Replace Throttle Busy Wait With Dequeue Signal And Byte Limit

Core/RenderManager.cs

@@ -5,7 +5,6 @@ using System.IO;
 using System.Threading;
 using System.Buffers;
 using System.Diagnostics;
-using System.Linq;
 using ADOFAIRenderer.API;
 using ADOFAIRenderer.Core.Data;
 using ADOFAIRenderer.Data;
@@ -81,8 +80,28 @@ internal class RenderManager : MonoBehaviour
     public static bool IsActive => Instance && Instance.stage >= RenderStage.OnScene;
     public static bool IsCapturingActive => Instance && Instance.stage >= RenderStage.Render;
 
+    /// <summary>
+    /// Maximum number of frames in ImageQueue. The main thread waits (throttles) while the queue holds more frames.
+    /// </summary>
     public const int MaxQueuedFrames = 64;
 
+    /// <summary>
+    /// Upper bound of the readback buffer bytes of the frames in ImageQueue (4 GiB).
+    /// The frame limit becomes min(MaxQueuedFrames, MaxQueuedBytes / bytes per frame) (see ComputeQueueFrameLimit),
+    /// so it only lowers the frame count at resolutions where MaxQueuedFrames frames exceed this size.
+    /// </summary>
+    public const long MaxQueuedBytes = 4L * 1024 * 1024 * 1024;
+
+    /// <summary>
+    /// Wait time of the throttle loop (OnPostFrame) for the render thread to take a frame from ImageQueue.
+    /// When it elapses without a dequeue, AsyncGPUReadback.WaitAllRequests() is called once so the readbacks
+    /// still in flight complete (the render thread may be waiting for one of them to fill a gap in ImageQueue).
+    /// </summary>
+    private const int ThrottleWaitMilliseconds = 10;
+
+    /// <summary>Frame limit of ImageQueue applied by the last throttle check (MaxQueuedFrames or lower).</summary>
+    public int QueueFrameLimit { get; private set; } = MaxQueuedFrames;
+        
     // Statistics
     public long framesSent;
     private readonly Stopwatch _renderStopwatch = new();
@@ -213,6 +232,8 @@ internal class RenderManager : MonoBehaviour
     // Thread and Queue State
     internal SortedList<long, CapturedFrame> ImageQueue = new();
     internal AutoResetEvent FrameAvailableEvent = new(false);
+    // Set by the render thread each time it takes a frame from ImageQueue; the throttle loop waits on it.
+    internal AutoResetEvent FrameDequeuedEvent = new(false);
     public Thread? RenderThread;
     public long passCount;
     public volatile bool endRequested;
@@ -439,13 +460,30 @@ internal class RenderManager : MonoBehaviour
         _holdFloorsSourceCount = floors.Count;
     }
 
+    /// <summary>
+    /// Frame limit of ImageQueue for the given capture texture: MaxQueuedBytes divided by the readback bytes of
+    /// one frame (YUV 4:2:0 when the GPU conversion is active, RGBA otherwise), within 4 to MaxQueuedFrames.
+    /// </summary>
+    private int ComputeQueueFrameLimit(RenderTexture texture)
+    {
+        var frameBytes = _frameReader?.GetFrameBufferBytes(texture.width, texture.height)
+                         ?? (long)texture.width * texture.height * 4;
+        if (frameBytes <= 0) return MaxQueuedFrames;
+
+        return (int)Math.Min(MaxQueuedFrames, Math.Max(4, MaxQueuedBytes / frameBytes));
+    }
+
     // ReSharper disable once Unity.IncorrectMethodSignature
     public void OnPostFrame(RenderTexture? texture)
     {
         if (Instance.stage < RenderStage.Render || !texture) return;
 
         // [Backpressure Throttling]
+        // The main thread waits while ImageQueue holds more than QueueFrameLimit frames. Only the render thread
+        // removes frames, and it signals FrameDequeuedEvent for each one. The wait has a timeout so that the loop
+        // re-checks the render thread state and pumps pending readback completions (see ThrottleWaitMilliseconds).
         isThrottling = false;
+        var frameLimit = QueueFrameLimit = ComputeQueueFrameLimit(texture!);
 
         while (Instance.stage >= RenderStage.Render)
         {
@@ -461,7 +499,7 @@ internal class RenderManager : MonoBehaviour
                 count = ImageQueue.Count;
             }
 
-            if (count <= MaxQueuedFrames) break;
+            if (count <= frameLimit) break;
 
             if (!isThrottling)
             {
@@ -469,7 +507,10 @@ internal class RenderManager : MonoBehaviour
                 throttleCount++;
             }
                 
-            AsyncGPUReadback.WaitAllRequests();
+            if (!FrameDequeuedEvent.WaitOne(ThrottleWaitMilliseconds))
+            {
+                AsyncGPUReadback.WaitAllRequests();
+            }
         }
         isThrottling = false;
 
@@ -649,6 +690,7 @@ internal class RenderManager : MonoBehaviour
         endRequested = false;
         endSuccessful = false;
         FrameAvailableEvent.Reset();
+        FrameDequeuedEvent.Reset();
 
         var renderFPS = Configuration.RenderFPS;
         var videoFPS = Configuration.VideoFPS;
@@ -712,7 +754,7 @@ internal class RenderManager : MonoBehaviour
             throw;
         }
 
-        var encoder = videoFFmpeg!;
+        var encoder = videoFFmpeg;
         RenderThread = new Thread(() => RunRenderThread(encoder))
         {
             IsBackground = true,
@@ -775,6 +817,7 @@ internal class RenderManager : MonoBehaviour
 
             if (TryDequeueNextFrame(out var frame))
             {
+                FrameDequeuedEvent.Set();
                 SendVideoFrame(videoFFmpeg, frame);
                 WriteAudioForFrame(videoFFmpeg);
                 continue;

UI/DebugMonitorUI.cs

@@ -164,7 +164,7 @@ internal class DebugMonitorUI : MonoBehaviour
             }
         }
             
-        sb.AppendLine($"• Queue: <b>{queueCount}</b> / {RenderManager.MaxQueuedFrames} frames");
+        sb.AppendLine($"• Queue: <b>{queueCount}</b> / {(RenderManager.Instance ? RenderManager.Instance.QueueFrameLimit : RenderManager.MaxQueuedFrames)} frames");
         sb.AppendLine($"• Rendered Frames: <b>{(RenderManager.Instance ? RenderManager.Instance.frameCount : 0)}</b>");
         sb.AppendLine($"• Frames Sent: <b>{(RenderManager.Instance ? Interlocked.Read(ref RenderManager.Instance.framesSent) : 0)}</b>");
             

Util/FrameReader.cs

@@ -73,6 +73,16 @@ namespace ADOFAIRenderer.Util
             }
         }
 
+        /// <summary>
+        /// Bytes of the readback buffer of one frame: Width x Height x 3 / 2 when the GPU conversion to NV12 / YUV420P
+        /// is used, Width x Height x 4 (RGBA) otherwise. The same size as ReadbackRequestState.Setup uses.
+        /// </summary>
+        public long GetFrameBufferBytes(int width, int height)
+        {
+            var pixels = (long)width * height;
+            return _useGPUConversion ? pixels * 3 / 2 : pixels * 4;
+        }
+
         public void RequestFrameCapture(Texture texture, long currentFrame)
         {
             if (!RenderManager.IsCapturingActive) return;