CPU optimisations

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);
+}