async improvement

src/Poe2Trade.Screen/DesktopDuplication.cs

@@ -83,16 +83,26 @@ public sealed class DesktopDuplication : IScreenCapture
 
         var mapped = _context.Map(_staging!, 0, MapMode.Read);
 
-        var mat = Mat.FromPixelData(h, w, MatType.CV_8UC4, mapped.DataPointer, (int)mapped.RowPitch);
+        // GPU copy is complete once Map returns — release DXGI frame immediately
+        // so the DWM compositor can recycle the buffer (~0.5ms hold vs ~2.5ms before).
+        srcTexture.Dispose();
+        resource.Dispose();
+        _duplication!.ReleaseFrame();
 
-        var duplication = _duplication!;
-        return new ScreenFrame(mat, () =>
+        // CPU copy from our staging texture (no DXGI dependency, ~1.5ms for 2560×1440)
+        var mat = new Mat(h, w, MatType.CV_8UC4);
+        unsafe
         {
-            _context.Unmap(_staging!, 0);
-            srcTexture.Dispose();
-            resource.Dispose();
-            duplication.ReleaseFrame();
-        });
+            var src = (byte*)mapped.DataPointer;
+            var dst = (byte*)mat.Data;
+            var rowBytes = w * 4;
+            var pitch = (int)mapped.RowPitch;
+            for (var row = 0; row < h; row++)
+                Buffer.MemoryCopy(src + row * pitch, dst + row * rowBytes, rowBytes, rowBytes);
+        }
+
+        _context.Unmap(_staging!, 0);
+        return new ScreenFrame(mat, () => mat.Dispose());
     }
 
     public unsafe Mat? CaptureRegion(Region region)