Enhance Profiler with pre-allocated sort buffers and frame management improvements

Nerfed.Runtime/Profiler.cs

@@ -110,12 +110,15 @@ public static class Profiler
     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)
         {
-            values = new double[Math.Max(8, capacity)];
+            int size = Math.Max(8, capacity);
+            values = new double[size];
+            sortBuffer = new double[size];
         }
 
         public void Add(double value)
@@ -138,21 +141,20 @@ public static class Profiler
             double sum = 0;
             double min = double.MaxValue;
             double max = double.MinValue;
-            double[] sorted = new double[count];
 
             int start = (index - count + values.Length) % values.Length;
             for (int i = 0; i < count; i++)
             {
                 double value = values[(start + i) % values.Length];
-                sorted[i] = value;
+                sortBuffer[i] = value;
                 sum += value;
                 min = Math.Min(min, value);
                 max = Math.Max(max, value);
             }
 
-            Array.Sort(sorted);
+            Array.Sort(sortBuffer, 0, count);
             int percentileIndex = (int)Math.Ceiling((count - 1) * 0.95d);
-            double p95 = sorted[Math.Clamp(percentileIndex, 0, count - 1)];
+            double p95 = sortBuffer[Math.Clamp(percentileIndex, 0, count - 1)];
             return new RollingLabelMetrics(sum / count, min, max, p95, count);
         }
     }
@@ -160,6 +162,7 @@ public static class Profiler
     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;
@@ -168,6 +171,7 @@ public static class Profiler
         {
             int size = Math.Max(8, capacity);
             durations = new double[size];
+            sortBuffer = new double[size];
             misses = new byte[size];
         }
 
@@ -192,7 +196,6 @@ public static class Profiler
             double sum = 0;
             double max = double.MinValue;
             int budgetMisses = 0;
-            double[] sorted = new double[count];
 
             int start = (index - count + durations.Length) % durations.Length;
             for (int i = 0; i < count; i++)
@@ -202,39 +205,44 @@ public static class Profiler
                 sum += value;
                 max = Math.Max(max, value);
                 budgetMisses += misses[at];
-                sorted[i] = value;
+                sortBuffer[i] = value;
             }
 
-            Array.Sort(sorted);
+            Array.Sort(sortBuffer, 0, count);
             int percentileIndex = (int)Math.Ceiling((count - 1) * 0.95d);
-            double p95 = sorted[Math.Clamp(percentileIndex, 0, count - 1)];
+            double p95 = sortBuffer[Math.Clamp(percentileIndex, 0, count - 1)];
             return new RollingThreadMetrics(sum / count, p95, max, count, budgetMisses);
         }
     }
 
-    public class Frame(uint frameCount)
+    public class Frame
     {
-        public uint FrameCount { get; } = frameCount;
-        public long StartTime { get; } = Stopwatch.GetTimestamp();
-        public long EndTime { get; private set; }
-
         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; }
 
         public IReadOnlyList<ScopeNode> RootNodes => rootNodes;
-
         public IReadOnlyDictionary<string, LabelMetrics> LabelMetrics => labelMetrics;
         public IReadOnlyDictionary<string, LabelMetrics> CategoryMetrics => categoryMetrics;
         public IReadOnlyDictionary<int, ThreadMetrics> ThreadMetrics => threadMetrics;
-        public long AllocatedBytesStart { get; } = GC.GetTotalAllocatedBytes(false);
+
+        // 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; } = GC.CollectionCount(0);
+        public int Gen0CollectionsStart { get; private set; }
-        public int Gen1CollectionsStart { get; } = GC.CollectionCount(1);
+        public int Gen1CollectionsStart { get; private set; }
-        public int Gen2CollectionsStart { get; } = GC.CollectionCount(2);
+        public int Gen2CollectionsStart { get; private set; }
         public int Gen0CollectionsEnd { get; private set; }
         public int Gen1CollectionsEnd { get; private set; }
         public int Gen2CollectionsEnd { get; private set; }
@@ -244,6 +252,31 @@ public static class Profiler
         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)
@@ -282,6 +315,7 @@ public static class Profiler
             labelMetrics.Clear();
             categoryMetrics.Clear();
             threadMetrics.Clear();
+            knownThreadIds.Clear();
             lock (rootNodesLock)
             {
                 for (int i = 0; i < rootNodes.Count; i++)
@@ -289,11 +323,12 @@ public static class Profiler
                     AccumulateLabelMetrics(rootNodes[i]);
                 }
 
-                foreach (ScopeNode rootNode in rootNodes)
+                for (int i = 0; i < rootNodes.Count; i++)
                 {
-                    for (int i = 0; i < rootNode.Children.Count; i++)
+                    ScopeNode rootNode = rootNodes[i];
+                    for (int j = 0; j < rootNode.Children.Count; j++)
                     {
-                        AccumulateThreadMetrics(rootNode.ManagedThreadId, rootNode.Children[i]);
+                        AccumulateThreadMetrics(rootNode.ManagedThreadId, rootNode.Children[j]);
                     }
                 }
             }
@@ -352,6 +387,7 @@ public static class Profiler
             else
             {
                 threadMetrics[threadId] = new ThreadMetrics(inclusiveMs, selfMs, 1, false);
+                knownThreadIds.Add(threadId);
             }
 
             for (int i = 0; i < node.Children.Count; i++)
@@ -368,10 +404,10 @@ public static class Profiler
                 return;
             }
 
-            int[] keys = threadMetrics.Keys.ToArray();
+            // knownThreadIds avoids Keys.ToArray() allocation
-            for (int i = 0; i < keys.Length; i++)
+            for (int i = 0; i < knownThreadIds.Count; i++)
             {
-                int key = keys[i];
+                int key = knownThreadIds[i];
                 ThreadMetrics metric = threadMetrics[key];
                 threadMetrics[key] = new ThreadMetrics(metric.InclusiveMs, metric.SelfMs, metric.Calls, metric.InclusiveMs > perThreadBudget);
             }
@@ -393,7 +429,7 @@ public static class Profiler
         internal void Reset(string label, string category, ulong tagMask, int managedThreadId, ScopeNode parent)
         {
             Label = label;
-            Category = string.IsNullOrWhiteSpace(category) ? DefaultCategory : category;
+            Category = string.IsNullOrEmpty(category) ? DefaultCategory : category;
             TagMask = tagMask;
             ManagedThreadId = managedThreadId;
             Parent = parent;
@@ -429,7 +465,6 @@ public static class Profiler
             return ((double)selfTicks) * 1000 / Stopwatch.Frequency;
         }
 
-        // Add a child node (used for nested scopes)
         internal ScopeNode AddChild(string label, string category, ulong tagMask)
         {
             ScopeNode child = RentNode(label, category, tagMask, ManagedThreadId, this);
@@ -448,13 +483,27 @@ public static class Profiler
     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<ThreadProfilerState> threadStates = new ThreadLocal<ThreadProfilerState>(() => new ThreadProfilerState(), 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();
@@ -474,13 +523,7 @@ public static class Profiler
 
     public static int CopyFramesTo(List<Frame> destination)
     {
-        destination.Clear();
+        return Frames.CopyTo(destination);
-        foreach (Frame frame in Frames)
-        {
-            destination.Add(frame);
-        }
-
-        return destination.Count;
     }
 
     public static IReadOnlyDictionary<string, RollingLabelMetrics> GetRollingLabelMetricsSnapshot()
@@ -524,7 +567,7 @@ public static class Profiler
             FinalizeCurrentFrame();
         }
 
-        currentFrame = new Frame(frameCount);
+        currentFrame = RentFrame(frameCount);
     }
 
     [Conditional("PROFILING")]
@@ -574,17 +617,14 @@ public static class Profiler
 
         if (scopes.Count == 0)
         {
-            // First scope for this thread (new root for this thread)
             int threadId = state.ThreadId;
             ScopeNode rootScopeNode = RentNode(GetThreadRootLabel(threadId), DefaultCategory, 0, threadId, null);
             scopes.Push(rootScopeNode);
             frame.AddRootNode(rootScopeNode);
         }
 
-        // Create a new child under the current top of the stack
         ScopeNode newScope = scopes.Peek().AddChild(label, category, tagMask);
-
+        scopes.Push(newScope);
-        scopes.Push(newScope);  // Push new scope to the thread's stack
     }
 
     [Conditional("PROFILING")]
@@ -628,6 +668,17 @@ public static class Profiler
         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))
@@ -658,17 +709,19 @@ public static class Profiler
             return;
         }
 
-        foreach (ThreadProfilerState state in threadStates.Values)
+        lock (registeredThreadStatesLock)
         {
-            Stack<ScopeNode> scopes = state.Scopes;
+            for (int i = 0; i < registeredThreadStates.Count; i++)
-            while (scopes.Count > 0)
             {
-                ScopeNode currentScope = scopes.Pop();
+                ThreadProfilerState state = registeredThreadStates[i];
-                currentScope.End();
+                Stack<ScopeNode> scopes = state.Scopes;
-            }
+                while (scopes.Count > 0)
+                {
+                    scopes.Pop().End();
+                }
 
-            scopes.Clear();
+                state.CaptureDecisions.Clear();
-            state.CaptureDecisions.Clear();
+            }
         }
 
         frame.End(FrameBudgetMilliseconds);
@@ -679,6 +732,8 @@ public static class Profiler
             {
                 ReturnNodeTree(evictedFrame.RootNodes[i]);
             }
+
+            framePool.Add(evictedFrame);
         }
 
         UpdateRollingWindows(frame);
@@ -691,7 +746,7 @@ public static class Profiler
     {
         lock (rollingWindowsLock)
         {
-            foreach (KeyValuePair<string, LabelMetrics> pair in frame.LabelMetrics)
+            foreach (KeyValuePair<string, LabelMetrics> pair in frame.LabelMetricsRaw)
             {
                 if (!rollingWindows.TryGetValue(pair.Key, out RollingWindow window))
                 {
@@ -708,7 +763,7 @@ public static class Profiler
     {
         lock (rollingWindowsLock)
         {
-            foreach (KeyValuePair<int, ThreadMetrics> pair in frame.ThreadMetrics)
+            foreach (KeyValuePair<int, ThreadMetrics> pair in frame.ThreadMetricsRaw)
             {
                 if (!rollingThreadWindows.TryGetValue(pair.Key, out RollingThreadWindow window))
                 {

Nerfed.Runtime/ProfilerVisualizer.cs

@@ -283,15 +283,22 @@ public static class ProfilerVisualizer
             return;
         }
 
-        bool isZoom = io.KeyCtrl;
+        if (!io.KeyCtrl && !io.KeyShift)
-        if (isZoom)
+        {
+            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, io.MouseWheel), 1f, 128f);
+            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);
@@ -303,7 +310,8 @@ public static class ProfilerVisualizer
         }
         else
         {
-            state.PanTicks -= io.MouseWheel * (visibleDurationTicks * 0.10d);
+            // Shift + scroll: horizontal pan
+            state.PanTicks -= wheel * (visibleDurationTicks * 0.10d);
             state.FollowLatest = false;
             userNavigated = true;
         }

Nerfed.Runtime/Util/BoundedQueue.cs

@@ -78,6 +78,21 @@ public class BoundedQueue<T> : IEnumerable<T>, ICollection, IReadOnlyCollection<
         }
     }
 
+    // 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()
     {
         T[] snapshot;