#include "OpenGLRenderPipeline.h"

#include "HealthTelemetry.h"
#include "OpenGLRenderer.h"
#include "RuntimeSnapshotProvider.h"
#include "VideoIOFormat.h"

#include <cstring>

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

OpenGLRenderPipeline::OpenGLRenderPipeline(
	OpenGLRenderer& renderer,
	RuntimeSnapshotProvider& runtimeSnapshotProvider,
	HealthTelemetry& healthTelemetry,
	RenderEffectCallback renderEffect,
	OutputReadyCallback outputReady,
	PaintCallback paint) :
	mRenderer(renderer),
	mRuntimeSnapshotProvider(runtimeSnapshotProvider),
	mHealthTelemetry(healthTelemetry),
	mRenderEffect(renderEffect),
	mOutputReady(outputReady),
	mPaint(paint),
	mOutputReadbackMode(ReadOutputReadbackModeFromEnvironment())
{
}

OpenGLRenderPipeline::~OpenGLRenderPipeline()
{
	ResetAsyncReadbackState();
}

bool OpenGLRenderPipeline::RenderFrame(const RenderPipelineFrameContext& context, VideoIOOutputFrame& outputFrame)
{
	const VideoIOState& state = context.videoState;

	const auto renderStartTime = std::chrono::steady_clock::now();
	glBindFramebuffer(GL_FRAMEBUFFER, mRenderer.CompositeFramebuffer());
	mRenderEffect();
	glBindFramebuffer(GL_READ_FRAMEBUFFER, mRenderer.CompositeFramebuffer());
	glBindFramebuffer(GL_DRAW_FRAMEBUFFER, mRenderer.OutputFramebuffer());
	glBlitFramebuffer(0, 0, state.inputFrameSize.width, state.inputFrameSize.height, 0, 0, state.outputFrameSize.width, state.outputFrameSize.height, GL_COLOR_BUFFER_BIT, GL_LINEAR);
	glBindFramebuffer(GL_FRAMEBUFFER, mRenderer.OutputFramebuffer());
	if (mOutputReady)
		mOutputReady();
	if (state.outputPixelFormat == VideoIOPixelFormat::V210 || state.outputPixelFormat == VideoIOPixelFormat::Yuva10)
		PackOutputFor10Bit(state);
	glFlush();

	const auto renderEndTime = std::chrono::steady_clock::now();
	const double renderMilliseconds = std::chrono::duration_cast<std::chrono::duration<double, std::milli>>(renderEndTime - renderStartTime).count();
	mHealthTelemetry.TryRecordPerformanceStats(state.frameBudgetMilliseconds, renderMilliseconds);
	mRuntimeSnapshotProvider.AdvanceFrame();

	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;
}

void OpenGLRenderPipeline::PackOutputFor10Bit(const VideoIOState& state)
{
	glBindFramebuffer(GL_FRAMEBUFFER, mRenderer.OutputPackFramebuffer());
	glViewport(0, 0, state.outputPackTextureWidth, state.outputFrameSize.height);
	glDisable(GL_SCISSOR_TEST);
	glDisable(GL_BLEND);
	glDisable(GL_DEPTH_TEST);
	glActiveTexture(GL_TEXTURE0);
	glBindTexture(GL_TEXTURE_2D, mRenderer.OutputTexture());
	glBindVertexArray(mRenderer.FullscreenVertexArray());
	glUseProgram(mRenderer.OutputPackProgram());

	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)
		glUniform1f(activeWordsLocation, static_cast<float>(ActiveV210WordsForWidth(state.outputFrameSize.width)));
	if (packFormatLocation >= 0)
		glUniform1i(packFormatLocation, state.outputPixelFormat == VideoIOPixelFormat::Yuva10 ? 2 : 1);

	glDrawArrays(GL_TRIANGLES, 0, 3);
	glUseProgram(0);
	glBindVertexArray(0);
	glBindTexture(GL_TEXTURE_2D, 0);
}

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;
}

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;
	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);

	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)
	{
		finishTiming();
		return false;
	}

	if (mAsyncReadbackBytes != requiredBytes
		|| mAsyncReadbackFormat != format
		|| mAsyncReadbackType != type
		|| mAsyncReadbackFramebuffer != framebuffer)
	{
		mAsyncReadbackFormat = format;
		mAsyncReadbackType = type;
		mAsyncReadbackFramebuffer = framebuffer;
		if (!EnsureAsyncReadbackBuffers(requiredBytes))
		{
			finishTiming();
			return false;
		}
	}

	AsyncReadbackSlot& slot = mAsyncReadbackSlots[mAsyncReadbackWriteIndex];
	if (slot.inFlight)
	{
		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, OutputReadbackTiming& timing)
{
	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 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);
		slot.inFlight = false;
		mAsyncReadbackReadIndex = (mAsyncReadbackReadIndex + 1) % mAsyncReadbackSlots.size();
		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);

	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);
}

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);
	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, 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, 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;
}

OpenGLRenderPipeline::OutputReadbackTiming OpenGLRenderPipeline::ReadOutputFrame(const VideoIOState& state, VideoIOOutputFrame& outputFrame)
{
	OutputReadbackTiming timing;

	if (mOutputReadbackMode == OutputReadbackMode::Synchronous)
	{
		if (outputFrame.bytes != nullptr)
		{
			ReadOutputFrameSynchronously(state, outputFrame.bytes, timing);
			CacheOutputFrame(outputFrame);
		}
		return timing;
	}

	if (mOutputReadbackMode == OutputReadbackMode::CachedOnly)
	{
		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();
	(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;
}