Performance chasing

2026-05-11 23:10:45 +10:00
parent c5cead6003
commit a434a88108
18 changed files with 1115 additions and 82 deletions
--- a/apps/LoopThroughWithOpenGLCompositing/gl/RenderEngine.cpp
+++ b/apps/LoopThroughWithOpenGLCompositing/gl/RenderEngine.cpp
@@ -18,6 +18,7 @@ RenderEngine::RenderEngine(
 	mRenderPass(mRenderer),
 	mRenderPipeline(mRenderer, runtimeSnapshotProvider, healthTelemetry, std::move(renderEffect), std::move(screenshotReady), std::move(previewPaint)),
 	mShaderPrograms(mRenderer, runtimeSnapshotProvider),
+	mHealthTelemetry(healthTelemetry),
 	mHdc(hdc),
 	mHglrc(hglrc),
 	mFrameStateResolver(runtimeSnapshotProvider)
@@ -546,7 +547,11 @@ bool RenderEngine::RequestOutputFrame(const RenderPipelineFrameContext& context,
 {
 	if (mRenderThreadRunning)
 	{
-		return TryInvokeOnRenderThread("output-render", [this, &context, &outputFrame]() {
+		const auto queuedAt = std::chrono::steady_clock::now();
+		return TryInvokeOnRenderThread("output-render", [this, &context, &outputFrame, queuedAt]() {
+			const auto startedAt = std::chrono::steady_clock::now();
+			const double queueWaitMilliseconds = std::chrono::duration_cast<std::chrono::duration<double, std::milli>>(startedAt - queuedAt).count();
+			mHealthTelemetry.TryRecordOutputRenderQueueWait(queueWaitMilliseconds);
 			mRenderCommandQueue.RequestOutputFrame({ context.videoState, context.completion });
 			RenderOutputFrameRequest request;
 			return mRenderCommandQueue.TryTakeOutputFrame(request) &&
--- a/apps/LoopThroughWithOpenGLCompositing/gl/RenderEngine.h
+++ b/apps/LoopThroughWithOpenGLCompositing/gl/RenderEngine.h
@@ -209,6 +209,7 @@ private:
 	OpenGLRenderPass mRenderPass;
 	OpenGLRenderPipeline mRenderPipeline;
 	OpenGLShaderPrograms mShaderPrograms;
+	HealthTelemetry& mHealthTelemetry;
 	HDC mHdc;
 	HGLRC mHglrc;

--- a/apps/LoopThroughWithOpenGLCompositing/gl/composite/OpenGLComposite.cpp
+++ b/apps/LoopThroughWithOpenGLCompositing/gl/composite/OpenGLComposite.cpp
@@ -164,6 +164,9 @@ error:

 void OpenGLComposite::paintGL(bool force)
 {
+	if (mRuntimeUpdateController)
+		mRuntimeUpdateController->ProcessRuntimeWork();
+
 	if (!force)
 	{
 		if (IsIconic(hGLWnd))
@@ -171,6 +174,12 @@ void OpenGLComposite::paintGL(bool force)
 	}

 	const unsigned previewFps = mRuntimeStore ? mRuntimeStore->GetConfiguredPreviewFps() : 30u;
+	if (!force && mVideoBackend && mVideoBackend->ShouldPrioritizeOutputOverPreview())
+	{
+		ValidateRect(hGLWnd, NULL);
+		return;
+	}
+
 	if (!mRenderEngine->TryPresentPreview(force, previewFps, mVideoBackend->OutputFrameWidth(), mVideoBackend->OutputFrameHeight()))
 	{
 		ValidateRect(hGLWnd, NULL);
@@ -261,6 +270,9 @@ bool OpenGLComposite::Start()
 	if (!mRenderEngine->StartRenderThread())
 		return false;

+	if (mRuntimeUpdateController)
+		mRuntimeUpdateController->ProcessRuntimeWork();
+
 	if (mVideoBackend->Start())
 		return true;

@@ -351,9 +363,6 @@ bool OpenGLComposite::RequestScreenshot(std::string& error)

 void OpenGLComposite::renderEffect()
 {
-	if (mRuntimeUpdateController)
-		mRuntimeUpdateController->ProcessRuntimeWork();
-
 	const RenderFrameInput frameInput = BuildRenderFrameInput();
 	RenderFrame(frameInput);
 }
--- a/apps/LoopThroughWithOpenGLCompositing/gl/pipeline/OpenGLRenderPipeline.cpp
+++ b/apps/LoopThroughWithOpenGLCompositing/gl/pipeline/OpenGLRenderPipeline.cpp
@@ -8,7 +8,9 @@
 #include <cstring>

 #include <chrono>
+#include <cstdlib>
 #include <gl/gl.h>
+#include <string>

 OpenGLRenderPipeline::OpenGLRenderPipeline(
 	OpenGLRenderer& renderer,
@@ -22,7 +24,8 @@ OpenGLRenderPipeline::OpenGLRenderPipeline(
 	mHealthTelemetry(healthTelemetry),
 	mRenderEffect(renderEffect),
 	mOutputReady(outputReady),
-	mPaint(paint)
+	mPaint(paint),
+	mOutputReadbackMode(ReadOutputReadbackModeFromEnvironment())
 {
 }

@@ -53,9 +56,22 @@ bool OpenGLRenderPipeline::RenderFrame(const RenderPipelineFrameContext& context
 	mHealthTelemetry.TryRecordPerformanceStats(state.frameBudgetMilliseconds, renderMilliseconds);
 	mRuntimeSnapshotProvider.AdvanceFrame();

-	ReadOutputFrame(state, outputFrame);
-	if (mPaint)
-		mPaint();
+	OutputReadbackTiming readbackTiming = ReadOutputFrame(state, outputFrame);
+	mHealthTelemetry.TryRecordOutputRenderPipelineTiming(
+		renderMilliseconds,
+		readbackTiming.fenceWaitMilliseconds,
+		readbackTiming.mapMilliseconds,
+		readbackTiming.copyMilliseconds,
+		readbackTiming.cachedCopyMilliseconds,
+		readbackTiming.asyncQueueMilliseconds,
+		readbackTiming.asyncQueueBufferMilliseconds,
+		readbackTiming.asyncQueueSetupMilliseconds,
+		readbackTiming.asyncQueueReadPixelsMilliseconds,
+		readbackTiming.asyncQueueFenceMilliseconds,
+		readbackTiming.syncReadMilliseconds,
+		readbackTiming.asyncReadbackMissed,
+		readbackTiming.cachedFallbackUsed,
+		readbackTiming.syncFallbackUsed);

 	return true;
 }
@@ -151,8 +167,9 @@ void OpenGLRenderPipeline::FlushAsyncReadbackPipeline()
 	mAsyncReadbackReadIndex = 0;
 }

-void OpenGLRenderPipeline::QueueAsyncReadback(const VideoIOState& state)
+bool OpenGLRenderPipeline::QueueAsyncReadback(const VideoIOState& state, OutputReadbackTiming& timing)
 {
+	const auto queueStartTime = std::chrono::steady_clock::now();
 	const bool usePackedOutput = state.outputPixelFormat == VideoIOPixelFormat::V210 || state.outputPixelFormat == VideoIOPixelFormat::Yuva10;
 	const std::size_t requiredBytes = static_cast<std::size_t>(state.outputFrameRowBytes) * state.outputFrameSize.height;
 	const GLenum format = usePackedOutput ? GL_RGBA : GL_BGRA;
@@ -161,8 +178,16 @@ void OpenGLRenderPipeline::QueueAsyncReadback(const VideoIOState& state)
 	const GLsizei readWidth = static_cast<GLsizei>(usePackedOutput ? state.outputPackTextureWidth : state.outputFrameSize.width);
 	const GLsizei readHeight = static_cast<GLsizei>(state.outputFrameSize.height);

+	const auto finishTiming = [&timing, queueStartTime]() {
+		const auto queueEndTime = std::chrono::steady_clock::now();
+		timing.asyncQueueMilliseconds += std::chrono::duration_cast<std::chrono::duration<double, std::milli>>(queueEndTime - queueStartTime).count();
+	};
+
 	if (requiredBytes == 0)
-		return;
+	{
+		finishTiming();
+		return false;
+	}

 	if (mAsyncReadbackBytes != requiredBytes
 		|| mAsyncReadbackFormat != format
@@ -173,30 +198,50 @@ void OpenGLRenderPipeline::QueueAsyncReadback(const VideoIOState& state)
 		mAsyncReadbackType = type;
 		mAsyncReadbackFramebuffer = framebuffer;
 		if (!EnsureAsyncReadbackBuffers(requiredBytes))
-			return;
+		{
+			finishTiming();
+			return false;
+		}
 	}

 	AsyncReadbackSlot& slot = mAsyncReadbackSlots[mAsyncReadbackWriteIndex];
-	if (slot.fence != nullptr)
+	if (slot.inFlight)
 	{
-		glDeleteSync(slot.fence);
-		slot.fence = nullptr;
+		finishTiming();
+		return false;
 	}

+	auto stageStartTime = std::chrono::steady_clock::now();
 	glPixelStorei(GL_PACK_ALIGNMENT, 4);
 	glPixelStorei(GL_PACK_ROW_LENGTH, 0);
 	glBindFramebuffer(GL_READ_FRAMEBUFFER, framebuffer);
 	glBindBuffer(GL_PIXEL_PACK_BUFFER, slot.pixelPackBuffer);
+	auto stageEndTime = std::chrono::steady_clock::now();
+	timing.asyncQueueSetupMilliseconds += std::chrono::duration_cast<std::chrono::duration<double, std::milli>>(stageEndTime - stageStartTime).count();
+
+	stageStartTime = std::chrono::steady_clock::now();
 	glBufferData(GL_PIXEL_PACK_BUFFER, static_cast<GLsizeiptr>(requiredBytes), nullptr, GL_STREAM_READ);
+	stageEndTime = std::chrono::steady_clock::now();
+	timing.asyncQueueBufferMilliseconds += std::chrono::duration_cast<std::chrono::duration<double, std::milli>>(stageEndTime - stageStartTime).count();
+
+	stageStartTime = std::chrono::steady_clock::now();
 	glReadPixels(0, 0, readWidth, readHeight, format, type, nullptr);
+	stageEndTime = std::chrono::steady_clock::now();
+	timing.asyncQueueReadPixelsMilliseconds += std::chrono::duration_cast<std::chrono::duration<double, std::milli>>(stageEndTime - stageStartTime).count();
+
+	stageStartTime = std::chrono::steady_clock::now();
 	slot.fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
+	stageEndTime = std::chrono::steady_clock::now();
+	timing.asyncQueueFenceMilliseconds += std::chrono::duration_cast<std::chrono::duration<double, std::milli>>(stageEndTime - stageStartTime).count();
 	slot.inFlight = slot.fence != nullptr;
 	glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);

 	mAsyncReadbackWriteIndex = (mAsyncReadbackWriteIndex + 1) % mAsyncReadbackSlots.size();
+	finishTiming();
+	return slot.inFlight;
 }

-bool OpenGLRenderPipeline::TryConsumeAsyncReadback(VideoIOOutputFrame& outputFrame, GLuint64 timeoutNanoseconds)
+bool OpenGLRenderPipeline::TryConsumeAsyncReadback(VideoIOOutputFrame& outputFrame, GLuint64 timeoutNanoseconds, OutputReadbackTiming& timing)
 {
 	if (mAsyncReadbackBytes == 0 || outputFrame.bytes == nullptr)
 		return false;
@@ -206,15 +251,24 @@ bool OpenGLRenderPipeline::TryConsumeAsyncReadback(VideoIOOutputFrame& outputFra
 		return false;

 	const GLenum waitFlags = timeoutNanoseconds > 0 ? GL_SYNC_FLUSH_COMMANDS_BIT : 0;
+	const auto waitStartTime = std::chrono::steady_clock::now();
 	const GLenum waitResult = glClientWaitSync(slot.fence, waitFlags, timeoutNanoseconds);
+	const auto waitEndTime = std::chrono::steady_clock::now();
+	timing.fenceWaitMilliseconds += std::chrono::duration_cast<std::chrono::duration<double, std::milli>>(waitEndTime - waitStartTime).count();
 	if (waitResult != GL_ALREADY_SIGNALED && waitResult != GL_CONDITION_SATISFIED)
+	{
+		timing.asyncReadbackMissed = true;
 		return false;
+	}

 	glDeleteSync(slot.fence);
 	slot.fence = nullptr;

 	glBindBuffer(GL_PIXEL_PACK_BUFFER, slot.pixelPackBuffer);
+	const auto mapStartTime = std::chrono::steady_clock::now();
 	void* mappedBytes = glMapBuffer(GL_PIXEL_PACK_BUFFER, GL_READ_ONLY);
+	const auto mapEndTime = std::chrono::steady_clock::now();
+	timing.mapMilliseconds += std::chrono::duration_cast<std::chrono::duration<double, std::milli>>(mapEndTime - mapStartTime).count();
 	if (mappedBytes == nullptr)
 	{
 		glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
@@ -223,7 +277,10 @@ bool OpenGLRenderPipeline::TryConsumeAsyncReadback(VideoIOOutputFrame& outputFra
 		return false;
 	}

+	const auto copyStartTime = std::chrono::steady_clock::now();
 	std::memcpy(outputFrame.bytes, mappedBytes, slot.sizeBytes);
+	const auto copyEndTime = std::chrono::steady_clock::now();
+	timing.copyMilliseconds += std::chrono::duration_cast<std::chrono::duration<double, std::milli>>(copyEndTime - copyStartTime).count();
 	glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
 	glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);

@@ -243,8 +300,26 @@ void OpenGLRenderPipeline::CacheOutputFrame(const VideoIOOutputFrame& outputFram
 	std::memcpy(mCachedOutputFrame.data(), outputFrame.bytes, byteCount);
 }

-void OpenGLRenderPipeline::ReadOutputFrameSynchronously(const VideoIOState& state, void* destinationBytes)
+bool OpenGLRenderPipeline::TryCopyCachedOutputFrame(VideoIOOutputFrame& outputFrame, OutputReadbackTiming& timing) const
 {
+	if (outputFrame.bytes == nullptr || outputFrame.height == 0 || outputFrame.rowBytes <= 0)
+		return false;
+
+	const std::size_t byteCount = static_cast<std::size_t>(outputFrame.rowBytes) * outputFrame.height;
+	if (mCachedOutputFrame.size() != byteCount)
+		return false;
+
+	const auto copyStartTime = std::chrono::steady_clock::now();
+	std::memcpy(outputFrame.bytes, mCachedOutputFrame.data(), byteCount);
+	const auto copyEndTime = std::chrono::steady_clock::now();
+	timing.cachedCopyMilliseconds += std::chrono::duration_cast<std::chrono::duration<double, std::milli>>(copyEndTime - copyStartTime).count();
+	timing.cachedFallbackUsed = true;
+	return true;
+}
+
+void OpenGLRenderPipeline::ReadOutputFrameSynchronously(const VideoIOState& state, void* destinationBytes, OutputReadbackTiming& timing)
+{
+	const auto readStartTime = std::chrono::steady_clock::now();
 	const bool usePackedOutput = state.outputPixelFormat == VideoIOPixelFormat::V210 || state.outputPixelFormat == VideoIOPixelFormat::Yuva10;

 	glPixelStorei(GL_PACK_ALIGNMENT, 4);
@@ -259,24 +334,78 @@ void OpenGLRenderPipeline::ReadOutputFrameSynchronously(const VideoIOState& stat
 		glBindFramebuffer(GL_READ_FRAMEBUFFER, mRenderer.OutputFramebuffer());
 		glReadPixels(0, 0, state.outputFrameSize.width, state.outputFrameSize.height, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, destinationBytes);
 	}
+	const auto readEndTime = std::chrono::steady_clock::now();
+	timing.syncReadMilliseconds += std::chrono::duration_cast<std::chrono::duration<double, std::milli>>(readEndTime - readStartTime).count();
+	timing.syncFallbackUsed = true;
 }

-void OpenGLRenderPipeline::ReadOutputFrame(const VideoIOState& state, VideoIOOutputFrame& outputFrame)
+OpenGLRenderPipeline::OutputReadbackTiming OpenGLRenderPipeline::ReadOutputFrame(const VideoIOState& state, VideoIOOutputFrame& outputFrame)
 {
-	if (TryConsumeAsyncReadback(outputFrame, 500000))
+	OutputReadbackTiming timing;
+
+	if (mOutputReadbackMode == OutputReadbackMode::Synchronous)
 	{
-		QueueAsyncReadback(state);
-		return;
+		if (outputFrame.bytes != nullptr)
+		{
+			ReadOutputFrameSynchronously(state, outputFrame.bytes, timing);
+			CacheOutputFrame(outputFrame);
+		}
+		return timing;
 	}

-	// If async readback misses the playout deadline, prefer a fresh synchronous
-	// frame over reusing stale cached output, then restart the async pipeline.
-	if (outputFrame.bytes != nullptr)
+	if (mOutputReadbackMode == OutputReadbackMode::CachedOnly)
 	{
-		ReadOutputFrameSynchronously(state, outputFrame.bytes);
+		if (TryCopyCachedOutputFrame(outputFrame, timing))
+			return timing;
+
+		if (outputFrame.bytes != nullptr)
+		{
+			ReadOutputFrameSynchronously(state, outputFrame.bytes, timing);
+			CacheOutputFrame(outputFrame);
+		}
+		return timing;
+	}
+
+	if (TryConsumeAsyncReadback(outputFrame, 500000, timing))
+	{
+		(void)QueueAsyncReadback(state, timing);
+		return timing;
+	}
+
+	if (TryCopyCachedOutputFrame(outputFrame, timing))
+	{
+		(void)QueueAsyncReadback(state, timing);
+		return timing;
+	}
+
+	// Bootstrap only: until the first async readback has produced cached output,
+	// use one synchronous readback so DeckLink has a valid frame to schedule.
+	if (outputFrame.bytes != nullptr && mCachedOutputFrame.empty())
+	{
+		ReadOutputFrameSynchronously(state, outputFrame.bytes, timing);
 		CacheOutputFrame(outputFrame);
 	}

 	FlushAsyncReadbackPipeline();
-	QueueAsyncReadback(state);
+	(void)QueueAsyncReadback(state, timing);
+	return timing;
+}
+
+OpenGLRenderPipeline::OutputReadbackMode OpenGLRenderPipeline::ReadOutputReadbackModeFromEnvironment()
+{
+	char* mode = nullptr;
+	std::size_t modeSize = 0;
+	if (_dupenv_s(&mode, &modeSize, "VST_OUTPUT_READBACK_MODE") != 0 || mode == nullptr)
+		return OutputReadbackMode::AsyncPbo;
+
+	const std::string modeValue(mode);
+	std::free(mode);
+	if (modeValue == "async_pbo")
+		return OutputReadbackMode::AsyncPbo;
+	if (modeValue == "sync")
+		return OutputReadbackMode::Synchronous;
+	if (modeValue == "cached_only")
+		return OutputReadbackMode::CachedOnly;
+
+	return OutputReadbackMode::AsyncPbo;
 }
--- a/apps/LoopThroughWithOpenGLCompositing/gl/pipeline/OpenGLRenderPipeline.h
+++ b/apps/LoopThroughWithOpenGLCompositing/gl/pipeline/OpenGLRenderPipeline.h
@@ -36,6 +36,13 @@ public:
 	bool RenderFrame(const RenderPipelineFrameContext& context, VideoIOOutputFrame& outputFrame);

 private:
+	enum class OutputReadbackMode
+	{
+		AsyncPbo,
+		Synchronous,
+		CachedOnly
+	};
+
 	struct AsyncReadbackSlot
 	{
 		GLuint pixelPackBuffer = 0;
@@ -44,15 +51,34 @@ private:
 		bool inFlight = false;
 	};

+	struct OutputReadbackTiming
+	{
+		double fenceWaitMilliseconds = 0.0;
+		double mapMilliseconds = 0.0;
+		double copyMilliseconds = 0.0;
+		double cachedCopyMilliseconds = 0.0;
+		double asyncQueueMilliseconds = 0.0;
+		double asyncQueueBufferMilliseconds = 0.0;
+		double asyncQueueSetupMilliseconds = 0.0;
+		double asyncQueueReadPixelsMilliseconds = 0.0;
+		double asyncQueueFenceMilliseconds = 0.0;
+		double syncReadMilliseconds = 0.0;
+		bool asyncReadbackMissed = false;
+		bool cachedFallbackUsed = false;
+		bool syncFallbackUsed = false;
+	};
+
 	bool EnsureAsyncReadbackBuffers(std::size_t requiredBytes);
 	void ResetAsyncReadbackState();
 	void FlushAsyncReadbackPipeline();
-	void QueueAsyncReadback(const VideoIOState& state);
-	bool TryConsumeAsyncReadback(VideoIOOutputFrame& outputFrame, GLuint64 timeoutNanoseconds);
+	bool QueueAsyncReadback(const VideoIOState& state, OutputReadbackTiming& timing);
+	bool TryConsumeAsyncReadback(VideoIOOutputFrame& outputFrame, GLuint64 timeoutNanoseconds, OutputReadbackTiming& timing);
 	void CacheOutputFrame(const VideoIOOutputFrame& outputFrame);
-	void ReadOutputFrameSynchronously(const VideoIOState& state, void* destinationBytes);
+	bool TryCopyCachedOutputFrame(VideoIOOutputFrame& outputFrame, OutputReadbackTiming& timing) const;
+	void ReadOutputFrameSynchronously(const VideoIOState& state, void* destinationBytes, OutputReadbackTiming& timing);
 	void PackOutputFor10Bit(const VideoIOState& state);
-	void ReadOutputFrame(const VideoIOState& state, VideoIOOutputFrame& outputFrame);
+	OutputReadbackTiming ReadOutputFrame(const VideoIOState& state, VideoIOOutputFrame& outputFrame);
+	static OutputReadbackMode ReadOutputReadbackModeFromEnvironment();

 	OpenGLRenderer& mRenderer;
 	RuntimeSnapshotProvider& mRuntimeSnapshotProvider;
@@ -60,6 +86,7 @@ private:
 	RenderEffectCallback mRenderEffect;
 	OutputReadyCallback mOutputReady;
 	PaintCallback mPaint;
+	OutputReadbackMode mOutputReadbackMode = OutputReadbackMode::AsyncPbo;
 	std::array<AsyncReadbackSlot, 3> mAsyncReadbackSlots;
 	std::size_t mAsyncReadbackWriteIndex = 0;
 	std::size_t mAsyncReadbackReadIndex = 0;
--- a/apps/LoopThroughWithOpenGLCompositing/runtime/presentation/RuntimeStatePresenter.cpp
+++ b/apps/LoopThroughWithOpenGLCompositing/runtime/presentation/RuntimeStatePresenter.cpp
@@ -95,6 +95,24 @@ JsonValue RuntimeStatePresenter::BuildRuntimeStateValue(const RuntimeStore& runt
 	outputRender.set("renderMs", JsonValue(telemetrySnapshot.backendPlayout.outputRenderMilliseconds));
 	outputRender.set("smoothedRenderMs", JsonValue(telemetrySnapshot.backendPlayout.smoothedOutputRenderMilliseconds));
 	outputRender.set("maxRenderMs", JsonValue(telemetrySnapshot.backendPlayout.maxOutputRenderMilliseconds));
+	outputRender.set("acquireFrameMs", JsonValue(telemetrySnapshot.backendPlayout.outputFrameAcquireMilliseconds));
+	outputRender.set("renderRequestMs", JsonValue(telemetrySnapshot.backendPlayout.outputFrameRenderRequestMilliseconds));
+	outputRender.set("endAccessMs", JsonValue(telemetrySnapshot.backendPlayout.outputFrameEndAccessMilliseconds));
+	outputRender.set("queueWaitMs", JsonValue(telemetrySnapshot.backendPlayout.outputRenderQueueWaitMilliseconds));
+	outputRender.set("drawMs", JsonValue(telemetrySnapshot.backendPlayout.outputRenderDrawMilliseconds));
+	outputRender.set("fenceWaitMs", JsonValue(telemetrySnapshot.backendPlayout.outputReadbackFenceWaitMilliseconds));
+	outputRender.set("mapMs", JsonValue(telemetrySnapshot.backendPlayout.outputReadbackMapMilliseconds));
+	outputRender.set("readbackCopyMs", JsonValue(telemetrySnapshot.backendPlayout.outputReadbackCopyMilliseconds));
+	outputRender.set("cachedCopyMs", JsonValue(telemetrySnapshot.backendPlayout.outputCachedCopyMilliseconds));
+	outputRender.set("asyncQueueMs", JsonValue(telemetrySnapshot.backendPlayout.outputAsyncQueueMilliseconds));
+	outputRender.set("asyncQueueBufferMs", JsonValue(telemetrySnapshot.backendPlayout.outputAsyncQueueBufferMilliseconds));
+	outputRender.set("asyncQueueSetupMs", JsonValue(telemetrySnapshot.backendPlayout.outputAsyncQueueSetupMilliseconds));
+	outputRender.set("asyncQueueReadPixelsMs", JsonValue(telemetrySnapshot.backendPlayout.outputAsyncQueueReadPixelsMilliseconds));
+	outputRender.set("asyncQueueFenceMs", JsonValue(telemetrySnapshot.backendPlayout.outputAsyncQueueFenceMilliseconds));
+	outputRender.set("syncReadMs", JsonValue(telemetrySnapshot.backendPlayout.outputSyncReadMilliseconds));
+	outputRender.set("asyncReadbackMissCount", JsonValue(static_cast<double>(telemetrySnapshot.backendPlayout.outputAsyncReadbackMissCount)));
+	outputRender.set("cachedFallbackCount", JsonValue(static_cast<double>(telemetrySnapshot.backendPlayout.outputCachedFallbackCount)));
+	outputRender.set("syncFallbackCount", JsonValue(static_cast<double>(telemetrySnapshot.backendPlayout.outputSyncFallbackCount)));

 	JsonValue recovery = JsonValue::MakeObject();
 	recovery.set("completionResult", JsonValue(telemetrySnapshot.backendPlayout.completionResult));
--- a/apps/LoopThroughWithOpenGLCompositing/runtime/telemetry/HealthTelemetry.cpp
+++ b/apps/LoopThroughWithOpenGLCompositing/runtime/telemetry/HealthTelemetry.cpp
@@ -212,6 +212,7 @@ void HealthTelemetry::RecordBackendPlayoutHealth(const std::string& lifecycleSta
 	std::size_t minReadyQueueDepth, std::size_t maxReadyQueueDepth, uint64_t readyQueueZeroDepthCount,
 	uint64_t readyQueuePoppedCount, uint64_t readyQueueDroppedCount, uint64_t readyQueueUnderrunCount,
 	double outputRenderMilliseconds, double smoothedOutputRenderMilliseconds, double maxOutputRenderMilliseconds,
+	double outputFrameAcquireMilliseconds, double outputFrameRenderRequestMilliseconds, double outputFrameEndAccessMilliseconds,
 	uint64_t completedFrameIndex, uint64_t scheduledFrameIndex, uint64_t scheduledLeadFrames,
 	uint64_t measuredLagFrames, uint64_t catchUpFrames, uint64_t lateStreak, uint64_t dropStreak,
 	uint64_t lateFrameCount, uint64_t droppedFrameCount, uint64_t flushedFrameCount,
@@ -232,6 +233,9 @@ void HealthTelemetry::RecordBackendPlayoutHealth(const std::string& lifecycleSta
 	mBackendPlayout.outputRenderMilliseconds = std::max(outputRenderMilliseconds, 0.0);
 	mBackendPlayout.smoothedOutputRenderMilliseconds = std::max(smoothedOutputRenderMilliseconds, 0.0);
 	mBackendPlayout.maxOutputRenderMilliseconds = std::max(maxOutputRenderMilliseconds, 0.0);
+	mBackendPlayout.outputFrameAcquireMilliseconds = std::max(outputFrameAcquireMilliseconds, 0.0);
+	mBackendPlayout.outputFrameRenderRequestMilliseconds = std::max(outputFrameRenderRequestMilliseconds, 0.0);
+	mBackendPlayout.outputFrameEndAccessMilliseconds = std::max(outputFrameEndAccessMilliseconds, 0.0);
 	mBackendPlayout.completedFrameIndex = completedFrameIndex;
 	mBackendPlayout.scheduledFrameIndex = scheduledFrameIndex;
 	mBackendPlayout.scheduledLeadFrames = scheduledLeadFrames;
@@ -251,6 +255,7 @@ bool HealthTelemetry::TryRecordBackendPlayoutHealth(const std::string& lifecycle
 	std::size_t minReadyQueueDepth, std::size_t maxReadyQueueDepth, uint64_t readyQueueZeroDepthCount,
 	uint64_t readyQueuePoppedCount, uint64_t readyQueueDroppedCount, uint64_t readyQueueUnderrunCount,
 	double outputRenderMilliseconds, double smoothedOutputRenderMilliseconds, double maxOutputRenderMilliseconds,
+	double outputFrameAcquireMilliseconds, double outputFrameRenderRequestMilliseconds, double outputFrameEndAccessMilliseconds,
 	uint64_t completedFrameIndex, uint64_t scheduledFrameIndex, uint64_t scheduledLeadFrames,
 	uint64_t measuredLagFrames, uint64_t catchUpFrames, uint64_t lateStreak, uint64_t dropStreak,
 	uint64_t lateFrameCount, uint64_t droppedFrameCount, uint64_t flushedFrameCount,
@@ -274,6 +279,9 @@ bool HealthTelemetry::TryRecordBackendPlayoutHealth(const std::string& lifecycle
 	mBackendPlayout.outputRenderMilliseconds = std::max(outputRenderMilliseconds, 0.0);
 	mBackendPlayout.smoothedOutputRenderMilliseconds = std::max(smoothedOutputRenderMilliseconds, 0.0);
 	mBackendPlayout.maxOutputRenderMilliseconds = std::max(maxOutputRenderMilliseconds, 0.0);
+	mBackendPlayout.outputFrameAcquireMilliseconds = std::max(outputFrameAcquireMilliseconds, 0.0);
+	mBackendPlayout.outputFrameRenderRequestMilliseconds = std::max(outputFrameRenderRequestMilliseconds, 0.0);
+	mBackendPlayout.outputFrameEndAccessMilliseconds = std::max(outputFrameEndAccessMilliseconds, 0.0);
 	mBackendPlayout.completedFrameIndex = completedFrameIndex;
 	mBackendPlayout.scheduledFrameIndex = scheduledFrameIndex;
 	mBackendPlayout.scheduledLeadFrames = scheduledLeadFrames;
@@ -289,6 +297,98 @@ bool HealthTelemetry::TryRecordBackendPlayoutHealth(const std::string& lifecycle
 	return true;
 }

+void HealthTelemetry::RecordOutputRenderQueueWait(double queueWaitMilliseconds)
+{
+	std::lock_guard<std::mutex> lock(mMutex);
+	mBackendPlayout.outputRenderQueueWaitMilliseconds = std::max(queueWaitMilliseconds, 0.0);
+}
+
+bool HealthTelemetry::TryRecordOutputRenderQueueWait(double queueWaitMilliseconds)
+{
+	std::unique_lock<std::mutex> lock(mMutex, std::try_to_lock);
+	if (!lock.owns_lock())
+		return false;
+
+	mBackendPlayout.outputRenderQueueWaitMilliseconds = std::max(queueWaitMilliseconds, 0.0);
+	return true;
+}
+
+void HealthTelemetry::RecordOutputRenderPipelineTiming(
+	double drawMilliseconds,
+	double fenceWaitMilliseconds,
+	double mapMilliseconds,
+	double readbackCopyMilliseconds,
+	double cachedCopyMilliseconds,
+	double asyncQueueMilliseconds,
+	double asyncQueueBufferMilliseconds,
+	double asyncQueueSetupMilliseconds,
+	double asyncQueueReadPixelsMilliseconds,
+	double asyncQueueFenceMilliseconds,
+	double syncReadMilliseconds,
+	bool asyncReadbackMissed,
+	bool cachedFallbackUsed,
+	bool syncFallbackUsed)
+{
+	std::lock_guard<std::mutex> lock(mMutex);
+	mBackendPlayout.outputRenderDrawMilliseconds = std::max(drawMilliseconds, 0.0);
+	mBackendPlayout.outputReadbackFenceWaitMilliseconds = std::max(fenceWaitMilliseconds, 0.0);
+	mBackendPlayout.outputReadbackMapMilliseconds = std::max(mapMilliseconds, 0.0);
+	mBackendPlayout.outputReadbackCopyMilliseconds = std::max(readbackCopyMilliseconds, 0.0);
+	mBackendPlayout.outputCachedCopyMilliseconds = std::max(cachedCopyMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueMilliseconds = std::max(asyncQueueMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueBufferMilliseconds = std::max(asyncQueueBufferMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueSetupMilliseconds = std::max(asyncQueueSetupMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueReadPixelsMilliseconds = std::max(asyncQueueReadPixelsMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueFenceMilliseconds = std::max(asyncQueueFenceMilliseconds, 0.0);
+	mBackendPlayout.outputSyncReadMilliseconds = std::max(syncReadMilliseconds, 0.0);
+	if (asyncReadbackMissed)
+		++mBackendPlayout.outputAsyncReadbackMissCount;
+	if (cachedFallbackUsed)
+		++mBackendPlayout.outputCachedFallbackCount;
+	if (syncFallbackUsed)
+		++mBackendPlayout.outputSyncFallbackCount;
+}
+
+bool HealthTelemetry::TryRecordOutputRenderPipelineTiming(
+	double drawMilliseconds,
+	double fenceWaitMilliseconds,
+	double mapMilliseconds,
+	double readbackCopyMilliseconds,
+	double cachedCopyMilliseconds,
+	double asyncQueueMilliseconds,
+	double asyncQueueBufferMilliseconds,
+	double asyncQueueSetupMilliseconds,
+	double asyncQueueReadPixelsMilliseconds,
+	double asyncQueueFenceMilliseconds,
+	double syncReadMilliseconds,
+	bool asyncReadbackMissed,
+	bool cachedFallbackUsed,
+	bool syncFallbackUsed)
+{
+	std::unique_lock<std::mutex> lock(mMutex, std::try_to_lock);
+	if (!lock.owns_lock())
+		return false;
+
+	mBackendPlayout.outputRenderDrawMilliseconds = std::max(drawMilliseconds, 0.0);
+	mBackendPlayout.outputReadbackFenceWaitMilliseconds = std::max(fenceWaitMilliseconds, 0.0);
+	mBackendPlayout.outputReadbackMapMilliseconds = std::max(mapMilliseconds, 0.0);
+	mBackendPlayout.outputReadbackCopyMilliseconds = std::max(readbackCopyMilliseconds, 0.0);
+	mBackendPlayout.outputCachedCopyMilliseconds = std::max(cachedCopyMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueMilliseconds = std::max(asyncQueueMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueBufferMilliseconds = std::max(asyncQueueBufferMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueSetupMilliseconds = std::max(asyncQueueSetupMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueReadPixelsMilliseconds = std::max(asyncQueueReadPixelsMilliseconds, 0.0);
+	mBackendPlayout.outputAsyncQueueFenceMilliseconds = std::max(asyncQueueFenceMilliseconds, 0.0);
+	mBackendPlayout.outputSyncReadMilliseconds = std::max(syncReadMilliseconds, 0.0);
+	if (asyncReadbackMissed)
+		++mBackendPlayout.outputAsyncReadbackMissCount;
+	if (cachedFallbackUsed)
+		++mBackendPlayout.outputCachedFallbackCount;
+	if (syncFallbackUsed)
+		++mBackendPlayout.outputSyncFallbackCount;
+	return true;
+}
+
 HealthTelemetry::SignalStatusSnapshot HealthTelemetry::GetSignalStatusSnapshot() const
 {
 	std::lock_guard<std::mutex> lock(mMutex);
--- a/apps/LoopThroughWithOpenGLCompositing/runtime/telemetry/HealthTelemetry.h
+++ b/apps/LoopThroughWithOpenGLCompositing/runtime/telemetry/HealthTelemetry.h
@@ -97,6 +97,24 @@ public:
 		double outputRenderMilliseconds = 0.0;
 		double smoothedOutputRenderMilliseconds = 0.0;
 		double maxOutputRenderMilliseconds = 0.0;
+		double outputFrameAcquireMilliseconds = 0.0;
+		double outputFrameRenderRequestMilliseconds = 0.0;
+		double outputFrameEndAccessMilliseconds = 0.0;
+		double outputRenderQueueWaitMilliseconds = 0.0;
+		double outputRenderDrawMilliseconds = 0.0;
+		double outputReadbackFenceWaitMilliseconds = 0.0;
+		double outputReadbackMapMilliseconds = 0.0;
+		double outputReadbackCopyMilliseconds = 0.0;
+		double outputCachedCopyMilliseconds = 0.0;
+		double outputAsyncQueueMilliseconds = 0.0;
+		double outputAsyncQueueBufferMilliseconds = 0.0;
+		double outputAsyncQueueSetupMilliseconds = 0.0;
+		double outputAsyncQueueReadPixelsMilliseconds = 0.0;
+		double outputAsyncQueueFenceMilliseconds = 0.0;
+		double outputSyncReadMilliseconds = 0.0;
+		uint64_t outputAsyncReadbackMissCount = 0;
+		uint64_t outputCachedFallbackCount = 0;
+		uint64_t outputSyncFallbackCount = 0;
 		uint64_t completedFrameIndex = 0;
 		uint64_t scheduledFrameIndex = 0;
 		uint64_t scheduledLeadFrames = 0;
@@ -161,6 +179,7 @@ public:
 		std::size_t minReadyQueueDepth, std::size_t maxReadyQueueDepth, uint64_t readyQueueZeroDepthCount,
 		uint64_t readyQueuePoppedCount, uint64_t readyQueueDroppedCount, uint64_t readyQueueUnderrunCount,
 		double outputRenderMilliseconds, double smoothedOutputRenderMilliseconds, double maxOutputRenderMilliseconds,
+		double outputFrameAcquireMilliseconds, double outputFrameRenderRequestMilliseconds, double outputFrameEndAccessMilliseconds,
 		uint64_t completedFrameIndex, uint64_t scheduledFrameIndex, uint64_t scheduledLeadFrames,
 		uint64_t measuredLagFrames, uint64_t catchUpFrames, uint64_t lateStreak, uint64_t dropStreak,
 		uint64_t lateFrameCount, uint64_t droppedFrameCount, uint64_t flushedFrameCount,
@@ -170,11 +189,46 @@ public:
 		std::size_t minReadyQueueDepth, std::size_t maxReadyQueueDepth, uint64_t readyQueueZeroDepthCount,
 		uint64_t readyQueuePoppedCount, uint64_t readyQueueDroppedCount, uint64_t readyQueueUnderrunCount,
 		double outputRenderMilliseconds, double smoothedOutputRenderMilliseconds, double maxOutputRenderMilliseconds,
+		double outputFrameAcquireMilliseconds, double outputFrameRenderRequestMilliseconds, double outputFrameEndAccessMilliseconds,
 		uint64_t completedFrameIndex, uint64_t scheduledFrameIndex, uint64_t scheduledLeadFrames,
 		uint64_t measuredLagFrames, uint64_t catchUpFrames, uint64_t lateStreak, uint64_t dropStreak,
 		uint64_t lateFrameCount, uint64_t droppedFrameCount, uint64_t flushedFrameCount,
 		bool degraded, const std::string& statusMessage);

+	void RecordOutputRenderQueueWait(double queueWaitMilliseconds);
+	bool TryRecordOutputRenderQueueWait(double queueWaitMilliseconds);
+
+	void RecordOutputRenderPipelineTiming(
+		double drawMilliseconds,
+		double fenceWaitMilliseconds,
+		double mapMilliseconds,
+		double readbackCopyMilliseconds,
+		double cachedCopyMilliseconds,
+		double asyncQueueMilliseconds,
+		double asyncQueueBufferMilliseconds,
+		double asyncQueueSetupMilliseconds,
+		double asyncQueueReadPixelsMilliseconds,
+		double asyncQueueFenceMilliseconds,
+		double syncReadMilliseconds,
+		bool asyncReadbackMissed,
+		bool cachedFallbackUsed,
+		bool syncFallbackUsed);
+	bool TryRecordOutputRenderPipelineTiming(
+		double drawMilliseconds,
+		double fenceWaitMilliseconds,
+		double mapMilliseconds,
+		double readbackCopyMilliseconds,
+		double cachedCopyMilliseconds,
+		double asyncQueueMilliseconds,
+		double asyncQueueBufferMilliseconds,
+		double asyncQueueSetupMilliseconds,
+		double asyncQueueReadPixelsMilliseconds,
+		double asyncQueueFenceMilliseconds,
+		double syncReadMilliseconds,
+		bool asyncReadbackMissed,
+		bool cachedFallbackUsed,
+		bool syncFallbackUsed);
+
 	SignalStatusSnapshot GetSignalStatusSnapshot() const;
 	VideoIOStatusSnapshot GetVideoIOStatusSnapshot() const;
 	PerformanceSnapshot GetPerformanceSnapshot() const;
--- a/apps/LoopThroughWithOpenGLCompositing/videoio/VideoBackend.cpp
+++ b/apps/LoopThroughWithOpenGLCompositing/videoio/VideoBackend.cpp
@@ -9,12 +9,14 @@
 #include <algorithm>
 #include <chrono>
 #include <cstring>
+#include <cmath>
 #include <windows.h>

 VideoBackend::VideoBackend(RenderEngine& renderEngine, HealthTelemetry& healthTelemetry, RuntimeEventDispatcher& runtimeEventDispatcher) :
 	mHealthTelemetry(healthTelemetry),
 	mRuntimeEventDispatcher(runtimeEventDispatcher),
 	mPlayoutPolicy(NormalizeVideoPlayoutPolicy(VideoPlayoutPolicy())),
+	mOutputProductionController(mPlayoutPolicy),
 	mReadyOutputQueue(mPlayoutPolicy),
 	mVideoIODevice(std::make_unique<DeckLinkSession>()),
 	mBridge(std::make_unique<OpenGLVideoIOBridge>(renderEngine))
@@ -80,6 +82,7 @@ bool VideoBackend::ConfigureInput(const VideoFormat& inputVideoMode, std::string
 bool VideoBackend::ConfigureOutput(const VideoFormat& outputVideoMode, bool externalKeyingEnabled, std::string& error)
 {
 	mPlayoutPolicy = NormalizeVideoPlayoutPolicy(mPlayoutPolicy);
+	mOutputProductionController.Configure(mPlayoutPolicy);
 	mReadyOutputQueue.Configure(mPlayoutPolicy);
 	if (mLifecycle.State() != VideoBackendLifecycleState::Configuring)
 		ApplyLifecycleTransition(VideoBackendLifecycleState::Configuring, "Configuring video backend output.");
@@ -101,7 +104,10 @@ bool VideoBackend::Start()
 	StartOutputCompletionWorker();
 	const bool started = mVideoIODevice->Start();
 	if (started)
+	{
+		StartOutputProducerWorker();
 		ApplyLifecycleTransition(VideoBackendLifecycleState::Running, "Video backend started.");
+	}
 	else
 	{
 		StopOutputCompletionWorker();
@@ -113,6 +119,7 @@ bool VideoBackend::Start()
 bool VideoBackend::Stop()
 {
 	ApplyLifecycleTransition(VideoBackendLifecycleState::Stopping, "Video backend stopping.");
+	StopOutputProducerWorker();
 	const bool stopped = mVideoIODevice->Stop();
 	StopOutputCompletionWorker();
 	if (stopped)
@@ -232,6 +239,12 @@ const std::string& VideoBackend::StatusMessage() const
 	return mVideoIODevice->StatusMessage();
 }

+bool VideoBackend::ShouldPrioritizeOutputOverPreview() const
+{
+	const RenderOutputQueueMetrics metrics = mReadyOutputQueue.GetMetrics();
+	return metrics.depth < static_cast<std::size_t>(mPlayoutPolicy.targetReadyFrames);
+}
+
 void VideoBackend::SetStatusMessage(const std::string& message)
 {
 	mVideoIODevice->SetStatusMessage(message);
@@ -288,27 +301,36 @@ void VideoBackend::HandleOutputFrameCompletion(const VideoIOCompletion& completi

 void VideoBackend::StartOutputCompletionWorker()
 {
-	std::lock_guard<std::mutex> lock(mOutputCompletionMutex);
-	if (mOutputCompletionWorkerRunning)
-		return;
+	{
+		std::lock_guard<std::mutex> lock(mOutputCompletionMutex);
+		if (mOutputCompletionWorkerRunning)
+			return;

-	mPendingOutputCompletions.clear();
-	mReadyOutputQueue.Clear();
-	mNextReadyOutputFrameIndex = 0;
-	mHasReadyQueueDepthBaseline = false;
-	mMinReadyQueueDepth = 0;
-	mMaxReadyQueueDepth = 0;
-	mReadyQueueZeroDepthCount = 0;
-	mOutputRenderMilliseconds = 0.0;
-	mSmoothedOutputRenderMilliseconds = 0.0;
-	mMaxOutputRenderMilliseconds = 0.0;
-	mOutputCompletionWorkerStopping = false;
-	mOutputCompletionWorkerRunning = true;
-	mOutputCompletionWorker = std::thread(&VideoBackend::OutputCompletionWorkerMain, this);
+		mPendingOutputCompletions.clear();
+		mReadyOutputQueue.Clear();
+		mNextReadyOutputFrameIndex = 0;
+		mHasReadyQueueDepthBaseline = false;
+		mMinReadyQueueDepth = 0;
+		mMaxReadyQueueDepth = 0;
+		mReadyQueueZeroDepthCount = 0;
+		mOutputRenderMilliseconds = 0.0;
+		mSmoothedOutputRenderMilliseconds = 0.0;
+		mMaxOutputRenderMilliseconds = 0.0;
+		mOutputFrameAcquireMilliseconds = 0.0;
+		mOutputFrameRenderRequestMilliseconds = 0.0;
+		mOutputFrameEndAccessMilliseconds = 0.0;
+		mLastLateStreak = 0;
+		mLastDropStreak = 0;
+		mOutputCompletionWorkerStopping = false;
+		mOutputCompletionWorkerRunning = true;
+		mOutputCompletionWorker = std::thread(&VideoBackend::OutputCompletionWorkerMain, this);
+	}
 }

 void VideoBackend::StopOutputCompletionWorker()
 {
+	StopOutputProducerWorker();
+
 	bool shouldJoin = false;
 	{
 		std::lock_guard<std::mutex> lock(mOutputCompletionMutex);
@@ -322,6 +344,40 @@ void VideoBackend::StopOutputCompletionWorker()
 		mOutputCompletionWorker.join();
 }

+void VideoBackend::StartOutputProducerWorker()
+{
+	std::lock_guard<std::mutex> lock(mOutputProducerMutex);
+	if (mOutputProducerWorkerRunning)
+		return;
+
+	mLastOutputProductionCompletion = VideoIOCompletion();
+	mLastOutputProductionTime = std::chrono::steady_clock::time_point();
+	mOutputProducerWorkerStopping = false;
+	mOutputProducerWorkerRunning = true;
+	mOutputProducerWorker = std::thread(&VideoBackend::OutputProducerWorkerMain, this);
+	mOutputProducerCondition.notify_one();
+}
+
+void VideoBackend::StopOutputProducerWorker()
+{
+	bool shouldJoin = false;
+	{
+		std::lock_guard<std::mutex> lock(mOutputProducerMutex);
+		if (mOutputProducerWorkerRunning)
+			mOutputProducerWorkerStopping = true;
+		shouldJoin = mOutputProducerWorker.joinable();
+	}
+	mOutputProducerCondition.notify_one();
+
+	if (shouldJoin)
+		mOutputProducerWorker.join();
+}
+
+void VideoBackend::NotifyOutputProducer()
+{
+	mOutputProducerCondition.notify_one();
+}
+
 void VideoBackend::OutputCompletionWorkerMain()
 {
 	for (;;)
@@ -351,6 +407,57 @@ void VideoBackend::OutputCompletionWorkerMain()
 	}
 }

+void VideoBackend::OutputProducerWorkerMain()
+{
+	for (;;)
+	{
+		{
+			std::unique_lock<std::mutex> lock(mOutputProducerMutex);
+			mOutputProducerCondition.wait_for(lock, OutputProducerWakeInterval());
+
+			if (mOutputProducerWorkerStopping)
+			{
+				mOutputProducerWorkerRunning = false;
+				return;
+			}
+		}
+
+		const RenderOutputQueueMetrics metrics = mReadyOutputQueue.GetMetrics();
+		RecordReadyQueueDepthSample(metrics);
+		const OutputProductionDecision decision = mOutputProductionController.Decide(BuildOutputProductionPressure(metrics));
+		if (decision.action != OutputProductionAction::Produce || decision.requestedFrames == 0)
+			continue;
+
+		VideoIOCompletion completion;
+		{
+			std::lock_guard<std::mutex> lock(mOutputProducerMutex);
+			if (mOutputProducerWorkerStopping)
+				continue;
+			completion = mLastOutputProductionCompletion;
+		}
+
+		const auto now = std::chrono::steady_clock::now();
+		if (mLastOutputProductionTime != std::chrono::steady_clock::time_point() &&
+			now - mLastOutputProductionTime < OutputProducerWakeInterval())
+		{
+			continue;
+		}
+
+		if (ProduceReadyOutputFrames(completion, 1) > 0)
+			mLastOutputProductionTime = std::chrono::steady_clock::now();
+	}
+}
+
+std::chrono::milliseconds VideoBackend::OutputProducerWakeInterval() const
+{
+	const double frameBudgetMilliseconds = State().frameBudgetMilliseconds;
+	if (frameBudgetMilliseconds <= 0.0)
+		return std::chrono::milliseconds(8);
+
+	const int intervalMilliseconds = (std::max)(1, static_cast<int>(std::floor(frameBudgetMilliseconds * 0.75)));
+	return std::chrono::milliseconds(intervalMilliseconds);
+}
+
 void VideoBackend::ProcessOutputFrameCompletion(const VideoIOCompletion& completion)
 {
 	RecordFramePacing(completion.result);
@@ -358,31 +465,69 @@ void VideoBackend::ProcessOutputFrameCompletion(const VideoIOCompletion& complet
 	const RenderOutputQueueMetrics initialQueueMetrics = mReadyOutputQueue.GetMetrics();
 	RecordReadyQueueDepthSample(initialQueueMetrics);
 	const VideoPlayoutRecoveryDecision recoveryDecision = AccountForCompletionResult(completion.result, initialQueueMetrics.depth);
+	{
+		std::lock_guard<std::mutex> lock(mOutputMetricsMutex);
+		mLastLateStreak = recoveryDecision.lateStreak;
+		mLastDropStreak = recoveryDecision.dropStreak;
+	}
+	{
+		std::lock_guard<std::mutex> lock(mOutputProducerMutex);
+		mLastOutputProductionCompletion = completion;
+	}
+	NotifyOutputProducer();

-	FillReadyOutputQueue(completion);
-	if (!ScheduleReadyOutputFrame())
+	if (!ScheduleReadyOutputFrame() &&
+		(ProduceReadyOutputFrames(completion, 1) == 0 || !ScheduleReadyOutputFrame()))
+	{
 		ScheduleBlackUnderrunFrame();
+	}
+	NotifyOutputProducer();
 	RecordBackendPlayoutHealth(completion.result, recoveryDecision);
 }

 void VideoBackend::RecordBackendPlayoutHealth(VideoIOCompletionResult result, const VideoPlayoutRecoveryDecision& recoveryDecision)
 {
 	const RenderOutputQueueMetrics queueMetrics = mReadyOutputQueue.GetMetrics();
+	std::size_t minReadyQueueDepth = 0;
+	std::size_t maxReadyQueueDepth = 0;
+	uint64_t readyQueueZeroDepthCount = 0;
+	double outputRenderMilliseconds = 0.0;
+	double smoothedOutputRenderMilliseconds = 0.0;
+	double maxOutputRenderMilliseconds = 0.0;
+	double outputFrameAcquireMilliseconds = 0.0;
+	double outputFrameRenderRequestMilliseconds = 0.0;
+	double outputFrameEndAccessMilliseconds = 0.0;
+	{
+		std::lock_guard<std::mutex> lock(mOutputMetricsMutex);
+		minReadyQueueDepth = mMinReadyQueueDepth;
+		maxReadyQueueDepth = mMaxReadyQueueDepth;
+		readyQueueZeroDepthCount = mReadyQueueZeroDepthCount;
+		outputRenderMilliseconds = mOutputRenderMilliseconds;
+		smoothedOutputRenderMilliseconds = mSmoothedOutputRenderMilliseconds;
+		maxOutputRenderMilliseconds = mMaxOutputRenderMilliseconds;
+		outputFrameAcquireMilliseconds = mOutputFrameAcquireMilliseconds;
+		outputFrameRenderRequestMilliseconds = mOutputFrameRenderRequestMilliseconds;
+		outputFrameEndAccessMilliseconds = mOutputFrameEndAccessMilliseconds;
+	}
+
 	mHealthTelemetry.TryRecordBackendPlayoutHealth(
 		VideoBackendLifecycle::StateName(mLifecycle.State()),
 		CompletionResultName(result),
 		queueMetrics.depth,
 		queueMetrics.capacity,
 		queueMetrics.pushedCount,
-		mMinReadyQueueDepth,
-		mMaxReadyQueueDepth,
-		mReadyQueueZeroDepthCount,
+		minReadyQueueDepth,
+		maxReadyQueueDepth,
+		readyQueueZeroDepthCount,
 		queueMetrics.poppedCount,
 		queueMetrics.droppedCount,
 		queueMetrics.underrunCount,
-		mOutputRenderMilliseconds,
-		mSmoothedOutputRenderMilliseconds,
-		mMaxOutputRenderMilliseconds,
+		outputRenderMilliseconds,
+		smoothedOutputRenderMilliseconds,
+		maxOutputRenderMilliseconds,
+		outputFrameAcquireMilliseconds,
+		outputFrameRenderRequestMilliseconds,
+		outputFrameEndAccessMilliseconds,
 		recoveryDecision.completedFrameIndex,
 		recoveryDecision.scheduledFrameIndex,
 		recoveryDecision.scheduledLeadFrames,
@@ -397,50 +542,93 @@ void VideoBackend::RecordBackendPlayoutHealth(VideoIOCompletionResult result, co
 		StatusMessage());
 }

-bool VideoBackend::FillReadyOutputQueue(const VideoIOCompletion& completion)
+std::size_t VideoBackend::ProduceReadyOutputFrames(const VideoIOCompletion& completion, std::size_t maxFrames)
 {
+	if (maxFrames == 0)
+		return 0;
+
+	std::lock_guard<std::mutex> productionLock(mOutputProductionMutex);
 	RenderOutputQueueMetrics metrics = mReadyOutputQueue.GetMetrics();
-	bool filledAny = false;
-	while (metrics.depth < mPlayoutPolicy.targetReadyFrames)
+	std::size_t producedFrames = 0;
+	while (producedFrames < maxFrames)
 	{
+		const OutputProductionDecision decision = mOutputProductionController.Decide(BuildOutputProductionPressure(metrics));
+		if (decision.action != OutputProductionAction::Produce)
+			break;
+
 		if (!RenderReadyOutputFrame(mVideoIODevice->State(), completion))
-			return filledAny;
-		filledAny = true;
+			break;
+		++producedFrames;
 		metrics = mReadyOutputQueue.GetMetrics();
 		RecordReadyQueueDepthSample(metrics);
 	}
-	return true;
+	return producedFrames;
+}
+
+OutputProductionPressure VideoBackend::BuildOutputProductionPressure(const RenderOutputQueueMetrics& metrics) const
+{
+	OutputProductionPressure pressure;
+	pressure.readyQueueDepth = metrics.depth;
+	pressure.readyQueueCapacity = metrics.capacity;
+	pressure.readyQueueUnderrunCount = metrics.underrunCount;
+	{
+		std::lock_guard<std::mutex> lock(mOutputMetricsMutex);
+		pressure.lateStreak = mLastLateStreak;
+		pressure.dropStreak = mLastDropStreak;
+	}
+	return pressure;
 }

 bool VideoBackend::RenderReadyOutputFrame(const VideoIOState& state, const VideoIOCompletion& completion)
 {
 	const auto renderStart = std::chrono::steady_clock::now();
 	VideoIOOutputFrame outputFrame;
+	const auto acquireStart = std::chrono::steady_clock::now();
 	if (!BeginOutputFrame(outputFrame))
 		return false;
+	const auto acquireEnd = std::chrono::steady_clock::now();

 	bool rendered = true;
+	const auto renderRequestStart = std::chrono::steady_clock::now();
 	if (mBridge)
 		rendered = mBridge->RenderScheduledFrame(state, completion, outputFrame);
+	const auto renderRequestEnd = std::chrono::steady_clock::now();

+	const auto endAccessStart = std::chrono::steady_clock::now();
 	EndOutputFrame(outputFrame);
+	const auto endAccessEnd = std::chrono::steady_clock::now();
+	const double acquireMilliseconds = std::chrono::duration_cast<std::chrono::duration<double, std::milli>>(acquireEnd - acquireStart).count();
+	const double renderRequestMilliseconds = std::chrono::duration_cast<std::chrono::duration<double, std::milli>>(renderRequestEnd - renderRequestStart).count();
+	const double endAccessMilliseconds = std::chrono::duration_cast<std::chrono::duration<double, std::milli>>(endAccessEnd - endAccessStart).count();
+
 	if (!rendered)
 	{
 		ApplyLifecycleTransition(VideoBackendLifecycleState::Degraded, "Output frame render request failed; skipping schedule for this frame.");
 		const double renderMilliseconds = std::chrono::duration_cast<std::chrono::duration<double, std::milli>>(
 			std::chrono::steady_clock::now() - renderStart).count();
-		RecordOutputRenderDuration(renderMilliseconds);
+		RecordOutputRenderDuration(renderMilliseconds, acquireMilliseconds, renderRequestMilliseconds, endAccessMilliseconds);
+		if (outputFrame.nativeFrame != nullptr)
+		{
+			static_cast<IUnknown*>(outputFrame.nativeFrame)->Release();
+			outputFrame.nativeFrame = nullptr;
+		}
 		return false;
 	}

 	const double renderMilliseconds = std::chrono::duration_cast<std::chrono::duration<double, std::milli>>(
 		std::chrono::steady_clock::now() - renderStart).count();
-	RecordOutputRenderDuration(renderMilliseconds);
+	RecordOutputRenderDuration(renderMilliseconds, acquireMilliseconds, renderRequestMilliseconds, endAccessMilliseconds);

 	RenderOutputFrame readyFrame;
 	readyFrame.frame = outputFrame;
 	readyFrame.frameIndex = ++mNextReadyOutputFrameIndex;
-	return mReadyOutputQueue.Push(readyFrame);
+	const bool pushed = mReadyOutputQueue.Push(readyFrame);
+	if (!pushed && outputFrame.nativeFrame != nullptr)
+	{
+		static_cast<IUnknown*>(outputFrame.nativeFrame)->Release();
+		outputFrame.nativeFrame = nullptr;
+	}
+	return pushed;
 }

 bool VideoBackend::ScheduleReadyOutputFrame()
@@ -516,6 +704,7 @@ void VideoBackend::RecordFramePacing(VideoIOCompletionResult completionResult)

 void VideoBackend::RecordReadyQueueDepthSample(const RenderOutputQueueMetrics& metrics)
 {
+	std::lock_guard<std::mutex> lock(mOutputMetricsMutex);
 	if (!mHasReadyQueueDepthBaseline)
 	{
 		mHasReadyQueueDepthBaseline = true;
@@ -532,14 +721,18 @@ void VideoBackend::RecordReadyQueueDepthSample(const RenderOutputQueueMetrics& m
 		++mReadyQueueZeroDepthCount;
 }

-void VideoBackend::RecordOutputRenderDuration(double renderMilliseconds)
+void VideoBackend::RecordOutputRenderDuration(double renderMilliseconds, double acquireMilliseconds, double renderRequestMilliseconds, double endAccessMilliseconds)
 {
+	std::lock_guard<std::mutex> lock(mOutputMetricsMutex);
 	mOutputRenderMilliseconds = (std::max)(renderMilliseconds, 0.0);
 	if (mSmoothedOutputRenderMilliseconds <= 0.0)
 		mSmoothedOutputRenderMilliseconds = mOutputRenderMilliseconds;
 	else
 		mSmoothedOutputRenderMilliseconds = mSmoothedOutputRenderMilliseconds * 0.9 + mOutputRenderMilliseconds * 0.1;
 	mMaxOutputRenderMilliseconds = (std::max)(mMaxOutputRenderMilliseconds, mOutputRenderMilliseconds);
+	mOutputFrameAcquireMilliseconds = (std::max)(acquireMilliseconds, 0.0);
+	mOutputFrameRenderRequestMilliseconds = (std::max)(renderRequestMilliseconds, 0.0);
+	mOutputFrameEndAccessMilliseconds = (std::max)(endAccessMilliseconds, 0.0);

 	PublishTimingSample("VideoBackend", "outputRender", mOutputRenderMilliseconds, "ms");
 	PublishTimingSample("VideoBackend", "smoothedOutputRender", mSmoothedOutputRenderMilliseconds, "ms");
--- a/apps/LoopThroughWithOpenGLCompositing/videoio/VideoBackend.h
+++ b/apps/LoopThroughWithOpenGLCompositing/videoio/VideoBackend.h
@@ -1,5 +1,6 @@
 #pragma once

+#include "OutputProductionController.h"
 #include "RenderOutputQueue.h"
 #include "VideoBackendLifecycle.h"
 #include "VideoIOTypes.h"
@@ -59,6 +60,7 @@ public:
 	bool KeyerInterfaceAvailable() const;
 	bool ExternalKeyingActive() const;
 	const std::string& StatusMessage() const;
+	bool ShouldPrioritizeOutputOverPreview() const;
 	void SetStatusMessage(const std::string& message);
 	void PublishStatus(bool externalKeyingConfigured, const std::string& statusMessage = std::string());
 	void ReportNoInputDeviceSignalStatus();
@@ -69,14 +71,20 @@ private:
 	void StartOutputCompletionWorker();
 	void StopOutputCompletionWorker();
 	void OutputCompletionWorkerMain();
+	void StartOutputProducerWorker();
+	void StopOutputProducerWorker();
+	void OutputProducerWorkerMain();
+	void NotifyOutputProducer();
+	std::chrono::milliseconds OutputProducerWakeInterval() const;
 	void ProcessOutputFrameCompletion(const VideoIOCompletion& completion);
-	bool FillReadyOutputQueue(const VideoIOCompletion& completion);
+	std::size_t ProduceReadyOutputFrames(const VideoIOCompletion& completion, std::size_t maxFrames);
+	OutputProductionPressure BuildOutputProductionPressure(const RenderOutputQueueMetrics& metrics) const;
 	bool RenderReadyOutputFrame(const VideoIOState& state, const VideoIOCompletion& completion);
 	bool ScheduleReadyOutputFrame();
 	bool ScheduleBlackUnderrunFrame();
 	void RecordFramePacing(VideoIOCompletionResult completionResult);
 	void RecordReadyQueueDepthSample(const RenderOutputQueueMetrics& metrics);
-	void RecordOutputRenderDuration(double renderMilliseconds);
+	void RecordOutputRenderDuration(double renderMilliseconds, double acquireMilliseconds, double renderRequestMilliseconds, double endAccessMilliseconds);
 	bool ApplyLifecycleTransition(VideoBackendLifecycleState state, const std::string& message);
 	bool ApplyLifecycleFailure(const std::string& message);
 	void PublishBackendStateChanged(const std::string& state, const std::string& message);
@@ -92,6 +100,7 @@ private:
 	RuntimeEventDispatcher& mRuntimeEventDispatcher;
 	VideoBackendLifecycle mLifecycle;
 	VideoPlayoutPolicy mPlayoutPolicy;
+	OutputProductionController mOutputProductionController;
 	RenderOutputQueue mReadyOutputQueue;
 	std::unique_ptr<VideoIODevice> mVideoIODevice;
 	std::unique_ptr<OpenGLVideoIOBridge> mBridge;
@@ -99,8 +108,17 @@ private:
 	std::condition_variable mOutputCompletionCondition;
 	std::deque<VideoIOCompletion> mPendingOutputCompletions;
 	std::thread mOutputCompletionWorker;
+	std::mutex mOutputProducerMutex;
+	std::condition_variable mOutputProducerCondition;
+	std::thread mOutputProducerWorker;
+	VideoIOCompletion mLastOutputProductionCompletion;
+	std::chrono::steady_clock::time_point mLastOutputProductionTime;
+	std::mutex mOutputProductionMutex;
+	mutable std::mutex mOutputMetricsMutex;
 	bool mOutputCompletionWorkerRunning = false;
 	bool mOutputCompletionWorkerStopping = false;
+	bool mOutputProducerWorkerRunning = false;
+	bool mOutputProducerWorkerStopping = false;
 	uint64_t mNextReadyOutputFrameIndex = 0;
 	uint64_t mInputFrameIndex = 0;
 	uint64_t mOutputFrameScheduleIndex = 0;
@@ -121,6 +139,11 @@ private:
 	double mOutputRenderMilliseconds = 0.0;
 	double mSmoothedOutputRenderMilliseconds = 0.0;
 	double mMaxOutputRenderMilliseconds = 0.0;
+	double mOutputFrameAcquireMilliseconds = 0.0;
+	double mOutputFrameRenderRequestMilliseconds = 0.0;
+	double mOutputFrameEndAccessMilliseconds = 0.0;
+	uint64_t mLastLateStreak = 0;
+	uint64_t mLastDropStreak = 0;
 	uint64_t mLateFrameCount = 0;
 	uint64_t mDroppedFrameCount = 0;
 	uint64_t mFlushedFrameCount = 0;
--- a/apps/LoopThroughWithOpenGLCompositing/videoio/VideoPlayoutPolicy.h
+++ b/apps/LoopThroughWithOpenGLCompositing/videoio/VideoPlayoutPolicy.h
@@ -30,5 +30,8 @@ inline VideoPlayoutPolicy NormalizeVideoPlayoutPolicy(VideoPlayoutPolicy policy)
 		policy.targetReadyFrames = 1;
 	if (policy.maxReadyFrames < policy.targetReadyFrames)
 		policy.maxReadyFrames = policy.targetReadyFrames;
+	const unsigned minimumOutputFramePoolSize = policy.targetPrerollFrames + policy.maxReadyFrames + policy.minimumSpareDeviceFrames;
+	if (policy.outputFramePoolSize < minimumOutputFramePoolSize)
+		policy.outputFramePoolSize = minimumOutputFramePoolSize;
 	return policy;
 }
--- a/apps/LoopThroughWithOpenGLCompositing/videoio/decklink/DeckLinkSession.cpp
+++ b/apps/LoopThroughWithOpenGLCompositing/videoio/decklink/DeckLinkSession.cpp
@@ -423,7 +423,6 @@ bool DeckLinkSession::AcquireNextOutputVideoFrame(CComPtr<IDeckLinkMutableVideoF
 		return false;

 	outputVideoFrame = outputVideoFrameQueue.front();
-	outputVideoFrameQueue.push_back(outputVideoFrame);
 	outputVideoFrameQueue.pop_front();
 	return outputVideoFrame != nullptr;
 }
@@ -448,6 +447,7 @@ bool DeckLinkSession::PopulateOutputFrame(IDeckLinkMutableVideoFrame* outputVide
 	frame.width = mState.outputFrameSize.width;
 	frame.height = mState.outputFrameSize.height;
 	frame.pixelFormat = mState.outputPixelFormat;
+	outputVideoFrame->AddRef();
 	frame.nativeFrame = outputVideoFrame;
 	frame.nativeBuffer = outputVideoFrameBuffer.Detach();
 	return true;
@@ -506,7 +506,10 @@ VideoPlayoutRecoveryDecision DeckLinkSession::AccountForCompletionResult(VideoIO
 bool DeckLinkSession::ScheduleOutputFrame(const VideoIOOutputFrame& frame)
 {
 	IDeckLinkMutableVideoFrame* outputVideoFrame = static_cast<IDeckLinkMutableVideoFrame*>(frame.nativeFrame);
-	return ScheduleFrame(outputVideoFrame);
+	const bool scheduled = ScheduleFrame(outputVideoFrame);
+	if (outputVideoFrame != nullptr)
+		outputVideoFrame->Release();
+	return scheduled;
 }

 bool DeckLinkSession::Start()
@@ -616,8 +619,18 @@ void DeckLinkSession::HandleVideoInputFrame(IDeckLinkVideoInputFrame* inputFrame
 	inputFrameBuffer->EndAccess(bmdBufferAccessRead);
 }

-void DeckLinkSession::HandlePlayoutFrameCompleted(IDeckLinkVideoFrame*, BMDOutputFrameCompletionResult completionResult)
+void DeckLinkSession::HandlePlayoutFrameCompleted(IDeckLinkVideoFrame* completedFrame, BMDOutputFrameCompletionResult completionResult)
 {
+	if (completedFrame != nullptr)
+	{
+		CComPtr<IDeckLinkMutableVideoFrame> reusableFrame;
+		if (completedFrame->QueryInterface(IID_IDeckLinkMutableVideoFrame, reinterpret_cast<void**>(&reusableFrame)) == S_OK &&
+			reusableFrame != nullptr)
+		{
+			outputVideoFrameQueue.push_back(reusableFrame);
+		}
+	}
+
 	if (!mOutputFrameCallback)
 		return;