BlueRoseNote/03-UnrealEngine/Rendering/RenderingPipeline/向往渲染系列文章阅读笔记/剖析虚幻渲染体系（11）- RDG.md

title, date, excerpt, tags, rating

title	date	excerpt	tags	rating
剖析虚幻渲染体系（11）- RDG	2024-02-04 21:42:54			⭐

前言

原文地址:https://www.cnblogs.com/timlly/p/15217090.html

概念

• FRDGAllocator：简单的C++对象分配器, 用MemStack分配器追踪和销毁物体。
• FComputeShaderUtils
  • Dispatch()：派发ComputeShader到RHI命令列表，携带其参数。
  • DispatchIndirect()：派发非直接的ComputeShader到RHI命令列表,，携带其参数。
  • AddPass()：派发计算着色器到RenderGraphBuilder， 携带其参数。
  • ClearUAV()：清理UAV。
  • AddCopyTexturePass()：增加拷贝纹理Pass。
  • AddCopyToResolveTargetPass()：增加拷贝到解析目标的Pass。
  • AddEnqueueCopyPass()：增加回读纹理的Pass。
  • AddPass()：无参数的Pass增加。
  • AddBeginUAVOverlapPass()/AddEndUAVOverlapPass()： 其它特殊Pass。
• FRDGResource：RDG资源并不是直接用RHI资源，而是包裹了RHI资源引用，然后针对不同类型的资源各自封装，且增加了额外的信息。
  • FRDGUniformBuffer、TRDGUniformBuffer
  • FRDGParentResource：一种由图跟踪分配生命周期的渲染图资源。可能有引用它的子资源(例如视图)
  • FRDGView
  • FRDGBuffer、FRDGBufferSRV、FRDGBufferUAV
  • FRDGViewableResource：一种由图跟踪分配生命周期的RPGResource。可能有引用它的子资源
    • FRDGTexture
  • FRDGView
    • FRDGUnorderedAccessView、FRDGTextureUAV
    • FRDGShaderResourceView、FRDGTextureSRV
• FRDGTextureDesc：创建渲染纹理的描述信息。
• FRDGPooledTexture：贴图池里的贴图资源。
• FRDGPooledBuffer：池化的缓冲区。
• FRHITransition：用于表示RHI中挂起的资源转换的不透明表面材质数据结构。
• FRDGBarrierBatch：RDG屏障批。
  • FRDGBarrierBatchBegin
  • FRDGBarrierBatchEnd
• FRDGPass
  • TRDGLambdaPass
  • FRDGSentinelPass：哨兵Pass，用于开始/结束。
• #FRDGBuilder

RDG基础类型

enum class ERDGBuilderFlags
{
	None = 0,

	/** Allows the builder to parallelize execution of passes. Without this flag, all passes execute on the render thread. */
	AllowParallelExecute = 1 << 0
};

/** Flags to annotate a pass with when calling AddPass. */
enum class ERDGPassFlags : uint16
{
	/** Pass doesn't have any inputs or outputs tracked by the graph. This may only be used by the parameterless AddPass function. */
	None = 0,

	/** Pass uses rasterization on the graphics pipe. */
	Raster = 1 << 0,

	/** Pass uses compute on the graphics pipe. */
	Compute = 1 << 1,

	/** Pass uses compute on the async compute pipe. */
	AsyncCompute = 1 << 2,

	/** Pass uses copy commands on the graphics pipe. */
	Copy = 1 << 3,

	/** Pass (and its producers) will never be culled. Necessary if outputs cannot be tracked by the graph. */
	NeverCull = 1 << 4,

	/** Render pass begin / end is skipped and left to the user. Only valid when combined with 'Raster'. Disables render pass merging for the pass. */
	SkipRenderPass = 1 << 5,

	/** Pass will never have its render pass merged with other passes. */
	NeverMerge = 1 << 6,

	/** Pass will never run off the render thread. */
	NeverParallel = 1 << 7,

	ParallelTranslate = 1 << 8,

	/** Pass uses copy commands but writes to a staging resource. */
	Readback = Copy | NeverCull
};

/** Flags to annotate a render graph buffer. */
enum class ERDGBufferFlags : uint8
{
	None = 0,

	/** Tag the buffer to survive through frame, that is important for multi GPU alternate frame rendering. */
	MultiFrame = 1 << 0,

	/** The buffer is ignored by RDG tracking and will never be transitioned. Use the flag when registering a buffer with no writable GPU flags.
	 *  Write access is not allowed for the duration of the graph. This flag is intended as an optimization to cull out tracking of read-only
	 *  buffers that are used frequently throughout the graph. Note that it's the user's responsibility to ensure the resource is in the correct
	 *  readable state for use with RDG passes, as RDG does not know the exact state of the resource.
	 */
	SkipTracking = 1 << 1,

	/** When set, RDG will perform its first barrier without splitting. Practically, this means the resource is left in its initial state
	 *  until the first pass it's used within the graph. Without this flag, the resource is split-transitioned at the start of the graph.
	 */
	ForceImmediateFirstBarrier = 1 << 2,
};

/** Flags to annotate a render graph texture. */
enum class ERDGTextureFlags : uint8
{
	None = 0,

	/** Tag the texture to survive through frame, that is important for multi GPU alternate frame rendering. */
	MultiFrame = 1 << 0,

	/** The buffer is ignored by RDG tracking and will never be transitioned. Use the flag when registering a buffer with no writable GPU flags.
	 *  Write access is not allowed for the duration of the graph. This flag is intended as an optimization to cull out tracking of read-only
	 *  buffers that are used frequently throughout the graph. Note that it's the user's responsibility to ensure the resource is in the correct
	 *  readable state for use with RDG passes, as RDG does not know the exact state of the resource.
	 */
	SkipTracking = 1 << 1,
	
	/** When set, RDG will perform its first barrier without splitting. Practically, this means the resource is left in its initial state
	 *  until the first pass it's used within the graph. Without this flag, the resource is split-transitioned at the start of the graph.
	 */
	ForceImmediateFirstBarrier = 1 << 2,

	/** Prevents metadata decompression on this texture. */
	MaintainCompression = 1 << 3,
};
ENUM_CLASS_FLAGS(ERDGTextureFlags);

/** Flags to annotate a view with when calling CreateUAV. */
enum class ERDGUnorderedAccessViewFlags : uint8
{
	None = 0,

	// The view will not perform UAV barriers between consecutive usage.
	SkipBarrier = 1 << 0
};
ENUM_CLASS_FLAGS(ERDGUnorderedAccessViewFlags);

/** The set of concrete parent resource types. */
enum class ERDGViewableResourceType : uint8
{
	Texture,
	Buffer,
	MAX
};

/** The set of concrete view types. */
enum class ERDGViewType : uint8
{
	TextureUAV,
	TextureSRV,
	BufferUAV,
	BufferSRV,
	MAX
};

inline ERDGViewableResourceType GetParentType(ERDGViewType ViewType)
{
	switch (ViewType)
	{
	case ERDGViewType::TextureUAV:
	case ERDGViewType::TextureSRV:
		return ERDGViewableResourceType::Texture;
	case ERDGViewType::BufferUAV:
	case ERDGViewType::BufferSRV:
		return ERDGViewableResourceType::Buffer;
	}
	checkNoEntry();
	return ERDGViewableResourceType::MAX;
}

enum class ERDGResourceExtractionFlags : uint8
{
	None = 0,

	// Allows the resource to remain transient. Only use this flag if you intend to register the resource back
	// into the graph and release the reference. This should not be used if the resource is cached for a long
	// period of time.
	AllowTransient = 1,
};

enum class ERDGInitialDataFlags : uint8
{
	/** Specifies the default behavior, which is to make a copy of the initial data for replay when
	 *  the graph is executed. The user does not need to preserve lifetime of the data pointer.
	 */
	None = 0,

	/** Specifies that the user will maintain ownership of the data until the graph is executed. The
	 *  upload pass will only use a reference to store the data. Use caution with this flag since graph
	 *  execution is deferred! Useful to avoid the copy if the initial data lifetime is guaranteed to
	 *  outlive the graph.
	 */
	 NoCopy = 1 << 0
};

enum class ERDGPooledBufferAlignment : uint8
{
	// The buffer size is not aligned.
	None,

	// The buffer size is aligned up to the next page size.
	Page,

	// The buffer size is aligned up to the next power of two.
	PowerOfTwo
};

/** Returns the equivalent parent resource type for a view type. */
inline ERDGViewableResourceType GetViewableResourceType(ERDGViewType ViewType)
{
	switch (ViewType)
	{
	case ERDGViewType::TextureUAV:
	case ERDGViewType::TextureSRV:
		return ERDGViewableResourceType::Texture;
	case ERDGViewType::BufferUAV:
	case ERDGViewType::BufferSRV:
		return ERDGViewableResourceType::Buffer;
	default:
		checkNoEntry();
		return ERDGViewableResourceType::MAX;
	}
}

using ERDGTextureMetaDataAccess = ERHITextureMetaDataAccess;

/** Returns the associated FRHITransitionInfo plane index. */
inline int32 GetResourceTransitionPlaneForMetadataAccess(ERDGTextureMetaDataAccess Metadata)
{
	switch (Metadata)
	{
	case ERDGTextureMetaDataAccess::CompressedSurface:
	case ERDGTextureMetaDataAccess::HTile:
	case ERDGTextureMetaDataAccess::Depth:
		return FRHITransitionInfo::kDepthPlaneSlice;
	case ERDGTextureMetaDataAccess::Stencil:
		return FRHITransitionInfo::kStencilPlaneSlice;
	default:
		return 0;
	}
}

FRDGBuilder

FRDGBuilder是RDG体系的心脏和发动机，也是个大管家，负责收集渲染Pass和参数，编译Pass、数据，处理资源依赖，裁剪和优化各类数据，还有提供执行接口。

重要函数：

• FindExternalTexture()：查找外部纹理, 若找不到返回null.
• RegisterExternalTexture()：注册外部池内RT到RDG, 以便RDG追踪之. 池内RT可能包含两种RHI纹理: MSAA和非MSAA。
• RegisterExternalBuffer()：注册外部缓冲区到RDG, 以便RDG追踪之.
• 资源创建接口：
  • CreateTexture()
  • CreateBuffer()
  • CreateUAV()
  • CreateSRV()
  • CreateUniformBuffer()
• 分配内存, 内存由RDG管理生命周期。
  • Alloc()
  • AllocPOD()
  • AllocObject()
  • AllocParameters()
• AddPass()
  • FRDGPassRef AddPass(FRDGEventName&& Name, const ParameterStructType* ParameterStruct, ERDGPassFlags Flags, ExecuteLambdaType&& ExecuteLambda); ：增加有参数的LambdaPass。
  • FRDGPassRef AddPass(FRDGEventName&& Name, const FShaderParametersMetadata* ParametersMetadata, const void* ParameterStruct, ERDGPassFlags Flags, ExecuteLambdaType&& ExecuteLambda); ：增加带有实时生成结构体的LambdaPass
  • FRDGPassRef AddPass(FRDGEventName&& Name, ERDGPassFlags Flags, ExecuteLambdaType&& ExecuteLambda);：增加没有参数的LambdaPass
• QueueTextureExtraction()：在Builder执行末期, 提取池内纹理到指定的指针. 对于RDG创建的资源, 这将延长GPU资源的生命周期，直到执行，指针被填充. 如果指定，纹理将转换为AccessFinal状态, 否则将转换为kDefaultAccessFinal状态。
• QueueBufferExtraction()：在Builder执行末期, 提取缓冲区到指定的指针。
• PreallocateTexture()/PreallocateBuffer()：预分配资源. 只对RDG创建的资源, 会强制立即分配底层池内资源, 有效地将其推广到外部资源. 这将增加内存压力，但允许使用GetPooled{Texture, Buffer}查询池中的资源. 主要用于增量地将代码移植到RDG.
• GetPooledTexture()/GetPooledBuffer()：立即获取底层资源, 只允许用于注册或预分配的资源。
• SetTextureAccessFinal()/SetBufferAccessFinal()：设置执行之后的状态。
• 变量
  • RDG对象注册表
    • FRDGPassRegistry Passes;
    • FRDGTextureRegistry Textures;
    • FRDGBufferRegistry Buffers;
    • FRDGViewRegistry Views;
    • FRDGUniformBufferRegistry UniformBuffers;

FRDGBuilder::Compile

RDG编译期间的逻辑非常复杂，步骤繁多，先后经历构建生产者和消费者的依赖关系，确定Pass的裁剪等各类标记，调整资源的生命周期，裁剪Pass，处理Pass的资源转换和屏障，处理异步计算Pass的依赖和引用关系，查找并建立分叉和合并Pass节点，合并所有具体相同渲染目标的光栅化Pass等步骤。

RDG开发

注册外部资源

如果我们已有非RDG创建的资源，可以在RDG使用么？答案是可以，通过FRDGBuilder::RegisterExternalXXX接口可以完成将外部资源注册到RDG系统中。下面以注册纹理为例：

// 在RDG外创建RHI资源.
FRHIResourceCreateInfo CreateInfo;
FTexture2DRHIRef MyRHITexture = RHICreateTexture2D(1024, 768, PF_B8G8R8A8, 1, 1, TexCreate_CPUReadback, CreateInfo);

// 将外部创建的RHI资源注册成RDG资源.
FRDGTextureRef MyExternalRDGTexture = GraphBuilder.RegisterExternalTexture(MyRHITexture);

需要注意的是，外部注册的资源，RDG无法控制和管理其生命周期，需要保证RDG使用期间外部资源的生命周期处于正常状态，否则将引发异常甚至程序崩溃。

如果想从RDG资源获取RHI资源的实例，以下代码可达成：

FRHITexture* MyRHITexture = MyRDGTexture.GetRHI();

用图例展示RHI资源和RDG资源之间的转换关系：

• FRHIResource =>FRDGBuilder::RegisterExternalXXX =>FRDGResource
• FRDGResource => FRDGResource::GetRHI => FRHIResource

提取资源

RDG收集Pass之后并非立即执行，而是延迟执行（包括资源被延迟创建或分配），这就导致了一个问题：如果想将渲染后的资源赋值给某个变量，无法使用立即模式，需要适配延迟执行模式。这种适配延迟执行的资源提取是通过以下接口来实现的：

• FRDGBuilder::QueueTextureExtraction
• FRDGBuilder::QueueBufferExtraction

// 创建RDG纹理.
FRDGTextureRef MyRDGTexture;

FRDGTextureDesc MyTextureDesc = FRDGTextureDesc::Create2D(OutputExtent, HistoryPixelFormat, FClearValueBinding::Black, TexCreate_ShaderResource | TexCreate_UAV);

MyRDGTexture = GraphBuilder.CreateTexture(MyTextureDesc, "MyRDGTexture", ERDGTextureFlags::MultiFrame);

// 创建UAV并作为Pass的shader参数.
(......)
PassParameters->MyRDGTextureUAV = GraphBuilder.CreateUAV(MyRDGTexture);
(......)

// 增加Pass, 以便渲染图像到MyRDGTextureUAV.
FComputeShaderUtils::AddPass(GraphBuilder, RDG_EVENT_NAME("MyCustomPass", ...), ComputeShader, PassParameters, FComputeShaderUtils::GetGroupCount(8, 8));

// 入队提取资源.
TRefCountPtr<IPooledRenderTarget>* OutputRT;
GraphBuilder.QueueTextureExtraction(MyRDGTexture, &OutputRT);

// 对提取的OutputRT进行后续操作.
(......)

RDG调试

RDG系统存在一些控制台命令，其名称和描述如下：

控制台变量	描述
r.RDG.AsyncCompute	控制异步计算策略：0-禁用；1-为异步计算Pass启用标记（默认）；2-开启所有使用compute命令列表的计算通道。
r.RDG.Breakpoint	当满足某些条件时，断点到调试器的断点位置。0-禁用，1~4-不同的特殊调试模式。
r.RDG.ClobberResources	在分配时间用指定的清理颜色清除所有渲染目标和纹理/缓冲UAV。用于调试。
r.RDG.CullPasses	RDG是否开启裁剪无用的Pass。0-禁用，1-开启（默认）。
r.RDG.Debug	允许输出在连接和执行过程中发现的效率低下的警告。
r.RDG.Debug.FlushGPU	开启每次Pass执行后刷新指令到GPU。当设置(r.RDG.AsyncCompute=0)时禁用异步计算。
r.RDG.Debug.GraphFilter	将某些调试事件过滤到特定的图中。
r.RDG.Debug.PassFilter	将某些调试事件过滤到特定的Pass。
r.RDG.Debug.ResourceFilter	将某些调试事件过滤到特定的资源。
r.RDG.DumpGraph	将多个可视化日志转储到磁盘。0-禁用，1-显示生产者、消费者Pass依赖，2-显示资源状态和转换，3-显示图形、异步计算的重叠。
r.RDG.ExtendResourceLifetimes	RDG将把资源生命周期扩展到图的全部长度。会增加内存的占用。
r.RDG.ImmediateMode	在创建Pass时执行Pass。当在Pass的Lambda中崩溃时，连接代码的调用堆栈非常有用。
r.RDG.MergeRenderPasses	图形将合并相同的、连续的渲染通道到一个单一的渲染通道。0-禁用，1-开启（默认）。
r.RDG.OverlapUAVs	RDG将在需要时重叠UAV的工作。如果禁用，UAV屏障总是插入。
r.RDG.TransitionLog	输出资源转换到控制台。
r.RDG.VerboseCSVStats	控制RDG的CSV分析统计的详细程度。0-为图形执行生成一个CSV配置文件，1-为图形执行的每个阶段生成一个CSV文件。

除了以上列出的RDG控制台，还有一些命令可以显示RDG系统运行过程中的有用信息。 vis列出所有有效的纹理，输入之后可能显示如下所示的信息：