mousePos.X >= min.X && mousePos.X <= max.X && mousePos.Y >= min.Y && mousePos.Y <= max.Y;
+        if (isHovered)
+        {
+            hovered = new HoverEntry(node, frame, frameIndex, depth, visibleStartTicks);
+            drawList.AddRect(min, max, headerHoveredColor, 3f, ImDrawFlags.None, 1.5f);
+        }
+
+        if (barWidth > MinTextWidth)
+        {
+            string label = node.Label;
+            float textWidth = ImGui.CalcTextSize(label).X;
+            if (textWidth + 8f <= barWidth)
+            {
+                drawList.AddText(new Vector2(min.X + 4f, min.Y + 2f), textColor, label);
+            }
+        }
+
+        for (int i = 0; i < node.Children.Count; i++)
+        {
+            RenderNode(
+                drawList,
+                node.Children[i],
+                frame,
+                frameIndex,
+                baseY,
+                depth + 1,
+                visibleStartTicks,
+                visibleDurationTicks,
+                originX,
+                originY,
+                width,
+                clipMinX,
+                clipMaxX,
+                clipMaxY,
+                ref hovered,
+                textColor,
+                headerColor,
+                headerHoveredColor,
+                selectedFrame);
+        }
+    }
+
+    private static Dictionary<int, int> BuildThreadDepths(IReadOnlyList<Profiler.Frame> frames, int startIndex, int endIndex)
+    {
+        Dictionary<int, int> threadMaxDepths = new Dictionary<int, int>(8);
+        for (int frameIndex = startIndex; frameIndex <= endIndex; frameIndex++)
+        {
+            foreach (Profiler.ScopeNode root in frames[frameIndex].RootNodes)
+            {
+                int maxDepth = 0;
+                for (int i = 0; i < root.Children.Count; i++)
+                {
+                    maxDepth = Math.Max(maxDepth, GetMaxDepth(root.Children[i], 0));
+                }
+
+                if (threadMaxDepths.TryGetValue(root.ManagedThreadId, out int currentMax))
+                {
+                    if (maxDepth > currentMax)
+                    {
+                        threadMaxDepths[root.ManagedThreadId] = maxDepth;
+                    }
+                }
+                else
+                {
+                    threadMaxDepths[root.ManagedThreadId] = maxDepth;
+                }
+            }
+        }
+
+        return threadMaxDepths;
+    }
+
+    private static int GetMaxDepth(Profiler.ScopeNode node, int depth)
+    {
+        if (node.Children.Count == 0)
+        {
+            return depth;
+        }
+
+        int maxDepth = depth;
+        for (int i = 0; i < node.Children.Count; i++)
+        {
+            maxDepth = Math.Max(maxDepth, GetMaxDepth(node.Children[i], depth + 1));
         }
         return maxDepth;
     }
+
+    private static uint BuildBarColor(string label, int depth, bool selectedFrame)
+    {
+        int hash = label.GetHashCode();
+        float hue = ((hash & 1023) / 1023f + (depth * 0.031f)) % 1f;
+        ImGui.ColorConvertHSVtoRGB(hue, 0.52f, selectedFrame ? 0.82f : 0.68f, out float r, out float g, out float b);
+
+        Vector4 frameBg = ImGui.ColorConvertU32ToFloat4(ImGui.GetColorU32(ImGuiCol.FrameBg));
+        Vector4 accent = new Vector4(r, g, b, 1f);
+        Vector4 mixed = Vector4.Lerp(frameBg, accent, 0.72f);
+        return ImGui.ColorConvertFloat4ToU32(mixed);
+    }
+
+    private static float ToScreenX(double ticks, double visibleStartTicks, double visibleDurationTicks, float startX, float width)
+    {
+        double normalized = (ticks - visibleStartTicks) / visibleDurationTicks;
+        return startX + (float)(normalized * width);
+    }
+
+    private static double TicksToMilliseconds(double ticks)
+    {
+        return ticks * TickToMs;
+    }
+
+    private static uint SetAlpha(uint color, float alpha)
+    {
+        Vector4 c = ImGui.ColorConvertU32ToFloat4(color);
+        c.W *= alpha;
+        return ImGui.ColorConvertFloat4ToU32(c);
+    }
+
+    private static uint LerpColor(uint a, uint b, float t)
+    {
+        Vector4 av = ImGui.ColorConvertU32ToFloat4(a);
+        Vector4 bv = ImGui.ColorConvertU32ToFloat4(b);
+        return ImGui.ColorConvertFloat4ToU32(Vector4.Lerp(av, bv, Math.Clamp(t, 0f, 1f)));
+    }
 }

Nerfed.Runtime/Systems/LocalToWorldSystem.cs

@@ -15,19 +15,17 @@ namespace Nerfed.Runtime.Systems
 {
     public class LocalToWorldSystem : MoonTools.ECS.System
     {
-        private readonly bool useParallelFor = true; // When having a low amount of transforms or when in debug mode this might be slower.
+        private readonly JobSystem jobs;
         private readonly Filter rootEntitiesFilter;
         private readonly Filter entitiesWithoutLocalToWorldFilter;
-        private readonly Action<int> updateWorldTransform;
+        private readonly Action<int> updateWorldTransformByIndex;
 
-        public LocalToWorldSystem(World world) : base(world)
+        public LocalToWorldSystem(World world, JobSystem jobs = null) : base(world)
         {
+            this.jobs = jobs ?? JobSystem.Default;
             rootEntitiesFilter = FilterBuilder.Include<LocalTransform>().Exclude<Child>().Build();
-            if (useParallelFor)
+            entitiesWithoutLocalToWorldFilter = FilterBuilder.Include<LocalTransform>().Exclude<LocalToWorld>().Build();
-            {
+            updateWorldTransformByIndex = UpdateWorldTransformByIndex;
-                entitiesWithoutLocalToWorldFilter = FilterBuilder.Include<LocalTransform>().Exclude<LocalToWorld>().Build();
-                updateWorldTransform = UpdateWorldTransformByIndex;
-            }
         }
 
         public override void Update(TimeSpan delta)
@@ -37,20 +35,18 @@ namespace Nerfed.Runtime.Systems
                 return;
             }
 
-            if (useParallelFor)
+            if (this.jobs.WorkerCount > 0)
             {
-                Profiler.BeginSample("ParallelFor.LocalToWorldCheck");
+                Profiler.BeginSample("LocalToWorldCheck");
-                // This check is needed because some entities might not have a LocalToWorld component yet.
+                // Structural pre-pass: ensure LocalToWorld exists on all entities before parallel writes.
-                // Adding this during the loop will break.
+                foreach (Entity entity in entitiesWithoutLocalToWorldFilter.Entities)
-                foreach (Entity entity in entitiesWithoutLocalToWorldFilter.Entities) {
+                {
                     Set(entity, new LocalToWorld(Matrix4x4.Identity));
                 }
                 Profiler.EndSample();
 
-                Profiler.BeginSample("ParallelFor.LocalToWorldUpdate");
+                Profiler.BeginSample("LocalToWorldUpdate");
-                // This should only be used when the filter doesn't change by executing these functions!
+                this.jobs.Dispatch(rootEntitiesFilter.Count, updateWorldTransformByIndex);
-                // So no entity deletion or setting/removing of components used by the filters in this loop.
-                Parallel.For(0, rootEntitiesFilter.Count, updateWorldTransform);
                 Profiler.EndSample();
             }
             else
@@ -66,22 +62,25 @@ namespace Nerfed.Runtime.Systems
 
         private void UpdateWorldTransformByIndex(int entityFilterIndex)
         {
-            Profiler.BeginSample("UpdateWorldTransformByIndex");
+            using ProfilerScope scope = new("UpdateWorldTransformByIndex");
             Entity entity = rootEntitiesFilter.NthEntity(entityFilterIndex);
             UpdateWorldTransform(entity, Matrix4x4.Identity);
-            Profiler.EndSample();
         }
 
         private void UpdateWorldTransform(in Entity entity, Matrix4x4 localToWorldMatrix)
         {
-            // TODO: Only update dirty transforms.
-            // If a parent is dirty all the children need to update their localToWorld matrix.
-            // How do we check if something is dirty? How do we know if a LocalTransform has been changed?
             if (Has<LocalTransform>(entity))
             {
                 LocalTransform localTransform = Get<LocalTransform>(entity);
                 localToWorldMatrix = Matrix4x4.Multiply(localToWorldMatrix, localTransform.TRS());
                 LocalToWorld localToWorld = new(localToWorldMatrix);
+#if DEBUG
+                if (!Has<LocalToWorld>(entity))
+                {
+                    throw new InvalidOperationException(
+                        $"Entity {entity} is missing LocalToWorld. Ensure the structural pre-pass runs before parallel dispatch.");
+                }
+#endif
                 Set(entity, localToWorld);
             }
 

Nerfed.Runtime/Util/BoundedQueue.cs

@@ -5,6 +5,7 @@ namespace Nerfed.Runtime;
 public class BoundedQueue<T> : IEnumerable<T>, ICollection, IReadOnlyCollection<T>
 {
     private readonly Queue<T> queue = null;
+    private readonly object syncLock = new object();
     private readonly int maxSize = 10;
     private T lastAddedElement;
 
@@ -16,58 +17,129 @@ public class BoundedQueue<T> : IEnumerable<T>, ICollection, IReadOnlyCollection<
 
     public void Enqueue(T item)
     {
-        queue.Enqueue(item);
+        Enqueue(item, out _);
-        if (queue.Count > maxSize)
+    }
-        {
-            queue.Dequeue(); // Remove the oldest element
-        }
 
-        lastAddedElement = item;
+    public bool Enqueue(T item, out T evictedItem)
+    {
+        lock (syncLock)
+        {
+            queue.Enqueue(item);
+            if (queue.Count > maxSize)
+            {
+                evictedItem = queue.Dequeue();
+                lastAddedElement = item;
+                return true;
+            }
+
+            evictedItem = default;
+            lastAddedElement = item;
+            return false;
+        }
     }
 
     public T Dequeue()
     {
-        return queue.Dequeue();
+        lock (syncLock)
+        {
+            return queue.Dequeue();
+        }
     }
 
     public T Peek()
     {
-        return queue.Peek();
+        lock (syncLock)
+        {
+            return queue.Peek();
+        }
     }
 
     public T LastAddedElement()
     {
-        return lastAddedElement;
+        lock (syncLock)
+        {
+            return lastAddedElement;
+        }
     }
 
     public void Clear()
     {
-        queue.Clear();
+        lock (syncLock)
+        {
+            queue.Clear();
+        }
     }
 
     public bool Contains(T item)
     {
-        return queue.Contains(item);
+        lock (syncLock)
+        {
+            return queue.Contains(item);
+        }
+    }
+
+    // Iterates the internal Queue<T> directly (struct enumerator, no allocation) under the lock.
+    public int CopyTo(List<T> destination)
+    {
+        lock (syncLock)
+        {
+            destination.Clear();
+            foreach (T item in queue)
+            {
+                destination.Add(item);
+            }
+
+            return destination.Count;
+        }
     }
 
     public IEnumerator<T> GetEnumerator()
     {
-        return queue.GetEnumerator();
+        T[] snapshot;
+        lock (syncLock)
+        {
+            snapshot = queue.ToArray();
+        }
+
+        return ((IEnumerable<T>)snapshot).GetEnumerator();
     }
 
     IEnumerator IEnumerable.GetEnumerator()
     {
-        return queue.GetEnumerator();
+        return GetEnumerator();
     }
 
     public void CopyTo(Array array, int index)
     {
-        ((ICollection)queue).CopyTo(array, index);
+        lock (syncLock)
+        {
+            ((ICollection)queue).CopyTo(array, index);
+        }
+    }
+
+    public int Count
+    {
+        get
+        {
+            lock (syncLock)
+            {
+                return queue.Count;
+            }
+        }
     }
 
-    public int Count => queue.Count;
     public int Capacity => maxSize;
-    public bool IsSynchronized => ((ICollection)queue).IsSynchronized;
+    public bool IsSynchronized => true;
-    public object SyncRoot => ((ICollection)queue).SyncRoot;
+    public object SyncRoot => syncLock;
-    int IReadOnlyCollection<T>.Count => queue.Count;
+
+    int IReadOnlyCollection<T>.Count
+    {
+        get
+        {
+            lock (syncLock)
+            {
+                return queue.Count;
+            }
+        }
+    }
 }