Use cv::swap and copyTo to eliminate redundant frame copies

Replace clone() with cv::swap() in video_tick for both background
and enhance filters. Swap is O(1) — it transfers buffer ownership
to the tick thread instead of deep-copying ~8 MB per frame.

In the render thread, use copyTo() instead of clone() for the stage
surface data. copyTo reuses the existing cv::Mat allocation when
dimensions match, avoiding a fresh heap allocation every frame.

Add a newFrameAvailable flag (protected by inputBGRALock) so the
tick thread only processes genuinely new frames and doesn't
re-process stale data after swap empties the shared buffer.

Signed-off-by: Xavier Ruiz <github@xav.ie>

Author: xav-ie

src/FilterData.hpp

@@ -31,6 +31,7 @@ struct filter_data : public ORTModelData, public std::enable_shared_from_this<fi
 	gs_stagesurf_t *stagesurface;
 
 	cv::Mat inputBGRA;
+	std::atomic<bool> newFrameAvailable{false};
 
 	std::atomic<bool> isDisabled{false};
 

src/background-filter.cpp

@@ -536,11 +536,11 @@ void background_filter_video_tick(void *data, float seconds)
 			// No data to process
 			return;
 		}
-		if (tf->inputBGRA.empty()) {
-			// No data to process
+		if (!tf->newFrameAvailable) {
 			return;
 		}
-		imageBGRA = tf->inputBGRA.clone();
+		cv::swap(imageBGRA, tf->inputBGRA);
+		tf->newFrameAvailable = false;
 	}
 
 	if (tf->enableImageSimilarity) {

src/enhance-filter.cpp

@@ -247,18 +247,18 @@ void enhance_filter_video_tick(void *data, float seconds)
 		return;
 	}
 
-	if (tf->inputBGRA.empty()) {
-		return;
-	}
-
 	// Get input image from source rendering pipeline
 	cv::Mat imageBGRA;
 	{
 		std::unique_lock<std::mutex> lock(tf->inputBGRALock, std::try_to_lock);
 		if (!lock.owns_lock()) {
 			return;
 		}
-		imageBGRA = tf->inputBGRA.clone();
+		if (!tf->newFrameAvailable) {
+			return;
+		}
+		cv::swap(imageBGRA, tf->inputBGRA);
+		tf->newFrameAvailable = false;
 	}
 
 	cv::Mat outputImage;

src/obs-utils/obs-utils.cpp

@@ -67,9 +67,10 @@ bool getRGBAFromStageSurface(filter_data *tf, uint32_t &width, uint32_t &height)
 		std::lock_guard<std::mutex> lock(tf->inputBGRALock);
 		// Create a temporary Mat that wraps the video_data pointer
 		cv::Mat temp(height, width, CV_8UC4, video_data, linesize);
-		// Clone the data to ensure tf->inputBGRA has its own copy
-		// This prevents use-after-unmap race condition
-		tf->inputBGRA = temp.clone();
+		// Copy frame data into tf->inputBGRA, reusing its allocation
+		// when dimensions match. Ensures we own the pixels before unmap.
+		temp.copyTo(tf->inputBGRA);
+		tf->newFrameAvailable = true;
 	}
 	gs_stagesurface_unmap(tf->stagesurface);
 	return true;