CPU optimisations

apps/LoopThroughWithOpenGLCompositing/LoopThroughWithOpenGLCompositing.cpp

@@ -531,7 +531,7 @@ LRESULT CALLBACK WndProc(HWND hWnd, UINT message, WPARAM wParam, LPARAM lParam)
 				if (!sInteractiveResize && pOpenGLComposite)
 				{
 					wglMakeCurrent(hDC, hRC);
-					pOpenGLComposite->paintGL();
+					pOpenGLComposite->paintGL(true);
 					wglMakeCurrent( NULL, NULL );
 					RaiseStatusControls(sStatusStrip);
 				}

apps/LoopThroughWithOpenGLCompositing/control/ControlServer.cpp

@@ -501,6 +501,9 @@ bool ControlServer::SendWebSocketText(SOCKET clientSocket, const std::string& pa
 
 void ControlServer::BroadcastStateLocked()
 {
+	if (mClients.empty())
+		return;
+
 	const std::string stateMessage = mCallbacks.getStateJson ? mCallbacks.getStateJson() : "{}";
 	for (auto it = mClients.begin(); it != mClients.end();)
 	{

apps/LoopThroughWithOpenGLCompositing/gl/OpenGLComposite.cpp

@@ -35,7 +35,7 @@ OpenGLComposite::OpenGLComposite(HWND hWnd, HDC hDC, HGLRC hRC) :
 		*mRuntimeHost,
 		[this]() { renderEffect(); },
 		[this]() { ProcessScreenshotRequest(); },
-		[this]() { paintGL(); });
+		[this]() { paintGL(false); });
 	mVideoIOBridge = std::make_unique<OpenGLVideoIOBridge>(
 		*mVideoIO,
 		*mRenderer,
@@ -156,8 +156,26 @@ error:
 	return false;
 }
 
-void OpenGLComposite::paintGL()
+void OpenGLComposite::paintGL(bool force)
 {
+	if (!force)
+	{
+		if (IsIconic(hGLWnd))
+			return;
+
+		const unsigned previewFps = mRuntimeHost ? mRuntimeHost->GetPreviewFps() : 30u;
+		if (previewFps == 0)
+			return;
+
+		const auto now = std::chrono::steady_clock::now();
+		const auto minimumInterval = std::chrono::microseconds(1000000 / (previewFps == 0 ? 1u : previewFps));
+		if (mLastPreviewPresentTime != std::chrono::steady_clock::time_point() &&
+			now - mLastPreviewPresentTime < minimumInterval)
+		{
+			return;
+		}
+	}
+
 	if (!TryEnterCriticalSection(&pMutex))
 	{
 		ValidateRect(hGLWnd, NULL);
@@ -165,6 +183,7 @@ void OpenGLComposite::paintGL()
 	}
 
 	mRenderer->PresentToWindow(hGLDC, mVideoIO->OutputFrameWidth(), mVideoIO->OutputFrameHeight());
+	mLastPreviewPresentTime = std::chrono::steady_clock::now();
 	ValidateRect(hGLWnd, NULL);
 	LeaveCriticalSection(&pMutex);
 }
@@ -314,15 +333,35 @@ void OpenGLComposite::renderEffect()
 	}
 	else if (mRuntimeHost)
 	{
-		if (mRuntimeHost->TryGetLayerRenderStates(mVideoIO->InputFrameWidth(), mVideoIO->InputFrameHeight(), layerStates))
-		{
-			mCachedLayerRenderStates = layerStates;
-		}
-		else
+		const unsigned renderWidth = mVideoIO->InputFrameWidth();
+		const unsigned renderHeight = mVideoIO->InputFrameHeight();
+		const uint64_t renderStateVersion = mRuntimeHost->GetRenderStateVersion();
+		const bool renderStateCacheValid =
+			!mCachedLayerRenderStates.empty() &&
+			mCachedRenderStateVersion == renderStateVersion &&
+			mCachedRenderStateWidth == renderWidth &&
+			mCachedRenderStateHeight == renderHeight;
+
+		if (renderStateCacheValid)
 		{
 			layerStates = mCachedLayerRenderStates;
 			mRuntimeHost->RefreshDynamicRenderStateFields(layerStates);
 		}
+		else
+		{
+			if (mRuntimeHost->TryGetLayerRenderStates(renderWidth, renderHeight, layerStates))
+			{
+				mCachedLayerRenderStates = layerStates;
+				mCachedRenderStateVersion = renderStateVersion;
+				mCachedRenderStateWidth = renderWidth;
+				mCachedRenderStateHeight = renderHeight;
+			}
+			else
+			{
+				layerStates = mCachedLayerRenderStates;
+				mRuntimeHost->RefreshDynamicRenderStateFields(layerStates);
+			}
+		}
 	}
 	const unsigned historyCap = mRuntimeHost ? mRuntimeHost->GetMaxTemporalHistoryFrames() : 0;
 	mRenderPass->Render(

apps/LoopThroughWithOpenGLCompositing/gl/OpenGLComposite.h

@@ -23,6 +23,7 @@
 #include <string>
 #include <vector>
 #include <deque>
+#include <chrono>
 
 class VideoIODevice;
 class OpenGLVideoIOBridge;
@@ -64,7 +65,7 @@ public:
 	std::string GetOscAddress() const;
 
 	void resizeGL(WORD width, WORD height);
-	void paintGL();
+	void paintGL(bool force = false);
 
 private:
 	void resizeWindow(int width, int height);
@@ -87,8 +88,12 @@ private:
 	std::unique_ptr<ShaderBuildQueue>		mShaderBuildQueue;
 	std::unique_ptr<RuntimeServices>		mRuntimeServices;
 	std::vector<RuntimeRenderState>			mCachedLayerRenderStates;
+	uint64_t								mCachedRenderStateVersion = 0;
+	unsigned								mCachedRenderStateWidth = 0;
+	unsigned								mCachedRenderStateHeight = 0;
 	std::atomic<bool>						mUseCommittedLayerStates;
 	std::atomic<bool>						mScreenshotRequested;
+	std::chrono::steady_clock::time_point	mLastPreviewPresentTime;
 
 	bool InitOpenGLState();
 	void renderEffect();

apps/LoopThroughWithOpenGLCompositing/gl/pipeline/OpenGLRenderPass.cpp

@@ -45,7 +45,7 @@ void OpenGLRenderPass::Render(
 	}
 	else
 	{
-		const std::vector<RenderPassDescriptor> passes = BuildLayerPassDescriptors(layerStates, layerPrograms);
+		const std::vector<RenderPassDescriptor>& passes = BuildLayerPassDescriptors(layerStates, layerPrograms);
 		for (const RenderPassDescriptor& pass : passes)
 		{
 			RenderLayerPass(
@@ -71,9 +71,9 @@ void OpenGLRenderPass::RenderDecodePass(unsigned inputFrameWidth, unsigned input
 	glBindVertexArray(mRenderer.FullscreenVertexArray());
 	glUseProgram(mRenderer.DecodeProgram());
 
-	const GLint packedResolutionLocation = glGetUniformLocation(mRenderer.DecodeProgram(), "uPackedVideoResolution");
-	const GLint decodedResolutionLocation = glGetUniformLocation(mRenderer.DecodeProgram(), "uDecodedVideoResolution");
-	const GLint inputPixelFormatLocation = glGetUniformLocation(mRenderer.DecodeProgram(), "uInputPixelFormat");
+	const GLint packedResolutionLocation = mRenderer.DecodePackedResolutionLocation();
+	const GLint decodedResolutionLocation = mRenderer.DecodeDecodedResolutionLocation();
+	const GLint inputPixelFormatLocation = mRenderer.DecodeInputPixelFormatLocation();
 	if (packedResolutionLocation >= 0)
 		glUniform2f(packedResolutionLocation, static_cast<float>(captureTextureWidth), static_cast<float>(inputFrameHeight));
 	if (decodedResolutionLocation >= 0)
@@ -96,7 +96,8 @@ std::vector<RenderPassDescriptor> OpenGLRenderPass::BuildLayerPassDescriptors(
 	// Flatten the layer stack into concrete GL passes. A layer may now contain
 	// several shader passes, but the outer stack still sees one visible output
 	// per layer.
-	std::vector<RenderPassDescriptor> passes;
+	std::vector<RenderPassDescriptor>& passes = mPassScratch;
+	passes.clear();
 	const std::size_t passCount = layerStates.size() < layerPrograms.size() ? layerStates.size() : layerPrograms.size();
 	std::size_t descriptorCount = 0;
 	for (std::size_t index = 0; index < passCount; ++index)

apps/LoopThroughWithOpenGLCompositing/gl/pipeline/OpenGLRenderPass.h

@@ -56,4 +56,5 @@ private:
 
 	OpenGLRenderer& mRenderer;
 	ShaderTextureBindings mTextureBindings;
+	mutable std::vector<RenderPassDescriptor> mPassScratch;
 };

apps/LoopThroughWithOpenGLCompositing/gl/pipeline/OpenGLRenderPipeline.cpp

@@ -4,6 +4,8 @@
 #include "RuntimeHost.h"
 #include "VideoIOFormat.h"
 
+#include <cstring>
+
 #include <chrono>
 #include <gl/gl.h>
 
@@ -21,6 +23,11 @@ OpenGLRenderPipeline::OpenGLRenderPipeline(
 {
 }
 
+OpenGLRenderPipeline::~OpenGLRenderPipeline()
+{
+	ResetAsyncReadbackState();
+}
+
 bool OpenGLRenderPipeline::RenderFrame(const RenderPipelineFrameContext& context, VideoIOOutputFrame& outputFrame)
 {
 	const VideoIOState& state = context.videoState;
@@ -62,9 +69,9 @@ void OpenGLRenderPipeline::PackOutputFor10Bit(const VideoIOState& state)
 	glBindVertexArray(mRenderer.FullscreenVertexArray());
 	glUseProgram(mRenderer.OutputPackProgram());
 
-	const GLint outputResolutionLocation = glGetUniformLocation(mRenderer.OutputPackProgram(), "uOutputVideoResolution");
-	const GLint activeWordsLocation = glGetUniformLocation(mRenderer.OutputPackProgram(), "uActiveV210Words");
-	const GLint packFormatLocation = glGetUniformLocation(mRenderer.OutputPackProgram(), "uOutputPackFormat");
+	const GLint outputResolutionLocation = mRenderer.OutputPackResolutionLocation();
+	const GLint activeWordsLocation = mRenderer.OutputPackActiveWordsLocation();
+	const GLint packFormatLocation = mRenderer.OutputPackFormatLocation();
 	if (outputResolutionLocation >= 0)
 		glUniform2f(outputResolutionLocation, static_cast<float>(state.outputFrameSize.width), static_cast<float>(state.outputFrameSize.height));
 	if (activeWordsLocation >= 0)
@@ -78,18 +85,195 @@ void OpenGLRenderPipeline::PackOutputFor10Bit(const VideoIOState& state)
 	glBindTexture(GL_TEXTURE_2D, 0);
 }
 
-void OpenGLRenderPipeline::ReadOutputFrame(const VideoIOState& state, VideoIOOutputFrame& outputFrame)
+bool OpenGLRenderPipeline::EnsureAsyncReadbackBuffers(std::size_t requiredBytes)
 {
+	if (requiredBytes == 0)
+		return false;
+
+	if (mAsyncReadbackBytes == requiredBytes && mAsyncReadbackSlots[0].pixelPackBuffer != 0)
+		return true;
+
+	ResetAsyncReadbackState();
+	mAsyncReadbackBytes = requiredBytes;
+	for (AsyncReadbackSlot& slot : mAsyncReadbackSlots)
+	{
+		glGenBuffers(1, &slot.pixelPackBuffer);
+		glBindBuffer(GL_PIXEL_PACK_BUFFER, slot.pixelPackBuffer);
+		glBufferData(GL_PIXEL_PACK_BUFFER, static_cast<GLsizeiptr>(requiredBytes), nullptr, GL_STREAM_READ);
+		slot.sizeBytes = requiredBytes;
+		slot.inFlight = false;
+	}
+	glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
+	mAsyncReadbackWriteIndex = 0;
+	mAsyncReadbackReadIndex = 0;
+	return true;
+}
+
+void OpenGLRenderPipeline::ResetAsyncReadbackState()
+{
+	FlushAsyncReadbackPipeline();
+	for (AsyncReadbackSlot& slot : mAsyncReadbackSlots)
+		slot.sizeBytes = 0;
+
+	if (mAsyncReadbackSlots[0].pixelPackBuffer != 0)
+	{
+		for (AsyncReadbackSlot& slot : mAsyncReadbackSlots)
+		{
+			if (slot.pixelPackBuffer != 0)
+			{
+				glDeleteBuffers(1, &slot.pixelPackBuffer);
+				slot.pixelPackBuffer = 0;
+			}
+		}
+	}
+
+	mAsyncReadbackWriteIndex = 0;
+	mAsyncReadbackReadIndex = 0;
+	mAsyncReadbackBytes = 0;
+}
+
+void OpenGLRenderPipeline::FlushAsyncReadbackPipeline()
+{
+	for (AsyncReadbackSlot& slot : mAsyncReadbackSlots)
+	{
+		if (slot.fence != nullptr)
+		{
+			glDeleteSync(slot.fence);
+			slot.fence = nullptr;
+		}
+		slot.inFlight = false;
+	}
+
+	mAsyncReadbackWriteIndex = 0;
+	mAsyncReadbackReadIndex = 0;
+}
+
+void OpenGLRenderPipeline::QueueAsyncReadback(const VideoIOState& state)
+{
+	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;
+	const GLenum type = usePackedOutput ? GL_UNSIGNED_BYTE : GL_UNSIGNED_INT_8_8_8_8_REV;
+	const GLuint framebuffer = usePackedOutput ? mRenderer.OutputPackFramebuffer() : mRenderer.OutputFramebuffer();
+	const GLsizei readWidth = static_cast<GLsizei>(usePackedOutput ? state.outputPackTextureWidth : state.outputFrameSize.width);
+	const GLsizei readHeight = static_cast<GLsizei>(state.outputFrameSize.height);
+
+	if (requiredBytes == 0)
+		return;
+
+	if (mAsyncReadbackBytes != requiredBytes
+		|| mAsyncReadbackFormat != format
+		|| mAsyncReadbackType != type
+		|| mAsyncReadbackFramebuffer != framebuffer)
+	{
+		mAsyncReadbackFormat = format;
+		mAsyncReadbackType = type;
+		mAsyncReadbackFramebuffer = framebuffer;
+		if (!EnsureAsyncReadbackBuffers(requiredBytes))
+			return;
+	}
+
+	AsyncReadbackSlot& slot = mAsyncReadbackSlots[mAsyncReadbackWriteIndex];
+	if (slot.fence != nullptr)
+	{
+		glDeleteSync(slot.fence);
+		slot.fence = nullptr;
+	}
+
 	glPixelStorei(GL_PACK_ALIGNMENT, 4);
 	glPixelStorei(GL_PACK_ROW_LENGTH, 0);
-	if (state.outputPixelFormat == VideoIOPixelFormat::V210 || state.outputPixelFormat == VideoIOPixelFormat::Yuva10)
+	glBindFramebuffer(GL_READ_FRAMEBUFFER, framebuffer);
+	glBindBuffer(GL_PIXEL_PACK_BUFFER, slot.pixelPackBuffer);
+	glBufferData(GL_PIXEL_PACK_BUFFER, static_cast<GLsizeiptr>(requiredBytes), nullptr, GL_STREAM_READ);
+	glReadPixels(0, 0, readWidth, readHeight, format, type, nullptr);
+	slot.fence = glFenceSync(GL_SYNC_GPU_COMMANDS_COMPLETE, 0);
+	slot.inFlight = slot.fence != nullptr;
+	glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
+
+	mAsyncReadbackWriteIndex = (mAsyncReadbackWriteIndex + 1) % mAsyncReadbackSlots.size();
+}
+
+bool OpenGLRenderPipeline::TryConsumeAsyncReadback(VideoIOOutputFrame& outputFrame, GLuint64 timeoutNanoseconds)
+{
+	if (mAsyncReadbackBytes == 0 || outputFrame.bytes == nullptr)
+		return false;
+
+	AsyncReadbackSlot& slot = mAsyncReadbackSlots[mAsyncReadbackReadIndex];
+	if (!slot.inFlight || slot.fence == nullptr || slot.pixelPackBuffer == 0)
+		return false;
+
+	const GLenum waitFlags = timeoutNanoseconds > 0 ? GL_SYNC_FLUSH_COMMANDS_BIT : 0;
+	const GLenum waitResult = glClientWaitSync(slot.fence, waitFlags, timeoutNanoseconds);
+	if (waitResult != GL_ALREADY_SIGNALED && waitResult != GL_CONDITION_SATISFIED)
+		return false;
+
+	glDeleteSync(slot.fence);
+	slot.fence = nullptr;
+
+	glBindBuffer(GL_PIXEL_PACK_BUFFER, slot.pixelPackBuffer);
+	void* mappedBytes = glMapBuffer(GL_PIXEL_PACK_BUFFER, GL_READ_ONLY);
+	if (mappedBytes == nullptr)
+	{
+		glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
+		slot.inFlight = false;
+		mAsyncReadbackReadIndex = (mAsyncReadbackReadIndex + 1) % mAsyncReadbackSlots.size();
+		return false;
+	}
+
+	std::memcpy(outputFrame.bytes, mappedBytes, slot.sizeBytes);
+	glUnmapBuffer(GL_PIXEL_PACK_BUFFER);
+	glBindBuffer(GL_PIXEL_PACK_BUFFER, 0);
+
+	slot.inFlight = false;
+	mAsyncReadbackReadIndex = (mAsyncReadbackReadIndex + 1) % mAsyncReadbackSlots.size();
+	CacheOutputFrame(outputFrame);
+	return true;
+}
+
+void OpenGLRenderPipeline::CacheOutputFrame(const VideoIOOutputFrame& outputFrame)
+{
+	if (outputFrame.bytes == nullptr || outputFrame.height == 0 || outputFrame.rowBytes <= 0)
+		return;
+
+	const std::size_t byteCount = static_cast<std::size_t>(outputFrame.rowBytes) * outputFrame.height;
+	mCachedOutputFrame.resize(byteCount);
+	std::memcpy(mCachedOutputFrame.data(), outputFrame.bytes, byteCount);
+}
+
+void OpenGLRenderPipeline::ReadOutputFrameSynchronously(const VideoIOState& state, void* destinationBytes)
+{
+	const bool usePackedOutput = state.outputPixelFormat == VideoIOPixelFormat::V210 || state.outputPixelFormat == VideoIOPixelFormat::Yuva10;
+
+	glPixelStorei(GL_PACK_ALIGNMENT, 4);
+	glPixelStorei(GL_PACK_ROW_LENGTH, 0);
+	if (usePackedOutput)
 	{
 		glBindFramebuffer(GL_READ_FRAMEBUFFER, mRenderer.OutputPackFramebuffer());
-		glReadPixels(0, 0, state.outputPackTextureWidth, state.outputFrameSize.height, GL_RGBA, GL_UNSIGNED_BYTE, outputFrame.bytes);
+		glReadPixels(0, 0, state.outputPackTextureWidth, state.outputFrameSize.height, GL_RGBA, GL_UNSIGNED_BYTE, destinationBytes);
 	}
 	else
 	{
 		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, outputFrame.bytes);
+		glReadPixels(0, 0, state.outputFrameSize.width, state.outputFrameSize.height, GL_BGRA, GL_UNSIGNED_INT_8_8_8_8_REV, destinationBytes);
 	}
 }
+
+void OpenGLRenderPipeline::ReadOutputFrame(const VideoIOState& state, VideoIOOutputFrame& outputFrame)
+{
+	if (TryConsumeAsyncReadback(outputFrame, 500000))
+	{
+		QueueAsyncReadback(state);
+		return;
+	}
+
+	// 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)
+	{
+		ReadOutputFrameSynchronously(state, outputFrame.bytes);
+		CacheOutputFrame(outputFrame);
+	}
+
+	FlushAsyncReadbackPipeline();
+	QueueAsyncReadback(state);
+}

apps/LoopThroughWithOpenGLCompositing/gl/pipeline/OpenGLRenderPipeline.h

@@ -1,8 +1,11 @@
 #pragma once
 
+#include "GLExtensions.h"
 #include "VideoIOTypes.h"
 
+#include <array>
 #include <functional>
+#include <vector>
 
 class OpenGLRenderer;
 class RuntimeHost;
@@ -26,10 +29,26 @@ public:
 		RenderEffectCallback renderEffect,
 		OutputReadyCallback outputReady,
 		PaintCallback paint);
+	~OpenGLRenderPipeline();
 
 	bool RenderFrame(const RenderPipelineFrameContext& context, VideoIOOutputFrame& outputFrame);
 
 private:
+	struct AsyncReadbackSlot
+	{
+		GLuint pixelPackBuffer = 0;
+		GLsync fence = nullptr;
+		std::size_t sizeBytes = 0;
+		bool inFlight = false;
+	};
+
+	bool EnsureAsyncReadbackBuffers(std::size_t requiredBytes);
+	void ResetAsyncReadbackState();
+	void FlushAsyncReadbackPipeline();
+	void QueueAsyncReadback(const VideoIOState& state);
+	bool TryConsumeAsyncReadback(VideoIOOutputFrame& outputFrame, GLuint64 timeoutNanoseconds);
+	void CacheOutputFrame(const VideoIOOutputFrame& outputFrame);
+	void ReadOutputFrameSynchronously(const VideoIOState& state, void* destinationBytes);
 	void PackOutputFor10Bit(const VideoIOState& state);
 	void ReadOutputFrame(const VideoIOState& state, VideoIOOutputFrame& outputFrame);
 
@@ -38,4 +57,12 @@ private:
 	RenderEffectCallback mRenderEffect;
 	OutputReadyCallback mOutputReady;
 	PaintCallback mPaint;
+	std::array<AsyncReadbackSlot, 3> mAsyncReadbackSlots;
+	std::size_t mAsyncReadbackWriteIndex = 0;
+	std::size_t mAsyncReadbackReadIndex = 0;
+	std::size_t mAsyncReadbackBytes = 0;
+	GLenum mAsyncReadbackFormat = GL_BGRA;
+	GLenum mAsyncReadbackType = GL_UNSIGNED_INT_8_8_8_8_REV;
+	GLuint mAsyncReadbackFramebuffer = 0;
+	std::vector<unsigned char> mCachedOutputFrame;
 };

apps/LoopThroughWithOpenGLCompositing/gl/renderer/GLExtensions.cpp

@@ -62,6 +62,8 @@ PFNGLGENBUFFERSPROC glGenBuffers;
 PFNGLDELETEBUFFERSPROC glDeleteBuffers;
 PFNGLBINDBUFFERPROC glBindBuffer;
 PFNGLBUFFERDATAPROC glBufferData;
+PFNGLMAPBUFFERPROC glMapBuffer;
+PFNGLUNMAPBUFFERPROC glUnmapBuffer;
 PFNGLBUFFERSUBDATAPROC glBufferSubData;
 PFNGLBINDBUFFERBASEPROC glBindBufferBase;
 PFNGLACTIVETEXTUREPROC glActiveTexture;
@@ -131,6 +133,8 @@ bool ResolveGLExtensions()
 	glDeleteBuffers = (PFNGLDELETEBUFFERSPROC) wglGetProcAddress("glDeleteBuffers");
 	glBindBuffer = (PFNGLBINDBUFFERPROC) wglGetProcAddress("glBindBuffer");
 	glBufferData = (PFNGLBUFFERDATAPROC) wglGetProcAddress("glBufferData");
+	glMapBuffer = (PFNGLMAPBUFFERPROC) wglGetProcAddress("glMapBuffer");
+	glUnmapBuffer = (PFNGLUNMAPBUFFERPROC) wglGetProcAddress("glUnmapBuffer");
 	glBufferSubData = (PFNGLBUFFERSUBDATAPROC) wglGetProcAddress("glBufferSubData");
 	glBindBufferBase = (PFNGLBINDBUFFERBASEPROC) wglGetProcAddress("glBindBufferBase");
 	glActiveTexture = (PFNGLACTIVETEXTUREPROC) wglGetProcAddress("glActiveTexture");
@@ -176,6 +180,8 @@ bool ResolveGLExtensions()
 			&& glDeleteBuffers
 			&& glBindBuffer
 			&& glBufferData
+			&& glMapBuffer
+			&& glUnmapBuffer
 			&& glBufferSubData
 			&& glBindBufferBase
 			&& glActiveTexture

apps/LoopThroughWithOpenGLCompositing/gl/renderer/GLExtensions.h

@@ -89,6 +89,11 @@
 #define GL_EXTERNAL_VIRTUAL_MEMORY_BUFFER_AMD	0x9160
 #define GL_SYNC_GPU_COMMANDS_COMPLETE			0x9117
 #define GL_SYNC_FLUSH_COMMANDS_BIT				0x00000001
+#define GL_ALREADY_SIGNALED						0x911A
+#define GL_TIMEOUT_EXPIRED						0x911B
+#define GL_CONDITION_SATISFIED					0x911C
+#define GL_WAIT_FAILED							0x911D
+#define GL_READ_ONLY							0x88B8
 
 typedef struct __GLsync *GLsync;
 typedef unsigned __int64 GLuint64;
@@ -100,6 +105,8 @@ typedef void (APIENTRYP PFNGLBINDBUFFERPROC) (GLenum target, GLuint buffer);
 typedef void (APIENTRYP PFNGLDELETEBUFFERSPROC) (GLsizei n, const GLuint *buffers);
 typedef void (APIENTRYP PFNGLGENBUFFERSPROC) (GLsizei n, GLuint *buffers);
 typedef void (APIENTRYP PFNGLBUFFERDATAPROC) (GLenum target, GLsizeiptr size, const GLvoid *data, GLenum usage);
+typedef GLvoid* (APIENTRYP PFNGLMAPBUFFERPROC) (GLenum target, GLenum access);
+typedef GLboolean (APIENTRYP PFNGLUNMAPBUFFERPROC) (GLenum target);
 typedef void (APIENTRYP PFNGLATTACHSHADERPROC) (GLuint program, GLuint shader);
 typedef void (APIENTRYP PFNGLCOMPILESHADERPROC) (GLuint shader);
 typedef GLuint (APIENTRYP PFNGLCREATEPROGRAMPROC) (void);
@@ -159,6 +166,8 @@ extern PFNGLGENBUFFERSPROC glGenBuffers;
 extern PFNGLDELETEBUFFERSPROC glDeleteBuffers;
 extern PFNGLBINDBUFFERPROC glBindBuffer;
 extern PFNGLBUFFERDATAPROC glBufferData;
+extern PFNGLMAPBUFFERPROC glMapBuffer;
+extern PFNGLUNMAPBUFFERPROC glUnmapBuffer;
 extern PFNGLBUFFERSUBDATAPROC glBufferSubData;
 extern PFNGLBINDBUFFERBASEPROC glBindBufferBase;
 extern PFNGLACTIVETEXTUREPROC glActiveTexture;

apps/LoopThroughWithOpenGLCompositing/gl/renderer/OpenGLRenderer.cpp

@@ -71,6 +71,9 @@ void OpenGLRenderer::SetDecodeShaderProgram(GLuint program, GLuint vertexShader,
 	mDecodeProgram = program;
 	mDecodeVertexShader = vertexShader;
 	mDecodeFragmentShader = fragmentShader;
+	mDecodePackedResolutionLocation = program != 0 ? glGetUniformLocation(program, "uPackedVideoResolution") : -1;
+	mDecodeDecodedResolutionLocation = program != 0 ? glGetUniformLocation(program, "uDecodedVideoResolution") : -1;
+	mDecodeInputPixelFormatLocation = program != 0 ? glGetUniformLocation(program, "uInputPixelFormat") : -1;
 }
 
 void OpenGLRenderer::SetOutputPackShaderProgram(GLuint program, GLuint vertexShader, GLuint fragmentShader)
@@ -78,6 +81,9 @@ void OpenGLRenderer::SetOutputPackShaderProgram(GLuint program, GLuint vertexSha
 	mOutputPackProgram = program;
 	mOutputPackVertexShader = vertexShader;
 	mOutputPackFragmentShader = fragmentShader;
+	mOutputPackResolutionLocation = program != 0 ? glGetUniformLocation(program, "uOutputVideoResolution") : -1;
+	mOutputPackActiveWordsLocation = program != 0 ? glGetUniformLocation(program, "uActiveV210Words") : -1;
+	mOutputPackFormatLocation = program != 0 ? glGetUniformLocation(program, "uOutputPackFormat") : -1;
 }
 
 bool OpenGLRenderer::ReserveTemporaryRenderTargets(std::size_t count, unsigned width, unsigned height, std::string& error)
@@ -217,6 +223,9 @@ void OpenGLRenderer::DestroyDecodeShaderProgram()
 		glDeleteProgram(mDecodeProgram);
 		mDecodeProgram = 0;
 	}
+	mDecodePackedResolutionLocation = -1;
+	mDecodeDecodedResolutionLocation = -1;
+	mDecodeInputPixelFormatLocation = -1;
 
 	if (mDecodeFragmentShader != 0)
 	{
@@ -238,6 +247,9 @@ void OpenGLRenderer::DestroyOutputPackShaderProgram()
 		glDeleteProgram(mOutputPackProgram);
 		mOutputPackProgram = 0;
 	}
+	mOutputPackResolutionLocation = -1;
+	mOutputPackActiveWordsLocation = -1;
+	mOutputPackFormatLocation = -1;
 
 	if (mOutputPackFragmentShader != 0)
 	{

apps/LoopThroughWithOpenGLCompositing/gl/renderer/OpenGLRenderer.h

@@ -70,6 +70,12 @@ public:
 	GLuint GlobalParamsUBO() const { return mGlobalParamsUBO; }
 	GLuint DecodeProgram() const { return mDecodeProgram; }
 	GLuint OutputPackProgram() const { return mOutputPackProgram; }
+	GLint DecodePackedResolutionLocation() const { return mDecodePackedResolutionLocation; }
+	GLint DecodeDecodedResolutionLocation() const { return mDecodeDecodedResolutionLocation; }
+	GLint DecodeInputPixelFormatLocation() const { return mDecodeInputPixelFormatLocation; }
+	GLint OutputPackResolutionLocation() const { return mOutputPackResolutionLocation; }
+	GLint OutputPackActiveWordsLocation() const { return mOutputPackActiveWordsLocation; }
+	GLint OutputPackFormatLocation() const { return mOutputPackFormatLocation; }
 	GLsizeiptr GlobalParamsUBOSize() const { return mGlobalParamsUBOSize; }
 	void SetGlobalParamsUBOSize(GLsizeiptr size) { mGlobalParamsUBOSize = size; }
 	void ReplaceLayerPrograms(std::vector<LayerProgram>& newPrograms) { mLayerPrograms.swap(newPrograms); }
@@ -101,9 +107,15 @@ private:
 	GLuint mDecodeProgram = 0;
 	GLuint mDecodeVertexShader = 0;
 	GLuint mDecodeFragmentShader = 0;
+	GLint mDecodePackedResolutionLocation = -1;
+	GLint mDecodeDecodedResolutionLocation = -1;
+	GLint mDecodeInputPixelFormatLocation = -1;
 	GLuint mOutputPackProgram = 0;
 	GLuint mOutputPackVertexShader = 0;
 	GLuint mOutputPackFragmentShader = 0;
+	GLint mOutputPackResolutionLocation = -1;
+	GLint mOutputPackActiveWordsLocation = -1;
+	GLint mOutputPackFormatLocation = -1;
 	GLsizeiptr mGlobalParamsUBOSize = 0;
 	int mViewWidth = 0;
 	int mViewHeight = 0;

apps/LoopThroughWithOpenGLCompositing/gl/shader/GlobalParamsBuffer.cpp

@@ -12,7 +12,8 @@ GlobalParamsBuffer::GlobalParamsBuffer(OpenGLRenderer& renderer) :
 
 bool GlobalParamsBuffer::Update(const RuntimeRenderState& state, unsigned availableSourceHistoryLength, unsigned availableTemporalHistoryLength)
 {
-	std::vector<unsigned char> buffer;
+	std::vector<unsigned char>& buffer = mScratchBuffer;
+	buffer.clear();
 	buffer.reserve(512);
 
 	AppendStd140Float(buffer, static_cast<float>(state.timeSeconds));

apps/LoopThroughWithOpenGLCompositing/gl/shader/GlobalParamsBuffer.h

@@ -3,6 +3,8 @@
 #include "OpenGLRenderer.h"
 #include "ShaderTypes.h"
 
+#include <vector>
+
 class GlobalParamsBuffer
 {
 public:
@@ -12,4 +14,5 @@ public:
 
 private:
 	OpenGLRenderer& mRenderer;
+	std::vector<unsigned char> mScratchBuffer;
 };

apps/LoopThroughWithOpenGLCompositing/runtime/RuntimeHost.cpp

@@ -841,6 +841,8 @@ bool RuntimeHost::PollFileChanges(bool& registryChanged, bool& reloadRequested,
 		}
 
 		reloadRequested = mReloadRequested;
+		if (registryChanged || reloadRequested)
+			MarkRenderStateDirtyLocked();
 		return true;
 	}
 	catch (const std::exception& exception)
@@ -884,6 +886,7 @@ bool RuntimeHost::AddLayer(const std::string& shaderId, std::string& error)
 	EnsureLayerDefaultsLocked(layer, shaderIt->second);
 	mPersistentState.layers.push_back(layer);
 	mReloadRequested = true;
+	MarkRenderStateDirtyLocked();
 	return SavePersistentState(error);
 }
 
@@ -900,6 +903,7 @@ bool RuntimeHost::RemoveLayer(const std::string& layerId, std::string& error)
 
 	mPersistentState.layers.erase(it);
 	mReloadRequested = true;
+	MarkRenderStateDirtyLocked();
 	return SavePersistentState(error);
 }
 
@@ -921,6 +925,7 @@ bool RuntimeHost::MoveLayer(const std::string& layerId, int direction, std::stri
 
 	std::swap(mPersistentState.layers[index], mPersistentState.layers[newIndex]);
 	mReloadRequested = true;
+	MarkRenderStateDirtyLocked();
 	return SavePersistentState(error);
 }
 
@@ -949,6 +954,7 @@ bool RuntimeHost::MoveLayerToIndex(const std::string& layerId, std::size_t targe
 	mPersistentState.layers.erase(mPersistentState.layers.begin() + static_cast<std::ptrdiff_t>(sourceIndex));
 	mPersistentState.layers.insert(mPersistentState.layers.begin() + static_cast<std::ptrdiff_t>(targetIndex), movedLayer);
 	mReloadRequested = true;
+	MarkRenderStateDirtyLocked();
 	return SavePersistentState(error);
 }
 
@@ -964,6 +970,7 @@ bool RuntimeHost::SetLayerBypass(const std::string& layerId, bool bypassed, std:
 
 	layer->bypass = bypassed;
 	mReloadRequested = true;
+	MarkRenderStateDirtyLocked();
 	return SavePersistentState(error);
 }
 
@@ -988,6 +995,7 @@ bool RuntimeHost::SetLayerShader(const std::string& layerId, const std::string&
 	layer->parameterValues.clear();
 	EnsureLayerDefaultsLocked(*layer, shaderIt->second);
 	mReloadRequested = true;
+	MarkRenderStateDirtyLocked();
 	return SavePersistentState(error);
 }
 
@@ -1024,6 +1032,7 @@ bool RuntimeHost::UpdateLayerParameter(const std::string& layerId, const std::st
 		const double previousCount = value.numberValues.empty() ? 0.0 : value.numberValues[0];
 		const double triggerTime = std::chrono::duration_cast<std::chrono::duration<double>>(std::chrono::steady_clock::now() - mStartTime).count();
 		value.numberValues = { previousCount + 1.0, triggerTime };
+		MarkRenderStateDirtyLocked();
 		return true;
 	}
 
@@ -1032,6 +1041,7 @@ bool RuntimeHost::UpdateLayerParameter(const std::string& layerId, const std::st
 		return false;
 
 	layer->parameterValues[parameterId] = normalized;
+	MarkRenderStateDirtyLocked();
 	return SavePersistentState(error);
 }
 
@@ -1079,6 +1089,7 @@ bool RuntimeHost::UpdateLayerParameterByControlKey(const std::string& layerKey,
 		const double previousCount = value.numberValues.empty() ? 0.0 : value.numberValues[0];
 		const double triggerTime = std::chrono::duration_cast<std::chrono::duration<double>>(std::chrono::steady_clock::now() - mStartTime).count();
 		value.numberValues = { previousCount + 1.0, triggerTime };
+		MarkRenderStateDirtyLocked();
 		return true;
 	}
 
@@ -1087,6 +1098,7 @@ bool RuntimeHost::UpdateLayerParameterByControlKey(const std::string& layerKey,
 		return false;
 
 	matchedLayer->parameterValues[parameterIt->id] = normalized;
+	MarkRenderStateDirtyLocked();
 	return SavePersistentState(error);
 }
 
@@ -1110,6 +1122,7 @@ bool RuntimeHost::ResetLayerParameters(const std::string& layerId, std::string&
 
 	layer->parameterValues.clear();
 	EnsureLayerDefaultsLocked(*layer, shaderIt->second);
+	MarkRenderStateDirtyLocked();
 	return SavePersistentState(error);
 }
 
@@ -1169,6 +1182,7 @@ bool RuntimeHost::LoadStackPreset(const std::string& presetName, std::string& er
 
 	mPersistentState.layers = nextLayers;
 	mReloadRequested = true;
+	MarkRenderStateDirtyLocked();
 	return SavePersistentState(error);
 }
 
@@ -1197,10 +1211,21 @@ bool RuntimeHost::TrySetSignalStatus(bool hasSignal, unsigned width, unsigned he
 
 void RuntimeHost::SetSignalStatusLocked(bool hasSignal, unsigned width, unsigned height, const std::string& modeName)
 {
+	const bool changed = mHasSignal != hasSignal ||
+		mSignalWidth != width ||
+		mSignalHeight != height ||
+		mSignalModeName != modeName;
 	mHasSignal = hasSignal;
 	mSignalWidth = width;
 	mSignalHeight = height;
 	mSignalModeName = modeName;
+	if (changed)
+		MarkRenderStateDirtyLocked();
+}
+
+void RuntimeHost::MarkRenderStateDirtyLocked()
+{
+	mRenderStateVersion.fetch_add(1, std::memory_order_relaxed);
 }
 
 void RuntimeHost::SetDeckLinkOutputStatus(const std::string& modelName, bool supportsInternalKeying, bool supportsExternalKeying,
@@ -1456,6 +1481,11 @@ bool RuntimeHost::LoadConfig(std::string& error)
 		const double configuredValue = maxTemporalHistoryFramesValue->asNumber(static_cast<double>(mConfig.maxTemporalHistoryFrames));
 		mConfig.maxTemporalHistoryFrames = configuredValue <= 0.0 ? 0u : static_cast<unsigned>(configuredValue);
 	}
+	if (const JsonValue* previewFpsValue = configJson.find("previewFps"))
+	{
+		const double configuredValue = previewFpsValue->asNumber(static_cast<double>(mConfig.previewFps));
+		mConfig.previewFps = configuredValue <= 0.0 ? 0u : static_cast<unsigned>(configuredValue);
+	}
 	if (const JsonValue* enableExternalKeyingValue = configJson.find("enableExternalKeying"))
 		mConfig.enableExternalKeying = enableExternalKeyingValue->asBoolean(mConfig.enableExternalKeying);
 	if (const JsonValue* videoFormatValue = configJson.find("videoFormat"))
@@ -1674,6 +1704,8 @@ bool RuntimeHost::ScanShaderPackages(std::string& error)
 			++it;
 	}
 
+	MarkRenderStateDirtyLocked();
+
 	return true;
 }
 
@@ -1840,6 +1872,7 @@ JsonValue RuntimeHost::BuildStateValue() const
 	app.set("oscPort", JsonValue(static_cast<double>(mConfig.oscPort)));
 	app.set("autoReload", JsonValue(mAutoReloadEnabled));
 	app.set("maxTemporalHistoryFrames", JsonValue(static_cast<double>(mConfig.maxTemporalHistoryFrames)));
+	app.set("previewFps", JsonValue(static_cast<double>(mConfig.previewFps)));
 	app.set("enableExternalKeying", JsonValue(mConfig.enableExternalKeying));
 	app.set("inputVideoFormat", JsonValue(mConfig.inputVideoFormat));
 	app.set("inputFrameRate", JsonValue(mConfig.inputFrameRate));

apps/LoopThroughWithOpenGLCompositing/runtime/RuntimeHost.h

@@ -56,6 +56,7 @@ public:
 	bool TryGetLayerRenderStates(unsigned outputWidth, unsigned outputHeight, std::vector<RuntimeRenderState>& states) const;
 	void RefreshDynamicRenderStateFields(std::vector<RuntimeRenderState>& states) const;
 	std::string BuildStateJson() const;
+	uint64_t GetRenderStateVersion() const { return mRenderStateVersion.load(std::memory_order_relaxed); }
 
 	const std::filesystem::path& GetRepoRoot() const { return mRepoRoot; }
 	const std::filesystem::path& GetUiRoot() const { return mUiRoot; }
@@ -64,6 +65,7 @@ public:
 	unsigned short GetServerPort() const { return mServerPort; }
 	unsigned short GetOscPort() const { return mConfig.oscPort; }
 	unsigned GetMaxTemporalHistoryFrames() const { return mConfig.maxTemporalHistoryFrames; }
+	unsigned GetPreviewFps() const { return mConfig.previewFps; }
 	bool ExternalKeyingEnabled() const { return mConfig.enableExternalKeying; }
 	const std::string& GetInputVideoFormat() const { return mConfig.inputVideoFormat; }
 	const std::string& GetInputFrameRate() const { return mConfig.inputFrameRate; }
@@ -80,6 +82,7 @@ private:
 		unsigned short oscPort = 9000;
 		bool autoReload = true;
 		unsigned maxTemporalHistoryFrames = 4;
+		unsigned previewFps = 30;
 		bool enableExternalKeying = false;
 		std::string inputVideoFormat = "1080p";
 		std::string inputFrameRate = "59.94";
@@ -135,6 +138,7 @@ private:
 	const LayerPersistentState* FindLayerById(const std::string& layerId) const;
 	std::string GenerateLayerId();
 	void SetSignalStatusLocked(bool hasSignal, unsigned width, unsigned height, const std::string& modeName);
+	void MarkRenderStateDirtyLocked();
 	void SetPerformanceStatsLocked(double frameBudgetMilliseconds, double renderMilliseconds);
 	void SetFramePacingStatsLocked(double completionIntervalMilliseconds, double smoothedCompletionIntervalMilliseconds,
 		double maxCompletionIntervalMilliseconds, uint64_t lateFrameCount, uint64_t droppedFrameCount, uint64_t flushedFrameCount);
@@ -179,6 +183,7 @@ private:
 	bool mAutoReloadEnabled;
 	std::chrono::steady_clock::time_point mStartTime;
 	std::chrono::steady_clock::time_point mLastScanTime;
-	std::atomic<uint64_t> mFrameCounter;
+	std::atomic<uint64_t> mFrameCounter{ 0 };
+	std::atomic<uint64_t> mRenderStateVersion{ 0 };
 	uint64_t mNextLayerId;
 };